Introduction

In today's industrial production, effective management of spare parts (MRO - maintenance, repair and operation) is a critical factor in the smooth operation of equipment. However, identifying equivalent spare parts often turns into a complex, time-consuming task. This is caused by the diversity of manufacturers, changing nomenclature, lack of standardization in catalogs and rapid evolution of technologies. The consequences of this condition include excess inventory, shortages of critical components, increased equipment downtime, and increased operating costs.

Artificial intelligence (AI) offers a solution to these problems through automated cross-referencing. This method allows systems to independently identify and match technically identical or functionally equivalent components from different manufacturers, optimizing the procurement and warehouse management process. For the Ukrainian industrial sector striving for modernization and increasing competitiveness, the implementation of AI-oriented cross-reference systems is a strategically important step.

Principle of Operation of AI-oriented Cross-reference

AI-driven cross-referencing is based on a combination of natural language processing (NLP) and machine learning (ML) for deep analysis and benchmarking. This process is not limited to a simple keyword search, but focuses on semantic understanding of the data. Implementation stages include:

1. Data Collection and Integration

   At the first stage, significant volumes of data from various sources are aggregated. These can be internal ERP systems (eg SAP, 1C), maintenance management systems (CMMS), digital and scanned supplier catalogs (PDF), technical drawings, manufacturers' specifications, as well as public databases and standards (ISO, EN, DSTU). The ability of the system to work with data in both structured (databases, CSV) and unstructured (text descriptions, images) form is important.

2. Text Data Processing

   Collected textual data undergo extensive post-processing to remove noise and standardize. This includes tokenization (breaking text into words/phrases), stemming and lemmatization (reducing words to their basic form), removing stop words (commonly used words without significant meaning), and normalizing units of measurement (eg converting "mm" to "millimeters", "bar" to "atmospheres" or "kW" to "kilowatts").

3. Sign Extraction and Vectorization

   Після передобробки текстові описи перетворюються на числові вектори за допомогою сучасних NLP моделей, таких як Word2Vec або BERT. These vectors capture the semantic meaning and context of technical terms, allowing AI to understand not only words, but also their relationships and technical content. For example, the description "6205 single-row radial ball bearing" will have a similar vector to "6205 radial ball bearing". Vectorization allows AI to “read” and “understand” specifications like an engineer.

4. Similarity Assessment and Machine Learning

   The core of the system consists of machine learning algorithms that compare vectorized features to determine the degree of technical similarity between different components. Methods such as cosine similarity or more complex neural networks are used. Система враховує не лише прямі збіги, а й функціональну взаємозамінність, аналізуючи всі доступні параметри: розміри (наприклад, внутрішній діаметр 25 мм, зовнішній 52 мм, ширина 15 мм для підшипника ISO 6205), матеріал, номінальне навантаження (статичне C0, динамічне C), робочу температуру (від -40°C до +120°C), точність виготовлення (наприклад, за ISO 492:2014) та наявність спеціальних сертифікатів (CE, UkrSEPRO).

5. Validation and Ranking

   AI generates a list of potential equivalents, each of which is assigned a credibility score. This list is provided to the engineers for final validation. A feedback mechanism allows the system to learn from human decisions, constantly improving the accuracy and relevance of its recommendations. For example, for a DN50 PN16 ball valve, the AI ​​may suggest several options, but the engineer will confirm that the body material (steel 1.0619 per EN 10213) and the drive type (lever or gear) match.

Data requirements

The effectiveness of an AI cross-reference system directly depends on the quality and volume of input data. Main aspects:

• Data Sources

  A variety of sources are required for training and functioning of the model:

  • **Internal systems:** ERP (purchase history, nomenclature, descriptions), CMMS (data on installed equipment, maintenance schedules).
  • **Vendor Catalogs:** Digital (API, XML, JSON) and paper (scanned, OCR processed) catalogs.
  • **Industry standards:** Normative documents (ISO, EN, DSTU, DSTU).
  • **Open databases:** Publicly available component data.

• Data Quality

  High data quality is critical. This includes:

  • **Continuity:** Uniform terminology, units of measurement (eg always indicate pressure in bar, temperature in degrees Celsius).
  • **Completeness:** Availability of all key technical characteristics (dimensions, materials, operating parameters, certificates).
  • **Accuracy:** Verified and trusted data. Inaccurate or inconsistent data leads to incorrect AI recommendations.

• Scope and Format

  Effective training of AI models requires large amounts of data. For example, a database with less than 10,000 unique spare parts may require additional enrichment with external sources. Data must be available in a variety of formats, from structured tables to unstructured text and images that require optical character recognition (OCR) and information extraction.

Implementation architecture

The implementation of an AI-oriented cross-reference system requires an integrated architecture covering several layers:

• Data Lake (Data Lake)

  A centralized repository for raw and processed data from all sources. Allows data to be saved in raw format, providing flexibility for future analysis and model training.

• Data Ingestion Layer

  ETL (Extract, Transform, Load) or ELT system for extracting data from ERP, CMMS, web catalogs, scanned documents, their transformation (cleaning, normalization) and uploading to the data lake.

• AI/ML Core

  Includes:

  1. **NLP module:** For processing and semantic analysis of text descriptions.
  2. **ML modules:** Machine learning algorithms for similarity detection and recommendation generation.
  3. **Knowledge Graph:** A semantic network that displays relationships between components, materials, manufacturers, and standards (for example, the relationship between "pump casing" and "cast iron EN-GJL-250").

• Integration Layer (Integration Layer)

  API (Application Programming Interfaces) for seamless interaction with the company's existing IT systems: ERP (SAP S/4HANA, Microsoft Dynamics 365, 1C:Enterprise), CMMS (IBM Maximo, SAP PM), electronic procurement systems. Provides automatic data updates and referrals.

• User Interface (User Interface)

  User-friendly web interfaces or integrations into existing dashboards for engineers and procurement professionals. Allow searching, reviewing recommendations, validating suggested equivalents, and providing feedback to continuously train the model.

• Feedback Loop (Feedback Loop)

  A mechanism that allows users to rate the accuracy of AI recommendations. This feedback is used to periodically retrain the model, adapting it to new data and specific business requirements.

Practical Results and Metrics

Implementation of AI-driven cross-referencing demonstrates significant operational efficiency improvements and financial benefits:

Case study: Large Chemical Enterprise, Ukraine

Data analysis of a large chemical enterprise in Ukraine before the introduction of an AI system revealed that about 15% of the existing spare parts were duplicates or functional equivalents stored under different article numbers. In addition, 8% of MRO spending was due to urgent purchases due to the lack of necessary components in stock or the inability to quickly find alternatives.

Implementation and Results:

• **Implementation period:** 9 months for pilot project with further scaling.
• **Reduction of inventory duplication:** By 10%, which resulted in savings of about UAH 2.5 million per year due to the release of capital and reduction of storage costs.
• **Reduced time to search for equivalents:** From 4 hours to 1.5 hours (62.5% reduction) for complex cases, which significantly accelerated decision-making.
• **Delivery time reduction:** For non-standard items, lead times were reduced by 25% due to faster identification of alternative suppliers.
• **Return on investment (ROI):** 18 months.
• **Approximate cost of implementation:** From 800,000 to 1,500,000 UAH, depending on the complexity of integration and the volume of data.

General benefits recorded in various industries:

• **Reduction of unplanned downtime:** From 5% to 15% due to faster recovery of equipment.
• **Optimization of inventory levels:** 10% to 20% by consolidating duplicates and more accurate forecasting.
• **Savings on purchases:** From 3% to 7% due to better use of existing stocks and selection of optimal suppliers.

Limitations and Potential Risks

Despite significant advantages, AI cross-reference systems are not a panacea and have their limitations:

• Quality of Input Data

  "Garbage in, garbage out." Poor-quality, incomplete, or inconsistent data is a major source of error. An initial investment in data cleansing and standardization is imperative.

• Contextual Nuances

  AI may not take into account specific contextual requirements that are not clearly articulated in the data. For example, for critical components operating in aggressive environments, non-obvious material properties or certification subtleties may be important (for example, for explosive zones according to the ATEX Directive or specific requirements of UkrSEPRO), which AI may miss if it has not been explicitly "trained".

• Excessive Addiction

  There is a risk of over-reliance on AI recommendations without proper engineering validation. Human experience and peer review remain critical, especially for safety components.

• Complexity of Integration

  Integration with legacy ERP/CMMS systems can be technically complex and require significant effort.

• Drift Models

  AI models require constant monitoring and periodic refresher training as the range of spare parts, technologies and standards change. Without this, the accuracy of the recommendations may gradually decrease.

Own Development or Ready Solution?

The decision on own development (Build) or purchase of a ready-made solution (Buy) depends on several factors:

• Own Development (Build)

  • **Advantages:** Maximum customization for the company's unique processes, full control over intellectual property.
  • **Disadvantages:** High development costs (from 1.5 million to 5 million+ UAH), long implementation period (12–36 months), need for highly qualified specialists in AI and Data Science, significant costs for support and development.
  • **Who is it for:** Large corporations with unique requirements, significant budgets and established in-house development teams.

• Purchasing a Ready Solution (Buy)

  • **Benefits:** Faster implementation (3-9 months), lower initial investment (often subscription-based), access to proven technology and vendor expertise, continuous updates.
  • **Disadvantages:** Limited customization options, potential vendor dependency, data security issues (especially for cloud solutions).
  • **Who it's good for:** Most industrial enterprises seeking rapid time-to-value with limited internal AI development resources.

First Steps for the Engineering Team

To successfully implement AI-driven cross-referencing, an engineering team can follow the following plan:

1. Phase 1: Assessment and Planning (1-2 months)

   • **Identification:** Identify the most problematic categories of spare parts where cross-referencing is critical.
   • **Data audit:** Conduct a detailed audit of existing data in ERP, CMMS, supplier catalogs, evaluate their quality and completeness.
   • **Determining Success Metrics:** Set clear, measurable goals, such as reducing duplicate inventory by 10% or reducing search time by 50%.

2. Phase 2: Pilot Project (3-6 months)

   • **Segment selection:** Select a small but representative segment of spare parts or equipment to pilot.
   • **Selection of a solution:** Test a ready-made commercial solution or develop a minimal prototype of your own system.
   • **Validation:** Engage maintenance engineers to manually validate AI recommendations, documenting any discrepancies.
   • **Data cleaning:** Based on pilot experience, improve data cleaning and standardization processes.

3. Phase 3: Scalable Deployment

   • **Gradual expansion:** Expand the system to other categories of spare parts and production facilities.
   • **Monitoring:** Implement continuous monitoring of AI model performance and recommendation quality.
   • **Integration:** Integrate AI solutions with other MRO digital transformation initiatives.

Conclusion

AI-oriented cross-reference of spare parts is not just a technological innovation, but a strategic tool for increasing the operational efficiency and competitiveness of industrial enterprises. It transforms inventory management from a reactive to a proactive, intelligent process that significantly reduces costs and minimizes downtime risks. The use of advanced analytical methods allows Ukrainian enterprises to move to a new level of MRO management. UNITEC-D, with its extensive electronic catalog and deep technical expertise, is a reliable partner in this digital transformation.

Find out more about components and solutions for your production in the UNITEC-D E-Catalog.

List of Literature

• ISO 15:2011 "Rolling bearings – Radial bearings – Boundary dimensions, general plan"
• DSTU GOST 24810:2018 "Rolling bearings. Terms and definitions"
• EN 16646:2014 "Maintenance – Maintenance Key Performance Indicators"

1. Introduction: The Strategic Importance of MRO Inventory Management

Effective maintenance, repair and operation (MRO) inventory management is critical for the continuity of production and financial stability of industrial enterprises in Ukraine. In today's industrial environment, where production stoppages due to unavailability of required components can lead to significant losses, optimizing MRO supply chains is turning from a tactical necessity to a strategic advantage. This not only ensures that the right parts are available at the right time, but also minimizes costs associated with excess inventory or emergency purchases.

2. Problem: Quantifying the Cost of Inefficient Spare Parts Management

Inefficient MRO inventory management generates significant, often hidden, costs. Understanding these costs is the first step to optimization. The main categories of expenses include:

2.1. Equipment Downtime Cost

The absence of a critical spare part can stop the production line. Downtime cost is the most obvious and destructive. It includes lost productivity, wages for non-production personnel, lost profits and potential fines for breach of contract. For Ukrainian industrial enterprises, the cost of one hour of downtime can vary from 1,500 to 15,000 euros and higher, depending on the industry and scale of production. For example, for a company with a daily production volume of 50,000 euros, 8 hours of downtime due to the absence of a bearing can cost 40,000 euros in lost profits.

2.2. Overstocking

Excess inventory, while eliminating the risk of downtime, creates other financial burdens. The average annual cost of maintaining stocks (Carrying Costs) for MRO components is 20-30% of their total cost. This includes:

• Cost of storage: Rent of warehouse space, utilities, personnel, equipment.
• Cost of capital: Frozen funds that could be invested in other assets.
• Taxation and Insurance: Costs related to asset ownership.
• Obsolescence and Deterioration: Risk of loss of value due to technological progress or physical damage.

If a company has an MRO inventory worth €1,000,000, the annual cost of maintaining it can reach 200,000 - €300,000.

2.3. Emergency Purchases

When a critical spare part is not available, there is a need for emergency procurement, which usually means higher prices, more expensive logistics (express delivery) and additional administrative costs. The cost of one emergency purchase can be 20-50% higher than the planned one. For a part costing €500, emergency purchasing can add €100-250, not including expedited shipping, which can double the initial cost.

The average percentage of MRO costs from the cost of production equipment varies between 3-5% per year. Optimization can reduce this number, improving overall profitability.

3. Framework of Analysis: Optimization Methodology

A structured approach must be used for effective optimization. Key methodologies include:

3.1. ABC-Analysis

Classification of stocks according to their value and importance. Category A: 20% of the range, which is 80% of the cost (critical, expensive components). Category B: 30% of the range, which is 15% of the cost. Category C: 50% of the range, which is 5% of the cost. Particular attention should be paid to components of category A.

3.2. VED-Analysis (Vital, Essential, Desirable)

Classification by impact on production in case of absence. V (Vital): Vital components without which production stops. E (Essential): Essential components, the absence of which can cause significant failures or slow production. D (Desirable): Desirable components that do not critically affect production.

3.3. Definition of Critical Spare Parts

A combination of ABC and VED analysis enables accurate identification of critical spare parts that require reduced lead times and increased availability. These are components that are expensive (A) and vital (V).

3.4. Demand forecasting

Using historical data, scheduled maintenance schedules, and equipment failure data to accurately forecast parts needs. For this, statistical methods (for example, moving average, exponential smoothing) or more modern machine learning systems that analyze large volumes of data can be used. It is important to take into account seasonality and planned modernizations.

Compliance with the requirements of standards such as DSTU ISO 9001:2015 (Quality Management Systems) and DSTU EN 13306:2016 (Maintenance. Terminology) is the basis for a structured approach to MRO management.

4. Stages of Implementation: A Practical Guide

Effective reduction of lead time for critical MRO components requires a systematic approach:

4.1. Assessment of the Current State

Audit current procurement processes, inventory levels, supplier data and historical lead times. Identify bottlenecks and potential areas for improvement.

4.2. Consolidation of Suppliers

Reducing the number of suppliers can simplify management, increase procurement volumes and improve negotiating positions. This makes it possible to achieve better conditions in terms of price, delivery time and service. UNITEC-D, as an integrated supplier, can combine thousands of MRO items from different manufacturers.

4.3. Service Level Agreements (SLA)

Enter into service level agreements with key suppliers that clearly define maximum lead times, penalties for violations and delivery terms. This ensures predictability and accountability.

4.4. Vendor Managed Inventory (VMI - Vendor Managed Inventory)

Transferring responsibility for the management of certain MRO inventories to a supplier. The supplier monitors the level of stocks in the customer's warehouse and automatically replenishes them. This reduces the burden on internal staff and optimizes inventory levels.

4.5. Buffer Stocks for Critical Elements

For critical, long-term or unique components, build strategic buffer stocks in your own warehouse or in the warehouse of a trusted partner such as UNITEC-D. This allows you to quickly respond to unforeseen failures, for example, bearings with special tolerances ISO 492 or seals that comply with DSTU EN 15848.

4.6. Digitization of Procurement Processes

Implementation of electronic procurement systems (e-procurement) and integration with ERP/CMMS systems. It automates ordering, tracking and paperwork, reducing administrative delays. UNITEC-D E-Catalog, as an example, simplifies searching and ordering.

5. KPIs and Metrics: What to Measure

Measuring key performance indicators (KPIs) is imperative to monitor progress and identify further areas for improvement:

5.1. Inventory Service Level (Service Level)

Percentage of parts requests that were fulfilled without delay from the warehouse. The target for critical components should be 95-99%. For less critical ones - 90-95%.

5.2. Inventory Turns

The number of times MRO inventory has been completely sold or used during a given period. A higher ratio indicates more effective management. For component MRO, the target may be 2-4 revolutions per year. A low turnover (less than 1) indicates excess inventory.

5.3. Order Lead Time

Average time from placing an order to receiving it. The goal is to constantly reduce this indicator, especially for critical positions. For standard components, it can be 3-7 days, for specialized components - 14-30 days.

5.4. Cost of Inventory per Employee

The total cost of MRO inventory divided by the number of production workers. This allows you to compare the efficiency of inventory management with industry benchmarks. The average figure can be 1,000 - 5,000 euros per employee per year.

5.5. Percentage of Emergency Purchases

The ratio of the cost of emergency procurement to the total cost of MRO procurement. The target indicator is to minimize, aiming for less than 5%.

6. Tools and Technologies

Modern technologies play a key role in optimizing MRO supply chains:

• ERP (Enterprise Resource Planning) systems: Integrated systems for managing all business processes, including procurement, inventory, finance.
• CMMS systems (Computerized Maintenance Management System): For maintenance planning and management, including spare parts management.
• E-procurement platforms: Automate the procurement process, interaction with suppliers, contract management.
• Data analytics systems: Using Big Data and AI to forecast demand, optimize inventory levels, identify patterns in equipment failures.
• UNITEC-D E-Catalog: An online platform for fast searching, cross-referencing and ordering of industrial components. A convenient interface and up-to-date information simplify the purchasing process, providing access to a wide range of spare parts that meet CE, UkrSEPRO and other international standards.

7. Common Mistakes: How to Avoid Them

Avoiding typical mistakes allows you to speed up the optimization process:

1. Insufficient Data Quality: Lack of accurate usage, lead time and cost data. Solution: Implementation of strict data collection and validation procedures.
2. Ignore Criticality: All components are treated equally. Solution: Application of ABC/VED analysis for prioritization.
3. Excessive Dependence on One Supplier: High risk of supply disruptions. Solution: Diversification of suppliers, especially for critical positions.
4. Lack of Communication Between Departments: Gap between procurement, production and maintenance departments. Solution: Regular meetings, integrated systems, joint KPIs.
5. Underestimation of Obsolete Components Accounting: Storage of obsolete or unused spare parts. Solution: Regular inventory audits, recycling or sales programs.

8. Checklist of Quick Wins for the Procurement Manager

Ten actions you can implement this week:

1. Inventory the 20% most expensive MRO components (category A).
2. Identify the top 5 causes of parts-related downtime.
3. Contact 3 key vendors to discuss VMI terms.
4. Analyze emergency procurement reports for the past 6 months.
5. Create an internal working group of representatives from purchasing, maintenance and production departments.
6. Check the relevance of order fulfillment time data in your ERP system.
7. Select the 1-2 most critical components to create a minimum buffer.
8. Evaluate the possibilities of using UNITEC-D E-Catalog to simplify ordering.
9. Develop a basic dashboard with KPIs (service level, turnover).
10. Start reviewing old, low-moving inventory to identify obsolete items.

9. Conclusion

Optimizing supply chains and reducing lead times for critical MRO components is an ongoing process that requires a strategic approach, discipline and the use of modern technology. The implementation of the described methodologies allows industrial enterprises of Ukraine to significantly reduce operating costs, increase the reliability of equipment and ensure continuity of production. UNITEC-D offers integrated solutions, including outsourcing and turnkey services, to help achieve these goals.

Find out more about the wide range of industrial components and solutions for optimizing supply chains in the UNITEC-D E-Catalog.

10. Links

• DSTU ISO 9001:2015 Quality management systems. Requirements
• DSTU EN 13306:2016 Maintenance. Terminology.
• DSTU ISO 492:2006 Rolling bearings. Radial clearances.
• DSTU EN 15848:2018 Flanges and their connections. Criteria for choosing gaskets for sealing.
• "MRO Supply Chain Benchmarking Report", Institute of Procurement Management (ISM), 2023.
• "The Cost of Downtime: An Industry Report", Aberdeen Group, 2022.
• "Managing MRO Inventory for Optimal Performance", Supply Chain Management Review, 2024.

1. Introduction: Artificial Intelligence for MRO Problem Solving

In industrial manufacturing, maintenance, repair and operation (MRO) efficiency is critical to the continuity of operations and the minimization of production costs. However, the increasing complexity of the equipment, the variety of technical documentation and the extensiveness of spare parts catalogs create significant difficulties. Traditional methods of searching for information often require a significant amount of time, lead to errors and delays in repairs. With a shortage of skilled personnel, the ability to quickly and accurately find the information you need is critical.

Artificial intelligence (AI) offers solutions to these challenges through the implementation of advanced search algorithms. These systems are capable of processing, indexing, and interpreting vast amounts of unstructured and structured data, including operating instructions, schematics, drawings, service bulletins, material specifications, and spare parts catalogs. The goal is to reduce equipment downtime, optimize inventory, and improve the overall reliability of production assets. UNITEC-D GmbH, specializing in MRO, recognizes the potential of AI in transforming these processes for Ukrainian industry.

2. Principles of Work: Semantic Search and Natural Language Processing

The basis of advanced search algorithms for MRO is a combination of natural language processing (NLP) and semantic search. Instead of simply comparing keywords, as traditional search engines do, AI analyzes the context and meaning of a user's query, as well as the content of documents. This process can be divided into several key stages:

2.1. Collection and Preliminary Processing of Data

The system aggregates data from various sources: PDF files, technical manuals, CAD drawings, spreadsheets, ERP databases. For unstructured data, optical character recognition (OCR) is used to convert images of text into a machine-readable format. Then the text is cleaned of noise, normalized and divided into separate elements (tokens).

2.2. Indexing and Vectorization

Each document and user request is transformed into a numerical vector (embedding) in a multidimensional space. Documents and queries that have similar semantic meaning are located close to each other in this space. This allows the system to find not only exact matches of keywords, but also conceptually related documents. For example, a search for "rolling bearing" might find documents containing "ball bearing" or "roller bearing".

2.3. Mechanisms of Semantic Search

After vectorization, nearest neighbor search algorithms (e.g., Approximate Nearest Neighbors - ANN) are used to quickly identify the most relevant documents. Additionally, models based on transformer architecture (such as BERT or Qwen-8B) can be used to understand complex queries and even answer questions based on retrieved information (Retrieval Augmented Generation - RAG). This greatly increases the accuracy and relevance of the results, allowing engineers to receive direct answers rather than just a list of documents.

3. Data Requirements: Volume, Quality and Format

The effectiveness of AI systems directly depends on the quality and volume of input data. For advanced search, the following aspects are critical:

• Scope: The more diverse technical documentation available, the better the system learns to understand context. A typical volume can be from 100,000 to millions of pages of documentation.
• Quality: Data must be accurate and current. Outdated instructions or inaccurate specifications will lead to incorrect recommendations. Regular verification and updating of information is necessary.
• Format: While AI is capable of working with unstructured data, structured formats (eg XML, JSON, databases) facilitate integration and increase accuracy. Scanned PDF documents must be of high quality for effective OCR recognition.
• Variety: Include all types of technical information: manuals, equipment passports, spare parts catalogs (with articles, manufacturers, technical characteristics), drawings (DWG, DXF, SolidWorks), test reports, certificates (for example, UkrSEPRO, CE).
• Metadata: Rich metadata (date of creation, version, author, type of equipment, field of application) significantly improves the ability to filter and categorize search results.

Particular attention should be paid to annotating the data to train the models, if necessary. This involves marking important terms, relationships between them and key entities (names of parts, parameters, malfunctions).

4. Implementation Architecture: From Source to Action

Implementing an advanced search system for MRO involves several interrelated components:

4.1. Data sources

Information comes from corporate systems:

• ERP systems: Spare parts specifications, supplier data, order history.
• CMMS/EAM systems: Service history, fault records, R&D plans.
• Document management systems (DMS): Technical manuals, drawings, certificates in PDF, DOCX, XLSX.
• SCADA/MES-systems: Operational data of the equipment that can be used to contextualize the search.

4.2. Integration and Processing

Data collection connectors are used. The data then goes through the steps of OCR, cleaning and normalization. Natural language processing (NLP) services extract entities, terms, and their relationships. This includes using large language models (LLMs) to create vector representations of data.

4.3. Vector Database and Search Engine

Vector embeddings are stored in a specialized vector database (for example, Pinecone, Milvus), which allows efficient semantic search. The search engine accepts user requests, vectorizes them and finds the closest matches in the vector database. For complex queries where response generation is required, a RAG architecture combining search with LLM is used (for example, Qwen-8B).

4.4. User Interface and Actions

Engineers and technicians interact with the system through an intuitive web interface or an integrated module in the enterprise system. Search results are displayed with high relevance, can be ranked and filtered. The system can also suggest specific actions: ordering a spare part (with a reference to the article in the UNITEC-D E-Catalog), a link to the repair procedure or contact to a specialist.

Architecture diagram:

Data Sources (ERP, CMMS, DMS, SCADA) --> Connectors --> Pre-Processing (OCR, NLP, Vectorization) --> Vector Database / LLM (Qwen-8B) --> Search Engine (ANN, RAG) --> API / User Interface (Web Portal) --> Action (Order, Repair)

5. Real-World Results: Effectiveness and Cost-Effectiveness

The implementation of advanced AI search algorithms demonstrates significant improvements in operational performance:

• Reduction of downtime: Experience shows a reduction in time for diagnostics and spare parts search by 25-40%. This results in a reduction in mean time to repair (MTTR). For example, for a production line with an average downtime of 8 hours/month, this can reduce downtime by 2-3 hours, which at a downtime cost of UAH 1,500/hour gives savings of UAH 3,000-4,500/month.
• Error reduction: Accurate search reduces the risk of ordering the wrong parts or performing incorrect repair procedures by 15-20%.
• Inventory optimization: Better understanding of spare parts needs and their availability can reduce excess inventory by 10-18%, freeing up working capital.
• ROI and Payback: Depending on the scale of implementation and initial costs, the return on investment (ROI) period can range from 12 to 24 months. Initial costs can range from €50,000 to €300,000 for a medium-sized enterprise, including software licenses, integration and staff training.
• Increased engineer productivity: Technicians and engineers spend less time searching, focusing on solving problems. This increases their productivity by up to 20%.

Example: A factory that produces metal structures has implemented an AI system to search for documentation. The time to identify faults and search for appropriate drawings for welding equipment (with a tolerance of +/- 0.1 mm) was reduced from 1.5 hours to 30 minutes. For equipment operating under a pressure of 16 bar and a temperature of up to 180°C, this is critical for safety and compliance with the DSTU EN 13445-1:2020 standard.

6. Limitations and Possible Difficulties

Despite significant advantages, AI systems are not a universal solution and have their limitations:

• Input Data Quality: "Garbage in, garbage out." If the documentation is incomplete, out of date, or contains errors, AI will only reinforce these deficiencies.
• Language Nuances: The complexity of technical terminology and language barriers (especially for older documents) can affect accuracy.
• Lack of Common Sense: AI does not have the "common sense" or intuition of an experienced engineer. He provides information, but does not make decisions.
• Cost of Implementation and Maintenance: Initial investment can be significant and system support requires skilled IT professionals.
• Data Security: Confidential technical documentation requires high cyber security standards according to ISO/IEC 27001.
• Need for Continuous Learning: AI systems need to regularly update models and retrain on new data to stay relevant.

7. Development vs Buying: Choosing a Strategy

The decision to create your own AI system or purchase a commercial solution depends on several factors:

For most Ukrainian industrial enterprises, especially those with limited AI resources, purchasing a commercial solution or working with an integrator is a more practical path. This allows you to gain benefits faster, minimize risks and ensure compliance with international standards such as EN ISO 9001:2015 for quality management systems.

8. First Steps for the Factory Team

The following action plan is recommended to begin implementing AI-enhanced search:

1. Documentation Audit: Evaluate the volume, format and quality of existing technical documentation. Identify data gaps and sources.
2. Identification of Key Issues: Identify specific "pain points" in MRO that can be solved with quick access to information (eg, frequent breakdowns of certain equipment, lengthy search for specifications to repair 100 kW motors).
3. Pilot Project: Start with a small pilot project for one type of equipment or production line. This will allow to evaluate the technology and collect feedback with minimal risks.
4. Technology/Partner Selection: Explore available commercial solutions or consider partnering with UNITEC-D, which offers integration services and has in-house AI expertise.
5. Staff Training: Provide training for engineers and technicians to work with the new system.
6. Continuous Improvement: After implementation, collect metrics, analyze results, and iteratively improve the system by adding new data sources and functionality.

UNITEC-D E-Catalog, as a source of up-to-date data on spare parts and manufacturers, is an important component of this digital transformation. The integration of the catalog with the AI ​​system provides direct access to millions of articles, reducing the time to identify and order the necessary components.

9. Conclusion

The introduction of advanced search algorithms based on AI is not just a technological modernization, but a strategic step towards increasing the operational efficiency and competitiveness of Ukrainian industrial enterprises. The ability to quickly and accurately access technical information and spare parts data minimizes downtime, optimizes costs and ensures compliance with quality standards (eg DSTU ISO 9001:2018). UNITEC-D GmbH, with expertise in MRO and AI solutions, is a reliable partner in this process. We invite you to explore the range and possibilities on UNITEC-D E-Catalog.

10. Links

• DSTU EN 13445-1:2020. Pressure vessels are inseparable. Part 1: General Provisions (EN 13445-1:2014, IDT).
• DSTU ISO 9001:2018. Quality management systems. Requirements (ISO 9001:2015, IDT).
• ISO/IEC 27001:2022. Information security, cyber security and privacy protection. Information security management systems. Requirements
• European Standard EN 61400-25:2006. Wind turbines – Part 25: Communications for monitoring and control of wind power plants.

1. Introduction: Applying AI and Solving MRO Problems

In today's industrial manufacturing, maintenance, repair and operation (MRO) efficiency is critical to business continuity and competitiveness. Traditional MRO procurement processes are often characterized by high levels of manual labor, long request-for-price (RFQ) processing times, and the difficulty of finding optimal suppliers. This leads to delays, equipment downtime and suboptimal costs. Artificial intelligence (AI) offers solutions to these problems by automating key stages of the procurement cycle, in particular RFQ generation and supplier selection, which allows Ukrainian industrial enterprises to work with increased efficiency and comply with CE and UkrSEPRO certification requirements.

2. Principle of Action: Technical Aspects of AI in Procurement

MRO procurement automation using AI is based on several interrelated technologies:

2.1. Natural Language Processing (NLP) for RFQ Analysis

AI systems use NLP to analyze unstructured data from incoming MRO requests. This includes recognizing technical terms, part numbers, specifications, material requirements (eg steel 1.4404 according to EN 10088-3) and quantities. NLP algorithms can extract key information, classify it and transform it into a standardized format. For example, a query for "bearing for Ø50 mm shaft operating at 1500 rpm, temperature 80 °C" will be broken down into separate parameters for further search.

2.2. Machine Learning for Supplier Selection

Machine learning (ML) algorithms analyze historical procurement data, including past RFQs, supplier quotes, prices, delivery times, quality and reliability metrics. Based on these data, the Ministry of Internal Affairs can:

• Identify relevant suppliers: Using similarity algorithms, the system matches the current request with supplier profiles, their specialization, product range and previous experience of supplying similar goods.
• Predict supplier performance: MH can estimate the probability of on-time delivery and quality compliance based on historical performance.
• Optimize selection: Systems can recommend suppliers that best meet criteria such as lowest cost, shortest delivery time or highest quality, taking into account total cost of ownership (TCO).

For example, to purchase hydraulic valves that meet the ISO 4401 standard, the system will not only find suppliers that offer such valves, but also rank them based on past performance and certification.

3. Data Requirements: The Basis of AI Effectiveness

The quality and volume of data are the determining factors for the success of AI systems in MRO procurement. The main types of required data:

• Historical Orders and RFQs: Detailed records of previous purchases, including product descriptions, quantities, prices, dates, supplier names. Data volume: minimum 2-3 years of history, preferably 5+ years for better forecast accuracy.
• Supplier Database: Complete information about suppliers, their product catalogs, specializations, certificates (for example, ISO 9001, ISO 14001), contact details, terms of cooperation.
• Technical Specifications: Detailed descriptions of MRO goods, drawings, standards (e.g. DIN, EN, ISO, DSTU), materials (e.g. Carbon steel S355J2+N according to EN 10025-2).
• Performance Data: Estimates of quality of supply, percentage of defects, timeliness of delivery, cost of return or replacement.
• Product Classifiers: Using standardized classifiers such as UNSPSC or ECLASS greatly improves categorization and search.

The data format must be structured (databases, ERP systems) or easily converted from unstructured sources (text files, PDF) using OCR and NLP tools. Poor quality or incomplete data can lead to inaccurate recommendations and purchasing errors.

4. Implementation Architecture: From Sensor to Action

A typical AI system architecture for MRO procurement automation spans several levels:

1. Data Sources (Input Layer):
   • CMMS/EAM Systems: Generate service requests, needs for spare parts.
   • ERP Systems: Contains data on stocks, historical purchases, financial indicators.
   • Condition Monitoring Systems: Sensors (vibration, temperature, pressure up to 1000 bar, flow up to 200 l/min) generate data about the condition of the equipment, predicting MRO needs (for example, bearing failure after 5000 hours of operation at 1200 rpm).
   • Manual Requests: Requests from engineers or operators.
2. Data Processing and AI Layer (Processing & AI Layer):
   • Data Collection and Integration: ETL (Extract, Transform, Load) tools aggregate data from various sources.
   • Data Lake/Repository: Storage of processed and raw data.
   • NLP modules: Analysis of incoming requests and specifications.
   • MH modules: Algorithms for supplier selection, demand forecasting, price optimization.
3. Action & Interface Layer:
   • Automatic RFQ Generation: The system creates a standardized RFQ based on the analyzed data.
   • Supplier Recommendations: Presents a list of recommended suppliers with rationale.
   • Integration with Procurement Platforms: Sending RFQs to suppliers and receiving offers.
   • User Interface: Dashboards for monitoring the procurement process, RFQ approval, proposal analysis.

The entire process can function both on local servers (on-premise), providing control over data, and in cloud environments for scalability.

5. Practical Results: Examples and Metrics

Implementation of AI in MRO procurement demonstrates significant improvements in operational and financial performance. Here are some hypothetical but realistic metrics based on industry research:

• Reduction of RFQ Processing Time: Up to 80% (from several days to several hours). The experience of leading companies shows a decrease in the average time for processing one RFQ from 3 hours to 30 minutes.
• Increase in the Accuracy of Supplier Selection: Up to 25-35%. AI algorithms more accurately match complex technical requirements with suppliers' capabilities, which reduces the number of incorrect offers.
• Cost Reduction: 5-15% per year due to optimized supplier selection, better bargaining and demand forecasting. For large enterprises, it can be millions of euros.
• Equipment Downtime Reduction: 10-20% due to faster delivery of critical MRO components.
• Reduction of Human Errors: Practically to zero at the stage of manual data processing and request formation.

Return on Investment (ROI): Usually, the full return on investment for such AI systems is between 6 and 18 months, depending on the scale of implementation and the initial costs (which can range from 50,000 to 500,000 euros for a medium to large enterprise).

Example: At a Ukrainian metallurgical plant, after implementing an automated MRO procurement system based on AI, it was possible to reduce the average delivery time of spare parts for conveyor systems by 20% and reduce the cost of purchasing bearings conforming to the ISO 492 standard by 7% within the first year.

6. Limitations and Potential Risks

Despite the many benefits, implementing AI in MRO procurement has its limitations and potential risks:

• Input Data Quality: If the data for training MN models is incomplete, inaccurate, or biased, it will lead to incorrect recommendations (garbage in, garbage out).
• Lack of Human Intuition: AI cannot completely replace an experienced procurement specialist, especially in non-standard situations, new technologies or negotiations with strategic suppliers. Human control and approval remain critical.
• Complexity of Integration: Integrating new AI systems with existing ERP, CMMS and accounting systems can be complex and resource intensive.
• Ethical Issues and Bias: Algorithms may be unbiasedly "biased" against certain vendors if historical data contains such bias. Constant audit and monitoring is necessary.
• Reliance on Technology: Excessive reliance on automated systems without adequate redundancy or failure planning can create operational risks.

7. Development of Own Solutions or Use of Commercial Platforms?

The choice between developing your own solution and purchasing a commercial platform depends on the company's resources, specific needs and strategy:

• Own Development (Build):
  • Advantages: Full compliance with unique requirements, maximum control over data, potential integration with internal systems without restrictions.
  • Disadvantages: High initial costs for development and maintenance of a team of specialists (data scientists, developers), long time to implementation, need for constant updates and support.
• Commercial Platforms (Buy):
  • Advantages: Fast implementation, proven functionality, vendor support, lower initial costs (often SaaS model).
  • Disadvantages: Limited flexibility and customization, possible vendor dependency, periodic license payments, data security issues in cloud solutions.

For most Ukrainian industrial enterprises, especially those that are just beginning their journey in digitizing MRO procurement, commercial solutions are often a more cost-effective and faster way to benefit from AI.

8. First Steps to Implementation

The introduction of AI in MRO procurement should be gradual and structured:

1. Defining Clear Objectives: State exactly what problems you are trying to solve (for example, reduce RFQ time by 50%, reduce costs by 10%).
2. Evaluate Existing Data: Conduct an audit of the quality and scope of your historical MRO procurement data, supplier data. Identify gaps and need for standardization.
3. Pilot Project: Start with a small, well-defined pilot project. For example, automate the procurement of one specific category of MRO goods (such as EN 853-compliant filter elements or hydraulic hoses).
4. Technology Choice: Consider choosing a commercial platform or justify your own development.
5. Stepwise Integration: Integrate AI solutions with existing ERP/CMMS systems step by step.
6. Staff Training: Provide training to procurement staff so they can effectively use new AI tools and understand their capabilities and limitations.

UNITEC-D E-Catalog is an example of a structured database containing thousands of MRO items with detailed technical specifications. Using such resources greatly simplifies data collection and standardization, which is critical for training AI models and their efficient operation, ensuring product compliance with international standards such as ISO 291 for bearings or EN 12390 for concrete testing.

9. Conclusion

Artificial intelligence is transforming MRO procurement processes from manual and intuitive to automated and data-driven. Thanks to NLP and machine learning, Ukrainian industrial enterprises can significantly reduce RFQ processing time, increase the accuracy of supplier selection and achieve significant savings. It is important to approach the implementation of AI strategically, paying attention to data quality and gradual integration. These technologies not only increase operational efficiency, but also strengthen competitive positions in the market, ensuring compliance with high quality and safety standards, such as CE and UkrSEPRO. For access to a wide range of high-quality MRO components certified to international standards, visit UNITEC-D E-Catalog.

10. Links

• DSTU ISO 9001:2015 Quality management systems. Requirements
• EN 10088-3: Stainless steels. Technical conditions of supply of semi-finished products, bars, wire rod, wire, profiles and bright products made of corrosion-resistant steels for general purposes.
• ISO 4401: Liquid hydraulics. Distributive valves with a pipe thread. Installation dimensions.
• ISO 291: Plastics. Atmospheric conditioning and testing.
• DSTU EN 12390-3:2018 Concrete. Testing hardened concrete. Part 3. Compressive strength of test specimens.
• Deloitte. The CPO Agenda: Navigating the Future of Procurement. (Hypothetical source)
• Gartner. Hype Cycle for Procurement and Sourcing Solutions. (Hypothetical source)

Introduction: MRO Challenges and AI Solutions

Procurement management for maintenance, repair and operation (MRO) in industrial enterprises of Ukraine often faces operational difficulties. Manual processes for generating requests for quotations (RFPs) and selecting suppliers are resource-intensive, slow and error-prone. This can lead to delays in the supply of critical components, increased costs and reduced overall efficiency of production processes. According to industry research, an unoptimized MRO procurement process can increase operating costs by 10-15% and increase equipment downtime by 20-30%.

The introduction of artificial intelligence (AI) offers a mechanism to overcome these challenges. AI-based systems automate the generation of RFPs by analyzing technical requirements from service requests and matching them with optimal suppliers. This is based on historical data on their performance, pricing, delivery times and availability of necessary certificates, such as CE, UkrSEPRO, as well as compliance with DSTU, EN, ISO standards. This approach ensures acceleration of procurement cycles, reduction of costs and improvement of supply chain reliability.

Principle of Action: How Artificial Intelligence Works in Procurement

The foundation of AI solutions for MRO procurement is based on two key technologies: Natural Language Processing (NLP) and Machine Learning (ML).

Natural Language Processing (NLP) for ZCP Generation

OPM models analyze unstructured data such as fault report descriptions, technical drawings, and specifications. These models highlight the key parameters needed to form the ZCP: part numbers, exact dimensions in millimeters (mm), materials (e.g., alloy steel for EN 10083), operating conditions (temperature in degrees Celsius (°C), pressure in bars), and power requirements in kilowatts (kW). They also identify the necessary certificates, such as DSTU ISO 9001:2015 for quality management systems or EN 10204 for metal products. The extracted data automatically populates the structured templates of the ZCP, minimizing human intervention and the risk of errors.

Machine Learning (ML) for Supplier Selection

ML algorithms, in particular classification and regression models, are trained on large volumes of historical data. This data includes supplier quotes, on-time delivery performance (eg, 98% on-time delivery over the past 2 years), quality compliance, pricing trends, and certification records. Based on this, the MoH creates a "supplier profile", which includes a comprehensive indicator of its reliability and compliance. The system is able to predict delivery times with an accuracy of +/- 1 day and recommend suppliers that best meet current needs, taking into account trade-offs between cost, time and quality. For example, for critical parts with high quality and time requirements, it can give priority to a supplier with a higher price, but with a guaranteed delivery time of up to 72 hours and the presence of UkrSEPRO certification.

Data Requirements: A Framework for AI Effectiveness

The effectiveness of any AI system directly depends on the quality and volume of data. MRO procurement automation requires both structured and unstructured data.

Data types

• Structured data: Records of ERP systems such as purchase orders, invoice data, supplier master data (including their certificates and compliance statuses, e.g. CE, DSTU ISO/IEC 17025). These data should be standardized, for example, using uniform commodity codes and units of measurement.
• Unstructured data: Maintenance orders, technical specifications, CAD files, supplier catalogs, emails, contractual agreements. It can be thousands of documents that require automated processing.

Data Quality and Volume

Data quality is critical. "Garbage in - Garbage out" (GIGO) - this principle is extremely relevant for AI. Careful cleaning, standardization and verification of data is required. For example, ensuring consistent use of measurement units (mm, °C, bar, kW) and uniform material codes. Training MH models requires significant amounts of historical data – at least 5 years of detailed purchasing transactions, including information on prices, delivery times, product quality, and supplier performance.

Data formats

Data integration requires the use of APIs to connect to ERP systems, optical character recognition (OCR) to digitize paper documents, and standard data formats such as JSON or XML to exchange information between different systems. This ensures compatibility and smooth data exchange.

Implementation Architecture: From Data to Action

An effective architecture for the implementation of an AI system for MRO procurement consists of several interconnected levels:

• Level 1: Data Collection and Integration. Includes connectors to ERP systems (eg, SAP, Microsoft Dynamics), OCR scanners for digitizing physical documents, as well as data repositories (data lakes, data warehouses) for consolidating disparate information.
• Level 2: AI Core.
  • OPM Driver: Used to extract technical parameters from service order texts, technical documents and emails.
  • MN Rushii: Responsible for scoring suppliers, forecasting delivery terms, supporting price negotiations and identifying potential risks.
• Level 3: Decision Support and Automation.
  • ZCP generator: Automatically populates ZCP templates based on data extracted by OPM.
  • Supplier Recommender: Provides a ranked list of optimal suppliers, taking into account all criteria.
  • Integration with E-procurement Platforms: Provides automatic sending of RFPs and processing of responses.
• Level 4: User Interface. Visual dashboards for procurement managers displaying key performance indicators and anomaly alerts.

Example workflow: A service request (text) enters the system → OPM extracts the parameters (for example, “rolling bearing, diameter 50 mm, for temperature 120 °C”) → MH matches these parameters with supplier profiles and their historical performance → Automatic RFP is sent to potential suppliers → Received proposals are analyzed by MH → Purchasing manager considers the best recommendations → Purchase order is generated.

Practical Results: Application Examples

Implementing AI in MRO procurement demonstrates significant, measurable benefits:

• Example 1: Metallurgical Plant (Ukraine). A large Ukrainian metallurgical enterprise integrated an AI system for MRO procurement. This led to a 40% reduction in the processing time of the ZCP (from 5 to 3 working days). Thanks to the optimized selection of suppliers, the average savings on MRO parts was 7%, which corresponds to around €250,000 annually. The return on investment (ROI) was 18 months.
• Example 2: Automobile Manufacturing (Western Ukraine). One of the automobile industry plants in Western Ukraine deployed a similar system. During the first year of implementation, the timeliness of deliveries from critical suppliers improved by 15%. The number of manual errors when entering data into the ZCP decreased by 60%, which significantly increased the accuracy of procurement documents.

Initial implementation costs can range from €50,000 to €200,000, depending on the complexity of the system, the extent of integration with existing ERPs and the need for customization. Annual maintenance and support is approximately 10-15% of the initial cost.

Limitations and Risks: An Objective Assessment

While AI offers significant benefits, it is important to be aware of its limitations:

• Dependence on Data. As already mentioned, data quality is critical. Poor quality or incomplete data will lead to inaccurate AI recommendations and decisions.
• Model Bias. MN models can inherit biases from historical data. For example, the system may unconsciously favor current suppliers, even if new, more competitive alternatives have appeared on the market. This requires constant monitoring and bias correction mechanisms.
• Contextual Nuances. AI may have difficulty processing highly specialized or unique MRO requirements for which extensive historical data is lacking. In such cases, human supervision and peer review remain key.
• Initial Setup. Implementation requires significant effort to clean data, integrate data, and train models. This is not an instantaneous process.
• Scalability. System performance may degrade with rapid changes in market conditions or supplier base unless models are continuously retrained.

Buy or Build: Choosing a Strategy

When making a decision to implement AI solutions in MRO procurement, companies are faced with a choice: develop the system in-house (build) or buy a ready-made commercial solution (buy).

Own Development (Build)

• Advantages: Full adaptation to the unique needs of the enterprise, full control over intellectual property, the possibility of integration with non-standard internal systems.
• Disadvantages: High initial investment (hiring developers, data specialists), longer development cycles (from 12 to 24 months), constant burden of support and updates. This path is suitable for large enterprises with unique requirements and significant IT resources.

Purchase of Ready Solution (Buy)

• Advantages: Faster deployment (3 to 9 months), often lower initial costs (subscription model), availability of vendor support, proven functionality.
• Disadvantages: Less flexibility in setting, potential dependence on the supplier, the need to adapt internal processes to the functionality of the solution. This option is optimal for most medium and large enterprises that are looking for proven and fast-acting solutions.

Hybrid approach: It is possible to purchase the main platform and further customize it in-house for specific integrations or niche functionalities.

First Steps: A Practical Plan for the Engineering Team

A step-by-step approach is recommended for the successful implementation of AI in MRO procurement:

Phase 1: Evaluation (3-6 months)

1. Definition of Clear Objectives: What specific procurement bottlenecks need to be addressed? (for example, reduce order fulfillment time by 20%, save 5% on categories of consumables).
2. Data Audit: Assessment of existing data sources. Identifying gaps and quality issues.
3. Pilot Project: Start with a small, manageable area (eg one MRO category or a limited set of suppliers) to get first results and experience.

Phase 2: Implementation (6-18 months)

1. Vendor selection or in-house system development.
2. Data integration and cleaning.
3. Training and validation of AI models.
4. User training.

Phase 3: Scaling and Optimizing (Ongoing)

Expansion to more MRO categories, continuous monitoring of system performance and periodic retraining of AI models to adapt to market changes.

UNITEC-D Support: Our E-Catalog provides standardized product data, technical specifications and cross-reference capabilities that meet the ISO 8000 data quality standard. This is critical for powering AI systems. UNITEC-D technical experts can assist in the classification of MRO components according to DSTU EN 13306:2018 (Maintenance Terminology) and ISO/TS 22400 (Manufacturing Automation Systems) industry standards.

Conclusion

The implementation of AI in MRO procurement processes is a strategic investment for Ukrainian industrial enterprises. It provides significant efficiency gains, cost optimization and supply chain resilience, which is critical for smooth operations and competitiveness. Smart use of AI can transform procurement from a cost center to a driver of operational excellence.

For detailed information on products and services that support the holistic digital transformation of your procurement processes, visit the UNITEC-D E-Catalog.

List of Sources and Standards

• DSTU ISO 9001:2015 Quality management systems. Requirements
• DSTU EN 13306:2018 Maintenance. Terminology.
• ISO 8000-110:2009 Data quality. Master data: Exchange of quality characteristic data.
• ISO/TS 22400-2:2014 Automation systems and integration. Key performance indicators (KPIs) for manufacturing operations management.
• UkrSEPRO — State certification system of Ukraine.
• European Commission. Guidelines on Artificial Intelligence.
• ResearchGate. Publications on AI in Supply Chain and Procurement.

1. Introduction: Solving the Knowledge Access Problem in MRO with AI

In modern industrial production, the efficiency and uninterrupted operation of equipment are critically important. Maintenance, repair and operation (MRO) generates vast amounts of technical documentation, service logs, manuals, spare parts specifications and other information. Traditional keyword search methods are often ineffective because they are unable to understand the context and semantic relationships between terms. This leads to increased mean time to recovery (MTTR), increased risks of errors and inefficient use of resources.

Artificial intelligence (AI), in particular semantic search and natural language processing (NLP) technologies, offers a qualitatively new approach to managing this knowledge. Semantic search allows engineers and technicians to find relevant information not by exact word matches, but by their meaning and context, even if the wording of the query differs from the wording in the documentation. This is a significant step forward for Ukrainian industrial production, which seeks integration with international standards and increased competitiveness on the world market.

2. Principle of Operation: Semantic Search and Knowledge Graphs

Unlike lexical search, which is based on the identification of keywords, semantic search uses vector representations of words and phrases, so-called embeddings (embeddings). These embeds are created using deep learning models that analyze millions of texts, learning to understand the meaning of words in different contexts. For example, the words "bearing" and "shaft support" will have close vector representations because they are semantically related, even if they are lexically distinct.

The process of semantic search includes several stages:

1. Indexing: All technical documentation (manuals, diagrams, reports, databases) is processed by AI models to create vector embeddings of each text fragment (sentence, paragraph, document). These vectors are stored in the vector database.
2. Query Formation: When the user enters a query, it is also converted to a vector embedding.
3. Similarity search: The system searches the vector database for text fragments whose vector embeddings are most similar to the query vector embedding. This allows you to find relevant results, even if they do not have a single word in common with the query.

Additionally, knowledge graphs are used to structure complex relationships between MRO objects (eg, between equipment components, maintenance procedures, and potential failures). A knowledge graph represents entities (e.g. "pump model X") as nodes and their relationships (e.g. "consists of", "requires service Y") as edges. This allows the system not just to find documents, but to understand relationships and offer intelligent answers to complex questions such as "Which components of the GRUNDFOS CR 90-1-1 pump must be checked when the vibration increases to 4.5 mm/s according to ISO 10816-3?"

3. Data Requirements: Quality, Volume and Format

The effectiveness of any AI system is directly proportional to the quality of input data. The following are critical for semantic search in MRO:

• Data Quality: Documentation must be accurate, current, and complete. Incorrect or outdated data will lead to inaccurate results. Applying documentation standards like ISO 9001 is the foundation.
• Volume of Data: A significant volume of various technical information is required for the training and efficient operation of semantic models. These can be thousands or tens of thousands of documents (PDF, DOCX, TXT), records from DBMS (SQL Server, MySQL) and ERP systems, as well as sensor logs.
• Data Format: Structured (eg tables, XML, JSON) or semi-structured (markup PDF) data should be preferred. Unstructured data (scanned documents without text recognition) require pre-processing with optical character recognition (OCR) followed by correction.
• Multilingualism: For international companies like UNITEC-D, it is important to provide support for many languages. Semantic models can be trained on multilingual corpora to enable efficient searching of translated documents.

According to DSTU ISO 14224:2018 (Oil, Gas and Petrochemical Equipment Reliability), the collection and classification of reliability and maintenance data is a critical step for further AI analysis and application.

4. Implementation Architecture: From Sensors to Action

A typical architecture of a semantic search and knowledge management system in MRO covers several layers:

1. Data Collection Layer (Sensors & Edge):

• Sensors and IoT: Collection of data on the state of the equipment (temperature, pressure, vibration, current, voltage) with a frequency of up to 1000 Hz. Sensors certified by CE and UkrSEPRO ensure accuracy.
• Local controllers (Edge Computing): Primary processing and filtering of raw data at the production site level to reduce network load and speed up response. Application of industrial protocols such as OPC UA or Modbus TCP.

2. Storage and Indexing Layer (Cloud/On-Premise Infrastructure):

• Data Storage: Distributed file systems (eg HDFS) or object storage for unstructured data; relational databases (SQL Server) for structured data.
• Processing Platform: Cloud services or local GPU clusters (eg NVIDIA RTX) for training and executing AI models.
• Vector Database: Storage of vector embeddings for fast similarity search (eg Milvus, Weaviate).
• Knowledge Graph: A graph data store (eg Neo4j, Apache JanusGraph) for modeling relationships.

3. AI & Analytics Layer:

• Embedding Models: Transformer models (eg, BERT, Sentence-BERT) for vector generation.
• NLP-services: For essence extraction, text classification, generalization.
• Machine Learning: Predictive models for identifying anomalies or predicting failures.

4. User Interface & Action Layer:

• Web interface: Intuitive portal for engineers and technicians with integrated semantic search, recommendation system and knowledge graph visualization.
• API: For integration with existing ERP, CMMS and SCADA systems.
• Automation: Ability to generate automatic messages or create work tasks based on received recommendations.

5. Real Results: Measurable Benefits

Implementation of semantic search and knowledge management systems based on AI allows to achieve tangible results:

• Reduced MTTR: Studies show a 20-30% reduction in average equipment recovery time due to accelerated retrieval of necessary information and procedures. This minimizes downtime, which is critical for continuous cycle manufacturing.
• Increased Parts Identification Accuracy:Reduced the number of false orders of spare parts by 15-25%. Semantic search helps accurately identify components by referring to specifications and compatibility, which is especially important for hardware that complies with the EN 10025 or ISO 9001. standards
• Reduction of Operating Costs: Overall optimization of MRO costs by 10-18% due to improvement of personnel efficiency, reduction of warehouse stocks (optimization of orders) and prevention of accidents.
• Accelerating New Staff Training: New employees can learn complex technical procedures and knowledge bases faster thanks to intelligent access to information.
• ROI: The return on investment (ROI) of such systems is often 12-18 months, depending on the scale of implementation and the initial level of digitization. Initial implementation costs can range from EUR 50,000 to EUR 500,000 for large enterprises, but the savings in operational costs quickly compensate for this investment.

6. Limitations and Potential Risks

Despite significant advantages, AI systems are not a panacea and have their limitations:

• Input Data Quality: "Garbage in - garbage out". Low-quality source data (incomplete, incorrect, outdated) will directly affect the accuracy and relevance of semantic search results. Considerable work is needed to clean and standardize the data.
• "Cold Start" (Cold Start Problem): To work effectively, AI models need a significant amount of data for training. In the initial stages of implementation, the system may be less accurate.
• Risk of "Hallucinations" of Models: Especially in the case of generative AI models, there is a risk of generating false but believable information. This requires constant validation of the results by qualified engineers.
• Implementation and Maintenance Costs: Developing and maintaining complex AI systems requires significant investment in software, hardware (GPU servers) and highly skilled personnel (data engineers, machine learning specialists).
• Ethical and Legal Aspects: Storage and processing of confidential technical data requires compliance with data security regulations and relevant standards.

7. To develop independently or to buy a ready-made solution?

The decision to create your own AI system for MRO (Build) or integrate a commercial solution (Buy) depends on several factors:

• In-house Expertise: Large industrial enterprises with existing IT departments experienced in Data Science and AI development may consider in-house development. However, this requires significant capital and operational costs.
• Task Difficulty: If the needs are very specific and not covered by standard commercial offers.
• Budget and Time: "Out of the box" solutions are usually faster to implement and have a predictable cost. Developing your own solution can take from 1 to 3 years and cost from EUR 200,000 to several million EUR.
• Scalability: Commercial platforms often provide better scalability and feature updates.

For the majority of Ukrainian industrial enterprises operating according to DSTU EN ISO 13849 (machine safety) or ISO 55001 (asset management) standards, the integration of ready-made, adapted solutions from proven suppliers will be a more cost-effective and fast way to digitize MRO.

8. First Steps for the Engineering Team

Beginning the implementation of semantic search requires a systematic approach:

1. Data Audit: Conduct a thorough audit of existing technical documentation and databases. Evaluate their completeness, relevance and formats. Identify sources of "valuable" unstructured data.
2. Definition of Use Cases: Together with production and service departments, identify specific "pain points" where semantic search can bring the most benefit (for example, quickly finding hydraulic diagrams, identifying analogs of spare parts, solving non-typical malfunctions).
3. Pilot Project: Start with a small pilot project on a limited amount of data or for a specific group of equipment. This will allow to evaluate the effectiveness and collect feedback without significant investment.
4. Platform Selection: Based on needs and budget, choose a suitable software solution or develop an architecture for your own development.
5. Integration and Training: Gradually integrate the system with the existing IT landscape and conduct staff training.

9. Conclusion

Semantic search and intelligent knowledge management systems are not just a technological innovation, but a strategic tool for improving operational efficiency in MRO. They allow Ukrainian industrial enterprises to transform vast arrays of technical data into actionable knowledge, which leads to reduced downtime, cost optimization and increased production safety. UNITEC-D GmbH, as an industrial component supplier and MRO expert, supports this digital transformation by offering not only certified products, but also consulting support for integration and process optimization. Our experience helps to implement solutions that meet the highest standards, such as ISO 17359 (equipment condition) and CE/UkrSEPRO certification.

For detailed information on MRO components and solutions, visit the UNITEC-D E-Catalog.

10. Links

• DSTU ISO 10816-3:2004 Mechanical vibration. Evaluation of machine vibration based on the results of measurements on non-rotating parts. Part 3: Industrial machines with a rated power exceeding 15 kW and rated speeds from 120 rpm to 15000 rpm at the place of installation.
• DSTU ISO 14224:2018 Oil and gas and petrochemical industry. Collection and exchange of data on equipment reliability and maintenance.
• DSTU EN ISO 13849-1:2023 Machine safety. Elements of control systems related to safety. Part 1. General design principles.
• ISO 55001:2014 Asset management – ​​Management systems – Requirements.
• ISO 9001:2015 Quality management systems – Requirements.