Optimization of TORO: Automatic Classification of Applications Using Natural Language Processing

Introduction: The Problem of Classification of Applications in TORO

In today's industry, where production efficiency is critical, maintenance, repair and operation (MRO) management plays a central role. Hundreds and sometimes thousands of maintenance, repair or spare parts requests are generated every day. The traditional manual classification, prioritization, and routing of these requests is a time-consuming and human-error-prone process. This results in response delays, misallocation of resources, increased equipment downtime and, as a result, financial losses.

To solve this problem, UNITEC-D GmbH proposes to consider the use of artificial intelligence technologies, in particular Natural Language Processing (NPM), to automate the classification of applications in TORO systems. This approach allows you to transform an unstructured textual description of the problem into structured data that can be instantly classified, prioritized and directed to the appropriate actors. This corresponds to the principles of Industry 4.0 and allows Ukrainian enterprises to increase operational efficiency and reliability of production processes.

Principle of Operation of the OPM for the Classification of Applications

Artificial intelligence is a branch of artificial intelligence that allows computers to understand, interpret, and generate human speech. In the context of the classification of TOO requests, the OPM system analyzes the textual description of the problem provided by the operator or technician and assigns it to one or more predefined categories. The process includes several key stages:

Collection and Preprocessing of Data: The system needs historical data - thousands or tens of thousands of already classified applications with text descriptions. The text is cleaned of unnecessary characters, typographical errors, stop words (for example, "and", "but", "y").
Tokenization and Vector Representations (Embeddings): The text is divided into separate words or phrases (tokens). Each token is converted into a numeric vector that reflects its semantic meaning and context in the sentence. Modern models such as BERT or Word2Vec create high-quality vector representations that allow the system to "understand" synonyms and context.
Training a Machine Learning Model: Based on these vector representations, a machine learning model (for example, a classifier based on deep neural networks or a support vector method) is trained. The teacher model matches text vectors to their correct categories, revealing patterns.
Classification: After training, when the system receives a new, unclassified request, it goes through the same processing steps. The trained model analyzes the vector representation of the new text and predicts the most likely category (eg, "Electrical fault", "Hydraulic leak", "Scheduled maintenance") with a certain level of confidence.

This approach allows the system not just to search for keywords, but to understand the content of the application, even if the wording is unclear or contains specific technical slang characteristic of a particular enterprise.

Data requirements

The quality and volume of data are critical to the successful implementation of an OPM system. Without proper data, any model will be ineffective. The main requirements include:

Volume of Historical Applications: For the initial training of the model, a minimum of 5,000 - 10,000 qualitatively classified applications are required. To achieve high accuracy, it is desirable to have 20,000 - 50,000 applications or more.
Quality of Text Descriptions: Descriptions should be as detailed and clear as possible. Descriptions that are unclear, too short, or filled with abbreviations that are not decoded in the system reduce accuracy. It is necessary to conduct an audit of the data and, if possible, to unify the terminology.
Manual Classification Accuracy: Historical applications must be correctly classified by humans. If the training data contains errors, the model will reproduce them. It is recommended to check and correct existing classifications.
Data Format: Data must be available in a structured format (for example, CSV, JSON, or directly from the STUOiR/SAUAP database). Each entry must contain a text description of the problem and the corresponding category (or categories). Additional fields such as asset ID, date, priority, performer can be used to enrich the model.
Privacy and Data Protection: According to the requirements of Ukrainian legislation (for example, the Law of Ukraine "On Protection of Personal Data") and international standards (ISO/IEC 27001), it is necessary to ensure proper protection of confidential information contained in applications.

The preparation of data is the most time-consuming stage of implementation of OPM. Investments in data quality pay off with significantly higher system accuracy and reliability.

System Implementation Architecture

Integration of the OPM system into the existing production landscape requires a well-thought-out architecture. A typical architecture includes the following components:

Operators/Sensors → TORO Management System (SUTOiR/SUAP) → OPM Module → Dispatching System/Database → Performers

Source of Applications: These can be operators who manually enter descriptions of problems into the system (for example, SAP PM, IBM Maximo, 1C:TORO), or automatic condition monitoring systems (Condition Monitoring), which generate notifications based on data from sensors (temperature, vibration, pressure according to EN ISO 10816-1).
Toro Management System (SUTOiR/SUAP): The company's existing system that collects, stores and manages applications. The OPM module integrates with this system to receive unclassified applications and return classified data.
OPM module: The central component that performs natural language processing. It can be deployed as a local server (On-Premise) to ensure maximum data security or as a cloud service (Cloud-based) for flexibility and scalability. The module includes:
- Input/Output interface: For interaction with SUTOiR/SUAP via API (Application Programming Interface).
- Pre-Processing subsystem: Text cleaning, tokenization.
- Subsystem of Vector Representations: Conversion of text into numerical vectors.
- Classification Model: Trained machine learning model.
Dispatch System and Database: Classified requests are automatically routed to the appropriate teams or individual technicians who specialize in this category of malfunctions. The information is stored in a database for further analysis and reporting.
Feedback and Retraining: A critical element. Technicians or TORO managers check the automatic classification. If the classification was incorrect, they correct it. This adjusted data is used to periodically retrain the model, improving its accuracy over time.

It is important to ensure that the integration complies with data security and information systems standards such as DSTU ISO/IEC 27001:2015.

Real Results and Economic Efficiency

The implementation of OPM for the classification of TORO applications demonstrates significant improvements in operational indicators:

Improve Classification Accuracy: Automated OPM systems achieve classification accuracy of 85-95%, which is significantly higher than manual classification, which often hovers around 60-70% due to fatigue, lack of standardization, and subjectivity.
Reduction of Application Processing Time: The time from receiving an application to its classification and routing is reduced from 15-30 minutes (with manual processing) to less than 1 minute. This speeds up the response to critical malfunctions.
Decreasing the Number of Incorrectly Routed Requests: Up to 20-40% of requests that were previously sent to the wrong team or specialist are now classified and routed correctly. This reduces the "rollover" of the application between departments.
Decreased Mean Time to Recovery (MTTR): Faster and more accurate dispatching leads to reduced MTTR, which directly impacts equipment downtime. A 10-15% drop in MTTR is typical.
Improved Data Analysis: Standardized automatic classification generates high-quality data for further root cause analysis, failure prediction and optimization of preventive maintenance schedules.
Cost Savings and Return on Investment (ROI): Due to the reduction of downtime, optimization of staff working time (up to 10-15% of the working time of maintenance managers), and increased efficiency of repair work, the return on investment is usually 12-24 months.

Applied Metrics:

Cost of the pilot project: 20,000 – 50,000 EUR (including data preparation and deployment of the base model).
Cost of full integration: 100,000 – 500,000 EUR (depends on complexity of integration, amount of data and need for customization).
Reduction of administrative costs: Up to 0.5 - 1.5 EUR for each processed application. With 10,000 applications per month, this results in savings of 5,000 - 15,000 EUR per month.

Limitations and Potential Problems

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

Garbage In, Garbage Out: If the training data is of poor quality, the model will produce inaccurate predictions. Poor descriptions, inconsistent terminology, or incorrect manual classification create fundamental problems.
Ambiguity and Context: Human language is complex. OPM models may have difficulty understanding highly ambiguous, sarcastic, or overly brief descriptions that require deep contextual understanding.
Domain Specificity: Models trained on general text corpora will not be efficient for TORO technical jargon. Training on industry-specific data is required. This requires significant effort in data collection and labeling.
Support and Retraining: Production processes and equipment change. New types of malfunctions appear, new assets, deadlines change. The OPM model requires regular monitoring and periodic retraining (for example, once every 3-6 months) on new data to maintain its relevance and accuracy.
Dependence on IT infrastructure: Deploying and operating OPM systems requires a robust IT infrastructure, computing resources, and possibly specialized machine learning expertise.

Build or Buy: Decision Analysis

When planning the implementation of an OPM system for the classification of applications, enterprises are faced with a dilemma: to develop a solution in-house (Build) or to purchase a commercial solution (Buy).

Development by own forces (Build):

Advantages:

Full customization for the unique needs of the enterprise.
Preservation of intellectual property.
Deep integration with existing internal systems without compromise.

Disadvantages:

High initial development and maintenance costs.
The need for a specialized data team (Data Scientists, ML Engineers).
Long implementation time (12-24 months).
The need for constant support and development.

Purchase of a Commercial Solution (Buy):

Advantages:

Faster deployment (3-6 months for basic integration).
Lower upfront costs (although overall license and support costs can be significant).
Support and updates from the vendor.
Often includes pre-trained models that require retraining on specific client data.

Disadvantages:

Less flexibility in customization.
Vendor lock-in is possible.
Data security issues if the solution is cloud-based.
Standard solutions may not fully match unique business processes.

For the majority of Ukrainian industrial enterprises, especially at the initial stage, a hybrid approach is optimal: the purchase of a ready-made platform that allows further training on own data and integration with existing EMS/ACS. This allows you to gain the benefits of speed of implementation and support while maintaining adaptability.

Practical Steps for Implementation

The following step-by-step plan is recommended for the company's engineering team considering the implementation of OPM:

Data Audit: Conduct a detailed analysis of the existing historical applications in your EMS/ACS. Evaluate the volume, quality of textual descriptions, and accuracy of the current classification. Determine if there is enough data to train the model.
Usage Scenario Definition: Start with a specific, high-volume category of claims where manual classification is the most problematic (for example, failures of a single type of equipment or a specific shop floor).
Formation of an Interdisciplinary Team: Create a working group that includes maintenance specialists, IT department, automation department and possibly operational personnel.
Pilot Project: Deploy a small pilot project. Use a limited data set and focus on one or two categories. This will allow testing the technology, evaluating its effectiveness and identifying potential problems without significant investment.
Selection of Solutions: Based on the results of the pilot project and market analysis, choose the most appropriate commercial solution or develop an internal development plan.
Integration: Ensure seamless integration of the OPM module with your existing EMS/SMS by following data exchange standards such as DSTU ISO/IEC 19505-1:2015 (UML).
Monitoring and Optimization: After implementation, continuously monitor classification accuracy, collect user feedback, and regularly retrain the model on new, adjusted data to maintain high performance.

UNITEC-D GmbH supports digital transformation strategies in industry by providing reliable and certified (CE, UkrSEPRO) spare parts and components that comply with EN and ISO standards. Our wide range of products, available through the electronic catalog, allows you to quickly find and order the necessary elements for any systems that have been diagnosed with the help of intelligent TORO systems. This ensures uninterrupted equipment operation and minimizes downtime, which is key to successfully realizing the benefits of OPM.

Summary

Automatic classification of maintenance requests using Natural Language Processing is a powerful tool for improving maintenance efficiency in Ukrainian industry. It allows to reduce the response time, optimize the allocation of resources and significantly improve the quality of data for further analysis. Although implementation requires investment in data and technology, the potential savings and increased operational reliability provide a quick payback.

To support your M&A initiatives and ensure the reliability of production processes, we invite you to familiarize yourself with our range of high-quality industrial spare parts and components. Visit the UNITEC-D E-Catalog.

Link

DSTU ISO 9001:2015 Quality management systems. Requirements
DSTU ISO/IEC 27001:2015 Information technologies. Methods and means of ensuring security. Information security management systems. Requirements
DSTU EN ISO 10816-1:2006 Vibration. Measurement and evaluation of machine vibration. General requirements.
EN ISO 13849-1:2015 Safety of machinery. Safety-related parts of control systems. General principles for design.
CE Marking Directives for Machinery (2006/42/EC).
Technical regulation of machine safety, approved by the resolution of the CMU dated January 30, 2013 No. 62 (corresponds to Directive 2006/42/EC).
DSTU 4163-2003 Design documentation system. Rules for design of text documents.