MRO Spare Parts Demand Forecasting Using Artificial Intelligence: Warehouse and Maintenance Optimization

Introduction

Spare parts management for Maintenance, Repair and Overhaul (MRO) represents a critical challenge for the manufacturing industry, particularly in the machine tool sector. The primary objective is to ensure the availability of essential components to minimize unplanned downtime, without incurring excessive costs due to immobile inventories. Traditionally, forecasts have been based on simple statistical methods or human experience, which are often insufficient to handle the complexity and variability of modern demand. The introduction of Artificial Intelligence (AI) and Machine Learning (ML) techniques offers an advanced approach to transform this practice, significantly improving forecast accuracy and optimizing spare parts inventory management.

How It Works

The effectiveness of AI spare parts demand forecasting lies in the ability to analyze complex correlations and hidden patterns within vast amounts of data, overcoming the limitations of linear models. The process begins with the collection of historical consumption data, which represents the core of any forecasting algorithm. Added to these are exogenous variables that directly or indirectly influence demand:

Historical consumption data: Detailed records of purchases and issues of spare parts from the warehouse, with daily or weekly granularity.
Machine operating data: Operating hours, work cycles, data from sensors (vibrations, temperature, pressure, amperage) that may indicate imminent failure or wear. For example, an abnormal increase in engine vibration can herald the need for a new bearing.
Maintenance plans: CMMS (Computerized Maintenance Management System) system logs which include scheduled preventive, predictive and corrective maintenance.
Production plans: Changes in production volumes or introduction of new products which may alter the workload of the machines and, consequently, the demand for spare parts.
External factors: Seasonality, market trends, technological obsolescence, supplier policies.

This data is processed via Machine Learning algorithms, such as advanced time series models (e.g. ARIMA, SARIMA, Prophet) or recurrent neural networks (e.g. LSTM) to capture complex and non-linear time dependencies. The model learns from historical patterns, identifying the most influential variables and quantifying their impact on future demand. For example, an algorithm can determine that the demand for a specific seal increases by 15% every 500 hours of operation of a centrifugal pump, especially if the fluid temperature exceeds 70°C. Validation of the model occurs by comparing the predictions with real data not used in the training phase, measuring error metrics such as the Mean Absolute Error (MAE) or the Root Mean Squared Error (RMSE). An MAE of less than 5% is often a realistic target for critical parts.

Data Requirements

The accuracy of AI predictions is directly proportional to the quality and completeness of the data available. The given requirements include:

Historical inventory and transaction data: At least 2-3 years of daily or weekly data on receipts, expenditures and stocks of each individual SKU (Stock Keeping Unit). This data typically comes from the ERP (Enterprise Resource Planning) system.
CMMS data: Logs of failures, maintenance interventions, repair times, MTBF (Mean Time Between Failures) and MTTR (Mean Time To Repair) for each asset and component.
Sensor data (IoT/OT): For critical assets, real-time or near-real-time data from industrial sensors (temperature, vibration, current, pressure). Compliance with the UNI CEI EN 62443 (series) standard for the cyber security of industrial control systems is essential.
Component data sheets: Manufacturer specifications, estimated useful life, optimal operating conditions.
Production and maintenance plans: Flexible and updated, essential for anticipating peaks or drops in demand.

Data quality is critical: incomplete, incorrect or inconsistent data can lead to biased predictions. It is necessary to implement rigorous data governance and data cleansing processes, possibly automated, to ensure the reliability of the input. Format normalization and standardization are also essential for integration between heterogeneous systems such as ERP, CMMS and IoT platforms.

Implementation Architecture

The implementation of an AI forecasting system for MRO spare parts requires a robust and integrated architecture, which can be summarized as follows:

Data Acquisition: IoT/OT sensors on machines (e.g. encoders, thermocouples) collect data in real time. SCADA, CMMS and ERP systems provide historical and transactional data.
Edge Computing (Optional): For preliminary analysis and filtering of high-frequency data directly in the factory, reducing latency and the volume of data to be transmitted.
Data Lake/Data Warehouse: A centralized repository (e.g. Data Lake based on Apache Hadoop or Data Warehouse on SQL Server) where all raw and pre-processed data is stored.
AI/ML platform: Dedicated environment for the development, training, validation and deployment of prediction models (e.g. TensorFlow, PyTorch, scikit-learn on cloud or on-premise infrastructure).
Inference Engine: The trained model generates future demand forecasts, often with confidence intervals.
Integration Layer: The forecast results are fed back into the management systems (ERP/CMMS) to automate purchase orders, optimize stock levels and plan maintenance. The integration must comply with interoperability standards, such as those defined by the EN ISO 22745.
User Interface (Dashboard): An intuitive dashboard allows warehouse and maintenance managers to view forecasts, monitor performance, set parameters and receive alerts on potential critical issues.

The entire system must be designed with scrupulous attention to industrial data security, in line with Directive (EU) 2022/2555 (NIS2) and CEI EN 62443 (series) standards.

Real Results

The benefits deriving from the adoption of an AI forecasting system in the management of MRO spare parts are tangible and measurable:

Inventory Reduction: Accurate forecasting can reduce warehouse inventory levels by 15-30%. This translates into a decrease in fixed capital costs and maintenance costs (insurance, obsolescence, space) by up to 20%. A company with a spare parts inventory value of 5 million Euros can achieve annual savings of 1 million Euros.
Stockout Minimization: The probability of stockouts for critical spare parts can be reduced by 50-75%, ensuring greater operational continuity. This means avoiding unexpected machine downtime, which in the machine tool sector can cost from 500 to 5,000 Euros per hour, depending on the complexity of the system and the production volume.
Maintenance Optimization: The availability of spare parts at the right time facilitates the implementation of predictive and preventive maintenance strategies, improving the efficiency of interventions and the useful life of assets.
Service Improvement: Faster response times and greater production reliability translate into better service to the end customer.

The return on investment (ROI) for projects of this type is estimated between 12 and 36 months, with an initial implementation cost ranging from 50,000 Euros for solutions based on existing platforms to over 500,000 Euros for complex custom developments, excluding software and hardware licensing costs. These values are highly dependent on the scale of the implementation and the pre-existence of a robust data infrastructure.

Limitations and Pitfalls

Despite the significant benefits, implementing AI forecasting systems has limitations and potential pitfalls that require careful management:

Data Quality and Availability: Incomplete, noisy or insufficient data seriously compromises the accuracy of forecasts. The data collection and cleaning phase is often the most expensive.
Rare or “Black Swan” Events: AI models struggle to predict events historically not represented in training data, such as unexpected catastrophic failures or large-scale supply chain disruptions (e.g. pandemics, conflicts). In these cases, human intervention based on experience and contingency plans is essential.
Model Complexity and “Black Box”: Some advanced AI models can be difficult to interpret, making it challenging for engineers to understand the “why” of a certain prediction. This can hinder trust and adoption.
Initial Cost and Skills: The investment in IT infrastructure, software and specialized personnel (Data Scientist, AI Engineer) can be significant. Lack of in-house expertise is a common holdback.
Overspecification/Overfitting: A model that is too complex for the amount of data available can "memorize" the noise of historical data instead of real patterns, failing miserably on new data.

It is essential to adopt an iterative and collaborative approach, which integrates the expertise of maintenance engineers with the predictive capabilities of AI.

Build vs Buy

The decision between developing an AI forecasting solution in-house (“Build”) or adopting an existing commercial platform (“Buy”) depends on several strategic and operational factors:

Internal Development (Build)

Advantages: Maximum customization for specific business needs, full control over intellectual property, deep integration with existing IT systems.
Disadvantages: Requires high internal skills in Data Science and AI/ML, long development times (typically 12-24 months for a mature solution), ongoing development and maintenance costs. Suitable for businesses with unique requirements or a significant competitive advantage from customization.

Purchase of Commercial Solutions (Buy)

Advantages: Faster implementation times (3-9 months), less need for internal AI skills (third-party vendors often manage model development), access to pre-built features and industry best practices, costs distributed in the form of licenses or SaaS.
Cons: Less flexibility and customization, vendor dependency, potentially high long-term licensing costs. Ideal for companies looking for a ready-to-use solution, with a focus on time-to-value.

UNITEC-D can support both approaches, providing UNI EN ISO 9001 certified spare parts and technical consultancy for the integration of components, whether for custom systems or for powering existing platforms.

How to Get Started

For a team of factory engineers in Italy, starting an AI parts demand forecasting project can follow these steps:

Preliminary Evaluation and Pilot Identification: Select a limited group of high-turnover or high-cost parts, or a specific plant, as a pilot project. This reduces complexity and allows you to quickly demonstrate value.
Data Audit: Conduct an in-depth analysis of the availability and quality of historical ERP, CMMS and sensor data. This includes identifying gaps, errors and establishing processes for their correction and ongoing management.
Stakeholder involvement: Secure management support and actively involve maintenance and warehouse staff. Their experience is crucial to validate predictions and ensure adoption.
Platform/Vendor Selection: Based on the “Build vs Buy” strategy, select the most suitable AI/ML tools or service providers. Consider solutions that ensure compatibility with industry standards.
Iterative Development and Validation: Implement the system in phases, testing and refining models with continuous feedback loops.
Training: Train staff on using the new system and understanding forecasts.

UNITEC-D, with its wide range of certified components and its expertise in the MRO sector, can be a strategic partner in this transformation, guaranteeing the supply of quality spare parts that power the predictive systems.

Conclusion

Integrating Artificial Intelligence into MRO parts demand forecasting is no longer a mere option, but a strategic necessity for manufacturing companies aiming for operational excellence. By transforming warehouse management from a cost center to a manufacturing enabler, you achieve significant reductions in costs and downtime. This approach, based on advanced data and algorithms, supports more efficient maintenance, greater operational resilience and, ultimately, a sustainable competitive advantage. Technological evolution requires a proactive review of MRO strategies to align with the principles of Industry 4.0.

Explore our offering of high-quality, certified components that support your digital transformation: UNITEC-D E-Catalog.

References

UNI EN ISO 9001:2015 - Quality management systems.
UNI CEI EN 62443 (series) - Cyber security for industrial automation and control systems.
Directive (EU) 2022/2555 (NIS2) - Measures for a high common level of cybersecurity in the Union.