1. Introduction
A modern industrial enterprise generates a huge amount of information about the condition of equipment, which comes in the form of technical applications. In MRO (Maintenance, Repair, and Operations) systems, the efficiency of processing this data is critically dependent on the speed and accuracy of classification. Traditional manual processing of applications by dispatchers causes delays, routing errors and reduced equipment availability. Classification automation using natural language processing (NLP) techniques solves this problem by transforming unstructured texts into structured data for ERP systems.
2. How it works
NLP algorithms analyze the text description of the fault provided by the operator or the monitoring system. The main process includes:
- Tokenization: Dividing the text into separate words or symbols.
- Embedding (Embedding): Transformation of words into vector representations, where words with similar meaning are close in multidimensional space.
- Classification: Using Transformer-type models (e.g. DeBERTa) that estimate the probability of an application belonging to a certain category (e.g. "electrical", "hydraulic", "mechanical").
The model learns from historical data, revealing patterns that cannot be grasped by humans with large volumes of data.
3. Data requirements
The following conditions are necessary to build a reliable classification system:
- Data quality: Availability of a historical archive of applications (minimum 10,000 - 50,000 records) that have already been classified by specialists.
- Cleaning: Removal of technical garbage, duplicates and spelling errors.
- Format: Data must be in a machine-readable format (SQL, CSV) with clear labeling of categories.
4. Implementation architecture
The system functions according to the following scheme:
- Source: osTicket or other application accounting system.
- Preprocessor: Service (Worker V8) that performs text cleaning and querying the AI model.
- Inference: The model classifies the text and returns a category tag with a confidence level.
- Action: The result is written to the SQL database (for example ERP Unitec), which automatically starts the corresponding work process.
5. Practical results
The implementation of automated classification demonstrates the following indicators:
| Indicator | The result |
|---|---|
| Time for primary classification | 40-60% reduction |
| Classification accuracy | 85-95% |
| Return on investment (ROI) | 6-12 months |
For example, at a medium-sized enterprise, the implementation of such a system made it possible to reduce MTTR (mean recovery time) by 15% due to the instant routing of applications to specialized specialists.
6. Limitations and risks
AI is not a one-size-fits-all tool. Main risks:
- Errors with new types of faults: If the model has not seen similar cases during training, it may make an error.
- Data Drift (Data Drift): Changing terminology or technical processes at the plant requires regular retraining of the model.
- Human factor: The need for a "human-in-the-loop" mechanism to verify applications with low AI confidence.
7. Purchase of a ready-made solution or own development
In-house development: Expedient if you have specific terminology and your own team of developers. Provides full control over data.
Commercial solutions: Suitable for a quick start. Advantage in support and ready integrators for popular ERP.
8. First steps
Recommended plan for the engineering team:
- Conduct an audit of historical applications for the past 2 years.
- Select a pilot plant or type of testing equipment.
- Provide access to the database for analysis (writing to PROD servers is prohibited).
- Develop a mechanism for verification of automatic tags by engineers.
9. Summary
Automating application classification is a first but critical step towards Industry 4.0. This allows engineers to focus their efforts on solving technical problems, rather than on bureaucracy. For efficient equipment maintenance and selection of spare parts, use UNITEC-D E-Catalog.
10. Links
- ISO 55001:2014 – Asset management. Management systems. Requirements.
- DIN 31051:2012 – Fundamentals of maintenance.
- Documentation on the design principles of NLP systems (according to industry software development standards).
