Knowledge integration for fault severity estimation
| Reference number | |
| Coordinator | Luleå tekniska universitet - Institutionen för system- och rymdteknik |
| Funding from Vinnova | SEK 4 474 684 |
| Project duration | September 2019 - September 2023 |
| Status | Completed |
| Venture | Strategic innovation programme for process industrial IT and automation – PiiA |
| Call | Digitization of industrial value chains |
Important results from the project
New concepts for analysing textual descriptions of machine faults and measurement data in condition monitoring systems for paper machines have been developed to automate time-consuming and frequent signal and alarm analysis tasks. Methods for improved decision support based on technical language analysis have been developed so that analysts can perform more in-depth analyses and more preventive maintenance with the aim of avoiding unplanned stops. The results are summarized in a licentiate thesis.
Expected long term effects
A plan for commercialisation that allows for the digitisation of domain knowledge described in text within the condition monitoring domain has been developed and is expected to create new opportunities for the development of decision support solutions, services, and business relationships. Methods for identifying cable and sensor faults have been developed which can reduce the analysts´ workload by up to 75%, freeing up time for in-depth analyses and preventive maintenance, thereby avoiding costly unplanned stops and enabling remanufacturing of large rolling element bearings.
Approach and implementation
The approach is based on machine learning. Large language models have been adapted for technical language within the condition monitoring area. The models are optimized using contrastive learning and can relate technical language to measurement data, for instance, to search for similar historical data/cases, describe a measured signal in plain text, or generate signals that correspond to a described machine damage. This creates new opportunities to automate the analysis of alarms and machine damage and to train analysts.