|
Abstract:
|
This work presents the development of a cognitive assistant for industrial maintenance at STMicroelectronics, using the Retrieval-Augmented Generation (RAG) methodology. Industrial maintenance is crucial to ensure the availability of machines, the safety of operators, and the sustainability of production processes. Traditionally, maintenance strategies such as corrective, preventive, and predictive are employed to optimize machine operation. However, the emergence of Industry 4.0 and 5.0 demands more advanced solutions, such as Prescriptive Maintenance (RxM), which integrates artificial intelligence to improve decision-making. In this context, the developed cognitive assistant employs the RAG technique, which combines generative language models, such as Generative pre-trained transformer (GPT), with the retrieval of specific information to enrich the system’s response capability. This method allows the assistant to access technical documents and maintenance records from STMicroelectronics, expanding its knowledge beyond the data with which it was originally trained. The RAG approach helps to mitigate problems of hallucinations and lack of interpretability common in purely generative models, providing more accurate and contextualized responses. The project demonstrates how the combination of generative artificial intelligence with specific information retrieval can create a powerful tool to support industrial maintenance, aligning with the Industry 5.0 principles of putting the human at the center of technological interactions. The evaluation method categorized responses into five levels: completely false (weight –5), partially false (weight –3), factually correct but incomplete (weight 1), fully correct (weight 3), and fully correct with useful reasoning (weight 5). This approach ensured robust assessment by penalizing hallucinations and rewarding high-quality responses. Results from four qualitative evaluation sessions showed the cognitive assistant, tested with 45 question-answer pairs, achieved an average score of 3.27, a normalized average of 82.67 indicating complete, correct and useful information on the majority of tested cases, and an average response time of 32.50 seconds, demonstrating its effectiveness in supporting industrial maintenance tasks. |
