UC 2.1 Maintenance Ticket Analysis
The progress of this use case is a joint endeavor between ABB and Schwermetall, two integral partners of the project. Their collaborative efforts represent a dynamic synergy aimed at refining and expanding the scope of the project’s capabilities. As ABB and Schwermetall continue to innovate and contribute their expertise, this ongoing partnership exemplifies their shared commitment to driving progress and achieving the project’s objectives.
This use case focuses on operators who are responsible for searching and identifying Technical Reports from all parts of the metal manufacturing plants. A semantic analyzer engine allows “Technical Reports (TR)” to be analyzed for their meaning and not just for the presence of certain keywords. By using large language models, we can improve the accuracy of the results and provide more relevant information to Users.
Impact: The semantic engine will allow users to search for documents and information to make domain knowledge at Schwermetall more accessible. Comprehensive knowledge management for the specific domain knowledge.
Desired Outcome: Develop a semantic analyzer engine for domain knowledge using large language models. More relevant and faster analysis of problems in the plant and easier familiarization and semantic assistance for new employees.
User Interface
A search page allows the semantic search for tickets that are similar to an entered text. The tickets found by the engine are presented in a sorted table.
The user can give feedback for each ticket via a colored dropdown menu. The explanation of the relevance of a selected ticket is expressed by the accentuation (highlighting) of the relevant words.
ML Model Summary
This use case relies on the technique of vector embeddings to solve semantic search queries and explanation demands.
A language Model, more precisely, a BERT (Bidirectional Encoder Representations from Transformers) model, pretrained on German Language, is used to create the embeddings for all available technical reports. The embeddings are stored in a database, which is deployed as part of the sematic search application.
Explainer Component Overview
The relevance scores for the individual words of a ticket text can be computed using the embeddings contained in the database when the ticket is compared to a query text. The prototype uses these relevance scores for the explanation visualization by colouring the word background using colour intensities that correspond to the weights.
Feedback Component Overview
The feedback component provides the option to rate a search result as a good, moderate, or poor match to the search query. Once user has rated a ticket the visibility of the value is supported by the background color of table cell showing the rating.
- Green: good match
- Yellow: moderate match
- Red: Poor Match