New Framework Boosts Transparency in Sensor Diagnostic AI Pipeline Search

Automated machine learning (AutoML) is increasingly used to design diagnostic pipelines for industrial sensors, reducing the need for manual configuration. However, current AutoML reports often only detail successful or fitted trials, omitting a significant number of generated candidates that were invalid, pruned, or otherwise not evaluated. This lack of transparency makes it difficult for reviewers to understand the full scope of the search process, including budget allocation and adherence to constraints. To address this, researchers propose "candidate-fate accounting," a novel audit framework designed for diagnostic search traces. This system meticulously records every candidate generated during the search, assigning each a specific "fate" such as invalid, pruned, or fitted. It uses hashing to merge repeated observations, performs legality checks, and provides rationales for budget decisions, ensuring a complete and auditable record. Experiments on three bearing-diagnostic datasets demonstrated the framework's effectiveness in identifying invalid candidates and revealing numerous candidates previously omitted from standard reports. This approach maintains competitive diagnostic performance while significantly enhancing the transparency and accountability of automated pipeline search.