CFM56 turbine trench-filler inspection using instance segmentation

This research addresses the imperative need within the aerospace industry for automated diagnostics of critical components, traditionally reliant on human expertise. The purpose is to introduce a novel methodology employing vision-based neural networks to automatically diagnose geometric nonconformities in aeronautical components, specifically focusing on turbine trench-filling elements. The study aims to contribute to the literature by configuring computer vision algorithms tailored to this specific application. The methodology encompasses three key stages: 1) Utilizing a computer vision system with a monochrome camera for image acquisition, 2) Training deep neural networks through transfer learning, and 3) Implementing a stage for automated analysis of non-conformities. Experimental results highlight the efficacy of the YOLO (You Only Look Once) in significantly enhancing automatic diagnosis. Quantitative results showed that the trained model can achieve an overall accuracy of 97.57%, with a specific accuracy reaching 97.89%. The recall rate reached 97.42%, emphasizing the model’s effectiveness. Thus, the presented methodology not only addresses the specific challenges of diagnosing turbine trench-filling components but also opens avenues for further research and development in the field of industrial artificial intelligence applied to aerospace manufacturing and diagnostics.