Process monitoring of thermal crimping systems using sensorisation and machine learning

During electrical machines manufacturing, quality control is often realised at the end of the line or at selected intervals. The manufacturing process for reliable terminations of machine windings is manual and error-prone, with up to 10% of the strands in terminations found to have incomplete connections. One solution lies in in-process monitoring and inspection of activities involved in the manufacturing. This research proposes the first framework for parallel bimodal quality prediction using real-time RGB images and thermal data from infrared images for process monitoring of thermally crimped wires. The framework demonstrated around 70% accuracy. The developed system simultaneously detects and classifies visible surface defects and hidden subsurface defects at a rate of 0.31 s per sample, while addressing key hindering issues associated with a single sensing modality. The non-destructive defect detection is achieved by combining visual and infrared thermal testing technologies. A Deep Learning Object Detector was trained on RGB images for surface defect identification, and two Long Short-Term Memory neural network models were designed and trained on time-series temperature profiles derived from infrared thermal images to predict hidden (subsurface) defects. Additionally, a novel crimping-score formula is proposed as a complementary diagnostic metric. This formula unifies the effects of individual crimping process parameters into a set of singular values, mapping them to sample quality. While independent of the proposed framework, this score provides manufacturers with an actionable path from parameter prescription to product verification, without interrupting the manufacturing process.