This thesis addresses the problem of anomaly detection in laser-welded windings of hairpin stators, a component in electric vehicle traction motors. Reliable inspection of weld quality is needed, but manual inspection is labor-intensive, subjective, and difficult to scale. The objective of this work was to develop and evaluate a system capable of automatically distinguishing between acceptable and defective welds using 3D point clouds acquired with high-resolution line laser scanners.
The proposed approach combines ideas from PatchCore and PointCore anomaly detection methods and leverages a part of Point-MAE architecture as a feature extractor. A synthetic data generation pipeline was designed to model weld bead geometries and simulate the scanning process, due to limited availability of real training data. The method builds a compact memory bank of descriptors representing normal welds, against which test samples' descriptors are compared to produce per-point and per-part anomaly scores.
Experiments were conducted on a dataset of real welds collected from two hairpin stators. Results show that the method achieves an AUROC = 0.897 and an average precision AP = 0.942. Qualitative analysis confirms that high anomaly scores correspond to visible defects such as craters, spatter, or missing wires. While the system successfully detects the majority of anomalies with TPR = 95.7%, it suffers from a high false positive rate of FPR = 67.7%, limiting its immediate applicability in industrial environments and their strict requirements.
Despite these limitations, the work demonstrates the feasibility of anomaly detection from 3D point clouds in this domain and provides a foundation for future improvements. Potential directions include more realistic synthetic data generation, incorporation of real training data, refined global anomaly scoring, and then later integration into production lines for real-time operation.
|
