Robust Unsupervised Anomaly Detection for Surface Defects Based on Stacked Broad Learning System

In modern industrial manufacturing, anomaly detection is critical for improving productivity, reducing maintenance costs, and optimizing product quality. Considering that in actual industrial application scenarios, the collected data in production are often polluted by random events or sensor failure, the high-quality ground-truth anomalous sample acquisition may be expensive and labor-intensive. Facing the variability of the production process and the scarcity and unpredictability of abnormal information, here we propose a robust stacked broad learning system (BLS) for unsupervised surface defect anomaly detection. This method relaxes the requirements of the basic BLS for training data labels, fully utilizes the residual characteristics of the stacked BLS, allocates different weighting factors according to the dispersion of the sample to eliminate the interference of outliers or noise in the process of extracting the feature information of the surface image, and maintains the stability and robustness of the model output based on improving detection accuracy. The accuracy and robustness of the proposed robust stacked BLS are evaluated on MVTec Anomaly Detection, a dataset for benchmarking anomaly detection, and Hardware Mechanical Parts Surface Defect, a dataset obtained from the production line. The experimental results indicate the excellent performance of the robust stacked BLS in unsupervised anomaly detection compared with representative deep learning anomaly detection methods.