Neural Network-Based Anomaly Detection Architecture for Bearing Fault Diagnosis

Rotating machinery plays a vital role in modern factory operations; however, the occurrence of component faults resulting from various factors can precipitate system failures and compromise product precision. To mitigate this challenge, we propose an anomaly detection methodology grounded in kernel density estimation (KDE), utilizing the median absolute deviation (MAD) bandwidth parameter to effectively discern unknown faulty signals, thereby alleviating the interference of unexpected data and minimizing performance degradation. Using edge computing devices, we introduce a comprehensive hardware-efficient anomaly detection architecture to overcome bandwidth constraints, data errors, and transmission delays. Demonstrating superior performance across diverse datasets and fault classification stages, our proposed anomaly detection method exhibits remarkable accuracy rates, achieving 94.68% under the National Yang Ming Chiao Tung University-CERES (NYCU-CERES) dataset and 96.59% under the Case Western Reserve University (CWRU) dataset for fault classification. Furthermore, it shows robust specificity and sensitivity, registering averages of 97% and 94.22% on the NYCU-CERES dataset, and 96.78% and 96.82% on the CWRU dataset, respectively. In addition, synthesis utilizing TSMC 40-nm process technology yields substantial reductions in area and power consumption by 25.9% and 16.5% compared to existing methods, respectively.