Speaker
Description
Unplanned breakdowns in industrial rotating machinery can cost millions per incident, making fault detection a critical maintenance strategy. This work presents an anomaly detection approach based on a 2D convolutional autoencoder (CAE) trained exclusively on healthy vibration data. The publicly available Mechanical Faults in Rotating Machinery (MFRM) dataset is used, comprising acceleration signals acquired under normal conditions and three fault types: unbalance, misalignment, and mechanical looseness. Vibration signals are first resampled into the angular domain via Computed Order Tracking (COT), then transformed into order spectrograms (OS) whose axes, harmonic orders and shaft revolutions, are directly meaningful to domain experts. A Log-Z standardisation enhances the sensitivity of the CAE to subtle spectral deviations. Trained to minimise reconstruction error on healthy OS only, the OS-CAE produces elevated mean squared error (MSE) when encountering faulty inputs. A threshold calibrated on the combined training and validation MSE distributions separates healthy from anomalous samples. Beyond binary detection, a per-order MSE profile of reconstructed spectrograms localises reconstruction failures at specific harmonic orders, providing interpretable visual cues that hint at the nature of the detected fault without relying on post-hoc explainability methods.