Lifecycle monitoring of non-stationary vibrations using autoencoder-based health indicators

This study presents a data-driven Predictive Maintenance (PdM) methodology for monitoring the condition of a rotating machinery operating in non-stationary conditions. Employing Normal Behaviour Modelling (NBM), the approach transforms 1D vibration signals into 2D time-frequency power spectrum images using Short-Time Fourier Transform (STFT). AutoEncoders (AE) and Convolutional AutoEncoders (ConvAE) are trained on these images to learn healthy system behaviour. The models were trained on extracts of signals containing frequency ranges of interest. Anomalies are identified by quantifying reconstruction errors with metrics from image quality assessment. Control charts are then used to establish degradation trends and anomaly thresholds. This AE framework offers a robust solution for machine health monitoring, effectively detecting anomalies across various frequency ranges.