An Auditory Measure for Anomaly Detection based on Auto-encoders

We aim at detecting anomalies of several hydro-turbines and electric generators in a power plant based on their auditory signals. Auto-encoders implemented with artificial neural networks are used for this task. For each device, an auto-encoder is trained to describe the audio properties of normal signals of the device. For inference, conventionally, residual spectra between input and prediction produced by auto-encoders are used for anomalies detection. The frame energies of the residual spectra are used for such detection; higher energies are used as indications of presence of anomalies. This approach does not fit well with the industry environment of this work. Audio signals of the devices have quite large variances. Frame energies of the residual spectra are influenced by those variances dramatically, making the conventional approach unable to make robust detection. To deal with this problem, we propose a measure called Peaks-to-Noise Ratio(PNR) to estimate the auditory energies(instead of the physical energies) of residual spectra to determine confidences of anomaly occurrences. Experiments showed that this measure is more robust than conventional ones against the energy variances of the residuals.