Anomalous sound detection with timbre separation

A computer-implemented method for training an anomalous behavior classifier, comprising:

• training a target autoencoder, using training data that represents normal operation of a target device, to generate a reconstructed target device waveform from a mixed input waveform;
• training a first environmental autoencoder, using training data that represents a first set of unrelated sound events, to generate a reconstructed unrelated waveform from a mixed input waveform;
• and
• training a neural network classifier, using reconstructed target device waveforms and unrelated waveforms based on the training data, to generate an output that classifies whether an input waveform represents anomalous behavior of the target device.

A system for detecting anomalous behavior, comprising:

• a hardware processor;
• a memory that stores computer program instructions, the computer program instructions executable by the hardware processor to cause the hardware processor to:
• reconstruct a waveform of a target device from an input waveform using a target autoencoder;
• reconstruct a waveform of unrelated sound events from the input waveform using an environmental autoencoder;
• classify the input waveform to determine that the input waveform is produced by anomalous behavior of the target device using a classifier, based on the reconstructed waveform of the target device and the reconstructed waveform of the unrelated sound events;
• and
• generate an automatic response to the anomalous behavior.