Machine hearing for industrial fault diagnosis

This paper proposes to apply a machine hearing framework for industrial fault diagnosis, which is inspired by humans' 'listening and diagnostic' capability in identifying machinery faults. The proposed method combines simplified human auditory functionalities with machine learning, aiming to model in a more biologically plausible way. It includes primarily using cochleagram to extract useful time-frequency information in sound signals-representing the cochlea filtering properties in human hearing. Then, a recurrent neural network with long short-term memory layers is constructed to learn and classify the cochleagrams for fault diagnosis-this is to incorporate memory elements in temporal information processing. The proposed method is validated with an experimental study on bearing fault diagnosis using acoustic measurements, while the developed machine hearing scheme could be beneficial to many industrial fault diagnosis applications, e.g., for aeronautical, automotive, marine, railway and manufacturing industry.