Rolling bearing plays an important role in rotary machines. In rotating machine fault diagnosis, two-channel signals that are recorded from bearing provide sufficient information. It is meaningful to integrate two-channel signals for improving the comprehensiveness and accuracy of status information extracted from the signals. Human auditory nerve system can integrate the binaural information through the mechanism of neurons oscillation and delivery of oscillation. In view of the aspects mentioned above, to simulate the operating mechanism of human binaural auditory system, a double-layer auditory nerve oscillator network (DLNON) model, whose inputs are two-channel vibration signals, is proposed for features extraction and faults diagnosis. By this model, independent component analysis (ICA) is used at the first place to reduce the correlation and increase the independence between the signals; then, outputs of ICA are processed by short-time Fourier transform (STFT) and spectral envelop to reduce the complexity of frequency structure and highlight the formant informant of signal. After that, two results of time–frequency envelop are processed, respectively, by the first-layer oscillator network to obtain synchronous oscillatory period function (SOPF). Finally, information of two SOPFs and two time–frequency envelops is integrated by the second-layer oscillator network, so as to simulate auditory masking effect. The synchronous oscillatory feature function (SOFF) reflected the feature of two-channel signals is obtained by calculating the oscillatory result (SOPF) of the second oscillator network. The performance of DLNON model is evaluated by experiments. The results show that this model can effectively extract fault features, and distinguish fault types, fault severity ratings.
