Novel synthetic index-based adaptive stochastic resonance method and its application in bearing fault diagnosis

Stochastic resonance (SR), which is characterized by the fact that proper noise can be utilized to enhance weak periodic signals, has been widely applied in weak signal detection. SR is a nonlinear parameterized filter, and the output signal relies on the system parameters for the deterministic input signal. The most commonly used index for parameter tuning in the SR procedure is the signal-to-noise ratio (SNR). However, using the SNR index to evaluate the denoising effect of SR quantitatively is insufficient when the target signal frequency cannot be estimated accurately. To address this issue, six different indexes, namely, power spectral kurtosis of the SR output signal, correlation coefficient between the SR output and the original signal, peak SNR, structural similarity, root mean square error, and smoothness, are constructed in this study to measure the SR output quantitatively. These six quantitative indexes are fused into a new synthetic quantitative index (SQI) via a back propagation neural network to guide the adaptive parameter selection of the SR procedure. The index fusion procedure reduces the instability of each index and thus improves the robustness of parameter tuning. In addition, genetic algorithm is utilized to quickly select the optimal SR parameters. The efficiency of bearing fault diagnosis is thus further improved. The effectiveness and efficiency of the proposed SQI-based adaptive SR method for bearing fault diagnosis are verified through numerical and experiment analyses.