The vibration signals of rotating machinery basic components (RMBC) have strong noise and strong time-varying (SNST) characteristics due to the influence of strong noise and strong interference signals, which leads to the problem of limited feature extraction ability of convolutional neural network for SNST time-series signals. Hence, a novel method is proposed using mode spectral array map (MSAM) and dual-channel adaptive scaled convolutional neural network (DASCNN) for RMBC fault diagnosis. First, multiresolution and spectral analyses are performed on the acquired RMBC time-series signals to obtain the spectral modal component signals (SMCS). The SMCS is imaged and array-ordered using Gramian summation angular field (GASF) to construct MSAM-GASF sample set. Second, dual branching CNN channels are constructed to learn different MSAM weight values for high-dimensional feature complementation and enhancement. And Zebra Optimization Algorithm (ZOA) is applied to select feature extraction network block convolution kernel optimally for adaptive scale feature extraction. Finally, a multi-fault experiment is performed on the RMBC. Furthermore, the method is applied to the experimental data to verify its effectiveness. Experiments show that the method proposed has a high fault recognition accuracy. Under strong noise and interference environment, it has higher stability and robustness.

Intelligent fault diagnosis; Rotating machinery basic components; Mode spectral array map; Convolutional neural network; Adaptive scale feature extraction