To tackle the challenge of convolutional neural network and its integration methods with dcnoising preprocessing methods struggling to effectively extract useful signal features amidst high noise environments and low-quality data, a deep convolutional neural network model based on the Geronimo-Hardin-Massopust multiwavelet decomposition (GHMMD-DCNN) is proposed. The model's concept revolves around deeply integrating the multiwavelet packet decomposition with the convolutional neural network. In other words, this involves the creation of multiple first-level multiwavelet decomposition layers to extract the low-frequency and high-frequency signal components, and these layers arc lined alternately with the convolutional layer. This approach enables the model to extract and learn the useful time-frequency information of the signal on a multiscale basis. The signal decomposition and the feature learning arc executed alternately, and robust feature extraction is realized even under strong noise conditions. Tests arc carried out using aerospace high-speed bearing vibration data under different working conditions. The results show that the proposed model is able to reach stable convergence quickly and the recognition accuracy surpasses 99. 9%. The proposed method showcases superior fault recognition accuracy and stability in the presence of significant noise interference compared to contrast methods, which demonstrates its excellent anti-noise ability. In the test of fewer training samples, the proposed method achieves an impressive average diagnosis accuracy of 91. 19% with only 60 training samples per class. This represents a 13. 19% enhancement over alternative methods, verifying the GHMMD-DCNN's exceptional low-sample generalization ability. © 2024 Xi'an Jiaotong University. All rights reserved.
