Fusion of enhanced ACGAN and MSACNN for small samples of rotating machinery fault diagnosis methods

Abstract Rotating machinery is widely employed in critical industries such as petrochemical, aerospace, and rail transportation, where its operational condition is vital for ensuring system safety and reliability. However, fault diagnosis in such machinery is challenged by the limited availability of data samples in real-world applications. Traditional diagnostic methods often fail to significantly improve fault detection accuracy under these conditions. To address this issue, this paper proposes a fault diagnosis method tailored for rotating machinery with limited sample sizes, combining enhanced auxiliary classifier generative adversarial networks (EACGAN) with a multi-scale attention convolutional neural network (MSACNN). First, a continuous wavelet transform is applied to convert the one-dimensional vibration signal into a time-frequency spectrogram, generating a time-frequency sample dataset. Next, the model structure and loss function are refined to develop an EACGAN capable of generating high-quality data samples for data augmentation. Finally, a multi-scale convolutional network is introduced, and in conjunction with the dual-pooling efficient channel attention (DP-ECA) mechanism, the MSACNN is designed to enhance diagnostic recognition accuracy. Experimental results from bearing and gear datasets show that the proposed method can accurately identify multi-modal faults in rotating machinery, even with limited sample sizes, significantly improving diagnostic performance.