An intelligent detection approach for multi-part cover based on deep learning under unbalanced and small size samples

Abstract

The problem of unbalanced and small samples is main challenge to the application of deep learning in fault detection of complex systems. To address this issue, this paper introduces an intelligent detection approach for multi-part cover (MPC) based on auto-encoder Wasserstein generative adversarial networks (AEWGANs) and structure adaptive adjustment convolution neural network (SAACNN). The proposed approach incorporates data augmentation techniques and a detection algorithm to enhance the accuracy of MPC detection. For the data enhancement, a novel AEGWAN model is proposed to enhance the correlation and reduce the difference between the generated samples and real samples, achieved by replacing the random noise vector in the traditional generative adversarial network (GAN) with hidden variables auto-encoded by real samples. In addition, the Wasserstein distance is utilized to substitute for the Kullback–Leibler divergence or Euclidean distance in traditional GAN as the objective function. This substitution helps ease the gradient disappearance and training instability in the training process. For the detection algorithm, although AEWGAN can expand the samples, there are still differences between the generated and real samples due to the limitations of the model. To further ease the effect of the difference for detection accuracy, a novel energy function constraint model is designed for a convolution neural network. On the basis of the new energy function constraint model, a novel SAACNN is created to adaptively select the optimal network structure, which speeds up network training progress and improves the detection accuracy. The effectiveness of the proposed approach is verified by experiments with other models, showcasing its superior capabilities in terms of data enhancement, denoising, and generalization.