A Multiscale Feature Learning Scheme Based on Deep Learning for Industrial Process Monitoring and Fault Diagnosis

Considering that industrial data exhibit nonlinearity, high dimensionality and inherent multiscale characteristics, this paper proposes an intelligent industrial process monitoring and fault diagnosis method based on the discrete wavelet transform and deep learning. First, the discrete wavelet transform is used to present the multiscale representation of the raw data. Second, a multiscale convolutional neural network is used to extract the features at each scale, and then the extracted multiscale features are fused by the long short-term memory network to further reduce useless information and retain useful information. Finally, softmax classification is performed. The proposed method has two advantages: 1) the hierarchical learning structure with multiple pairs of convolutional and pooling layers can effectively learn nonlinear, high-dimensional fault features; and 2) the multiscale feature learning scheme can capture complementary diagnosis information at different scales. Detailed comparative studies between the proposed method and conventional methods have been carried out through the Tennessee Eastman benchmark process and the p-xylene oxidation reaction process.