Design and application of unsupervised convolutional neural networks integrated with deep belief networks for mechanical fault diagnosis

To overcome the limitations of manual features and obtain the operating characteristics of the equipment in complex operation processes, different deep learning models have been utilized for industrial data, improving classification accuracy yet causing some other limitations meanwhile. In this paper, a deep hybrid model named Stochastic Convolutional and Deep Belief Network (SCDBN), which assembles unsupervised CNN with DBN, was proposed based on the modified CNN. By adding unsupervised components to deep learning methods, proposed model aims to promote the condition that the features extracted by supervised model rely on the training samples and own poor generality. The works of this paper mainly focus on: 1) STFT is used as a signal preprocessing method to create the input of CNN. 2) By means of stochastic kernels and averaging processing, unsupervised CNN is built for general rather than optimal features. 3) A hybrid model is constructed which combines unsupervised CNN with DBN to realize the goal of fault diagnosis. The proposed model is validated by bearing dataset from experiment, and the result has shown that it can obtain high accuracies in diagnosis. Additionally, the features extracted by unsupervised CNN present obvious divergence among different types of bearing data and good generalization ability compared with time domain features.