Feature-Aligned Stacked Autoencoder: A Novel Semisupervised Deep Learning Model for Pattern Classification of Industrial Faults
Xinmin Zhang, Hongyi Zhang, Zhihuan Song
Abstract
Autoencoder is a widely used deep learning method, which first extracts features from all data through unsupervised reconstruction, and then fine-tunes the network with labeled data. However, due to the limited number of labeled data samples, the network may lack sufficient generalization ability and is prone to overfitting. This article proposes a new semisupervised deep learning method called feature-aligned stacked autoencoder (FA-SAE). FA-SAE takes advantage of the unlabeled data during the fine-tuning process by aligning the feature of both labeled and unlabeled data. In FA-SAE, a new training loss function is designed by integrating the Sinkhorn distance measure of the difference between the features extracted from labeled and unlabeled data through the neural network into the cross-entropy classification loss. The effectiveness of the proposed FA-SAE is verified through its application to two industrial processes, and the application results demonstrated that the proposed FA-SAE has better generalization ability and higher fault classification accuracy as compared to the state-of-the-art methods.