Sparse Filtering With Adaptive Basis Weighting: A Novel Representation Learning Method for Intelligent Fault Diagnosis

Although representation learning (RL) has achieved great success in intelligent fault diagnosis, existing RL methods still have two deficiencies: 1) all learned bases are employed for fault diagnosis, which may degrade the computational efficiency and diagnosis accuracy and 2) it is unable to know which bases are more useful or less useful for fault diagnosis. In this work, we present a novel RL method, namely, sparse filtering with adaptive basis weighting (SFABW) whose architecture is a three-layer neural network. The first and the last two layers are responsible for basis learning and basis weighting, respectively. We formulate a loss function to model such architecture and develop an iterative algorithm to minimize it, also we prove the convergence of this algorithm in theory. Through optimizing the whole network, we are able to obtain a group of bases together with their weights simultaneously. A subset of top-ranked bases with great weights is retained while the rest bases are discarded. The experimental results on a motor bearing dataset and a gear dataset have demonstrated the effectiveness of our method.