Activation Control Networks with Gradient Sparse Regularization for Mechanical Fault Diagnosis

Mechanical fault diagnosis has become increasingly important as the popularity and complexity of mechanical devices increase. Deep neural networks-based mechanical fault diagnosis methods perform well in terms of accuracy, but their poor interpretability makes them under-utilized in real industrial scenarios. To address this problem, this work proposes activation control networks with gradient sparse regularization. Specifically, firstly, a large number of interpretable features are constructed based on the monitoring signals, and then the activation control networks with gradient sparse regularization is used to adaptively select the effective features and give the feature importance ranking, which can be used as the basis for fault mechanism analysis. Also, the decision-making process of the activation control networks is well interpretable. Experimental results show that the proposed method has favorable fault diagnosis accuracy and interpretability.