A reusable decoder network penalized by smooth group lasso and its applications to large-scale fault diagnosis of machinery

Representation learning approaches have achieved great success in fault diagnosis of large-scale mechanical data, among which the popular auto-encoder method has developed a series of effective variants. In the existing variants, the encoder network is re-employed to encode feature representations of the data, while the decoder network is directly discarded after training, leading to a regrettable waste of computational resources. Instead of proposing advanced variants of the auto-encoder, this paper explicitly penalizes the decoder network with group lasso, thereby transforming waste into treasure. Specifically, the group lasso constrains the column vectors of the decoder network’s weight matrix at the group level, making them reusable for feature selection. Moreover, a smooth function is utilized to approximate the group lasso to prevent numerical oscillations when computing the gradients. The simulated data and experimental gear data are sequentially used to verify the effectiveness of the smooth group lasso through investigations on two representative auto-encoder variants. The results show that the decoder network penalized by smooth group lasso can be re-utilized to guide selection of a subset of key features for training a classifier, exhibiting an extraordinary feature selection capability.