A reinforcement neural architecture search convolutional neural network for rolling bearing fault diagnosis

The complexity of machinery makes accurate identification of rolling bearing fault signals difficult. Convolutional neural networks (CNNs) have made some progress, but they rely on the expertise of the network designer and the iterative process of optimizing numerous parameters. Therefore, there is an urgent need to develop a method that reduces the threshold for designing CNNs for a given task. In this article, we propose a reinforcement neural architecture search CNN to address this problem. Firstly, we design a neural architecture search algorithm that can generate different types of sub-networks specifically for fault diagnosis tasks. Secondly, we execute a reinforcement learning-based search strategy to discover promising sub-networks. Furthermore, we enhance the performance of the sub-network by improving the optimizer and training parameters. We conduct extensive experiments using two different types of datasets and verify that the proposed method’s fault classification capability is superior to existing methods.