Multi-scale dilated convolutional auto-encoder network for weak feature extraction and health condition detection

Health condition detection is an essential method in ensuring the equipment operates safely, which relies on effective feature learning technology. However, traditional feature extraction methods need to design manual feature extractors based on considerable and costly prior knowledge, and it is difficult to achieve robust feature representation. Deep convolutional neural networks often use fixed-size convolutional kernels or filters with a limited receptive field, which can struggle to capture the widely distributed weak signals. To address these limitations, an unsupervised multi-scale dilated convolutional auto-encoder network is designed to capture weak feature representation and execute health condition detection. The method extracts deep features by constructing a multi-scale dilated convolutional auto-encoder, which can expand the range of receptivity without adding several parameters and provide abundant feature representations. Furthermore, the support vector data description is utilized to perform health condition detection, which only requires samples worked in the normal state for model training. To showcase the efficiency of the proposed approach, a set of comparative experiments is performed on the actual dataset of rolling bearings. It can be contended that the proposed method for extracting weak fault features has the potential to enhance the detection results in practical industries.