DL-MSCNN: a general and lightweight framework for fault diagnosis with limited training samples

Accurately and intelligently detecting faults is critical to implement smart manufacturing regarding cost reduction and manufacturing system reliability. Recent boosting deep learning technology has been widely used for fault diagnosis (FD) and obtained great success. However, most of them require collecting sufficient data to train the model while it is difficult and even not available sometimes due to the system faults’ variability and randomness. To overcome this issue, this article presents a novel and effective deep model named dual-level multi-scale convolution neural network (DL-MSCNN) for FD with limited training samples. Primarily, it utilizes dual-level multi-scale schemes to extract rich hidden features from raw signals for accurate FD. Firstly, the raw signals are preprocessed by differ-scale average window functions to obtain its grain-level scale representation. Secondly, each grain-level scale feature is fed into a multi-scale cascade CNN (implemented with different kernels) to mine its deeper and detail-level scale representation. Various detail-level scale representations are fused as the fusion features of each grain-level scale CNN (GL-SCNN) representation. Besides, one multi-task supervised scheme (MTSS) is developed to supervise each detail-level scale CNN (DL-SCNN) to learn thoroughly. Finally, the comprehensive features are integrated from all detail-level representations for FD. The authors designed, trained, and tested the proposed method on three real data sets. The comparative analysis confirmed the proposed method’s effectiveness with scarce training samples. In addition, we also proved that the MTSS has a good generality for FD on other multi-scale structures.