Intelligent Prototype Generation Method On Machine Fault Diagnosis Through Compressing Large Dataset

In modern industries, machine condition monitoring data have been available for improved maintenance. While big data generally benefits intelligent fault diagnosis performance, the significantly increased data amount inevitably poses high requirements for storage and computation. As a consequence, it is very difficult for the fault diagnosis model to be updated and applied efficiently. In order to address this issue, an intelligent prototype generation method is proposed in this paper. First, a deep convolutional neural network is used as the main framework, which is pre-trained to generate data prototype from large amounts of raw data. Next, the loss of raw data and prototypes is calculated in high-level spatial representations. The model weights are fixed and the prototypes are updated correspondingly. Furthermore, considering the interference of noise and further expanding the variety of the raw dataset, an in-process data augmentation method is proposed, which improves the robustness of prototypes. The performance of the proposed method is validated on a practical dataset collected from a gear failure simulation system. The results show that the proposed method has the ability to compress a large dataset into a lightweight one with prototypes, while achieving similar diagnosis accuracy. In this way, the data storage and computation requirements can be largely relaxed for industrial applications.