An interpretable deep feature aggregation framework for machinery incremental fault diagnosis

Abstract

Existing data-driven Intelligent Fault Diagnosis (IFD) models are trained on an ideal closed set that includes all fault modes. However, new fault modes continue to emerge throughout the service life of mechanical equipment. The current IFD method struggles with learning new diagnostic knowledge. Furthermore, these black-box models also lack interpretability, making the diagnosis results difficult to understand and trust. This study develops an interpretable deep feature aggregation (IDFA) framework. IDFA decouples feature extractors into a shallow feature extraction module and a scalable deep feature extraction module. The former embeds interpretable wavelet kernels as filters into the backbone network layer to help learn interpretable features from vibration signals. The latter will be continuously expanded based on incremental tasks, and enable the model to learn new diagnostic capabilities effectively. Additionally, a Local Grad-CAM method is devised to assist incremental models in achieving better feature localization and visual interpretation of input data. Validated on a benchmark dataset and simulated transmission systems, IDFA method has superior diagnosis performance and explainable capability in incremental diagnostic tasks, providing a new solution for interpretable incremental updates of IFD models.