Feature hybrid fusion-based fault diagnosis of multi-scale and multi-stage industrial processes

The multi-stage production workflows of modern industrial processes pose significant challenges in establishing fault diagnosis models. In addition, the conventional methods struggle to address the inter-stage dependencies and intra-stage multi-scale pattern representations. To this end, this paper proposes a Fault Diagnosis model for Multi-Stage industrial processes (MSFD) with hybrid domain knowledge and cosine similarity fusion strategies, which can discover a comprehensive plant-wise representation through aggregating the stage-wise multi-scale features of multivariate time-series (MTS) process data. Specifically, the MSFD framework consists of two core components. Firstly, an MTS feature extraction (MTS-FE) module independently captures the multi-scale features of the process data within each production stage on local patterns and broader trends. This module can ensure a more refined stage-wise representation. Following this module, a hybrid feature fusion (HFF) module is proposed based on two distinct strategies – domain knowledge and cosine similarity – to integrate the stage-wise features into a unified representation. This dual strategy bridges semantic process constraints of domain knowledge-guided feature weighting with statistical data patterns cosine similarity-based cross-stage dependency learning, enabling a holistic and context-aware fault diagnosis across sequential workflows. The experiments are conducted on a benchmark dataset and a real-world dataset to demonstrate the effectiveness and superiority of the proposed approach.