Multi‐Modal Feature Fusion‐Based Method for Acoustic Signal Analysis of Fastener Loosening in Saturable Reactors

The loosening of fasteners in converter valve saturable reactors presents a significant and often concealed operational hazard. To address the deficiencies in effective feature extraction from acoustic‐vibration signals and their incomplete correlation, this study proposes a novel fault diagnosis methodology for these fasteners based on multi‐modal feature fusion (MFF). Initially, the Gramian Angular Summation Field (GASF) extracted time‐domain features reflecting global information, whereas the continuous wavelet transform (CWT) obtained time‐frequency domain features capturing local characteristics. These were subsequently integrated within a 2D feature fusion framework through vertical channel‐wise concatenation. Secondly, the methodology employed a gated recurrent unit (GRU) to fuse the temporal sequences of acoustic signals, which were then combined with the 2D features using convolutional neural networks (CNNs) to achieve progressive multi‐source feature fusion from local to global scales. Furthermore, the model incorporated a multi‐head attention (MA) mechanism to specifically enhance fault‐indicative features, culminating in a deep fusion of time‐domain, frequency‐domain and time‐frequency domain information. Finally, this study acquired acoustic data using a bespoke single‐valve‐layer experimental platform, supplemented by field data from the ± 800 kV Qingnan converter station. These datasets facilitated the analysis of overall acoustic‐vibration signals from saturable reactor fasteners under both normal and faulty conditions. The proposed MFF model achieved a maximum detection accuracy of 96% for noncontact loosening faults. This work highlights the feasibility of acoustic‐based fault detection in saturable reactor fasteners and provides practical guidance for enhancing predictive maintenance and operational safety in HVDC converter stations.