Diagnosis of Loosening Faults in Transformer Windings and Cores Based on Temporal Dynamically Weighted Liquid Neural Network (TDW-LNN)

Diagnosing transformer winding and core looseness is critical for ensuring the safe and stable operation of power systems. However, existing methods often struggle with significant noise interference in vibration signals, redundant multi-channel information, and the challenges of modeling long-sequence temporal evolution. In this paper, we propose a temporal dynamic weighted liquid neural network (TDW-LNN) diagnosis method to address these issues. Our approach integrates a multi-channel adaptive weight (MC-AW) module to suppress channel redundancy, a temporal segmented attention (TS-Attn) module that leverages prior knowledge of the 50 Hz fundamental frequency, and an Optimized liquid neuron dynamic modeling module that employs a gated composite activation function to capture temporal dynamics effectively. We evaluated TDW-LNN against architectures using CNN, GRU, and Transformer backbones while retaining the MC-AW and TS-Attn modules. The TDW-LNN model outperformed these alternatives in both diagnostic accuracy and robustness. Across different load conditions, it achieved an average test accuracy exceeding 97%.