Band level similarity reduction and deep dual signal fusion for reliable diagnosis of power insulator defects.

Effective monitoring of electrical insulators is critical for maintaining power system reliability, because insulator defects can trigger partial discharges and service interruptions. However, state-of-the-art diagnostic methods are computationally inefficient, require extensive hyperparameter tuning and perform poorly on complex scenarios-particularly perforated insulators without contamination (Signal d). This study addresses these limitations by proposing a novel, hyperparameter-free framework that applies a Similarity Reduction (SR) technique to remove redundant frequency bands through bandpass filtering and feature extraction, then fuses classifier outputs from acoustic (0-20 kHz) and ultrasonic (21-250 kHz) signal components using MFCC features (which outperform FFT, wavelet and time-domain alternatives). By reducing inter-signal redundancy via SR, the framework enables FFT features to be effectively used and allows a transition from a complex CNN-RNN architecture to a lightweight CNN while maintaining high performance. The resulting SRMF + CNN model achieved perfect classification (100% true positive and true negative rates) and dramatically shorter training times, with about a 20 times speedup on perforated cases and a 10 times speedup on contaminated cases compared with the SRMF + CNN-RNN baseline; it also resolved previous failures in Signal d classification, achieving 96.8% accuracy for perforation and 96.9% for contamination. Comparative benchmarks confirm that the SRMF approach outperforms XGBoost, Random Forest and MLP classifiers in both accuracy and efficiency, without increasing sensitivity to class imbalance. These results highlight the SRMF framework as a scalable, generalisable solution for real-time, embedded insulator diagnostics in modern power systems.