A Study on an Intelligent Fault Diagnosis Method for Substations Based on the Improved VMD-AKELM Algorithm

Abstract

Addressing the shortcomings of traditional substation fault diagnosis methods—such as weak resistance to noise interference, redundant feature extraction, low recognition accuracy for small-sample faults, and the inability to diagnose multiple coupled faults—this paper proposes an intelligent substation fault diagnosis algorithm based on improved variational mode decomposition (Improved VMD) and attentionweighted kernel extreme learning machine (AKELM). Firstly, adaptive penalty factor and mode number optimisation strategies are introduced to enhance the traditional VMD, thereby achieving noise reduction and effective purification of mode components in substation equipment vibration and current fault signals, whilst avoiding endpoint effects and mode aliasing. Secondly, sample entropy and power spectrum entropy features of each optimal mode component are extracted to construct a highly discriminative set of fault features. Finally, a channel attention mechanism is introduced to weight the kernel extreme learning machine, thereby strengthening the weight of key fault features, suppressing interference from redundant features, and enabling the accurate classification of multiple fault types, such as substation short circuits, earthing faults, equipment ageing and poor contacts. Validation experiments were conducted using a simulated substation fault dataset and a measured dataset. The results show that, compared with traditional VMD-ELM, SVM, CNN and other diagnostic models, the improved algorithm presented in this paper increases the accuracy of fault diagnosis by 6.2%–12.5%, the accuracy of small-sample fault recognition by 15.3%, and reduces the average diagnosis time by 41.2%. It exhibits stronger antiinterference and real-time performance, and can be effectively adapted to intelligent fault diagnosis scenarios in substations operating under complex conditions.