This paper proposes a fault diagnosis method based on an improved dung beetle optimization algorithm (IDBO) to solve the problem of low accuracy of mechanical fault diagnosis of high-voltage circuit breakers. Tent chaotic mapping, a golden sine strategy, and adaptive t-distribution perturbation are incorporated to optimize a deep hybrid kernel extreme learning machine (DHKELM).

Firstly, this paper takes a TY-1S-12/630-16 single-phase vacuum high-voltage circuit breaker as the research object and builds a platform for collecting high-voltage circuit breaker closing vibration signals. Five operating conditions are simulated: normal state, cushion spring fatigue, base looseness, insulator looseness, and drive shaft jam. The laser vibrometer’s sampling time and frequency are set to 1 000 ms and 78 125 Hz. 60 groups of samples for each condition of the high-voltage circuit breaker are collected, totaling 300 sets of samples.

Secondly, the successive variational modal decomposition (SVMD) is used to decompose the acquired signals, the seven IMF components with different center frequencies are obtained after decomposition, and the power spectral entropy of each IMF component is extracted to construct the feature vector matrix. Data dimensionality reduction of the feature vectors is carried out using the t-distribution-stochastic neighborhood embedding algorithm (t-SNE) to obtain 300 by 3-dimensional feature vectors. After dimensionality reduction by t-SNE, the samples of the same state show clear clustering characteristics, while the samples of different states are separated in the mapping results of t-SNE. Hence, the problems of information redundancy and high- dimensional data are avoided.

Then, by introducing three optimization strategies-fusion Tent chaotic mapping, golden sine strategy, and adaptive t-distribution perturbation, the improved dung beetle optimization (IDBO) algorithm is proposed. The IDBO algorithm optimizes the parameters of the DHKELM for constructing the IDBO-DHKELM high-voltage fault diagnosis model. The unimodal and multimodal functions from the CEC2005 test suite are selected for performance testing. The improved IDBO algorithm is compared with traditional PSO, WOA, and DBO algorithms, verifying its superior convergence speed, optimization precision, and stability in finding the optimal solution.

Finally, a platform is built to simulate mechanical failures of high-voltage circuit breakers. The fault diagnosis results show that the proposed method’s fault diagnosis accuracy reaches 98.33%, and the average accuracy of the classification of the DHKELM model is improved by 11.67%, 5.83%, and 3.33%, respectively, compared with that of the traditional SVM, ELM, and CNN models. The DHKELM model improves the average classification accuracy by 9.16% and 7.5% compared with PSO-DHKELM and DBO-DHKELM models, and the precision rate, recall rate, and F_1-score are greatly improved.