Fault Diagnosis in Mechanical Power Transmission System via Improved Particle Swarm Optimisation of Multilayer perceptron classifiers

Abstract

Mechanical power transmission systems are important components in rotating machinery, and its unexpected failures can cause severe economic and safety losses. Accurate and early fault diagnosis is therefore essential to ensure system reliability and enable predictive maintenance. This study introduces a novel fault diagnosis framework that integrates a multilayer perceptron (MLP) with Improved Particle Swarm Optimisation (IPSO) for enhanced classification performance. The method incorporates multi-domain feature extraction, combining statistical, spectral, and nonlinear descriptors, and addresses data imbalance using the ADASYN algorithm. IPSO is employed for hyperparameter tuning and classifier weight optimisation, overcoming premature convergence issues of conventional PSO through adaptive inertia adjustment and random mutation strategies. Experimental validation on the Case Western Reserve University (CWRU) bearing dataset demonstrates the effectiveness of the proposed approach. The model achieves a classification accuracy of 95.3% on training data and 92.4% on testing data, with consistently high precision, recall, and F1-scores across multiple fault categories. Notably, the approach shows robustness against imbalanced conditions and noisy signals, particularly in challenging fault classes such as Ball_014_1 and OR_014_6_1. Comparative ablation studies further highlight the contribution of IPSO-driven optimisation in improving diagnostic accuracy. These results confirm that the proposed MLP–IPSO framework provides a reliable, scalable, and generalisable solution for intelligent fault diagnosis in mechanical power