Fault Diagnosis of the Electromechanical Coupling System in a Hybrid Heavy-Duty Tractor Based on an HHO-BP Neural Network

Mechanical faults in the hybrid power coupling box of heavy-duty tractors significantly affect power transmission efficiency, system stability, and operational safety. The dual-row planetary gear set used in hybrid electromechanical coupling structures operates under high torque, complex vibration, and long-term cyclic loading, which frequently induces localized failures such as missing teeth, wear, root cracks, and broken teeth. Traditional signal-processing-based diagnostic methods show limited sensitivity to weak and transient features, while Back Propagation (BP) neural networks are prone to local minima due to initialization sensitivity. Therefore, this study aims to develop a high-accuracy fault-diagnosis method capable of identifying multiple mechanical faults in hybrid coupling boxes through improved feature extraction, dimensionality reduction, and global optimization of neural network parameters.The proposed method integrates Harris Hawks Optimization (HHO) with a BP neural network to construct an HHO-BP diagnostic framework. Vibration signals were captured from the coupling box using piezoelectric acceleration sensors at 10 kHz, yielding 420 sets of raw vibration samples. The signals were subjected to wavelet-packet denoising, normalization, and three-level decomposition using the “dmey” wavelet. Eight sub-band energy features were extracted and further normalized. Principal Component Analysis (PCA) reduced the original 50-dimensional feature set to 23 dimensions, retaining primary discriminative information while suppressing redundancy. A three-hidden-layer BP network (128-256-128 neurons, Swish activation; Softmax output for 5 fault categories) served as the classifier. HHO, configured with a population size of 50 and a maximum of 100 iterations, optimized four key parameters including the hidden-layer scaling coefficients and an attention coefficient. Comparative models included the traditional BP neural network and Particle Swarm Optimization-BP (PSO-BP) neural network. The models were evaluated using cross-entropy loss, accuracy, precision, recall, F1-score, confusion matrices, Receiver Operating Characteristic (ROC) curves, and threshold sensitivity analysis.Experimental results demonstrate that the proposed HHO-BP model markedly outperforms BP and PSO-BP in diagnostic accuracy, convergence stability, and robustness. Wavelet-packet decomposition revealed that different fault types exhibit distinct energy-distribution patterns: normal signals concentrated in the second and third frequency bands, missing teeth primarily in the second band, broken teeth in the first and second bands, root cracks predominantly in the fourth band, and wear faults mainly in the second band. These differences confirmed the validity of energy-distribution features as discriminative indicators.Across the three models, the BP model showed substantial misclassification—such as 9.8% of broken-tooth samples misidentified as missing teeth and 11.3% of missing-teeth samples wrongly predicted as broken teeth. The PSO-BP model reduced several error rates, reflecting improved optimization capability. However, the HHO-BP model achieved the highest performance on all metrics. The overall classification accuracy reached 98.26%, improving upon BP and PSO-BP by 6.36% and 5.54%, respectively. Category-level precision for the HHO-BP model ranged from 89.2% to 96.5%, recall ranged from 84.8% to 94.1%, and F1-scores ranged from 90.4% to 93.0%, with wear faults achieving 100% precision and normal conditions achieving 99.5% accuracy. Confusion-matrix analysis confirmed that the HHO-BP model greatly reduced cross-category misclassification, especially for the difficult-to-separate missing-tooth and broken-tooth classes.The training loss curve showed a distinct rapid decline around the 67th iteration, indicating that HHO successfully escaped local minima and guided the BP network toward the global optimum. The final cross-entropy loss approached 0.02, lower than both PSO-BP and BP. ROC results showed that the HHO-BP model achieved the highest Area Under the Curve (AUC) values across all five categories, demonstrating superior discriminative ability. This study presents an effective and robust intelligent diagnostic method for identifying mechanical faults in the hybrid coupling box of heavy-duty tractors. By integrating wavelet-packet energy features, PCA dimensionality reduction, and HHO-based parameter optimization, the HHO-BP model overcomes the inherent limitations of traditional BP networks. The method delivers significant improvements in diagnostic accuracy, convergence speed, and classification stability under identical feature conditions. Achieving 98.26% accuracy and demonstrating superior precision, recall, and F1-scores across five typical fault types, the HHO-BP model proves highly suitable for mechanical-fault diagnosis in electromechanical coupling systems. The approach provides a reliable theoretical and technical foundation for early fault detection and predictive maintenance in hybrid agricultural machinery, supporting the safer and more efficient operation of modern hybrid tractors.