Fault diagnosis of the electromechanical coupling system in a hybrid heavy-duty tractor based on HHO-BP neural network

Mechanical faults can significantly dominate the power transmission, stability, and operational safety in the hybrid power coupling box of the heavy-duty tractors. The dual-row planetary gear set has been widely used in the hybrid electromechanical coupling structures, particularly under high torque, complex vibration, and long-term cyclic loading. However, the local failures can be frequently induced during operation, such as missing teeth, wear, root cracks, and broken teeth. Conventional signal-processing diagnosis is often sensitive to the weak and transient features. While Back Propagation (BP) neural networks are prone to local minima, mainly due to the initialization sensitivity. In this study, a high-accuracy fault-diagnosis framework was developed to identify the multiple mechanical faults in the hybrid coupling boxes, according to the feature extraction, dimensionality reduction, and global optimization of the neural network parameters. The Harris Hawks Optimization (HHO) was integrated with a BP neural network in order to construct an HHO-BP diagnostic framework. Vibration signals were captured from the coupling box using piezoelectric acceleration sensors. The 420 sets of the raw vibration samples were obtained at 10 kHz. The signals were then subjected to the wavelet-packet denoising, normalization, and three-level decomposition using the “Dmey” wavelet. Eight features of the sub-band energy were extracted and further normalized after processing. Principal Component Analysis (PCA) reduced the original 50-dimensional feature set into 23 dimensions, thus retaining primary discriminative information with less redundancy. A three-hidden-layer BP network (128-256-128 neurons, Swish activation; Softmax output for 5 fault categories) served as the classifier. The HHO was also configured with a population size of 50 and a maximum of 100 iterations. The key parameters were optimized, including the hidden-layer scaling coefficients and an attention coefficient. The conventional BP and Particle Swarm Optimization-BP (PSO-BP) neural networks were compared on the cross-entropy loss, accuracy, precision, recall, F1-score, confusion matrices, Receiver Operating Characteristic (ROC) curves, and threshold sensitivity analysis. Experimental results demonstrate that the HHO-BP model markedly outperformed the BP and PSO-BP in terms of diagnostic accuracy, convergence stability, and robustness. Wavelet-packet decomposition revealed that the different fault types also exhibited the energy-distribution patterns: The normal signals were concentrated in the second and third frequency bands, the missing teeth primarily in the second band, broken teeth in the first and second bands, the root cracks predominantly in the fourth band, and wear faults mainly in the second band. The energy-distribution features also served as the indicators. Among the three models, the BP model showed substantial misclassification, such as 9.8% of broken-tooth samples misidentified as the missing teeth and 11.3% of missing-teeth samples wrongly predicted as the broken teeth. The PSO-BP model reduced several error rates, indicating improved optimization. By contrast, the HHO-BP model achieved the highest performance on all metrics. The overall classification accuracy reached 98.26%, which was improved by 6.36 and 5.54 percentage points, respectively, compared with the BP and PSO-BP. Category-level precision for the HHO-BP model ranged from 89.2% to 96.5%, recall from 84.8% to 94.1%, and F1-scores from 90.4% to 93.0%, with the wear faults of 100% precision and normal conditions of 99.5% accuracy. Confusion-matrix analysis confirmed that the HHO-BP model greatly reduced the cross-category misclassification, especially for the difficult-to-separate missing-tooth and broken-tooth classes. The training loss curve showed a rapid decline around the 67th iteration. The HHO successfully escaped the local minima in order to guide the BP network toward the global optimum. The final cross-entropy loss was approached 0.02, lower than both PSO-BP and BP. The ROC results showed that the HHO-BP model achieved the highest values of the Area Under the Curve (AUC) across all five categories, indicating superior performance. An effective and robust intelligent diagnosis was presented to identify the mechanical faults in the hybrid coupling box of the heavy-duty tractors. According to the wavelet-packet energy features, PCA dimensionality reduction, and HHO-based parameter optimization, the HHO-BP model overcomes the inherent limitations of the conventional BP networks. The diagnostic accuracy, convergence speed, and classification stability were obtained under identical feature conditions. The HHO-BP model was highly suitable for the mechanical-fault diagnosis in electromechanical coupling systems. The finding can provide a reliable theoretical and technical foundation for the early fault detection and maintenance in the hybrid agricultural machinery, thus supporting the safer and more efficient operation of the modern hybrid tractors.