Gearbox Fault Diagnosis Based on Two-Stream Convolutional Neural Network and Improved Least Squares Support Vector Machine

The present study aims to address the dual challenges of insufficient samples and the difficulty of extracting features in gearbox fault detection. To this end, an intelligent diagnostic framework is proposed, which is based on a Two-Stream Convolutional Neural Network, the Black-winged Kite Optimization Algorithm, and a Least Squares Support Vector Machine. The time-frequency domain multi-scale features of vibration signals are extracted through the dual-channel structure of Two-Stream CNN, combined with the dynamic optimization of network parameters and key feature weights by BKA algorithm, to achieve efficient feature adaptive screening under small sample conditions; the strong generalized classification ability of LSSVM is further utilized to construct an end-to-end fault recognition model. Experiments demonstrate that in scenarios characterized by extreme data scarcity, with a training set ratio of 3: 7, the model demonstrates a remarkable accuracy of 98.94%, which is 15.72% higher than that of the conventional 1D-CNN. Additionally, the model exhibits a diagnostic accuracy of 95.97% and 96.12% for cross-condition tasks, representing an enhancement of 1. 8 5% to 2.35% over models that lack fused BKA. The method has been shown to significantly reduce the dependence of manual feature engineering through the multimodal feature fusion and parameter dynamic optimization mechanism. It provides an innovative solution for high-precision and robust fault detection of industrial gearboxes under complex and variable operating conditions.