Gearbox Fault Diagnosis in Variable Speed Conditions Using Sensor Fusion and Deep Transfer Learning-based Multi-Channel CNN

This paper presents a novel neural network structure designed for the classification of individual and combined faults in wind turbine gearboxes. This structure features a deep multi-channel convolutional neural network (MC-CNN) that facilitates simultaneous processing of images captured by sensors in different orientations. Therefore, the developed fault diagnoser utilizes a broader range of features for training the network, thereby enhancing the accuracy of fault classification. Subsequently, by incorporating a generalized loss function based on Maximum Mean Discrepancy (MMD)—a distance metric for comparing probability distributions—along with Deep Domain Confusion (DDC) transfer learning, the proposed method not only detects fault labels but also adapts to features extracted from vibration images at varying speeds. This adaptability preserves the network's functionality under varying load conditions in wind turbines and ensures its effectiveness in low-data scenarios. For testing and evaluating the performance of this method, a wind turbine laboratory testbed was constructed to generate real data in various fault classes. The results demonstrate that the developed method achieves an average accuracy of ۹۹.۹۵% on validation data (used in transfer learning) and ۸۵.۱۵% on test data (have no role in the training process), confirming its superior performance compared to existing approaches.

کلیدواژه ها:

wind turbines ، Fault Diagnosis ، Vibrations ، Gearbox faults ، Neural Network ، Deep convolutional network ، Deep Domain Confusion Transfer Learning ، Maximum Mean Discrepancy Loss Function