Research On Fault Diagnosis Method Of Wind Turbine Gearbox Bearing Based On Data-driven

In the past ten years,with the rapid development of the wind power industry,the safety of wind turbine has become critical,and safety engineering has become the focus of wind turbine research and development.The gearbox is the core component of the wind turbine,consisting of gears,rotors,bearings and other components.Affected by factors such as load and temperature changes,the bearings among them are prone to failure,causing economic losses to the wind farm,and even more serious safety incidents.Therefore,wind turbine gearbox bearings health monitoring during the operation is of great significance for improving operational reliability,maintaining power generation efficiency,and reducing maintenance costs.Under the support of big data,this paper studies the data-driven diagnosis method according to the working conditions,and realizes the automatic detection of fan wind turbine gearbox bearing faults.The details are as follows:In constant working condition,traditional diagnosis methods rely too much on manual feature extraction and require strong professional background knowledge,while Convolution Neural Network（CNN）that automatically extracts features tends to ignore the timing relationship of vibration signal and lose key information.In this paper,multi scale convolution long short-term memory model is proposed.We firstly study how to learn local correlation and global context information from raw vibration signals for better fault identification.The network learns local features through two parallel large convolution kernel and small convolution kernel networks.In addition,the element product is used to fuse local features,which are then sent to the LSTM layer to further learn the global context.Finally,the final features generated by the entire network will contain both local features and long-term context dependencies.In addition,for the problem of poor interpretability of neural networks,we explain from the perspective of feature visualization.Experiments on the CWRU dataset prove that compared with CNN-based and single scale CNN-LSTM networks,this model has great improvements in automatic extraction of fault features,diagnostic accuracy,and robustness.In variable working condition,because the deep learning fault diagnosis method does not carry out domain alignment,while the transfer learning domain adaptive method only aligns the two domains globally without corresponding category alignment,the performance of fault diagnosis is poor.To solve the above problem,a double level confusion adversarial domain adaptation network（CADAN）is proposed in this paper.The network is composed of a feature generator,two label classifiers and an auxiliary classifier.The network uses source domain samples to help the two task classifiers learn.At the same time,an adversarial learning objective function based on two-level domain confusion loss is constructed on the auxiliary classifier.Through adversarial training,the feature generator is driven to generate category aligned features.Experiments on two datasets show that the proposed method can realize the alignment of corresponding categories between two domains,reduce the distribution difference between source domain and target domain,achieve high diagnosis accuracy under variable working conditions,and its performance is better than the latest five domain adaptation fault diagnosis methods.