Research On The Fault Diagnosis Method For Gears Of Wind Turbine Gearboxes Based On Adversarial Network

The gearbox is one of the key components of the wind turbine transmission system,which is subjected to large dynamic loads in the actual operation process.The complex operating environment easily leads to the failure of key components such as sun gears and planetary gears,and even causes safety accidents,so the research on the fault diagnosis of gears of wind turbine gearboxes is of great practical significance.Due to the complex operating environment of wind turbine gearboxes,the complex and variable loads,and the difficulty in obtaining fault samples,which lead to unbalanced data samples,asymmetric data features,small sample capacity,and insufficient sample labels,the data set imbalance problem and the small sample data problem become the two main reasons limiting the gear fault diagnosis and classification of wind turbine gearboxes.A fault diagnosis method for wind turbine gearboxes based on adversarial networks is proposed in this paper,and the main research contents are as follows.(1)A fault diagnosis method based on the adversarial network for the gears of wind turbine gearboxes with unbalanced data sets is proposed to address the problem that the fault characteristics are not obvious and difficult to diagnose due to unbalanced data sets.To begin,the fault signal is characterized by binary vectorized coding,and kurtosis is used as a perceptron to evaluate the fitness of the coded individual.Then,the coding string crossover of genetic probabilities is performed on individuals in the next-generation population by the expectation value of the deterministic selection operator.After that,a Gaussian approximation variation is performed based on the signed mean of the normal distribution to produce individuals with higher fitness to proceed to the next iteration.Finally,a logistic regression auxiliary classifier is used to establish nonlinear decision boundaries and provide a gradient penalty for the discriminator to complete the fault diagnosis and classification tasks in an independent fully connected layer.(2)A small-sample fault diagnosis method based on the adversarial network is proposed for the small-sample problem of the insufficient number of wind turbine gearbox fault samples or missing sample labels.First,a weak learner is built to optimize the efficiency of deep learning by converting the parameter space into a function space to achieve optimal convergence.Then,the nonlinear decision boundary is established by the K-Nearest Neighbor algorithm,and the Mahalanobis distance is introduced to measure the density distribution of the probability space continuously.Finally,the two-stream convolutional networks were constructed to score and fuse the data,providing decision support for signal stacking and diagnosis.The experimental results show that the proposed method can learn the characteristics of signals from small sample data sets,predict the distribution of fault signals,and provide an effective data stack for small sample data sets of wind turbine gearboxes,thus improving the diagnosis accuracy and classification accuracy.(3)To deeply explore the multidimensional fault feature information of wind turbine gearboxes and maximize the feature enhancement and sample expansion,a quantum adversarial network model is proposed.First,the data of the wind turbine gearbox is mapped into the feature space corresponding to the quantum polymorphs to construct the quantum state signal,and the features of the signal are represented in the reduced density matrix,and the output is mapped through a tensor network.Then,the probability distribution of the quantum state signal features is compressed and sampled to retain the most informative qubit signal.Finally,the entanglement entropy is minimized by using the classification label as a feature quantity,and the signal of the quantum state is compressed and spatially replicated so the resulting tensor is decomposed and adjusted to discriminate and output in a zero-sum game by adversarial training of the network.The experiments show that the proposed method can improve fault diagnosis and generalization capability by data mining of gear signals.