Research On Fault Diagnosis Of Complex Electromechanical System By Deep Transfer Learning

To ensure the safe and reliable operation of complex mechanical equipment,in recent years,domestic and foreign scholars have developed a variety of equipment fault intelligent diagnosis models based on traditional machine learning and deep learning algorithms.In practical industrial applications,the monitoring data collected under different working conditions of the equipment often obey different distributions,making it difficult for the intelligent diagnosis model trained under the original working conditions(source domain)to be suitable for new working conditions(target domain).Besides,labeling the collected monitoring data relies on the expert knowledge and experience in the specific field,which is time-consuming and labor-intensive.Therefore,after the working condition of mechanical equipment changes to a new one(target domain),it is very difficult to obtain labeled data,and it is impossible to directly train the intelligent diagnosis model in the target domain.Due to the different distributions of data and the lack of labeled data,the traditional intelligent diagnosis model is not suitable for fault diagnosis of mechanical equipment under variable working conditions.In this thesis,for different scenarios of deficiency of labeled data after changing the working condition of complex electromechanical system,two different intelligent diagnosis algorithms are proposed in combination with the deep transfer learning.First,based on the traditional convolutional neural network(CNN)model and the idea of fine-tuning transfer in the field of transfer learning,this thesis proposes a deep feature transfer diagnosis method(DFTD),which realizes the end-to-end diagnosis process in the case of small amount of labeled data in the target domain.In this method,the original time-domain vibration signals are transformed into frequency-domain by using the Fast Fourier Transform(FFT),and the CNN model is trained to further extract the effective features and classify the data.In the target domain,only a small number of label samples are needed to train the classifier of the model to transfer it to be suitable for the target domain.The combination of CNN model and transfer training not only improves the generalization performance of the model,but also saves computational resources.In view of the case that there are no labeled samples in the target domain,this paper proposes an unsupervised dynamic transfer adversarial learning algorithm(DTAL)based on the transfer learning idea of dynamic adversarial.The algorithm is composed of feature extractor,label classifier,global discriminator and local discriminator.To quantitatively analyze the adaptive relative importance of edge distribution and conditional distribution,a dynamic adversarial factor is designed to adjust the loss weight between the global discriminator and the local discriminator.To be able to extract more discriminative data features,an improved feature extractor is developed to process one-dimensional mechanical vibration signals.The combination of dynamic adversarial network structure and improved feature extraction overcomes the transfer problem caused by different data distributions between the source domain and the target domain,thereby obtaining better mechanical fault diagnosis performance under different working conditions.In this thesis,the proposed DFTD and DTAL methods are verified on two kinds of key components of typical rotating machinery systems(wind turbine gearbox and bearing),and compared with the state-of-the-art methods in the industry in terms of feature quality and diagnostic accuracy.The experimental results suggest that the two proposed methods obtain better accuracy and robustness for the diagnosis of mechanical equipment faults under different working conditions.