Research On Fault Diagnosis Method Of Rotating Machinery Based On Deep Transfer Learning

Rotating machinery is one of the most common industrial machinery equipment.Precise monitoring and diagnosis of rotating machinery can bring huge economic benefits to enterprises,not only ensuring the normal operation of equipment,but also greatly improving the production efficiency of enterprises while reducing maintenance and replacement costs.This thesis takes rotating machinery as the research object.Due to the variable operating conditions of rotating machinery and the distribution differences between data under different conditions,deep transfer learning technology is used as the core research method to study the fault diagnosis method of rotating machinery based on deep transfer learning.(1)A diagnosis method employing a multi-feature space adaptive network is proposed to address the issue of poor diagnosis results in rotating machinery faults caused by variations in data distribution resulting from different measuring points.The method is applied to diagnose the faults of harmonic reducers.Firstly,the vibration signals of the harmonic reducer are subjected to continuous wavelet transform to construct a timefrequency map that describes its operational characteristics.Secondly,the data measured by sensors at different locations are divided into multiple source domain and target domain data,which are then mapped into a common sub-network to obtain their common features.Each pair of source and target domain data is mapped into different feature spaces to obtain fault feature representations for different measuring point locations.Then,the adaptive network is used to automatically align the specific feature distribution in the target domain with the knowledge learned in the source domain to learn multiple domaininvariant representations.Finally,the domain-specific decision boundary is utilized to align the output of the classifier and effectively solve the problem of data distribution differences caused by sensor positions.Through the diagnostic experiment of industrial robot harmonic reducer,the recognition accuracy of 99.72% is achieved at 30% full speed,which verifies the effectiveness and feasibility of this method.(2)A diagnosis method is proposed for rotating machinery faults under various operating conditions,aiming to address the issue of poor diagnostic accuracy caused by varied operating conditions in practical engineering applications.The method improves the Transformer and utilizes an encoder structure to automatically extract features.The encoder contains self-attention modules and position coding modules,which can retain the spatial information of data to obtain more comprehensive fault feature representations.The domain adaptation structure is added to the encoder to reduce distribution differences caused by changes in working conditions.The average recognition accuracy was 98.87%and 99.08% respectively on the rolling bearing data set collected in the laboratory and the public data set,indicating that the method can obtain better diagnosis results under different rotation speeds and different loads.(3)A semi-supervised fault diagnosis model for rotating machinery is proposed based on an improved deep masked autoencoder to address the issue of severe shortage of labeled training samples in practical engineering applications.Firstly,the data is processed using continuous wavelet transform.Then,the masked autoencoder processes the unlabeled source domain data samples by masking,and the encoder learns the features of the unmasked parts.The decoder generates a large enough self-supervised sample set from the original data,and the autoencoder serves as a regularization constraint to complete the encoder’s pre-training.The deep fault features are learned by fine-tuning the network model trained on the small amount of labeled fault data.Finally,the domain adaptation module is utilized to reduce the distribution differences between the source and target domains,and an end-to-end rolling bearing fault diagnosis is achieved by the softmax classification layer.The experimental verification on the rolling bearing data set collected in the laboratory shows that the recognition accuracy of 94.58% can be achieved when the number of tags is 8,indicating that the method is feasible in the case of insufficient tag samples.At the conclusion of the thesis,a summary and synthesis of the research work were provided.Additionally,suggestions and prospects for future research directions.