Methodologies For Fault Diagnosis Of Rotary Machine Based On Transfer Learning

For ensuring the the safe and efficient operation of rotating machinery system with timely maintenance,condition monitoring and fault diagnosis on these key components,such as gears and bearings,are necessary.Hence,fault diagnosis has become one of the key technologies for modern industrial production.With the development of artificial intelligence,machine learningbased intelligent diagnosis technologies have been widely applied in the field of fault diagnosis.In most situations,sufficient training feature samples as well as the independent and identical distribution between training and testing feature samples are required to ensure effective implementation of conventional machine learning methods.Thus,the inevitable indirect measurement problem of target mechanical componet as well as variable working conditions can hinder the accuracy of traditional machine learning-based diagnostic mode and make fault diagnostic task more and more challenging.Transfer learning,by contrast,allows training and testing fault candidates to have different distributions,and can cope well with learning process across them.Technically,transfer learning aims at applying the knowledge learned from relevant task to help build an accurate diagnosis model for target task,which provides a new research direction for fault detection of rotating machinery.Based on thsese,this thesis utilizes transfer learning technology to train robust models for effective fault diagnosis with less labeled target feature samples or under varying working conditions.The main research contents and contributions are as follows:(1)The fundamental principles of transfer learning is introduced,where the aspects of“how to transfer” and “what to transfer” are emphatically investigated.Meanwhile,based on the characteristics of vibration signal under various working conditions,the feasibilities and necessities of utilizing transfer learning are presented,which lays a theoretical foundation for the subsequent transfer learning-based diagnostic strategies.(2)In order to handle the problem of insufficient training samples with supervised information in target task,a transfer learning-based framework,termed as modified least squares support vector machine,is developed by learning from source domain data to help build a precise target diagnostic model.The model is constructed by adding the function that control empirical risk of source domain features into least square support vector machine framework under the principle of structural risk minimization.Moreover,inspired by the work of multitask learning,an improvement of this model(called model parameter transfer-based joint transfer model)is proposed by minimizing the discrepancies of separating hyperplanes between source and target domains,and then transferring both shared and domain-specific parameters over tasks to make use of source domain data to assist target diagnostic task.Experimental results show that the modified least squares support vector machine-based fault diagnostic model can improve the diagnostic accuracy under reasonable help of source domain features,furthermore,joint transfer model is expected to be a valuable technology to boost practical performance of mechanical fault diagnosis with less labeled target training data.(3)To make least squares support vector machine-based fault diagnostic model trained only by source domain features be suitable for target task,a local weighted-based enhanced large margin projection model for domain adaptation is proposed.Here,the white cosine similarity criterion is introduced to compute the weighting factors of all features for the sake of measuring the influence of individual feature sample to global distributional discrepancies.By embedding these weights into projected maximum mean discrepancy during the training process of least squares support vector machine,the enhanced large margin projection model is constructed for fault diagnosis.This model measures and further minimizes the distributional discrepancies to make the two domain features follow the same distribution in reproducing kernel hilbert space.Experimental results show the involvement of sample weighting factor can further reduce domain discrepancies,besides,the enhanced large margin projection model has superiority in utilizing source domain features to assist target fault diagnostic task.(4)For the problem of instability caused by employing different source domains,a novel joint dual-probabilistic latent semantic analysis model,a kind of generative model that can excavate latent features,is proposed for fault diagnosis under variable working conditions.This model can generate the domain-specific latent features and then construct a mapping matrix between source and target domain-specific latent features to deal the problem of low-quality shared latent features caused by large distributional discrepancies.Meanwhile,the Fisher kernel-based least square support vector machine is used to improve classification accuracy.For successful application of probabilistic latent semantic analysis model in the area of fault diagnosis,the bag of fault words model is proposed to extract feature vectors and fault words.The experimental results show the bag of fault words model can provide fault features with strong distinguishing ability among different fault types,the combination of joint dualprobabilistic latent semantic analysis and Fisher kernel-based least squares support vector machine can achieve relatively higher diagnostic accuracy,and furthermore,multiple source domain features can contribute to boosting the stability of diagnostic model.