Fault Diagnosis Of Time Series Unlabeled Data Based On Self-supervised Contrastive Learning

With the continuous development of mechanical systems in the direction of complexity and intelligence,the intelligent equipment industry has increasingly highlighted the urgent need for the safe and reliable operation of the system.The state detection and fault diagnosis of the mechanical system and key components are of great significance for ensuring stable operation,improving efficiency and avoiding accidents.Deep learning has been widely used in fault diagnosis,and its efficient feature extraction capability has significantly improved the accuracy of diagnosis.However,the number of fault samples is less than the normal state in the mechanical system,which will result in an imbalance of failure data.At the same time,the data obtained is usually unlabeled fault data,and the annotation of data always requires a lot of manpower and time costs.When encountered with the problem of data imbalance and lack of labeled data,the traditional intelligent fault diagnosis model will seriously degrade.It’s difficult to achieve highprecision monitoring and fault diagnosis tasks.Aiming at these difficulties and problems in the research,this paper mainly includes the following contents.Aiming at the problem that one-dimensional vibration signals are susceptible to noise and the imbalance of fault samples,this paper combines wavelet transform and generative adversarial network as the data preprocessing method,and designs a fault data expansion approach.Through continuous wavelet transform,the time domain and frequency domain features are extracted from one-dimensional vibration signal and converted into time-frequency imaged to avoid the problem that failure features are not obvious.For the difficulty of fault diagnosis in unsupervised mode,a self-supervised contrastive learning method Sim CLR is proposed to construct an unlabeled fault diagnosis model,and the validity of the model is verified.The effects of different data augmentations,different self-supervised encoder networks,and different project head modules on feature extraction and classification results are analyzed through experiments.The results show that the use of appropriate data augmentation,encoder network,and the use of non-linear projection head can all improve the diagnostic accuracy.In view of the difficulty of extracting some fault features,the self-supervised contrastive learning Sim CLR has limitations in identifying these types of faults.By combining with the data preprocessing,a multi-instance contrastive learning method for fault diagnosis,WT-MICLe,is proposed.By adding additional comparative learning methods,the accuracy of fault diagnosis in unsupervised mode is further improved.For the semi-supervised case,part of the labeled data is used to assist classification through the task-agnostic and the task-specific way.By conducting multiple experiments,the effectiveness of the method is verified through the bearing data sets of Case Western Reserve University and Paderborn University.