Research On Online Anomaly Detection Algorithm Of Streaming Data Based On Deep Transfer Learning

In industrial fields such as machinery manufacturing and intelligent transportation,typical functional components such as rolling bearings have been exposed to harsh working environments such as high temperature,variable speed,and heavy load for a long time,and are prone to various faults.Early fault detection of typical components has become a key technique of Prognostics and Health Management（PHM）.In recent years,with the rapid development of sensor techniques,the status monitoring data of components has begun to show the characteristics of online collection and sequential arrival of streaming data.Real-time and online status anomaly detection of typical components is helpful to achieve non-stop state early fault warning to avoid serious accidents.Due to the restriction of the characteristics of streaming data,the above-mentioned anomaly detection problems still have the following challenges: 1)Improve the accuracy of the detection results,and avoid the decrease in detection accuracy due to low signal-to-noise ratio and insufficient data;2)Improve the reliability of the detection model,and avoid repeated threshold setting and model optimization in the online detection process;3)Improve the robustness of the detection model to avoid false alarms;Aiming at the above three problems,this paper takes rolling bearing,a typical mechanical equipment support element,as the application object,focuses on the characteristics and requirements of early fault detection of bearings,and takes effective use of fault information in different working conditions and even different sources of auxiliary data as the entry point.With the core goal of improving the sensitivity of early fault features,a number of deep transfer learning models have been constructed to solve the problems of late online detection results and high false alarm rate caused by insufficient online data and inconspicuous early bearing fault.The main work and contributions are as follows:（1）Aiming at the data dependence problem of traditional anomaly detection methods,an online anomaly detection algorithm based on deep model transfer is proposed.First,merge original vibration signal,FFT frequency spectrum data and Hilbert-Huang Transform marginal spectrum data into three-channel data.Secondly,use convolutional layer parameters of the VGG-16 model pre-trained on the Image Net image dataset as transfer object,the deep transfer features of bearing data are extracted by model fine-tuning method on basis of the deep model constructed by a small amount of online bearing data,and the online detection model is built.Experimental verification is carried out on the IEEE PHM Challenge2012 dataset,and the results show that the method proposed in this paper can effectively detect early fault while significantly reducing false alarms.（2）Aiming at the problem of low stability of existing deep feature domain adaptation methods,an online anomaly detection algorithm based on joint adversarial training is proposed.This method uses the idea of multi-perspective transfer learning,and uses the feature level and model level for joint adversarial training to transfer information from offline streaming data to online streaming data,and extract public feature representations of offline and online data.First of all,in order to obtain accurate data status labels in the whole life degradation sequence,a robust state division method is proposed by introducing the prior information of degradation process into the anomaly detection process of isolated forest,so as to achieve accurate division of normal state and early fault state under noise interference.On this basis,a new deep domain adaptation method is proposed.The algorithm uses the Domain Adversarial Neural Network（DANN）as basic model,uses the state division results as labels,adopts the mechanism of joint adversarial training,and performs feature adaptation and model adaptation at the same time,effectively improves the public feature representation effect of different working condition data,and builds an end-to-end online detection model.Comparison experiments on IEEE PHM Challenge 2012 dataset and XJTU-SY dataset show that,compared with existing representative early fault detection and diagnosis methods,the proposed method has better detection accuracy and lower false alarm number,which can provide robust detection results for online anomaly detection.（3）Aiming at the problem of late detection results caused by noise interference,on the basis of the model built in step（2）,a joint adversarial training deep domain adaptation model based on the attention mechanism is further constructed.The model uses the idea of multi-level transfer,and on the basis of joint adversarial training to obtain the overall common feature representation,the channel attention mechanism is introduced to amplify the early fault locally to obtain the feature representations that are more sensitive to early fault.First,the isolated forest algorithm is used to accurately assess the normal state and early fault state under noise interference.Then,a joint adversarial training DANN network is constructed using the results of state assessment as labels to improve the effectiveness of public feature representation on the overall level of different working conditions data.Finally,the channel attention mechanism is introduced to extract attention features with stronger early fault characterization ability,and to build an end-to-end online detection model.Comparison experiments on the IEEE PHM Challenge 2012 dataset and XJTU-SY dataset show that compared with the existing representative methods,the method in this paper has significant improvements in the abnormal alarm position and the number of false alarms,and is more suitable for anomaly detection problem of noisy data.As a conclusion,from the perspective of deep transfer learning,the research results of this paper provide a new solution for intelligent online detection of bearing early fault.The detection results are robust,stable,and accurate.Therefore,it has significant theoretical research value and actual engineering application value.