Research On Online Anomaly Detection Method Of Streaming Data Based On Deep Feature Representation

Rolling bearings,as the key supporting components of mechanical equipment,are prone to various failures due to long-term operation under complex working conditions,which leads to the deterioration of the working conditions of mechanical equipment.With the rapid development of sensing technology,non-stop failure detection of sequential,large,fast and continuous arrival of streaming data during the operation of mechanical equipment has become particularly important,with clear academic value and application requirements.However,for the problem of online abnormal detection of rolling bearings,the following problems still exist: 1)The detection model cannot adapt to the slight fluctuations of online data,so that the detection model misinterprets the normal sample as abnormal,thereby causing a high false alarm rate;It is more difficult to extract weak signals under the background of strong noise,and due to factors such as working environment and equipment conditions,the data distribution is different.The detection model trained based on offline working condition data performs poorly on online working condition data,thus Reduce the robustness of the detection model.Based on this,this paper theoretically introduces incremental learning methods,attention mechanisms,and transfer learning methods based on the needs and characteristics of online abnormal detection of rolling bearings,respectively to adapt to irregular fluctuations in online data and extract cross-bearing bearings under strong noise background Starting with the online anomaly detection feature,the goal is to reduce the false alarm rate of the test results and improve the online anomaly detection effect of the bearing.The main work and contributions are as follows:(1)In response to the problem that traditional anomaly detection methods cannot adapt to slight fluctuations in online data,this paper proposes an incremental weighted support vector data description(IW-SVDD)algorithm for online rolling bearings abnormal detection.First,train an initial SVDD detection model based on the existing online data,and pre-detect the online data that arrives sequentially.Second,in order to make the detection model adapt to the slight fluctuations of online data,this paper designed a sample state determination State Determination(SSD)strategy,which divides online data into four states: first occurrence of abnormality,continuous appearance of abnormality,disappearance of abnormality,and repeated occurrence of abnormality,and the corresponding weight is given to the sample according to the corresponding state;then,the pre-test results are used Anomalous samples that violate the KKT condition are replaced by the same amount of the earliest samples in the training set of the original detection model,so that the training set is updated online and the model is retrained;finally,the online data is re-trained by the retrained SVDD detection model Re-test and get the test result.The comparison experiment on the IEEE PHM 2012 Challenge data set shows that the IW-SVDD model can effectively reduce the false alarm rate of the detection results while ensuring the accuracy of the detection.(2)Aiming at the problem that the fault feature extraction is difficult under strong noise background and the data distribution is inconsistent across operating conditions,this paper proposes a deep transfer learning method for online abnormal detection of rolling bearings,which is used for online abnormal detection of rolling bearings under cross operating conditions.First,the monitoring signal is processed into a three-channel form consisting of the original signal-marginal spectrum-spectrum;secondly,by adding filters of different sizes to the residual attention module,and using convolution-deconvolution to reconstruct the input Information,a multi-scale residual attention module is constructed to extract attention features with greater ability to represent abnormal data;finally,based on the extracted attention features,a cross-entropy and maximum mean difference are constructed Regularize the constrained loss function to achieve domain adaptation,and use the stochastic gradient descent algorithm to optimize the parameters of the network model,and finally build an end-to-end anomaly detection model.Experimental results on the IEEE PHM 2012 Challenge data set show that,compared with the five representative anomaly detection and diagnosis methods,this method can effectively reduce false alarms without delaying the alarm time.This work is not only applicable to rolling bearings but can also be extended to other types of rotating machinery.It provides new solutions for online health management and condition monitoring of various rotating machinery,and has significant theoretical value and practical engineering application value.