Research On Unsupervised Learning-based Identification Method For Abnormal Vibration Signals Of Fiber Optic Bragg Grating Array In Smart Metro

During subway operation,long-term interactions between wheels and tracks,tracks and roadbeds can lead to varying degrees of damage to the subway tunnel.Therefore,it is essential to safeguard the operational safety of subway tunnels.With the development of distributed fiber optic Bragg grating(FBG)technology,it has become possible to achieve full-time and entire area monitoring of the structure along the subway line.The FBG sensing array can monitor the vibration response excited by the subway train passing through the monitoring area.Waveform characteristics of the vibration response are closely related to the monitoring area and the state of the train.Therefore,monitoring the abnormal vibration response is significant for early warning of the structure’s status in the monitoring area and the passing train.As an important component of a smart subway,the subway structural health monitoring system can collect a large number of vibration responses in a stable state of the structure during long-term operation.This kind of consistent vibration response can be regarded as a normal sequence.In contrast,anomalous sequences have various forms and random distribution,which makes it is very difficult to build a specific scale sample library of abnormal signals with labels.Therefore,the common supervised classification method has many restrictions to identify anomalous signals and can only be applied in a narrow range.Thus,for the correlation of vibration sequence waveform features in time,a prediction network model is used to learn the waveform features of normal vibration sequences.An unsupervised-learning-based method for detecting abnormal vibration sequences is proposed based on the prediction error as the criterion for discriminating abnormal sequences.The main research content is as follows:(1)An unsupervised-learning-based process for detecting anomalous vibration sequences is designed.The reconstruction error between the predicted and real sequences is used as the judging criterion.A complete anomaly detection process is designed based on the reconstruction error,including calculating anomaly signal indicators and anomaly threshold selection strategy.The LSTM network as the prediction network model is used to pre-test of anomaly sequence detection for the structural health monitoring vibration dataset with similarity to the vibration response of subway bed.The rationality of the unsupervised learning-based anomaly vibration sequence detection principle and the reliability of the anomaly detection process are analyzed.The analysis for poor results in identifying abnormal vibration sequences reveals that the simple LSTM network prediction model has a low nonlinear fitting and generalization ability for normal sequences containing different noise levels.(2)A CNN-LSTM-AM based prediction network model is proposed to address the phenomenon that the prediction network model has insufficient sequences’ nonlinear fitting and generalization ability.The effects of the primary hyperparameters,such as the number of convolutional kernels,LSTM stacking layers,the number of LSTM cells,and the prediction delay,are revealed to establish the optimal network prediction effect on a uniform pre-test dataset.To evaluate the performance of the established optimal network models,it is compared with RNN networks,LSTM networks,and AE network models.The effects of the four anomaly indicators on identifying anomalous data are quantified based on the AUC metric.Finally,the advantages of the CNN-LSTM-AM network are verified by the AUC metrics in detecting abnormal signals on the selected pre-test dataset.(3)Under the excitation of the subway,the characteristics of structural vibration sequences,waveform characteristics,and causes of abnormal vibration sequences are analyzed.An endpoint detection method based on common statistical features is proposed to intercept the structural vibration sequences under the excitation of the subway.The CNN-LSTM-AM method that has been verified in the pre-test is used to process the vibration sequence data of subway beds.Aiming at the insufficiency of the samples with abnormal labels collected,simulating three types of anomaly sequences with different degrees of abnormality according to the characteristics of real abnormal sequences.For three types of abnormal sequences,the performance of different abnormal index calculation methods is analyzed and evaluated by the AUC combined with TPR and FPR.The proposed anomaly sequence detection method and other typical anomaly sequence detection methods are evaluated by Accuracy and F1-score indicators.The results show that the anomaly detection method based on CNN-LSTMAM can significantly improve the recognition of partially continuous sequences when they deviate from the normal state.