| Distributed acoustic sensing(DAS)system is based on phase sensitive optical time domain reflectometer(?-OTDR)and widely used in safety monitoring.In real environment,the time-varying and disturbance of vibration sources always exist,which leads to the unknown distortion and impact of DAS sensing signals more easily than in quiet environment or laboratory environment.This means that the actual vibration patterns of signals in real environment are easily covered by the vibration patterns of other interfering vibration sources,and the signal characteristics are easily blurred or even erased by the characteristics of other interfering vibration sources,making it difficult to identify the vibration source in this complex environment with time-varying and multiple vibration sources simultaneously,and the recognition rate needs to be improved.In order to solve this problem,more abundant and more typical features must be extracted to support the pattern analysis of vibration sources.But the existing models cannot fully automatically and intelligently extract structural features and time series features at the same time.Therefore,this thesis proposes a method to automatically and intelligently perform multi-scale structure feature extraction and temporal sequence mining of DAS sensor signals at the same time – a composite algorithm based on improved multi-scale convolution neural network and hidden Markov model(m CNN-HMM).It can mine richer features to support pattern analysis and realize DAS vibration source recognition in complex real environments.The specific work of this thesis is as follows.(1)The existing structural feature extraction methods are optimized to achieve multi-scale structural feature extraction.The existing multi-scale convolutional neural network(MS-CNN)is improved,and the proposed multi-scale convolutional neural network(m CNN)can extract more abundant multi-scale structural features of DAS sensor signals,avoiding the feature omission problem caused by MS-CNN due to ignoring the deep connection between structural features at different scales.(2)On the basis of structural feature extraction,the time series mining link is improved.The Hidden Markov model(HMM)is added on the basis of m CNN to further mine the temporal relationship of event evolution,extract the time series features of DAS sensing signals and realize the vibration source classification.The addition of HMM algorithm on the basis of m CNN fills the lack of time series mining and realizes the simultaneous extraction of structural features,time series features and the vibration source identification of DAS sensing signals.(3)By integrating the m CNN and HMM models,the m CNN-HMM algorithm is implemented and evaluated from the two aspects of feature resolution and classification effect.The training set and test set of one-dimensional vibration events are established using the DAS sensing signals collected in the field.m CNN HMM model construction is realized by training the m CNN model and the HMM model of various vibration events to optimality on the training setthe best.The m CNN-HMM model is compared with artificial+HMM,CNN-HMM and MS-CNN-HMM on the test set from two aspects: a general qualitative,quantitative,and visual analysis method is proposed to evaluate the features,which provides a scientific basis for the confidence level of recognition results.Multiple classification performance indicators are calculated to evaluate the classification effect.Finally,it is concluded that the average prediction probability distance of the structural features extracted by the m CNN-HMM algorithm proposed in this thesis reaches 0.203 on the test set,which is greater than the other three models,and has higher feature resolution.The average classification accuracy reaches 98.1% on the test set after further increasing the time series feature extraction,and the DAS vibration source recognition rate is better than the other three models. |
PDF Full Text Request