Research Of Seismic Signal Classification Based On Entropy Feature Extraction And Graph Neural Network

Earthquake signal classification research refers to the technical study of classifying and identifying earthquake signals based on waveform characteristics extracted from seismic records.It is an important component of seismology research and holds significant theoretical and practical value.The most commonly used methods for earthquake signal classification and recognition include frequency spectrum analysis,wavelet transforms,and machine learning algorithms.However,during the collection of seismic signals,challenges such as complex environmental conditions and irrelevant interference often result in low signal-to-noise ratios.Addressing the issue of noise interference and improving the accuracy of earthquake signal classification is an urgent problem in seismology research and related applications.Signal entropy is a measure that describes the complexity of a time series and possesses strong resistance to noise.It has achieved remarkable success in applications such as medical signal classification and mechanical fault diagnosis.Graph neural networks(GNNs)have become popular deep learning methods in recent years due to their fast computation and strong feature extraction capabilities.Particularly,GNNs excel in handling graph data structures with strong relational connections,enabling the extraction of hidden information.Based on the existing literature,no application has been found that combines entropy feature extraction methods with graph neural networks for earthquake signal classification research.Therefore,this paper proposes a novel approach for earthquake signal classification by combining entropy feature extraction and graph neural networks.The main content is as follows:(1)Explored four methods for feature extraction of signal entropy: approximate entropy,permutation entropy,fuzzy entropy,and singular value decomposition entropy.Waveform features were extracted from two types of seismic events: natural earthquakes and artificial explosions.A detailed comparative analysis was conducted,revealing that the four entropy values of artificial explosion signals were significantly higher than the corresponding entropy values of natural earthquake signals,indicating their distinguishability.Additionally,a non-basis function signal decomposition method called ensemble empirical mode decomposition(EEMD)was further employed to decompose seismic signals from the two event types.Experimental studies revealed that the signal entropy feature extraction of the third component(IMF3)obtained from EEMD exhibited better classification boundaries,thereby enhancing the accuracy of distinguishing natural earthquake events from artificial explosion events.(2)Constructed graph data structures for natural earthquake signals and artificial explosion signals,and designed a graph attention neural network model for earthquake source type classification.Firstly,considering that earthquake signals are simple time-series signals with strong connections between adjacent sampling points,a novel classification method using a graph attention neural network model was proposed.By utilizing the graph attention neural network to capture crucial information from time-series signals,the accuracy of earthquake source type classification was further improved.Experimental results demonstrated that the proposed classification method based on entropy feature extraction and graph attention neural network model outperformed traditional classification methods,and it exhibited robustness against certain levels of noise interference.(3)Based on graph neural networks,an improved graph neural network model incorporating random walk was proposed.Earthquake signals are one-dimensional timeseries signals,and the graph data structure constructed from these one-dimensional signals exhibits a linear structure where the central node has at most two adjacent nodes.Consequently,during the training process using graph neural networks,it may not fully extract the hidden information beneficial for classification from earthquake signals.To address this issue,this study introduced the K-order neighborhood aggregation graph convolutional network model to enhance the feature extraction capability of nodes for earthquake signals.Additionally,considering the long duration of observed earthquake signal waveforms compared to the significantly shorter duration of effective seismic waveforms corresponding to specific events,the presence of abundant invalid signals or noise in the earthquake signals was acknowledged.Therefore,a random walk graph sampling algorithm was further introduced to optimize the classification results of the model.Experimental results demonstrated the effectiveness of this improved method and its superiority over current state-of-the-art models in comparative experiments.