Study And Implementation Of Sample Skew Oriented Infrasound Event Classification Algorithm

Infrasound is sound waves below 20 Hz,which was generated by many events,such as typhoons,lightning,and chemical explosions.Infrasound is not easy to attenuate,so the occurrence of specific events can be detected and identified by monitoring the infrasonic signal.However,due to the different frequencies of events in different categories,there are many samples in the typhoon and micro-pressure categories in the dataset,but few samples in the chemical explosion and lightning categories,resulting in the problem of sample skew in the infrasound event classification dataset.Moreover,the categories of infrasound events need to be labeled by professionals,resulting in a small overall sample size of the infrasound event classification dataset.First,to solve the problem of sample skew,this thesis adopts three methods to expand the sample size of two types of small sample data,and compares the effectiveness of the three data expansion methods of resampling,superimposed noise and random slice in the task of infrasound event classification.The LSTM-based classification model was trained with the augmented dataset to evaluate the effect of augmentation methods.Experimental results show that random slices have the best effect,which can effectively improve the classification accuracy of few-sample categories and improve the average accuracy of model classification.Secondly,this thesis proposes the MIFformer based on multi-source information fusion.the MIFformer divides the encoder into two stages:an independent encoder and a fusion encoder.The independent encoder is used to capture the correlation of the time dimension of a single sequence,and fuse the information of different time steps through the self-attention mechanism to solve the delay problem implicitly.Fusion encoders are used to fuse information between different sequences,resulting in more accurate results than a single source.Use the method of sharing parameters to reduce the number of parameters of the model and alleviate the problem of a small sample size of the data set.To verify the effectiveness of the MIFformer model,this thesis compares the accuracy of the five models of SVM,LSTM,BiLSTM,Transformer,and MIFformer on the infrasound event classification dataset.Experimental results show that MIFformer has the best effect.To verify the versatility of the MIFformer model in time series data classification tasks,this thesis tested 9 different data sets of multivariate time series classification tasks,and found that MIFformer performed better than or equal to the latest methods in 5 tasks,with an average accuracy increased by 2.75%.Subsequently,this thesis proposes a self-supervised learning task applicable to infrasonic signals:data correlation between multiple sequences.A pre-trained model is trained using unlabeled infrasound data and this self-supervised learning task.And then fine-tuned with supervised learning using the infrasonic event classification dataset.The experimental results show that the use of pre-training and fine-tuning can effectively alleviate the problem of small sample size of labeled data in the infrasound event classification dataset.Finally,based on the above algorithm results,this thesis develops an infrasound event classification subsystem,including real-time event classification and historical event classification functions.