A Lightning Whistler Speech Recognition Algorithm Based On Single-Channel Blind Source Separation

The Electric Field Detector(EFD)of the CSES-01 satellite has detected numerous Lightning Whistler(LW)events,most of which are interfered by other low-frequency disturbances.To overcome the influence of these low-frequency disturbances on LW recognition results and achieve fast,efficient and accurate identification of LW from massive observation data,this paper proposes a CSES-01 satellite electric field lightning whistler speech recognition algorithm based on single-channel blind source separation(SC-BSS).The relevant results have important significance for further analyzing the spatiotemporal variation of space weather lightning events.The algorithm includes two parts: the model training stage and the model application stage.In the model training stage,the EFD data is preprocessed to obtain speech segments,and then the Mel-Frequency Cepstral Coefficients(MFCC)are extracted as representation features.Finally,the MFCC features are input into the Long Short-Term Memory(LSTM)network,and the LW recognition model is trained using a supervised learning method.In the model application stage,the speech segments of the EFD are first subjected to single-channel blind source separation to obtain two different sub-segments.Then,the MFCC features of each sub-segment are extracted and input into the LSTM network to obtain the classification results of each sub-segment.Finally,the two classification results are combined by decision fusion.Experiments were conducted on a highly interfered dataset in August 2019,and it was found that the accuracy of this method increased by 17.2% compared to the improved method;the recall rate increased by 62.2%;the F1-score increased by 50%;the AUC(Area under Curve)increased by 4.3%;and the time cost was 2.655 seconds,which increased by 0.415 seconds compared to the original method.The experimental results show that this algorithm can effectively improve the noise resistance of the EFD-based LW speech recognition algorithm and achieve efficient and accurate identification of LW from electric field data at a lower computing cost.