Earthquake Event Detection Algorithm Based On Neural Network Models For Smartphones

Earthquake Early Warning(EEW)is a means of earthquake disaster defense.Traditional EEW involves deploying high-density seismic monitoring networks in earthquake-prone areas to construct an Earthquake Early Warning System(EEWS).With the development of Micro-Electro-Mechanical Systems(MEMS)technology and the widespread use of smartphones,utilizing smartphones for EEW has shown great potential for development.To date,multiple smartphone-based EEW Apps such as MyShake and Earthquake Network have been developed,which use the built-in MEMS sensors of smartphones to record acceleration data and analyze it.Based on this,the effective identification of earthquake events has become an important criterion for evaluating the performance of such smartphone-based EEW Apps.However,due to the uncertainty of the location of the smartphone,it can be affected by human activities,and the accuracy of the built-in sensor is not as high as that of professional equipment.Moreover,the noise level of the collected data itself is relatively high,which poses great challenges to earthquake event detection on mobile devices.In order to improve the accuracy of earthquake event detection on mobile devices,this paper proposes two deep learning algorithms:convolutional neural networks and fully connected neural networks.By training,validating and testing these models,a high-performance model for rapid and accurate earthquake event recognition is obtained.First,data collection is carried out,including noise data when the phone is in a stationary state,non-earthquake data under different human activities,and simulated earthquake event data by adding phone noise data to historical strong seismic records in Japan.The feasibility of this approach is demonstrated by verifying the stability of the noise data(by calculating multiple statistical features and drawing histograms and probability distribution graphs)and comparing the waveforms,amplitude spectra,power density spectra,etc.of earthquake data before and after adding noise.Then,the collected raw data is preprocessed to obtain different feature parameters,and the original data set is divided.On this basis,by comparing and analyzing different feature parameter combinations,the best feature combination that optimizes model performance is determined,forming the required feature data set for subsequent use.Finally,the obtained feature data set is input into different types and structures of models designed,and evaluated and compared through 10-fold cross-validation to determine the model structure with the best performance.The results showed that both neural networks can effectively distinguish between human activities and earthquake events,with high accuracy and short processing times.The vast majority of constructed models had recognition accuracy exceeding 98%,and the precision rate for earthquake event identification and recall rate for non-earthquake events could both reach 99%.This indicates that introducing neural network earthquake event recognition algorithms on mobile devices can effectively improve the accuracy of earthquake event identification and greatly reduce the amount of data transmitted to the server and transmission time,further reducing the processor load on the server and providing technical support for increasing effective warning time.