Research On Data Realtime Sensing Methods On Edge Computing Microseismic Detection

With the development of big data and Internet of Things technologies,wireless data acquisition technology has shown an explosive growth trend.The current mainstream technology is to upload the collected data to the cloud for storage,but this way of processing leads to a large amount of invalid data in the data stored in the cloud,resulting in a waste of cloud storage resources.To solve this problem,this thesis takes coal mine microseismic data acquisition as the research background,and identifies the data at the edge of data acquisition,thus reducing the proportion of invalid data and saving storage resources and communication bandwidth for the cloud and the terminal.The main challenges of microseismic signal recognition at the edge are the accuracy,real-time performance and parameter scale of the algorithm.For example,the traditional STA/LTA algorithm has good real-time performance but low accuracy,while the microseismic signal recognition algorithm deployed in the cloud has high accuracy but poor real-time performance and large parameter scale,which cannot meet the demand of microseismic signal recognition at the edge.Based on this,this thesis mainly studies the design and optimization methods of microseismic signal recognition algorithm on low-power edge platform.The main work of this thesis is summarized as follows:Firstly,the microseismic signal itself and its background noise are analyzed in time domain,frequency domain and time-frequency domain,and the difficulties of microseismic signal recognition are summarized.At the same time,a data set expansion method based on wavelet transform is proposed to deal with the small sample problem in microseismic signal data.Secondly,a microseismic signal recognition algorithm based on artificial feature extraction is established.By extracting 19 features from time domain and frequency domain of microseismic signal to form feature vector,and then compressing feature vector using PCA algorithm,feature vector dimension is reduced to 3 while retaining99.9% of original feature vector information.The compressed feature vector is input into the ensemble learning model GBDT for learning and training,and finally a model PCA-GBDT with an accuracy of 92.92% and a real-time performance of 18.97201 ms in local environment is obtained.Thirdly,a microseismic signal recognition algorithm based on automatic feature extraction is established.Firstly,based on the idea of feature pyramid model in computer vision field,a convolutional network model FPN-CNN is established.Considering the multi-scale features of microseismic signal itself,FPN-CNN is parallelized and pruned to obtain model PFPN-CNN.PFPN-CNN has an accuracy of94.21% and a real-time performance of 172.75023 ms in local environment.Fourthly,quantization-aware training is performed on PFPN-CNN to convert model parameters from float32 to int8.Because the mapping relationship between parameters is perceived online during model training,the accuracy of converted model only drops by 0.98%,while real-time performance improves to 4.62091 ms.The quantized model is deployed on edge platform for microseismic signal acquisition for experiment and analysis.PFPN-CNN has real-time performance of176.014727 ms and data compression ratio of 9.84.While PCA-GBDT model has real-time performance of 121.92761 ms and compression ratio of 9.42.