Intelligent recognition and classification of microseismic events based on machine learning techniques

Abstract

Microseismic monitoring system plays an important role in the monitoring, early warning, and prevention of mining-induced ground pressure disasters. These systems integrate functions such as collecting, locating, analyzing, and interpreting seismic activities induced by microcracks within rock masses. However, with the generation of a large amount of monitoring data, the rapid, accurate, and real-time identification of different types of microseismic events has become a fundamental requirement for disaster prevention and control, as well as for the construction of smart mines. This paper proposed different automatic identification and classification models for microseismic events using machine learning technology, based on data mining and analysis. The aim is to improve the efficiency and accuracy of microseismic data analysis, thereby providing a solid foundation for geostress disaster management and the advancement of smart mining systems. Firstly, microseismic data collected by monitoring systems from three different mines in Shaanxi Province, China, were processed into raw waveform images, with each event consisting of six sub-graphs forming a sample graph. Based on expert experience and manual identification, three sample databases including four types of events—mining microseisms, blasting, drilling, and noise—were established, resulting in diverse datasets. Subsequently, this paper employed various advanced algorithms and models to automatically extract features from different waveform images and construct an intelligent identification system for microseismic events. Specifically, methods combining Histogram of Oriented Gradients (HOG) features with Shallow Machine Learning (SML), Convolutional Neural Networks (CNN), and transfer learning-based deep learning models such as ResNet-18, MobileNet-V2, and Inception-V3 were selected. Experiments were conducted using the three sample databases, and the classification performance and recognition accuracy of different models were compared. The results showed that on the test dataset A, the overall accuracy of the HOG-SVM, MS-CNN, ResNet-18, MobileNet-V2, and Inception-V3 models reached 0.971, 0.974, 0.981, 0.982, and 0.987, respectively. Comparative analysis of the models revealed that deep learning models, especially Inception-V3, outperformed others in terms of accuracy, demonstrating the potential of deep learning in classifying microseismic events. The HOG-SVM method demonstrated the fastest processing efficiency. The MS-CNN model achieved an effective balance between recognition efficiency and classification accuracy. This study introduces an innovative, efficient, and precise approach for intelligently identifying microseismic events. It offers a comparative analysis of machine learning methods, aiding users in choosing the right algorithms for their tasks. The research expands beyond microseismic and blasting event identification to include drilling and noise events, enhancing the intuitive and precise recognition of waveforms. The models' adaptability across various mining data showcases their potential to boost mine safety and operational intelligence in real-world scenarios. The application of machine learning methods and computer vision technology helps achieve intelligent recognition and classification of microseismic events in the microseismic monitoring system of mines. This enables the rapid and accurate generation of classification results, effectively reducing the workload and misjudgment rate of manual identification of microseismic events. At the same time, it provides interpretable evidence for the mine disaster warning system and timely alerts for potential seismic activities.