Research And Implementation Of Infrasound Event Detection Algorithm Based On Machine Learning

Infrasound waves can travel long distances and skirt obstacles due to their low absorption and attenuation rates.Therefore,infrasound monitoring is used to detect natural disasters and is one of the monitoring methods for the CTBT.Infrasound event detection is crucial for monitoring,and detection accuracy is a measure of system performance.To improve the infrasound event detection algorithm,this article proposes various improvement methods based on the characteristics of the infrasound detection algorithm.Firstly,different noise reduction techniques are compared to choose a suitable technique for reducing the signal-to-noise ratio.Secondly,reinforcement learning is used to optimize the hyperparameters of the STA/LTA algorithm.Then,the Euclidean distance clustering is employed to reduce false positive events in the Fisher detection results.Finally,a real-time and historical infrasound event detection system is designed and developed based on the infrasound detection algorithm.The following details the work done:1)This article focuses on improving the signal-to-noise ratio of infrasound monitoring.To achieve this,the article performs noise reduction on a real infrasound event dataset collected from coastal areas and compares the effectiveness of different noise reduction methods.The experimental results show that low-pass filtering can effectively remove noise signals above 20Hz.Spectral subtraction achieves better signal enhancement effects,especially when actual noise signals are present.Additionally,wavelet thresholding can separate low-frequency and highfrequency components,effectively removing high-frequency noise in infrasound.Therefore,in different scenarios,the optimal noise reduction method can be selected based on the characteristics of the noise to improve the accuracy and reliability of infrasound monitoring.2)This article proposes a hyperparameter optimization algorithm based on reinforcement learning Q-learning algorithm to improve the accuracy of infrasound event detection and address the issue of selecting hyperparameters for different array environments and infrasound event detection algorithms.The algorithm regards the selection of hyperparameters of the STA/LTA algorithm as a sequential decisionmaking problem.The intelligent agent determines the next hyperparameter vector by choosing actions with a certain probability and uses the searched hyperparameter vector to run the infrasound detection algorithm and calculate the detection accuracy as a reward value.By continuously repeating the process and maintaining the policy table,the algorithm achieves an automated setting process and obtains the optimal hyperparameter vector that adapts to the environment and the maximum cumulative reward value.In order to improve the training efficiency and stability of hyperparameter optimization,this paper proposes to use asynchronous Q-Learning algorithm to replace the traditional Q-Learning algorithm for hyperparameter optimization.Asynchronous Q-Learning algorithm can train multiple agents at the same time,making the learning process more efficient,reducing training time,and improving the efficiency of hyperparameter optimization.The experimental results show that the detection effect and stability of the algorithm are better than Bayesian optimization on the infrasound data collected in coastal areas,which proves the feasibility of Q-Learning algorithm in the hyperparameter optimization of infrasound event detection algorithm.At the same time,it also proves that compared with the traditional Q-Learning algorithm,the asynchronous Q-Learning algorithm can improve the training efficiency and stability,so as to better optimize the hyperparameter and improve the accuracy and efficiency of infrasound event detection.3)To reduce the false alarm rate of the Fisher detection algorithm,this article proposes a similarity clustering algorithm based on Euclidean distance that effectively reduces the number of false noise signals.It sets a lower Fisher threshold,calculates parameters of significant infrasound events,and clusters them based on similarity rules.The algorithm effectively reduces the false noise signal in the infrasound data of Tonga volcano eruption4)To achieve engineering applications of infrasound detection algorithms,the article implements an infrasound event detection system that effectively detects events by utilizing real-time and historical data to differentiate natural events from background noise.The system also enables real-time infrasound event warning and obtain historical infrasound event data.