微震监测预警对保障矿井安全具有重要意义，微震信号降噪和P波初至的准确拾取是微震监测结果可靠性的基础。通过观察海量微震信号，发现单个微震信号的噪声段具有良好的重复性，由此创新性提出基于预测数据的信号降噪思路。具体地，构建了基于自适应增强(Adaptive Boosting, Adaboost)策略提升长短期记忆网络(Long Short-Term Memory, LSTM)的微震信号预测模型，提出了基于模型预测数据与观测数据之差的微震信号降噪方法，研发了长短时窗均值比(Short-Time Average/Long-Time Average, STA/LTA)与赤池信息准则(Akaike Information Criterion, AIC)联合的P波初至拾取方法(S/L-AIC法)，并使用P波初至拾取误差评估和方法比较不同降噪信号和拾取效果。含噪Ricker子波理论测试和耿村煤矿微震数据应用均表明，Adaboost_LSTM模型对于噪声具有很好的拟合性，而对于未进行神经网络训练的微震有用信号拟合性较差，且Adaboost_LTSM模型的信号预测和降噪效果均优于LSTM模型的结果。基于Adaboost_LTSM模型的预测数据几乎能全部去掉微震信号噪声，其降噪效果显著优于小波低频系数重构结果，对非平稳信号的适应性明显增强。小波和Adaboost_LSTM降噪信号能明显提升微震信号P波初至拾取效果，且Adaboost_LSTM降噪信号的P波初至拾取效果更优。S/L-AIC法的P波初至拾取效果优于STA/LTA法和AIC法，兼具了STA/LTA法的稳定性和AIC法的准确性特点，降噪信号S/L-AIC法P波初至拾取误差整体在10 ms以内。综上，矿山微震信号降噪和P波初至拾取方法能为矿山微震监测预警提供重要保障，可尝试推广至天然地震信号降噪和P波初至拾取。

Microseismic early warning is of great significance for ensuring mine safety, where a good denoising and accurate P-wave arrival picking of a microseismic signal is fundamental to the reliability of microseismic monitoring. By observing a large amount of microseismic signals, the noise segments of an individual microseismic signal were discovered to exhibit a good repeatability. Innovatively, a signal-denoising approach was proposed based on prediction data. Specifically, a microseismic signal prediction model was built that enhances the Long Short-Term Memory (LSTM) with the Adaptive Boosting (Adaboost) strategy. Then, a method for microseismic signal denoising based on the difference between model predictive data and observational data was developed. Furthermore, a method for P-wave arrival time picking was proposed, that combines the Short-Time Average/Long-Time Average (STA/LTA) ratio with the Akaike Information Criterion (AIC) (S/L-AIC method). Additionally, the noise reduction and P-wave arrival time picking performance was evaluated by the total cost function of P-wave arrival picking errors. Both the synthetic tests of noisy Ricker wavelet and the microseismic data application of the Gengcun coal mine indicate that the Adaboost_LSTM model has excellent noise fitting capabilities but poor fitting for useful microseismic signals that haven’t undergone neural network training. Furthermore, the signal prediction and noise reduction effects of the Adaboost_LSTM model surpass those of the LSTM model. The Adaboost_LSTM model effectively removes noise from microseismic signals, outperforming the wavelet-based low-frequency coefficient reconstruction methods and significantly enhancing the P-wave arrival characteristics. The wavelet and Adaboost_LSTM denoised signals can improve the P-wave arrival picking results of microseismic signals, and the Adaboost_LSTM denoised signal shows a superior performance. The P-wave arrival picking using the S/L-AIC method is more effective than that using the STA/LTA and AIC methods alone, combining the stability of STA/LTA method with the accuracy of AIC method. Overall, the P-wave arrival picking error of the denoised signals using the S/L-AIC method remains generally within 10 ms. In conclusion, the microseismic signal denoising and P-wave arrival picking methods provide a significant support for mine microseismic monitoring and early warning. Furthermore, this approach has the potential for extending to the denoising and P-wave arrival time picking of natural earthquake signals.