Fiber-Optic Sensing and CNN–LSTM Time-Series Model with Dynamic Path Generation for Coal-Mine Ventilation Wind-Speed Prediction

Abstract

This article proposes a time-series prediction algorithm model based on fiber optic sensing and deep learning to address the problem of insufficient accuracy in wind speed monitoring and prediction of coal mine ventilation systems under complex working conditions. We have developed a mechanism for fiber optic deployment and signal transmission, designed a real-time monitoring and data acquisition platform, and achieved structured processing of mine wind speed data through feature extraction. At the model level, an improved long short-term memory network and convolutional neural network fusion prediction framework are introduced to model wind speed time series, and combined with dynamic prediction path generation algorithm to enhance prediction robustness. The model forecasts 12 hours ahead using 24-hour inputs, with three CNN layers, two LSTM layers, and attention. The experiment used 120-day data from 48 fiber-optic sensors at 20 Hz, yielding 1.2×10⁸ records plus 950k equipment samples and 17k event logs. After anomaly correction and normalization preprocessing, traditional ARIMA, BP neural network models were compared using metrics such as root mean square error, mean absolute error, and coefficient of determination (R²). Training used a 70/20/10 split with Adam (lr = 0.001, batch = 64); results averaged over 30 runs significantly outperformed ARIMA and BP (p < 0.01). Results showed MAE 0.18 m/s (95% CI: 0.16–0.20), RMSE 0.23 (95% CI: 0.21–0.26), R² 0.94 (95% CI: 0.92–0.95), and delay 1.2 s (95% CI: 1.1–1.3), confirming robustness under complex conditions. The ablation experiment further validated the contribution of feature extraction and dynamic path module to overall performance. The research conclusion shows that the model can effectively improve the wind speed monitoring and prediction level of coal mine ventilation systems, and provide a feasible technical path for intelligent scheduling and safety warning. The results are based on field-deployed data, with 95% confidence intervals reported for all metrics.