Abstract: To overcome the limited ability of existing methods to detect trend-type pseudo data, this study develops a recognition method integrating grey relational analysis (GRA) with a hybrid model. GRA is employed to select key environmental factors, such as temperature, humidity, wind speed, and air pressure, which are highly correlated with methane concentration. Combined with wavelet threshold de-noising, support vector machine (SVM) anomaly detection, and heartbeat data-compression mechanisms, a multi-stage data cleaning chain is established. Subsequently, an auto regressive integrated moving average (ARIMA) model and a long short-term memory (LSTM) network are connected in series. The LSTM captures nonlinear temporal features, while ARIMA performs residual trend correction. A dynamic root mean square error (RMSE) threshold then automatically identifies pseudo data. The experimental results show that the proposed model has a recognition accuracy of 87%, RMSE of 0.09, mean absolute error (MAE) of 0.04, and an average response time of only 33 ms on 500 samples. In the detection of trend-type pseudo data, the pseudo data recognition rate is 85% and the data recovery rate is 93%. Compared with auto encoder (AE) and CNN-LSTM, this model achieves a recognition rate close to 96% in limited samples, with faster convergence of loss values and the lowest mean square error in multiple types of actual mine scenarios. The MSE of high gas working face is about 0.43, significantly lower than the 0.66 and 0.75 of the control model. The ablation experiment further validates the key roles of GRA feature screening, ARIMA trend correction, and RMSE dynamic threshold.