Applying feature extraction of acoustic emission and machine learning for coal failure forecasting

Recently, with increasing mining scale, intensity, and depth, the geological and mining conditions in coal mines are becoming more complicated; therefore, it has resulted in a more difficult situation of coal mine dynamic hazards, including rockburst, coal and gas outburst etc. Dynamic hazards are now posing a serious threat to the safety of coal mining. The precise forecasting of dynamic hazards is significant to their effective control. The acoustic emission (AE) monitoring technique is an effective geophysical monitoring and early warning method which can effectively reveal the characteristics and laws of coal and rock failure under loading. It has been successfully applied in the laboratory and engineering fields. To deeply analyze the characteristics of AE signals in the process of coal-rock damage and failure, thus, to help realize the precise monitoring and early warning of coal mine dynamic hazards, this study first conducted a uniaxial compression test on coal samples in the laboratory, and at the meantime, synchronously collected the full waveform data of AE and the loading data in the entire process of coal failure. Subsequently, using the feature extraction technique in the field of automatic speech recognition, this study extracted the Mel-frequency cepstral coefficient (MFCC) of AE and used it as the sample feature; the stress state of the coal sample was defined as the ratio of the current load the sample bore to its peak load and was employed as the sample label; a model for coal failure state forecasting was established by adopting machine learning methodology. Finally, the model’s forecasting accuracy was evaluated using the five-fold cross-validation method; the influence of different MFCC combinations as sample features on the forecasting accuracy of the model was discussed. The results show that MFCC can well characterize the failure state of coal samples. This parameter behaves in regular variation with increasing loading and shows the law of an obvious sudden increase or sudden decrease or increase followed by a sudden decrease when the loading exceeds 80% of the coal sample’s peak load. The established model can be well used to forecast coal failure state. The accuracy (ACC), true positive rate (TPR), true negative rate (TNR), and area under the curve (AUC) of the model forecasting reach 88.61%, 72.34%, 93.16%, and 0.93, respectively. Machine learning methodology can fully use MFCC features of AE and can identify essential sample features that are difficult to identify with the human eyes. Significant and key features included in the samples are the keys to the high forecasting accuracy of the model. TPR, TNR, and AUC of the model forecasting would be significantly influenced if crucial features were excluded from the samples. Adding features with low importance to the samples has little influence on the forecasting result of the model. This study’s results can provide a reference for further improving the prediction and early warning of coal and rock dynamic hazards.