Application of artificial intelligence in coal mine ultra-deep roadway engineering—a review

Abstract

The deep integration of computer field and coal mining field is the only way to coal mine intellectualization. A variety of artificial intelligence tools have been applied in open-pit and shallow coal mines. However, with the geometric increase of coal demand, the contradiction between supply and demand is becoming more and more serious, and the exploitation of resources from shallow layer (> 600 m) has become an inevitable trend. Well then, as a new engineering scene, the harsh conditions of “three high and one disturbance” seriously threaten the safety of personnel. The superposition of complex mining environment makes the number of input factors increase sharply, which leads to the application of artificial intelligence in deep coal mine roadway engineering. The guidance is not mature, the construction of various databases is missing, and there are still some problems in universality and applicability. To this end, this paper starts with the introduction of the operating characteristics of various artificial intelligence tools, and conducts a comprehensive study of the relevant high-level articles published in top journals. It systematically sorts out the research progress that has successfully solved the five directions of rock mechanics strength, surrounding rock stability, rock-burst, roof fall risks and micro-seismic events. While objectively evaluating the comprehensive performance of different tools, it also expounds its own views on key research and results. Literature review shows that whether as a development tool or a comparative model, the application of ANN in the field of coal mining is more than 98%, and it performs extremely well in the direction of surrounding rock stability and roof fall risk, with an accuracy rate of more than 90%. As the most mature direction of AI tools application, rock mechanical strength has experienced the development process from “SVM → ANN → DL → XGBoost → RF”. The dataset is from small samples (< 100) to big samples (> 1000), and the R2 of tree-based models can be stabilized at more than 95%. The research on rock-burst prediction mainly focuses on field factors and micro-seismic monitoring data. Whether it is a small sample or a large-scale data model, the accuracy of BN remains above 85%. The prediction and evaluation of micro-seismic events is a new direction in recent years. The image processing and application of CNN is extremely important. The signal recognition and classification accounts for more than 90%, and the research potential of source location needs to be further explored. In general, the nature of the rock itself is the first choice for almost all influencing factors. At the same time, the update iteration of monitoring methods (micro-seismic, ground sound, separation, deformation, etc.) expands the development of the database, making it possible to obtain the data due to threat to life and cost of equipment, which is very difficult to obtain before. In the process of parameter selection at the input end, the method of combining lithology conditions, geological environment and monitoring data will gradually become the first choice for research. Finally, in the follow-up work collation and on-the-spot investigation, it mainly focuses on the existing problems of deep coal mines, explores the application potential of artificial intelligence in deep coal mine roadway engineering, puts forward the possible research focus and challenging problems in the future, and gives its own opinions.