Neural Network of Roof Cutting Blasting Parameters Based on Mines with Different Roof Conditions

The design and construction of roof cutting and blasting is a key part of the roof cutting pressure releasing gob-side entry retaining (RCPRGER) technology. In the existing research, the blasting parameters of roof cutting have been primarily determined by field tests. However, the disadvantages of field tests include a complicated process, which hinders the succession of related procedures, and an unstable roof cutting effect. Therefore, in this work, the authors attempt to use a mathematical analysis method to simplify the design process of the key parameters of roof cutting blasting. First, the mechanics process mechanism of surrounding rocks with roof cutting and pressure releasing is investigated, and the stress evolution process of the surrounding rock is divided into the following six stages: original rock stress state, excavation stress state, supporting stress state, roof cutting stress state, premining stress state, and postmining stress state. Furthermore, the relationship between roof cutting and entry retaining from the perspective of Mohr’s stress circle is discussed. Next, using four typical mines, including the Tashan, Yuanlin, Jinfeng, and Hengyuan coal mines, as examples, the existing design methods of roof cutting and blasting, geological data characteristics of each mine, distribution rule of roof cutting connectivity rate, and explosive charge structure of roof cutting blasting are summarized. Based on these results, the logic of roof cutting blasting design is obtained, the key indices affecting blasting design are determined, and the blasting design is defined as a complex fuzzy problem with multiple factors. Finally, based on the study of the above mechanics mechanism and blasting rule, a three-layer back propagation (BP) neural network, including six input units, nine hidden units, and three output units, is developed with the four typical mines as the sample space. This neural network realizes the rapid determination of the three key parameters pertaining to sealing length, blasthole spacing, and the explosive charge weight of a single hole. Through training, the calculation error is less than 0.48%, which considerably simplifies the design process of the blasting parameters. The charge structure parameters can also be determined according to this method. At present, the construction of this neural network has the shortcomings of limited sample space. This problem can be overcome by supplementing the sample size in the subsequent