Objective As one of the most important primary energy resources, coal needs to be mined in large quantities every year. In this process, the mass proportion of biological coal gangue in raw coal is more than 10%. It needs to be separated from raw coal, which is called coal washing. As an important traditional energy field in China, the intelligent construction of coal industry is directly related to the process of the economic and social intelligence. Therefore, it is of great significance for the intelligent construction of coal mines to develop modern coal gangue separation technologies. Coal gangue is identified by machine learning algorithm, and then manipulators or pneumatic devices are used for the separation of coal gangue from coal, which is called intelligent separation of coal gangue. At present, coal gangue recognition based on an imaging technology has attracted more and more attention, but the recognition can only realize whether there is coal gangue, but cannot locate the coal gangue. In addition, the locating speed of coal gangue is directly related to the production efficiency of coal mines. This paper designs a coal gangue detection model based on the framework of YOLO v4, which aims to realize the faster identification and location of coal gangue, and provides the position information of coal gangue more quickly for the intelligent sorting system. In order to reduce the interference of environmental visible light, coal gangue is identified and located based on the multispectral imaging technology. In order to reduce redundant information and further improve detection speed, three pseudo RGB images are selected from 25 bands of multispectral images by using the optimal index factor theory.

Methods Firstly, the collected multispectral data of coal and gangue in the laboratory and the selected three bands from 25 bands of multispectral images were used to form the pseudo RGB images by using the best index factor theory. The training data, validation data, and test data were 460, 65, and 115 spectral images, respectively. Second, a lightweight coal gangue detection model was designed based on multispectral imaging, which is named YOLO mg. The maximum depth of its backbone network is 23 convolution layers. In order to achieve stable and fast training of YOLO mg without using the pre-training weights, three learning rates, which are 0.00025, 0.0005 and 0.0001, are used, respectively. Finally, the model YOLO mg was applied to the study of coal gangue detection based on the pseudo RGB images, and the redundant bounding box was filtered by the non-maximum suppression combined with confidence threshold.

Results and Discussions The rapid identification and location of coal gangue based on multispectral imaging are studied. The maximum depth of the backbone network of YOLO mg is 23 layers (Fig. 4). OIF is used for multispectral band selection, the maximum value is 11.138, and the corresponding combination band selection is [7, 12, 23] (Table 2). The combination band is used for the designed YOLO mg training. Three learning rates are set and trained for 100 epochs. The convergence speed of the designed model is fast, and the verification set loss decreases from the initial 1500.60 to the minimum of 6.81. In addition, compared with training loss and verification loss, YOLO mg is not over learning (Fig. 5 and Table 3).The redundant prediction boundary box is filtered by the non-maximum suppression combined with confidence threshold. Under the input resolution of 204 pixel × 204 pixel, the average accuracy of coal gangue detection is 91.91%, and there are only four false positive (FP), which effectively reduces the wrong positioning of coal gangue in the prediction results (Table 4 and Fig. 7). Compared with the advanced YOLO v4, M2Det, and YOLO v3, the YOLO mg designed here maintains an excellent detection accuracy, and the positioning speed of coal gangue is the fastest. The 115 combined band spectral images only take 1.255 s (Table 5).

Conclusions The intelligent coal washing technology is an important part of the realization of coal mine modernization construction. The multispectral imaging technology is applied to the study of intelligent separation of coal gangue. The optimal combination band is selected by OIF, and a detection model is designed to realize the location of coal gangue. Firstly, the multispectral imaging technology is used to identify and locate coal gangue. Three optimal combination bands are selected from 25 bands for coal gangue detection research by the OIF theory, which can effectively reduce the redundant multispectral information. Second, a fast detection model is designed. Without using the pre-training weights, the model can be trained quickly and stably by setting three learning rates. Finally, the detection time of 115 spectral images by YOLO mg designed in this paper is only 1.255 s, which is the least time-consuming in the comparison models, so that the rapid detection of coal gangue is realized and the design purpose is achieved. The lightweight detection model of YOLO mg designed here aims at the specific scientific problems of coal gangue localization based on multispectral imaging and provides a guidance for the lightweight model design of positioning problems in other fields.