Abstract: In order to improve the operating efficiency of belt conveyors through variable frequency speed regulation, it is necessary to detect the coal flow of belt conveyor. The existing deep learning-based coal flow detection methods for belt conveyors are difficult to achieve a balance between model lightweight and classification accuracy. There are few researches on the impact of imbalanced channel weight distribution on detection accuracy in the feature extraction process. In order to solve the above problems, a lightweight multi-scale cross channel attention coal flow detection network is proposed. The network consists of a feature extraction network and a classification network. The lightweight residual network ResNet18 is used as the feature extraction network, and on this basis, the coal flow channel attention (CFCA) subnetwork is introduced. The CFCA subnetwork uses multiple one-dimensional convolutions with different kernel sizes, and stacks the output of one-dimensional convolution to capture cross channel interaction relationships at different scales in the feature map. It achieves the reassignment of feature map weights, thereby improving semantic expression capability of the feature extraction network. The classification network consists of three fully connected layers, which take the output of the vectorized feature extraction network as input and perform nonlinear mapping on it. It ultimately obtains the probability distribution of three types of results: "little coal", "moderate coal", and "much coal". By transforming the coal flow detection problem into an image classification problem, the problem of frequent frequency conversion and speed regulation of belt conveyors caused by excessive fluctuations in instantaneous coal flow is avoided. It improves stability of belt conveyor operation. The experimental results show that the ResNet18+CFCA network improves classification accuracy by 1.6% compared to the ResNet18 network, with almost no increase in network parameters and computational complexity. It can distinguish foreground information in images more effectively and accurately extract coal flow features.