Abstract: The unsafe behavior of miners is a key factor influencing the safe production of underground coal mines. The recognition of such behaviors is crucial for achieving intelligent monitoring and enhancing safety in underground operations. Currently, deep learning-based methods for recognizing unsafe behaviors of miners require a large amount of labeled data to train, and data labeling is resource-intensive. Semi-supervised learning-based recognition methods can reduce the labeling cost of miners’ images effectively, but most mainstream semi-supervised learning methods employ Exponential Moving Average (EMA) to update the teacher model conservatively, which results in a slower learning rate for the teacher model in early stages. As a consequence, the quality of the generated pseudo-labels is low, which impacts the training performance considerably. To address this issue, a semi-supervised recognition algorithm based on dynamic EMA is designed to identify miners’ unsafe behavior. Combined with the idea of exponential decay, the weight parameters in EMA are set to vary with the training batch dynamically, which is adaptive to different stages of training. In addition, the mine environment is dim and fuzzy, it is difficult to extract miners’ information, and will aggravate the inconsistency between the classification task and the positioning task of the recognition model, affecting the recognition accuracy. To solve this problem, Efficient Local Attention (ELA) is integrated into the feature pyramid network, and Efficient Local Attention Feature Pyramid Network (ELA-FPN) is constructed to improve the salience of miners’ information. In order to further enhance the consistency between classification and location of miners’ unsafe behaviors, a Feature Alignment Detection Head (FA-Head) is designed to map the location features to the classification features for achieving alignment between classification features, location features and improving the recognition effect on miners’ behaviors. Experiments show that when 10% labeled data is utilized in the miners’ unsafe behavior dataset, the recognition accuracy of the proposed algorithm for the unsafe behaviors of miners reaches 71.008% without increasing the complexity of the model. Compared with the mainstream Unbiased teacher v1, Unbiased teacher v2, Consistent teacher, Dense teacher and ARSL, the results are improved by 5.33%, 1.76%, 2.08%, 1.24% and 0.40%, respectively, and outperformed other state of the arts on the former comparison algorithms under different supervision ratios. It can be concluded that the proposed algorithm is superior to the current mainstream semi-supervised learning method in the task of miners’ unsafe behavior recognition, and has a good recognition effect while the annotation cost is lessened effectively.