Abstract: Anti-vibration hammer is an important electrical equipment in overhead transmission line system. It is of great significance to prevent the periodic fatigue failure of overhead line due to wind blowing. In aerial images, the vibration hammer is small in size, different in shape, complex and changeable in background, and difficult to detect. To solve this problem, fully one-stage convolution target detection network(FCOS) is used to detect the vibration hammer of overhead transmission lines. In order to improve the detection accuracy, each feature point of FCOS feature extraction layer is regarded as a random variable, and its random distribution is expressed by the combination of each order center moment. On this basis, a spatial attention mechanism based on the central moments of each order is proposed to accurately describe the weight distribution of image features. The experimental results show that the average detection accuracy of the improved FCOS is higher than the original FCOS at different thresholds, when the threshold is 0.5, the average detection accuracy reaches 94.9%. At the same time, the average detection accuracy of this method under different threshold values is much better than other mainstream attention mechanisms.