Currently, training load monitoring is mainly divided into vision-based motion monitoring and wearable sensor-based motion monitoring. Vision-based motion monitoring tends to have a poor monitoring range and is affected by the environment, which makes it difficult to carry out long-term accurate monitoring and at the same time violates privacy. Wearable sensor-based motion monitoring is not affected by the above factors, this paper combines the advantages of the two, and proposes a training load monitoring method for athletes based on sensor technology and motion image analysis, which can be used for motion monitoring anytime and anywhere. In traditional wearable IMU-based motion monitoring algorithms, a large number of features usually need to be extracted for recognition, however, the extraction of features often requires specialized domain knowledge, and if the extracted features are not suitable it will lead to difficulties in improving the accuracy of the algorithm. Therefore, this paper proposes a two-stage neural network motion monitoring algorithm to identify periodic and non-periodic motions separately, which can effectively reduce the complexity of the network and also improve the accuracy of the recognition of each motion. In addition, this paper proposes a data enhancement algorithm based on acceleration data, which solves the problem of fewer data samples in some datasets, greatly increases the number of samples without re-collecting data, and is more suitable for end-to-end neural network training to further improve the accuracy of the algorithm recognition, and the results of the simulation experiments show that it can be applied to the actual situation.