Federal learning algorithm-based safe operation early warning method and device

The invention discloses a safe operation early warning method and device based on a federated learning algorithm, the safe operation early warning device comprises a wearable device provided with a camera, the wearable device is used as a client to interact with a back-end server, and the method comprises the following steps: collecting video data in real time, extracting motion features of an object in a video, and sending the motion features to a server; inputting a local behavior recognition model; acquiring a behavior recognition result output by the local behavior recognition model, and if no unknown behavior is found, uploading the action characteristics of the object in the video and the corresponding behavior recognition result to a server; if unknown behaviors are found, a local action recognition model is trained according to the unknown behaviors, and the model is uploaded to a server after gradient encryption; and waiting and acquiring risk information and/or model parameters issued by the server,