The feature relationship of the motion state of CNC machine tools is very complex. Realizing the prediction of the future operation state of CNC machine tools can tap the potential abnormal emergencies of machine tools and enhance the stability of machine tool processing. In view of the problem of poor adaptability and low accuracy of prediction under dynamic label of machine tool state and differential distribution data, an adaptive hybrid deep learning model was established to predict machine tool state by combining time series feature relationship and model fusion method. Firstly, by combining the nearest neighbor classifier, an adaptive updating rule based on weight accumulation was designed, and a state prediction model with data adaptability was established. On this basis, an optimization strategy of feature distance metric based on center loss function was proposed, and a comprehensive decision loss function was constructed to ensure model fusion effectively. Based on a combination convergence criterion, the BBPT method was used to train the model, and the test data was verified . The experimental results showed that the model can adapt dynamic label and differential distribution data. The prediction of the state category of CNC machine tools had strong antiinterference, fast response and high accuracy, and can better meet the requirements of machine tool state classification and prediction. The prediction accuracy and real-time performance were significantly compared with BP and LSTM classification networks, and the shortest prediction time was only 100ms in GPU mode.