Reinforcement learning unit matching recurrent neural network system and training and prediction method thereof

Abstract

The invention relates to a reinforcement learning unit matching recurrent neural network system and a training and prediction method thereof. The system includes recurrent neural networks, a reinforcement learning unit, and a monotone trend recognizer. The recurrent neural network includes input and output, and a plurality of hidden layers. The hidden layers include a plurality of hidden layer nodes. The monotonous trend recognizer judges the trend state of a time sequence constructed by input, and the reinforcement learning unit selects a recurrent neural network with the number of hidden layers and the number of hidden layer nodes matched with the change rule of the hidden layer nodes according to the trend state of the input time sequence. According to the invention, the trend state ofthe input time sequence is judged. The trend state, different hidden layer numbers and hidden layer node numbers are used for representing the state and action of the Q value set respectively, the optimal action is selected to be executed according to the updated Q value set, the generalization capability and convergence speed of the RLUMRNN are enhanced, and the residual life prediction method applied to the rotary machine is high in precision.

Description

technical field [0001] The invention relates to the technical field of neural networks, in particular to a reinforcement learning unit matching cycle neural network system and its training and prediction methods. Background technique [0002] Rotating machinery is one of the most widely used components in civil and national defense mechanical equipment. During long-term operation, rotating machinery will gradually degrade, and the remaining life will gradually decrease. The occurrence of failures often brings catastrophic accidents, resulting in major economic loss and serious social impact. For rotating machinery, the quality of the equipment can be improved through optimization of the design and manufacturing process, but it is still difficult to guarantee that there will be no failures during service. Even under the same working conditions, the rotating machinery of the same type and the same type has a large dispersion of life data due to different operating environment...

AI Technical Summary

Problems solved by technology

For PF, the resampling stage will cause loss of sample validity and diversity, leading to sample impoverishment phenomenon

Since the kernel function type and kernel parameters of SVR are still difficult to set accurately, the prediction results are also uncertain

There is no mature theoretical guidance for the selection of the number of hidden layers and the number of nodes of the artificial neural network. Generally, the selection is based on experience, which leads to unsatisfactory prediction accuracy and calculation efficiency of the model.

Images