A reinforcement learning unit matching recurrent neural network for the state trend prediction of rolling bearings. (October 2019)

• Record Type: Journal Article
• Authors: Li, Feng; Chen, Yong; Wang, Jiaxu; Zhou, Xueming; Tang, Baoping
• Is Part Of: Measurement. Volume 145(2019)
• Page Start: 191
• Page End: 203
• Publication Date: 2019-10
• Subjects: Reinforcement learning unit matching recurrent neural network -- Reinforcement learning -- Singular spectral entropy -- State trend prediction -- Rolling bearings; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2019.05.093
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A reinforcement learning unit matching recurrent neural network is proposed. A novel state trend prediction method using RLUMRNN for rolling bearings is proposed. A singular spectral entropy is used as a measure of degradation state. The effectiveness of the proposed state trend prediction method is demonstrated. Abstract: This paper proposes a novel neural network, called a reinforcement learning unit matching recurrent neural network (RLUMRNN), with the aim of resolving the problem that the generalization performance and nonlinear approximation ability of typical neural networks are not controllable, which is caused by the experience-based selection of the hidden layer number and hidden layer node number. In the proposed RLUMRNN, the monotone trend discriminator is constructed by using the least squares linear regression method for dividing the whole state degradation trend of rolling bearings into the following three kinds of monotonic trend units: ascending unit, descending unit and stationary unit. Moreover, by virtue of reinforcement learning, the recurrent neural network (RNN) with the hidden layer number and hidden layer node number fitted to a corresponding monotone trend unit is selected to enhance the generalization performance and nonlinear approximation ability of RLUMRNN. Additionally, three monotonic trend units and different hidden layer and node numbers are respectively used to represent the status and action of the Q value table, and a new reward function associated with the RNN's output errors is constructed to clarify the purpose of reinforcement learning. This makes the RNN's output errors smaller, which avoids the blind search of Agent (i.e., decision function) in the update of the Q value table and improves the convergence speed of RLUMRNN. By taking advantage of RLUMRNN in the generalization performance, nonlinear approximation ability and convergence speed, a new state trend prediction method for rolling bearings is proposed. In this prediction method, the moving average singular spectral entropy is first used as the state degradation feature of rolling bearings, and then the feature is input into RLUMRNN to accomplish the state trend prediction of rolling bearings. The examples of the state trend prediction for double-row roller bearings demonstrate the higher prediction accuracy and higher calculation efficiency of the proposed method.