Abstract: To solve the problems of poor prediction accuracy and low computational efficiency of the existing artificial intelligence-based prediction methods in state trend prediction of rolling bearings,a novel state trend prediction method was proposed based on Reinforcement Learning Unit Matching Recurrent Neural Network (RLUMRNN).The moving average singular spectral entropy was used as the state degradation feature of rolling bearing,and then the feature was input to RLUMRNN to accomplish the state trend prediction.In RLUMRNN,the monotone trend discriminator was constructed by using the least square linear regression method for dividing the whole state degradation trend of rolling bearing into the following three kinds of monotonic trend units: ascending unit,descending unit and stationary unit;by virtue of reinforcement learning,the RNN with the hidden layer number and hidden layer node number fitted to corresponding monotone trend unit was selected to enhance the nonlinear approximation ability and generalization performance of RLUMRNN.Besides,three monotonic trend units and different hidden layer and node numbers were respectively used to represent the status and action of Q value table,and a new reward function associated with RNN output errors was constructed to clarify the purpose of reinforcement learning,which made the output error of RNN smaller,avoided the blind search of agent (decision function) in the update of Q value table and improved the convergence speed of RLUMRNN.The example of state degradation trend prediction for double row roller bearing demonstrated the higher prediction accuracy and higher calculation efficiency of the proposed method.

Key words: reinforcement learning unit matching recurrent neural network, reinforcement learning, singular spectral entropy, state trend prediction, rolling bearing

摘要： 为了解决当前人工智能预测方法在滚动轴承状态趋势预测中预测精度较差、计算效率较低的问题,提出基于强化学习单元匹配循环神经网络(RLUMRNN)的滚动轴承状态趋势预测新方法。先采用滑动平均奇异谱熵作为滚动轴承状态退化特征,再将该特征作为RLUMRNN的输入完成滚动轴承状态趋势预测。在RLUMRNN中,利用最小二乘线性回归法构造单调趋势识别器,将轴承整体的状态退化趋势分为上升、下降、平稳3种单调趋势单元,并通过强化学习为每一种单调趋势单元选择一个隐层数和隐层节点数与其相适应的循环神经网络,从而改善了RLUMRNN的非线性逼近能力和泛化性能;用3种单调趋势单元和不同隐层数、隐层节点数分别表示Q值表的状态和动作,并构造关于循环神经网络输出误差的新型奖励函数,以明确强化学习的目标,从而减小循环神经网络的输出误差,避免在Q值表更新过程中使Agent(即决策函数)盲目搜索,提高了RLUMRNN的收敛速度。通过双列滚子轴承状态趋势预测实例验证了该方法具有较高的预测精度和计算效率。

关键词: 强化学习单元匹配循环神经网络, 强化学习, 奇异谱熵, 状态趋势预测, 滚动轴承