A new method for predicting the degradation of rotating machinery is proposed based on the reinforcement learning 3-state combined long and short time memory neural network (RL-3S-LSTMNN). The degradation of a rotating machinery is divided into three monotonic parts, naming stationary unit, descending unit and ascending unit, by a monotone trend discriminator based on the least square linear regression method. Moreover, by virtue of reinforcement learning, the proposed method select LSTMNNs with the number of hidden layers and hidden layer nodes number suitable for each monotone trend unit, which improves the generalization performance and nonlinear approximation ability of RL-3S-LSTMNN to obtain a higher prediction accuracy. Besides, different hidden layer and node numbers and three monotonic trend units represent the action and status of the Q table, and the output error of LSTMNN is associated with the update of the Q table to avoid the blind search of agent (i.e., decision function) and accelerate the convergence rate of RL-3S-LSTMNN. Accordingly, the higher computational efficiency can be obtained for the proposed prediction method. A state degradation trend prediction of rolling bearing demonstrates the effectiveness of the proposed method.