Study on the fault diagnosis method of ship main engine unbalanced data based on improved DQN

Abstract

The Deep Q_Network (DQN) algorithm in reinforcement learning is introduced to main engine fault diagnosis to improve the accuracy and efficiency of fault diagnosis by using the optimized DQN network algorithm to compensate for the lack of data imbalance for unbalanced fault data that are close to the real situation. The optimization of the DQN network algorithm is reflected in three aspects: firstly, the ɛ-greedy algorithm is optimized using the Upper Confidence Bound (UCB) algorithm, which makes the algorithm achieve a better balance between experience and exploratory in the selection of fault types; secondly, the fully connected network of the basic DQN is optimized using the triple-formed layer CNN network layer is optimized to improve the algorithm operation efficiency; meanwhile, the reward function for unbalanced data is set according to the balance rate, and the problem of reward value bias and local optimum for small amount of data is considered, so that the optimized DQN network algorithm gets improved accuracy in fault diagnosis of unbalanced data. Finally, the optimized DQN network, the base DQN network, the DCNN, and the ResNet18 are run for diagnosis on the unbalanced data set. Compared with other algorithmic networks, the optimized DQN improved 5.18%∼18.58% in accuracy. The results show that the DQN algorithm model can be applied with main engine unbalanced data fault diagnosis, and the improved DQN algorithm achieves good results in the efficiency and stability of diagnosis.