A New Fault Diagnosis Method Based on Improved DQN for Cutting Tools

The practices of fault diagnosis present challenges in obtaining sensitive fault characteristics of tool system leading to poor fault diagnosis accuracy and jeopardizing equipment safety. To address above problems, an improved Deep Q Network (DQN) deep reinforcement learning fault diagnosis method is proposed. The new method utilizes a one-dimensional wide convolutional neural network to fit the Q network, with one-dimensional vibration signals and fault types serving as action ensemble inputs. Meanwhile, the ε-greedy strategy guides decision action and feedback reward is employed. The agent in the method uses time difference error (TD-error) priority experience playback enhancing stability and convergence. The algorithm continuously interacts with the decision to maximize the reward and reach to the optimal strategy fault diagnosis results. The model is applied to the cutting tools worn test bench dataset and achieves an accuracy of 99.08%, which can be used for fast and effective fault diagnosis. The results demonstrate the high fault diagnosis accuracy and generality of the improved DQN model, providing potential for enhancing equipment safety and efficiency.