Reinforcement Learning Fault Diagnosis Method Based on Less Tag Data

Abstract

Vibration signals are often used in the fault diagnosis of rotating machinery. However, due to the influence of complex environment, environmental noise is often doped, and the diagnostic accuracy is reduced. The traditional deep self-encoder is used in the noise reduction process of rotating machinery fault diagnosis. The pooling model is poor and easy to lead to over-fitting problems, and deep learning training needs a large number of labeled data. Therefore, this paper proposes a reinforcement learning fault diagnosis method based on less label data. The random pooling is used to replace the pooling layer of the original convolutional self-encoder, and the exponential linear unit (ELU) is used to replace the original activation function to enhance the convolutional self-encoder. A large number of unlabeled samples are used for training, and then the deep reinforcement learning is used for network fine tuning. The experimental results of the sensor data collected by the fault diagnosis test bench show that the method used has a good improvement in denoising ability and feature extraction ability, and the recognition accuracy and stability are better than traditional convolutional autoencoder and traditional machine learning methods.