Research On Intelligent Fault Diagnosis Method Of Rotating Machinery Components With Few Samples

With the rapid development of industrial technology,the automation and complexity of mechanical equipment has been greatly improved.In order to avoid great loss,the requirement of equipment state diagnosis is getting higher and higher.Rotating mechanical parts bearing and gear as the joints of mechanical equipment,in the safe and normal operation of equipment play an extremely important role.It is an effective method to diagnose the fault by analyzing the vibration signal of rotating machinery parts.However,under actual operating conditions,most of the vibration data collected are unlabeled,despite the large volume of vibration data.It is not practical to collect and labeled enough failure data for every situation.Therefore,it is an important problem to establish a reliable fault classification model based on a small amount of vibration data.Therefore,this dissertation will study the fault diagnosis of rotating machinery parts with only a small amount of label data based on deep learning,including the following two points:(1)A deep adversarial convolutional neural network based on semi-supervised learning(SACNN)is proposed.In this method,the data are transformed into images by preprocessing,and the time-frequency information of the original data is obtained through gray-scale image analysis.A large number of fake data generated by generator and unlabeled true vibration data is used in the model to learns the overall distribution of data by judging the authenticity of the input.Three regular terms for different inputs are designed to impose constraints on the parameters of the discriminator to enhance the model’s learning of data features.The data collected from the motor bearing of the Case Western Bearing Data Center and the Lab-built experimental platform datasets were used to identify the faults of the bearing under actual conditions.The experiment results show that the proposed method has higher diagnostic accuracy than traditional deep models in multi-group small datasets of different capacities.The effectiveness of the model provides a new solution to the problem of big vibration data and few labels in fault diagnoses.(2)A cross-category mechanical fault diagnosis based on deep few-shot learning network(CFDM)is proposed.This method constructs a five-layer convolutional Siamese neural network to extract fault features from example pairs.We define two parameters for feature discrepancy metric into the novel loss function to maximize the inter-category distances while minimizing the intra-category distances,such that the CFDM is able to learn the accurate classification boundaries between fault features of the sample pairs.We conduct experiments on two benchmark datasets and one self-built testbed dataset with respect to one or five samples of the new component.The superiority of this method is verified by eight groups of experiments.