Research On Fault Diagnosis Method Of Gas Regulator Under Limited Data Condition

China is a large country consuming natural gas,and the annual consumption of gas is also increasing.The safety issue in the process of gas transportation is an important issue to be solved in the development of the natural gas industry.Gas pressure regulator is an important part of today’s urban gas transmission equipment.Any failure or damage that occurs in the process of its work may cause gas leakage,which leads to a great deal of casualties and economic losses.Therefore,the research on fault diagnosis of pressure regulator is of great significance.At present,with the rapid development of monitoring technology and industrial automation,a large amount of monitoring data has been collected and applied,so the datadriven fault diagnosis method has been widely concerned and used in industrial fault diagnosis.However,the current data-driven fault diagnosis research is facing industrial problems such as shortage of data marker information,difficulties in collecting failure sample data,and even inability to collect data.Traditional data-driven fault diagnosis methods cannot perform well in the face of these problems.Generative diffusion model and transfer learning model are gradually introduced into the field of fault diagnosis because of their unsupervised learning ability and strong domain migration characteristics.This paper combines the domain transfer learning of generative diffusion model with the above-mentioned problems in the fault diagnosis of gas regulator,and studies the fault diagnosis methods.The main research contents are as follows:(1)Based on the Semi-supervised Generation Adversarial Network,a semi-supervised fault diagnosis model of pressure regulator based on the self-training Generation Adversarial Network is proposed by using the extended self-training method of duplicate tags.Firstly,the network structure against the network design generation model is extracted based on deep convolution generation,and the discriminator uses a stacked discriminator model with shared weights.Then,a self-training algorithm is designed to predict the class labels of unlabeled samples using the trained initial classifier.Finally,the samples that meet the requirements are expanded into labeled sample sets to retrain,and the final classifier is saved to build a complete regulator fault diagnosis model.(2)Based on the transfer learning and meta-learning theory,a fault diagnosis method with few samples based on parameter optimization and feature measurement is designed,which is called model unknown matching network fault diagnosis model.First,traditional datasets are used to train network parameters in the training domain,then scenario tasks are used to modify network parameters for cross-domain migration.In the classification module,matching networks are used to measure the similarity between sample features and target samples to complete the classification.Parameter optimization networks and feature measurement networks use the same thematic feature extraction layer,while modifying the model.Diagnostic results between fault datasets show that this method is superior to other baseline fault diagnosis methods to reduce the differences between datasets due to different devices.Therefore,this method has universality and engineering application value.(3)A zero-sample fault diagnosis model based on joint mapping of fault attributes is designed for the zero-sample fault diagnosis with zero target number in the industrial fault diagnosis of gas regulator.Firstly,according to the fault sample category of the gas regulator,the fault attribute characteristics are designed according to the location,cause and performance of the fault,and the zero sample dataset of the gas regulator is established.Secondly,fake samples of unknown faults are synthesized using the fault attribute features and the conditions.Finally,this chapter designs a joint mapping diagnostic model based on sample instance supervision and sample category supervision,and uses the training mapping module of synthetic samples and known class samples to build a zero-sample fault diagnosis model for gas regulator.