Intelligent Few-Shot Diagnosis Method For Rotating Machinery

Rotating machinery is an important part of modern industry,and with the rapid development of the economy,there are requirements for safe,efficient,continuous and high-quality operation of rotating machinery.For fields such as industrial production,energy and power,and rail transportation,the breakdown of a component of rotating machinery will lead to a decline in its overall performance,reduced operational stability and even endanger personal safety.It is important to monitor and diagnose the operating status of rotating machinery which is the core equipment in the unit to reduce unplanned downtime caused by malfunction,so as to guarantee the safety of industrial personnel and equipment,ensuring the production efficiency of process industry and reducing the production cost of enterprises.With the development of sensing technology and information technology,the society is gradually entering the digital and intelligent "big data" industrial Internet era of rotating machinery,however,it still faces the few-shot problem with more normal data and less fault data or even missing.At present,the fault diagnosis technology based on machine learning needs to rely on a large amount of data of various fault states to complete the construction of intelligent diagnosis models,so it cannot be effectively applied in the case of few-shot with only normal state data available,resulting in the weak generalization ability of the existing intelligent diagnosis methods to specific equipment.In this paper,we propose a framework of intelligent diagnosis method for rotating machines with small samples,and build an instantiated intelligent diagnosis model for each machine by combining fault mechanism and machine learning technology,which solves the problem of difficult diagnosis due to the lack of faults of the equipment to be tested.The research in this paper consists of three major parts: the construction of a diagnostic model based on homologous migration,the construction of a caseless fault identification model based on occurred fault samples and the construction of an acoustic diagnostic model based on heterologous migration,as follows.(1)The research proposes two methods of constructing diagnostic model instantiation based on normal state data of the equipment to be tested and guided by the fault mechanism.In order to solve the problem that the diagnosis model cannot be constructed under the condition that only normal state data d are available for the equipment to be tested,an adaptive diagnosis with fault mechanism(ADFM)method is proposed.The fault samples are generated based on the health state data,which is based on the fault mechanism and transmission path law.The frequency domain features that can reflect the fault category are proposed to reduce the difference between the generated and real samples.The sample generation process has been greatly streamlined to ensure sample quality as well as to improve the efficiency of diagnosis in few-shot conditions.To address the problem that the ADFM method relies on manual experience in generating samples,a multi-source data and adversarial network hybrid driven mechanical intelligent diagnosis(MSGAN-MCGM)method is proposed.The MSGAN-MCGM method is guided by the fault mechanism and accomplishes the migration of fault components in multi-source fault data with the help of machine statistical learning techniques,and the potential information behind the mechanism and data is effectively utilized.The proposed homologous migration-based sample construction method on the one hand solves the difficulty of establishing a diagnostic model relying on real fault data of the device to be tested.On the other hand,the construction of the instantiation of the diagnostic model for a specific device is ensured.The accuracy of the proposed method is verified by experimental data and industrial data,and the results show that the proposed method has significant advantages compared to existing methods under the few-shot conditions with only normal status data available.(2)The Convolutional Neural Networks and Deep Convolutional Generative Adversarial Network(CNN-DCGAN)method is proposed to build an intelligent diagnostic model for unknown faults based on DCGAN.To address the problem that the diagnostic model can only recognize the occurred faults contained in the training set due to the missing fault samples,the diagnostic model is trained with the occurred fault samples and achieves the intelligent recognition of occurred faults and no-case faults by introducing confidence probability.In addition,a self-renewal strategy of the diagnostic model based on new fault samples is developed to improve the expansion and self-learning capability of the model while retaining the recognition effect of the diagnostic model on the original occurred faults.The experimental data and industrial data verify the advantages of the proposed method for the recognition of various types of faults,and provide strong support for the engineering application of intelligent diagnostic models under the conditions of small samples of limited types.(3)The acoustic diagnosis model construction method based on vibration sound migration under heterogeneous source and heterogeneous domain samples is proposed.The sound-based intelligent diagnosis method is studied for industrial scenarios where vibration sensors are difficult to install.Considering that the application range of sound monitoring is smaller compared with vibration,and there is a deficiency that both the source domain equipment and the equipment to be tested are missing fault sound data,the research on the generalization technology of diagnosis model based on heterogeneous source migration is carried out,and the construction method of acoustic diagnosis model of the equipment to be tested based on the vibration fault data of other equipment in the indoor complex sound field environment is formed.In-depth analysis of the composition,attenuation and signal characteristics of rotating machinery acoustic signals,clarify the differences between vibration and acoustic signals,and propose the Migration Diagnosis of Different Devices from Vibration to Acoustic(MD3VA)based on the single microphone acoustic signal enhancement method.Vibration to Acoustic(MD3VA)method based on the source domain device vibration data.The experimental data and industrial data analysis results show the effectiveness of the method,and the proposed method greatly reduces the cost of intelligent diagnosis of rotating machinery and broadens the prospect of engineering application of acoustic diagnosis technology in industrial production under small sample conditions.Finally,the research content of this paper is summarized and further research directions are proposed.