Research On Key Technologies Of Rotating Machinery Early Fault Diagnosis

Rotating machinery is a kind of key equipment widely used in the fields of aerospace,electric power,transportation,petrochemical,and national defense industry.Its working environment is usually harsh and the working conditions are complex.In the long-term operation process,the mechanical performance continues to deteriorate and the failure occurs frequently.Once the fault occurs,it would not only seriously restrict the production efficiency of enterprises,reduce the quality of products,affect the market competitiveness of enterprises,make enterprises suffer huge economic losses,and even cause irreparable casualties.Based on the vibration monitoring,a large number of data and parameter images containing running and condition information can be obtained,such as spectrum image,3D spectrum image,order image,etc.How to use them to extract features for abnormality detection and fault severity degree determination timely and accurately,which is of great significance to improve the reliability and stability of the rotating machinery.Currently,the studies mainly focus on the fault diagnosis with obvious features,uncorrelated data,balanced samples and complete information,and these research results can accurately extract fault features and determine fault types.However,with the development of maintenance mode from passive to active,the demands for fault diagnosis technology with intelligent and practicality are growing continually.In the modern industrial system,the diagnosis methods need to be able to detect the early fault of machinery in time and adaptively,so as to provide reliable basis for the later maintenance,while the existing methods are difficult to meet these requirements.Therefore,in order to solve the problems and to consider the characteristics of early faults such as weak features,coupling and incomplete feature information,it has important academic significance and engineering application value to carry out the research on key technologies of early fault diagnosis of rotating machinery.This study mainly includes the following aspects.(1)Research on early fault diagnosis method of rotating machinery based on vibration signal analysisThe periodic vibration signals with multiple frequency components are generated while rotating machinery running and these signals are generally divided into two categories:1)the vibrations related only to the elasticity of rotating parts;2)the vibrations related to the damage of rotating parts.Therefore,the rotating machinery fault diagnosis can be effectively achieved based on vibration signal analysis.However,due to the complexity of industrial process and the harsh working environment,the vibration signals generated by rotating machinery are usually non-stationary,non-linear and contain strong background noise.It is difficult to mine the early and weak fault features hidden in the original signals directly.Considering the deep autoencoder with the excellent performance on automatic feature extraction,a fault diagnosis method based on the improved deep sparse autoencoder(SAE)network is proposed in this paper.With the average identification accuracy of more than 99%,the early fault pattern recognition and severity determination of rotating machinery(whether loaded or not)are achieved.(2)Research on multiple faults diagnosis method of rotating machinery with data correlation conditionThe components of modern rotating machinery are interrelated and interacted with each other,leading to strong uncertainty and non-linear characteristics of the whole system.In addition,the parameters representing the running condition have the characteristics of wide variety,high-dimensional sparsity,hardly quantify and distinguish.Consequently,the monitored data reflecting the running mechanism and condition parameters of machinery are massive and correlated,which makes early multiple faults widely existing in all kinds of rotating machinery.However,the existing fault diagnosis methods for a single device,subsystem and sub unit are difficult to discover the correlation between components,resulting in a high misdiagnosis rate and missed diagnosis rate for multiple faults.Therefore,by analyzing and processing the correlation data,and considering the performance on adaptive feature extraction and data dimensionality reduction of deep SAE,a fault diagnosis method based on correlation analysis and improved SAE network is proposed,achieving the early multiple faults diagnosis of loaded rotating machinery with the average identification accuracy of more than 99.4%.(3)Research on fault diagnosis method of rotating machinery with incomplete data conditionCurrentlly,the design of fault diagnosis methods for rotating machinery is mostly based on the assumption that the class distribution is roughly balanced and the data obtained are complete.Generally,it is assumed that the number of all kinds of health condition samples used for training is approximately equal,and the data forming the samples are not missing.Actually,most of the monitored data are normal condtion data,while a small amount of fault condition data.In addition,due to the sensor fault,communication line and human factors,the signals collected in real industrial system may be incomplete,leading to the loss of information hidden in the signals.The current situation of incomplete data greatly affects the effectiveness and accuracy of fault diagnosis methods.Therefore,by designing a variable-scale resampling strategy and a multi-level denoising strategy,a fault diagnosis method based on ensemble fusion autoencoder network is proposed to effectively achieve rotating machinery fault diagnosis with the condition of data imbalanced and data partially missing,and the average identification accuracy is more than 99.69%(1)Research on early fault diagnosis method of rotating machinery based on vibration bispectrum recognitionIn fact,the characteristic signals of rotating machinery early fault are very weak and sparse In addition,the amplitude of background noise is much larger than the characteristic signals and almost distributed in the whole frequency band,leading to the lowly SNR(Signal-to-Noise Ratio).Compared with the 1D vibration data,the bispectrum with strong denoising ability contains more information about the running condition of machinery.Therefore,considering the excellent performance of the bispectrum and the ESRGAN(Enhanced Super-Resolution Generative Adversarial Networks),firstly,based on bispectrum analysis technology,the vibration bispectrum images of rotating machinery are obtained,and then convert them into 2D gray-scale images with less storage demands.Secondly,based on ESRGAN,a super-resolution reconstruction strategy is designed to achieve the gray-scale images clarity and improve the image resolution.Thirdly,an improved convolutional neural network(CNN)classification model for vibration bispectrum recognition is proposed,with the average identification accuracy of 99.99%and 100%,the early fault signals of different types of rolling bearings are detected.Finally,with a more intuitive and understandable way,the early fault diagnosis of rotating machinery is achieved.