Intelligent Diagnosis Of Rotating Machinery Fault Degree Insensitivity Under Data Imbalance Distribution

Rotating machinery has been widely used in the industrial era,but due to the complex operating environment of variable speed and variable load for a long time,it is very prone to failures and the initial failure is extremely difficult to detect.The classification of its life state can be generally divided into three stages: data acquisition,qualitative diagnosis of the fault site,and quantitative assessment of performance degradation.In this paper,the key problems encountered in the three stages are studied and explored by combining theory and experiment.In the data acquisition stage,a Time-Frequency Feature Oversampling(TFFO)method is used to combine a deep learning method with a Convolutional Neural Network(CNN)to address the problem that deep learning methods usually require abundant data,but in practice the fault samples are few,resulting in data imbalance.The balanced dataset is constructed by simultaneously expanding the time-domain signal and time-frequency domain features,and a comprehensive data expansion is performed from different dimensions.SMOTE oversampling is also performed on the denoised features after time-frequency conversion to generate high-quality samples,thus effectively solving the problem of low quality of generated data.In the qualitative analysis stage of the fault site,a Probabilistic Slicing Cumulative Projection feature(PSCP)is proposed to address the problem that the current diagnosis method often treats a discrete fault degree as a separate category,which leads to a model that cannot adapt to the complex scenario of dynamic changes of the actual fault degree.Featuresbased qualitative diagnosis method of bearing fault degree insensitivity.Firstly,the Probability box(P-box)theory is used to classify the data with different fault levels in the same fault site into the same class of fault states.Then the Nuisance Attribute Projection(NAP)algorithm is used to project the features to generate the PSCP matrix.Finally,fault diagnosis is achieved by CNN.A single or missing fault degree dataset is constructed to simulate a realistic situation for analysis,showing that the proposed method is still highly accurate when the training and test data belong to different fault degrees.In the quantitative assessment stage of failure degree,a quantitative assessment method of performance degradation based on the Multivariate State Estimation Technique(MSET)reconstructed model overall optimization is proposed for the phenomenon that the previous single-domain index cannot accurately describe the failure trend and the reconstructed model combining multi-domain features has information redundancy.First,multiple time-domain and frequency-domain features,AR model coefficients,and three-layer wavelet packet Renyi entropy are extracted to form a high-dimensional multi-domain feature vector,while the highdimensional features of the health state are constructed into a historical observation matrix D.Then,the extracted high-dimensional features and D are simultaneously and jointly optimized by Genetic Algorithm(GA),so as to achieve feature preference and overall adaptive optimization of the reconstructed evaluation model.Finally,the performance degradation evaluation curve is constructed by using the cosine similarity.The results of the analysis of several bearings’ whole-life data show that the proposed method has certain validity and reliability.In summary,the analysis results of several experiments in this paper show that the involved method can effectively carry out the research of intelligent diagnosis of rotating machinery fault degree insensitivity under data imbalance distribution,which has certain cutting-edge and engineering application value.