Feature Dimension Reduction And Adaptive Extraction Methods For Planetary Gearbox Fault Diagnosis

Planetary gearboxes are widely used in aeronautical and industrial field.However,their components are prone to damage and malfunction due to the intricate kinematic condition,poor operating circumstrances and complex alternating loads.Therefore,it is necessary to study the fault diagnosis of planetarty gearbox.The components of the vibration signal of the planetary gearbox are complex,and it is difficult to identify the fault features manually through signal analysis.To solve this problem,this thesis studies the fault diagnosis methods of planetary gearbox from the perspective of pattern recognition.How to extract features of good stability,sensitivity and regularity,and use them to diagnose planetary gearbox fault is the focus of this thesis.In this thesis,we take the planetary gearbox as the specific research object.Different fault disgnosis methods of gear and planet bearing have been proposed according to the actual problems.First,fault diagnosis methods based on traditional feature extraction are studied.After feature space construction,manifold learning and sparse filtering are utilized to further extract low dimensional sensitive features due to the high-dimensionality and non-linearity of the feature space,and the parameter estimation problems involved in manifold learning are solved.Second,fault diagnosis method based on adaptive feature extractin is studied.The main contents and innovations of this thesis are as follows:1.Extraction and application of multi-domain manifold features:stability,sensitivity and regularity of features are different for different systems.In order to ensure the richness of features,statistical features of time domain and frequency domain are extracted for the vibration signals of the planetary gearbox.Besides,local mean decomposition is used to decompose the signal,and the energy features of time-frequency distribution are extracted.Manifold learning is utilized to further extract low dimensional sentive features,and a novel parameter estimation method is proposed:a parameter estimation method based on improved pseudo-nearest neighbor,which is able to estimate the target dimension and the size of neighborhood simutaneously.Finally,a probabilistic neural network classifier is used to establish a fault diagnosis model,and the proposed methoed is verified experimentally.2.Sparse optimization of energy features of time-frequency distribution:it is found that the stability and sensitivity of energy features of time-frequency distribution are relatively good,and they are expected to be used alone for fault diagnosis.Variable mode decomposition method is utilized to decompose the signal for the merit of high efficiency,and the energy features of time-frequency distribution are extracted.Considering the fact that diffenret menafold learning methods have different effects on feature processing,in order to further extract sensitive features,the sparse filtering method is used instead of menafold learning to optimize the feature space,which is free from parameter estimation of target dimension and size of neiborhood,and it only needs to give the number of features that need to be learned,which simplifies the process of extracting sensitive features.Finally,a fault diagnosis model is constructed by combining with a probabilistic neural network classifier.The proposed method is verified experimentally.3.Sparse classification based on dictionary learning:traditional feature extraction methods are widely used and work well.However,they rely on expertise,and feature extraction lacks self-adaptation.The dictionary learning method can initialize the dictionary with the original signal samples,and adaptively extract features from the original signal.Each atom in the dictionary is a typical feature.Based on dictionary learning,a sparse classification method is proposed.Under certain sparsity constraints,the minimum reconstruction error is determined,and the samples are classified for fault diagnosis.The entire model is free from designing a specific classifier,which has a simple structure and high intelligence,and can be implemented by using traditional dictionary learning and sparse decomposition methods.The parameters used in dictionary learning and the influence of different parameters on the fault diagnosis accuracy of the model are analyzed through experimental signals,which provide reference for the parameter determination during dictionary learning,and finally the method is verified experimentally.4.Feature extraction and application of time-frequency map based on convolutional neural network:In variable conditions,the vibration signal has obvious non-stationarity,and the characteristic frequencies are time-varying.The traditional analysis methods of time-domain and frequency-domain are difficult to extract the fault characteristic frequency effectively.When the traditional time-frequency analysis methods are used for analysis,there are problems such as low time-frequency accuracy and false component interference,which makes identification of characteristic frequencies difficult.To solve this problem,a feature extraction method of time-frequency map based on convolutional neural network is proposed.This method gives full play to the advantages of convolutional neural networks in image processing in order to adaptively extract the characteristics of time-frequency map,without considering the problems of low time-frequency accuracy and false component interference in traditional time-frequency analysis.The method is verified experimentally.