Research On Intelligent Diagnosis Method Of Compound Fault Of Rotating Machinery Based On Tensor Decomposition

Machinery industry is the basic and strategic pillar industry of national economic development.The importance of mechanical equipment is further determined in the outline of the national 14 th Five Year Plan: The R&D and the technological innovation of the key large-scale mechanical equipment occupy the core position.Among the mechanical equipment,rotating machinery is the main power source and core of various mechanical equipment,especially large equipment(such as coal mine underground coal mining,crushing,transportation and other equipment).The fault types of rotating machinery also account for the vast majority of mechanical equipment faults.The normal operation of mechanical equipment mainly depends on the reliability and failure rate of rotating machinery.Therefore,it is of great significance to study the intelligent diagnosis method of rotating machinery fault.There are many research methods on condition monitoring,detection and fault diagnosis of rotating machinery in the existing literature,such as neural network method,support vector machine method,genetic algorithm,Hilbert-Huang transform and other methods.Although these theoretical research methods are relatively mature,due to the poor working environment of actural mechanical equipment,the different signals generated in use interfere with each other,and these signals have complex coupling characteristics and strong nonlinearity,which makes the accuracy of fault feature extraction insufficient,can not reflect the fault type.Moreover,the fault cannot be diagnosed at the initial stage of the fault.These problems in actual fault diagnosis and monitoring of rotating machinery need to be solved urgently.Aiming at the problems of serious signal interference,insufficient accuracy of feature extraction and difficult detection of initial faults in the process of actual fault diagnosis,condition monitoring and detection of rotating machinery,the rolling bearings and gearboxes used in coal mining machinery are taken as the research objects.Based on tensor decomposition method,the existing intelligent optimization algorithm,wavelet transform and neural network method are improved,respectively.Furthermore,the methods of signal denoising,feature extraction and initial fault diagnosis of rotating machinery in complex environment are proposed.The main contents of this paper are as follows:(1)The problems of the characteristics of heterogeneous,mixed and different dimensions of the data collected in fault diagnosis are studies.The commonly used tensor decomposition methods(CP decomposition,Tucker decomposition,HOSVD decomposition,HT decomposition,TT decomposition,TC decomposition,and TTr1 decomposition,etc.)cannot effectively eliminate the invalid information in the tensor.In the decomposition,it is necessary to select an appropriate number of rank 1 terms.If the value is small,it will lead to the loss of useful information of the original tensor.On the other hand,if the value is selected too much,it will lead to the decrease of computational efficiency.In order to improve the feature extraction capability and anti-interference capability to noise of TTr1 decomposition,mine the global information in the tensor and remove redundant features of tensor,an adaptive filtering truncation tensor reconstruction method is proposed.The method can calculate the global singular values according to the reconstruction parameters,comprehensively measure its contribution to the original tensor,adaptively set the filtering factor,remove the global singular values with low contribution to the original tensor,eliminate the interference information,and retain the important structural information of the tensor.(2)The complexity and nonstationarity of the vibration signal under the complicated working conditions are studied.The principal components of the original signal collected by multi-channel monitoring come from the rotating equipment,the remaining components contain useless information such as environmental noise,and the collected multi-channel signals have similarities in structure and content in the same vibration direction,it indicates that the principal components contain fault information of the equipment.In order to effectively extract the principal components of the collected signal,combining with the tensor decomposition method,which can capture the structural similarity and content similarity of high-dimensional data and mine the potential information to the greatest extent,a feature extraction method of rotating machinery based on TTr1 FS is proposed.The tensor data represented by three different dimensions of time,frequency and channel are established by continuous wavelet transform method.The data are decomposed by TTr1 to obtain the left singular value matrix,singular value matrix and right singular value matrix,respectively.A positive optimizer intelligent algorithm based on probability density function is used to find the optimal value of objective function reconstruction parameters.Using the tensor reconstruction method of adaptive filtering truncation,combining with the left singular value matrix,singular value matrix and right singular value matrix,the reconstruction tensor is obtained.Finally,for the reconstruction tensor,the time-domain signals of different channels are obtained by using continuous wavelet inverse transform.The feasibility and advantages of the proposed method are verified by comparing experiments with other methods.(3)Traditional intelligent diagnosis methods mainly depend on the prior knowledge.In the face of massive heterogeneous data,the features of the extracted vibration signal usually contain useless noise and measurement errors,making it difficult to obtain distinguishable data.To solve these problems,combined with the advantages of tensor decomposition and reconstruction method in mining potential information of fault signals to improve the accuracy of fault identification,an intelligent diagnosis method based on tensor low rank decomposition of synchronous extraction transformation is proposed.A three-dimensional high-order tensor is constructed by synchronous extraction transformation.The high-order tensor is decomposed using TTr1,and an intelligent optimization algorithm is used to solve the optimal reconstruction function.The tensor is reconstructed to obtain a new tensor.Finally,the Alex Net neural network is used to train and test the intelligent diagnosis of rotating machinery by using the 10 fold cross verification method,which improves the accuracy and efficiency of signal feature extraction.Public data sets and experimental data are used to verify the advantages of the proposed method.The results show that the proposed method has great advantages in the fault diagnosis of rotating machinery such as rolling bearings.(4)Rotating machinery will have complex mechanical behavior when the initial failure occurs.In the unbalanced mechanical environment,the other parts will appear fatigue,which will further develop into weak faults and significant faults.At the same time,their vibration signals are coupled with each other.In the case of strong noise environment and complex tramsimssion path,the weak faults are submerged by noise,and further the phenomena of missed diagnosis or misdiagnosis occur.Aiming at these problems,combined with the characteristics of the tensor decomposition that can mine the potential information of signal and extract irrelevant noise,a tensor decomposition fault diagnosis method based on composite multi-scale sample entropy is proposed.By combining tensor decomposition and empirical mode decomposition,tensor decomposition and ensemble empirical mode decomposition,tensor decomposition and complementary empirical mode decomposition,and tensor decomposition and variational modal decomposition,composite multi-scale sample entropy is used to mine the impact signals representing the faults of rotating machinery in the intrinsic mode function for experimental comparison.Experiments proved the effectiveness of this method.