Feature Space Entropy Weighted Nuisance Attribute Projection And Its Application In Mechanical Fault Diagnosis

The fault diagnosis technology of mechanical equipment is of great significance to the actual engineering safety assurance,where in key point is how to extract effective fault features.Traditional feature space projection methods rely on reducing the dimension of the feature vector to extract sensitive features contained in the fault signal,which are mainly used for fault feature extraction under constant operating conditions.In the face of actual operation process of mechanical equipment,operating conditions are often complex and changeable.Mechanical equipment operates under constant or various operating conditions,and the signals collected under various operating conditions are more complex and nonlinear.However,when dealing with mechanical vibration signals under various operating conditions,the fault features are masked by redundant interference,which makes traditional feature space projection methods fail in extracting fault features.Besides,multi-channel signals have the advantages of more comprehensive information collection and more complete reflection of equipment status,which gain much attention from researchers in the current research on mechanical fault diagnosis.Nonetheless,multiple sensors are used in the multi-channel signal acquisition process.The redundant interference is caused between different acquisition channels,which makes it difficult to accurately extract fault features.Toward the above problems,this dissertation systematically researches the key technologies based on feature space nuisance attribute projection method,including feature extraction,feature space projection and multivariate signal processing.Moreover,this dissertation proposes robust fault diagnosis methods,which effectively solves the redundant interference caused by various operating conditions and multiple acquisition channels.Above methods provides an effective solution for the fault diagnosis of mechanical equipment under various operating conditions,and their main contents are briefly described as follows:1.Aiming at the problem in nuisance attribute projection(NAP)that the interference is either taken into consideration in whole,or not considered at all,which will inevitably lead to information loss,entropy weighted nuisance attribute projection(EWNAP)is proposed.The proposed method constructs an improved weighted matrix by quantitatively estimating the interference degree of each operating condition,and solves the "bipolar problem" in NAP.Then,the proposed method is combined with back propagation(BP)neural network to eliminate redundant interference caused by various operating conditions,and achieves accurate fault diagnosis under various operating conditions.Experiments using public datasets and data collected from the experimental bench show that the proposed method has the advantage of high diagnosis accuracy under various operating conditions.2.Aiming at the problem that processing high-dimensional features extracted by EWNAP,entropy weighted nuisance attribute projection-orthogonal locality preserving projection(EWNAP-OLPP)is proposed.The proposed method uses orthogonal locality preserving projection(OLPP)to compress high-dimensional features processed by EWNAP into low-dimensional space,to eliminate redundant interference,and mine and retain sensitive features hidden in high-dimensional space.The sound signal samples processed by the proposed method are input into the BP neural network,and a fault diagnosis model is constructed under various operating conditions.Experiments show that the proposed method can accurately extract sensitive features related to the nature of the fault,and improve the accuracy and efficiency of the fault diagnosis model under various operating conditions.3.Aiming at the problem that the traditional multivariate signal processing method cannot accurately extract essential features due to redundant interference between acquisition channels when processing multi-channel signals,this dissertation proposes a multi-channel signal processing method based on EWNAP and adaptive projection intrinsically transformed multivariate empirical mode decomposition(APIT-MEMD).The method first uses APIT-MEMD to decompose the multi-channel signal and calculate the sample entropy value of intrinsic mode functions of each channel,and the feature space containing the main fault information is obtained.Then,redundant interference between channels is eliminated by using EWNAP to process the obtained feature space,and accurate essential features of multi-channel signals are obtained.The experimental results show that the proposed multi-channel signal processing method can effectively eliminate the redundant interference caused by the acquisition channel,and improve the accuracy of multi-channel signal feature extraction.4.Aiming at the problem that existing redundant interference caused by operating conditions and acquisition channels when processing multi-channel signals under various operating conditions,a multi-channel signal fault diagnosis method based on EWNAP,APIT-MEMD and weighted extreme learning machine(WELM)is proposed.The samples processed by APIT-MEMD and EWNAP are input into the proposed WELM,and features the most relevant to the nature of the fault are obtained through the processing of the weighted matrix.Finally,the samples are trained and tested to obtain fault classification results.The public data set and the multi-channel data collected from the experimental bench are used in the experiment.The experimental results show that the multi-channel signal fault diagnosis method can extract accurate fault features from multi-channel signals under various operating conditions,and achieve high-precision fault classification.