Research On Fault Diagnosis Method Based On Morphological Filtering And Manifold Learning

With the proposal of "Manufacturing Power" and other strategies,Chinese mechanical equipment gradually tends to be systematized and complicated,and its failure modes become more complex and diverse.The rolling bearing is one of the most common and easily damaged parts of industrial equipment.If it fails,it may lead to spindle fracture,transmission failure,and other situations,and then lead to production stagnation,resulting in incalculable losses,and even can cause serious safety accidents.Therefore,the diagnosis and prediction of the operating characteristics,fault type,location,damage degree of rolling bearing has become one of the hot spots of current research.As a non-linear system,the vibration signal of mechanical equipment can be used to describe the operation state of the system,so the vibration signal analysis method is usually used to study and analyze the mechanical equipment fault.However,in the complex operating environment,the bearing vibration signal presents the characteristics of non-linear and non-stationary.In addition,the influence of environmental noise and other factors makes the extraction of bearing fault features more difficult.The traditional signal processing technology has certain limitations in calculation efficiency and accuracy.Therefore,this paper adopts morphological filtering and manifold learning algorithm,taking rolling bearing as the research object,focusing on the filtering preprocessing of bearing fault vibration signal,the construction of fault eigenvector space and its dimensionality reduction,the training of fault diagnosis model and other aspects,and finally solves the key problems of signal preprocessing,feature extraction,feature dimensionality reduction and so on in mechanical equipment fault recognition ? Firstly,the intrinsic time scale decomposition(ITD)and linear local tangent space are proposed Based on the combination of alignment and lltsa,a new method of fault diagnosis is proposed,which is based on adaptive generalized morphological filtering and lltsa,and based on generalized morphological difference filtering(gdif)and auto encoder network(an).The specific research content and innovation points mainly include the following:(1)A fault diagnosis method combining ITD with LLTSA is proposed to deconstructs and reconstructs the bearing vibration signal,which can filter out the noise,extract the high-dimensional feature vector set and realize the dimension reduction.Finally,thebearing fault diagnosis model based on the extreme learning machine(ELM)is trained.The recognition rate of the model is 99.21%.The experiment result shows that the method extracts successfully the effect fault vibration signal with low noise and decrease the redundancy caused by high-dimensional features,and realize the effective recognition of the bearing running state.(2)To extract high-dimensional features of bearing fault vibration signals with non-stationary and non-linear features from the background of strong noise,solve the problem of signal redundancy of high-dimensional feature vectors,and improve the accuracy of the model,a fault diagnosis method based on adaptive generalized morphological filtering and LLTSA is proposed.The accuracy of ELM model trained by this method is as high as 99%.Experiments show that the ELM bearing state monitoring model built by this method can effectively filter out the noise and reduce the redundancy caused by high-dimensional features,realize the effective analysis and diagnosis of the running state of the bearing.(3)To further solve the problem that the traditional noise reduction method has many control parameters and large calculation amount,a method of fault diagnosis based on generalized morphological difference filtering(GDIF)and auto-encoder network(AN)is proposed.Firstly,the GDIF is used to denoise the vibration signal,and the low-dimensional essential manifold is obtained from the high-dimensional characteristics of the signal through AN manifold learning algorithm,to alleviate the dimension disaster of the high-dimensional characteristics.The results of bearing experiments in open data-sets show that this method can effectively reduce the computational complexity,filter out useless information and interference noise in signals,and reduce the dimension of features through AN,which can effectively improve the classification accuracy of ELM model to 98.04%.