Fault Diagnosis Of Bearing Based On Nonlinear Weighting Dual-kernel Extreme Learning Machine

After the rapid development of railways in recent years,the Passenger-dedicated Line operation mileage exceeded 15,000 kilometers last year,and the cargo transportation volume reached 55% of the national cargo transportation volume,the railway density ranks first in the world,which also raises people's demand for safe operation of high-speed railway.As a mechanical component with large load,bearing faults plays a vital role in train operation.In the past ten years,the number of train derailments caused by bearing damages and peeling faults has averaged 50 times per year all over the world.Therefore,it is of great practical significance to monitor,predict and detect the faults of locomotive bearings.Due to the low signal-to-noise ratio of the rolling bearing,the fault signal is often overwhelmed by noise during the diagnosis process,and the useful information is difficult to obtain.Therefore,the general time domain analysis method cannot obtain the current fault information in the bearing signal and judge the bearing fault state accurately in time,which may easily lead to missed inspection or maintenance delay.The method of regular inspection has a high detection cost,so a prompt and reliable inspection is the main target of current research.The neural network classification method has a high fit with bearing diagnosis,and the performance of which has been improved year by year with rapid development of the computer hardware.Kernel Extreme Learning Machine(K-ELM)is very suitable for dealing with such problems because of its unique large-scale parallel structure,distributed storage and strong adaptability and pattern recognition capabilities.To processing rolling bearing vibration signal,we could extract features by using time-frequency analysis and Multi-scale Permutation Entropy(MPE)feature extraction methods.Then,Compressing the feature by using principal component analysis(PCA)and Distance Evaluation Technique(DET).Finally,introducing the compressed feature as input into the neural network classifier for training,a corresponding bearing signal fault diagnosis model can be established.However,Extreme Learning Machine still has some disadvantages as a classifier.The number of hidden layer nodes in the ELM needs to be artificially set,and the accuracy and stability of the classification are worse than other single hidden layer feed-forward neural networks.K-ELM has higher requirements for the kernel function selection.The singlekernel function is difficult to fully learn the nonlinear samples,and the generalization is still insufficient.A MPE-based multi feature extraction method and nonlinear weighted combination Dual-Kernel Extreme Learning Machine(DK-ELM)is proposed and its feasibility has been proved.Firstly,the time,frequency and entropy parameters of bearing signals in different fault states are calculated to form a series of dimensionless features.Then,the high-dimensional eigenvectors are calculated by using the dual-kernel function and input into the DK-ELM to establish the bearing signal state classification model.So as to complete the classification task of bearing signals in different states..Finally,the bearing fault diagnosis model is visualized,and the trained parameters of different bearing models are classified and pre-stored,so as to quickly identify the state of unknown bearing signals in working conditions,greatly enhance the practicality of the algorithm.The above method was verified by diagnosing the open source bearing data of the Western Reserve University and the working condition data of locomotive.The experiment's results show that the classification model of DK-ELM has higher accuracy than SVM,ELM and single-core extreme learning machines.Moreover,the DK-ELM has better classification ability for the case of a smaller number of samples.