Research On Pattern Identification And Classification Visualization Of Typical Faults From Rotating Machinery

With the continuous advancement of rotating machinery in the direction of advanced intelligence,how to improve the decision-making level of intelligent fault diagnosis is our latest challenge.Fault pattern identification is the most critical step in fault diagnosis.It can effectively and accurately excavate abundant equipment fault information,find the distribution area and boundary of different types of sample groups in the feature space,and provi de the most important reference and guidance for the subsequent construction of decision knowledge base.Therefore,it is the research to train a classifier that can effectively identify faults and has a certain generalization ability and fault tolerance,pay attention to the uncertainty of failure occurrence(collectively,randomness and ambiguity),and the random amount in the fault data.Which were key issues that wer concerned.This study focuses on mechanical fault classification,with specific concern s as follows: the improvement of base classifier differences when ensemble learning is used for fault identification,the expression and calculation of uncertainty knowledge in fault data,and the visual expression in fault classification process.The spec ific research contents and the research results are as follows :(1)Aiming at the problem of low difference between base classifiers in Bagging ensemble learning,an improved ensemble neural network algorithm was proposed and applied to fault pattern identification of double-span rotor.That is,the characteristics of low-dimensional rotor fault data set were disturbed by Relief-F algorithm and improved roulette method,and the training sets were disturbed by Bagging algorithm.After two perturbations,the training sets input to each base classifier had differences in feature space and sample space,which made the trained base classifier have higher differences and made the final classification result more reliable.Finally,the identification performance of the BP network-based double disturbance ensemble learning method was verified on the low-dimensional double span rotor fault data set.The results showed that the double disturbance method could significantly improve the identification ability of BP neural network.(2)Aiming at the problems of the uncertainty and non-stationarity of the bearing vibration signals and the slow convergence speed and poor stability of the BP neural network learning algorithm,a method for identifying rolling bearing failur e modes based on cloud models and ensemble extreme learning machines was proposed.The preprocessed signals were clouded to generate signal clouds of rolling bear ings in different states,Three parameters that determine the distribution of the signal cloud were extracted: expectation,entropy,and superentropy.These three parameters were used as characteristics.Finally,the fault feature data set was normalized and sent to the ensemble extreme learning machine for identification.The research results showed that the cloud-ensemble extreme learning machine method could effectively achieve bearing fault pattern recognition.Compared with traditional neural network recognition methods,this method had higher recognition rate and stability,and the integrated extreme learning machine was resistant to noise.(3)In order to make full use of the information contained in massive data and effectively identify bearing faults,a cloud theory method was used to map the bearing fault data with its corresponding fault type,and established the cloud distribution of various characteristics of rolling bearings in different states.Based on the model,a cloud judgment knowledge base for bearing failure was constructed.At the same time,the Relief-F algorithm was introduced to determine the weight coefficients of each feature of the training set.Combined with the cloud distribution membership coefficient,a calculation method for the final membership of the sample for bearing failure was proposed.The comparison of the cl assification accuracy of the cloud classification knowledge base established by different numbers of training samples proved that the method had the ability to learn the data;the comparison method was used to compare the classification method with the com monly used classification method on the test samples containing noise.The results proved the superiority of this classification method in terms of noise immunity.Cloud theory can visualize the expression of semantic concepts,unify ambiguity and randomness in data,and has broad prospects in the field of fault identification in the context of mechanical big data.Subsequent research will focus on exploring the use of cloud models to derive virtual samples to achieve imbalanced learning and using cloud model to achieve fault data dimensionality reduction.