Research On Fault Diagnosis Method Of High Speed Automata Based On Transient Impulse Response Signal

For weapons that are subjected to cyclic stress for a long time,high-speed impact,work under high temperature,high pressure or harsh environment,cracks are prone to occur in their parts.It is an important method to diagnose the fault of high-speed impact weapons by testing and analyzing the vibration signal generated by them.However,the working environment of high-speed impact weapons is very bad,there are many moving parts in them,and the components of the vibration signal are complex.There are not only the determined signal components based on the transient impact signal,but also a large number of random signal components containing high noise.These make the effective crack information of the vibration signal carrying parts weak,and it is impossible to extract the characteristics of the vibration signal by using a single classical theory.At present,there are few theories for extracting the characteristics of the impact transient signal.The classical frequency-domain processing methods and the time-frequency-domain processing methods are often unable to directly and effectively process this kind of complex vibration signal with the impact transient signal as the main component.Based on these classical theories and combined with the new theory,the effective features of the impact transient signal can be extracted better.In addition,machine learning and integrated learning are needed for the extracted features in diagnosis,so the learning ability of the learner and the integrated theory adopted should also be studied.In this paper,a high-speed gun was taken as a typical representative to study the method of diagnosing the fault of high-speed impact weapon based on the impact transient signal,and to diagnose the fault of high-speed automaton of the high-speed gun.Specifically,three typical cracks of locking mechanism of high-speed automata were diagnosed.Firstly,the impact transient signal was preprocessed to eliminate the invalid signal segment by using the method of motion form decomposition.Secondly,the research was carried out from three aspects: extracting the signal characteristics,optimizing the learner and integrating the learner.Finally,the fault of high-speed automatic machine was diagnosed by using the integrated learner.In the aspect of extracting the features of impulse transient signals,the integral upper limit transformation method,probability density function method,the concept of information expression force and the concept of information difference degree to extract the features of signals were proposed in this paper.The extremum problem of information expression force of discrete system,and the method of obtaining differential entropy and differential expression force of signal when only the discrete value of upper limit function of probability density integral is known,but not the probability density function were also studied.Based on the proposed method and concept,ergodicity index,power spectrum function and Mallat discrete wavelet analysis,six groups of features were extracted for the impulse transient signal,among which the energy ratio feature group based on Mallat discrete wavelet analysis is the control group.The analysis of the impact transient signal and its features shows that:(1)The signal is nonstationary,but it can be analyzed by the way of processing stationary signal.Its frequency component is complex,so it is difficult to extract the effective features directly from the time spectrum,Hilbert amplitude spectrum and edge spectrum of the spectrum,continuous wavelet analysis.(2)The integral upper limit transformation method proposed in this paper can change the intensity of each frequency component of the signal.After the signal is processed by the integral upper limit transformation,the frequency spectrum of the same state signal shows a certain rule.The quality of the extracted integral spectrum feature group is better than that of the control feature group when learning based on the fuzzy neural network.(3)The probability density function method proposed in this paper deals with the complex impact transient signal in the time domain with the statistical concept.The frequency amplitude feature group extracted based on the intrinsic mode functions IMF1 and IMF2 and the probability density function method shows the best quality among all feature groups when learning based on the support vector machine,which fully confirms the effectiveness of the probability density function method.(4)The quality of differential expression based on spectrum is better than that of differential expression based on integral upper limit transformation.The quality of differential expression based on Hilbert Huang instantaneous frequency is better than that of differential expression based on spectrum extraction and differential entropy.At the same time,differential expression and differential entropy have the same ability of indicating signal state.(5)The concept of information difference degree is reasonable,but the quality of information difference degree is not good,which reflects the complexity of signal from the side again.In the aspect of optimizing and integrating the learners,the Moore Penrose inverse Newton algorithm to train the fuzzy neural network was proposed in this paper.The fuzzy neural network with five layers structure was constructed,and tested with random features.For the extracted feature groups,the fuzzy neural network and support vector machine were used to learn,and the sub-learning were obtained.The sub-learner was integrated hierarchically by Bayes integration theory to get the integrated learner.The integrated learner was used in this paper to diagnose three typical cracks in the locking mechanism of the high-altitude machine gun.At the same time,the cross-validation method was used to train and test the learner.The results show that:(1)Moore-Penrose inverse Newton method has better convergence and convergence speed than Levenberg-Marquardt algorithm.(2)Under the condition of small sample size,the prediction ability of support vector machine is higher than that of fuzzy neural network.Although the fuzzy neural network is in over-learning state,increasing the number of nodes of the fuzzy neural network can improve the prediction ability.(3)The Bayes ensemble theory adopted in this paper can effectively improve the prediction accuracy of the learner.In this paper,the fault of high-firing machine gun can be diagnosed with 83.93% prediction accuracy.Considering the experimental conditions,the diversity of working methods and the high difficulty of fault diagnosis,this prediction accuracy is still very ideal.