Research On Fatigue Life Prediction Of Medium Carbon Steel Based On Neural Network

With the rapid development of scientific level,the application of large-scale machinery and equipment in people’s life has also increased.The condition monitoring of large equipment is closely related to people’s life.For example,automobiles,trains,ships,airplanes,they bring a lot of convenience to our life and travel,but because their assembly often requires tens of thousands of parts,any part damage will affect the operation of the machine,and even endanger life.In order to keep the equipment stable for a long time,their quality assurance is particularly important.Based on the tensile test of medium carbon steel fatigue specimens,the acoustic emission signals collected by acoustic emission sensors are processed to extract the characteristics of acoustic emission signals,and then the residual life of different batches of medium carbon steel is estimated.The main research work in this paper is as follows:The first is the experimental part.The material selected for the experiment is Q-235 low carbon steel.Before the beginning of the experiment,black paint was sprayed on the middle part of the specimen to increase the infrared radiation rate and reduce the infrared reflection of the specimen surface.Acoustic emission sensor is installed in the symmetrical center of the specimen to capture the acoustic emission signal generated when the specimen initiation crack.Acoustic emission signal is processed by preamplifier.When the difference between high and low levels is greater than 14,the acquisition card will continuously collect the acoustic emission signal.Otherwise,the acquisition card will collect the data once in 15 minutes,one second at a time,and the frequency of the fatigue testing machine is 10 HZ.At the same time,Flir-A325sc infrared thermal imager was used to photograph the temperature of the specimens and record the temperature of the specimens every second.The sampling period frequency of the experiment is 100,000.It is set to collect 10 cycles of AE signals per second.The sampling point is 1 million.The data is too large to be pretreated.The acoustic emission signals are extracted from the time domain,and the seven time domain features of ringing count,absolute energy,root mean square,signal variance,absolute mean,root square amplitude and information entropy are extracted respectively.The same trend is observed for different batches of samples.Observing the image of AE signal at 6 periods,extracting the frequency domain features of AE signal,and extracting the five features of AE signal gravity center frequency,mean square frequency,root mean square frequency,frequency variance and frequency standard deviation respectively by signal power spectrum transformation.The BP neural network model is established,and the input data is the feature extracted above,and the output data is the residual life of medium carbon steel.By analyzing and comparing the training effects of different hidden layer nodes and different transfer functions,the first transfer function is logsig,the second transfer function is purelin,and the network structure is 12-12-1.At the same time,considering that BP neural network may not be able to guarantee the global search,PSO optimization algorithm is used to select the network parameters,which are mapped to the network,and the generalization ability of the network is compared and analyzed.Wavelet neural network includes two types:discrete and compact.The discrete wavelet neural network mainly decomposes the data through wavelet transform.In this paper,the acoustic emission data is decomposed by wavelet multi-resolution.The extracted wavelet coefficients are composed of vector groups.The square sum of the whole vector group is made.Four input features are extracted.Then the acoustic emission data are decomposed by wavelet multi-resolution decomposition.It carries out life assessment.The compact wavelet neural network mainly replaces the activation function with the wavelet basis.The wavelet basis selected in this paper is Morlet.Finally,two kinds of wavelet neural networks are used to evaluate the samples,and the prediction results of several different kinds of neural networks are analyzed and compared.The results show that the compact wavelet neural network has better prediction performance.