Research On The State Discrimination Method Of Structural Parts Based On Acoustic Emission Signal

Among various nondestructive testing methods,acoustic emission testing technology has become the first choice in the field of real-time monitoring of equipment status because of its ability of dynamic detection of material and structure damage signals.As a passive detection method,acoustic emission detection method has low requirements for the proximity of the tested object,high sensitivity,wide frequency range of the detected signal,and can detect the early damage of the tested object.In order to realize the dynamic monitoring of the health status of mechanical equipment,it is the premise to obtain the damage signal,but,more importantly,how to analyze the obtained signal is more worthy of attention.With the current signal feature extraction methods becoming more and more complete,the basis of judging damage is more abundant.However,this trend is accompanied by a significant increase in the number and dimensions of signal data.Therefore,this paper regards the signal data of AE detection as a big data type,establishes a mathematical model based on the improved radial basis function neural network(RBFNN),and verifies the feasibility of this method through the AE experiment of 45 steel.The main research contents and conclusions are as follows:(1)the RBFNN model belongs to a special type of feedforward neural network,which has consistent approximation for nonlinear continuous functions and is especially suitable for nonlinear classification problems.Considering that the variable selection of RBFNN is essentially a kind of noise reduction process,an improved RBFNN(ST-RBFNN)based on soft threshold is proposed,which is similar to the widely used dropout method in the model training process.Through the verification of numerical experiments,ST-RBFNN effectively reduces the over fitting of the model and greatly improves the accuracy of the validation data set.(2)according to the optimization theory,the first-order optimization algorithm can be applied to the mathematical model with convex loss function to obtain the global optimal solution.Therefore,the loss function of RBFNN is proved by convex function,so that the stability of its solution can be guaranteed under the first-order optimization algorithm.In order to make the iterative search direction more precise and the result more accurate,the Adam algorithm is improved.A first-order optimization algorithm based on compound gradient,compound gradient method(C-Adam),is proposed and its convergence is proved.(3)the performance of C-Adam algorithm is tested by three groups of open datasets.The test results show that this method has great advantages over other common methods in convergence speed and accuracy.In order to improve the expansibility of C-Adam and its compatibility with penalty function,this paper proposes specific improvement measures for ridge regression penalty problem,and explores other forms of weight attenuation in the algorithm.(4)the acoustic emission signals of static tensile test and fatigue test of 45 steel are collected,and the mathematical model based on experimental data are constructed by using STRBFNN combined with C-Adam.The experimental results show that the fitting accuracy of the two experimental data is over 99.84%,the verification accuracy of static tensile test is 99.31%,and the verification accuracy of fatigue test is 99.80%...