Research On Dynamic Compensation Method For High-g Acceleration Sensor

High-g acceleration sensor is the core device of the penetration system,and its dynamic performance directly affects the performance of the test system.Due to the high-g accelerometer structure characteristics,the measured frequency limit is restricted,the effective bandwidth of the working frequency can not completely cover the acceleration signal,which leads to the penetration test system dynamic error.In order to reduce the error of the test system and improve the accuracy of the test system,it is necessary to make dynamic compensation of the high-g acceleration sensor and broaden its operating frequency band.In this paper,a series of researches were carried out on the dynamic compensation methods of sensors.Firstly,for linear compensation method,the compensation effect based on zero pole phase method was simulated and analyzed from two angles of modeling error and input noise.The results showed that the amplitude error is not related to the frequency bandwidth after dynamic compensation.Secondly,for nonlinear compensation methods,the compensation methods based on BP and RBF neural networks were studied.Taking the sensor output signal as the prototype,a training sample of neural network is constructed to calibrate the excitation source signal of the system,that was to say,the sensor input signal was the expected output of the network.In the course of training,because of the weakness of the training accuracy of the weights and the long training steps,the compensation process time is longer,and the compensation system band was shorter and the compensation result was not satisfactory.An improved variable metric method(L-BFGS)combined with radial basis function(RBF)neural networks was proposed,and the Calman with better real-time performance was added to thecompensation system.Filtering algorithm to reduce the influence of input noise.Through the comparison of three kinds of compensation model of neural network training,analysis results showed that with global stability and improved RBF network compensation algorithm with variable scale method,less influence of input noise,better compensation effect etc.Finally,the radial basis function(RBF)neural network combined with the improved variable scale method(L-BFGS)was applied to the AYZ-4-120 K high-g acceleration sensor.The dynamic calibration of the dynamic calibration data was first carried out by using a high impact station calibration system.The training input samples AYZ-4-120 K sensor output data after pretreatment as the compensation model,the expected sample output standard sensor as the compensation model,another set of calibration data as test data,verify the correctness of the neural network training model.The results showed that the operating band of the AYZ-4-120 K sensor was compensated from the original 7.18 KHz to 44.3KHz,which proves the effectiveness of the RBF network compensation algorithm combined with the improved variable metric method(L-BFGS).