Calibration Technology And Application Of Current Sensor Based On Improved BP Neural Network

In response to accidents caused by the imbalance of carbon brush current caused by long-term operation of large generators in the power industry,online monitoring of carbon brush current and temperature is proposed.However,the measurement of current sensors is severely affected by temperature and environment,resulting in measurement accuracy deviation.Therefore,a combined optimization method(M-RMSP)is used for calibration.This dissertation introduces the topology and basic principles of BP neural network,and conducts performance testing on the network.In response to the problems of local minima,single learning rate,and slow convergence speed in BP neural networks,the momentum gradient descent method and adaptive learning rate algorithm are proposed for optimization,and the M-RMSP optimization algorithm is further proposed.The M-RMSP optimization method is used to fit a nonlinear function.Through comparative analysis of simulation,it is found that the algorithm improved by momentum Gradient descent and Adaptive learning rate method improves the stability of the network,but the Rate of convergence of the network is still slow.The combination of the two algorithms can improve the stability and Rate of convergence of the network,and the Learning rate can also be changed adaptively.The M-RMSP optimization algorithm is used to fit different nonlinear functions and target accuracies.The experimental results show that the algorithm can meet the requirements of the fitting problem.Finally,collect the test data of WCS1500 Hall current sensor,simulate using the improved BP network of M-RMSP algorithm,and compare and analyze the experimental results.And use the improved M-RMSP algorithm for sensor calibration experiments to measure the measured values of the sensor.Compare and analyze the output measured at different temperatures with the output measured by BP calibrated and uncalibrated sensors.A series of experiments have shown that the improved M-RMSP algorithm performs significantly better than traditional BP neural networks.This method can be applied to engineering practices such as carbon brush current monitoring systems for large steam turbine generators.