Research And Implementation Of High Precision Compensation Algorithm For Crystal Oscillator Based On Neural Network

As the clock frequency source of all kinds of electronic equipment and instruments,the crystal oscillator with high precision and high stability determines the synchronization accuracy of clock communication in military industry,base station,aerospace and other fields,and directly affects the synchronization operation of various electronic components in the system.With the rapid development of 5G technology,communication base station presents a trend of miniaturization and lightening,so the requirements of crystal oscillator miniaturization,low power consumption and low cost are increasingly obvious.In the miniaturized crystal oscillator represented by temperature compensated crystal oscillator(TCXO),it is often necessary to traction the output frequency of different temperature points,so as to avoid the large influence of environmental temperature change on the frequency output of the crystal oscillator,so as to achieve the purpose of compensation.In the existing scheme,usually scheme for analog and digital compensation adopted by the combination of two kinds of scheme to realize the temperature characteristic curve of ascension,and the corresponding algorithm often has certain limitations,under the environment of high and low temperature compensation and compensation value theory usually there is a certain gap,the accuracy of TCXO cannot reach ideal,In order to achieve the ideal accuracy in both high and low temperature conditions,compensation often needs to be repeated,resulting in an increase in compensation costs.In this paper,the simulation compensation and digital compensation algorithms of TCXO crystal oscillator are improved,and the simulation compensation and digital compensation scheme based on neural network is proposed.The accuracy of TCXO is improved on the existing basis,which is used to achieve the flow scheme of reducing compensation cost and increasing compensation accuracy.First of all,starting from the existing compensation process of TCXO,analog and digital compensation by crystal oscillator tuning data to control the output frequency,the generated data sets in two kinds of compensation phase,with depth within DNN and RBF neural network RBF neural network training,to ensure that the revised compensation data can more in line with the actual product needs.Secondly,the compensation problem in the simulation compensation stage is abstracted into the actual multiple linear regression problem,and the output frequency of TCXO at room temperature is taken as the nominal frequency.In the simulation compensation stage,several modification parameters and temperature characteristic curves of the same type of crystal oscillator are modeled and trained.In the actual process,the unified model is established through DNN forward propagation and reverse gradient update,and the unified compensation scheme is output according to the generated model.Compared with the traditional compensation scheme,it can effectively save time and determine the generation of the unified scheme.Finally,the digital compensation phase compensation problem,by means of RBF neural network fitting for compensating the offset voltage required to generate and temperature range of the compensation voltage interpolation is lacking,compared with the traditional compensation scheme,the proposed scheme can more effectively to ensure that the compensation of the curve generated by TCXO can satisfy the real operating environment to run,The frequency difference of the crystal oscillator at different temperatures is shrunk and corrected so as to achieve the best temperature compensation.In order to ensure that the scheme designed in this paper can meet the needs of improving the actual accuracy as much as possible,the corresponding compensation modification data need to be written into the crystal oscillator for verification after the neural network training.Finally,in the experimental stage,by writing the data of analog compensation and digital compensation respectively,the experiment verifies that the compensation algorithm based on neural network can obtain higher compensation effect than the traditional compensation algorithm.In the analog compensation stage,the compensation accuracy can reach 0.2ppm under wide temperature condition.In the digital compensation stage,the wide temperature condition can reach 0.05 ppm.