Long short-term memory network of machine learning for compensating temperature error of a fiber optic gyroscope independent of the temperature sensor

We present an artificial intelligence compensation method for temperature error of a fiber optic gyroscope (FOG). The difference from the existing methods is that the compensation model finally determined by this method only uses the FOG's data to complete the regression prediction of the temperature error and eliminate the dependency on the temperature sensor. In the experimental stage, the proposed method performs temperature experiments with three varying trends of temperature heating, holding, and cooling and obtains sufficient output data sets of the FOG. Taking the output time series of the FOG as the input sample and based on the long short-term memory network of machine learning, the training, validation, and test of the model are completed. From the two perspectives of network learning ability and the improvement degree of the FOG's performance, four indicators, including root mean square error, error cumulative distribution function, FOG bias stability, and Allan variance analysis are selected to evaluate the performance of the compensation model comprehensively. Compared with the existing methods using temperature information for prediction and compensation, the results show that the error compensation method without temperature information proposed can effectively improve the accuracy of the FOG and reduce the complexity of the compensation system. The work can also provide technical references for error compensation of other sensors.