A Hybrid GA-FNN Optimization Framework for Enhancing Temperature-Invariant Performance of MEMS Gyroscopes

The MEMS gyroscope is a core component of the inertial navigation system in precision-guided weapons. It is a micro-mechanical angular rate sensor realized using Micro-Electro-Mechanical Systems (MEMS) technology [1]. Among them, the quad-mass MEMS gyroscope, due to its fully symmetric structure, can better suppress common-mode signal interference. Addressing the significant variation in the output of quad-mass MEMS gyroscopes across the full temperature range, a temperature compensation method based on the GA-FNN fusion algorithm is proposed. This method leverages the powerful global search and swarm intelligence optimization capabilities of the Genetic Algorithm (GA). By setting the mean squared error (MSE) of the validation set fitting as the penalty fitness for evolution, and using the hidden layer structure of the Feedforward Neural Network (FNN) as the optimization target, the algorithm adaptively obtains the optimal neural network structure through iterative selection, crossover, and mutation. The weights between layers are updated using the Levenberg-Marquardt algorithm. Experiments show that within the temperature range of -40°C to 60°C, the average fitting residual of this method is 0.8234. After compensation, the bias instability is reduced from the initial 12.0249°/h to 3.2960°/h, and the bias stability is improved from 4357.05°/h to 13.11°/h. Compared with the 124.00°/h achieved by the traditional polynomial fitting compensation method, this represents a 9.46-fold enhancement. The performance improvement is significant, and the temperature hysteresis effect is also notably mitigated. This demonstrates that the proposed method effectively optimizes the bias performance of the quad-mass MEMS gyroscope over the full temperature range.