In this paper, the structure and working principle of four-mass vibration MEMS gyroscope (FMVMG) are introduced, and the working modes of FMVMG are simulated and analyzed. On the basis of this, an improved noise reduction method based on interval local mean decomposition (ILMD) and parabolic tracking time-frequency peak filtering (PTTFPF) is proposed. PTTFPF can resample the signal along a parabolic path and select the optimal filtering trajectory, but there is still a contradiction, choosing a short window length may lead to good signal amplitude retention, but the random noise reduction effect is not good, while choosing a long window length may lead to serious amplitude attenuation, but the random noise reduction effect is better. In order to achieve a better balance between effective signal amplitude preservation and random noise reduction, the ILMD method was used to improve PTTFPF. First, the original signal was decomposed into product functions (PFs) by local mean decomposition (LMD) method, and the sample entropy (SE) of each PF was calculated. The PFs are divided into three different components. Then, short window PTTFPF is used for useful PF and long window PTTFPF is used for mixed PF, noise PF is directly removed. Then the final signal is reconstructed. Finally, the denoised useful PF and mixed PF are reconstructed to obtain the final signal. The proposed ILMD-PTTFPF algorithm was verified by temperature experiments. The results show that the denoising performance of the ILMD-PTTFPF algorithm is better than that of traditional wavelet threshold denoising and Kalman filtering.

This paper presents an improved empirical modal decomposition (EMD) method to eliminate the influence of the external environment, accurately compensate for the temperature drift of MEMS gyroscopes, and improve their accuracy. This new fusion algorithm combines empirical mode decomposition (EMD), a radial basis function neural network (RBF NN), a genetic algorithm (GA), and a Kalman filter (KF). First, the working principle of a newly designed four-mass vibration MEMS gyroscope (FMVMG) structure is given. The specific dimensions of the FMVMG are also given through calculation. Second, finite element analysis is carried out. The simulation results show that the FMVMG has two working modes: a driving mode and a sensing mode. The resonant frequency of the driving mode is 30,740 Hz, and the resonant frequency of the sensing mode is 30,886 Hz. The frequency separation between the two modes is 146 Hz. Moreover, a temperature experiment is performed to record the output value of the FMVMG, and the proposed fusion algorithm is used to analyse and optimise the output value of the FMVMG. The processing results show that the EMD-based RBF NN+GA+KF fusion algorithm can compensate for the temperature drift of the FMVMG effectively. The final result indicates that the random walk is reduced from 99.608°/h/Hz1/2 to 0.967814°/h/Hz1/2, and the bias stability is decreased from 34.66°/h to 3.589°/h. This result shows that the algorithm has strong adaptability to temperature changes, and its performance is significantly better than that of an RBF NN and EMD in compensating for the FMVMG temperature drift and eliminating the effect of temperature changes.

In this paper, a fault identification algorithm combining a signal processing algorithm and machine learning algorithm is proposed, using a four-mass vibration MEMS gyroscope (FMVMG) for signal acquisition work, constructing a gyroscope fault dataset, and performing the model training task based on this dataset. Combining the improved EWT algorithm with SEResNeXt-50 reduces the impact of white noise in the signal on the identification task and significantly improves the accuracy of fault identification. The EWT algorithm is a wavelet analysis algorithm with adaptive wavelet analysis, which can significantly reduce the impact of boundary effects, and has a good effect on decomposition of signal segments with short length, but a reconstruction method is needed to effectively separate the noise signal and effective signal, and so this paper uses multiscale permutation entropy for calculation. For the reason that the neural network has a better ability to characterize high-dimensional signals, the one-dimensional signal is reconstructed into a two-dimensional image signal and the signal features are extracted. Then, the constructed image signals are fed into the SEResNeXt-50 network, and the characterization ability of the model is further improved in the network with the addition of the Squeeze-and-Excitation module. Finally, the proposed model is applied to the FMVMG fault dataset and compared with other models. In terms of recognition accuracy, the proposed method improves about 30.25% over the BP neural network and about 1.85% over ResNeXt-50, proving the effectiveness of the proposed method.