NIGWO-iCaps NN: A Method for the Fault Diagnosis of Fiber Optic Gyroscopes Based on Capsule Neural Networks

When using a fiber optic gyroscope as the core measurement element in an inertial navigation system, its work stability and reliability directly affect the accuracy of the navigation system. The modeling and fault diagnosis of the gyroscope is of great significance in ensuring the high accuracy and long endurance of the inertial system. Traditional diagnostic models often encounter challenges in terms of reliability and accuracy, for example, difficulties in feature extraction, high computational cost, and long training time. To address these challenges, this paper proposes a new fault diagnostic model that performs a fault diagnosis of gyroscopes using the enhanced capsule neural network (iCaps NN) optimized by the improved gray wolf algorithm (NIGWO). The wavelet packet transform (WPT) is used to construct a two-dimensional feature vector matrix, and the deep feature extraction module (DFE) is added to extract deep-level information to maximize the fault features. Then, an improved gray wolf algorithm combined with the adaptive algorithm (Adam) is proposed to determine the optimal values of the model parameters, which improves the optimization performance. The dynamic routing mechanism is utilized to greatly reduce the model training time. In this paper, effectiveness experiments were carried out on the simulation dataset and real dataset, respectively; the diagnostic accuracy of the fault diagnosis method in this paper reached 99.41% on the simulation dataset; the loss value in the real dataset converged to 0.005 with the increase in the number of iterations; and the average diagnostic accuracy converged to 95.42%. The results show that the diagnostic accuracy of the NIGWO-iCaps NN model proposed in this paper is improved by 13.51% compared with the traditional diagnostic methods. It effectively confirms that the method in this paper is capable of efficient and accurate fault diagnosis of FOG and has strong generalization ability.