KE-DDPM: Kinematics-Enhanced Denoising Diffusion Probabilistic Model for Inertial Measurement Unit Anomaly Detection

Inertial Measurement Unit is a key sensor of the aerospace vehicles. Undetected anomalies can cause severe attitude and navigation errors. Existing model-based and deep learning anomaly detection methods often ignore the kinematic relationship between IMU outputs and attitude. Moreover, they also impose a high computational cost on edge computing devices within flight control systems. To address these problems, this paper proposes a kinematics enhanced denoising diffusion probabilistic model named KE-DDPM for unsupervised IMU anomaly detection. First, a lightweight temporal encoder and decoder map IMU sequences into a low dimensional latent space where a diffusion model is trained under a joint loss that combines standard noise prediction with an attitude kinematic residual based on quaternion integration. In the detection process, an anomaly score is constructed from both reconstruction error and kinematic residual, providing physically interpretable indicators. Experiments on a satellite simulation dataset and a UAV real flight dataset show that KE-DDPM improves accuracy and F1 score and reduces false alarm rate compared with representative anomaly detection methods. A real-time implementation on a typical edge computing device demonstrates that the method satisfies real-time constraints while maintaining anomaly detection performance.