FSFDS: enhancing flight sensor fault diagnosis via diffusion and self-attention networks

Aircraft fault diagnosis relies on flight sensor data, with accuracy critical for safe operation. Deep neural networks (DNNs) have shown success in many fields, but their application in fault prediction faces challenges: 1) DNN-generated fault data differs significantly from real faults; 2) existing DNN-based classification models exhibit suboptimal accuracy. This paper proposes a flight sensor fault diagnosis system (FSFDS) using diffusion and attention networks. We enhance data quality with a diffusion model, introducing a scoring function for improved score matching. The generated data undergoes manual annotation and trains an attention-based diagnostic model with a weight-sharing multi-twin neural network to increase training samples. The attention mechanism extracts parameter relationships from time series, boosting accuracy. Deploying the model on an FPGA achieves high energy efficiency. Experiments show FSFDS improves diagnostic accuracy by 15.3% and speeds inference 14.23× over CPUs and 2.08 × over GPUs.