This study introduces a novel fluxgate current sensor with a compact, ring-shaped configuration that exhibits improved performance through the integration of magnetization residence times and neural networks. The sensor distinguishes itself with a unique magnetization profile, denoted as M waves, which emerge from the interaction between the target signal and ambient magnetic interference, effectively enhancing interference suppression. These M waves highlight the non-linear coupling between the magnetic field and magnetization residence times. Detection of these residence times is accomplished using full-wave rectification circuits and a Schmitt trigger, with a digital output provided by timing sequence detection. A dual-layer feedforward neural network deciphers the target signal, exploiting this non-linear relationship. The sensor achieves a linearity error of $0.054\%$ within a measurement range of 15 A. When juxtaposed with conventional sensors utilizing the residence-time difference strategy, our sensor reduces linearity error by more than 40-fold and extends the effective measurement range by $150\%$. Furthermore, it demonstrates a significant decrease in ambient magnetic interference.

Accurate and efficient data-driven prediction of embedded airborne load separation trajectories and attitudes can not only significantly improve the safety of the separation process but also substantially reduce reliance on costly aerodynamic simulations and wind tunnel testing. This paper proposes a real-time condition fusion Transformer (RCF-Transformer) model for predicting the trajectory and attitude after load separation. Using wind-tunnel datasets of separation events obtained from Captive Trajectory Simulation (CTS), the model encodes historical sequence information while dynamically injecting real-time input conditions measured at the moment of separation into the decoder. Masked multi-head self-attention and cross-attention mechanisms are employed for collaborative learning, enabling multi-step, multi-output prediction of three-axis position and attitude. Experimental results show that, for a multi-step prediction horizon of up to $T=5$, the proposed model achieves an overall prediction accuracy of 95.28%. Furthermore, error-structure analyses based on Theil’s inequality coefficient decomposition, confidence intervals, and F-tests of residual variances demonstrate that the residuals are dominated by nonsystematic, high-frequency fluctuations and that the performance gains over the strongest baseline are statistically significant. These results indicate that the proposed method is highly stable and robust, providing an efficient and scalable data-driven solution for safety monitoring and decision support during the initial separation of airborne loads.