【摘要】 针对低成本微惯性测量单元在旋转弹姿态估计中存在的精度不高、噪声敏感及环境适应性差等关键问题，提出了一种基于无迹卡尔曼滤波(UKF)与长短期记忆网络、卷积神经网络(CNN)及注意力机制的UKF-LSTM-CNN-Attention融合模型(ULCA)。该模型通过多模态信息融合，显著提升了旋转弹在复杂环境下的姿态估计精度和鲁棒性。在方法设计上，首先，利用UKF处理系统的非线性状态方程，保证基础估计精度。然后，引入长短期记忆网络捕捉姿态变化的时序动态特征，结合卷积神经网络提取传感器数据的空间局部特征。最后，通过注意力机制自适应地加权关键信息，有效抑制噪声干扰。为全面验证算法性能，设计了系统的仿真实验，在不同工况下，对比分析了ULCA模型与传统UKF以及扩展卡尔曼滤波的性能差异。结果表明：ULCA模型在滚转角、俯仰角和偏航角估计中较传统算法平均降低了59.89%的误差率，且在强噪声环境下表现出更强的鲁棒性。理论分析和实验验证表明，ULCA融合模型有效解决了传统滤波算法在复杂环境下的建模偏差问题，在提高姿态估计的精度和适应性方面取得了显著进展。更多还原

【Abstract】 To address the critical challenges of insufficient accuracy, noise sensitivity and poor environmental adaptability in low-cost micro inertial-measurement-unit(MIMU)for spinning projectile attitude-estimation, a novel fusion model named UKF-LSTM-CNN-Attention(ULCA)was proposed. The ULCA model integrates unscented Kalman filtering(UKF),long short-term memory networks, convolutional neural networks(CNN)and attention mechanisms to achieve multimodal information fusion, significantly improving attitude estimation accuracy and robustness under complex conditions. In the methodology, UKF was first employed to handle the nonlinear system dynamics, ensuring baseline estimation precision. LSTM was then introduced to capture temporal dependencies in attitude variations, while CNN extracted spatial-local features from sensor data. Finally, the attention mechanism adaptively weighted critical information to suppress noise interference effectively. To comprehensively evaluate the algorithm, systematic simulations were conducted under various operational conditions, comparing ULCA with traditional UKF and extended Kalman filter(EKF). Results demonstrate that the ULCA model reduces the average estimation error by 59.89% in roll, pitch and yaw angle estimation compared to traditional algorithms, exhibiting superior robustness in high-noise environments. Theoretical analysis and experimental validation show that the ULCA fusion model effectively addresses modeling biases of conventional filters in complex scenarios, achieving remarkable progress in precision and adaptability for attitude estimation.更多还原