Abstract:

Addressing the issue of data anomalies in radar countermeasure reconnaissance signals caused by susceptibility to interference in complex electromagnetic environments， this paper proposes a multidimensional temporal intelligent fusion network （MDTFusionNet）. The architecture integrates temporal convolutional network（TCN）， long short-term memory（LSTM）， and self-attention mechanisms with a gated network that dynamically adjusts module weights through a gated network， and optimizes model robustness by combining weight sparsity constraints. This constructs an anomaly detection model capable of not only capturing short-term fluctuations in radar signals and grasping long-term trends but also dynamically focusing on critical pulse information. To verify its effectiveness， multi-layer perceptron（MLP）， TCN， and LeNet networks are used for comparison， and evaluations are conducted from the perspectives of loss function and accuracy. Experimental results show that the loss function outliers of MDTFusionNet are significantly smaller than those of traditional models， and the accuracy is higher. Ablation studies further confirm that each component in MDTFusionNet has its respective role， enabling better learning of temporal data feature distributions and accurate anomaly detection， verifying its superiority and practicality in radar countermeasure reconnaissance signal anomaly detection.

Key words: temporal convolutional network（TCN）, long short-term memory（LSTM）, gated network, radar countermeasure reconnaissance, pulse description word （PDW）

摘要：

针对雷达对抗侦察信号在复杂电磁环境中易受干扰导致的数据异常问题，提出多维时序智融网络（MDTFusionNet）。网络架构先融合了TCN，LSTM及自注意力机制，又通过门控网络动态调整模块权重，并结合权重稀疏性约束优化模型鲁棒性，构建了一个既能捕捉雷达信号短期波动、把握长期趋势，又能动态关注关键脉冲信息的异常检测模型。为验证其有效性，采用MLP，TCN及LeNet网络与其进行对比，并从损失函数和准确性评估。实验结果表明，MDTFusionNet的损失函数异常值显著小于传统模型，准确率更高。消融实验进一步证明MDTFusionNet中每个模块都有其各自作用，能更好地学习时序数据特征分布，准确检测异常，验证了其在雷达对抗侦察信号异常检测中的优越性和实用性。

关键词: 时间卷积网络（TCN）, 长短期记忆（LSTM）, 门控网络, 雷达对抗侦察, 脉冲描述字