WiFi fingerprinting is a widely used approach for indoor localization, but its effectiveness is limited by signal instability and the challenges of real-time inference. To address these issues, we propose a dual-mode localization framework that combines a sequence-based long short-term memory (LSTM) model with a lightweight multilayer perceptron (MLP) trained through knowledge distillation. This framework supports both sequential and single-snapshot RSS inputs and reformulates localization as a block-wise classification task to enhance robustness. Experiments on two large-scale public datasets show that the distilled MLP model achieves more than 20% improvement in localization error compared to a non-distilled baseline, while maintaining high floor-prediction accuracy. This allows for fast, efficient inference on devices with limited computational resources while maintaining high accuracy levels. The dual-mode design enables adaptive selection based on input availability, which offers a flexible and practical solution for real-world indoor positioning in dynamic environments.

中文翻译：

---

通过时间学习和知识蒸馏进行室内定位的双模框架

WiFi 指纹识别是一种广泛使用的室内定位方法，但其有效性受到信号不稳定性和实时推理挑战的限制。为了解决这些问题，我们提出了一种双模定位框架，该框架将基于序列的长短期记忆（LSTM）模型与通过知识蒸馏训练的轻量级多层感知器（MLP）相结合。该框架支持顺序和单快照 RSS 输入，并将本地化重新表述为按块分类任务，以增强鲁棒性。在两个大规模公共数据集上的实验表明，与未蒸馏基线相比，蒸馏 MLP 模型的定位误差提高了 20%以上，同时保持了较高的底层预测精度。这允许在计算资源有限的设备上进行快速、高效的推理，同时保持高精度水平。双模设计可实现基于输入可用性的自适应选择，为动态环境中的真实室内定位提供了灵活实用的解决方案。