The demand for a ubiquitous and accurate indoor localization service is continuously growing. Cellular-based systems are a good candidate to provide such ubiquitous service due to their wide availability worldwide. One of the main barriers to the accuracy of such services is the large number of models of cell phones, which results in variations of the measured received signal strength (RSS), even at the same location and time. In this paper, we propose OmniCells++, a deep learning-based system that leverages cellular measurements from one or more training devices to provide consistent performance across unseen tracking phones. Specifically, OmniCells++ uses a novel approach to multi-task learning based on LSTM encoder–decoder models that allows it to learn a rich and device-invariant RSS representation without any assumptions about the source or target devices. OmniCells++ also incorporates different modules to boost the system’s accuracy with RSS relative difference-based features and improve the deep model’s generalization and robustness.

Evaluation of OmniCells++ in two realistic testbeds using different Android phones with different form factors and cellular radio hardware shows that OmniCells++ can achieve a consistent median localization accuracy when tested on different phones. This is better than the state-of-the-art indoor cellular-based systems by at least 148%.

对无处不在且准确的室内定位服务的需求不断增长。由于基于蜂窝的系统在全球范围内广泛可用，因此它们是提供这种无处不在的服务的理想选择。此类服务准确性的主要障碍之一是大量型号的手机，这会导致测量的接收信号强度 (RSS) 发生变化，即使在相同的位置和时间也是如此。在本文中，我们提出了OmniCells++，这是一种基于深度学习的系统，它利用来自一个或多个训练设备的蜂窝测量数据，在不可见的跟踪电话上提供一致的性能。具体来说，OmniCells++使用基于 LSTM 编码器 - 解码器模型的多任务学习新方法，使其能够学习丰富且设备不变的 RSS 表示，而无需对源或目标设备进行任何假设。OmniCells++还结合了不同的模块，通过基于 RSS 相对差异的特征来提高系统的准确性，并提高深度模型的泛化性和鲁棒性。

OmniCells++在两个实际测试台中使用不同外形尺寸和蜂窝无线电硬件的不同 Android 手机对OmniCells++ 的评估表明，在不同手机上进行测试时，OmniCells++可以实现一致的中值定位精度。这比最先进的室内蜂窝系统好至少 148%。