CSI-fingerprinting based human indoor localization in noisy environments using time-invariant CNN

Hyogo Hiruma

Research & Development Group
Hitachi, Ltd.

Indoor localization in noisy environments

Indoor localization, the process of determining precise indoor locations, is crucial for enhancing productivity, safety, and user experience in various sectors, including construction, industry, and retail. In Hitachi, we use indoor localization technology to efficiently collect field data for our Worksite-Augmenting Metaverse system, directly supporting frontline workers with information with spatial accuracy. Traditional methods of indoor localization, such as trilateration, which depend on measuring Wi-Fi signal strength (RSSI), require clear sightlines and numerous access points, making them impractical in obstacle-rich environments. Fingerprinting techniques using Wi-Fi Channel State Information (CSI) have emerged as promising alternatives, as CSI provides unique signal patterns for each location. However, real-world conditions often produce noisy, fluctuating CSI data, reducing accuracy (Fig. 1). In this blog I would like to introduce a new method designed to handle noisy CSI signals effectively.

Time-invariant CNN model

Our proposed localization system incorporates a novel Time-Invariant Convolution (TI Conv) Block within a modified CNN structure (Fig. 2). CNNs, or Convolutional Neural Networks, are a type of neural network commonly used to automatically learn meaningful features from data by applying convolution operations. They excel at extracting spatial and temporal patterns, making them well-suited for processing CSI signal data. In our model, the CNN initially extracts basic CSI features through standard convolutional layers and advanced residual and attention-enhanced blocks. These layers effectively capture local and global CSI patterns, which are essential for distinguishing locations.

The core innovation of our method lies in the TI Conv Block, designed specifically to address the noisy and fluctuating nature of real-world CSI data. We observed that CSI measurements tend to fluctuate significantly over time due to environmental interference and signal reflections, making it critical to selectively extract meaningful features from temporal sequences rather than relying on continuous or stable patterns. Unlike traditional convolutional methods that rely heavily on temporal consistency, the TI Conv Block focuses on the frequency of occurrence of specific CSI patterns rather than the order in which the patterns appear. This frequency-based extraction enables the model to robustly recognize locations even when data is highly inconsistent. The TI Conv Block thus ensures high accuracy and stability by selectively emphasizing meaningful patterns while disregarding temporal noise.

Experiment at indoor construction environment

We evaluated our model's performance in a realistic indoor construction environment characterized by significant signal interference from metal structures. CSI data were collected at multiple points in a 14m × 14m area using a Wi-Fi router as the transmitter and a mobile receiver (Raspberry Pi 4B). We systematically collected training data at predefined points, and gathered testing data at intermediate locations.

The experiments focused particularly on verifying robustness against noisy CSI data. We compared our method against traditional CNN and RNN-based localization methods using standard performance metrics such as localization accuracy (average Euclidean distance error) and stability (variance of predictions). Our proposed model demonstrated superior accuracy, consistently outperforming all comparative models. Notably, our model exhibited a remarkably low prediction variance, clearly indicating its robustness in handling fluctuating, noisy data. These results confirm that our frequency-focused TI Conv Block effectively maintains high accuracy and stability in challenging real-world scenarios.

Conclusion

In this blog, we introduced a robust indoor localization method leveraging a novel TI Conv Block, specifically designed to address the challenges posed by noisy, real-world CSI data. This approach shows great promise for industries where precise indoor localization directly contributes to operational efficiency, worker safety, and better resource management. By reliably mitigating the effects of signal interference and inconsistent data, the technology can streamline processes, enhance safety protocols, and support better decision-making onsite. We invite readers to delve into the original paper for a comprehensive explanation of the technical details and learn about potential future developments.

Acknowledgements

I would like to acknowledge my co-authors, co-workers from Hitachi Plant Construction: Yuichi Yashiro, Toshihiro Sugajima and Fumio Hatori, with whom this research was conducted.