Session D6: Navigation Using Environmental Features

Extended LTE Based Fingerprinting Positioning for Emergency Applications by utilizing Seq2seq Model with Beam-Search Inference
Sun Sim Chun, Jung Ho Lee, Ju-Il Jeon, Jin Ah Kang, Young-Su Cho, Electronics and Telecommunications Research Institute
Alternate Number 2

Positioning technology for emergency rescue must provide accurate location information in any space, indoors or outdoors. Since timing-based positioning technology must ensure LOS (Line of Sight) to ensure accuracy, there is a problem in that countless RF signal sources must be installed indoors or in complex urban areas. The Fingerprinting method, which estimates location based on a database built by utilizing the strengths of both the straight wave component and the multipath component of the signal, is more suitable for emergency positioning technology because it is relatively less sensitive to the multipath problem of RF signals. WiFi Fingerprinting has the disadvantage of requiring database updates due to frequent location changes of the AP (Access Point) and that it cannot be used in out-of-town areas where there is no AP. In this paper, the goal is to develop emergency rescue positioning technology using Fingerprinting technology and applying the LTE network among mobile communication technologies capable of communicating in both indoor and outdoor areas.
Features of the LTE signal applicable to Fingerprinting include information such as CID (Cell ID), PCI (Physical Cell ID), and RSRP. In the current LTE environment, only CID and PCI information for the subscribed telecommunication company can be obtained. However, if you can obtain CID and PCI information for other unsubscribed telecommunications companies, more accurate positioning is possible because the resources available for positioning increase. Therefore, in this paper, we propose a deep-learning-based pseudo cell generation technology that estimates the CID information of other telecommunication service providers by using the acquired CID information of a specific telecommunication service provider and improves the positioning accuracy by utilizing the redundancy characteristics of the wireless space.
In order to take advantage of the redundancy characteristics of wireless space, instead of LTE data information from only one telecommunication service provider, LTE information from three service providers, which are all the service providers available in Korea, is used. A pseudo cell composed of spatial redundancy can be predicted using information from all the service providers. It is possible to predict various pseudo cells using LTE fingerprinting information, and this research team has already proposed predicting the CIDs of the remaining two service providers using the LTE CID of one service provider. When we conducted a test to predict the CIDs of two service providers using the LTE CID of one service provider using the SGNS (Skip-Gram Negative Sampling) model [1] proposed by this research team, the similarity top 5 (when CIDs of two telecommunication companies are included in the selected top 5), the prediction result showed an accuracy of 95%. However, when experimented with the condition of similarity top 2 (when CIDs of two telecommunication companies are included in the selected top 2), the performance was very poor with an accuracy of 54%.
In this paper, to improve the performance of similarity top 2, we propose replacing the existing SGNS model with the Seq2Seq model. The Seq2Seq model consisted of an encoder-decoder structure and used attention, input feeding, and teacher forcing functions. The encoder and decoder are composed of stacked RNN (or LSTM). The Seq2Seq model uses the LTE CID of one service provider as an input and performs learning under the condition of predicting the CIDs of two carriers. During inference, the LTE CID of one service provider are used as the input of the encoder, and the context vectors generated by the encoder are used as the input of the decoder. The output of the decoder predicts the CIDs of the two service providers with the highest probability by applying a greedy search method. The data used in the experiment was collected three times in total in Seocho 1-dong, Seocho-gu, Seoul. A mobile data-collection device was attached to the vehicle to collect LTE wireless information from all the domestic telecommunication service providers: KT, SKT, and LGU+. Among the collected data, only about 12,000 Serving Cell information was extracted and used for the experiment. 10% was used for inference and the remaining 90% was used as training data. In the inference process, KT CID excluding redundancy were used as input values, and the cumulative count of the selected learning data was calculated to determine the ground truth for the inference data based only on the case with the highest cumulative count. Accuracy was calculated using the formula, that is, Accuracy = (number of accurate top2 similarity test) / (number of KT CIDs). When using the dataset obtained through the above process, the prediction accuracy for similarity top 2 was 77%, which did not yield satisfactory results. This is because the Seq2Seq model has an auto-regrsssive characteristic in which the current state is determined depending on the past state, so the methods for learning and inference are different, and the greedy search method is judged to be insufficient to reflect this characteristic. Therefore, in this paper, we applied the beam search method instead of greedy search to the inference part of the model, and as a result, the prediction accuracy of similarity top 2 was improved to 93.75%. The advantage of the beam-search method over greedy-search method is that it is performed by selecting candidates with high probability at each inference point (stage) as many as the number of beams (k), and this is repeated until the decoders of all batches receive input, and the cumulative probability is obtained by multiplying the conditional probabilities of the selected nodes at the time of inference.
In conclusion, in this paper, the accuracy of similarity top 2 was improved from 54% to 93.75% by using the Seq2Seq learning model, which can compensate for the shortcomings of existing fingerprinting-based positioning technology [2], and a learning method using the redundancy characteristics of wireless space. Improvement in the performance of inference results was possible from the following three factors. First, in order to take advantage of the redundancy characteristics of wireless space, LTE information from three service providers, which constructs pseudo cell and are all ones available in Korea, is used instead of the LTE data information of only one service provider. Second, the existing SGNS (Skip-Gram Negative Sampling) model was changed to the Seq2Seq model. Lastly, a beam search method suitable for Seq2Seq was applied to the inference part of the model instead of the commonly used greedy search.
Reference
[1] Chun, S., Jeon, J., Kang, J., Cho, Y. , 2023 IPNT, Prediction Method for Pseudo LTE Cell Information Based on Skip-Gram Model, Nov 1-3 2023, Shinwa World, Jeju, Korea, https://ipnt.or.kr/2023proc.php
[2] Noh, H., Oh, Y., Lee, N., Shin, W., A Survey of Deep Learning-Assisted Indoor Localization with Wi-Fi Fingerprinting: Current Status and Research Challenges, The Journal of Korean Institute of Communications and Information Sciences '21-05 Vol.46 No.05, https://doi.org/10.7840/kics.2021.46.5.848