Exploring the Feasibility of Automated Data Standardization using Large Language Models for Seamless Positioning

Accurate seamless positioning is paramount in the era of ubiquitous computing and the Internet of Things (IoT), enabling critical applications such as navigation, asset tracking, and location-based services [1, 2, 3]. The proliferation of IoT devices has exponentially increased the demand for precise and reliable positioning systems. Researchers have explored various indoor positioning techniques, including Bluetooth Low Energy (BLE) beacons [4], Ultra-Wideband (UWB) ranging [5], and inertial sensor-based dead reckoning [6]. Each method presents unique strengths and limitations regarding accuracy, coverage, scalability, and cost.

BLE beacons provide a cost-effective solution with reasonable accuracy but require dense deployment for high precision [4]. UWB offers high accuracy and low latency, suitable for applications needing precise location information, such as industrial asset tracking [5], but is generally more expensive and has a limited range compared to BLE-based systems. Inertial sensor-based dead reckoning relies on accelerometers, gyroscopes, and magnetometers to estimate position changes but suffers from cumulative errors over time, necessitating periodic calibration with other positioning methods [6].

Fusing multiple positioning technologies has become a promising approach to enhance performance and robustness, especially in complex urban environments [7]. Sensor fusion leverages the complementary strengths of different technologies to provide a more accurate and reliable positioning solution. For instance, combining BLE beacon data with inertial sensor data can compensate for the weaknesses of each method, resulting in improved accuracy and robustness.

However, integrating diverse positioning sensors poses significant challenges due to heterogeneous data formats and the need for sophisticated algorithms to handle uncertainties, noise, and interdependencies among different data sources [8]. Traditional sensor fusion systems often rely on manual feature engineering and domain expertise, limiting their scalability and adaptability to new sensor types and environments [9].

Recent advancements in artificial intelligence (AI) and natural language processing (NLP), particularly Large Language Models (LLMs) like GPT-4-0613, offer promising solutions for the standardization of heterogeneous sensor data. These models have demonstrated remarkable capabilities in understanding and generating human-like text, inspiring researchers to explore their potential in other domains. Leveraging LLMs for automated data standardization can significantly reduce manual intervention and enhance scalability.

The flowchart of the proposed feasibility study, depicted in Fig. 1, illustrates the iterative standardization and validation process essential for enhancing seamless positioning systems.

The study begins with collecting unstandardized sensor data from sources like smartphones, IoT devices, and UWB tags. This heterogeneous data is processed by the Intelligent Data Standardization Module (IDSM), leveraging a fine-tuned Large Language Model (LLM) to automate standardization. The IDSM segments incoming data by sensor type and normalizes complex elements (e.g., timestamps to UNIX nanoseconds). The standardized data is formatted according to predefined specifications, as shown in Table I.

Unit tests ensure the accuracy and integrity of the standardized data before progressing. Following standardization, the Transformation Rule Generation Module (TRGM) creates transformation rules, generating scripts for new sensor data standardization, enhancing scalability. The scripts undergo unit testing for functionality and reliability.

Finally, the Extended Kalman Filter (EKF) integrates standardized data from multiple sensors, improving the system’s precision and robustness. Covariance matrices represent uncertainties, which are manually specified but may require adaptive approaches for real-world scenarios.

The IDSM’s primary objective is to transform heterogeneous sensor data $\mathcal{D}$ into a standardized format $\mathcal{S}$. The raw data collected from various sensors is represented as:

where $d_{i}$ denotes the data from the $i$-th sensor. The standardization process is expressed as:

The IDSM was fine-tuned using a curated dataset with 100 complex examples, addressing specific challenges in sensor data standardization. The standardized data schema structure is detailed in Table I. Table II highlights various edge cases and scenarios.

The IDSM was developed and fine-tuned using the Azure platform and the GPT-4-0613 model. Azure’s robust infrastructure provided the computational resources necessary for handling large datasets and performing iterative validations efficiently. This integration allowed for seamless scaling and deployment of the IDSM, ensuring it could handle real-time data processing requirements.

The unit tests validate the standardized dataset $\mathcal{S}$ against predefined JSON schemas for various sensor types. The validation function is expressed as:

where $\nu$ indicates whether $\mathcal{S}$ conforms to schema $\sigma$, and $e$ contains details of validation errors. The iterative process ensures $\mathcal{S}$ conforms to the schema through multiple cycles if necessary, up to a maximum of five iterations, based on empirical evidence and practical considerations.

Most datasets (24 out of 30) required only one iteration for successful validation. Two datasets required up to four iterations, while four datasets did not achieve successful validation within the iteration limit of five. This rapid initial convergence supports five iterations as an optimal limit.

These results underscore the robustness and effectiveness of the IDSM in standardizing diverse sensor data in most cases, ensuring high data quality and consistency. However, the datasets that did not validate within the iteration limit highlight areas for potential improvement in handling more complex data entries.

The TRGM automates deriving transformation rules as detailed in Table III for JSON structures using the GPT-4-0613 model. It converts input JSON files into a specified output format, reducing manual intervention. The process is represented as:

where $\mathcal{S}$ is the standardized data and $\mathcal{I}$ is the input JSON structure. The transformation rules are then used to create a ”Transformation Script for Standardization” $\mathcal{T}$, ensuring consistency in data transformation tasks.

Table III shows the schema structure for the transformation rules used by the TRGM.

The unit tests for the TRGM validate the accuracy of the transformation scripts. The validation function is described as:

where $\nu$ indicates whether the output from $\mathcal{T}$ matches the standardized data $\mathcal{S}$, and $e$ contains error details. The iterative validation process refines the transformation rules through successive refinements until the correct output is achieved.

The Sensor Fusion module employs an Extended Kalman Filter (EKF) to integrate data from multiple sensors, providing accurate, real-time estimates of 3D positions and velocities [11].

The study highlights the efficacy of the proposed feasibility study for real-time automated data standardization using large language models, particularly in enhancing seamless positioning. The Intelligent Data Standardization Module (IDSM) achieved near-zero loss and full accuracy in both training and validation phases, demonstrating robust learning capabilities and reliable data standardization across diverse sensor inputs. The Transformation Rules Generation Module (TRGM) significantly reduced the manual effort required for script generation, enhancing productivity and minimizing human intervention. In terms of positioning accuracy, the data fusion approach outperformed traditional standalone methods, achieving a Root Mean Square Error (RMSE) of 0.35 meters and a Mean Absolute Error (MAE) of 0.25 meters. This improvement is critical for applications necessitating precise and reliable location data. The synergy of IDSM’s high accuracy and TRGM’s automation capabilities presents a powerful solution for sensor data processing, leading to more reliable data and improved decision-making. Despite promising results, the reliance on predefined schemas and manually specified covariance matrices may limit system adaptability in dynamic environments. Additionally, the controlled evaluation setting may not fully capture real-world complexities. Future research should refine this approach for dynamic and complex environmental adaptation and integrate emerging technologies for robustness. Investigate models beyond GPT-4-0613 to understand performance variations, study computational costs and assess real-time versus offline operations.

This research is supported by the University Grants Committee of Hong Kong under the scheme Research Impact Fund on the project R5009-21 “Reliable Multiagent Collaborative Global Navigation Satellite System Positioning for Intelligent Transportation Systems”.