WiLoc: Massive Measured Dataset of Wi-Fi Channel State Information with Application to Machine-Learning Based Localization

1. Problem Statement

Machine Learning (ML)-based localization using Channel State Information (CSI) is a promising alternative to traditional methods (like GPS or Time-of-Arrival) which often fail in indoor environments or require expensive hardware. However, the development of robust, high-accuracy ML localization algorithms is currently hindered by a lack of large-scale, realistic, and diverse datasets.

Existing public datasets suffer from severe limitations in three critical dimensions:

• Scale: They typically cover only one or two buildings with a small number of Access Points (APs) (e.g., <200 APs) and limited User Equipment (UE) locations.
• Diversity: They lack variety in environmental types (e.g., mixing indoor/outdoor, different building structures) and AP configurations.
• Data Volume: They often contain insufficient data points to prevent overfitting or to effectively train complex deep learning models and transfer learning strategies.

Synthetic data (ray-tracing) is an alternative but often fails to capture real-world complexities like diffraction and scattering accurately. Consequently, there is an urgent need for a massive, real-world measured dataset to train and validate ML algorithms.

2. Methodology

The authors conducted a massive, three-month precision measurement campaign at the University of Southern California (USC) campus to create the WiLoc dataset.

A. Data Collection Strategy

• Transmitters (TXs): Instead of deploying dedicated, calibrated transmitters (which is cost-prohibitive and logistically impossible at scale), the authors utilized the existing USC campus Wi-Fi network. This provided thousands of naturally deployed APs in diverse environments (indoor corridors, outdoor streets).
  • Constraint: To avoid interference and unpredictable transmission times, the system captured only beacon frames (transmitted every ~100ms) which are standard, non-beamforming, and regular.
• Receiver (RX): A custom-built cart equipped with a Universal Software Radio Peripheral (USRP) (Software-Defined Radio).
  • Advantage: Unlike consumer Wi-Fi cards that scan channels sequentially (missing data), the USRP digitized the entire 100 MHz ISM band simultaneously, capturing continuous CSI from all APs on channels 1, 6, and 11 without gaps.
  • Precision: The cart used a dual-phase optical encoder to track movement distance with high precision, synchronized with the RF data. This avoided the drift issues common in Inertial Measurement Units (IMUs).
• Mobility: The cart moved along straight-line trajectories (indoor corridors and outdoor streets) at controlled speeds (approx. 0.45 m/s).
  • Indoor: 16 buildings, 180 corridors (1.5–3m wide).
  • Outdoor: >30 streets, varying widths (3–15m).

B. Data Processing

• Signal Separation: Post-processing involved applying circular frequency shifts and digital low-pass filters to isolate specific channels (1, 6, 11).
• Packet Extraction: Legacy Short/Long Training Fields (L-STF/L-LTF) were used to identify packet starts and extract CSI.
• Labeling: Beacons were filtered by BSSID/ESSID to ensure only campus APs were included. Timestamps were matched with cart position data to create "CSI – Location – AP" triplets.

3. Key Contributions

The paper's primary contribution is the release of the WiLoc dataset, which is orders of magnitude larger than any existing public CSI dataset.

Dataset Statistics:

• UE Locations: >12 million labeled locations (6.0M indoor, 6.5M outdoor).
• Access Points: >3,000 unique APs (1,417 indoor, 2,739 outdoor).
• Coverage:
  • Indoor: 16 different buildings with varying floor plans and 180 corridors.
  • Outdoor: >30 streets covering ~450,000 m².
• Metadata: Includes channel index, timestamp, distance, speed, heading, AP labels, coordinates (X, Y, Floor), RSSI, Noise, SINR, and repetition indices.

Technical Innovations:

• Simultaneous Multi-Channel Capture: Overcoming the "channel sweeping" limitation of standard sniffers.
• Scalable Deployment: Leveraging existing infrastructure to achieve massive AP diversity without the cost of dedicated TX deployment.
• High-Precision Ground Truth: Using optical encoders and laser levels to ensure accurate location labeling, distinct from the noisy data of smartphone-based collection.

4. Results and Validation

The authors validated the dataset and demonstrated its utility through two main case studies:

A. Dataset Quality Verification

• Temporal Stability: Even with people moving in the environment, CSI magnitude correlation remained above 95% for 80% of cases, confirming the stability of the measurements.
• Beacon Consistency: ~85% of beacon intervals were nominal (102.4 ms), with less than 5% of cases losing more than one beacon.

B. Machine Learning Performance

• Supervised Learning (Indoor):
  • Using a Feed-Forward Neural Network (4 hidden layers, 512 neurons), the model achieved Root Mean Squared Error (RMSE) comparable to state-of-the-art results.
  • AP Count Impact: Localization accuracy improved significantly as the number of APs increased up to a threshold of ~5 APs. Beyond this, accuracy degraded slightly due to data sparsity (fewer locations see >5 APs simultaneously).
• Transfer Learning:
  • Cross-Floor: Pre-training on data from multiple floors (1–4) allowed the model to achieve high accuracy on a target floor (5) with very few target samples (<1,000).
  • Cross-Building: Pre-training on multiple buildings (10–15) improved performance on a target building (16) compared to single-building pre-training. However, cross-building transfer showed higher RMSE than cross-floor transfer, highlighting the difficulty of generalizing across different structural environments.
  • Key Finding: Large, diverse source datasets significantly reduce the amount of target data required for fine-tuning, making ML localization feasible in data-scarce environments.

5. Significance

• Standardization: WiLoc establishes a new benchmark for the ML localization community, moving beyond small-scale, single-building experiments to large-scale, real-world validation.
• Algorithm Development: It enables the training of more robust algorithms that can generalize across different environments, a critical step for commercial deployment.
• Transfer Learning: The dataset proves that diverse source data is the key to solving the "data scarcity" problem in new environments, facilitating the adoption of ML localization in real-world scenarios where collecting massive new datasets is impractical.
• Open Science: By making this massive dataset publicly available, the authors provide a foundational resource that allows researchers to compare algorithms fairly and accelerate progress in wireless localization.

Limitations Noted:

• AP locations inside rooms are confidential (NLOS only for indoor), preventing Time-of-Arrival (ToA) based algorithm testing.
• Measurements were taken on straight lines, not grids.
• The dataset represents a specific city's architectural style and does not include human-handheld device variations (e.g., phone model differences).