TL;DR
This paper proposes a novel method for classifying Global Navigation Satellite System (GNSS) signals as Line-of-Sight (LOS), Multipath (MP), or Non-Line-of-Sight (NLOS) using correlator-level measurements fed into a Convolutional Neural Network (CNN). This approach achieves state-of-the-art performance, outperforming traditional machine learning methods and offering a promising self-contained solution for improving GNSS accuracy in challenging urban environments.
What Problem Does This Paper Address?
Accurate positioning is crucial for applications like autonomous vehicles and location-based services. However, in urban canyons with tall buildings, GNSS signals are often blocked (NLOS) or reflected (MP), significantly degrading position accuracy. To mitigate these errors, it's essential to first identify the type of signal reception. Existing methods often rely on manually extracted signal features like signal strength, which can be insufficient. This paper tackles the challenge of reliably classifying GNSS signal reception types (LOS, MP, NLOS) by leveraging deeper, more informative measurements from the signal tracking process itself.
Methods
The core of this paper's contribution lies in its novel approach to signal classification. Instead of relying on traditional signal attributes, the authors propose using correlator-level measurements as input to a Convolutional Neural Network (CNN).
Here's a breakdown of the methodology:
- Correlator-Level Measurements: The process begins with a Vector Tracking (VT) loop, a standard technique in GNSS receivers for signal acquisition and tracking. The VT loop generates outputs from multiple correlators: an "early" (E), "prompt" (P), and "late" (L) correlator. These correlators measure the correlation between the incoming signal and locally generated replicas at different code phases. The paper argues that the outputs of these correlators, which capture the detailed shape and amplitude of the correlation function, contain richer information about the signal's reception characteristics than simpler signal attributes.
- Convolutional Neural Network (CNN): A CNN is then employed to automatically learn and extract features from these correlator-level measurements. The CNN architecture consists of:
- Convolutional Layers: These layers apply filters to the input data (correlator outputs) to detect local patterns and features.
- Max-Pooling Layers: These layers down-sample the feature maps, reducing dimensionality and making the network more robust to small variations.
- Fully Connected Layers: These layers take the extracted features and perform the final classification, outputting the probability of the signal being LOS, NLOS, or MP.
This approach moves beyond manual feature engineering, allowing the CNN to discover optimal features directly from the raw signal processing outputs.
Datasets
The authors conducted a field test to collect a GPS L1 intermediate frequency (IF) dataset.
- Dataset Collection: The data was collected at a site surrounded by tall buildings, a typical urban environment where NLOS and MP signals are prevalent.
- Signal Processing and Labeling: The collected IF data was processed using a software-defined receiver (SDR) with a VT loop. Multiple correlators with fine code chip spacing (0.01) were used to generate the correlator output curves. These curves were then labeled as LOS, NLOS, or MP. The labeling was achieved by comparing satellite visibility extracted from a 3D map and ephemeris data with the signal processing results. Specifically, certain satellites (PRN 3 and 22) were identified as NLOS, while others (PRN 23 and 27) were identified as multipath signals.
- Training and Testing: The dataset was split into training and testing sets. 50 seconds of data were processed for both training and testing, with 10,000 samples selected from the testing dataset for evaluation.
Key Insights and Ideas
- Correlator Outputs as Rich Features: The paper strongly advocates for using correlator-level measurements as input for signal classification. This is a significant departure from methods relying solely on signal strength or other derived metrics, as correlator outputs directly reflect the signal's interaction with the environment.
- CNNs Excel at Feature Extraction: The success of the CNN demonstrates its power in automatically learning complex patterns from the correlator outputs, eliminating the need for manual feature engineering. This is particularly beneficial for distinguishing between the subtle differences in signal characteristics caused by multipath and non-line-of-sight conditions.
- MP Signals are More Complex: The results indicate that MP signals are the most challenging to classify, likely due to the superposition of direct and reflected signals, leading to more complex correlator output curves. This suggests that more sophisticated models or features might be needed for highly accurate MP detection.
- Simpler CNNs Can Be Effective: For LOS and NLOS classification, a CNN with two convolutional layers proved sufficient, suggesting that overly complex architectures might not always be necessary and could even lead to slight performance degradation.
Key Results
The proposed CNN-based method using correlator-level measurements achieved state-of-the-art performance in classifying GNSS signal reception types.
- Overall Accuracy: The best-performing CNN configuration (Cov1D (2,2), Cov1D (4,2), MaxPooling (2,2)) achieved classification accuracies of 85.38% for NLOS, 84.91% for MP, and 88.48% for LOS on the testing dataset.
- Comparison with Traditional Methods: The CNN significantly outperformed traditional machine learning classifiers like kNN and SVM, which used manually selected features (e.g., signal-to-noise ratio, pseudo-range residual). For instance, the CNN achieved higher accuracies across all classes compared to the best kNN and SVM results presented.
- Performance Breakdown: LOS signals were classified with the highest accuracy, while MP signals had the lowest, consistent with the insight that MP signals are more complex.
Limitations and Future Work
The authors acknowledge several limitations and propose future research directions:
- Dataset Size and Diversity: The current dataset, while sufficient for demonstrating the method's efficacy, is limited in size and collected from a single site. This could impact the generalizability of the model. Future work should focus on collecting larger, more diverse labeled datasets.
- Environmental Variations: The current model might not perform as well in environments with different reflection and refraction characteristics than those encountered during data collection. Transfer learning could be explored to enhance the classifier's versatility.
- Noise in Correlator Outputs: The raw GNSS signals and correlator outputs contain noise. The paper did not implement any de-noising operations. Applying de-noising techniques to the correlator outputs could potentially improve classification accuracy.
- Advanced Deep Learning Models: While CNNs are effective, exploring other advanced deep learning architectures like Capsule Neural Networks (Cap-NN) could lead to further performance improvements.
- Generative Models for Data Augmentation: Generating synthetic datasets using Generative Adversarial Networks (GANs) or employing few-shot learning techniques could help overcome the labor-intensive nature of creating large labeled datasets.
Summary updated