Previous Abstract Return to Session B3 Next Abstract

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Dense Autoencoder-Based Anomaly Detection for GNSS CORS Measurement Quality Monitoring
Rene Manzano, Elena Tai, National Autonomous University of Mexico; Kyle O'Keefe, University of Calgary
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 11:03 a.m.

Continuous Operating Reference Stations (CORS) provide correction data and raw measurements enabling centimeter-level accuracy in real-time positioning applications. CORS networks are classified in two categories, publicly available networks operated by governmental and research entities, and commercial networks managed by private organizations [1]. The generated correction data is widely used in real-time kinematic (RTK), precise point positioning (PPP), and network-based solutions [2]. However, the integrity of these measurements must be continuously ensured, as faults may impact the computed position of the station and propagate through dependent positioning systems, potentially affecting numerous users simultaneously. Traditional monitoring strategies rely on fixed threshold tests applied independently to parameters such as code residuals, signal-to-noise ratio (SNR), or cycle-slip indicators [3]. While effective for detecting gross failures, these approaches may fail to detect subtle, multivariate, or slowly evolving anomalies that manifest as correlated deviations across several measurements simultaneously. Moreover, the implementation of real-time anomaly detection algorithms is challenging due to the non-linear patterns exhibited by the parameters analyzed in the measurements. Conventional CORS quality control methods typically employ statistical hypothesis testing based on fixed thresholds and Gaussian assumptions. Such approaches treat each observable independently and assume stationary error distributions. However, GNSS measurement errors often exhibit non-linear correlations between pseudorange residuals, code-minus-carrier (CMC) combinations, satellite geometry, and SNR. These multivariate dependencies are difficult to capture using classical threshold-based monitoring or linear statistical models.
Recent research in GNSS quality assessment has explored machine learning methods for multipath detection, interference monitoring, and integrity analysis. In [4], a deep learning approach for classifying multipath ranging error from a GNSS receiver correlation function is used. Similarly, a supervised machine learning based method for detecting non-line-of-sight multipath using GNSS signal correlation output from the receiver is shown in [5]. Many of proposed approaches rely on supervised learning techniques that require labeled fault data and a priori signal correlation information. However, in CORS infrastructure monitoring, labeling fault conditions is neither practical nor efficient, given the rarity and variability of anomalous events. Unsupervised methods such as autoencoders address this limitation by learning the nominal station behavior directly from normal-operation data, enabling effective anomaly detection without requiring labeled fault examples.
Autoencoders are a class of artificial neural networks designed to learn compact and informative representations of data in an unsupervised manner [6], and consist of two main components, an encoder that maps the input data into a lower-dimensional latent representation, and a decoder that reconstructs the original input from this compressed encoding. By minimizing the reconstruction error between the input and the output, the network learns the intrinsic structure and correlations present in the training data without requiring labeled examples [7]. Autoencoders have proven effective for anomaly detection tasks, where the model is trained on data representing nominal system behavior. In [8], a GNSS interference monitoring approach employing different hybrid autoencoder models using SNR observations is proposed. The authors show the effectiveness of the autoencoder, combining it with recurrent neural network models for jamming and spoofing event detection. In order to consider temporal correlation in GNSS anomaly data detection, the combination of Long Short-Term Memory (LSTM) networks with autoencoders is made in [9]. The results obtained from the combination of these machine learning models are comparable with the equivalent supervised method’s 95% accuracy.
In this paper, we propose a Dense Autoencoder neural network for unsupervised anomaly detection in CORS stations. Unlike conventional univariate monitoring techniques, the proposed method simultaneously analyzes six measurement features:
1. Pseudorange residuals computed using the known fixed station coordinates as a position constraint
2. SNR
3. Code-minus-carrier
4. Cycle-slip indicators
5. Satellite elevation
6. Satellite azimuth
The autoencoder is trained exclusively using nominal data. By minimizing reconstruction error, the network learns the intrinsic multivariate structure of healthy CORS behavior, implicitly modeling correlations between measurement geometry, signal strength, and residual patterns.
In Figure 1, a diagram of the proposed autoencoder model is shown. The input data is a six-dimensional vector, composed of the the pseudorange residuals, the SNR, CMC, cycle slip indicators, satellite elevation and satellite azimuth, which is compressed progressively through two encoder layer of 64 and 32 nodes, employing a Rectified Linear Unit (ReLU) activation functions into a compact representation of the original vector. The compress vector is then expanded back to its original dimension, reconstructing the original vector. The model is trained using nominal-condition measurements which results on an internal multivariate structure of nominal data. This enables the identification of any measurement that deviates from the nominal behavior.

Previous Abstract Return to Session B3 Next Abstract