M

machine learning

Federated learning (FL) enables multiple devices to collaboratively train a global model while maintaining data on local servers. Each device trains the model on its local server and shares only the model updates (i.e., gradient weights) during the aggregation step. A significant challenge in FL is managing the feature distribution of novel, unbalanced data across devices. In this paper, we propose an FL approach using few-shot learning and aggregation of the model weights on a global server. We introduce a dynamic early stopping method to balance out-of-distribution classes based on representation learning, specifically utilizing the maximum mean discrepancy of feature embeddings between local and global models. An exemplary application of FL is orchestrating machine learning models along highways for interference classification based on snapshots from global navigation satellite system (GNSS) receivers. Extensive experiments on four GNSS datasets from two real-world highways and controlled environments demonstrate that our FL method surpasses state-of-the-art techniques in adapting to both novel interference classes and multipath scenarios.

Indoor positioning based on 5G data has achieved high accuracy through the adoption of recent machine learning (ML) techniques. However, the performance of learning-based methods degrades significantly when environmental conditions change, thereby hindering their applicability to new scenarios. Acquiring new training data for each environmental change and fine-tuning ML models is both time-consuming and resource-intensive. This paper introduces a domain incremental learning (DIL) approach for dynamic 5G indoor localization, called 5G-DIL, enabling rapid adaptation to environmental changes. We present a novel similarity-aware sampling technique based on the Chebyshev distance, designed to efficiently select specific exemplars from the previous environment while training only on the modified regions of the new environment. This avoids the need to train on the entire region, significantly reducing the time and resources required for adaptation without compromising localization accuracy. This approach requires as few as 50 exemplars from adaptation domains, significantly reducing training time while maintaining high positioning accuracy in previous environments. Comparative evaluations against state-of-the-art DIL techniques on a challenging real-world indoor dataset demonstrate the effectiveness of the proposed sample selection method. Our approach is adaptable to real-world non-line-of-sight propagation scenarios and achieves an MAE positioning error of 0.261 meters, even under dynamic environmental conditions.

Developed in the context of the SIRIOS project

We developed a hardware setup that captures short, wideband snapshots in both E1 and E6 GNSS bands. This setup is mounted to a bridge over a motorway. The setup records 20 ms raw IQ snapshots triggered from the energy with a sample rate of 62.5 MHz, an analog bandwidth of 50MHz and an 8 bit bit-width. At certain frequencies the GPS/Galileo or GLONASS signals can easily be seen as a slight increase in the spectrum. Note that experts manually analyzed the datastreams by thresholding CN/0 and AGC values. Manual labeling of these snapshots has resulted in 11 classes: classes 0 to 2 represent samples with no interferences, distinguished by variations in background intensity, while classes 3 to 10 contain different interferences. The challenge lies in adapting to positive class labels with only a limited number of samples available. We partition the dataset into a 64% training set, 16% validation set, and a 20% test set split (balanced over the classes).

Interference signals degrade and disrupt Global Navigation Satellite System (GNSS) receivers, impacting their localization accuracy. Therefore, they need to be detected, classified, and located to ensure GNSS operation. State-of-the-art techniques employ supervised deep learning to detect and classify potential interference signals. Here, literature proposes ResNet18 and TS-Transformer as they provide the most accurate classification rates on quasi-realistic GNSS signals. However, employing these methods individually, they only focus on either spatial or temporal information and discard information during optimization, thereby degrading classification accuracy. This paper proposes a deep learning framework that considers both the spatial and temporal relationships between samples when fusing ResNet18 and TS-Transformers with a joint loss function to compensate for the weaknesses of both methods considered individually. Our real-world experiments show that our novel fusion pipeline with an adapted late fusion technique and uncertainty measure significantly outperforms the state-of-the-art classifiers by 6.7% on average, even in complicated realistic scenarios with multipath propagation and environmental dynamics. This works even well (F-β=2 score about 80.1%), when we fuse both modalities only from a single bandwidth-limited low-cost sensor, instead of a fine-grained high-resolution sensor and coarse-grained low-resolution low-cost sensor. By using late fusion the classification accuracy of the classes FreqHopper, Modulated, and Noise increases while lowering the uncertainty of Multitone, Noise, and Pulsed. The improved classification capabilities allow for more reliable results even in challenging scenarios.