Abstract
Training robust classifiers for sensor networks is often hindered by the scarcity of high-quality labels and the failure of models to generalize across different sensor locations. This paper introduces a framework that combines contrastive learning with an "information graph" to incorporate system context—such as temporal overlaps between sensors—into the training process. By treating labels as high-dimensional random vector "anchors," the model learns sensor-invariant features. Evaluation on a seismic dataset demonstrates that this method maintains high accuracy (85%) even when labels are restricted to a single station, significantly outperforming standard supervised baselines (50%).
Introduction
In real-world sensor networks, such as seismic arrays for natural hazard monitoring, obtaining comprehensive labels is difficult and expensive. Furthermore, models trained on specific sensors often fail to generalize to new locations due to overfitting on local environmental artifacts (e.g., subsurface properties). This paper addresses these challenges by using the physical context of the sensor network itself to compensate for missing labels and improve cross-sensor robustness.
Methodology
The proposed method integrates system context into training through two primary components: an information graph and a contrastive learning framework.
The Information Graph and Contrastive Learning
Instead of treating data points in isolation, the authors construct an information graph where nodes represent signal segments. Edges are established based on domain knowledge—specifically, if two segments overlap in time but originate from different stations or channels. This captures the physical reality that a single event, such as an earthquake, is recorded by multiple sensors simultaneously with varying signatures.
The model, utilizing a ResNet18 encoder, is trained via contrastive learning to pull connected nodes (different views of the same event) closer in a latent space while pushing disconnected nodes apart.
Annotation Anchors
To integrate sparse labels, the authors represent class labels as fixed, high-dimensional, L2-normalized Gaussian random vectors. These act as "anchor nodes" in the graph. Data segments with known labels are connected to these anchors, pulling them toward specific regions of the latent space and allowing supervised and self-supervised learning to occur within a unified mathematical framework.
Experimental Evaluation
The method was evaluated using a seismic dataset from Illgraben, Switzerland, consisting of 8 seismometers (3 channels each). The dataset includes three event types: earthquakes, slope failures, and noise. The study focused on two scenarios: a fully annotated setting and a "weakly labeled" setting where training labels were restricted to only one of the eight stations to test cross-sensor generalization.
Results and Discussion
The proposed ResNet18+IG+anchors method demonstrated superior performance across all test cases. In the fully annotated setting, it achieved 93.9% accuracy, outperforming both Random Forest (86.4%) and standard ResNet18 (91.3%).
The most significant breakthrough occurred in the one-station training scenario. While the standard ResNet18 baseline failed to generalize (dropping to 50.0% accuracy), the proposed method maintained a robust 85.0% accuracy. By forcing the model to map different sensor readings of the same event to the same latent area, the framework inherently learns to ignore sensor-specific artifacts, achieving true sensor invariance.
Conclusion
This research demonstrates that physical system context is a powerful substitute for explicit labels. By leveraging the inherent constraints of sensor networks through an information graph, models can achieve high robustness and generalization even with minimal human annotation. Future work may explore automated graph construction and application to other multi-sensor modalities like IoT or audio arrays.
Summary updated