Graph Regularized AutoFuse: Robust Sensor Fusion With Noisy Labels

Abstract

Manufacturing defects, wear, and operational conditions pose a huge risk for single-sensor-based sensing systems. The evolution of sensor technology and computing has led to the emergence of multisensor fusion systems, offering robust and improved performance. However, the effectiveness of the existing multisensor fusion methods is heavily reliant on the availability of labeled data. This challenge intensifies when known labels are corrupted by noise, which is quite common in practical scenarios. To address these issues, this letter introduces the graph regularized autoencoder-based multisensor fusion framework (GrAutoFuse). GrAutoFuse utilizes autoencoders to learn representations from individual sensors and combines them for robust classification within a semi-supervised learning framework. Unlike other semi-supervised methods, this approach can identify noisy labels, perform label estimation and correction through label propagation on a graph that captures correlations between different sensors. Here, we present a joint optimization formulation for learning sensor-specific representations, fused representations, and a classifier by estimating missing and correcting noisy labels. This results in a robust fusion model for classification. Experimental results on two datasets from different domains illustrate the generalizability and superior performance of GrAutoFuse compared to state-of-the-art methods, showcasing its effectiveness in handling missing and noisy labels.