GCC-GAN: Graph-Context-Conditioned GAN for Heterogeneous Sensor Dynamic Compensation

Abstract

To address the low efficiency of traditional one-to-one modeling for dynamic compensation in complex transient testing scenarios involving large numbers of heterogeneous sensors with different dynamic error mechanisms, this article proposes a heterogeneous sensor dynamic compensation method based on a graph-context-conditioned generative adversarial network (GCC-GAN) to explore the consensus response characteristics across heterogeneous sensors. First, according to the dynamic characteristics of transient signals, each signal is divided into multiple view segments, including prestep, poststep, and full-signal segments, so as to construct a unified representation learning mechanism that jointly captures local dynamic features and global evolutionary trends. Then, a K-nearest neighbor (KNN) graph is constructed in the feature space based on node-feature similarity. Through cross-sample neighborhood aggregation, the consensus response characteristics of the sample set are extracted to form a graph-level context vector that represents the commonality of the sample set, thereby reducing the interference of individual-sample noise, local distortion, and operating-condition differences on feature learning. Finally, the graph-level context vector is embedded into the generative adversarial compensation network as a conditional prior, guiding the generator to achieve local detail restoration and dynamic error suppression while preserving the overall signal morphology. The proposed method is validated on multiple types of sensor data, including multirange piezoresistive pressure sensors, strain-gauge six-axis force/torque sensors, armored thermocouples, and inertial measurement units (IMUs). Experimental results demonstrate that GCC-GAN enables unified compensation for heterogeneous sensors and achieves higher compensation accuracy.