Abstract
This work develops the underpinnings of self-supervised placement-awarerepresentation learning given spatially-distributed (multi-view and multimodal)sensor observations, motivated by the need to represent external environmentalstate in multi-sensor IoT systems in a manner that correctly distills spatialphenomena from the distributed multi-vantage observations. The objective ofsensing in IoT systems is, in general, to collectively represent an externallyobserved environment given multiple vantage points from which sensoryobservations occur. Pretraining of models that help interpret sensor data musttherefore encode the relation between signals observed by sensors and theobservers' vantage points in order to attain a representation that encodes theobserved spatial phenomena in a manner informed by the specific placement ofthe measuring instruments, while allowing arbitrary placement. The worksignificantly advances self-supervised model pretraining from IoT signalsbeyond current solutions that often overlook the distinctive spatial nature ofIoT data. Our framework explicitly learns the dependencies between measurementsand geometric observer layouts and structural characteristics, guided by a coredesign principle: the duality between signals and observer positions. Wefurther provide theoretical analyses from the perspectives of informationtheory and occlusion-invariant representation learning to offer insight intothe rationale behind our design. Experiments on three real-worlddatasets--covering vehicle monitoring, human activity recognition, andearthquake localization--demonstrate the superior generalizability androbustness of our method across diverse modalities, sensor placements,application-level inference tasks, and spatial scales.