Location based services have gained significant popularity among end users and are now inevitably becoming part of new wireless infrastructures and emerging business processes. Deep learning artificial intelligence methods have proved as highly effective for wireless fingerprinting localization based on extensive indoor radio measurement data.
Nevertheless, as these methods grow in complexity, their computational requirements and consequently energy consumption become significantly burdensome, both during the training phase and in subsequent operations. This concern is further exacerbated when considering the number of mobile users is estimated to exceed 7.48 billion by the end of 2025. Assuming that the networks serving these users will need to conduct only one localization per user per hour on average, the machine learning models employed for these calculations would be compelled to execute 65 * 10^12 predictions per year. Furthermore, when factoring in the tens of billions of other interconnected devices and applications that heavily rely on more frequent location updates, it becomes evident that localization will make a substantial contribution to carbon emissions unless more energy-efficient models are developed and adopted.
This motivated our work on a new deep learning architecture for indoor localization that achieves superior energy efficiency compared to existing state-of-the-art methods, with minimal performance degradation. To comprehensively evaluate the performance of our proposed model, we conducted a detailed analysis, revealing that it generates a only 58% of the carbon footprint while maintaining 98.7% of the overall performance compared to state of the art models external to our group. Additionally, we elaborate on a methodology for assessing the complexity of the DL model, enabling the estimation of its CO2 footprint during both training and operational phases.
|
