Sequential recalibration of wireless sensor networks with (stochastic) gradient descent and mobile references

Abstract

Sensors can suffer from aging drift over their lifetime, thus showing changes in sensitivity and baseline, which then leads to unreliable measurement results and affects digital trust. By performing Monte Carlo simulations, this work evaluated the opportunity to use gradient descent, in combination with reliable mobile nodes as references, for the recalibration of wireless sensor networks in the context of smart cities. In the implemented simulations, sensor aging was treated as a deterministic process while encounters between sensors and mobile references occurred in a stochastic manner. Furthermore, by exchange of data during several encounters between sensors and mobile references, a recalibration data set was constructed. In addition, encounters were weighted according to their recency – more important when being more recent. The results presented here are promising, showing that even few recent encounters might be sufficient for recalibration in order to provide once again trustworthy measurement results.

Keywords

• Concept drift
• Digital trust
• Gradient descent
• Machine learning
• Smart city
• Wireless sensor networks