Real-Time Fine-Grained Air Quality Sensing Networks in Smart City: Design, Implementation and Optimization

Zhiwen Hu, Zixuan Bai, Kaigui Bian, Tao Wang, Lingyang Song

TL;DR

This paper presents a novel system for real-time, fine-grained air quality monitoring in smart cities. It tackles the challenge of limited sensor resources by developing optimized strategies for sensor power control (using deep Q-learning) and sensor placement (using a genetic algorithm) to minimize data inference errors and create accurate air quality maps.

What Problem Does This Paper Address?

Air pollution is a major global health concern, and accurate, real-time monitoring is crucial for public health and urban planning. Traditional official monitoring systems are expensive, sparse, and can have significant latency. While low-cost sensors offer a solution for denser coverage, they are often power-limited and their data needs to be intelligently processed to create a comprehensive air quality map. The core challenge is to effectively utilize a limited number of power-constrained sensors to infer air quality across an entire area, minimizing both measurement and inference errors, especially in complex urban environments. This involves deciding when sensors should collect data (power control) and where they should be deployed (location selection) to maximize the accuracy of the resulting air quality map.

Methods

The authors propose a four-layer architecture for their air quality sensing system: sensing, transmission, processing, and presentation. The key innovations lie in the optimization strategies for power control and location selection:

  • Power Control: This problem is framed as a Markov Decision Process (MDP). For a single sensor, an optimal dynamic programming approach is used. For multiple sensors, a deep Q-learning (DQN) algorithm is employed to approximate the value function, enabling efficient decision-making for when each sensor should wake up and collect data to conserve power while maintaining data quality.
  • Location Selection: This is addressed using a genetic algorithm. The process involves:
  • Feature Space Creation: Locations are analyzed based on their statistical correlation (using parameters derived from collected data) to create a high-dimensional feature space.
  • Clustering: K-means clustering is used to group locations based on their feature space representation, aiming to select sensors from diverse locations.
  • Genetic Algorithm: A genetic algorithm then explores various combinations of sensor locations, using genetic coding, recombination, and mutation to iteratively improve the selection and minimize the overall joint error.

Datasets

The authors utilized a dataset collected from 30 deployed sensing devices at Peking University (PKU) over approximately seven months. This dataset primarily contains PM2.5 concentration values. The data was used to extract statistical properties for modeling measurement and inference errors, and for training and evaluating the proposed algorithms.

Key Insights and Ideas

  • Interdependence of Optimization: The paper highlights that power control and location selection are interdependent problems. Optimizing one can significantly impact the effectiveness of the other.
  • Value of Diverse Data: Selecting sensor locations that are spatially diverse (less correlated) is crucial for improving the accuracy of inferred air quality data.
  • Deep Q-Learning for Power Management: DQN proves effective in learning complex, dynamic power control policies for multiple sensors, adapting to changing environmental conditions and energy constraints.
  • Genetic Algorithm for Optimal Placement: A genetic algorithm provides a robust method for exploring the vast search space of sensor locations, leading to a near-optimal deployment strategy.
  • Gaussian Model for Inference Error: The authors demonstrate that a Gaussian model can effectively approximate measurement and inference errors, providing a tractable way to formulate the optimization problem.

Key Results

  • Performance Gain: The proposed strategies (deep Q-learning for power control and genetic algorithm for location selection) showed significant performance gains in simulations, leading to a reduction in the average joint error of the air quality map.
  • Single-Device Power Control: The optimal power control for a single device achieved an approximate 6% performance gain compared to a uniform sensing strategy.
  • Multi-Device Power Control: The DQN approach for multi-device power control demonstrated a much more promising performance than linear approximation or random control strategies.
  • Location Selection Effectiveness: The genetic algorithm with initial clustering significantly improved the starting point and convergence speed of the location selection process.
  • Combined Strategy: The combination of the optimized power control and location selection strategies yielded the best overall performance, outperforming random approaches for both.

Limitations and Future Work

  • Model Simplification: The paper acknowledges that the Gaussian model for inference error, while useful for tractability, might not perfectly capture all real-world complexities, especially in scenarios with low air quality values where quantization accuracy can be an issue.
  • Computational Complexity: While the proposed methods are more efficient than brute-force approaches, the computational complexity for multi-device power control and location selection can still be significant, especially with a large number of sensors and locations.
  • Environment Assumption: The power control strategy relies on the assumption that the statistics of air quality transitions are known. Adapting to completely unknown or rapidly changing environmental statistics could be a future research direction.
  • Scalability: While the authors state their solutions are scalable, further investigation into their performance with extremely large-scale smart city deployments would be beneficial.

Summary updated

Abstract

Driven by the increasingly serious air pollution problem, the monitoring of air quality has gained much attention in both theoretical studies and practical implementations. In this paper, we present the architecture, implementation and optimization of our own air quality sensing system, which provides real-time and fine-grained air quality map of the monitored area. As the major component, the optimization problem of our system is studied in detail. Our objective is to minimize the average joint error of the established real-time air quality map, which involves data inference for the unmeasured data values. A deep Q-learning solution has been proposed for the power control problem to reasonably plan the sensing tasks of the power-limited sensing devices online. A genetic algorithm has been designed for the location selection problem to efficiently find the suitable locations to deploy limited number of sensing devices. The performance of the proposed solutions are evaluated by simulations, showing a significant performance gain when adopting both strategies.
BibTeX
@article{hu2018real,
  author = {Zhiwen Hu and Zixuan Bai and Kaigui Bian and Tao Wang and Lingyang Song},
  title = {{Real-Time Fine-Grained Air Quality Sensing Networks in Smart City: Design, Implementation and Optimization}},
  journal = {IEEE Internet of Things Journal},
  year = {2018},
  url = {https://openreview.net/forum?id=hCVXM1fe8y},
}
Figure