Using Machine Learning Methods to Provision Virtual Sensors in Sensor-Cloud

Mingzheng Zhang, Liang-Min Wang, Shuming XiongSensors 2020 · DOI

TL;DR

This paper proposes a novel machine learning-based approach for provisioning virtual sensors in sensor-cloud environments, specifically for agricultural IoT applications. By classifying sensors based on their historical data trends and then clustering similar sensors, the method efficiently selects representative physical sensors to form virtual sensors, leading to reduced energy consumption and extended network lifetime while maintaining data accuracy.

What Problem Does This Paper Address?

The proliferation of sensor-cloud technology offers immense potential for applications like precision agriculture by aggregating data from numerous wireless sensor networks (WSNs) and leveraging cloud computing for storage and processing. However, a key challenge is "virtual sensor provisioning" – efficiently selecting and activating physical sensors to fulfill user requests for specific data. Traditional methods often activate too many sensors, leading to excessive energy consumption and reduced network lifespan, or rely solely on spatial correlation, which can be insufficient given the dynamic nature of environmental data. This paper addresses the need for a more intelligent and energy-efficient virtual sensor provisioning strategy that considers the temporal trends of sensor data, not just its spatial proximity.

Methods

The proposed scheme employs a two-step machine learning approach:

  1. Classification based on Changing Trends: Physical sensors are first classified into groups based on their historical data trends. This is achieved using a linear regression model. For each sensor's data, the model estimates the trend (increasing or decreasing) for each measured parameter. A flag is generated for each parameter indicating its trend, and these flags form a trend signature for the sensor. Sensors with similar trend signatures are grouped together. This step is crucial because sensors with similar current values might have diverging future trends, necessitating their separation.
  2. Clustering within Classes: Within each classified group, the k-means clustering algorithm is applied. The goal is to group sensors with high measurement similarity. The k-means algorithm iteratively assigns sensors to clusters and recalculates cluster centroids to minimize the sum of squared errors (SSE) between sensors and their assigned centroids.
  3. Representative Sensor Selection: From each final cluster, a single representative physical sensor is selected. These selected sensors then form the virtual sensor that serves the user's request. The remaining sensors in the cluster are put into a dormant state to conserve energy.

Datasets

The paper uses the Intel Lab Data dataset, which contains environmental data collected by 54 physical sensors every 31 seconds. For their experiments, they selected 50 sensors with sufficient data and focused on two parameters: temperature and humidity. No new benchmarks were introduced; instead, the proposed scheme is compared against existing methods like LEACH and a data aggregation scheme from reference [12].

Key Insights and Ideas

  • Temporal Trends Matter: The most significant insight is the recognition that simply clustering sensors based on current spatial proximity or measurement similarity is insufficient. Understanding the changing trends of sensor data over time is critical for accurate and efficient virtual sensor provisioning, especially in dynamic environments like agriculture.
  • Two-Stage ML Approach: Combining linear regression for trend classification with k-means for similarity clustering provides a robust mechanism to group sensors that are not only similar in value but also behave similarly over time.
  • Energy Efficiency through Dormancy: By selecting only a few representative sensors per cluster and putting the rest into a dormant state, the scheme drastically reduces energy consumption, a critical factor for WSN longevity.
  • Virtualization as a Service: The paper reinforces the concept of virtual sensors as a service (Se-aaS), abstracting the complexity of physical sensor management from the end-user.

Key Results

  • Energy Efficiency: The proposed scheme demonstrated significant energy savings. In one comparison, it achieved a 75.59% decrease in energy consumption compared to the data aggregation scheme by activating only 11 physical sensors out of 50.
  • Network Lifetime: The energy savings directly translate to a longer network lifetime. The paper shows that their scheme has the longest network lifetime compared to LEACH and the data aggregation scheme, lasting significantly more rounds.
  • Data Accuracy: Despite using fewer sensors, the scheme maintained or improved data accuracy. For temperature and humidity, the proposed scheme showed an increase in data accuracy of 55.11% and 50.78%, respectively, compared to the data aggregation scheme. The paper quantifies this using metrics like $\delta_{10}$, $\delta_{14}$, $\delta_{19}$, and overall $\Delta_{os}$.

Limitations and Future Work

  • Heterogeneous Environments: The authors acknowledge that virtual sensor provisioning under heterogeneous environments (e.g., different sensor types, communication protocols) is an area for future research.
  • Representative Sensor Selection: While the paper uses k-means to select representative sensors, further refinement of the selection algorithm for representative sensors could be explored.
  • Scalability: The paper focuses on a static network topology. Investigating the scheme's performance in dynamic or mobile network scenarios could be a future direction.
  • Dataset Limitations: The Intel Lab Data dataset, while used, had issues with missing values and varying record lengths, requiring pre-processing. Future work might involve testing on more comprehensive and cleaner datasets.

Summary updated

BibTeX
@article{zhang2020using,
  author = {Mingzheng Zhang and Liang-Min Wang and Shuming Xiong},
  title = {{Using Machine Learning Methods to Provision Virtual Sensors in Sensor-Cloud}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20071836},
}
Figure