Adaptive Compressive Sensing and Data Recovery for Periodical Monitoring Wireless Sensor Networks

Jian Chen, Jie Jia, Yansha Deng, Xingwei Wang, Abdol-Hamid Aghvami2018 · DOI

TL;DR

This paper introduces an adaptive compressive sensing (CS) framework for wireless sensor networks (WSNs) that dynamically adjusts sampling rates to match changing signal sparsity. This approach significantly improves data recovery accuracy and network lifetime compared to traditional methods by efficiently handling dynamic signal characteristics.

What Problem Does This Paper Address?

Wireless Sensor Networks (WSNs) are crucial for monitoring various environments, but their limited battery power necessitates energy-efficient data gathering. Traditional methods often involve transmitting raw data, which is energy-intensive. Compressive Sensing (CS) offers a promising solution by allowing signals to be reconstructed from fewer samples than traditional Nyquist-Shannon sampling, provided the signal is sparse. However, existing CS techniques typically assume a static or slowly changing signal sparsity. This assumption is problematic for many real-world monitoring applications (e.g., anti-fire forest monitoring) where signal sparsity can change rapidly due to environmental conditions. Ignoring these dynamic changes leads to degraded recovery performance and inefficient energy usage. This paper addresses this gap by proposing an adaptive CS approach that can effectively handle signals with unknown and dynamically varying sparsity.

Methods

The core of this work is an adaptive compressive sensing framework. The key innovations lie in two main components:

  • Adaptive Sampling Scheme: Instead of using a fixed sampling rate, the system dynamically adjusts the sampling rate at the beginning of each monitoring period. This is achieved by:
  • Re-sampling: A small number of additional measurements are taken.
  • Sparsity Determination: The current sparsity of the signal is estimated by comparing the recovered signal from the initial measurements with the signal recovered from the re-sampled data. A difference exceeding a threshold indicates a change in sparsity.
  • Sampling Rate Adjustment: A new sampling rate is calculated based on the estimated sparsity, ensuring sufficient samples for accurate recovery while minimizing overhead. This process is designed to be energy-efficient, requiring only a few re-sampling iterations.
  • Adaptive Signal Recovery Algorithm: To handle signals with unknown sparsity, the paper proposes an improved Sparsity Adaptive Matching Pursuit (SAMP) algorithm. This algorithm incorporates:
  • Adaptive Step Size: The step size used in the greedy pursuit algorithm is dynamically adjusted. Two approaches are proposed: linear and non-linear decrement weight factors. These adaptive step sizes aim to balance convergence speed and recovery accuracy, preventing overestimation and improving the likelihood of finding the true signal support.

The data gathering process itself follows a compressive sampling approach where sensor nodes transmit compressed measurements, and the sink node performs the reconstruction. The adaptive sampling and recovery are primarily handled at the sink node, with only the determined sampling rate broadcast back to the sensor nodes to save energy.

Datasets

The paper uses real-world datasets for evaluation:

  • Ocean Temperature Data: Collected by the National Oceanic and Atmospheric Administration (NOAA) at a specific location (5 N, 95 W) on February 23, 2012. This data was used to demonstrate sparsity analysis using Discrete Wavelet Transform (DWT).
  • RSSI Measurement: Data from a smartphone's Received Signal Strength Indicator (RSSI) measurement from a real environment. This was used to demonstrate sparsity analysis using the K-SVD algorithm with an overcomplete dictionary, as DWT was not suitable for this data.

The authors also mention simulations involving a 200-node wireless sensor network with a fixed sink node in a 500x500 area, with initial energy set to 10,800 J.

Key Insights and Ideas

  • Dynamic Sparsity is a Real Problem: The paper strongly emphasizes that assuming static sparsity in WSN monitoring is a significant limitation for many practical applications. The rapid changes in environmental conditions directly impact signal sparsity.
  • Adaptive Sampling is Crucial for Efficiency: By dynamically adjusting the sampling rate based on estimated sparsity, the system can achieve accurate recovery without the overhead of over-sampling or the performance degradation of under-sampling. This is key to balancing energy consumption and data quality.
  • SAMP Can Be Improved: The standard SAMP algorithm, while useful for unknown sparsity, can be slow to converge or may not always find the optimal solution. The proposed adaptive step size variation in their SAMP variant offers a way to improve both convergence speed and recovery accuracy.
  • Sink-Centric Adaptation Saves Node Energy: By performing most of the adaptive calculations (sparsity estimation, sampling rate determination) at the sink node and only broadcasting the resulting sampling rate, the energy burden on individual sensor nodes is minimized.

Key Results

  • Improved Recovery Performance: The proposed adaptive sampling scheme consistently achieved better signal recovery compared to fixed sampling methods, especially when signal sparsity changed over time. Figures 10 and 11 show that the adaptive method accurately tracks sparsity and maintains a lower Mean Squared Error (MSE) for recovery.
  • Network Lifetime: The adaptive CS approach demonstrated a significantly longer network lifetime compared to traditional raw data gathering. While comparable to fixed sampling or intelligent sampling in some scenarios (Figure 13), it offers better recovery performance.
  • Energy Efficiency: Figure 12 shows that the adaptive algorithm achieved more balanced energy consumption among sensor nodes compared to traditional methods.
  • Faster Convergence: The proposed adaptive SAMP algorithm showed faster convergence times (around 0.438s) compared to traditional SAMP with a fixed step size (0.542s for step=1, 0.1s for step=3, but the latter might be less robust). Table 1 and Figures 14 and 15 illustrate competitive relative error and SNR performance.

Limitations and Future Work

  • Computational Overhead at Sink: While sensor node energy is conserved, the adaptive sampling and recovery process at the sink node will incur higher computational load. The paper doesn't deeply explore the scalability of the sink's processing capabilities.
  • Threshold Selection: The paper mentions a predefined threshold ($T_{th}$) for detecting sparsity changes. The optimal selection of this threshold might be application-dependent and could require further investigation.
  • Environmental Factors: While simulations considered environmental factors, a more extensive study on how different types of environmental dynamics (e.g., sudden vs. gradual changes) affect performance could be beneficial.
  • Future Work: The authors suggest further research into optimizing the adaptive step size determination and exploring more advanced dictionary learning techniques for even better sparsity representation. They also hint at the potential for extending this framework to more complex WSN scenarios.

Summary updated

BibTeX
@misc{chen2018adaptive,
  author = {Jian Chen and Jie Jia and Yansha Deng and Xingwei Wang and Abdol-Hamid Aghvami},
  title = {{Adaptive Compressive Sensing and Data Recovery for Periodical Monitoring Wireless Sensor Networks}},
  year = {2018},
  doi = {10.3390/s18103369},
}
Figure