Distributed signal decorrelation in wireless sensor networks using the sparse matrix transform

Leonardo R. Bachega, Srikanth Hariharan, Charles A. Bouman, Ness B. ShroffProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2011 · DOI

TL;DR

This paper introduces the "vector SMT," a novel distributed transform for wireless sensor networks (WSNs) that decorrelates vector measurements from multiple sensors. By enabling joint processing of these correlated measurements, it significantly improves anomaly detection accuracy compared to processing them independently.

What Problem Does This Paper Address?

Wireless sensor networks (WSNs) often involve sensors collecting data that, when aggregated, forms a high-dimensional signal. In many real-world scenarios, such as a network of cameras monitoring the same environment, the measurements from individual sensors are correlated. This correlation can hinder effective signal processing, particularly for tasks like anomaly detection, where combining information from all sensors is crucial. Traditional methods often struggle with distributed processing and handling these correlations efficiently. This paper addresses the challenge of decorrelating these high-dimensional, correlated vector measurements in a distributed manner within a WSN to improve anomaly detection performance.

Methods

The core contribution is the vector Sparse Matrix Transform (vector SMT), an extension of the previously developed Sparse Matrix Transform (SMT). The SMT is a decorrelating transform that approximates a full-rank covariance estimate using a sequence of sparse Givens rotations. The vector SMT adapts this by applying it to pairs of vector measurements from different sensors.

Here's a breakdown of the approach:

  • Vector Measurements: Instead of scalar readings, each sensor produces a vector measurement (e.g., principal components of an image).
  • Distributed Decorrelation: The vector SMT decorrelates these vector measurements sequentially, pairing up sensor outputs and applying a decorrelating transform. This process continues until all sensor vectors are decorrelated.
  • Greedy Optimization: Similar to the original SMT, the vector SMT uses a greedy optimization approach to find the optimal sequence of rotations that minimizes a cost function related to the log-likelihood of the data.
  • Anomaly Detection Framework: The decorrelated signals are then used in a statistical anomaly detection test. The paper proposes using the volume of an ellipsoid defined by the detection statistic as a metric for evaluating detector performance.

Datasets

The paper does not explicitly mention using pre-existing datasets. Instead, it describes simulation experiments using synthetic data:

  • Moving Sphere: A 3D sphere moves along a trajectory, captured by two cameras. Typical and anomalous configurations are generated.
  • 3D Cloud of Spheres: A cloud of 30 spheres is generated, with "typical" clouds being hollow and "anomalous" clouds being dense. This scenario uses 14 camera views.

These simulations are used to train detectors (Ntrain = 100 samples) and test them (Ntest = 200 samples, with 100 typical and 100 anomalous).

Key Insights and Ideas

  • Joint Processing is Crucial for Correlation: The most significant insight is that ignoring correlations between sensor measurements (independent processing) severely degrades anomaly detection performance. Joint processing, enabled by the vector SMT, is essential for distinguishing subtle anomalies.
  • Vector SMT Captures Inter-Sensor Dependencies: The vector SMT's ability to decorrelate pairs of vector measurements effectively captures the dependencies between different sensors' viewpoints. This is visualized by the paired eigen-images in the joint processing case.
  • Distributed Nature is Key for WSNs: The method is designed for WSNs, implying that the decorrelation can be performed in a distributed fashion, reducing the need for a powerful central node to collect all raw data.
  • Visualizing Sensitivity: The use of generalized eigenvalues and eigen-images provides a powerful way to understand which dimensions (or combinations of sensor views) are most sensitive to anomalies and how the processing method affects this sensitivity.

Key Results

The paper demonstrates significant improvements in anomaly detection accuracy through simulations:

  • Moving Sphere Experiment: The vector SMT-based joint processing method achieved much higher accuracy, as evidenced by both ROC curves and log-volume plots, compared to independent processing. Independent processing was largely unable to distinguish typical from anomalous samples.
  • 3D Cloud of Spheres Experiment: Similar to the moving sphere, joint processing with the vector SMT showed substantially better anomaly detection results than independent processing, highlighting its effectiveness for more complex multi-view scenarios.
  • Qualitative Findings: Eigen-images from the joint processing case show pairs of features corresponding to different camera views, indicating that the transform learns to correlate information across sensors. In contrast, independent processing yields eigen-images that are specific to individual sensors.

Limitations and Future Work

  • Limitations:
    • The paper relies on simulations with synthetic data. Real-world application with actual sensor data is not presented.
  • The computational cost and communication overhead of the distributed vector SMT are not deeply analyzed, though the authors mention this as a future direction.
  • Future Work:
    • Analyze the communication costs associated with the vector SMT in a distributed setting.
  • Conduct experiments with real application data to validate the method's performance in practical scenarios.

Summary updated

Abstract

In this paper, we propose the vector SMT, a new decorrelating transform suitable for performing distributed anomaly detection in wireless sensor networks (WSN). Here, we assume that each sensor in the network performs vector measurements, instead of a scalar ones. The proposed transform decorrelates a sequence of pairs of vector sensor measurements, until the vectors from all sensors are completely decorrelated. We perform simulations with a network of cameras, where each camera records an image of the monitored environment from its particular viewpoint. Results show that the proposed transform effectively decorrelates image measurements from the multiple cameras in the network. Because it enables joint processing of the multiple images, our method provides significant improvements to anomaly detection accuracy when compared to the baseline case when we process the images independently.
BibTeX
@inproceedings{bachega2011distributed,
  author = {Leonardo R. Bachega and Srikanth Hariharan and Charles A. Bouman and Ness B. Shroff},
  title = {{Distributed signal decorrelation in wireless sensor networks using the sparse matrix transform}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2011},
  doi = {10.1117/12.887549},
}

Loading related content...

Figure