Optimization of Sensor Configurations for Fault Identification in Smart Buildings

Naveed Ahmad, Malcolm Egan, Jean-Marie Gorce, Jilles Steeve Dibangoye, Frédéric Le MouëlICASSP 2023 · DOI

TL;DR

This paper tackles the challenge of optimizing sensor configurations in smart buildings for effective fault detection while minimizing energy consumption. They introduce a novel codesign framework that jointly optimizes sensor settings (like sampling rate and compression) and the fault detection classifier, demonstrating significant tradeoffs between accuracy and power usage.

What Problem Does This Paper Address?

Predictive maintenance is crucial for smart buildings, enabling early detection of system degradation (e.g., in HVAC). This relies on data from numerous sensors. However, many sensors are battery-powered and have limited lifespans, necessitating energy-efficient operation. Traditional approaches often optimize either sensor configurations or the fault detection algorithm in isolation. This paper addresses the gap by proposing a method to jointly optimize both, recognizing that the effectiveness of a fault detection classifier is intrinsically linked to the data it receives, which is determined by the sensor configurations. This integrated approach aims to achieve reliable fault detection without excessively draining sensor batteries.

Methods

The core of this paper is a codesign bi-level optimization framework. This framework formalizes the problem of finding the optimal sensor configuration (defined by activity, sampling rate, and data compression) and the corresponding fault detection classifier.

  1. Master Classifier: A K-KNNG classifier is trained on "normal operation" data from all sensors at their highest fidelity. This classifier is used to label a separate test dataset, which may include "abnormal" events.
  2. Induced Classifiers: For each potential sensor configuration (a combination of sensor activity, sampling rate, and compression), an "induced classifier" is trained. These induced classifiers are then evaluated using the test data labeled by the master classifier.
  3. Codesign Optimization: The problem is formulated as a bi-level optimization problem. The outer level optimizes the sensor configurations (x, n, Tc) to minimize a weighted sum of total sensor power consumption and classifier inaccuracy. The inner level involves training and evaluating the induced classifier for each configuration.
  4. Algorithm: To handle the discrete and computationally intensive nature of the problem, they propose a stochastic optimization approach (Algorithm 1). This algorithm iteratively samples sensor configurations and updates parameters to find an efficient tradeoff between power and accuracy.

Datasets

The paper uses the Twin House Experiment dataset. This dataset comprises temperature measurements and heating power data from 8 sensors across various rooms in two standalone houses over 41 days. The data includes periods of "normal operation" (heating on) and "abnormal operation" (heating off). The authors divide this dataset into three zones (D1, D2, D3) to study the impact of different training data durations on the optimization process.

Key Insights and Ideas

  • The Power of Codesign: The most significant insight is that jointly optimizing sensor configurations and classifiers leads to better tradeoffs than optimizing them separately. This acknowledges the interdependence between data acquisition and data analysis.
  • Efficient Tradeoffs: The proposed algorithm can rapidly find sensor configurations that achieve high fault detection accuracy while significantly reducing battery power consumption.
  • Impact of Training Data: The choice of training data (specifically, its duration) can influence the resulting sensor configurations, potentially leading to higher battery power consumption if the training data covers a longer period with more active sensors.
  • Discrete Optimization Approximation: Approximating the discrete optimization problem with a smooth version and then using stochastic gradient descent is a practical approach to tackle the complexity.

Key Results

The numerical results from the Twin House Experiment demonstrate the effectiveness of the proposed algorithm.

  • The algorithm rapidly improves both accuracy and battery power over iterations.
  • The algorithm successfully identifies sensor configurations that offer a good balance between accuracy and power. For instance, they show that the battery power consumed by optimal configurations can be significantly lower than that of a full-sampling, zero-compression setup (e.g., around 0.3-0.4 µW compared to 0.7 µW).
  • While accuracy levels are generally high across different training data zones (D1, D2, D3), the battery power required can vary, with D2 (a longer training period) potentially leading to higher power consumption due to more active sensors.

Limitations and Future Work

  • Computational Complexity: While the stochastic approach is efficient, the bi-level optimization problem remains computationally intensive, especially as the number of sensors and possible configurations increases.
  • Generalizability: The study is based on a specific use case (heating system in the Twin House Experiment). Further validation on diverse smart building systems and environments would be beneficial.
  • Future Work: The authors suggest exploring more sophisticated models for sensor power consumption and investigating the impact of different classifier types and optimization algorithms. They also hint at the potential for dynamic adaptation of sensor configurations in real-time based on changing building conditions.

Summary updated

BibTeX
@inproceedings{ahmad2023optimization,
  author = {Naveed Ahmad and Malcolm Egan and Jean-Marie Gorce and Jilles Steeve Dibangoye and Frédéric Le Mouël},
  title = {{Optimization of Sensor Configurations for Fault Identification in Smart Buildings}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10097223},
}
Figure