Smart Buildings: State-Of-The-Art Methods and Data-Driven Applications

Cheng Fan, Fu Xiao, Huilong WangDigital 2024 · DOI · OpenAlex

Introduction

The "Data Revolution" is fundamentally altering the landscape of building management. Traditionally, buildings have been managed through manual oversight and rigid, rule-based systems that struggle to adapt to changing conditions. However, the rise of the Internet of Things (IoT) has flooded Building Automation Systems (BAS) with high-volume data. This paper reviews how advanced machine learning—specifically transfer, semi-supervised, and self-supervised learning—is being leveraged to transform these structures into intelligent, automated systems capable of sophisticated fault detection, occupancy sensing, and energy-efficient control.

The Data Landscape in Smart Buildings

While the volume of data available from modern buildings is immense, it is often "dirty," characterized by missing values and outliers. Furthermore, a chronic "label shortage" exists; while sensors record temperatures and flow rates, they rarely record the ground-truth labels (e.g., "this timestamp indicates a fan failure") required for standard supervised learning.

Typical multi-source data structure in Building Automation Systems (BAS).

As shown in the data landscape above, BAS data typically spans energy consumption, physical operating parameters, and environmental conditions. Extracting actionable intelligence from this multi-source, unlabeled data is the primary challenge facing the industry today.

A Framework for Data Preprocessing

To address the noise and complexity of raw BAS data, the authors propose a rigorous five-step preprocessing framework. This pipeline ensures that data is of sufficient quality for advanced analytical models.

The five-step data preprocessing framework for building data.
  1. Data Cleaning: Enhancing quality through missing value imputation (e.g., KNN, mean) and outlier detection (e.g., GESD, k-means).
  2. Data Reduction: Reducing dimensions via sampling or feature extraction to manage computational load.
  3. Data Scaling: Normalizing data into similar ranges using techniques like max-min normalization or z-score standardization.
  4. Data Transformation: Ensuring compatibility with algorithms through numerical or categorical encoding (e.g., one-hot encoding).
  5. Data Partitioning: Dividing data into subsets for training and validation using supervised or unsupervised methods.

Advanced Learning Paradigms for Fault Detection

To overcome the labeling bottleneck, researchers are moving toward advanced learning paradigms:

  • Transfer Learning: This involves migrating knowledge from "information-rich" buildings (those with labeled data) to "information-poor" ones. A creative application involves using t-SNE to convert tabular sensor data into 2D images, allowing researchers to utilize powerful pre-trained Computer Vision models (CNNs) for fault diagnosis.
  • Semi-supervised Learning: This approach uses "self-training," where a model initially trained on a small set of labels generates "pseudo-labels" for the remaining unlabeled data.
  • Self-supervised Learning: This paradigm uses "pretext tasks," such as Denoising Autoencoders (DAEs), to learn the inherent structure of building data without requiring any manual labels.

Human Behavior Identification and Optimal Control

Beyond mechanical systems, smart buildings must understand their occupants. By utilizing environmental sensors (CO2, light, noise) and wearable Inertial Measurement Units (IMUs), systems can infer occupancy and thermal comfort. For instance, identifying human gestures (like rubbing hands) via IMU data has reached an accuracy of 86.2%, providing a non-invasive way to gauge occupant needs.

For system control, the paper compares Model Predictive Control (MPC) and Reinforcement Learning (RL). While MPC uses physics-based models to plan ahead, it often faces complex optimization hurdles. RL offers high adaptability through trial and error but carries risks during its initial learning phase in real-world environments.

Key Results and Future Directions

The application of these methods has yielded significant improvements:

  • Fault Detection (FDD): Transfer learning boosted diagnosis accuracy by 8.18% to 12.63% in data-scarce scenarios, while semi-supervised methods saw increases of over 20%.
  • Occupancy Detection: Using Extreme Gradient Boosting (XGB) on multi-source data achieved 97.09% accuracy.

Despite these gains, limitations remain regarding sensor robustness and the scalability of "plug-and-play" models. Future research must focus on making AI models resilient to sensor failure and developing "Safe RL" methods that eliminate the risks associated with initial trial-and-error learning.

Summary updated

BibTeX
@inproceedings{fan2024smart,
  author = {Cheng Fan and Fu Xiao and Huilong Wang},
  title = {{Smart Buildings: State-Of-The-Art Methods and Data-Driven Applications}},
  booktitle = {Digital},
  year = {2024},
  doi = {10.1007/978-3-031-48161-1\_3},
}
Figure