Advancing Fault Detection in HVAC Systems: Unifying Gramian Angular Field and 2D Deep Convolutional Neural Networks...

Area of Science:

• Building Science and Engineering
• Artificial Intelligence in HVAC
• Data-Driven Fault Diagnosis

Background:

• HVAC systems are crucial for building efficiency and comfort, but faults degrade performance.
• Existing data-driven fault detection methods struggle with complex HVAC dynamics during occupancy.
• Real-time fault detection during active operation is highly valuable for capturing dynamic interactions.

Purpose of the Study:

• To develop and evaluate an advanced HVAC fault detection framework for real-time operational scenarios.
• To leverage simulated HVAC data and novel deep learning techniques for improved fault identification.
• To enhance the robustness and reliability of HVAC fault detection systems.

Main Methods:

• Developed an HVAC fault model using HVACSIM+ dynamic simulation with 194 sensor signals.
• Utilized Gramian Angular Fields (GAF) to transform time-series sensor data into 2D images.
• Employed 2D Convolutional Neural Networks (2DCNNs) for automated feature extraction and fault classification.

Main Results:

• The GAF-2DCNN framework achieved an overall accuracy of 97% for HVAC fault detection.
• Individual fault detection demonstrated precision, recall, and F1 scores exceeding 90%.
• Outperformed traditional methods like SVM, RF, and 1D-CNNs in accuracy and reliability.

Conclusions:

• The integrated approach using HVACSIM+ data and GAF-2DCNN offers a robust solution for HVAC fault detection.
• This method effectively captures hidden temporal relationships in sensor data for accurate fault diagnosis.
• The framework provides a significant enhancement in the reliability of detecting substantial HVAC faults during operation.