Session: 02-04: Building Performance Analysis and Simulation

Paper Number: 137242

137242 - Interpreting Convolutional Neural Network Model Developed for Building Fault Detection and Diagnosis Using Layer-Wise Relevance Propagation 

Abstract: 

Deep Learning (DL) models have gained much popularity in fault detection and diagnosis (FDD) of complex systems with high dimensional data in recent years. This is mainly due to their ability to extract hidden characteristics of data, without any need for prior feature selection, and achieve high diagnosis accuracies. However, despite their promising results, they are categorized as black-box models, meaning it is extremely difficult to interpret their decision-making mechanisms and results. This lack of interpretability greatly affects the reliability of these models for stakeholders and their applications in real-world scenarios.

To deal with this issue, multiple visualization techniques, as a branch of explainable Artificial Intelligence (XAI), have been developed to explain how these models make predictions, ensuring the accountable and transparent use of DL. While these techniques are well-established in image recognition and Natural Language Processing, their application has not been investigated much in the field of Heating, Ventilation, and Air-Conditioning (HVAC) systems’ FDD. Most of the published studies in this field have trusted the high accuracies of their developed DL models without investigating them further. More specifically, no current paper has investigated the application of visualization techniques for Air Handling Units (AHU) FDD.

To address this gap, this study first develops a Convolutional Neural Network (CNN) model, a type of deep neural network with high feature extraction ability, that can detect and diagnose the three faults, namely cooling coil valve stuck fully open fault, outdoor air damper stuck fully closed fault, and supply duct leakage at a rate of 20% fault, implemented to an AHU system controlled by ASHRAE-Guideline 36. Using 5-week simulated data of a medium-sized office building, the developed CNN model, with optimized hyperparameters using the grid search method, demonstrates more than 95% accuracy in isolating the fault, indicating the effectiveness of the model for its intended task.

Next, a modified version of Layer-wise Relevance Propagation (LRP) visualization technique is applied to the developed model. LRP backpropagates the output classification through the layers of the network to assign relevance scores to each of the input features. With this approach, the influential features in the decision-making of the CNN model are found, which are considered as diagnosis criteria. The results of this study show that the developed model can find the true features that are justifiable by domain experts, making the model highly reliable in its task. Lastly, using this visualization method, the reason behind misdiagnosis cases can also be interpreted for further investigation and improvement.

Presenting Author: Naghmeh Ghalamsiah Drexel University

Presenting Author Biography: Naghmeh Ghalamsiah is a PhD student in Architectural Engineering at Drexel University. She earned her Bachelor of Science in Architectural Engineering from the University of Tehran, Iran. Her research is focused on Automated Fault Detection and Diagnosis, Deep Learning applications, and the development of Energy-efficient Buildings.

Authors:

Naghmeh Ghalamsiah Drexel University
Jin Wen Drexel University