Unsupervised domain adaptation for HVAC fault diagnosis using contrastive adaptation network

Naghmeh Ghalamsiah

doi:10.17918/00011117

Faults in heating, ventilation and air-conditioning (HVAC) systems can lead to excessive energy waste, poor indoor environment, equipment failures, and higher maintenance costs. This highlights the critical need for developing fault detection and diagnosis (FDD) tools, especially data-driven FDD strategies, which could be more cost-effective and accurate compared with rule-based FDD tools. However, the effectiveness of data-driven FDD approaches largely relies on the availability and quality of training data that accurately represent the operational characteristics of systems under both fault-free and faulted conditions. Due to the high cost and practical challenges of obtaining training data (especially under faulted conditions) from real systems, data from simulation models and/or similar systems are often used instead. This might lead to significant distribution discrepancies between training data and real system data (i.e., test dataset) due to variations in outdoor conditions, control strategies, internal loads, and occupant behavior. Transfer learning (TL) offers a potential solution to this problem by leveraging rich labeled data from one domain, known as the source domain, to diagnose faults in a shifted domain, called the target domain. Since collecting even limited labeled data in real systems is labor-intensive and costly, this study focuses on a branch of TL, known as unsupervised domain adaptation (UDA). UDA aims to transfer knowledge from a labeled source domain to a fully unlabeled target domain with differing data distribution. While convolutional neural network (CNN)-based approaches remain most popular within the reported UDA methods for HVAC FDD, there exists a few research gaps: firstly, their performance is limited due to the simplistic architecture or structure defined for the CNN part. Additionally, the critical step of properly preparing HVAC data to best be utilized by CNN-based models for feature extraction and fault diagnosis is often overlooked. This research addresses the above-mentioned research gaps through the following approach: it utilizes the contrastive adaptation network (CAN) algorithm, originally successful in image classification, to overcome the specific challenges faced by current domain adaptation algorithms in HVAC systems. Furthermore, temporal causal discovery framework (TCDF), a causality-based framework for discovering causal relationships in time series data, is implemented in the data processing step to meet the requirements of convolutional networks, where spatially closer features are more likely to be correlated. The results on air handling unit (AHU) datasets demonstrate that the CAN algorithm effectively facilitates domain adaptation in the absence of target labels and that the feature reordering process reduces the training time and the number of loops required for convergence.

Unsupervised domain adaptation for HVAC fault diagnosis using contrastive adaptation network

Files and links (1)

Abstract

Metrics

Details

Unsupervised domain adaptation for HVAC fault diagnosis using contrastive adaptation network

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media