Fault detection diagnostic for HVAC systems via deep learning algorithms

• Published in: Energy and buildings Vol. 250; p. 111275
• Main Authors: Taheri, Saman, Ahmadi, Amirhossein, Mohammadi-Ivatloo, Behnam, Asadi, Somayeh
• Format: Journal Article
• Language: English
• Published: Lausanne Elsevier B.V 01.11.2021; Elsevier BV
• Subjects: Accuracy; Air conditioning; Algorithms; Computer applications; Computing time; Configurations; Deep learning; deep recurrent neural network; Embedding; Empirical analysis; Fault detection; Fault detection diagnosis; HVAC; HVAC equipment; hyperparameter optimization; Learning algorithms; Machine learning; Neural networks; Optimization; Recurrent neural networks; Regularization; Training; deep recurrent neural network; machine learning; hyperparameter optimization; Fault detection diagnosis
• ISSN: 0378-7788; 1872-6178
• DOI: 10.1016/j.enbuild.2021.111275

Because of high detection accuracy, deep learning algorithms have recently become the focus of increased attention for fault detection diagnostic (FDD) analysis of heat, ventilation, and air conditioning (HVAC) systems. Among all the machine learning algorithms in the field, deep recurrent neural networks (DRNNs) are being widely used since they are capable of learning the complex, uncertain, and temporal-dependent nature of the faults. However, embedding DRNN in FDD applications is still subject to two challenges: (I) a bespoke DRNN configuration, out of conceivably infinite DRNN architectures, is not explored on the task of FDD for HVAC systems; (II) Hyperparameter optimization, which is a computationally expensive task due to its empirical nature, is not investigated. In this respect, seven DRNNs configurations are introudecd and tuned that can automatically detect faults of different degrees under the faulty and normal conditions. Then, a comprehensive study of hyperparameters is conducted to optimize and compare all the proposed configurations based on their accuracy and training computational time. By searching through different hidden layers and layer sizes, optimization methods, model regularization, and batching, the ultimate DRNN model is selected out of more than 200 experiments. All the training configuration files are publicly available. Also, a comparison is made between the proposed DRNN model and two other advanced data-driven techniques, namely, random forest (RF) and gradient boosting (GB). The final DRNN model outperforms RF and GB regression by a significant margin.