Causal Machine Learning Model for Hvac System-Based Preventive Maintenance: A Causal Discovery for Pre-Failure Detection and Explanation

Abstract

Preventive maintenance for the Heating, Ventilation, and Air-Conditioning systems (HVAC systems) aims to monitor and detect early-stage failures, benefiting thermal comfort satisfaction and energy consumption. Still, it is a complex and uncertain factor, and technicians must understand the cause of the problem to interpret the pre-failure events. This study addresses this concern with a causal machine learning model, encoding HVAC systems behavior, indoor ambiance, and outdoor environment based on random variables and modeling their causal representation concerning pre-failure based on the Structural Causal Model (SCM). It employs d-separation and d-connection to justify causal-and-effect relationships with the model and applies the expectation maximization (EM) algorithm to fit model parameters given observational data. The causal model is verified using a do-operator to rationalize the sound explanations (d-separated and d-connected) and employing Odds Ratio and Confidential Interval to prove their statistical strength. The results showed that the predicted outcomes and their explanations can encode human-like interpretation and achieve high causal significance aligned with real-world pre-failure events. It is suitable for causal-based decision-making systems for HVAC system preventive maintenance.

Keywords: causal inference, causal artificial intelligence, Internet of Things, Building Engineering, Thermal comfort, Energy Conservation