Summary
Diagnosing problems in heating and cooling systems can be tricky because different methods need different kinds of information or human help. The authors study a new type of AI model called a decision model that can suggest possible faults and how likely they are without needing specific training data. They tested this model on real HVAC fault cases to see which problems it can detect, what information it requires, and how well it works when conditions like seasons or buildings change. They found that the model works well when it has detailed physical data but struggles with faults that need broader context, and although its confidence scores help, they still need adjusting.
What this means in practice
- •For building maintenance teams: Rank HVAC fault candidates using decision model probabilities to assist operators in real-time diagnosis across different buildings and seasons.
- •For industrial equipment monitoring teams: Integrate decision models with physical feature data to identify simple faults in machinery where operating context is less critical.
Abstract
Artificial intelligence supports building operations in several forms, each with its own barrier. Expert rules must be tuned for every system, supervised models need labeled data that buildings rarely record, and language models return free text that requires human-in-the-loop checking, since their stated confidence is unreliable. A newer kind of pretrained model, here called a decision model, returns a probability for every allowed answer, so one model could serve many decisions without training. This study answers three open questions for fault diagnosis in heating, ventilation, and air-conditioning systems: which decisions such a model can make, what input it needs, and whether its probabilities hold when conditions change. On 128 fault days from four public datasets of real equipment, faults are graded by the reasoning their diagnosis demands, with data given raw, as physical features, or with Brick topology. The decision model Jev, open language models, and a supervised model face nine tests that change season, control configuration, or building. Given physical features, Jev and the larger open model diagnosed faults whose evidence one feature carries, but not faults that need operating context. Under shift they kept their accuracy and calibration, while the supervised model lost 0.33 macro-F1 yet led or tied within a building. Their probabilities still needed correction, and detection was weak. The study maps which faults a decision model can diagnose and from what input, and supports a division of work in which code computes the physics and the model ranks candidate faults for an operator.