Continuous Monitoring and Automated Fault Detection and Diagnosis of Air-Handling Units

Introduction

Air-Handling Units (AHUs) are highly customized equipment. The regulations concerning AHUs are increasingly becoming strict to meet higher energy efficiency and ventilation goals, which adds to the complexity inherent with customised equipment. This upsurge in complexity increases the need for continuous maintenance and monitoring of AHUs. However, such programs are difficult to implement due to the shortage of skilled personnel. Therefore, continuous monitoring and Fault Detection and Diagnosis (FDD) processes need to be automated, referred as AFDD. Despite the plethora of research on AFDD, there are limited real-life applications. Adding to this, the available solutions are either unreliable, unaffordable, and/or not scalable. Surveying through the literature on FDD tools that are commercially deployed or under development revealed that these tools rely on a combination of expert rules or first principles. Rules-based approaches are heavily reliant on sensed information and expert knowledge. This makes the maintenance of such tools unsustainable. Further, such tools carry very limited ability to prevent significant energy wastage. It is estimated that up to 30% of energy could be saved through the effective use of data collected with continuous monitoring systems. To realise this, highly sensitive (ability to diagnose condition) diagnosis models can be trained with Artificial Intelligence (AI) based approaches that are scalable and have lesser reliance on expert knowledge and sensors. Through this project, an AFDD tool that incorporates these approaches has been developed. It is generally recommended to treat the fault detection and diagnosis processes separately as in this way, it is easier to overcome challenges associated with implementation. For example, supervised machine learning algorithms require annotated fault labels. In the designed tool, these practical limitations have been overcome and completely automated fault detection and diagnosis have been implemented. For fault detection, a widely used machine learning algorithm called XGBoost is deployed, whilst for fault diagnosis, Bayesian network-based probabilistic models are deployed. Typically, these models are inhibitive due to the complexity associated with their development. For adding interpretability to these models, frameworks, such as (i) Shapley Additive Explanations (SHAP) for fault detection and (ii) 4-Symptoms and 3-Faults (4S3F) for fault diagnosis are included. The 4S3F framework is a generalizable framework that supports the development of diagnostic Bayesian networks using piping and instrumentation (P&I) diagrams prepared during HVAC design. Using these P&I schemes, the developed Bayesian networks have been discretised for cooling and heating mode operations of the AHU. This way the adopted diagnosis approach translates HVAC domain knowledge and remains in sync with building practitioner’s approach toward fault diagnosis. To understand the reliability and generalizability of the proposed FDD strategy, 5 case studies have been utilized with a diverse operational environment, weather conditions, and AHU configurations. Additionally, the developed diagnosis models are validated under multiple fault scenarios experimentally induced in two building environments. The specificity of the deployed diagnosis models exceeds 90% with samples collected through long-term test procedures exceeding 60 days. A fault condition diagnosed with a highly specific model imputes a very high likelihood of fault presence. Further, the developed diagnosis models not only distinguish between fault presence and absence conditions but can also accurately isolate root causes. The diagnosed outcomes are supported by visual evidence delivered through a web-based interactive user interface. It’s been noted that visualizations in the available continuous monitoring tools are not sufficient to aid visual diagnostic procedures. Therefore, a lot more time is consumed in preparing data than using it for preventing energy wastage. In the developed AFDD tool, this aspect has been taken care of through visualisations that promise to support building practitioners and take a step forward toward human-in-the-loop diagnostics. Further, a fault library is also realized using a lightweight database structure to manage taxonomy, prior fault probabilities and recommendations for fault correction. The developed scheme is useful for building practitioners to manage information on faults easily and consistently. Upon deployment, it is estimated that nearly 33% chiller of energy waste can be prevented using the AFDD tool. A financial plan to roll out the tool commercially for non-residential buildings in the Netherlands carrying large HVAC installations has been developed. It’s projected that a SaaS business providing such a tool is financially viable. Under all forecasted revenue scenarios profitability can be realised in a window of 5 years from inception.