Background & Summary

Air Handling Units (AHUs) within Heating, Ventilation, and Air Conditioning (HVAC) systems primarily regulate and circulate conditioned air to maintain optimal indoor environmental conditions, including temperature, humidity, ventilation, and air quality1,2,3. Automated Fault Detection and Diagnosis (AFDD) for AHUs involves continuous and automated monitoring of multiple system components, including sensors, actuators, and controllers, to rapidly detect any abnormal or faulty conditions.

In recent decades, numerous data-driven AFDD methods have been extensively studied, with supervised machine learning approaches receiving particular attention due to their high accuracy and reliability4,5,6,7. However, these supervised methods heavily depend on extensive and accurately labeled datasets for training and validation. Examples of such datasets include the 1043-RP dataset developed for chiller systems8,9 and the 1312-RP dataset focusing on laboratory-based AHU systems10. Additionally, researchers have commonly created simulation datasets using widely recognized tools such as HVACSIM + and Modelica11,12,13,14.

Although simulation and laboratory datasets offer ease of accessibility and controlled conditions, real operational data obtained from actual building environments are crucial. Real-world operational data authentically capture complex system behaviors, environmental variations, and unforeseen events that simulations or laboratory experiments might not fully replicate15. Despite advancements, there remain significant gaps in the availability of comprehensive real operational datasets, particularly in specialized building contexts. To date, only limited research specifically addresses real operational data for office buildings, with a predominant focus on developing operational guidelines rather than extensively providing real-world datasets16.

To bridge this gap, the primary objective of this research is to develop extensive, real operational datasets from three distinct building types—a large-scale office, an auditorium, and a hospital. These buildings were specifically chosen because they represent typical urban environments, enhancing the transferability and generalizability of the developed AFDD methods to similar facilities. Researchers can utilize these datasets to explore, validate, and publish novel methods, while developers can leverage them to refine, implement, and enhance machine learning and diagnostic models specifically designed for AFDD in AHU systems.

Methods

The real operational dataset collected from various building types utilizes AHUs equipped with Constant Air Volume (CAV) in HVAC systems. While a consistent methodological framework guided the overall process of data collection and preparation, specific procedures were adapted to accommodate the unique operational characteristics and configurations of each building. These steps included: (1) customized raw data collection tailored to the specific requirements of each building type; (2) removal of missing or duplicated data points within each dataset; (3) systematic labelling of operational conditions and fault types at the target class level; and (4) dataset splitting into training, validation, and test subsets designed according to the distinct data characteristics of each building.

Raw data collection

Raw data collection was conducted across three distinct building types: an auditorium, a hospital, and an office. Each building type utilized multiple AHUs, varying in number, specifications, and sensor configurations specifically tailored for AFDD across different rooms or spaces. Initially, each sensor recorded data at one-minute intervals. For practical, hourly-scale analysis, these high-frequency measurements were aggregated into hourly intervals. This aggregation significantly reduced storage requirements and simplified data management challenges typically associated with high-frequency datasets in real-world applications.

However, aggregating data into hourly intervals introduces certain considerations for AFDD method development. Operational statuses and valve positions, for example, might fluctuate or modulate multiple times within an hour, and such detailed variations may not be fully captured when data are aggregated. Although this does not diminish the overall usefulness of the dataset, AFDD developers should carefully consider these implications when applying or interpreting hourly aggregated data.

All sensors utilized for AFDD were calibrated annually according to standard procedures specified by their respective manufacturers. Calibration procedures, detailed below for each sensor type, ensured consistent data accuracy and reliability. The sensors employed for FDD are detailed in Table 1, including their descriptions and measurement units with sensor accuracy. Although sensor types were consistent across buildings, the total number of sensors varied slightly depending on each building’s unique characteristics. More detailed dataset descriptions for each building type are provided below.

  • Auditorium

    The data collection was conducted at the Sejong Arts Center (21 Gungnipbangmulgwan-ro, Sejong-si), covering an area of 16,186 m² across floors B1 to 6 F. The dataset was collected over a period of one year, from January 1 to December 31, 2022. The facility was equipped with 13 AHUs, each designed with cooling and heating capabilities tailored to specific spaces, as detailed in Table 2. The AHU system incorporated 15 distinct sensor variables, with sensor placement illustrated schematically in Fig. 1. Data from these AHUs were recorded hourly and continuously throughout the entire year, resulting in a comprehensive dataset comprising 113,880 hourly records.

    Fig. 1
    figure 1

    Schematic diagrams of sensor placement for AHUs in the auditorium.

    Full size image
    Table 1 Description of used sensors with their units.
    Full size table
    Table 2 Specifications of installed AHUs within the auditorium facility.
    Full size table

  • Hospital

    The data collection was conducted at the New Wing of the National Cancer Center Hospital (323 Ilsan-ro, Ilsandong-gu, Goyang-si, Gyeonggi-do), which covers an area of 18,900 m2 extending from basement level 2 to the 5th floor. The building was completed in October 2020, and data collection took place continuously over one year, from January 1 to December 31, 2022. The facility included eight AHUs, each equipped with cooling and heating coils, as detailed in Table 3. A total of nine sensor variables were employed, with their locations illustrated schematically in Fig. 2. Hourly data collection throughout the entire year resulted in a comprehensive dataset comprising 70,080 hourly records.

    Table 3 Specifications of installed AHUs within the hospital facility.
    Full size table
    Fig. 2
    figure 2

    Schematic diagrams of sensor placement for AHUs in the hospital.

    Full size image

  • Large-scale office

    The data collection was performed at a large-scale office building located at 48 Gwacheon-daero 7na-gil, Gwacheon-si, Gyeonggi-do, covering an area of 50,966 m² from basement level 4 to the 6th floor. The dataset was collected continuously from December 1, 2023, to November 30, 2024. The facility employed 20 AHUs, each equipped with cooling and heating coils, as detailed in Table 4. A total of 18 sensor variables were utilized, with their schematic locations illustrated in Fig. 3. The data collection yielded 175,532 hourly entries, which slightly exceeds the theoretical total of 175,200 hourly records (calculated as 20 AHUs × 365 days × 24 hours). This difference is likely attributed to additional operational logging or monitoring events captured during data collection.

    Table 4 Specifications of installed AHUs within the office facility.
    Full size table
    Fig. 3
    figure 3

    Schematic diagrams of sensor placement for AHUs in the office.

    Full size image

Table 5 summarizes the collected raw datasets for each building type, including the total area, data collection period, number of AHUs, number of sensors, and total data points recorded. Although each dataset has a clearly defined total number of data points, it may still contain some missing sensor variables. Additionally, as mentioned earlier, the dataset from the office building includes 175,532 hourly records, slightly exceeding the theoretically expected 175,200 entries (calculated as 20 AHUs × 24 hours/day × 365 days). This discrepancy indicates that additional operational logging or data recording activities might have occurred. In the following chapter, the collected datasets will be thoroughly reviewed, checked for completeness, and preprocessed to ensure accuracy and consistency for further analysis.

Table 5 Summary of AHU data collection across different building types.
Full size table

Removal of missing, and duplicated data

Initially, data from multiple AHUs across three buildings were collected hourly for one year (365 days). The theoretical data points were calculated based on the total number of AHUs, days in a year, and hourly intervals. However, discrepancies were found when comparing actual collected data with theoretical expectations, as shown in Table 6.

Table 6 Comparison between theoretical and actual data points and identified data quality issues.
Full size table

The auditorium and hospital datasets presented fewer actual data points than theoretically expected, indicating missing values in some independent variables. Specifically, the Auditorium dataset had 505 missing entries, while the Hospital dataset had 4,032 missing entries. These missing data points likely resulted from sensor malfunctions, temporary data logging interruptions, or communication errors during data collection17.

Conversely, the office dataset contained 332 more entries than theoretically expected. Upon closer examination, these excess entries were identified as duplicated data points caused by logging or synchronization errors. After removing these duplicates, the office dataset matched the theoretical data expectation.

Annotation of target class

Seven operational conditions were initially defined for annotation, comprising one normal operation and six fault types: supply fan fault, cooling pump fault, heating pump fault, return air temperature sensor fault, supply air temperature sensor fault, and valve position fault.

While universal guidelines are available to distinguish normal from faulty operational conditions, specific numerical ranges are not consistently defined for all fault types. According to Wang et al.16, certain faults like the return air temperature sensor fault and the supply air temperature sensor fault have widely accepted numerical criteria. In contrast, faults such as valve position issues generally rely on broader guidelines without explicit numerical ranges.

In practice, numerical fault criteria must be tailored to specific AHU installations due to variations in system configurations, manufacturer specifications, and operational contexts. An important observation made during annotation is that typically, the return air temperature is higher than the supply air temperature. However, specific operational conditions can cause this relationship to reverse. For instance, during periods of low occupancy or inactivity—particularly in office and auditorium buildings—interior spaces can naturally cool, resulting in the return air temperature falling below the supply air temperature. Additionally, when cooler outdoor air is introduced through free cooling (economizer mode) without mechanical cooling, the return air temperature may also become relatively lower.

In this study, six HVAC experts, each with over 20 years of engineering experience, applied ASHRAE-based guidelines to empirically establish appropriate numerical ranges suitable for typical hospital AHU systems. Table 7 details the established criteria and numerical ranges for each condition.

Table 7 Criteria and numerical ranges for annotating operational and fault conditions.
Full size table

Annotations were manually entered into an Excel program by four annotators, each possessing over 20 years of HVAC FDD expertise and backgrounds in mechanical engineering. To enhance reliability and predictive accuracy, an independent team of two experts cross-checked a randomly selected subset of the annotated data.

The operational conditions and overall fault distributions for each building type are visualized in Fig. 4. Each hourly data point was treated individually; therefore, a continuous fault lasting multiple hours was counted as multiple occurrences (one per hour). Similarly, continuous normal operation was also counted as multiple occurrences, one for each hour. The analysis shows that hospitals exhibit greater operational reliability compared to auditoriums and offices. Auditoriums have a relatively balanced distribution between normal and faulty conditions, whereas office buildings display a significant number of faults, indicating a critical area for targeted improvement.

Fig. 4
figure 4

Distribution of operational conditions (normal and fault condition) by building type.

Full size image

Table 8 provides the distribution of each fault category across datasets, highlighting differences arising from the varied occurrence of specific fault types. Notably, despite being part of the annotation guidelines, certain faults were absent in specific buildings:

  • The cooling pump fault and heating pump fault were not identified in the hospital and office buildings, indicating stable pump operation aligned with system demands throughout the monitoring period.

  • The cooling supply temperature fault was rarely identified, occurring only once within the hospital building. Conversely, this fault was entirely absent in both the auditorium and office buildings.

  • These observations reflect variations in operational reliability and highlight differences in fault profiles across different building types and system configurations.

Table 8 Distribution of annotated fault types across building types.
Full size table

Data split

After completing the annotation of the entire dataset, the data were randomly partitioned into three subsets: training (60%), validation (20%), and test (20%). In machine learning, the training set is typically used exclusively for model training, the validation set is employed for hyperparameter tuning and selecting the best-performing models, and the test set provides an unbiased evaluation of the final model’s performance on unseen data18. However, traditional random splitting can lead to smaller classes becoming disproportionately underrepresented19,20. To address this issue, stratified sampling was employed to maintain balanced representation across all subsets.

Notably, certain faults either did not occur during the data collection period or occurred very infrequently, presenting challenges for robust machine learning model training and evaluation. Consequently, these minor classes were excluded from the data-splitting process. Specifically, seven categories—one normal condition and six fault conditions—were defined for fault detection and diagnosis as follows:

  • Normal condition: “Normal”

  • Return Air Temperature Sensor Fault: “RATSF”

  • Supply Air Temperature Sensor Fault: “SATSF”

  • Supply Fan Fault: “SFF”

  • Valve Position Fault: “VPF”

  • Cooling Pump Fault: “CPF”

  • Heating Pump Fault: “HPF”

Stratified sampling preserves the proportional representation of each class within each dataset subset, ensuring that classes, particularly those with fewer instances, are neither overrepresented nor underrepresented. This approach reduces the risk of bias, supports stable model training, and enhances the reliability and validity of subsequent evaluations. Detailed distributions of each dataset across training, validation, and test subsets are provided in Table 9.

Table 9 Distribution of training, validation, and test datasets by building type and fault condition.
Full size table

Data Record

Each sensor continuously monitors and records its corresponding variable, generating a detailed record of each AHU’s operational status along with labeling for relevant faults. These sensors are integrated into the Building Management System (BMS), enabling continuous monitoring and control of AHU operations. Initially, data from each sensor variable was individually stored in csv file format, directly exported from the BMS, which manages and monitors HVAC systems across the building. The integrated datasets for each building type have been compiled and stored in the Figshare repository21. Details of the stored files, including the number of AHUs, sensors, operating conditions, and filenames, are provided in Table 10.

Table 10 Stored AHU dataset information by building type.
Full size table

Technical Validation

Because imputation methods rely heavily on assumptions about how data becomes missing—and violations of these assumptions can introduce significant biases—this study excluded data entries if more than one independent variable was missing. Specifically, to improve the robustness and accuracy of machine learning models, missing data were intentionally not imputed, as artificial imputation can potentially distort underlying data patterns and negatively impact predictive performance22,23.

To further enhance data quality and reliability, each AHU condition across all building types was meticulously labeled according to clearly defined guidelines and numerical ranges outlined in Table 7. Annotation was conducted manually by four annotators, each possessing at least 20 years of specialized experience in HVAC FDD and comprehensive mechanical engineering expertise. Their extensive experience significantly reduced potential labeling errors, improved annotation accuracy, and ensured consistency throughout the process.

An illustrative example demonstrating the labeling approach using auditorium data for each condition is provided in Table 11. In this table, it is important to note that certain rows labeled as “normal condition” reflect special but valid operational scenarios. For instance, the first row, which shows low Supply Air Temperature (17.12 °C) combined with a higher cooling supply temperature (~26 °C) and inactive cooling pumps, indicates the system operating in economizer mode (free cooling), directly utilizing cooler outdoor air. The partially open valve position (48.08%) is part of the AHU’s airflow modulation strategy to efficiently maintain indoor conditions without mechanical cooling. The second row represents transient or transitional operational states, where the higher supply air temperature (26.77 °C) compared to the set point (22 °C), along with the fully open valve position (100%) but inactive heating pumps, reflects specific AHU control logic during temporary operational adjustments or transitions between operating modes.

Table 11 Example of annotated sensor data for seven operational conditions.
Full size table

Annotators directly recorded corresponding fault labels into an Excel spreadsheet. Additionally, to ensure high-quality annotations and reliable predictive outcomes, an independent verification team comprising two additional experts cross-checked randomly selected annotated samples. This rigorous cross-validation step was instrumental in identifying and correcting annotation inconsistencies, further strengthening dataset quality.