An ensemble-based faults detection and diagnosis approach for determining faults severities at whole-building level

Figures & data

Table 1. Summary of previous studies on ensemble approaches for detecting and diagnosing building system faults.

Reference	Approach	System	Faults	Severity
(Han et al. 2020)	Majority voting [consists of KNN, SVM, RF]	Centrifugal chiller system	Reduced evaporator water flow Reduced condenser water flow Refrigerant overcharge Refrigerant leakage Condenser fouling Non-condensable gas Excess oil	Single
(Yao, Li, and Gao 2022)	XGBoost, RF, LightGBM	Chiller	Reduced condenser water flow Reduced evaporator water flow Refrigerant leakage Refrigerant overcharge Excess oil Condenser fouling Non-condensable gas	Multiple
(Chakraborty and Elzarka 2019)	XGBoost	Water-cooled centrifugal chiller Packaged A/C unit VAV	Coefficient of performance reduction VAV motor efficiency and fan total efficiency reduction	Single
(Ebrahimifakhar, Kabirikopaei, and Yuill 2020)	Bagging, AdaBoost, XGBoost	Packaged A/C unit	Refrigerant undercharge Refrigerant overcharge Compressor valve leakage Liquid-line restriction Condenser airflow reduction Evaporator airflow reduction Non-condensable gas	Single
(McHugh, Isakson, and Nagy 2019)	Bagged Trees, Gradient Boosted Trees, AdaBoost Trees	Air handling unit (AHU)	Steam and chilled water valve leakage	Multiple

Figure 1. Stacking classifier framework with a k-fold cross-validation schematic.

Figure 2. Layout of the applied methodology.

Figure 3. An illustration of a 3D model of the considered building.

Table 2. Building construction and materials information.

Construction	Materials	U-value (W/m^2.K)
Exterior wall	200 mm concrete block (CMU), 30 mm cement/plaster/mortar on both sides with painted finish	1.344
Interior partition	200 mm concrete block (CMU), 30 mm cement/plaster/mortar on both sides with painted finish	1.344
Window	Doubled glazed with 3 mm thickness generic LoE clear glass (both sides) and 6 mm air gap Aluminium frame extruded profile 3.2 mm	2.470
Roof	50 mm cement/plaster/mortar, 300 mm concrete slab, 50 mm sand and gravel, 20 mm clay tile	0.469

Table 3. Building's window-to-wall ratio with respect to building orientation.

	Total	North (315° to 45°)	East (45° to 135°)	South (135° to 225°)	West (225° to 315°)
Gross Wall Area (m^2)	1105.8	306.99	245.95	307.05	245.81
Window Opening Area (m^2)	71.32	13.46	5.21	17.2	35.45
Gross Window-Wall Ratio (%)	6.45	4.38	2.12	5.6	14.42

Figure 4. Packaged terminal air conditioner (PTAC).

Table 4. System information and EnergyPlus basic input.

Item	Details
Zones	Men's hall, Library, Elderly women's room, Women's Floor, toilets, ablution spaces, library, funeral room, Muazzin's and an Imam's rooms.
Cooling system capacity (KW) (Spilt unit: floor-standing/ wall-mounted)	13.2 /4.53
Cooling set point (C°)	23
Cooling size factor	1.15
Solar distribution	Full exterior with reflections
Indoor surface heat transfer convection algorithm	TARP: variable natural convection based on the temperature difference
Outside surface heat transfer convection algorithm	DOE-2: correlation from measurements by Klems and Yazdanian for rough surfaces
Heat balance algorithm	CTF (Conduction Transfer Functions)
Timestep for the simulation	6
Ground reflectance value	0.2
Internal gains	From people, lights and other equipment such as water coolers and fridge
Zones infiltration (design flowrate calculation)	Flow/zone(modified by a schedule fraction, temperature difference and wind speed)

Figure 5. Simulated and measured monthly energy consumption.

Figure 6. Average energy consumption per day.

Figure 7. The simulated faults classes with different severities.

Table 5. Building’s simulated faults.

Systems	Faults	Introduced severity	Severity	Ranking
HVAC	Thermostat offset (TO)	TO1: ΔT =  + 1 °C	ΔT ≤ 2°C	Low
		TO2: ΔT =  + 2 °C
		TO3: ΔT =  + 3 °C	2°C < ΔT ≤ 3°C	Medium
		TO4: ΔT =  + 4 °C	ΔT > 3 °C	High
		TO5: ΔT =  + 5 °C
	Refrigerant undercharge (RU)	RU1: Leakage = 10%	Leakage ≤ 10%	Low
		RU2: Leakage = 20%	10% < Leakage ≤ 20%	Medium
		RU3: Leakage = 30%	Leakage > 20%	High
	Dirty filter (DF)	DF1: ΔP =  + 50%	ΔP ≤ 60%	Low
		DF2: ΔP =  + 60%
		DF3: ΔP =  + 70%	60% < ΔP ≤ 70%	Medium
		DF4: ΔP =  + 80%	ΔP > 70%	High
		DF5: ΔP =  + 90%
	Condenser fouling (CF)	CF1: Blockage = 50%	Blockage ≤ 60%	Low
		CF2: Blockage = 60%
		CF3: Blockage = 70%	60% < Blockage ≤ 70%	Medium
		CF4: Blockage = 80%	Blockage > 70%	High
		CF5: Blockage = 90%
Exhaust Fan	Exhaust Fan-Motor-Worn-Out (EXF)	EXF1: Fan Efficiency = −4%	−4% ≥ Efficiency drop ≤ −8%	Low
		EXF2: Fan Efficiency = −8%
		EXF3: Fan Efficiency = −12%	−8% < Efficiency drop ≤ −16%	Medium
		EXF4: Fan Efficiency = −16%	Efficiency drop > −16%	High
		EXF5: Fan Efficiency = −20%
Lighting Control	Occupancy sensor time delay (LCD)	LCD1: Sensor delay = 20 min	20 min ≥ Delay ≤ 30 min	Low
		LCD2: Sensor delay = 30 min
		LCD3: Sensor delay = 40 min	30 min < Delay ≤ 40 min	Medium
		LCD4: Sensor delay = 50 min	Delay > 40min	High
		LCD5: Sensor delay = 60 min

Figure 8. A schematic illustration of k-fold cross-validation on a dataset.

Figure 9. Time series decomposition components of one-year hourly energy consumption.

Figure 10. Importance of features in training data.

Figure 11. DT classification accuracy with the utilized number of features.

Table 6. Summary of main tuning parameters of each model.

Methods	Parameters	Values	Optimal
Features Number		[1,2, 3, … ..16,17]	17
LR	Regularization parameters ‘C’	[1,2,3,4 …  … .,99,100]	26
	LR Solver	[newton-cg, lbfgs, liblinear, sag, saga]	lbfgs
DT	Criterion	[gini, entropy, log_loss]	entropy
	Max_features	[auto,sqrt, log2, None]	None
RF	Criterion	[gini, entropy, log_loss]	entropy
	Estimators number	[100,250,500,750,1000,1250,1500,1750,2000]	1750
SVC	Kernel type	[Linear, poly, sigmoid, rbf]	rbf
	Regularization parameters ‘C’	[1,2,3,4 …  … .,99,100]	98
	gamma	[scale, auto]	scale
	Decision function shape	[ovo, ovr]	ovo
MLPC	Activation function	[ Identity, logistic, tanh, relu]	relu
	Solver	[lbfgs, sgd, adam]	adam
	Hidden layers number	[1,2,3,4,5]	5
GBC	Loss function	[log_loss, deviance, exponential]	deviance
	Estimators number	[100,250,500,750,1000,1250,1500,1750,2000]	1750
	Criterion	[friedman_mse,squared_error, mse]	mse
	Max_features	[auto, sqrt, log2, None]	None
XGBoost	Estimators number	[100,250,500,750,1000,1250,1500,1750,2000]	500
	Colsample_bytree	[0.1,0.2,0.3 …  … .0.8,.09,1.0]	0.3
	Max_depth	[1,2, 3, … ..,13,14,15]	15

Figure 12. The performance of classification models in classifying each fault type in the validation dataset.

Figure 13. The performance of classification models in classifying each fault type in an unseen dataset.

Figure 14. The performance of classification models in classifying the fault type and severity in the validation dataset.

Figure 15. The performance of classification models in classifying the fault type and severity in an unseen dataset.

Table 7. A comparison of the model’s accuracy in relation to the utilization of data decomposition components.

	Fault type classification
	DT	SVC	MLPC	SKC	RF	GBC	XGBoost
Without Decomposition	80%	59%	66%	83%	87%	87%	89%
With Decomposition	91%	71%	78%	92%	95%	95%	97%
	Fault severity classification
Without Decomposition	60%	44%	49%	60%	72%	71%	75%
With Decomposition	86%	59%	67%	81%	91%	93%	95%

Figure 16. A comparison of average classification accuracy with the runtime in minutes.

Figure 17. Classification accuracy of the proposed models under different noise levels.

Table

List of Abbreviations
A/C	Air conditioning
AdaBoost	Adaptive boosting
AHU	Air handling unit
ANN	Artificial neural network
BEMS	Building energy management systems
CBM	Condition-based maintenance
CF	Condenser fouling
CO2	Carbon dioxide
DF	Dirty filter
DX	Direct expansion
DNN	Deep neural network
DT	Decision tree
EDM	Ensemble diagnostic model
EU	European Union
EXF	Exhaust fan-motor-worn-out
FFD	Fault detection and diagnosis
FN	False negative
FP	False positive
GBC	Gradient boosting classifier
HC	Healthy condition
HVAC	Heating ventilation air conditioning
KNN	k-nearest neighbours
KSA	Kingdom of Saudi Arabia
LCD	Occupancy sensor time delay
LightGBM	Light gradient boosting machine
LR	Logistic regression
LSTM	Long short-term memory
MLPC	Multi-layer perceptron classifier
PIAs	Physical industrial assets
PPM	Planned preventive maintenance
PTAC	Packaged terminal air conditioner
R	Residual
RF	Random forest
RM	Reactive maintenance
RU	Refrigerant undercharge
S	Seasonal components
SKC	Stacking classifier
SNN	Shallow neural networks
Std	The standard deviation
Sum	The summation
SVC	Support vector classifier
SVM	Support vector machines
T	Trend cycles
TN	True negatives
TO	Thermostat offset
TP	True positives
TRNSYS	Transient system simulation tool
UK	United Kingdom
VAV	Variable-air-volume
XGBoost	Extreme gradient boosting

Table

List of Nomenclature
$\mathrm{\varnothing }(\mathbf{\text{x}}{)}_{\mathbf{\text{i}}}$	Non-linear transformation that converts input data to a high-dimensional feature space in SVC
C	The cost parameter that denotes the trade-off between computational complexity and training error in SVC
${\mathit{P}}_{\mathit{j}}({\mathit{x}}_{\mathit{i}})$	Probability of fault class j in LR
T	The number of leaves in a decision tree in XGBoost
w	Weight vector in SVC
${\mathit{y}}_{\mathit{i}}$	Actual value XGBoost
$\widehat{{\mathit{y}}_{\mathit{i}}}$	Prediction value XGBoost
α	The intercept in LR
β	The coefficients that are linked to the explanatory variable in LR
$\mathit{\gamma }$	Leaf's complexity XGBoost
λ	Scaling parameter for the penalty's severity XGBoost
ξ	The slack variable to deal with misclassification in SVC
$\mathbf{\Omega }(\mathbf{\text{f}})$	Regularization that penalizes the model's complexity of XGBoost

Han, H., Z. Zhang, X. Cui, and Q. Meng. 2020. “Ensemble Learning with Member Optimization for Fault Diagnosis of a Building Energy System.” Energy Build. (Netherlands) 226: 97–110. [Online]. http://doi.org/10.1016/j.enbuild.2020.110351.

 Web of Science ®Google Scholar

Yao, W., D. Li, and L. Gao. 2022. “Fault Detection and Diagnosis Using Tree-Based Ensemble Learning Methods and Multivariate Control Charts for Centrifugal Chillers.” Journal of Building Engineering 51. https://doi.org/10.1016/j.jobe.2022.104243.

 Web of Science ®Google Scholar

Chakraborty, D., and H. Elzarka. 2019. “Early Detection of Faults in HVAC Systems Using an XGBoost Model with a Dynamic Threshold.” Energy and Buildings 185: 326–344. https://doi.org/10.1016/j.enbuild.2018.12.032.

 Web of Science ®Google Scholar

Ebrahimifakhar, A., A. Kabirikopaei, and D. Yuill. 2020. “Data-driven Fault Detection and Diagnosis for Packaged Rooftop Units Using Statistical Machine Learning Classification Methods.” Energy and Buildings 225. https://doi.org/10.1016/j.enbuild.2020.110318.

 PubMed Web of Science ®Google Scholar

McHugh, M. K., T. Isakson, and Z. Nagy. 2019. “Data-driven Leakage Detection in air-Handling Units on a University Campus.” ASHRAE Transactions 125: 381–388.

 Google Scholar

Data availability statement

The data that support the findings of this study are available from the corresponding author, [AYK], upon reasonable request.