Chicken Swarm Optimization with Deep Learning Based Packaged Rooftop Units Fault Diagnosis Model

: Rooftop units (RTUs) were commonly employed in small commercial buildings that represent that can frequently do not take the higher level maintenance that chillers receive. Fault detection and diagnosis (FDD) tools can be employed for RTU methods to ensure essential faults are addressed promptly. In this aspect, this article presents an Optimal Deep Belief Network based Fault Detection and Classification on Packaged Rooftop Units (ODBNFDC-PRTU) model. The ODBNFDC-PRTU technique considers fault diagnosis as a multi-class classification problem and is handled using DL models. For fault diagnosis in RTUs, the ODBNFDC-PRTU model exploits the deep belief network (DBN) classification model, which identifies seven distinct types of faults. At the same time, the chicken swarm optimization (CSO) algorithm-based hyperparameter tuning technique is utilized for resolving the trial and error hyperparameter selection process, showing the novelty of the work. To illustrate the enhanced performance of the ODBNFDC-PRTU algorithm, a comprehensive set of simulations are applied. The comparison study described the improvement of the ODBNFDC-PRTU method over other recent FDD algorithms with maximum accuracy of 99.30% and TPR of 93.09%.


Introduction
Fault detection and diagnostic (FDD) models have gained significant interest in the area of spaceconditioning equipment as an approach to reducing energy consumption and refrigerant emissions and providing more reliable comfort. Air conditioning equipment utilizes a substantial amount of energy in the United States. In commercial buildings, mostly medium-sized and small buildings, packaged rooftop unit (RTU) system becomes the most typical type of air-conditioning system. RTUs were employed in 46% of all commercial buildings and serve over 60% of commercial building floor space in the United States [1][2][3]. The main energy utilization linked with RTUs was beyond 2.6 quads yearly. Thus, a minor enhancement in efficacy or part load functioning of such units could cause important minimizations of carbon emissions and energy utility [4]. Apart from discovering methods for improving the part load performances, there is a necessity to improve the persistence of RTU functions. For instance, an air-side economizer could offset an important fraction of the cooling necessities in dry and mild weather conditions [5,6]. But several researchers proved that a substantial fraction of economists has control over RTUs that do not work as envisioned [7]. Along with the operational perplexities with air-side economists, several researchers stated issues with the refrigerantside functioning as well. Fig. 1 showcases the sample infrastructure of RTU. Though the advanced controller grants a substantial development in part load efficacy, it does not assure the persistence of functions [8][9][10]. To ensure persistence, there comes a necessity for automatic fault detection and diagnosis (AFDD) [11]. RTU diagnoses (refrigerant-side and air-side) were typically positioned offline through a collection of data from the RTUs, which were compiled for framing automatic mechanisms or utilizing retrofit packages [12]. Diagnosis utilizing offline approaches [13] was of limited utility because many RTUs do not install every sensor essential for diagnosis, particularly a refrigerant-side diagnosis. Using a retrofit monitoring and diagnostics package could ease the sensor's complexities [14]. But it is costly to install an AFDD method [15]. To meet the cost, DOE has funded the expansion of a low-cost monitoring and diagnostic approach [16]. The low-cost method minimizes the requirement to sense by utilizing 2 (outdoor-air temperature and unit power) or 3 sensors. The low-cost approach could identify the degradations in RTU operations over time; however, it contains confines in detecting the real cause of the fault [17].
Many FDD techniques formulated for packaged air conditioners are rule related means a set of professional rules which is based on the physical traits of the mechanism were implemented to measure data [18,19]. Rules are implemented for functioning parameters like observable conduct or suction superheat [9]. Several rule-related FDD tools were effective, but there comes a challenge. A common difficulty for the FDD techniques is differentiating unfaulted and faulted operations in unusual operating circumstances. Typical rule-related approaches imply an exceeded threshold, and choosing this threshold suitably becomes difficult [20,21].
Zhang et al. [22] formulated a new analysis method that broadly assesses sensor fault on sensor choosing and FDD accuracy. Monte Carlo simulations can be implemented for dealing with many stochastic sensor inaccuracies and offer probabilistic scrutiny fallouts of the effect of sensor inaccuracies on FDD accuracy and sensor selection. This method emphasizes the net effect of the fault state over a complete sensor set. The advanced method is utilized for operation stage sensors maintenance and prior-stage sensor model. Bezyan et al. [23] introduce a new technique for diagnosing and detecting many faults in the air handling unit (AHU) of the HVAC mechanism of an institutional building during heating seasons. The presented algorithm guides the search for faults through the operation flow and information among sensors. Support vector regression (SVR) methods, advanced from the building automation system (BAS) trend dataset, forecasted air temperatures of 2 targeted sensors in normal functioning circumstances without familiar issues.
Woradechjumroen et al. [24] devise a new fault-detecting approach for fault HVAC functions based on issues in repetitive functions and the excess energy utility of the refrigeration mechanism. Four steps were advanced as a new unfavourable communication approach for identifying abnormal HVAC functions related to identifying energy signs. Zonal and outdoor air temperatures (OAT and ZAT) were used to specify common area functions for rooftop units (RTUs). An FD method was projected related to RTU outliers utilizing plots of ZAT and OAT vs. the refrigeration method's energy utilisation related to fixed 10% changes from the indoor relative humidity range. Rogers et al. [25] offer a technique for obtaining chief functioning parameters for air conditioning systems (system efficiency, airflow rate, refrigerant mass flow, and cooling capacity) in a way, i.e., more suitable for the residential sectors. The devised air-side sensing approach needs fewer sensors and is an important benefit for residential split mechanisms due to the necessity of no outdoor unit sensor.
Granderson et al. [26] offer a primary public dataset with ground-truth data on the absence and presence of constructing faults. This data spans a range of seasons and operational circumstances and involves many construction systems kinds. In [27], the difficulty of the VAV technique demonstrating is dispersed with hierarchical modelling structure, and system level to element scale FDD strategy for VAV mechanisms utilizing integrated quantitative, residual, and qualitative methods was advanced. The 3 layers of the modelling structure, like the component layer, system layer, and unit layer, adopt distinct modelling and FD approaches for improving the FDD's reliability and accuracy.
This study presents an Optimal Deep Belief Network based Fault Detection and Classification on Packaged Rooftop Units (ODBNFDC-PRTU) model. The ODBNFDC-PRTU technique considers fault diagnosis a multi-class classification problem and is handled using DL models. For fault diagnosis in RTUs, the ODBNFDC-PRTU model exploits the deep belief network (DBN) classification model, identifying seven distinct types of faults. At the same time, the chicken swarm optimization (CSO) algorithm-based hyperparameter tuning technique is utilized for resolving the trial-and-error hyperparameter selection process. A comprehensive set of simulations is implemented to depict the improved performances of the ODBNFDC-PRTU method.
The rest of the study is organized as follows. Section 2 offers the proposed model, and Section 3 discusses the performance validation. Finally, Section 4 draws the concluding remarks.

Materials and Methods
In this study, a new ODBNFDC-PRTU model was introduced to identify and classify faults in PRTUs. The ODBNFDC-PRTU approach considers fault diagnosis as a multi-class classification problem and is handled using DL models. For fault diagnosis in RTUs, the ODBNFDC-PRTU model uses the DBN classification model, which identifies seven distinct types of faults. At the same time, the CSO algorithm-based hyperparameter tuning technique is utilized for resolving the trial-and-error hyperparameter selection process.

Data Collection
Faults in RTU methods are separated into 2 classes: soft faults and hard failures. The hard failures take place rudely and reason RTU to stop operating. The soft faults reduce system efficiency; however, they allow the continued function of the method. Hard failures like compressor failure often occur from RTU methods and are costly to repair [28]. But, it can be simply identified and analyzed with an inexpensive sensor. For the sample, the compressor failure is simply identified and detected by verifying a compressor's outlet and inlet temperatures. The hard failures were classically produced by extended function periods with soft faults. The concentration of this examination is on detecting and diagnosing the subsequent soft faults: (i) evaporator airflow reduction (EA), (ii) refrigerant undercharge (UC), (iii) non-condensable gas (NC), (iv) refrigerant overcharge (OC), (v) compressor valve leakage (VL), (vi) liquid-line restriction (LL), and (vii) condenser airflow reduction (CA).

Fault Detection and Classification
In this study, the DBN method was employed to detect and classify faults. Based on biological neural networks, a DBN is a probability generation module that has been proposed [29]. The DBN infers the sampling distribution for the input dataset based on the joint likelihood distribution. The whole network generates an inherent feature dataset with maximal probability by fine-tuning the weight coefficients between the network layers. Fig. 2 showcases the infrastructure of DBN, which is generally composed of a classifier and restricted Boltzmann machine (RBM) a. A classifier is a crucial component in accomplishing fault detection in a DBN. In this study, different kinds of faults should be identified. Consequently, an RBM and a Softmax classification are selected to construct the DBN. Unlike the conventional binary classifiers, a Softmax classification is preferable for classifying many conflicting events. RBM is an undirected graph method that is utilized for resolving the value of the variable θ to fit the provided training dataset and the extracted features.
The RBM task was utilised to fit the input-trained dataset, design the optimum parameter θ , and perform feature extraction. The variable θ is learned in the training subset to exploit the logarithmic probability function: (1) The aim is to resolve the optimum variable θ * to attain the partial derivative of log P(V (t) |θ) for W ij , a i , b j , and other variables. Assume that θ was a variable value of θ, the logarithmic probability function concern θ as ∂L ∂θ Because the sample count of T was identified, the partial derivative of the logarithmic likelihood function for the connection weight W ij , the offset a i of the visible unit, and the offset b j of the hidden element is formulated as P(h|V (t) , θ) and P(v, h|θ). P(h|V (t) , θ) indicates a hidden likelihood distribution of training instance V (t) ; P(v, h|θ) was a joint likelihood function for a provided state (v, h) as follows In Eq. (4), E(v, h|θ) denotes the energy function of RBM, and Z (θ ) indicates the normalization factor.

Hyperparameter Tuning
The CSO technique was utilised in this work for optimal modification of the hyperparameters related to the DBN method. The concept for the basic CSO algorithm is shown in the following [30]: 1) Chicken swarm comprises distinct groups. Each group has a pair of hen, chick, and dominant roosters. 2) According to the fitness value estimated for individuals in this chicken swarm, an individual with a good fitness value is applied as a rooster, whereas some individuals with poor fitness value are applied as a chick. Other individuals are applied as hens. 3) Given that the several roosters in a chicken swarm as RN, several hens as HN, several chicks as CN, and several chick mothers as MN. The individual hen is arbitrarily chosen as the chick mother from each group. 4) Afterwards, for each G iteration, the individual is re-determined based on the fitness value and is applied as a rooster. An individual is applied as a hen, and other individuals are applied as a chick. 5) Within each population group, the hen follows the rooster to discover food and compete arbitrarily with other individuals for food. An individual with a better fitness value in the population is more likely to receive food.
Chicks, roosters, and hens in the chicken swarm utilize different position upgrading equations, and it is shown in the following Consider R represents the individual rooster, N 0, σ 2 denotes a Gaussian distribution with mean and variance, σ 2 , and it is given below.
Now, ε is applied to prevent zero division error, and ε shows the minimal constant. i and k ∈ [1, 2, . . . , RN] indicate the rooster index arbitrarily chosen from the rooster groups, where i do not often correspond to k. f i and f k characterize the fitness value of i-th and k-th roosters as follows Here, rand characterizes a uniformly distributed arbitrary number within [0, 1]. R signifies that the individual was a rooster. H denotes that the individual was a hen. r1 ∈ [1, 2, . . . , RN] denotes the rooster index, viz., i-th hen groupmate, where r2 ∈ [1, 2, . . . , RN + HN] characterizes the chicken index (hens or roosters), viz., chosen arbitrarily from the swarm. r1 isn't often corresponding to r2, as follows Here, m denotes the chick mother indices in the i-th group, C shows the individual is a chick, and H shows the individual is a hen. FL signifies a uniformly distributed arbitrary number within [0, 1].
The CSO system grows a fitness function (FF) for accomplishing higher classifier performances. It explains a positive numeral for representing optimal performances of candidate results. The minimized classifier error rate was supposed to be that FF is providing in Eq. (12).
the number of misclassified samples total number of samples * 100.
A clear precision-recall analysis of the ODBNFDC-PRTU algorithm on the test dataset is depicted in Fig. 9. The figure signifies the ODBNFDC-PRTU technique has resulted in enhanced values of precision-recall values in all classes.  A brief ROC analysis of the ODBNFDC-PRTU method on the test dataset is depicted in Fig. 10. The outcomes represented the ODBNFDC-PRTU algorithm exemplified its ability to categorize distinct classes on the test dataset.  Table 5 reports an overall comparison study of the ODBNFDC-PRTU model with recent models interms of accu y and TPR [30]. Fig. 11 demonstrates a brief accu y investigation of the ODBNFDC-PRTU approach with existing methods. The figure implied the LR and RF approaches had exposed poor results with minimal accu y of 80.01% and 82.18%, respectively. Then, the QDA, LDA, BA, AD, and XGB models have resulted in slightly enhanced accu y of 83.27%, 83.95%, 84.5%, 85.36%, and 85.79%, respectively. Following, the SVM model has obtained a considerable accu y of 89.08%. Although the KNN model has reached a reasonable accu y of 93.85%, the ODBNFDC-PRTU model has shown enhanced results with a minimal accu y of 99.30%.  Figure 11: Accu y analysis of ODBNFDC-PRTU approach with existing methodologies Fig. 12 validates a brief TPR study of the ODBNFDC-PRTU approach with existing techniques. The figure implied the LR and RF algorithms have exhibited poor results with minimum TPR of 80.25% and 80.01% correspondingly. Next, the QDA, LDA, BA, AD, and XGB models have resulted in slightly enhanced TPR of 86.46%, 82.74%, 83.13%, 86.86%, and 88.48%, correspondingly. Then, the SVM algorithm acquired a considerable TPR of 83.43%. Although the KNN approach has gained a reasonable TPR of 82.78%, the ODBNFDC-PRTU technique has exposed enhanced results with a minimal TPR of 93.09%. From these results and discussion, it is assumed that the ODBNFDC-PRTU model has accomplished enhanced performance over other models.

Conclusion
In this paper, a new ODBNFDC-PRTU technique was introduced for the identification and classification of faults in PRTUs. The presented ODBNFDC-PRTU technique considered fault diagnosis as a multi-class classification problem and is handled by the use of DL models. For fault diagnosis in RTUs, the ODBNFDC-PRTU model makes use of the DBN classification model, which identifies seven distinct types of faults. At the same time, the CSO algorithm-based hyperparameter tuning technique is utilized for resolving the trial-and-error hyperparameter selection process. For showing the enriched performances of the ODBNFDC-PRTU model, a comprehensive set of simulations is applied. The comparison study reported the improvement of the ODBNFDC-PRTU technique over other recent FDD methodologies with maximum accuracy of 99.30% and TPR of 93.09%. In future, data handling and feature selection algorithms are incorporated to enhance the classification results.
Funding Statement: This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2020-0-00107, Development of the technology to automate the recommendations for big data analytic models that define data characteristics and problems).

Conflicts of Interest:
The authors declare that they have no conflicts of interest to report regarding the present study.