MD-MARS: Maintainability Framework Based on Data Flow Prediction Using Multivariate Adaptive Regression Splines Algorithm in Wireless Sensor Network

The demand for Wireless Sensor Networks is increasing day by day because of their diverse nature. Due to the limited energy, it is a complex task to retract the sensor node after deployment. So, there is a requirement for network maintainability before the deployment phase for its smooth working. It is achieved in three phases: hardware of the sensor node, communication and external environmental phase. This paper focuses on network maintainability in the communication phase. A novel framework MD-MARS is presented to enhance the network maintainability. This framework is classified into three phases namely analysis of performance parameters, data flow optimization and maintainability evaluation. In the initial phase, the performance parameter is analyzed using NS2 simulator. The next phase deals with data flow optimization using a machine learning algorithm. It reduces congestion and enhances network performance. The proposed algorithm is finely tuned to different degrees using the Grid Search approach to achieve the highest accuracy. The best model is selected based on accuracy and minimizes the prediction error. This algorithm predicts with the highest accuracy of 99.83%, lowest being 21.17%. Maintainability is achieved in the last phase using the total time taken to optimize the data flow. Several observations of repair time are determined for the best-tune model during the prediction of optimized data flow. These observations are used to calculate the mean time to repair, standard deviation, probability density function, maintainability and repair rate. The maximum maintainability achieved in this paper is 97.67% at a repair time of 26.07 milliseconds.


I. INTRODUCTION
The rapid growth of smart sensors has enabled Wireless Sensor Networks (WSNs) usage in diverse applications such as health care, industrial monitoring, urban monitoring, The associate editor coordinating the review of this manuscript and approving it for publication was Guangjie Han . environmental monitoring, prediction of natural disasters, military surveillance, Internet of Things (IoT), flora and fauna [1], [2], [3], [4]. Maintainability is a significant parameter of Quality of Service (QoS) which is mandatory in a real-time environment [5], [6]. It is defined as ''the ability of system under given conditions of use, to be retained in, or restored to, a state in which it can perform a required function, when maintenance is performed under given conditions and using stated procedures and resources'' [7]. FIGURE 1 represents an anatomy of maintainability along with its threats, attributes and definitions [8]. Threats are unlawful activities that main aim is to breach information. It can be accidental or intentional which leads to network failure. Maintainability is not a single measure, it is evaluated based on several attributes such as performance, fault tolerance, availability, reliability, safety and security. Network performance deals with the quality of the network [9]. Fault tolerance is the capability of the network to work smoothly in the presence of faults and is correlated to reliability [10], [11]. Availability is ''the ability of a network to be in a state to perform a required function at a given instant of time within a given time interval; assuming that the external resources if required, are provided''. Reliability is defined as ''a measure of the continuity of correct service'' [12]. Safety deals with the harmless state of the network. Security is ''a process to design an activity to protect the network from unauthorized users''. It provides integrity, confidentiality and authenticity [13]. Integrity belongs to the consistency of information; confidentiality deals with sensitive information and authenticity is the process of user vetting through their credentials. It provides several essential and non-essential services based on applications. These services are performed across the network such as fault resistance, fault recognition and fault recovery. It is a challenging task to maintain the good functioning of WSNs due to the stochastic behavior of these networks [14], [15], [16]. The unreliable nodes can cause network failure and affect the network maintainability. Here, QoS needs serious efforts to make the network reliable and efficient with optimal packet delivery time with the minimum possible failure rate. Maintainability can be achieved from two perspectives: Time based parameters and Economic based parameters [17]. Time based parameters deal with network-specific layers. It is evaluated using optimal data flow. Mean time to repair (MTTR), standard deviation, probability density function (PDF) and repair rate is considered time-based parameters of maintainability. Economic based parameters deal with user-specific layers. It comprises direct maintenance cost/hour and direct maintenance manhours/hour [18]. These issues have attracted researchers to work on maintainability parameter in WSN. In this paper, maintainability is estimated at the designing phase of the network which improves the maintainability in the operational phase. In this paper, it is evaluated based on performance attributes in terms of time-based parameters such as Mean Time to Repair (MTTR) and repair rate.
The maintainability depends on various factors demonstrated in FIGURE 2. It is divided into three phases as follows [19]: Hardware of Sensor Nodes: This phase deals with the designing of hardware sensor nodes. The energy of the sensor nodes is limited and depleted during working operation. The maintainability of the network is directly affected by power failure. Generally, simple nodes are designed for WSN due to hardware cost. There is a maximum chance of simple node failure in an adverse environment.
Communication: The various factors affect communication such as data flow, topology, packet loss and transmission delay. When multiple nodes transmit information simultaneously in a limited bandwidth, then a collision will occur which increases the data loss rate and leads to poor maintainability. Due to the self-organizing characteristic of WSN stable topology should be considered [20]. Adjusting the data flow mechanism reduces packet collision and packet loss.
External Environment: WSN works in harsh environments such as outer space, wild forests, oceans and volcano regions where external factors affect the normal working of the network such as geographical factors, weather factors and electromagnetic interference.
The system workflow is shown in FIGURE 3. In the initial phase, sensor nodes are deployed based on Relative Identification and Direction-Based Sensor Routing (RIDSR) topology. This topology is reliable and energy-efficient [21], [22]. The primary data is generated using NS-2.35 simulator based on eleven performance parameters [12]. These parameters are analyzed at the network layer. In the next step, the reliability of network is examined. If the working of the network is smooth then there is no need for data flow optimization. But if performance is hampered then the prediction of optimized data flow is performed using Multivariate Adaptive Regression Splines (MARS) algorithm. It includes the basic MARS model and the hyperparameters based MARS model. Hyperparameter models are selected based on the Grid Search approach. Hyperparameters are tuned at different degrees such as degree 1, degree 2 and degree 3. The best-tune model is selected based on the highest accuracy to achieve optimized data flow. The total time taken for optimal DF is known as time to repair or restore time. Time to repair observations are collected for the best-tuned model. These observations estimate the MTTR value. MTTR further evaluates the standard deviation. MTTR and standard deviation are used to  determine the probability density function (PDF). It is used to calculate the maintainability across the WSN. Maintainability and PDF are used to estimate the repair rate.

A. CONTRIBUTION
This research work emphasizes data flow parameter (DF) that improves the network performance in terms of maintainability. This DF optimization minimizing the count of fault and regulates the flow of data [23]. It maintains the good functioning of a network where sensor is deployed in specific topology. ML algorithm provide generalized solution for data optimization using accuracy.
The acronyms are listed in TABLE 1 for the readability of the researchers.
The contribution of the paper is summarized as: • MD-MARS framework is presented to improve network maintainability.in quantitative manner. Protocol Name (PN). DF is act as the target parameter and the other ten parameters are used as input parameters.
• MARS model is used to optimize the DF parameter and collects the samples of repair time during DF optimization.
• The repair time determines the maintainability using normal distribution method.

B. ORGANIZATION
The paper is organized as: Section I presents the introduction to network maintainability and also describes the acronyms in tabular format. Section I-A elaborates on related work in detail. Section II presents an efficient framework MD-MARS for maintainability analysis. It includes the formulation of maintainability, dataset and MARS ML algorithm. The performance evaluation and detailed discussion have been described in Section III. Section IV presents findings of the research work and Section V concludes the article with future research directions.

II. RELATED WORK
Several prediction models are used by researchers to estimate maintainability. It can be determined in various aspects such as software maintainability as well as network maintainability. Software maintainability is crucial for the success of software. It is predicted using distinct ML approaches [23]. It includes fault correction inclusion of new code and removal of obsolete code [24]. An imbalanced dataset generates low maintainability due to biased predictions. The safe-level-SMOTE approach is used to preprocess the unbalanced dataset before software maintainability prediction [25]. Code smell issue is addressed in an article [26]. It is a confusing, complex and unstructured code of the software. This code is identified by a fuzzy genetic based automatic refactoring approach. Naïve ayes classifier corrects the software component and reduces the fault rate [27]. Dependency Injection mechanism is used to improve maintainability [28]. Network maintainability is performed by Fault Management Framework (FMF). It includes fault identification, tolerance and recovery mechanism [29]. These faults can cause failure to occur in the network. Different type of fault is identified by distinct mechanisms that achieve network reliability [30]. Reliability is correlated with maintainability and availability [31], [32]. Network maintainability depends on several factors such as communication and external factors [33]. Network maintainability is determined from two perspectives: economic and time-based parameters. Economic parametersbased maintainability is evaluated using a Bayesian network [15]. Time based maintainability is achieved in terms of repair rate. The data discretization mechanism predicts the optimal data flow [34]. An optimized data rate provides a congestion-free environment to improve communication [35]. It maintains network reliability. The above articles have a specific framework or model based on their problem formulation. Our proposed framework provides a generalized efficient framework to enhance maintainability using data flow optimization. The previous literature is summarized in tabular form as TABLE 2.

III. PROPOSED FRAMEWORK FOR MAINTAINABILITY
Reliable network performance is mandatory for the good functioning of the network [46], [47]. The presence of fault occurrence is more prone in WSNs as compared to traditional wireless networks. This fault generates an error that leads to network failure and depletes network performance [48]. Maintainability is a parameter which is affected by an accidental failure. Here, MD-MARS framework is proposed to enhance the network performance shown in FIGURE 4. The  proposed framework is divided into three phases: network parameters analysis using simulation, data flow optimization using MARS algorithm and maintainability achieved using time to repair.

A. ANALYSIS OF NETWORK PARAMETERS
This is the initial phase of the network where sensors are randomly deployed based on RIDSR topology to provide reliable and energy-efficient routing in a static environment [22]. In RIDSR, the sensing area is divided into sectors. A base station provides a unique sector ID to every sector based on the quadrant name and provides an estimated distance from the base station. In each sector, the manager node is present. Each sensor node is represented with its hop ID. If the manager node and sensor node are present in the same sector with communication range, then hop ID 1 is assigned to the sensor node. It determines the shortest distance for routing with less complexity. This network configuration is performed using NS-2.35 simulator. The simulation parameters are shown in TABLE 3. The channel type used for communication is wireless. Two-Ray Ground is used as a radio-propagation model for routing from source to destination node. Omnidirectional is selected as the antenna model. DropTail and CMUPriQueue are used in Interface Queue. As the name VOLUME 11, 2023  suggests, DropTail drops packets when the interface queue is full. CMUPriQueue is used to prioritize the packets. The maximum packets interface queue is set to 150. The number of nodes considered in the network design is between 5 to 50. 1000 m × 1000 m area is selected for simulation. Dynamic Source Routing (DSR) and Ad-hoc On-Demand Distance Vector (AODV) protocols are used for routing. AODV and DSR are reactive and demand-driven routing protocols that initiate the discovery of a route on demand [49]. The reactive protocols outperform as compared to proactive protocols. It generates less routing overhead and consumes minimum resources. The data flow range lies between 0.1 to 10Mbps and the simulation time is set to 20 milliseconds. Based on these simulation parameters dataset is constructed [12]. It comprises 10,000 records with eleven performance parameters. The description of these parameters is shown in TABLE 4. The values of SP, RP, RA, RO, APL, PDR, PF and TH are measured using simulation. These values are evaluated using trace files generated during simulation. These files are analyzed by awk scripts which are used for generating reports. The values of PN, NN and DF are userspecific. The dataset sample is represented in TABLE 5. This dataset is free from missing and redundant values.
The performance parameters are used to optimize DF to achieve maintainability. Ten parameters are used as an input variable and DF behaves as the target variable. These parameters are correlated with each other and their diagrammatic representation of correlation is shown in FIGURE 5. The correlation is used to determine the linear relationship between two variables or features of a dataset. It is calculated as: where r is correlation coefficient, x represents actual value,x depicts the mean of actual values, y indicates predicted value, y depicts the mean of predicted values and NI represents the count of instances.
Here, a matrix is created based on Pearson Correlation that determines the relationship between pair of parameters [50], [51], [52]. Pearson Correlation Matrix (PCM) of the dataset is represented in FIGURE 6. It measures the relationship between variables of the dataset. It comprises two properties strength and direction. Strength deal with the linear relationship and direction deals with the direct or inverse relationship among variables. The range of r lies between +1 to −1. The + sign represents a positive or direct relationship between parameters and the -sign denotes an inverse relationship between parameters. The strength depends on the value of r. If the value of r lies between 0.7 to 1, then it belongs to a strong relationship. A value less than 0.3 shows weak relationship. If the value of r is 0 then it means there is no linear relationship between parameters. DF is strongly associated with the PF variable. When DF is optimized then it will minimize PF across the network [12].   maximum degree of interaction and count of retained terms are the hyperparameters of the MARS Model. These parameters are tuned at different degrees to achieve an optimal result. Tuning is performed using a Grid Search Algorithm that builds every combination of hyperparameters and estimates each model. At last, it provides the best combination of hyperparameters to reduce prediction error. The different degrees of interaction such as degree 1, degree 2 and degree 3 are applied to MARS training that provides a distinct trained model. The best-tune model is the best-fit model that is selected based on accuracy for the prediction of optimal DF. The total time taken during DF optimization is used in phase 3.

1) MULTIVARIATE ADAPTIVE REGRESSION SPLINES ALGORITHM (MARS)
MARS algorithm is a flexible regression model that resolves the complex problem of high dimensionality on a large scale. This model uses lesser variables to maintain a relationship between dependent and independent variables [53], [54], [55]. It utilizes recursive auto-regressive and projection tracking methods. MARS comprise multiple spline functions which are known as basis function (B F ).
where z is the output variable dependent on the u parameter.
Here, e represents an error vector (1 X h). The main aim of MARS algorithm is to perform regression analysis on a dataset using B F . It fits the best model for a non-linear relationship among the parameters. It splits the dataset into smaller regions and each region behaves as a linear function. B F is applied at each region for best fit. MARS model is represented using basis function B F . Each B F comprises hinge functions and each hinge function is represented as a knot. where In MARS formula,ẑ is the predicted value of the DF parameter. α k represents as a coefficient of a k th spline function, B k (U) acts as a spline function and N depicts the count of nodes. The value of S N is either a positive one (+1) or a negative one (−1). A positive one represents the function in the right direction and a negative one shows function in the left direction. v(N) is the independent variable's identifier. t N is the node's position. Here B k (U) can be a single or multiple spline function.
Eqautions (9) and (10) show that there is a knot present due to the presence of two linear models. When the linear model increases then the knot will be increased. But there is a requirement for an optimal knot which gives the best results.
Algorithm 1 illustrates the MARS Algorithm. DF is an output variable and the other 10 parameters are used as an input variable. MARS algorithm includes three steps to achieve the output namely Forward Pass, Backward Pass and Model Selection. A forward mechanism is used to construct the B F that is required for optimum results. It splits the dataset and applies the spline function to each region for best fit and determine new B F . Due to the large count of B F in a Forward Pass generates an overfitted model. This issue is resolved by using a Backward Pass mechanism. It eliminates duplicates B F and provides the best fit model with accuracy. Finally, the best model is selected from many models based on the accuracy parameter.

2) FORWARD PASS ALGORITHM
In basis iteration (BI) zero, the first pair of B F is constant in the MARS algorithm i.e. B F0 = 1 and every base iteration generates two B F when BI is greater than one (BI>1).
where B p (u) shows the previous iteration B F and u v behaves as an input variable. Here, t is the spline base's node. At every iteration, the constructed model accuracy is affected by the node's position. The total estimated time of the MARS model depends on the count of nodes. It is not feasible to calculate each input data that can cause a very small distance to be created between adjacent nodes. This issue can be avoided by considering minimum step size D for every input variable. Its formula is defined as: The range of b lies between 0.01 to 0.05 (0.05>b>0.01). Model accuracy will not be affected after defining the minimum step size. It will become fast as well as minimize calculation time to build a model. This model takes care of interaction among distinct functions and improves the model's accuracy. During the building model, B F is continuous increases till the maximum value (K max ).
Algorithm 2 represents the forward pass algorithm. The important point about this algorithm is that it produces a large count of B F that will generate results as overfitting.

3) BACKWARD PASS ALGORITHM
The Forward Iteration Pass constructs a large count of B F that produces overfit results and increases the model's complexity. It doesn't provide a generalized solution. This problem is solved by using Backward Pass Algorithm. It is a pruning process that deletes B F which are generated during a forward pass based on Generalized Cross-Validation (GCV) criterion.
It is used to penalized model complexity.
where P(K) represents a penalty function. best fit B f . 8 endif 9 end for 10 end while K is denoted as a count of B F , B represents a matrix of K X L, trace (B(B T B) −1 B T )+1 is used as a count of effective coefficient in the model and a depicts as a coefficient of penalty.The optimal MARS model is selected based on the value of GCV. The model which has a minimum GCV value, that model is selected as the optimal MARS model. At last, the formula of MARS model is achieved as: Algorithm 3 represents Backward Pass Algorithm. It provides a generalized solution to predict the output with accuracy.

4) MODEL SELECTION
It is defined as the process to determine the best-fit model for the predictive modelling problem. It can be selected based on performance evaluation metrics such as correlation, coefficient of determination, root mean square error and accuracy. 1) Correlation (r): It determines the relationship among all features of a dataset. The formula of correlation is represented as equation (1).

2) Coefficient of Determination ( R 2 ): The primary result of MARS is evaluated as a coefficient of determination.
The range of R 2 is between 0 to 1 (1> R 2 >0). If the value of R 2 is towards zero then it defines the failure of the MARS model. But if its value is near 1 then it shows the best fit to the MARS model. Mathematically, it is represented as the square of correlation. It is estimated as: (iii) Root Mean Square Error (RMSE): RMSE represents the total amount of error that occurs between the predicted and actual output. The formula of RMSE is defined as: where y is the predicted output variable, x is the actual output variable and NI is the total count of instances.
(iv) Accuracy: Accuracy determines the closeness between the actual and predicted value. The best model is selected for a specific problem based on accuracy [36]. It is estimated as: where y i is the predicted target output, x i is actual output, NI indicates the count of instances and er depicts an acceptable error. The selected MARS model provides samples of the restoration or repair time of the network. The total time taken to optimize the DF parameter is known as repair time or restore time. This parameter is used in phase 3 to evaluate maintainability.

C. MAINTAINABILITY EVALUATION
The observations of repair time are collected during optimization shown in TABLE 6.
These values determine the MTTR, standard deviation, probability density function (PDF), maintainability and repair rate. The mathematical formulation of each parameter is discussed below:

1) MATHEMATICAL FORMULATION OF MAINTAINABILITY
''Maintainability is the probability that a failed system or component will be restored or repaired to a specified condition within a specified period of time when maintenance is performed in accordance with prescribed procedures'' [56]. It is estimated using the normal distribution method. This method is used for the analysis of maintainability for straightforward repair actions. Straightforward action deals with simple removal and replacement tasks. The formula of maintainability is expressed as [8]:   Here, g(t) indicates the probability density function (PDF) or repair density function. It is defined as ''the probability of faulty system or component is repaired to normal condition in δt'' [56]. It is calculated as: MT ct i is the individual maintenance action (repair time). The average maintenance of n observation is known as Mean Time To Repair (MTTR) and is calculated as: The median time to repair is equal to MTTR due to the symmetry of normal distribution and is expressed as: The standard deviation of n observations during maintenance action is: The maximum time to repair is defined as it the maximum time required to complete all maintenance actions for a specific percentage. It is expressed as: (24) φ is the value of the normal distribution function for the percentage of maintainability function. It evaluates the maintainability.
The repair rate is defined as ''it is the conditional probability that the component or system is repaired to specified function in (t, t+ δt) when the component or system fails at time t'' [57]. It is expressed as:

MD-MARS framework evaluates network maintainability.
Here, the dataset is generated using NS2 simulator [12]. Simulation parameters are discussed along with dataset features in Section II. MARS algorithm with different tuning parameters is applied using R Studio for optimal data flow. FIGURE 8 represents the fitting of MARS model. In model selection, the x-axis represents the count of retained terms, the y-axis(right) indicates the count of used predictors and the y-axis (left) shows GRSq and RSq. GRSq is a GCV R 2 which is represented as a solid black line. It is the actual value of the data flow parameter. RSq is a predicted value of the data flow parameter and is represented as a red dotted line. 10-fold Cross-Validation generates optimal target output using tuning of hyperparameters. There are two significant hyperparameters namely the count of retained terms and the degree of interaction. These hyperparameters are tuned to three degrees based on the Grid Search algorithm. In our results, the cross-validated RMSE of three degrees is represented in FIGURE 9. The predicted RMSE value at degree=1 is 0.26, degree=2 is 0.13 and degree=3 is 3.3. This shows that degree=2 gives the best results in terms of RMSE. TABLE 7 shows the comparison of performance in terms of RMSE, correlation, R 2 and accuracy. The correlation of MARS without tuning is 0.99, R 2 is 0.99, RMSE value is 0.26 and accuracy is 96%. At degree=1, the performance parameter value is the same without tuning performance parameter. At degree 2, the correlation is 1, R is 1, RMSE is 0.13 and accuracy is 99.83. At degree 3, the correlation is 0.03, R is 0, RMSE is 3.3 and accuracy is 21.17.
The final model is selected based on the accuracy of the predicted model. Here, degree 2 shows the maximum accuracy of 99.83 as compared to other tuning degrees 1 and 3. The MARS model with tuning (degree=2) is selected as the final model in the proposed framework. The final MARS model includes four predictors of the dataset such as PF, RA, RP and TH. These predictors show their importance as compared to other predictors. It is selected based on the impact on GCV and Residual Sum of Squares (RSS) values.  GCV is used for the trade-off between model complexity and best fit. RSS is a technique used to determine the level of variance that occurs in the error term. The value of RSS defines the level of model fitness. For high value, it is a poor fit and for low value, it is a better fit. But if the value of RSS is zero, it defines the model as a perfect fit. FIGURE 10 shows that all four predictors have little change when the count of terms is increased.
When accuracy is compared with existing techniques such as random forest, weka lazy model, conditional inference tree, ensemble model with equal weight [12], KNN [58], Bayesian [59], Ensemble Classifier [60] and enhanced deep reinforcement learning [61], it will show that the proposed framework outperforms well shown in TABLE 8. FIGURE 11 shows the scatter plot relation between the actual and predicted MARS model with and without tuning parameters. FIGURE 11(a) and 11(b) show variations between the actual and predicted output. FIGURE 11(c) represents the best-fit relation between the predicted and actual model where the tuning parameter is degree 2. FIGURE 11(d) shows the model is poorly fit at degree 3.
The total time duration to achieve optimized data flow is termed repair or restore time. The sample of repair time is collected during optimization at degree 2. It is expressed in terms of milliseconds. This repair time is used to evaluate the probability density function (PDF), normal distributed function z(t), maintainability MT(t) and repair rate µ(t) shown in  TABLE 9.
The graphical representation of the probability density function, cumulative density function or maintainability and repair rate with respect to repair time is shown in FIGURE 12, 13 and 14. FIGURE 12 represents a plot of the normal distribution pdf of the repair time to optimize the data flow. FIGURE 13 represents a graphical representation of maintainability with respect to repair time.
When maintainability is compared with the existing approaches then it performs well shown in TABLE 10. The proposed framework achieved maximum maintainability of 97.67% at a repair time of 26.07 milliseconds. FIGURE 14 represents the repair rate is directly proportional to the repair time. It is increasing with increasing repair time.

V. FINDING DISCOURSE
The proposed framework evaluates network maintainability quantitively using the normal distribution method. NS2 simulation scenario is created for a dataset. RIDSR topology is used for deployment which enables energy-efficient data delivery from sensor nodes to a base station and also enhances the network lifetime. Eleven performance parameters are analyzed using a trace file with awk scripts. DF is the core parameter of the network that affects the network's maintainability. It is used as a target variable. DF parameter range lies between 0.1 MB to 10 MB. The high value of the DF parameter increases the congestion in the network due to the limited bandwidth. The packet loss ratio is increased and leads to network failure. Similarly, the lower value of the DF takes more time to reach the destination. The optimized DF maintains the smooth functioning of the network. It is optimized using the MARS algorithm. The variation of the MARS model is used to optimize the DF parameter such as the basic model and the hyperparameter-tuned model at different degrees. Model selection with degree 2 shows the best results based on accuracy. It shows that there is very less error between actual and predicted DF. It provides good results as compared to existing literature. The best tuned MARS model provides distinct observations of repair time.  Repair time is the core metric of maintainability. It uses the normal distribution method to achieve maximum maintainability. The main novelty of this research work is: • Maintainability is evaluated through DF optimization using the MARS algorithm.
• Normal distribution method is used to estimate the maintainability in a quantified manner.

VI. CONCLUSION AND FUTURE DIRECTION
Quality assurance is a popular and mandatory buzzword for real-time applications. It is a challenging task to configure the network that enhances QoS metrics and improves the network performance. It depends on various metrics such as throughput, maintainability, reliability, latency, PDR and availability. This paper addresses the issue of network maintainability in WSNs. It can be estimated from three perspectives such as communication, external environment factors and hardware of sensor node. Here, the research work is emphasized communication. The presence of faults degraded the network's performance. So, there is a need for an efficient framework for the smooth functioning of the network that improves maintainability. MD-MARS framework is presented to enhance the network maintainability for communication. It includes three phases namely analysis of performance parameters, data flow optimization and maintainability evaluation. In the first phase, RIDSR topology is created using NS2 simulator. Based on simulation parameters, a dataset is generated. It comprises eleven performance parameters. These parameters are analyzed using trace files and awk scripts. The next phase deals with data flow optimization using MARS algorithm. It reduces collision and packet loss across the network. The different MARS models are applied to the dataset such as the basic model and tuned model. The hyperparameters are tuned using the Grid Search algorithm. The best-tuned model is used as the final model. The hyperparameters tuned at degree 2 behaves as the best tuned model to optimize data flow with an accuracy of 99.83 percent. MARS algorithm performs well as compared to other ML algorithms such as random forest, weka lazy model, ensemble model with equal weight and enhanced deep reinforcement learning. The total time to optimize data flow is known as repair time. The last phase deals with the repair time which is a core metric to evaluate maintainability. Repair time determines the MTTR, standard deviation, probability density function, maintainability and repair rate using normal distribution method. The maximum maintainability evaluated is 97.67 percent at a repair time of 21.17 milliseconds. MD-MARS framework approach shows better results in the comparison of existing approaches such as maintainability achieved using virtuality reality, analytic network process and super position degree using Bayesian, AHP and fuzzy. The proposed work has some limitations. It is limited to a small scale and provides the best results for a single parameter. The future direction of this research work is: • The more network parameters can be evaluated using the proposed framework such as availability and security.
• This work emphasizes a small scale which can be extended to a large scale using new protocols. DEEPALI GUPTA is currently working as a Professor of research with the Chitkara University Research and Innovation Network (CURIN), Chitkara University, Punjab, India. She specializes in software engineering, cloud computing, the IoT, and genetic algorithms. She has worked with undergraduate and postgraduate students and research scholars throughout her career and plans to continue to involve students in her research and eager to participate in projects and guide independent student's research. She has published more than 120 research papers in national and international journals and conferences. Based on these areas, she has guided many Ph.D. and M.E. scholars. She has worked at various administrative positions such as the principal, the head (CSE), the dean academics, an IBM (Spoc), a remote centre coordinator (IITB), a coordinator for IITB spoken tutorial, an executive committee member at the Computer SAPNA JUNEJA received the master's and Ph.D. degrees in computer science and engineering from M. D. University, Rohtak, in 2010 and 2018, respectively. She is currently a Professor with the Department of Computer Science, KIET Ghaziabad, India. She has more than 18 years of teaching experience. Her broad area of research is software reliability of embedded systems. Her areas of interests include software engineering, computer networks, operating systems, database management systems, and artificial intelligence. She has guided several research Thesis of UG and PG students in computer science and engineering. She is editing book on recent technological developments. She is having several papers in various international journals of repute and various patents as well. Her current area of research is block chain technology, machine learning, and artificial intelligence.

ACKNOWLEDGMENT
ALI NAUMAN received the M.Sc. degree in wireless communications from the Institute of Space Technology, Islamabad Pakistan, in 2016, and the Ph.D. degree in information and communication engineering from Yeungnam University, Republic of Korea, in 2022. He is currently working as an Assistant Professor with the Department of Information and Communication (ICE), Yeungnam University, Republic of Korea. He has contributed to five patents and authored/coauthored three book chapters and more than 20 technical articles in leading journals and peer-reviewed conferences. His research interests include artificial intelligence-enabled wireless networks for tactile healthcare, multimedia, and industry 5.0. His research interests also include resource allocation for 5G and beyond-5G (B5G) networks, device-to-device communication (D2D), the Internet of Everything (IoE), URLLC, tactile internet (TI), and artificial intelligence (AI).
AMENA MAHMOUD received the master's degree in virtual reality specification from the Computer Science Department, Helwan University, Egypt, and the Ph.D. degree in bioinformatics from the Computer Science Department, Mansoura University, Egypt. She is currently an Assistant Professor with the Department of Computer Science beside being the Vice Director of the Quality Assurance Center, Kafrelsheikh University, Egypt. She is currently a Researcher in computer science and interested in bioinformatics, machine learning, and other topics such as pattern recognition, image processing, and natural language processing.