Appended Virtual Resource Migration Scheme for Application-Specific IoT Backhauled Wireless Sensor Networks

In the Internet of Things (IoT)-based Wireless Sensor Network (WSN) applications, resource virtualization provides a virtual representation of the physical resources so that multiple applications can run on the same resource without interference [1]. The users profit from the optimum exploitation of resources, reducing the costs for the physical resource owner [2]. Virtualization makes real-time changes of parameters possible, ensuring scalability and flexibility, while abstraction simplifies the deployment and management of complex [3] IoT systems. It will ensure that the reconfiguration of virtual resources en-roots fault handling and load balancing efficiently [4]. Virtualization, hence, ensures that the fundamental principle regarding scalability and flexibility is met. The framework allows load balancing efficiently since the virtual resources can easily be reconfigured [5]. Virtualization also provides energy-efficient operations since it allows energy to be fed only when needed. It thus makes efficient resource management in IoT WSNs feasible [6].

Resource migration is a process of adjusting resources to the changing demands of the user. Processing power, storage, and bandwidth are adjusted correspondingly for optimum performance [7]. Resources are migrated from one node to another to ensure efficiency within the network. Effective management minimizes downtime and loss of data, hence facilitating the delivery of services uninterrupted [8]. The network dynamically responds to varied loads that vary, thus avoiding bottlenecks. Predictive algorithms initiate resource transfers to avoid congestion [9]. The distribution of workload through load balancing across the network avoids overloads on certain nodes. The process allows the network to respond in a dynamic manner due to varying loads [10]. Predictive algorithms of resource needs are helpful in proactively initiating transfers. It ensures resources are at the right place at the right time. Resource migration also facilitates load balancing efficiently across a network [11].

Machine learning (ML) can enhance resource virtualization and migration by analyzing past data to project future network conditions. Such a type of optimization in resource allocation determines the trends of patterns in traffic and resource use and proactively readjusts resources [12, 13]. Fault detection is improved, and it initiates resource migration for the maintenance of stability. It achieves energy efficiency by predicting low-activity times and downsizing resource use to prolong battery life in the case of sensor nodes [14]. Adaptive learning provides optimization based on past experiences in the network. The trends of traffic and resource utilization have been identified through machine learning [15]. Such inference readjusts the resources. Predictive ability in that respect ensures proactive management. That will ensure that resources are available where needed and when required. Machine learning also ensures that the system is energy efficient. These are some of the intelligent methods that enable efficient and adaptive resource management in IoT WSNs [16, 17]. The contributions of the article are:

To discuss different works related to virtualization, resource allocation, utilization, etc., for IoT-WSN and individual networks, with their methodology

To propose a novel demand-appended virtual expansion scheme using federated learning (FL) to improve the decisions on node/ resource expansion and migration, considering the connectivity and application demand factors

To obtain the performance results of the proposed scheme using application response, virtualization rate, migration ratio, response time, and resource utilization metrics

To verify the above-metric performance by providing a comparative analysis with the selected methods disclosed in Section 2

The article’s contribution is: Section 2 discusses the related works, followed by the proposed scheme and self-analysis in Section 3. Section 4 presents the experimental scenario and metric-based comparative analysis. In Section 5, the conclusion, limitations, and future scope of the proposal are presented.

The proposed scheme is designed to increase the duration of application support based on user and resource connectivity inputs. The connected users and resource services and network processing are required from the IoT-based wireless sensor nodes (WSN), i.e., the elements of a WSN, namely the processing unit, sensing unit, power unit, and communication unit that are observed in random time intervals. The proposed scheme is presented in Figure 1.

Figure 1. Proposed Demand-appended Virtual Expansion and Migration (DVEM) scheme

The actual goal of this proposed scheme is to reduce the battery drain, power drops, power consumption, network routing, network speed, transmission media, network topology, scalability, and fault tolerance in analyzing individual resource constraints of detected applications. The challenging task is the virtualization and migration of discrete resource constraint-detected applications with the previous knowledge of resource constraints. The application requirements are stored as records from the previously identified instances.

3.1 Users and resources connectivity analysis

The appropriate demands of the applications are identified through wearable sensor nodes over the IoT platform. The observations from WSNs are classified as individual and groups of users and resources based on their nodes' connectivity. In this model, the battery drain is detected as an instance of wrong observations of node availability, node death, and node energy level.

Initial Modeling of user and resource connectivity, the variable $U_c(\mathbb{t})$ and $R_c(\mathbb{t})$ signifies the sequence of users and resources connectivity detected in a given time interval, such that the application demands $A D(\mathbb{t})$ is given as:

$A D(\mathbb{t})=U_c(\mathbb{t})-\mathbb{E} * R_c(\mathbb{t})$             (1)

$\arg \min _{\mathbb{t}} \sum \mathbb{E}(\mathbb{t}) \forall U_c(\mathbb{t})+R_c(\mathbb{t})$             (2)

In the above Eqs. (1) and (2), $\mathbb{E}$ is the error-causing factor, and the objective of reducing the impacting factors using the constraint $\forall U_c(\mathbb{t})+R_c(\mathbb{t}) \in A D(\mathbb{t})$ is defined. The connectivity is observed from two instances, such as users ( $U_c$ ) and resources ( $R_c$ ) . Therefore, the total connectivity is expressed as $\mathrm{C}=U_c+R_c$ such that the resource connectivity is detected between the user connectivity. Where $\mathbb{N}$ symbolizes the number of nodes in a particular network, then $U_c=(\mathbb{N} \times C)-R_c$ is a single-node observation that is classified cluster node. Eqs. (1) and (2), the application demand at any interval $\mathbb{t}$ is computed, for which the objective of error minimization is set. The error confining condition in Eq. (2) considers the user connectivity and their demand for resources. This computation relies on precise requests and responses between the service providers and users. For a balanced response, the target of $\mathbb{E}$ is to be satisfied as defined above. Let $\operatorname{APS}\left(U_c\right)$ and $\operatorname{APS}\left(R_c\right)$ represent the application support/ response for all connected users and resources observed from $\mathbb{N}$ in different time intervals, and $\mathbb{E}$ are detected in all $R_c$ that is formulated as:

$\operatorname{APS}\left(U_c\right)=\mathbb{N} * U_c: U_c(\mathbb{t})+R_c(\mathbb{t}), \forall \mathbb{E}=0$            (3)

And,

$\operatorname{APS}\left(R_c\right)=\frac{R_c}{\mathbb{N}}: \mathbb{E}-U_c(\mathbb{t})+R_c(\mathbb{t}), \forall \mathbb{E} \neq 0$            (4)

As per Eqs. (3) and (4), the application responses observed from both connectivity are represented as $\left(\mathbb{N} \times U_c\right)$ and $\left(\frac{R_c}{\mathbb{N}}\right)$ are mapped to identify the application demands based on resource constraints. Now, based on the application support as in the above equation, Eq. (1) is rewritten as:

$\begin{gathered}A D(\mathbb{t})=\operatorname{APS}\left(U_c\right)-\operatorname{APS}\left(R_c\right)=\left[\mathbb{N} * U_c: U_c(\mathbb{t})+\right. \\ \left.R_c(\mathbb{t})\right]-\left[\frac{R_c}{N}: \mathbb{E}-U_c(\mathbb{t})+R_c(\mathbb{t})\right]\end{gathered}$            (5)

For the above-expanded application demands, the continuous $\operatorname{APS}\left(U_c\right) \in \mathrm{C}$ is validated on facing the first resource connectivity as in Eq. (5). This validation is performed to identify influencing factors based on virtual expression to the migration process using the FL process. Eq. (5) considers the user request and response-based demand update such that the influence of $\mathbb{E}$ through retained response or failed requests is identified. If the identification is true, then the imbalance between $U_c$ and $R_c$ requires additional migration/ virtualization support. The application demand modeling is illustrated in Figure 2.

Figure 2. Application demand modeling before virtualization

The initial modeling of the $A D(\mathbb{t})$ is defined using the maximum request and response in $\mathbb{t}$. Being a conventional, pervasive access-based platform, IoT provides $U_c(\mathbb{t})$ satisfaction with $R_c(\mathbb{t})$. However, the $\mathbb{N}$ is the deciding factor for $A D(\mathbb{t})$ such that $A P S\left(U_c\right) \vdots \equiv A P S\left(R_c\right)$ under $R_c(\mathbb{t})$ allocation. This $\mathbb{E}$ is defined (identified) by the energydrain/node unavailability between successive /random $\mathbb{t}$ if $A D(\mathbb{t})$ is true. Therefore, the initial demand model relies on the $C$ check for the $\mathbb{N}$ that defines $\operatorname{APS}\left(U_c\right)$ and $\operatorname{APS}\left(R_c\right)$ such that $\mathbb{E}$ is true/false (Refer to Figure 2).

The virtual expression to migration is performed using the available users and resources through FL. For this process, the sequence of node connectivity for all the resources $\mathbb{N} \in R_c$ is expressed as:

$\mathbb{N}\left(R_c\right)=\left(1-\frac{U_c}{\mathrm{~N}}\right) U_{c_{i-1}}+\frac{R_c}{\mathrm{~N}} \sum_{i=1}^{\mathrm{t}}\left(1-\frac{R_c}{\mathrm{~N}}\right)^{i-1} R_{c_{i-1}}$           (6)

In Eq. (6), $U_{c_{i-1}}$ is the previous knowledge of the connected user's service information and $t_s$ is the previous data of connected resource services and their responses using WSNs in the IoT environment. Therefore, based on the services of connected nodes, $\operatorname{AD}(\mathbb{t})=\operatorname{APS}\left(U_c\right)-\operatorname{APS}\left(R_c\right)\left[1-\mathbb{N}\left(R_c\right)\right]$ is the outcome without influencing factors. The node connectivity plays a vital role in defining the resource availability between users and service providers. Based on migration/ overload, the connectivity feature varies for which $\mathbb{I}$ assessment is made. Besides the connectivity problem over the resource allocation is validated to confine the error in response. The virtual expression $\left(V E \times P_{U_c}\right)$ and $\left(V E \times P_{R_c}\right)$ for connected users and resources in the first stage is expressed as:

$\begin{gathered}V E x P_{U_c}=\sum_{i \in \mathbb{t}}\left[\mathbb{N}\left(R_c\right) \times U_c(\mathbb{t})+R_c(\mathbb{t})\right]_i- \\ \operatorname{APS}\left(U_c\right) . \mathbb{t}\end{gathered}$             (7)

$\begin{gathered}\operatorname{VExP}_{R_c}=\sum_{i \in \mathbb{t}} \mathbb{N}\left(R_c\right)_i\left\{\left[1-\mathbb{N}\left(U_c\right)_i\right] \times\right. \\ \left.\operatorname{APS}\left(U_c\right)\right\}_i-\operatorname{APS}\left(R_c\right) \cdot \mathbb{t}\end{gathered}$             (8)

Based on Eqs. (7) and (8), the virtual expression is based on the application support observed from the IoT system for the consecutive process. In this initial proposed modeling process, the computation of $V E \times P_{U_c}, V E \times P_{R_c}, \operatorname{APS}\left(U_c\right)$ and $A P S\left(R_c\right)$ are the serving inputs for the FL process. The consecutive support from the application helps to easily identify the error occurrence by checking node availability, node inactivity, and node energy level between the connected nodes. The need for a virtual expansion decision is presented in Figure 3.

Figure 3. Virtual expansion decision process

There are two possible cases for expanding the virtualization process: $\left(V E \times P_{U_C}\right)$ and $\left(V E \times P_{R_C}\right)$. The decisions on these possibilities are verified under connectivity, resource, and demand cases. If $\mathrm{C}=$ true, then the allocation is performed at $\mathbb{t}$ provided APS is true in increasing the allocation sequences. If the response tallies $A D(\mathbb{t})$  modeled, then $V E \times P_{R_c}$ is the expansion demand. In the case of resource condition, if $A P S$ is false, then $V E \times P_{U_c}$ is the virtual expansion required (Figure 3). This FL process is discussed in the section below.

3.2 FL process for virtualization

In the virtualization process, FL is used for identifying the correctness of battery energy, node availability, and node energy, and detecting energy-drained nodes. In this model, the number of nodes and instances may vary based on the service requests and responses. For instance, the FL process performs two types of modeling, namely sequence migration and battery drain identification. In the migration process, the no-node availability, no-node energy, and the occurrence of dead nodes are identified using a virtual expression. In this proposed model, the connectivity inputs for node migration are based on the connected users and resources in different time intervals. The resource constraint of $U_c(\mathbb{t})+R_c(\mathbb{t}) \in C$ is processed based on their application demands and the occurrence of migration.

In the grouping process, the user and resource connectivity and the time instances are differentiated independently by using the virtualization concept. The virtualization process is performed for both connectivity, after which the IoT system FL process is used to migrate the current node. The virtualization output sequence $\mathbb{V}_{\mathbb{t}}$ is expressed using Eqs. (9)-(10) as:

$\mathbb{V}_{\mathbb{t}}=\mathbb{N}\left(U_c\right)_{\mathbb{t}}-\mathbb{N}\left(R_c\right)_{\mathbb{t}}-\operatorname{APS}\left(U_c\right)_{\mathbb{t}-1}$           (9)

$\mathbb{M}_{\mathbb{t}}=\mathbb{N}\left(R_c\right)_{\mathbb{t}}+\operatorname{APS}\left(U_c\right)_{\mathbb{t}-1}-\operatorname{APS}\left(R_c\right)_{\mathbb{t}-2}$           (10)

The virtualization processes generate two outputs, namely appendable and removable or taken away from the instances. In Eq. (10), $\mathbb{M}$ signifies the migration sequence that is represented as ($\mathbb{M}_1$ to $\mathbb{M}_{\mathbb{t}}$). Here, the grouping is pursued through virtual expression based on the impacting factors present in WSNs. The resource constraint of $\mathbb{t} \in \mathbb{V}$ must not be equal to $\mathbb{t} \in \mathbb{M}$ is the best decision for appropriate grouping. If the occurrence of resource connectivity is the initial service, then migration is performed using application support. In this model, the associated virtual expression and migration are dependable such that $U_c=\left\{U_c \cup \operatorname{APS}\left(R_c\right)\right\}$ and $R_c=\left\{\mathbb{V}_{\mathbb{t}} \cap \operatorname{APS}\left(R_c\right)\right\}$ is grouped independently to improve the IoT system. The FL process for virtualization is illustrated in Figure 4. In Figure 4, illustrated above, the resource and usercase validations using FL are presented. In both cases, $\mathbb{t}$ is the $A D$ mapping time common to $U_c$ and $R_c$. Under both instances of the verification is a common form $\mathbb{M}_1$ to $\mathbb{M}_t$ such that $C$ decides if it is to be virtualized between the user and IOT or IOT and resources. In the $\mathbb{N}\left(R_c\right)_{\mathbb{t}}$, to $\left[\mathbb{V}_{\mathbb{t}} \cap\right. \left.\operatorname{APS}\left(R_c\right)\right]$ occurs between $\mathbb{t}$ and $(\mathbb{t}-1)$ whereas $\left[U_c \cup\right. \operatorname{APS}\left(R_c\right)$] occurs between ($\mathbb{t}-1$) and ($\mathbb{t}-2$). These sequence reductions follow the continuous allocation across various $\mathbb{t}$. The common sequences are free from $\mathbb{V}_{\mathbb{t}}$ whereas the remaining are virtualized.

Figure 4. Virtualization process using federated learning (FL)

The common sequences are free from $\mathbb{V}_{\mathbb{t}}$ whereas the remaining are virtualized. Therefore, from the virtualized sequence, the training is iterated to ensure FL decisions satisfy the constraints in $\left(V E x P_{U_c}\right)$ and $\left(V E \times P_{R_c}\right)$. The FL system is distributed to account for two different sets of inputs ($\mathbb{t}, R_c, U_c$) and ($A D, R_c, \mathbb{t}$). The $R_c$ and t are the connecting factors to decide if $U_c$ is required or $A D$ is detected. If $U_c \geq A D$, then new requests are input, failing which requires a migration. Depending on the outputs in $(\mathbb{t}+1)$ and its continuous intervals, the resource allocation or user demand (pending) is computed. In the allocation case, the demand acceptance for processing is high, whereas in the pending case, the migration recommendations are. Therefore, the training is performed for resource and user cases separately using $\mathbb{V}_t$ observed under classified Eqs. (8) and (9) respectively. This update is made in all the $(\mathbb{t}+2)$ interval to ensure $\mathbb{I}$ is less. The aggregation is performed from the WS node inputs, such as connectivity, energy level, and response provided. The change in any of the above requires a training update to enhance the accuracy of computations. The migration of the initial allocation is performed, from which the nodes are alone grouped for further processing. After the dependable process, the application support of a single user is compared with the set of users based on $\operatorname{APS}\left(R_c\right)$. In this process, the $\operatorname{APS}\left(R_c\right)$ and the node migration serves as the input for the training set in different $\mathbb{t}$ instances that are classified under resource constraints. First, the virtual expression is represented using a non-linear matrix representation, where the $i$ and $j$ in the matrix symbolize the row and column vectors based on the resource and user connectivity. If $R_c$ and $U_c(\mathbb{t})+R_c(\mathbb{t})$ represents the training set for migration, i.e., if $V E \times P_{U_c}< V E \times P_{R_c}$, then the matrix 1 value is 1 . Instead, if $V E \times P_{U_c}> V E \times P_{R_C}$ is true, then the inverted values are used to represent the inverted matrix. From the matrix representation, the new sequence of node availability is further distinguished and grouped under the virtualization concept, where $\mathbb{t} \in U_c$. In this model, the node's dead occurrence due to high power consumption, battery drain, and other reasons is identified.

Based on the above matrix representation, the resource constraints of $V E \times P_{U_C}>V E \times P_{R_C}$ is marked as " 1 " whereas $V E \times P_{U_C}<V E \times P_{R_C}$ is marked as "0" for the addon process. If the presence of 1 is detected in any particular node, then the migration is recommended through the FL process. Hence, the classification of connected resources from connected users initiates the virtualization and migration process as in Eqs. (9) and (10). Now, the resource constraints based on their application demands are verified using the FL process and the nodes from the initial allocation. The FL process of user and resource connectivity validation occurs in the proposed model for reducing battery drain. In this process, the application support is decided for the dependable process based on the above matrix representation.

3.3 Migration process

In the migration process, the above-mentioned training set $\mathbb{T}$ serves as the input for the appendable process with the extracted resource connectivity required for accurate migration. In the initial retrieval process, the migrated WSNs are segregated for correlation verification. In this verification, if there is any battery drain or power drops occur, then $V E \times P_{U_c}$ is computed as in Eqs. (7) and (8). Instead, for $V E \times P_{R_C}$, the resource constraint sequence is updated, and the nodes are migrated. Here, the FL process and migration are different from the IoT system, as expressed in:

$\mathbb{V}_{\mathbb{t}}=\mathbb{T} \times\left(U_c(\mathbb{t})+R_c(\mathbb{t})\right)_i+\operatorname{APS}\left(R_c\right)_{\mathbb{t}-1}+\mathbb{E}(\mathbb{t})$            (11)

And,

$\mathbb{V}_{\mathbb{t}}=\operatorname{APS}\left(U_c\right)_{\mathbb{t}-1}+\mathbb{T} \times R_{c_{\mathbb{t}}}$            (12)

The above Eqs. (11) and (12) formulate the occurrence of resource connectivity in different time intervals, and the intermediate occurrence of user connectivity in resource connectivity observation leads to high EC and battery drain. In this model, the $U_c$ in $R_c$ satisfies $\mathbb{V}_{\mathrm{t}}=0$ is the previous solution for individual users, where the migration for $U_c$ leads to zero. The equations presented in Eqs. (11) and (12) are different from those of Eq. (9) with the accumulation of resource connectivity and resource constraint factors. Depending on $U_c$ and $R_c$, the imbalance modifies the $\mathbb{V}_t$ aided by the $F L$ process. In the process, $\mathbb{V}_t \cap A P S\left(R_c\right)$ and $U_c \cup A P S\left(R_c\right)$ are the two computations to categorize $\mathbb{V}_t$ for resource allocation. Hence, this influencing factor is not considered in the virtualization process. In both the connectivity processing, the error-causing factors are increased before the node is migrated, as:

$\begin{gathered}\mathbb{M}_{\mathbb{t}}=\sum_{i=1}^{\mathrm{t}} \mathbb{T}\left(U_c(\mathbb{t})\right)+R_c(\mathbb{t})_{i-1}-\left[\frac{R_c}{\mathbb{N}}: \mathbb{E}-\right.\left.\operatorname{APS}\left(R_c\right)_{\mathbb{t}-1}+\operatorname{APS}\left(R_c\right)_{\mathbb{t}-2}\right]\end{gathered}$             (13)

Figure 5. Migration process illustration

The migration process is illustrated in Figure 5 above for $\mathbb{V}_{\mathbb{t}} \in \mathbb{E}_{\mathbb{t}}$ and $\mathbb{V}_{\mathbb{t}} \in \mathbb{T}$ cases. The first case concerns the varying user demands that overflow the available resources. The alternating case concerns the varying $\mathbb{N}$ state in various $\mathbb{t}$. Using the available sequences, the FL classification is used to satisfy $\mathbb{V}_1=0$ for both $U_c$ and $R_c$ under ( $V E \times \left.P_{R_c}\right)$ and $\left(V E \times P_{U_c}\right)$ respectively. Therefore, the reverse migration in $\mathbb{M}_{\mathbb{t}-2}$ is performed if $A D(\mathbb{t}) \leq$ user demand and $\operatorname{APS}\left(U_c\right): \equiv \operatorname{APS}\left(R_c\right)$ is true. Contrarily, if the case fails C , then $\mathbb{M}_{\mathbb{t}}$ is reversed until $\mathbb{N}$ (new) is available. This case is active until the user is served with a new $\mathbb{N}$ between IoT and $R_c$. In this scheme, the battery drain is addressed in either user or resource connectivity and grouped under connected users, and resources are migrated in the previous allocations. The WSNs' influencing factors identification is given as follows:

$\begin{gathered}\mathbb{E}=\mathbb{N} \times \mathbb{T}-A P S\left(R_c\right) *\left(V E x P_{R_c}\right)+\mathbb{M}_{\mathbb{t}}(\mathbb{T})- \\ \left(\frac{\mathbb{V}_{\mathbb{t}}}{\mathbb{N}}\right) \mathbb{C}+A P S\left(U_c\right)_{\mathbb{t}}+A P S\left(R_c\right)_{\mathbb{t}}\end{gathered}$              (14)

Eq. (14) computes the occurrence of battery drain, high EC, and other influencing factors in WSNs. This identification of impacting factors is used to process the following requests and responses with the probability of migration in either user or resource connectivity. In this model, if $\mathbb{E}=0$ then the neighbor node is visited for further resource allocation. Similarly, if the condition of $V E \times P_{U_c}>V E \times P_{R_c}$ is satisfied by the particular node, then $\mathbb{E}=0$. Another $\mathbb{E}$ based assessment is defined in Eq. (14), considering the $\mathbb{N}$ and $\operatorname{APS}\left(U_c\right)$. In the redefining phase of $\mathbb{M}(\mathbb{T})$, the change in $\mathbb{V}_t$ identifies its impact on performance degradation. This equation is provided to verify the objective satisfaction throughout the virtualization and migration processes. Resource virtualization is performed in $\mathbb{V}_t$ is categorized as presented in Eqs. (8) and (9). The $U_c$ and $R_c$ demands finalize the migration based on the connectivity. This refers to the twofold decisions of $\mathbb{V}_t$ and $R_c$ based allocations such that the $F L$ identifies the synchronous $\Uparrow$ for $A D$ and $U_c$. In this case, the migration for reducing $\mathbb{E}$ is the target based on which virtualization (increasing/ decreasing) is mode. The FL decisions are made for opting for virtualization/ migration based on the response generated and resource allocations. The FL's training update maximizes the chances of $\mathbb{E}=0$ that eventually balances the migration and virtualization functions. Therefore, the migration is not performed, which means the virtual expression of allocated resources is observed for all node data. This observation of resource connectivity is valid until it satisfies the user's application demands. Instead, the occurrence of influencing factors continuously migrates both connected users and resources until it is stabilized either for users or resources.

3.4 Self-analysis

Based on the discussions above, a few self-analysis process is discussed below. In the analysis presented in Figure 6, the $\mathbb{E}$ from $U_c$ and $\mathbb{N}$ are considered.

In the proposed scheme, FL identifies the three possibilities for $\mathbb{E}$ under C. This includes $\mathrm{C}=0$ (least), $\mathrm{C}=1$ (high), and $0<\mathrm{C}<1$ that replicates the $\mathbb{E}$. The $\mathbb{E}$ is accounted for from $R_c$ and $U_c$ under distinguishable $\mathbb{t}$. The FL separates $\left[\mathbb{V}_{\mathbb{t}} \cap \operatorname{APS}\left(R_c\right)\right]$ and $\left[U_c \cup \operatorname{APS}\left(R_c\right)\right]$ to identify and migrate $\mathbb{E}$ in both ends. This process is repeated until $A D(\mathbb{t})$ is satisfied by the applications to ensure that maximum resource allocations are performed. Besides, the C variations are precisely identified to decide if the augmentation of $\mathbb{M}_{\mathbb{t}}$ is mandatory in satisfying $A D(\mathbb{t})$ (Figure 6). Following this $\operatorname{APS}\left(U_c\right)$ and $\operatorname{APS}\left(R_c\right)$ for the same time intervals is analyzed in Figure 7.

The $\operatorname{APS}\left(U_c\right)$ and $\operatorname{APS}\left(R_c\right)$ analysis presented in the above Figure 7 is required to ensure $\mathbb{N}$ allocation/reallocation along with $\mathbb{t}$. This ensures maximum resource utilization in any $\mathbb{t}$ for which $R_c$ satisfies $A D(\mathbb{t})$. Besides, the $U_c$ demands are met in $\mathbb{t}$ that does not exceed to $(\mathbb{t}+1)$ interval. The FL addresses the problem of $\mathbb{E}$ by identifying $V E \times P_{U_c}>V E \times P_{R_C}$ such that the mapping is optimal. If the IoT platform provides an optimal allocation through $\mathbb{V}_{\mathbb{t}}$ then $\mathbb{M}_{\mathbb{t}}$ is also concurrently satisfying the appended $A D(\mathbb{t})$. Thus, the processes are expected to be concurrent as aligned by the virtualization and migrations recommended by the FL. In the following analysis, $V E \times P_{U_c}$ and $V E \times P_{R_c}$ are considered in the same $\mathbb{t}$ as in Figure 7.

Figure 6. $\mathbb{E}$ from $U_c$ and $\mathbb{N}$

Figure 7. $\operatorname{APS}\left(U_c\right)$ and $\operatorname{APS}\left(R_c\right)$ analysis

The analysis presented in Figure 8 above showcases the $V E \times P_{R_c}$ and $V E \times P_{U_c}$ rates for different $\mathbb{t}$. As $\mathbb{t}$ increases, the appendable $\mathbb{V}_{\mathbb{t}}$ and $\mathbb{M}_{\mathbb{t}}$ varies with new demands and resource availability. The optimal case is the $\mathbb{M}_t$ that ensures high resource utilization, and suppresses $C$. In the proposed scheme, the FL is independent of $R_c$ and $U_c$ processes such that the successive resource allocations in $\mathbb{t}$, $(\mathbb{t}-1)$, and $(\mathbb{t}-2)$ are prominent in deciding the appendability of $\mathbb{V}_{\mathbb{t}}$ and $\mathbb{M}_{\mathbb{t}}$. Therefore, the successive verifications of C and distinguishable $\mathbb{E}$ are useful in maximizing utilization. The p -value assessment for $\mathbb{M}$ chance under $\mathbb{E}=0$ and $\mathbb{E} \neq 0$ cases is presented in Figure 9.

Figure 8. Analysis of $V E \times P_{R_c}$ and $V E \times P_{v_c}$

Figure 9. p-value assessment

The p -value assessment for $\mathbb{M}$ chances under $\mathbb{E}=0$ and $\mathbb{E} \neq 0$ is considered. If the objective of $\mathbb{E}$ minimization fails, the chances of migration and resource virtualization are high. Therefore, considering two consecutive intervals for $\mathbb{t}=0$ to $\mathbb{t}=0.2$, the increasing and decreasing demands are considered. The rest of the interval is conventional to satisfy $U_c=R_c$ condition. The FL is also trained using this case for multiple intervals, such that $\mathbb{V}_t$ (classified) defines the case of APS ($R_c$) for the maximum request demands. This reduces the chances of migration satisfying the objective of $\mathbb{E}=0$ (Figure 9).