Energy-Efficient Edge Intelligence in IoT Environment Using Cross-Layer Bio-Inspired Optimization with Deep Learning Framework

The rapid evolution of the IoT has fundamentally changed data driven applications in many industries, including agriculture, healthcare, smart cities, and industry 4.0. In fact, given the potential scale and future of IoT a key enabling technology for Internet of Things is the Wireless Sensor Networks (WSNs) which enables a network of collaborating sensor nodes to collect, process, and transmit data. The majority of sensor nodes are energy-constrained and unfortunately do not have frequently replacing batteries. Therefore, routing in WSNs along with effective task management are indispensable for networks to maximize their lifespan while maintaining quality of service (QoS) especially across large-scale IoT. To address the energy consumption and enabling effective routing, bio-inspired and evolutionary algorithms are intriguing options to consider for energy optimization and adaptive routing. For instance, bio-inspired and evolutionary strategies [1] are based on natural phenomena including swarm intelligence, genetic evolution, and predator-prey dynamics can be simple yet effective. Another example of a bio-inspired hybrid optimization algorithm showed prominent results for clustering, load balancing, and energy efficiency in WSNs [2]. Similarly, bio-inspired neural networks demonstrated abilities for optimal cluster head selection that resulted in efficient energy distributions and improved scalability within a low power wireless personal area network [3]. Moth-Flame Optimization (MFO) and variants are also one of many metaheuristics and provided a great deal of flexibility in dynamic environments by maintaining a controlled balance between exploration and exploitation of the optimization space [4].

Few researchers have found that metaheuristic clustering protocols outperform the heuristic-based clustering schemes in terms of energy consumption and packet delivery performance [5]. The papers strongly suggest that evolutionary computing should be included when quantifying the performance of WSNs. As routing optimization matured, started investigating the application of edge and fog computing to enhance efficiency in IoT ecosystems. Deep learning is a promising approach for real-time analytics that can be performed using IoT data is an extremely energy intensive computing technology. Recent efforts to benchmark the energy efficiency of deep learning across models and edge devices continue to emerge [6, 7]. These articles outline the trade-offs between accuracy, latency, and energy efficiency. The need for adaptive frameworks for both computation and communication overhead. Task offloading is another important domain where energy efficiency has implications for IoT performance. Recent advances in reinforcement learning (RL) and deep reinforcement learning (DRL) frameworks can optimize the migration of computations among IoT devices, fog nodes, and cloud infrastructure. The use of Deep Q-Networks (DQNs) demonstrated promising results for latency-aware, energy-efficient offloading strategies. Here, the computational load is balanced while minimizing delay [8]. Fuzzy logic enhanced DRL approaches demonstrate further adaptability for IoT application within a fog-cloud three-tier architecture [9]. Earlier works also recognize the critical aspects of an energy-scheduling while identifying that issues to ensure Quality of Service [10].

Despite the progress in the literature, there remain significant gaps. A systematic review of energy-aware resource management techniques in fog-enabled IoT highlights the absence of a unified framework that accounts for communication, computation, and routing optimization at the same time [11]. To make further progress, we also find that many of the existing techniques are domain or application-specific, or static, and research in these areas neglects the unpredictable nature of IoTs. There are current constraints that require an efficient evolutionary computing framework that combines multi-objective energy-aware routing and adaptive cluster head selection with intelligent task offloading. For an efficient evolutionary computing framework, bio-inspired algorithms should be utilized for exploration, reinforcement learning for variability, and multi-objective optimization to balance energy efficiency, latency, and reliability.

This paper addresses the above challenges by contributing the following:

• An innovative evolutionary computing Bio-Inspired Deep Learning and Edge–Cloud (BIO-DLEC) framework for routing and energy-aware clustering in IoT-based WSNs.
• Identifying an optimized route and task offloading selection using bio-inspired optimization and reinforcement learning.
• A multi-objective fitness function including residual energy, latency, link reliability, and cost for offloading.
• Detailed simulations and comparisons with clustering, routing and offloading protocols.

The structure of the remainder of the paper is detailed as follows: In Section 2, we review related literature in bio-inspired optimization, deep learning compression for edge devices and fog-cloud offloading. In Section 3, we describe the proposed BIO-DLEC framework along with the mathematical model and optimization process. In Section 4, we present the experimental protocol and details. In Section 5, we conclude with a summary and future research.

This paper introduces Bio-Inspired Deep Learning and Edge–Cloud synergy (BIO-DLEC) a novel all-encompassing scheme expected to lower the energy dissipation, alleviate the inference delay, and optimize the lifetime of IoT WSNs. Our holistic framework incorporates the following three synergistic components.

• Hybrid Whale–Grey Wolf Optimization Algorithm (WOA–GWO) for clustering providing equal energy distribution and adjustable selection of cluster heads (CHs),
• Bio-inspired routing mechanism allocating energy-aware and congestion-free nodes for transmission,
• Deep learning enabled edge cloud collaboration delivering adaptive inference at the edge respective locations and uploading complex tasks to the cloud incrementally to improve latency and network congestion.

The proposed BIO-DLEC framework designed in three phases as given below:

• WSN clustering and routing
• Energy-aware DL compression performed on edge nodes
• Adaptive fog–cloud offloading

Prior work has been treated each of the layers independently from an optimization perspective. Here, BIO-DLEC has the advantage that it considers type of topology, model size, and placement of computation. These will co-evolve with the energy and traffic dynamics. The overall design architecture is shown in Figure 1 including all stages.

Figure 1. Design diagram of the proposed model

We model a system with $N$ battery-powered nodes $\left\{n_1, \ldots \ldots, n_N\right\}$ deployed over an area A , with a fixed sink at $P_s$. Each node $n_i$ has residual energy at some decision epoch $t$. Eq. (1) identifies the group of sensor nodes occupying the IoTenabled WSN. Each node, $n_i$ represents a single sensor with its own position coordinates, available energy and computational capacity, while $N$ is the total number of nodes. This notation will be the base of the network model, because all equations that follow (clustering, routing, and energy optimization) fundamentally rely on the usefulness of the nodes and the interactions of the nodes. In the case of the BIODLEC model we will consider that the sensor nodes are heterogeneous, where they are able to sense, and provide limited edge intelligence functions.

Nodes $=\left\{n_1, n_2, \ldots, n_N\right\}$     (1)

As shown in Eq. (2), the Euclidean distance between two nodes, $n_i$ and $n_j$, is defined by their coordinates, $\left(x_i, y_i\right)$ and $\left(x_j, y_j\right)$. Distance is an important component of a communication cost metric, since energy dissipation is a function of transmission distance. By effectively incorporating distance into the WOA–GWO hybrid optimization routine, BIO-DLEC creates energy efficient clusters and identifies transmission paths by ensuring that smaller distance transmissions prevail over longer distance scenarios, keeping energy expenditure to a minimum.

$d_{i j}=\sqrt{\left(x_i-x_j\right)^2+\left(y_i-y_j\right)^2}$     (2)

Eq. (3) defines the energy required to transmit a data packet. The first term, $E_{\text {elec}} \cdot k$, captures the energy cost of the electronic circuit ($k$ in bits) operating on a packet and the second term, $E_{\text {amp}} . k . d_n$, captures the additional cost introduced by the power amplifier, transmitting the packet a distance $d$. In this case, $n$ is the path-loss exponent (commonly 2 in free space or 4 in multi-path). In the case of BIO-DLEC, the energy consumed shapes cluster-head selection and affects routing because reducing distance while requiring equal use of amplifiers will maximize the life of the network.

$E_{t x}(k, d)=E_{\text {elec}} \cdot k+\epsilon_{\text {amp }} \cdot k \cdot d_n$      (3)

Eq. (4) provides the energy consumed to receive a data packet of size k bits. Unlike transmission, which makes use of a power amplifier, the energy consumed on reception is restricted to what is consumed by electronic circuitry. This is especially relevant for CHs which are collecting information from more than one-member node. The reception model in BIO-DLEC also ensures equitable cluster-head rotation through the use of WOA-GWO optimization, thus eliminating nodes from draining energy faster than the others.

$E_{r x}(k)=E_{\text {elec }} \cdot k$      (4)

Eq. (5) outlines the remaining energy of node $n_i$ at time $t+1$. The amount of energy available to the node at time $t$, designated as $E_i(t)$ is reduced by the transmission energy $\left(E_{t x}\right)$, the reception energy $\left(E_{r x}\right)$, and the energy consumed by deep learning computations. This energy consumes a portion of the energy costing by the nodes running inference at the edge. It is an important part of ensuring there is balance between communication costs and computation costs. The framework makes adaptation decisions based on communication (sending) a task locally or utilizing the fogcloud layers to process the task, while also ensuring there is sustainability sooner not later on battery capacity, while latency is kept to a minimum even at peak times.

$E_i(t+1)=E_i(t)-E_{t x}-E_{r x}-E_{D L}$      (5)

3.1 Clustering and routing with hybrid WOA–GWO

The model looks at a hybrid optimization model which combines the WOA) with the GWO, to concurrently solve clustering and routing in an IoT supported WSN. The target of the model is to use less energy, maximize the life of the network, and improve routing efficiency, so large IoT installations are more feasible over the long term. Our model stands alone from past models that looked at clustering and routing as separate processes by examining them as integrated optimization problems simultaneously. The value in our model, is partitioning the network logically into two layers, whilst the clustering layer can look at selecting the best CH, and the routing layer solved the least energy path from CH to sink node. Together these two layers of optimization worked towards finding a balance across communications load across the network whilst minimizing redundancy and maximizing the available residual energy reserves consumed during communication. We formed clusters based on minimizing the amount of radio energy consumed whilst providing nodes that are facing large inference loads some refuge from high radio loads. We assign ranks to candidate CHs using the BIO-DLEC fitness as given in Eq. (6).

$F_i=\alpha E_i(t) E \max +\beta Q_i-\frac{\gamma E_{D L, i}}{E_{D L, \max}}$      (6)

Let $Q_i$ denote the link quality, while $\alpha, \beta, \gamma>0$ represent the weights assigned to energy, connectivity, and DL cost, respectively. Eq. (10) is imposed as a constraint on the cluster geometry in order to limit the cluster radius, denoted as $R_c$, and the number of members in the cluster, denoted by $N_c$. Its purpose is to ensure bounded intra-cluster hop lengths for load balancing. Eq. (3) gives the transmission energy of a k-bit payload when the distance d is $E_{t x}(k, d)$, while Eq. (4) is its reception energy $E_{r x}(k)$. With DL energy, these describe node-level consumption in Eq. (5), which will be fed back into the optimizer every epoch. Fusing WOA with GWO avoids premature convergence and retains diversity. The encircling/exploitation step in Eq. (7) of WOA fine-tunes around the current best $X^*(t)$, while the leadership average step in Eq. (8) of GWO favors exploration by utilizing $\alpha$, $\beta$, and $\delta$ leaders. Hybrid update is shown in Eq. (9).

$X(t+1)=X^*(t)-A \cdot\left|C \cdot X^*(t)-X(t)\right|$      (7)

$X(t+1)=\frac{X_\alpha+X_\beta+X_\delta}{3}$      (8)

$X_{n e w}=\lambda X_{W O A}+(1-\lambda) X_{G W O}, \lambda \in[0,1]$      (9)

It strikes a balance between both linearly anneal $\lambda$ in relation to the epochs spend in the subsequent exploratory and exploitative phases. Each candidate solution additionally encodes each CH index, member assignments, and Tx-power schedules. The metric from our fitness function given in Eq. (6) finds and returns the fitness of each candidate and the best candidate will be pushed to the network for the next epoch.

$R_c \leq R_{max}$ and $N_c \leq N_{max} $       (10)

Once advanced clusters are declared, members send unicast packets to their CHs. CHs relay down to the sink. The paths can be shortened based on $Q_i$ and the cost of routing. The schedules and Tx powers will need to be inherited from the selected candidate given in formula (10). When: (i) the mean residual energy in any cluster node sinks below a threshold, (ii) link quality degrades below a threshold, (iii) DL compression or offloading changes the energy profile of each node, so the optimizer will rerun.

3.2 Energy-aware deep learning compression

The deployed backbone with different layers has a few parameters at precision with uncompressed rate. BIO-DLEC performed structured pruning on each layer, followed by post-training quantization to INT8, with a resulting compressed size. The storage size of the deep learning model operating in the IoT-WSN structure is defined in Eq. (11). Here, $W_l$ is the number of weights in layer B stands for the bit-width precision, and L is the number of layers. The summation takes into consideration all the layers of the model, and therefore gives the total size and storage requirements of the deep learning model. This is important in BIO-DLEC because model size affects memory, transmission overhead from offloading, and device limitations for edge nodes with constrained resources.

$S_{\text {model }}=\sum_{l=1}^L\left(W_l \cdot B\right)$      (11)

In Eq. (12), the total active number of weights in a layer $l$ after applying dropout regularization is presented. The dropout rate of this layer is given by $p_l$, whereas $W_l$ is the previous number of weights within that layer. Dropout reduces the number of active weights, which also improves generalization performance, reduces computational burden of real-time inference. In BIO-DLEC, this equation captures the trade-off between accuracy and efficiency, since dropout prevents overfitting but also lower number of operations triggered at the sensor or fog layer, which in turn, should reduce dissipated energy.

$h^{(l)}=f\left(W^{(l)}\left(r^{(l-1)} \odot h^{(l-1)}\right)+b^{(l)}\right)$     (12)

Here, $h^{(l)}$ is the output activation, $h^{(l-1)}$ is the input activation. The effective weights $W^l$, and $b^{(l)}$ bias vector. In BIO-DLEC, this equation quantifies the fact that the compression of the model directly results in reduced memory footprint, reduction of the transmission overhead for offloading, and reduced energy (cost at edge devices) for the particular inference tasks under consideration while ensuring acceptable accuracy measurements.

$S_{\text {comp }}=\sum_{l=1}^L W_l^{\prime} \cdot B^{\prime}$      (13)

Eq. (14) indicates the energy consumed during deep learning inference. In this expression, $M A C_s$ refers to the total number of multiply accumulate (MAC) operations necessary for the compressed model, $P_{o p}$ is the energy used per MAC operation, and $\phi$ is a conversion factor for efficiency. This demonstrates that inference energy does not simply depend on a measurement of algorithmic complexity, but includes the operational characteristics of the implementation. The model detailed in BIO-DLEC, produces inference energy $E_{D L}$ that is tracked in each node's residual energy balance as represented in Eq. (5), and allows the framework to determine when local inference is no longer efficient from an energy perspective and adaptive offloading can be accomplished.

$E_{D L}=\phi \cdot M A C s \cdot P_{o p}$     (14)

Here, $P_{o p}$ is energy per multiply-accumulate on the target device and $\phi$ reflects any memory/activation effects measured offline. The aggressiveness of reducing prune channels adapts to runtime conditions given in Eq. (15).

$P_l(t)=f\left(\frac{E_i(t)}{E_{\max }}, \frac{\tau_{\text {req }}}{\tau_{\text {obs }}}\right)$      (15)

To decrease the amount of pruning where there is little residual energy or if the observed latency $\tau_{\text {obs }}$ exceeds the requirement $\tau_{r e q}$. Quantization by default is INT8, and if energy reaches a critical low, the agent may activate additional channel-sparsity constraints to further decrease the number of MACs. The alteration of accuracy is bounded via inequality (16), with $A c c_{\text {min}}$ for each application. If this continual adjustment is violated, then pruning can be rolled back and offloading preferred.

$A c c_{c o m p} \geq A c c_{min} $      (16)

Since $E_{D L}$ is appeared directly in the CH fitness given in Eq. (6), nodes with heavier models are considered de-preferred for a CH role-unless compensated for by higher reservoir energy, or more favorable links. On the other hand, as pruning represents a reduction in $E_{D L}$, the node fitness increases, leading to improved equity in CH rotation without impacting lifetime.

3.3 Adaptive offloading

By combining these four reality dimensions, the state vector $S_t$ provides a complete overview of the system state, allowing the BIO-DLEC optimization strategy (hybrid WOA-GWO with DL-based inference) to dynamically adapt the clustering, routing, and offloading policies. In this regard, Eq. (17) modifies the BIO-DLEC in a context-aware and dynamic manner to make sure that the decisions made in the real-time IoT-WSN environment are also dynamic. Eq. (17) defines the state vector for the proposed BIO-DLEC framework at a given time step $t$. This multi-dimensional state consists of the relevant parameters when making decisions in clustering the nodes, routing packets, and offloading tasks. Here, $E_i(t)$ is the residual energy of node $i$ at time $t$, as previously defined in Eq. (5). Including this term makes sure that decisions are energyaware to avoid placing excessive burden on nodes with an insufficient level of energy. $L_{i j}(t)$ is the link quality measure between nodes $i$ and $j$ that captures the distance (via Eq. (2)) and the dynamic wireless channel condition. A strong link would ensure improved PDR and reduced retransmission costs. $fog(f)$ is the fog node utilization at time $t$, representing the current processing load at the fog layer. Including this metric allows for maintaining low latency and a balanced distribution of computation by ensuring that task offloading wouldn't overload fog servers. The state vector is given in Eq. (17),

$S_t=\left[E_i(t), L_{i j}(t), U_{\text {fog }}(t), \rho_{\text {traffic}}(t)\right]$     (17)

Eq. (18) describes the action space of the BIO-DLEC decision-making. Each time step t, the system can take one of three actions, (i) Local execution at the sensor node, (ii) Offloading to the fog layer, or (iii) Offloading to the cloud. This discrete action space allows for an adaptive inference deployment policy for the BIO-DLEC decision-making, while balancing energy consumption, latency and accuracy. The flexibility provided by separate execution domains means that a resource-constrained sensor node does not become a limiting factor or bottleneck, while also providing performance scaling through fog and cloud execution domains.

$A_t \in\{$Local,Fog,Cloud$\}$     (18)

The normalized multi-objective formulation preserves interpretability and consistency in relative priorities $\alpha: \beta: \gamma$ across experiments and promotes stability in convergence when heterogeneous tasks (prediction, anomaly detection, routing) are integrated. Eq. (19) will be displayed as the corrected multi-objective loss function.

$Q(\Theta)=\frac{\alpha}{\alpha+\beta+\gamma} L_{\text {pred }}+\frac{\beta}{\alpha+\beta+\gamma} L_{\text {anom }}+\frac{\gamma}{\alpha+\beta+\gamma} L_{\text {route }}+\lambda\|\Theta\|_2^2$      (19)

Here, $L_{\text {pred}}$ is fill-level prediction loss, $L_{\text {anom}}$ is the anomaly detection or waste-type classification loss, $L_{\text {route}}$ is the routing optimization cost represented as a differentiable surrogate of route distance or energy consumption, along with $\lambda\|\Theta\|_2^2$ as an L2 regularization term to avoid overfitting and stabilize weights updates.

Eq. (20) represents the reward function for reinforcement learning. The reward at time $t$ is a weighted sum of three factors, which are: (i) the inference Accuracy, weighted by $\delta$; (ii) the Energy consumed, penalized by $\eta$; and (iii) Latency, penalized by $\kappa$. This represents the goals of the system design - maximizing accuracy while minimizing energy dissipation and delay. Tuning the coefficients $\delta$, $\eta$, and $\kappa$ allows the framework to prioritize application-specific requirements, e.g., energy-constrained sensing vs. latency-critical industrial monitoring, etc.

$R_t=\delta \cdot$ Accuracy $-\eta \cdot$ Energy Used $-\kappa \cdot$ Latency      (20)

Eq. (21) represents the total latency incurred by performing an offload task. It consists of three components: (i) transmission delay, which is effective by size of input $S_{\text {input}}$ and transmission bandwidth factor $B_{i j}$ between nodes $i$ and $j$; (ii) execution delay $\tau_{\text {exec}}$ by the fog or clouds server; and (iii) return delay $\tau_{\text {return}}$ to receive the results back to the requester. This ensures that BIO-DLEC directly measures if an offload produces a net latency decrease versus local execution.

$\tau_{\text {offload}}=B_{i j} S_{\text {input}}+\tau_{\text {exec}}+\tau_{\text {return}}$     (21)

Eq. (22) describes the local inference delay when tasks are executed on the IoT device itself. Here, MACs represent the number of multiply-accumulate operations that the corresponding compressed deep learning model will take, and $f_{\text {CPU}}$ represents the CPU frequency of this IoT device. This formulation relates computational complexity to execution time so that the system can compare $\tau_{\text {local}}$ with $\tau_{\text {offload}}$ and determine the best execution option.

$\tau_{\text {local}}=\frac{M A C s}{f_{C P U}}$      (22)

In the BIO-DLEC model, the accuracy-based decision rule as defined by inequality (23) is as follows. In the equation, $A c c_{\text {comp}}$ is the inference accuracy achieved by the compressed model executed locally (using the compressed model from Eqs. (11)-(14), while $A c c_{\text {mintext}}$ is the minimum accuracy specified by the accuracy requirements of the application. If the locally compressed model using compressed dynamics model does not reach that sufficient threshold of accuracy, the preference would be offloading the task associated with the decision support system requested to a fog or cloud server with less stringent computations and restrictions. This would be particularly important in instances where the application is health or safety monitoring system and the lightweight local inference might not be good enough for proper decision making.

$A c c_{\text {compressed}}<A c c_{\text {mintext}} \rightarrow$ prefer $\frac{f o g}{\text {cloud}}$      (23)

Inequality (24) includes a fog resource constraint. In this case $U_{f o g}$ indicates the current utilization of the fog server at time $t$ and $U_m$ is the maximum utilization threshold rate. If the utilization of the fog server exceeds this threshold, then the BIO-DLEC framework may automatically disallow offloading to fog nodes and select either local execution (if in the consideration set mode) or offloading to the cloud. This protects from overflowing the fog server and therefore ensure service is acceptable with importance to ensuring capacity to maintain low latency and fairness across multiple users.

$U_{f o g}>U_{max} \rightarrow$ avoid fog      (24)

3.4 Cross-layer optimization

BIO-DLEC is designed with a cross-layer optimization approach that will have a combined optimization methodology that recognizes the systems physical, network and application layers to maximize computing energy efficiency and sustainability while being deployed in IoT-WSN environments. Ultimately unlike a conventional design that optimizes layers in isolation, as we do in BIO-DLEC, which optimizes all three layers in unison while considering the system as a whole through the objective function. The weighting coefficients $\omega_1$, $\omega_2$, and $\omega_3$ are tunable for the system designer to consider the requirements of the applications. The overall objective is the multi-criteria program as given in Eq. (25):

$\min _{A_t, p_l} \omega_1 E_{\text {total}}+\omega_2 \tau_{\text {total}}-\omega_3$Accuracy      (25)

Here, $E_{\text {total}}$ and $\tau_{\text {total}}$ are cumulative sensing, inference, and communication components over the epoch, and $\omega_1, \omega_2, \omega_3>0$, are application priorities e.g., safety critical latency or lifetime. Constraints include cluster geometry, fog utilization, and accuracy floor. In practice, BIO-DLEC does so indirectly as a hybrid metaheuristic for clustering/scheduling, an adaptive compression controller for $p_l$, and the Q-learning agent for $A_t$; they each share telemetry each epoch.