Energy-Efficient IoT Network Routing Model Based on Multi-Layer Clustering and Deep Learning

Internet of Things (IoT) is the current era of smart things connected with each other for specific purpose of communication. The rise of IoT has given rise to the deployment of large networks of heterogeneous, resource constrained devices. These devices, which are commonly used to support applications including smart cities, industrial automation, and environmental monitoring, are typically limited by power, and thus energy efficiency is a key issue. The design of efficient routing protocols is critical to guarantee long network life time and consistent data forwarding in such energy-constrained environments [1].

These conventional routing schemes like Low-Energy Adaptive Clustering Hierarchy (LEACH) and HEED [2] use clustering approach as a strategy for data aggregation at cluster heads to minimize power usage of sensor nodes when data is to be sent to a BS. Despite the fact that such protocols can lead to some improvements due to their energy efficiency, they lack in dynamic IoT environments when dealing with a changing topologies and traffic patterns. Furthermore, single-layered clustering schemes figure out that a tight distribution of the energy load throughout the network may not be achieved, and we will rather experience premature node death and an eventual degradation in network performance.

Some clustering-based protocols have been presented for extending network lifetime and improving energy efficiency. Among these, LEACH and Hybrid Energy-Efficient Distributed Clustering (HEED) [2] are well-known. LEACH adopts random rotation of cluster head to balance energy consumption, while HEED takes account of residual energy and communication cost in cluster head election. Although they are effective in static environment or under small-scaling considerations, they showed some limitations.

In LEACH and HEED, only localized decision-making, and heuristic-based methods are employed and they could not adjust themselves based on historic or network wide context. The optimal clusters are prone to computationally expensive and less efficient in the case of nodes’ density growth. Dynamic topologies, varying energy levels and data traffic conditions in practice way off the performance of operation as compared to typical IoT deployments. These protocols work in a flat or single-cluster hierarchy causing excessive inter-cluster communication overhead and fast death of nodes near sink nodes [3].

The recent progress in the field of machine learning [4], in particular deep learning (DL), has provided and is still providing ways to take advantage of more sophisticated mechanisms to improve the routing decisions in IoT networks. Deep learning models (e.g., CNNs and GRUs) have demonstrated the ability of learning spatial and temporal patterns in network data, which can facilitate more intelligent and adaptive routing strategies. CNN [5] can capture spatial features, while the GRU can model temporal dependence, so the combination is well in line for the dynamic IoT environment.

In this paper, we develop a hierarchical clustering and deep learning based hybrid routing scheme for making efficient routing decisions in IoT networks. In the multi-layer clustering, layering the network into hierarchical clusters can balance the energy load and improve the scalability. The CNN-GRU model uses effective parameters (residual energy, node distance, link quality, and traffic load) to predict the optimal routing paths to react to dynamic IoT networks. Nowadays, with the fast growth of Internet of Things (IoT) the need for intelligent energy-aware routing and clustering algorithms is more and more pressing. Conventional routing schemes could not cope with the dynamics and heterogeneity of IoT networks and leave nodes with battery exhaustion and ineffective communication. In this regard higher level machine learning and deep learning techniques are used for provision of self-adaptive, context ware and energy-efficient routing and clustering.

In STM environments, GNNs in particular have shown interesting potential for clustering in Social IoT context as they enable a context-driven node clustering based on social and mobility behaviors rather than pure geographic vicinity. This leads to a considerable gain in routing efficiency and scalability in dense IoT settings [6]. At the same time, the combination of Blockchain and deep learning models has been investigated to realize secure and tamperproof routing decisions. Apart this hybrid technique minimizes the energy utilization and provide confidence and transparency of multi-hop communication paths in WSN [7]. In the context of deep learning for unsupervised feature learning, multi-layer neural networks (e.g., deep learning networks such as what has been recently employed in the DeepCluE model) have demonstrated that deep architectures can be effectively applied to improve the results of clustering on high-dimensional data. These models leverage the ensemble learning scenario to smooth the prediction among different layers, which shed light on clustering techniques in IoT [8].

Additionally, for the joint optimization of energy demand and computation overhead in mobile edge IoT, hybrid clustering algorithm-based models combining with the CNN referred to as the K-Means have been introduced [9]. These methods utilize CNNs for spatial feature learning and take advantage of clustering to ease the complexity of routing. Likewise, energy-efficient computation and clustering methodologies specifically designed for heterogeneous smart city deployment were proposed based on multi-metric optimization. Based on these models, node capabilities, energy status, and data type to enhance network resilience and lifetime as a whole is introduced [10]. Taken together, these developments highlight a common thread: the intertwining of deep learning, intelligent device co-location and energy-economical computation to mitigate the urgent scalability and sustainability concerns of today’s IoT. Based on this, our proposed work adopts a multi-level clustering approach combined with a CNN-GRU deep learning model to improve routing efficiency, network lifetime, and computational latency across various IoT environments. Research contributions are:

• Multi-layer Clustering Architecture Design: We propose a multi-level clustering scheme, where sensor nodes are divided into several layers to achieve efficient data collection and transmission by balancing energy consumption within the network.
• CNN-GRU Routing Model: We design a CNN-GRU-based model to extract spatial features by using CNNs followed by temporal features utilizing Gated Recurrent Units, which assists in making intelligent energy-aware routing decisions.
• Full Performance Discussion: We show by simulation with NS-3 and for synthetic IoT traces that we achieve better performance in terms of energy consumption, packet delivery ratio and end-to-end delay than conventional routing protocols.

The rest of the paper is organized as follows: In Section 2, we review the related work on energy-efficient routing and DNN applications in IoT. The system model and problem formulation and the proposed approach is detailed in Section 3, which includes the multi-layer clustering structure and the CNN-GRU model. The experimental setup and results are provided in Section 4. Finally, the conclusion and future work is presented in Section 5.

The goal of the proposed model is to improve energy efficiency, latency, and packet loss in IoT enabled WSN, by using multi-layer clustering and hybrid deep learning model (CNN-GRU). The structure is intended to be dynamic to cope, as the nodes can move, and the energy levels and the density of network nodes can change. Using CNNs for spatial pattern extraction and GRUs to learn temporal sequences, the model enables smart routing decisions based on live network states.

3.1 Network assumptions

In this paper, several major assumptions are considered to formulate the IoT-based WSN capable of implementing the proposed multi-layer clustering and CNN-GRU routing mechanism. The network is considered to be composed of a number of NNN sensor nodes which are randomly deployed in a two-dimensional fixed area. These nodes generate a static or semi-static topology so that the capability of moving the nodes is restricted and almost null. A BS (sink node) located at the centre or periphery of the network is responsible for collecting data from the sensor nodes; It is also assumed that the BS has access to unlimited energy and computing resources.

All sensor nodes are considered in identical level of hardware and initial energy. Every node is powered by a battery, low-power radios, onboard computation (microcontroller), and environmental sensors (e.g., temperature). The energy initially distributed uniformly is depleted at different rates due to different roles the nodes play. Furthermore, all nodes are assumed to be location-aware through GPS modules or with the aid of localization algorithms for optimized clustering and routing decisions. The communication model among the nodes is symmetric wireless communication model, meaning that if A can reach B, then B can reach A. Correspondingly, we use classical first order radio model for the radio energy model in which energy is proportional to the size of the data packet and square of transmission distance. The communication is considered to be without noise, and with the absence of collisions and MAC-level interference, so that the emphasis is on routing and energy efficiency.

Regarding network operation, nodes are periodically grouped into clusters based on criteria like remaining energy and communication range. A Cluster Head (CH) dynamically elected for each cluster has the role of evenly distributing the level of energy consumption in the network. In the multi-layer clustering approach, existing networks are partitioned into a secondary (i.e., formation) hierarchy and sub-CHs aid the primary CH to collect and send data, to save energy and reduce the delay in transferring data. These CHs also perform the CNN-GRU hybrid model that captures the spatio-temporal features of the network state to make smart and adaptive routing.

The sensing operation is assumed to be periodic, where each node sends packets of certain size at regular time instants. Data is transmitted for the first time to the CH, and is aggregated there and sent to the base station. The environmental uncertainties such as signal fading, node power-off and weather interferences are not modeled in this phase for simplicities. The choice of the model architecture is obviously extendable, and can incorporate this aspect in the future.

3.2 Multi-layer clustering algorithm

The Multi-Layer Clustering Algorithm is a state-of-the-art hierarchical data aggregation and routing protocol that aims to improve energy efficiency, scalability, and network longevity in IoT and WSNs. Contrary to the flat clustering approaches, where nodes constitute a single layer of clusters with direct communication to the sink/base station, multi-layer clustering arranges nodes in several layers of clusters. Each layer plays a specific role in the transmission of data, the distribution of energy and the cluster management, where they are organized in a tree-based or pyramid-like topology.

Sensor nodes in the bottom layer are organized as primary cluster according to nearest or signal strength. From each of these clusters, a Cluster Head (CH) is elected, typically based on factors like residual energy, node centrality and historical load. The CH is the role to gather data from the participating nodes in its cluster, aggregate or process the data and forward the aggregated information to the higher echelons. The aggregation of data leads to less redundant information, which reduces communication overhead and the cost of energy.

The intermediate layers are formed by secondary/tertiary CHs acting as aggregators for multiple lower-layer CHs. These mid-level CHs are dynamically chosen or pre-assigned according to the network size and topology. They are responsible for routing decisions, more data aggregation and packet relays to the base station. This hierarchical architecture not only decreases the transmission distance of a single node (to save its energy) but also can balance the load and solve the hotspot problem that the nearer node to the base station would consume energy more quickly.

At the highest layer, a number of high capacity nodes or a centralized sink node are in charge of gathering and processing the ultimate aggregate data. This layer could also contain AI models or cloud servers for decision making, anomaly detection, or analytics. Let N is the total number of nodes, A is deployment area, L is the number of cluster layers, Ci is the number of clusters in the layer I, and CH is the cluster head. For all in N, node assignment to cluster assign to cᵢ, is defined as Eq. (1).

$d\left(n_k, C H_1^j\right)=\min \left(d\left(n_k, C H_1^j\right)\right)$          (1)

The cluster head selection is as given in Eq. (2).

$C H_{ {Score }\left(n_k\right)}=\alpha *\left(\frac{E_{ {res }}\left(n_k\right)}{E_{max }}\right)+\beta *\left(1-\left(\frac{ { Distance }\left(n_k, B S\right)}{D_{max }}\right)\right)$              (2)

The Energy consumption model is considered as Eq. (3) for transmitting the k bits and Eq. (4) for the receiving k bits.

$\begin{gathered}E_{t x(k, d)}= \left\{\begin{array}{c}\text { If } d<d^0, \quad k * E_{\text {elec }}+k * \varepsilon_{f s} * d^2 \\ \text { Else, } k * E_{\text {elec }}+k * \varepsilon_{m p} * d^4\end{array}\right.\end{gathered}$                (3)

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

Here, $E_{ {elec }}$ shows the energy to run the transmitter/receiver circuit, $\varepsilon_{f s}$ is free space model amplifier constant, $\varepsilon_{m p}$ is multipath model amplifier constant, and d⁰ is $\sqrt{\left(\frac{\varepsilon_{f s}}{\varepsilon_{m p}}\right)}$.

The total energy the consumption in each layer is estimated as Eq. (5), where is the energy dissipated during the data collection.

$E_{ {layer }_{\mathrm{i}}}=\Sigma\left[E_{ {tx }_{\left(k, { distance }_{ {to }_{C H}}\right)}}+E_{r x(k)}+E_{{agg }}\right]$            (5)

Total delay occurs at each layer is estimated as Eq. (6), where,  $T_{t x_{\mathrm{i}}}$ is the transmission delay at i^th layer, $T_{ {queue }_{{i}}}$ is the queuing delay, and $T_{ {process }_{\mathrm{i}}}$ data fusion delay.

$T_{ {total }}=\Sigma\left\lceil T_{t x_{\mathrm{i}}}+T_{ {queue }_{\mathrm{i}}}+T_{ {process }_{\mathrm{i}}}\right\rceil$               (6)

The energy optimization function is given in the Eq. (7) subjected to where residual energy is greater than the threshold and cluster head communication is less than the average distance in the network.

$E_{\text {Obj }}=\left(\frac{E_{ {total }}}{N}\right)+\gamma * T_{ {total }}$                (7)

3.3 CNN-GRU hybrid model

The CNN-GRU hybrid model [18] is developed for the purpose of intelligent and energy-efficient routing in IoT-based WSNs. In this model, we get the best of both worlds by employing the ability of CNNs to extract spatial features and the power of GRUs to recognize temporal patterns. The model is deployed at the CH level and the aggregation node data is processed to achieve the optimal next hop for forwarding the data with the ultimate aim to minimize the energy, latency, and packet loss.

The CNN module is used to learn spatial feature correlations in the temporal dimension. The first layer is a 1D convolutional layer with 32 filters with (3,1) kernel size, plus activation function of ReLU. This structure is beneficial to capture temporal patterns locally in time gto improve the generalization and stability of the model, we utilize batch normalization and dropout with a rate of (0.2). Then a max pooling layer having a pool size of (2,1) is employed to reduce the temporal dimensionality and to obtain dominant features. The output would then be a small feature map (4,32) that capture the filtered spatial features over time.

The feature maps are then flattened after the CNN block and passed into GRU to capture long-term temporal dependencies and dynamic behavior of nodes. The GRU layer has 64 hidden units and tanh as its activation function. Dropout (0.3) with recurrent dropout (0.2) are used for regularization. Unlike standard LSTMs, GRUs are computationally more effective in controlling gating operations, e.g., in real-time routing decisions in sparse IoT networks.

The CNN module performs the duty of learning the spatial features from the input, which contains the following node attributes: residual energy, distance with neighbors, received signal strength (RSSI), queue length, and hops to the base station. They are represented as structure input vectors and processed using 1D CNN layers. The convolutional filters will discover important patterns and correlations among these metrics, e.g. low residual energy and high RSSI value, which is an indication that the node should be depleted and not to be chosen as a relay. Such spatial representation is an intermediate result that encodes raw input features into a high-level representation, which would be used as an input of the temporal modeling. let the input vectorat time t is given as Eq. (8).

$X_t=\left[E_t, L_t, {RSSI}_t, \Delta E_t, D_t\right]$               (8)

Here, $E_t$ is the current residual energy, L_t is the coordinates of the location, RSSL_t is the received signal strength indication, $\Delta E_t$ is energy change rate, Dt is distance between base station and the cluster head. The input is passed through a CNN layer for local feature extraction as given in Eq. (9).

$Z_t={ReLU}\left(W_c * X_t+b_c\right)$              (9)

Here, W_c weights of convolutional filters, b_c is the bias value, and rectified linear unit activation function is used here as ReLU. After the CNN stage, the feature maps are input into the GRU. GRUs are a variant of Recurrent Neural Network (RNN) which is well-suited to sequential data and provide a computationally cheaper alternative to Long Short-Term Memory (LSTM) networks. The GRU receives time-series information from each node, which represents how network metrics evolve over time. For example, it can discover that there is a continuous decrease in energy of a node while being selected as a CH god or that packet delay increases at certain times. GRUs updating and resetting gates can selectively remember or forget part of the input, to automatically remove less informative information or help to ignore noise and preserve useful information, which is indispensable for the network to remember the former states effectively without losing the ability to forget it as shown in Equations from (10) to (14).

$h_t=G R U\left(Z_t, h_{t-1}\right)$            (10)

$Z_t={sigmoid}\left(W_z * Z_t+U_z * h_{t-1}+b_z\right)$              （11）

$r_t={sigmoid}\left(W_r * Z_r+U_r * h_{t-1}+b_r\right)$        （12）

$\tilde{h}_t=tanh \left(W_h * Z_h+U_h *\left(r_h \circ h_{t-1}\right)+b_h\right)$             (13)

$h_t=\left(1-z_h\right) \circ h_{t-1}+z_h \circ \tilde{h}_t$            (14)

Here, W and U are learnable weights, $Z_t$ and $r_t$ are update and reset gates, and h_t is the hidden state. A fully connected layer and softmax classifier receive the output final hidden state of the GRU, and score potential next-hop candidates according to predicted reliability, transmission cost, and a global utility. The node having highest score is chosen for data forwarding. This rational decision making guarantees that routing is not only adaptive but (long-term) optimal for network sustainability as shown in Eq. (15).

$Y_t={Softmax}\left(W_o * h_t+b_o\right)$                (15)

Here, Y_t is the routing decision to the next hop node or the cluster, and  W_O, b_o are the weights and bias of output layer. The model also trained with loss function given as Eq. (16).

$L_{ {total }}=L_{ {class }}+\lambda * L_{ {energy }}$             (16)

Here, $L_{ {class }}$ is the categorical class entropy for routing decisions, $L_{ {energy }}$ is the penalty for choosing high energy dissipation paths, and $\lambda$ is the regularization factor. The objective of the paper is to minimize the network energy consumption while maximizing the data delivery accuracy as shown in Eq. (17).

${Minimize}(E)=\left(\frac{E_{ {total }}}{N}\right)+\alpha * L_{ {total }}+\beta * T_{ {total }}$                (17)

Here, total energy consumed is $E_{ {total }}$, N number of nodes, $T_{ {total }}$ is total network delay, $\alpha, \beta$ are weigh balancing factors are considered to achieve the objective. It also brings some benefits of combining CNN and GRU for WSN routing in IoT. It is a light-weight and effective scheme based on the consideration of the spatial and temporal trends on the behavior of a node. This hybrid strategy improves the system's capacity to lie congested and battery drained paths, and increases overall network performance and prolongs network life. Furthermore, the model enables online processing such as in smart environments where timely adaptive routing strategies are needed.