Application of Deep Neural Networks in Multi-Hop Wireless Sensor Network (WSN) Channel Optimization

The proposed approach consists of three main components: (i) feature extraction using convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, (ii) reinforcement learning-based decision-making for adaptive routing and channel allocation, and (iii) an optimization mechanism for energy-efficient network performance. Figure 1 illustrates the overall workflow.

Figure 1.

Proposed deep learning-based channel and routing optimization framework. Feature extraction captures spatial-temporal variations, reinforcement learning optimizes policy decisions, and final routing and channel selection are dynamically adjusted.

To effectively capture spatial and temporal variations in network state, we employ a hybrid deep learning architecture combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. CNNs extract spatial dependencies among sensor nodes, while LSTMs model the temporal evolution of network conditions, enabling the system to adapt dynamically to changes in channel quality, interference, and traffic load.

At each time step, the network state is represented as a multi-channel tensor S_t ∈ ℝ^M×N×C, where: - M and N correspond to the spatial dimensions of the network topology. - C represents different input features, such as signal-to-noise ratio (SNR), residual node energy, interference level, and traffic congestion.

The CNN module applies multiple convolutional layers to extract spatial patterns from S_t. The output feature map is given by: Ft=σ(Wc∗St+bc) where: - W_c and b_c are the convolutional filter weights and biases, - ∗ denotes the convolution operation, - σ(·) is a nonlinear activation function (e.g., ReLU).

To enhance feature learning, batch normalization and max pooling operations are applied: Ft'=maxω×ωBatchNorm(Ft)

This ensures stable training and reduces computational complexity.

The extracted spatial features F_t′ are then processed by an LSTM network to capture temporal dependencies. The LSTM updates its hidden state according to: ht=tanh(WhFt'+Uhht−1+bh) where: - h_t is the hidden state at time t, - W_h, U_h, b_h are trainable parameters.

The LSTM outputs a context-aware embedding H_t, which encodes long-term dependencies in network conditions.

The final feature representation Z_t is obtained by concatenating CNN-extracted spatial features and LSTM-extracted temporal dependencies: Zt=Concat(Ft',Ht)

This hybrid feature embedding is then passed to the reinforcement learning agent for adaptive routing and channel allocation.

To dynamically optimize routing and channel allocation in multi-hop wireless sensor networks (WSNs), we formulate the problem as a reinforcement learning (RL) task. The RL agent continuously learns from the network environment and selects optimal transmission policies to maximize network efficiency while minimizing interference and energy consumption. The decision-making process is modeled as a Markov Decision Process (MDP), where the RL agent interacts with the environment through iterative exploration and exploitation.

The optimization problem is represented as a tuple (S, A, P, R, γ), where: - S is the state space, representing the current network conditions. - A is the action space, consisting of routing decisions and channel selection strategies. - P(s′ | s, a) is the state transition probability, defining the likelihood of moving to state s′ after taking action in state s. - R(s, a) is the reward function, quantifying the effectiveness of an action in optimizing network performance. - γ ∈ (0,1) is the discount factor, controlling the importance of future rewards.

The state s_t at time t captures the real-time network conditions, including: - Residual energy of sensor nodes. - Link quality metrics such as signal-to-noise ratio (SNR) and interference levels. - Network traffic load and congestion indicators. - Hop count and latency metrics.

The state representation is obtained from the CNN-LSTM feature extraction module, which provides a spatiotemporal embedding of the network topology.

At each time step, the RL agent selects an action a_t from the action space A, which includes: - Selecting the next-hop node for packet forwarding. - Assigning an appropriate communication channel to minimize interference.

The agent learns an adaptive routing and channel allocation policy, balancing data transmission efficiency and network longevity.

The agent’s objective is to maximize a cumulative reward function that considers throughput, energy efficiency, and latency: R(st,at)=αTt−βEt−γLt where: - T_t represents the network throughput at time t. - E_t denotes the total energy consumption of the network. - L_t is the end-to-end latency. - α, β, γ are weighting coefficients that balance different optimization objectives.

A penalty term is added if the selected action leads to excessive congestion or energy depletion in critical nodes: R(st,at)←R(st,at)−δCt where C_t represents congestion at time t and δ is a penalty coefficient.

The RL agent updates its policy using Q-learning, where the action-value function Q(s, a) is iteratively updated based on observed rewards: Q(st,at)←Q(st,at)+η(rt+λmaxa'Q(st+1,a')−Q(st,at)) where: - η is the learning rate. - λ is the discount factor. - max_a′Q(s_t+1, a′) represents the estimated future reward for the best possible action in the next state.

The policy π(a|s) is derived from the Q-values using an ϵ-greedy strategy, where the agent explores with probability ϵ and exploits the best-known action with probability 1 – ϵ.

To enhance learning stability and prevent catastrophic forgetting, experience replay is employed. A replay buffer stores past experiences (s_t, a_t, r_t, s_t+1), and mini-batches are sampled for training. The Q-learning update is applied to these batches, improving sample efficiency and convergence.

Figure 2 illustrates the reinforcement learning-based routing and channel allocation process.

Figure 2.

Reinforcement learning-based policy optimization process. The RL agent iteratively updates policies based on observed network states and reward feedback.

The reinforcement learning-based approach enables dynamic and adaptive optimization of routing and channel allocation in multi-hop WSNs. By leveraging Q-learning with experience replay, the model continuously improves its decision-making policy, leading to enhanced network throughput, energy efficiency, and reduced congestion. The next section presents experimental evaluations demonstrating the effectiveness of the proposed method.

The RL-based optimization updates policy parameters using the Q-learning algorithm: Q(st,at)←Q(st,at)+η(rt+λmaxa'Q(st+1,a')−Q(st,at)) where Q(s_t, a_t) is the action-value function, η is the learning rate, and λ is the discount factor.

The learned policy is periodically fine-tuned using experience replay and policy gradient updates.

The proposed method integrates deep learning with reinforcement learning for optimizing channel allocation and routing in multi-hop WSNs. CNN and LSTM networks extract spatial-temporal features from network states, while an RL agent dynamically learns optimal transmission policies. The approach balances throughput, energy efficiency, and latency through a carefully designed reward function and policy update mechanism. Experimental validation of this method will be presented in the following section.

To simulate a realistic multi-hop WSN, a network topology consisting of 100 sensor nodes randomly deployed over a 500m × 500m area is used. Each node has a limited transmission range of 50m, ensuring that multi-hop communication is necessary for data delivery to a central sink node. The simulation runs for 5000 time steps, during which packets are generated following a Poisson distribution. The data packets vary in size from 64 bytes to 512 bytes to simulate heterogeneous network traffic.

The proposed deep reinforcement learning (DNN-RL) model is implemented using a hybrid architecture combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks for feature extraction. The RL agent is trained using the Q-learning algorithm with an experience replay buffer of 10⁴ state-action pairs. The learning rate is set to η = 0.001, and the discount factor is λ = 0.9.

For comparison, two baseline methods are used: - Q-RL: A traditional reinforcement learning approach based on tabular Q-learning without deep feature extraction. - Heuristic Method (HM): A conventional routing and channel allocation strategy that selects the shortest path with the least congested channel.

Each experiment is repeated 10 times with different random seeds to ensure statistical reliability.

This experiment measures the network throughput and end-to-end latency under different traffic loads. Throughput is defined as the number of successfully delivered packets per second, while latency represents the average time required for a packet to reach the sink node.

Figure 3 illustrates the network throughput for increasing packet arrival rates. The proposed DNN-RL model consistently achieves higher throughput compared to the baseline methods. At moderate traffic loads (30 packets/sec), DNN-RL achieves a throughput of approximately 6.5 Mbps, which is 12% higher than Q-RL and 30% higher than HM. Under high traffic conditions (80 packets/sec), the performance gap widens as DNN-RL efficiently adapts to congestion by dynamically adjusting channel allocation.

Figure 3.

Comparison of network throughput for different methods. The proposed DNN-RL model maintains higher throughput under increasing network loads.

The results of Experiment 1 indicate that the proposed DNN-RL approach significantly improves network throughput and reduces latency compared to traditional methods. As shown in Figure 3, the throughput of the DNN-RL model remains consistently higher across different network loads. When the traffic load reaches 50 packets per second, DNN-RL achieves a throughput of approximately 9.1 Mbps, which is 11% higher than Q-RL (8.2 Mbps) and 25% higher than HM (7.3 Mbps). The performance gap becomes more pronounced under higher loads, where the adaptive channel selection of DNN-RL prevents congestion and maximizes available bandwidth.

Latency measurements further validate the efficiency of the proposed method. Under high traffic conditions, DNN-RL reduces the average packet delay by 20% compared to Q-RL and 35% compared to HM. This improvement is attributed to the model’s ability to dynamically reallocate channels and reroute packets in response to congestion hotspots. In contrast, the heuristic method follows a static shortest-path routing approach, leading to bottlenecks in high-load scenarios.

Overall, these results demonstrate that deep reinforcement learning effectively optimizes multi-hop WSN performance by adapting to real-time network conditions. By learning optimal transmission policies, the proposed approach maintains high throughput and minimizes latency, ensuring reliable data delivery in dynamic environments.

Energy efficiency is a critical factor in WSNs, as sensor nodes typically operate on limited battery power. This experiment evaluates the energy consumption per successfully delivered packet and estimates network lifetime under different optimization strategies. The goal is to determine how effectively the proposed DNN-RL method reduces energy usage compared to traditional approaches.

Table 1 presents the average energy consumption per packet and the estimated network lifetime for each method. The results indicate that the proposed DNN-RL model achieves the lowest energy consumption per packet, consuming only 0.12 mJ, which is 20% lower than Q-RL and 30% lower than the heuristic method (HM). The prolonged network lifetime observed in DNN-RL confirms its effectiveness in optimizing power-aware routing and channel allocation.

To visualize the energy consumption trends, Figure 4 illustrates the total energy depletion over time for different methods. The heuristic method (HM) exhibits the steepest decline, indicating rapid battery depletion. In contrast, DNN-RL maintains a more gradual energy depletion rate, allowing for a significantly longer operational period.

Figure 4.

Energy depletion over time for different optimization methods. The proposed DNN-RL approach maintains a more gradual depletion rate, ensuring extended network lifetime.

The experimental results demonstrate that the proposed deep reinforcement learning-based optimization significantly reduces energy consumption compared to traditional methods. The heuristic method (HM) leads to faster battery depletion due to its rigid routing decisions, which fail to adapt dynamically to network conditions. The Q-RL model achieves moderate improvements, but it lacks the deep feature extraction capabilities of DNN-RL, which enable more efficient power-aware routing and adaptive channel selection.

DNN-RL prolongs network lifetime by approximately 47% compared to HM and 10% compared to Q-RL. The gradual depletion rate in Figure 4 confirms that the reinforcement learning-based model effectively balances energy-efficient transmissions, optimizing power consumption across the network.

These findings highlight the importance of deep learning in enhancing the sustainability of WSN deployments. By leveraging real-time decision-making, the proposed model ensures that energy resources are utilized optimally, preventing premature node failures and extending network longevity.

This experiment tests the model’s robustness against dynamically changing interference and varying traffic conditions. External interference sources are introduced randomly, and the packet delivery ratio (PDR) is monitored.

Experiment 3 evaluates the resilience of different optimization methods against dynamic interference and fluctuating traffic loads. Figure 5 illustrates that the proposed DNN-RL approach maintains a significantly higher packet delivery ratio (PDR) under increasing interference levels compared to Q-RL and HM. At moderate interference levels, DNN-RL achieves a PDR of approximately 94%, whereas Q-RL and HM degrade to 90% and 85%, respectively. As interference intensifies, the performance gap widens, with HM dropping below 65% PDR at the highest interference levels, while DNN-RL still maintains over 80% delivery reliability.

Figure 5.

Packet delivery ratio (PDR) under varying interference levels. The proposed method maintains a higher PDR, demonstrating robustness against interference.

These findings suggest that deep reinforcement learning enhances the robustness of WSNs by dynamically adapting routing and channel selection policies. The traditional heuristic approach suffers significant performance degradation in highly dynamic environments because it lacks adaptability, leading to excessive packet loss and inefficient channel utilization. In contrast, the DNN-RL model continuously learns from changing network conditions, allowing it to mitigate the impact of interference and maintain stable network performance.

The results confirm that the proposed method is well-suited for real-world WSN deployments where environmental conditions and traffic loads are unpredictable. By leveraging real-time decision-making, DNN-RL ensures reliable communication even in challenging network conditions, making it a promising solution for mission-critical applications such as industrial monitoring and emergency response systems.