Research on OTA Task Scheduling and Adaptive Fault-Tolerance Algorithm Under Cloud-Edge Collaboration

1. Introduction

Over-the-air (OTA) task scheduling plays a critical role in cloud–edge collaborative systems, particularly in industrial environments where stringent constraints on latency, energy consumption, and workload balance must be satisfied simultaneously [1,2]. In such systems, OTA tasks—including software updates, configuration synchronization and control-message dissemination—are often executed across heterogeneous edge nodes with limited resources and variable network conditions [3]. Inefficient scheduling can lead to excessive delays, uneven resource utilization, or cascading failures, directly degrading system reliability and operational performance in industrial applications. Recent research has highlighted the growing importance of cloud-native and cross-domain OTA architectures that can adapt to heterogeneous edge environments while maintaining predictable performance and operational safety [4]. These studies emphasize that OTA systems must move beyond static orchestration and support adaptive scheduling mechanisms capable of handling dynamic workloads, network uncertainty, and fault conditions in real deployments [5]. However, while architectural flexibility has improved, the scheduling layer itself remains a key bottleneck in achieving robust and scalable OTA performance across diverse industrial edge infrastructures.

Traditional OTA scheduling approaches, such as first-in-first-out (FIFO) policies or static resource allocation, are widely used due to their simplicity and low computational overhead [6]. Nevertheless, these methods are inherently reactive and lack the ability to adapt to changing system states. When faced with fluctuating network latency, bursty task arrivals, or partial node failures, static schedulers often exhibit degraded performance and instability. Their inability to learn from prior scheduling outcomes further limits their effectiveness in complex, large-scale cloud–edge environments [7,8]. To address these limitations, reinforcement learning (RL) has been increasingly explored as a promising approach for adaptive task scheduling in distributed systems. By learning scheduling policies through interaction with the environment, RL-based methods can dynamically adjust decisions based on observed system states, such as queue length, node load and network conditions [9]. Existing studies have demonstrated the potential of RL to reduce task latency or energy consumption compared with heuristic baselines [10,11]. However, most prior work focuses on optimizing a single objective and is evaluated in relatively small-scale settings with simplified or stable network assumptions. In practical industrial OTA scenarios, scheduling decisions must balance multiple, often competing objectives. Minimizing delay alone may lead to excessive energy consumption or severe load imbalance across edge nodes, while energy-focused optimization can increase response time and reduce system responsiveness [12]. Moreover, industrial edge systems frequently operate under non-ideal communication conditions, where network latency and packet loss fluctuate over time. Few existing studies provide systematic evaluation of scheduling robustness under such adverse conditions, particularly in the context of large-scale OTA task dissemination [13]. Another limitation of current learning-based schedulers lies in their convergence behavior and global optimization capability [14]. While deep reinforcement learning excels at local decision-making and long-term policy adaptation, it may converge to suboptimal solutions when the action space is large or the reward landscape is highly non-linear [15]. Evolutionary algorithms, such as genetic algorithms, offer complementary strengths by performing global search and maintaining population diversity, but they typically lack real-time adaptability [16]. The integration of learning-based and evolutionary optimization strategies remains underexplored in OTA scheduling research, especially for cloud–edge collaborative systems.

In this study, a hybrid OTA task scheduling framework is proposed that combines deep Q-learning with genetic algorithms to jointly optimize latency, energy efficiency, and workload balance in cloud–edge industrial environments. The deep Q-learning component enables adaptive, state-aware scheduling decisions under dynamic network and workload conditions, while the genetic algorithm enhances global optimization and prevents premature convergence. Extensive experiments are conducted on a simulated industrial edge platform with 50 edge nodes and 120 OTA tasks. The proposed framework reduces average task delay by 23.9% and energy consumption by 18.5%, while significantly improving load balance, as reflected by a reduction in the coefficient of variation from 0.42 to 0.17. Even under adverse network conditions with 300 ms communication latency, task completion rates remain above 99.2%, outperforming baseline scheduling models by 27.6%. These results demonstrate the robustness and practical potential of the proposed framework, offering a scalable and adaptive solution for large-scale, delay-sensitive OTA task scheduling in real-world industrial cloud–edge systems.

2. Materials and Methods

2.1. Sample and Study Area Description

This study involved 50 industrial edge computing nodes distributed across four smart factory zones in Jiangsu Province, China. Each node featured heterogeneous processing capabilities, including ARM-based and x86-based microcontrollers. A total of 120 OTA update tasks were prepared, covering firmware, security patches, and model parameter tuning modules. The selected environments exhibited high ambient electromagnetic interference and variable network quality to simulate realistic deployment constraints. Sampling was conducted during three time windows per day over five consecutive days to capture peak and off-peak operation conditions.

2.2. Experimental Design and Control Setup

Two groups were constructed for comparative analysis: the experimental group used the proposed hybrid scheduling method combining Deep Q-Networks (DQN) and genetic algorithms, while the control group employed traditional FIFO and static partitioning strategies. Both groups received identical tasks and node configurations. The experimental design accounted for scheduling fairness, network congestion levels, and energy profiles. Baseline performance thresholds were determined from industrial standards and prior benchmarks. Control trials were repeated five times under identical task loads to ensure stability and statistical reliability.

2.3. Measurement Procedures and Quality Control

Key metrics included task scheduling delay (ms), energy consumption (J), task success rate (%), and load distribution uniformity (COV). Delay was recorded using embedded timestamp hooks synchronized via GPS across all edge nodes. Energy metrics were captured using the INA219 current sensors interfaced with each node’s power supply line. Success rate was determined by verifying OTA hash integrity upon completion. All measurements were repeated thrice per trial and cross-verified by independent logging units. Quality control included time synchronization drift monitoring (tolerance <5 ms) and voltage calibration checks before each experiment cycle.

2.4. Data Processing and Model Formulations

Data preprocessing involved outlier filtering using a 1.5×IQR method and normalization to [0,1] range. Task scheduling effectiveness was evaluated using the following delay-weighted fairness index [17]:

$$F_d={\displaystyle \frac{(\sum _{i=1}^n\frac{1}{D_i}{)}^2}{n\text{⋅}\sum _{i=1}^n\frac{1}{D_i^2}}}$$

where ${\mathrm{D}}_{\mathrm{i}}$ denotes the delay of the $\mathrm{i}$-th node. Energy efficiency was further assessed using a normalized consumption ratio [18]:

$$E_r={\displaystyle \frac{E_{actual}-E_{min}}{E_{max}-E_{min}}}$$

where ${\mathrm{E}}_{\mathrm{actual}}$ is the measured energy, and ${\mathrm{E}}_{\min }$, ${\mathrm{E}}_{\max }$ represent baseline extremes. Data analysis was conducted using Python 3.11 and SciPy 1.11.1.

2.5. Fault Simulation and Network Perturbation Strategy

To evaluate robustness, controlled fault injection was applied to 20% of nodes per trial, including simulated hardware resets and intentional packet drops. Communication latencies were emulated by introducing synthetic delays of 100–300 ms using NetEm on each edge node's virtual interface. Link failure conditions were induced with 5% random disconnection probability per node per cycle. The system’s adaptive fault tolerance performance was assessed by comparing task rescheduling frequency, cumulative delay, and update completion integrity.

3. Results and Discussion

3.1. Scheduling Delay Reduction

The combined reinforcement-learning and genetic-algorithm scheduler achieved an average job dispatch delay reduction of 23.9 % compared to baseline approaches. In the testbed comprised of 50 edge nodes and 120 update tasks, the mean delay dropped from 312 ms (baseline) to 237 ms (proposed method). The reduction supports the effectiveness of learning-driven allocation under high-latency conditions. Figure 1 displays the delay distributions for both the baseline and optimized scheduler. Compared with earlier work in cloud-edge task scheduling [19] that reported delay reductions in the 10–15% range for smaller node clusters, this study demonstrates stronger performance gains in a significantly larger and more heterogeneous industrial environment. The learning-based algorithm successfully adapts to variable network delays (up to 300 ms) and high task volumes.

3.2. Energy Consumption and Load Balancing

Energy consumption under the scheduling model dropped by 18.5% compared to the static allocation baseline. Simultaneously, the coefficient of variation (COV) for node load dropped from 0.42 to 0.17, indicating more even distribution of update tasks across nodes. This outcome suggests that integrating genetic-based mutation of scheduling policies helps balance workload and thus lower idle energy overhead. Prior studies often focused either on delay or energy, but seldom both, the present results highlight that joint optimization is feasible under real workloads [20].

3.3. Task Success and Robustness Under High Latency

When communication latency was artificially increased to 300 ms, the task success rate of the proposed scheduler remained at 99.2%. This contrasts with 71.6% success under the FIFO/static baseline—indicating a 27.6% improvement. The system maintained high reliability even under adverse conditions, a result typically not shown in earlier edge scheduling studies which rarely present high-latency robustness data [20,21]. The hybrid scheduler dynamically assigned tasks to nodes with current favourable conditions, avoiding time-outs and failures.

3.4. Practical Deployment Considerations and Limitations

Although the algorithm performed well across the 50-node testbed, some limitations were observed. In setups involving older hardware with significantly lower computational capacity, the custom scheduling overhead introduced up to 8% extra dispatch delay compared to modern devices. Moreover, memory footprint of the genetic algorithm module presented constraints in resource-tight nodes. Figure 2 illustrates the node-class stratification of success rates under varying hardware tiers. These limitations suggest that while the algorithm is scalable, real-world deployments must account for device heterogeneity and possibly include fallback simpler schedulers [22,23]. Future work should investigate reducing algorithmic overhead on low-end nodes and extending the approach to thousands of nodes.

4. Conclusion

This study proposed a structured method for scheduling over-the-air (OTA) updates in edge systems across different domains. The method integrates latency-sensitive reinforcement learning with adaptive priority adjustment. Experimental results show that this approach improves task scheduling efficiency and increases completion rates under unstable network conditions. Compared to traditional rule-based methods, the proposed model adjusts to real-time network load changes and latency limits without relying on predefined update rules. This feature addresses a key issue in time-critical distributed applications, including autonomous driving and industrial control. The main contribution lies in the system’s ability to operate efficiently under variable latency without requiring manual tuning. However, its performance in real-world network environments with complex topologies still needs further validation. In the future, the model will be extended to support multi-cloud systems. It will also be optimized for energy efficiency to meet sustainable computing goals.