IJCNC 01

AI-Driven Multi-Agent System for QOS Optimization in 6g Industrial Networks

Ndidi Nzeako Anyakora M.I.E.E.E and Cajetan
M. Akujuobi, P.E., S.M.I.E.E.E., F.I.A.A.M.

Center of Excellence for Communication Systems Technology Research (CECSTR),
Roy G. Perry College of Engineering, Prairie View A & M University

ABSTRACT

The emergence of sixth-generation (6G) wireless technology will unlock unprecedented capabilities for Industrial Internet of Things (IIoT) networks by enabling terabit-per-second data rates, sub-millisecond latency, and extreme reliability. These advances will support mission-critical applications such as realtime robotics, autonomous manufacturing, and immersive automation. This paper presents an AI-driven Multi-Agent System (MAS) for real-time Quality of Service (QoS) anomaly detection and adaptive network optimization in 6G industrial environments. The MAS integrates three cooperating agents: a Monitoring Agent for telemetry collection, an AI-based Anomaly Detection Agent using Isolation Forest and deep Autoencoders, and a Reinforcement Learning Optimization Agent employing Proximal Policy Optimization (PPO) to self-tune network parameters. Experiments conducted on a Firecell 5G Standalone testbed emulating 6G conditions demonstrate the system’s effectiveness. The MAS reduced average latency by ≈40%, increased throughput by 15–20%, and lowered packet loss by up to 70% compared to static management baselines. These results validate the MAS’s ability to maintain consistent QoS under dynamic industrial workloads. Key contributions include: (1) a unified MAS architecture for closed-loop QoS control, (2) integration of hybrid AI models for anomaly detection and adaptive optimization, and (3) real-world testbed validation bridging 5G SA and 6G-IIoT research. For access to the code, data, and experimental results, visit our GitHub repository (Didilish/AI_Driven_MAS_For_Anomaly-Detection-QoS-Optimization-6G-IIOT).

KEYWORDS

6G, Industrial IoT, QoS, multi-agent systems, reinforcement learning, anomaly detection, 5G Standalone,
real-time control

1.INTRODUCTION

The deployment of sixth-generation (6G) wireless networks is expected to transform industrial IoT applications by delivering unmatched performance. Building on 5G advancements, 6G targets a tenfold increase in peak data rates—reaching terabits per second—and a substantial reduction in communication latency to the sub-millisecond range [1]. These capabilities are crucial for advanced IIoT use cases such as real-time process control, autonomous robotics, and immersive telepresence in smart factories. According to the ITU IMT-2030 framework [2], 6G will enable new scenarios such as integrated sensing and communication, hyper-reliable lowlatency communication, and AI-enhanced connectivity. This vision aligns with key performance targets: ultra-low latency (0.1 ms), extreme reliability (10⁻⁷), and dense device connectivity (up to 10⁹ devices per km²). Achieving and maintaining Quality of Service (QoS) at these scales is essential, as industrial efficiency depends on meeting stringent requirements for latency, throughput, and availability [1].

However, ensuring such QoS in practice poses significant challenges. 6G IIoT networks will integrate vast numbers of heterogeneous devices and technologies, including terrestrial and satellite links, while supporting highly diverse applications. These conditions increase the risks of congestion, interference, and rapid context switching that can degrade performance.
Moreover, 6G’s use of the sub-terahertz spectrum introduces new propagation constraints, demanding innovative solutions for reliable connectivity [2]. Traditional centralized network management approaches struggle to react quickly to such complexity, motivating the shift toward intelligent, autonomous control.

Artificial Intelligence (AI) and, specifically, Multi-Agent Systems (MAS) provide a promising foundation for adaptive network management. A MAS comprises multiple intelligent agents that can perceive their environment, make decisions, and act autonomously while coordinating with one another. This decentralized structure allows local adaptation and scalability advantages
critical for distributed IIoT networks. Deploying cooperative agents throughout the network (at base stations, gateways, or edge devices) enables continuous monitoring and rapid response to anomalies such as congestion or node failure. Previous studies show that AI-driven algorithms can significantly reduce latency and packet loss by learning optimal control policies for scheduling and resource allocation [3].

The Firecell Labkit testbed offers an effective experimental platform for evaluating these approaches. It functions as a portable “network-in-a-box,” providing an open-source 5G Standalone (SA) testbed with a configurable core, radio unit, and UE support [5]. Its programmable environment allows fine-grained control of bandwidth, frequency, and telemetry capture making it ideal for emulating 6G-like behaviour. Researchers can collect RRC, MAC, and IQ samples to assess throughput, latency, and packet error rates in real time [6][7]. Leveraging this platform, our research integrates AI-driven MAS intelligence into the Firecell testbed to analyze performance and optimize QoS under realistic industrial conditions.

2 .LITERATURE REVIEW

This section reviews existing research related to QoS optimization, AI-driven network management, and Multi-Agent Systems (MAS) in industrial 6G environments. It summarizes progress in 6G IIoT technologies, identifies current challenges, and highlights the role of AIbased solutions that motivate the proposed framework.

2.1. QoS Requirements and Advances in 6G IIoT

The vision for 6G wireless networks emphasizes extreme performance targets, including terabitper-second data rates and sub-millisecond latency, to support mission-critical IIoT applications such as real-time robotics, remote surgery, and large-scale automation. Achieving end-to-end massive Ultra-Reliable Low-Latency Communication (mURLLC) requires advanced, AI-driven
QoS mechanisms that combine dynamic slicing, edge computing, and predictive orchestration [4][8]. As a compact, open-source private mobile network, it enables experiments on factory automation and wireless control. Previous studies using the Firecell testbed demonstrated its effectiveness in evaluating performance indicators such as throughput and latency, as well as its utility for identifying security vulnerabilities and benchmarking industrial wireless systems [9].

2.2. Challenges in Maintaining QoS

Despite innovation, the use of the THz spectrum will introduce propagation and interference issues. Ensuring cybersecurity, integrating with legacy industrial infrastructure, and achieving energy-efficient deployment of dense small cells remain key concerns [4]. These call for intelligent, adaptive control systems to optimize performance while managing cost and complexity

2.3. AI and MAS in Network Management

Multi-Agent Systems (MAS) offer decentralized control, scalability, and context-awareness. Each agent can monitor, analyze, or act autonomously while cooperating with others. MAS has been used in smart factories and vehicular networks to reduce congestion and improve service continuity [8]. In 6G IIoT, MAS will provide real-time performance adaptation across network layers by distributing the management load.

2.4. Machine Learning for Anomaly Detection

Machine learning models like Isolation Forests and Autoencoders can learn normal traffic behavior and flag deviations as anomalies. These models have been deployed in smart city systems for detecting failures in sensors, traffic flows, and IoT endpoints. Recent work explores agent-based anomaly detection, where distributed agents each analyze their local data streams
for early fault detection [10].

2.5. Reinforcement Learning for QoS Optimization

Reinforcement Learning (RL) allows agents to learn optimal policies by interacting with the network environment. In IIoT, RL can adjust parameters like routing, scheduling, or power levels to reduce latency and congestion. Multi-agent RL (MARL) extends this to collaborative optimization across network slices or nodes. Park et al. demonstrated MARL’s effectiveness in network slicing for 6G edge computing environments [11].

2.6. Emerging Trends

Recent proposals emphasize the need for Quality of AI Service (QoAIS) metrics such as generalization, robustness, and explainability. These complement QoS to evaluate the AI models powering network optimization. Integration of LLM-based agents, federated learning, and context-aware orchestration is emerging to enhance autonomy and scalability [8][12].

2.7. Summary of Related work

Table 1.Summary of Related work.

3 .PROBLEM STATEMENT

Industrial IoT (IIoT) networks face increasing challenges in meeting the stringent requirements of modern automation and manufacturing systems [14]. Traditional network management approaches are often static, centralized, and unable to adapt to rapidly changing network conditions, heterogeneous devices, and time-sensitive workloads [15]. With the emergence of 6G, these limitations are amplified as networks must simultaneously support massive connectivity, ultra-low latency, and high throughput for mission-critical applications [16].

A key challenge lies in maintaining consistent Quality of Service (QoS) across diverse industrial applications. For instance, control systems in robotics demand sub-millisecond latency, while data aggregation systems prioritize throughput and reliability. Conventional rule-based or threshold-based approaches struggle to balance these competing demands under dynamic conditions such as traffic surges, interference, or equipment failure. In contrast, AI-based mechanisms allow the system to learn complex relationships between network metrics dynamically, rather than relying on static thresholds. Machine learning models can identify subtle performance degradations before they manifest as service disruptions. This capability is essential for 6G IIoT systems, where rapid adaptation and predictive control are required to maintain continuous QoS[17]. Hence, incorporating AI into network management is not optional but a necessity to achieve self-optimization and resilience in high-density industrial environments [4][5][8].

Artificial Intelligence (AI) offers a more adaptive solution. Unlike static management systems, AI techniques, particularly Multi-Agent Systems (MAS), enable decentralized intelligence. Each agent can monitor, reason, and act locally while collaborating globally, allowing faster and more context-aware decision-making. For example, an AI-driven agent can detect latency spikes in real time and coordinate bandwidth adjustments before service degradation occurs. Previous studies show that AI-based control reduces response time and packet loss compared to heuristic or rule-driven systems [4][8][11]. Comparative studies such as [11] and [19] demonstrate that while static or threshold-based controllers can maintain QoS within limited operating conditions, they lack the adaptability required for heterogeneous 6G environments. Our proposed MAS builds upon these findings by introducing distributed intelligence—allowing each agent to react to context changes locally, leading to faster stabilization and more resilient QoS recovery under variable load.

This research proposes an AI-driven MAS framework to provide dynamic, real-time QoS management in 6G-enabled IIoT networks. The framework integrates three cooperating agents: a Monitoring Agent, an Anomaly Detection Agent, and a Reinforcement Learning–based Optimization Agent. Together, they form a continuous feedback loop that senses performance, identifies anomalies, and autonomously adjusts network parameters to restore optimal QoS.

To ensure realism, the proposed system is implemented and evaluated using the Firecell 5G Standalone (SA) Labkit, testbed which emulates 6G-like conditions through programmable control of traffic patterns, bandwidth allocation, and latency injection [18]. This setup enables direct comparison between the MAS framework and traditional static management approaches, demonstrating the measurable improvements achieved through AI-driven adaptation.

4.METHODOLOGY

To address the challenge of real-time QoS management in 6G-enabled IIoT networks, we designed a Multi-Agent System (MAS) composed of three cooperating agents, each endowed with distinct AI capabilities. Together, these agents form a closed feedback loop that continuously monitors network performance, detects anomalies, and optimizes configuration parameters in response. The Firecell Labkit testbed serves as the experimental platform, providing a 5G/6G core network and radio environment capable of generating traffic and capturing fine-grained telemetry such as RRC transitions, MAC-layer metrics, and IQ samples[5][9][18].

4.1. Monitoring Agent (QoS Data Collector)

The Monitoring Agent continuously extracts network telemetry from the Firecell Labkit testbed, gathering metrics such as latency, packet-loss rate, throughput, and signal quality. It parses system logs through Labkit testbed APIs and maintains a real-time view of network health. When any metric exceeds a predefined threshold for instance, latency above X ms or throughput below Y Mbps the agent issues alerts that trigger analysis by the Anomaly Detection Agent. This agent functions as the eyes of the MAS, ensuring constant situational awareness of network performance [7][9].

4.2. Anomaly Detection Agents (AI “Detective”)

The Anomaly Detection Agent functions as an AI detective its role is to investigate the network’s health by examining continuous data streams from the Monitoring Agent. Each batch of incoming metrics: latency, throughput, packet loss, and channel quality presents a case that the agent must analyze. Using unsupervised learning techniques such as the Isolation Forest algorithm and a deep Autoencoder, the agent first learns the patterns of normal network behaviour and then identifies deviations from those patterns. The Isolation Forest is trained on historical data representing normal operating conditions and assigns an anomaly score to each new observation; any instance with a score above the defined threshold is labelled anomalous [10]. To capture more subtle irregularities, the deep Autoencoder reconstructs expected metric values and measures the reconstruction error; large deviations indicate abnormal behaviour [10]. This process enables the agent to detect complex, nonlinear interactions that static thresholds cannot capture. For example, a temporary latency increase with stable throughput might signal scheduling inefficiency rather than congestion. By learning these multidimensional relationships directly from data, the agent distinguishes between such patterns and flags only meaningful anomalies. Once an anomaly is detected and categorized, such as a throughput collapse, latency spike, or packet-loss burst, the information is transmitted to the Optimization Agent for corrective action. This adaptive, learning-based analysis replaces brittle rule sets with selfimproving models capable of evolving alongside the network’s dynamics. The inclusion of AI is essential here because deterministic rules cannot represent the complex dependencies among latency, throughput, and loss in real industrial traffic. The Anomaly Detection Agent’s machinelearning core continuously refines its understanding of these dependencies without human supervision, enabling proactive and reliable anomaly detection in highly dynamic 6G environments [5][8][10][12].

4.3. Optimization Agent (Reinforcement Learning Controller)

When an anomaly is detected, the Optimization Agent selects and applies corrective actions using Reinforcement Learning (RL). The environment state includes current metrics and contextual load indicators. Actions include adjusting scheduling priorities, reallocating bandwidth, and tuning transmission power. We implement a Deep Q-Network (DQN) for discrete control and extend with Proximal Policy Optimization (PPO) for policy-gradient updates where finer control is needed. The reward encourages lower latency and loss and higher throughput. Over iterative interaction, the agent learns policies that restore QoS efficiently [11][13]. As experience accrues, the controller transitions from reactive correction to proactive