Self-Evolving Multi-Agent Network for Industrial IoT Predictive Maintenance
Authors: Rebin Saleh, Khanh Pham Dinh, Balázs Villányi, Truong-Son Hy
Organizations: Department of Electronics Technology, Faculty of Electrical Engineering and Informatics, Budapest University of Technology and Economics, Műegyetem rkp. 3, H-1111 Budapest, Hungary · DataScienceWorld.Kan, Hanoi, Vietnam · Department of Computer Science, The University of Alabama at Birmingham, 1720 University Blvd, Birmingham, AL 35294 USA
Industrial IoT predictive maintenance requires systems capable of real-time anomaly detection without sacrificing interpretability or demanding excessive computational resources. Traditional approaches rely on static, offline-trained models that cannot adapt to evolving operational conditions, while LLM-based monolithic systems demand prohibitive memory and latency, rendering them impractical for on-site edge deployment. We introduce SEMAS, a self-evolving hierarchical multi-agent system that distributes specialized agents across Edge, Fog, and Cloud computational tiers. Edge agents perform lightweight feature extraction and pre-filtering; Fog agents execute diversified ensemble detection with dynamic consensus voting; and Cloud agents continuously optimize system policies via Proximal Policy Optimization (PPO) while maintaining asynchronous, non-blocking inference. The framework incorporates LLM-based response generation for explainability and federated knowledge aggregation for adaptive policy distribution. This architecture enables resource-aware specialization without sacrificing real-time performance or model interpretability. Empirical evaluation on two industrial benchmarks (Boiler Emulator and Wind Turbine) demonstrates that SEMAS achieves superior anomaly detection performance with exceptional stability under adaptation, sustains prediction accuracy across evolving operational contexts, and delivers substantial latency improvements enabling genuine real-time deployment. Ablation studies confirm that PPO-driven policy evolution, consensus voting, and federated aggregation each contribute materially to system effectiveness. These findings indicate that resource-aware, self-evolving 1multi-agent coordination is essential for production-ready industrial IoT predictive maintenance under strict latency and explainability constraints.
Figures & tables
Fig. 1: High-level self-evolving hierarchical MAS architecture for IIoT predictive maintenance.
Fig. 2: Detailed agent-level architecture showing all subagents, communication protocols, and feedback loops.
Characteristic
Specification
Dataset Size
Total samples
10,000 operational cycles
Training set
8,000 samples (80%)
Test set
2,000 samples (20%)
Sampling strategy
Stratified by anomaly class
Feature Space
TABLE I: Boiler Emulator dataset characteristics and configuration.
Characteristic
Specification
Dataset Size
Total samples
500 records (after balancing from 1000+)
Training set
400 samples (80%)
Test set
100 samples (20%)
Feature Space
Total dimensions
42 features
TABLE II: Wind Turbine IIoT dataset characteristics and configuration.
System
Dataset
F1
Precision
Recall
ROC-AUC
Δ F1
Latency (ms)
Baseline1
Boiler
0.4871
0.3225
0.9949
0.4099
0.0000
1923.0
Baseline2
Boiler
0.4489
0.3140
0.7911
0.3969
-0.0550
1594.3
SEMAS
Boiler
0.4873
0.3737
0.7114
0.6118
-0.0239
1.22
Baseline1
Wind Turbine
0.9440
0.9219
0.9672
0.3705
0.0000
455.9
Baseline2
Wind Turbine
0.9349
0.9205
0.9508
0.2634
+0.0606
286.2
SEMAS
Wind Turbine
0.9571
0.9371
0.9781
0.7583
-0.0166
0.30
TABLE III: Performance comparison across datasets
Fig. 3: Comprehensive performance comparison across all systems and metrics. SEMAS achieves superior F1-scores (Boiler: 0.4873, Wind Turbine: 0.9571), highest ROC-AUC values (0.6118 and 0.7583), and dramatically lower latency (1.22ms and 0.30ms) compared to Baseline1 (1923ms and 456ms) and Baseline2 (1594ms and 286ms). The multi-metric visualization demonstrates SEMAS’s dual advantage in both accuracy and computational efficiency.
Fig. 4: F1-score evolution across iterations for Baseline1, Baseline2, and SEMAS on both datasets. SEMAS demonstrates stable convergence and positive learning on challenging data.
Parameter
It. 1
It. 2
It. 3
Change
F1-Score
0.5031
0.4795
0.4792
-4.7%
Contamination ( ρ )
0.3200
0.3100
0.3300
+3.1%
Consensus threshold ( τ )
0.7327
0.8000
0.8000
+9.2%
Weight w1 (B1)
0.4243
0.4290
0.4333
+2.1%
Weight w2 (B2)
0.5757
0.5710
0.5667
-1.6%
TABLE IV: SEMAS policy parameter evolution on Boiler dataset across three iterations, with corresponding F1-scores.
System
Boiler (ms)
Wind (ms)
Memory (GB)
Baseline1
1923.0
455.9
2.1 (est.)
Baseline2
1594.3
286.2
2.4 (est.)
SEMAS
1.22
0.30
3.8 (est.)
TABLE V: Computational efficiency comparison (averages across all iterations).
Configuration
F1
Impact
Prec.
Op. Acc.
SEMAS (Full)
0.4956
Baseline
0.3712
82%
w/o PPO optimization
0.4782
-3.5%
0.3501
81%
w/o consensus voting
0.4634
-6.5%
0.3045
79%
w/o federated aggregation
0.4856
-2.0%
0.3614
80%
w/o LLM response
0.4956
+0% (F1)
0.3712
41%
TABLE VI: Ablation study on Boiler dataset: per-component contributions to accuracy and operator acceptance. The last row reports an orthogonal trust metric rather than detection accuracy.
Dataset
Comparison
Δ F1
t
p
Interpretation
Boiler
SEMAS vs. BL1
+0.0002
0.08
0.94
Not significant
Boiler
SEMAS vs. BL2
+0.0384
4.21
<0.001
Highly significant
Wind Turbine
SEMAS vs. BL1
+0.0131
2.76
<0.01
Significant
Wind Turbine
SEMAS vs. BL2
+0.0222
3.12
<0.005
Highly significant
TABLE VII: Paired t -tests comparing SEMAS against baselines. Significance threshold: p<0.05 .
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Justification
Valid Range
Contamination ( ρ )
0.32
Matches Boiler anomaly prevalence (36.8%)
[0.25, 0.35]
Window size ( W )
100 ts
10 seconds at 10Hz sampling rate
[50, 200]
Initial w1,w2
0.42, 0.58
Equal weighting; PPO learns optimal split
[0.3, 0.7]
PPO clip ratio ( ϵ )
0.2
Standard RL literature value
[0.1, 0.3]
Appendix
TABLE VIII: TABLE VIII Hyperparameter justification and sensitivity analysis.
Model
Hyperparam.
Value
Isol. Forest
Est. ( ntrees )
200
Contam. ( ρ0 )
0.32
Max samples/tree
256
Rand. seed
42
Training
Unsupervised
One-Class SVM
Kernel
RBF
Appendix
TABLE IX: Training configuration for anomaly detection models.
Component
Parameter
Value/Spec.
RUL Model
Arch.
LSTM [64,32,32,1] (ReLU)
Input
(Xt−10:t,RULt)
Batch
32
Epochs
10 (conv. obs.)
Opt.
Adam ( α=10−3 )
Loss
MAE
Appendix
TABLE X: Training configuration for RUL prediction and PPO policy optimization.
Large language model (LLM) agents have shown promise in automating complex data-analysis workflows, but their reliable deployment remains challenging in high-stakes industrial scenarios. Industrial anomaly detection (IAD) is essential for manufacturing quality, safety, and efficiency, yet existing LLM-based IAD agents mainly focus on execution while under-exploiting strategy formulation. Consequently, they struggle to handle heterogeneous modalities in a unified and cost-effective manner. Inspired by the DMAIC quality-management framework, we propose DMAIC-IAD (DMAIC-inspired Agentic Industrial Anomaly Detection), a "Plan First, Judge Later" multi-agent system that aligns LLM agents with structured industrial problem-solving. DMAIC-IAD distills heterogeneous references into standardized operating procedures (SOPs) before strategy generation, and introduces a pre-trained execution-free judge model to rank candidate strategies without costly runtime trials. Extensive experiments across four modalities show that DMAIC-IAD improves average detection performance over applicable agentic baselines by 37.76%.
Yongzi Yu, Ao Li, Le Wang +4
1The Hong Kong University of Science and Technology (Guangzhou) · 2The Hong Kong University of Science and Technology · 3Shanghai University of Finance and Economics +2
Providing reliable predictive maintenance is a critical industrial AI service essential for ensuring the high availability of manufacturing devices. Existing deep-learning methods present competitive results on such tasks but lack a general service-oriented framework to capture complex dependencies in industrial IoT sensor data. While Transformer-based models show strong sequence modeling capabilities, their direct deployment as robust AI services faces significant bottlenecks. Specifically, streaming sensor data collected in real-world service environments often exhibits multi-scale temporal correlations driven by machine working principles. Besides, the datasets available for training time-to-failure predictive services are typically limited in size. These issues pose significant challenges for directly applying existing models as robust predictive services. To address these challenges, we propose MsFormer, a lightweight Multi-scale Transformer designed as a unified AI service model for reliable industrial predictive maintenance. MsFormer incorporates a Multi-scale Sampling (MS) module and a tailored position encoding mechanism to capture sequential correlations across multi-streaming service data. Additionally, to accommodate data-scarce service environments, MsFormer adopts a lightweight attention mechanism with straightforward pooling operations instead of self-attention. Extensive experiments on real-world datasets demonstrate that the proposed framework achieves significant performance improvements over state-of-the-art methods. Furthermore, MsFormer outperforms across industrial devices and operating conditions, demonstrating strong generalizability while maintaining a highly reliable Quality of Service (QoS).
Jiahui Zhou, Dan Li, Ruibing Jin +5
School of Software Engineering, Sun Yat-Sen University, China · Agency for Science, Technology and Research (A*STAR), Singapore · School of the Art Institute of Chicago, Chicago +1
Industrial IoT sensors produce continuous, high-volume data streams that predictive maintenance needs to analyse in real time. Cloud-centric architectures send all raw telemetry to central servers, which costs latency, bandwidth and resilience. This paper presents a hybrid cloud-edge architecture for manufacturing predictive maintenance built on AWS IoT Core and AWS IoT Greengrass v2, with local machine learning inference at the edge and selective cloud synchronization. We deploy two configurations on real AWS infrastructure and compare them: a cloud-only baseline (Configuration A) and the proposed hybrid edge architecture (Configuration B). We measure end-to-end inference latency, bandwidth consumption and inference accuracy. Configuration B cuts cloud-bound bandwidth by 86.8% (63.6 KB to 7.2 KB per 200-reading window) and decision latency by 256x (23.1 ms cloud round trip to 0.09 ms local inference). At the edge, with no cloud involvement, it keeps an anomaly-class F1 score of 0.84 and 91% recall. The evaluation also instruments system uptime and Stream Manager queuing under cloud disconnection; the quantitative disconnection experiment is left as planned work. Our review of the literature points to four gaps that this work addresses: the lack of empirically validated Greengrass v2 deployments for predictive maintenance, unquantified bandwidth savings from selective edge synchronization, unmeasured resilience under cloud disconnection, and the open MLOps feedback loop at the edge.