DUA-MQTT & MAC-GC – Comprehensive Study Notes Paper title: “DUA-MQTT: A Distributed High-Availability Message Communication Model for the Industrial Internet of Things”. Authors & affiliations: Anying Chai, Wanda Yin, Mengjia Lian*, Yunpeng Sun, Chenyang Guo, Lei Wang, Zhaobo Fang. Core purpose:Address rising IIoT data-volume, protocol heterogeneity, and real-time requirements. Propose two complementary contributions: DUA-MQTT → a distributed, high-availability OPC UA + MQTT communication architecture. MAC-GC → an information-modelling method that transforms unstructured industrial text into OPC UA-compliant nodes. Key outcomes (vs. traditional baseline):End-to-end latency ↓ 28.6 %. Throughput ↑ 22.2 %. NER-based modelling metrics: Precision = 0.9701 , Recall = 0.9601 , F 1 = 0.9651. \text{Precision}=0.9701,\;\text{Recall}=0.9601,\;F_1=0.9651. Precision = 0.9701 , Recall = 0.9601 , F 1 = 0.9651. Keywords: Industrial IoT, OPC UA, MQTT, distributed systems, information modelling. Introduction & Context Manufacturing is moving toward higher automation, digitalization, intelligence. IIoT interconnects sensors, PLCs, machines → enables real-time monitoring & scheduling, but brings:Massive device counts ⇒ skyrocketing concurrent messages. Diverse protocols ⇒ interoperability headaches. Data mixture (structured numeric + unstructured text) ⇒ modelling complexity. Traditional C/S communication & single-broker MQTT suffer:Concurrency bottleneck, low throughput, high latency. Single-point failure risk. Limited ability to parse/structure free-text maintenance logs, manuals, alarms, etc. OPC UA & MQTT – Strengths & Weaknesses OPC UARich information model, platform independent, semantic consistency. Centralized server trend → poor scaling in huge, high-frequency environments. MQTTLightweight publish/subscribe, tiny header, works well on low-bandwidth/high-delay links. Classic setup uses one broker → bottleneck under IIoT scale. Combining them: Common practice employs a centralized OPC UA–MQTT proxy; still inherits single-broker issues. Research Gaps Identified Need a distributed, fault-tolerant broker system to sustain 100-1000+ publishers. Need automated conversion of unstructured device text into OPC UA nodes to avoid manual modelling. Proposed Contribution ①: DUA-MQTT Communication Model High-level goal: “Same semantics of OPC UA, throughput of distributed MQTT”. Entire stack contains 4 cooperating modules:Data analysis – classifies incoming traffic into structured vs. unstructured.Unstructured data processing – runs MAC-GC NER pipeline (details later).OPC UA address space manager – inserts/updates nodes (Objects, Variables, References) in real-time.Protocol integration & message routing – performs OPC UA→MQTT conversion, route optimisation, topic pub/sub control. Visualised in Figure 1 (not reproduced). Distributed Multi-Agent Strategy (Core of DUA-MQTT) Replaces single broker with multiple collaborating agents (brokers) → loop-free, self-electing overlay tree. Construction logic:Agents broadcast PINGREQ packets that piggy-back control info: root-IP, ability value Z Z Z P P P Ability value formula :Z = α S + β R Z=\alpha S + \beta R Z = α S + β R S S S R R R α , β \alpha,\beta α , β Highest Z Z Z Path-cost metric based on measured RTT:RTT = T < e m > r e s p o n s e − T < / e m > r e q u e s t \text{RTT} = T<em>{response} - T</em>{request} RTT = T < e m > r es p o n se − T < / e m > r e q u es t Agents dynamically recompute routes when nodes join/leave/fail → adaptive topology. Message delivery respects MQTT QoS levels 0/1/2; fallback blocking connections ensure reliability; Last-Will messages forwarded network-wide. Algorithm 1 (pseudo-code) summarises lifecycle: initial ability calc → root election → topology maintenance → QoS forwarding → fault recovery. Motivation: Industrial manuals, tickets, logs contain device functions, attributes, status; must be transformed into OPC UA schema. Model composition = MacBERT encoder + BiGRU interaction + CRF inference . Three-stage flow:Encoding layer – 12-layer Transformer; pre-training tasks:Whole-Word Masking (WWM) preserves Chinese phrase semantics. Sentence-Order Prediction (SOP) captures discourse relations. Self-attention equations:Q = X W < e m > Q , K = X W < / e m > K , V = X W < e m > V Q = XW<em>Q,\;K = XW</em>K,\;V = XW<em>V Q = X W < e m > Q , K = X W < / e m > K , V = X W < e m > V A t t e n t i o n ( Q , K , V ) = softmax ( Q K T d < / e m > k ) V Attention(Q,K,V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d</em>k}}\right)V A tt e n t i o n ( Q , K , V ) = softmax ( d < / e m > k Q K T ) V Interaction layer – bidirectional GRU with update/reset gates:r < e m > t = σ ( W < / e m > r [ h < e m > t − 1 , x < / e m > t ] + b < e m > r ) r<em>t = \sigma(W</em>r[ h<em>{t-1}, x</em>t ]+b<em>r) r < e m > t = σ ( W < / e m > r [ h < e m > t − 1 , x < / e m > t ] + b < e m > r ) z < / e m > t = σ ( W < e m > z [ h < / e m > t − 1 , x < e m > t ] + b < / e m > z ) z</em>t = \sigma(W<em>z[ h</em>{t-1}, x<em>t ]+b</em>z) z < / e m > t = σ ( W < e m > z [ h < / e m > t − 1 , x < e m > t ] + b < / e m > z ) h ~ < e m > t = tanh ( W < / e m > h [ r < e m > t ∘ h < / e m > t − 1 , x < e m > t ] + b < / e m > h ) \tilde{h}<em>t = \tanh(W</em>h[ r<em>t\circ h</em>{t-1}, x<em>t ]+b</em>h) h ~ < e m > t = tanh ( W < / e m > h [ r < e m > t ∘ h < / e m > t − 1 , x < e m > t ] + b < / e m > h ) h < e m > t = ( 1 − z < / e m > t ) ∘ h < e m > t − 1 + z < / e m > t ∘ h ~ t h<em>t = (1-z</em>t)\circ h<em>{t-1} + z</em>t\circ \tilde{h}_t h < e m > t = ( 1 − z < / e m > t ) ∘ h < e m > t − 1 + z < / e m > t ∘ h ~ t Inference layer (CRF) – imposes global label consistency; objective:P ( Y ∣ X ) = exp ( ∑ < e m > i = 1 n ( A < / e m > y < e m > i − 1 , y < / e m > i + S < e m > i , y < / e m > i ) ) ∑ < e m > Y ′ exp ( ∑ < / e m > i = 1 n ( A < e m > y < / e m > i − 1 , y < e m > i ′ + S < / e m > i , y < e m > i ′ ) ) P(Y|X)=\frac{\exp\left(\sum<em>{i=1}^{n}(A</em>{y<em>{i-1},y</em>i}+S<em>{i,y</em>i})\right)}{\sum<em>{Y'}\exp\left(\sum</em>{i=1}^{n}(A<em>{y</em>{i-1},y<em>i'} + S</em>{i,y<em>i'})\right)} P ( Y ∣ X ) = ∑ < e m > Y ′ e x p ( ∑ < / e m > i = 1 n ( A < e m > y < / e m > i − 1 , y < e m > i ′ + S < / e m > i , y < e m > i ′ ) ) e x p ( ∑ < e m > i = 1 n ( A < / e m > y < e m > i − 1 , y < / e m > i + S < e m > i , y < / e m > i ) ) L < / e m > C R F = − log P ( Y ∣ X ) \mathcal{L}</em>{CRF} = -\log P(Y|X) L < / e m > C R F = − log P ( Y ∣ X ) Output BIO-style labels mapped to OPC UA nodes (Table 5 mapping: OBJ→Object, COM→Reference, ATT/VAL→Variable nodes/values). Generates XML files → imported into UA-Server; verified with UaExpert tree view. Experimental Setup Hybrid lab mimicking “perception–edge–cloud”:Field: wireless sensor nodes. Edge: Industrial PC running DUA-MQTT conversion; three Raspberry Pi 4 brokers (4 GB RAM each). Cloud: Xeon server (64 GB RAM) for storage, analytics, MAC-GC inference. Network: mixed wired/wireless; publishers varied 10→500; QoS 0/1/2 tested. Dataset for NER: SIGHAN-2005 Chinese corpus (~2 000 texts, 150 k vocab). Model hyper-params: 12 layers, hidden 768, GRU 128, epoch 20, lr 1 × 10 − 5 1\times10^{-5} 1 × 1 0 − 5 Throughput (messages/sec) vs #publishers (Fig. 7):DUA-MQTT sustains ≈220 msg/s at 300–500 publishers (QoS 0/1) vs <180 for centralized OPC-MQTT. QoS 2 lower but still superior to baseline. End-to-end delay (ms) vs #publishers (Fig. 8):OPC-MQTT delay balloons to ≈14 000 ms at 500 pubs (QoS 2). DUA-MQTT grows much slower; remains low in QoS 0/1 thanks to distributed load-balancing. Net improvement: latency ↓28.6 %, throughput ↑22.2 % relative to baseline. Ablation study:BiGRU-CRF → P = 0.8671 , R = 0.8599 , F 1 = 0.8635 P=0.8671, R=0.8599, F_1=0.8635 P = 0.8671 , R = 0.8599 , F 1 = 0.8635 MacBERT-CRF → P = 0.9355 , R = 0.9238 , F 1 = 0.9294 P=0.9355, R=0.9238, F_1=0.9294 P = 0.9355 , R = 0.9238 , F 1 = 0.9294 MAC-GC (full) → P = 0.9701 , R = 0.9601 , F 1 = 0.9651 P=0.9701, R=0.9601, F_1=0.9651 P = 0.9701 , R = 0.9601 , F 1 = 0.9651 Cross-model comparison:BERT-CRF 0.903 F1; BERT-BiLSTM-CRF 0.919; RoBERTa-BiLSTM-CRF 0.932; MAC-GC leads by sizeable margin. Qualitative gains:Better long-distance dependency capture. Higher robustness to ambiguous industrial jargon. Faster convergence within 20 epochs. Annotation System & OPC UA Integration Custom BIO labels: B-OBJ/I-OBJ, B-COM/I-COM, B-ATT/I-ATT, B-VAL/I-VAL, O. Converts recognised triplets into XML conforming to OPC UA NodeSet schema → auto-import into server; UaExpert shows nodes → proves feasibility for plug-and-play semantic integration. Conclusions & Significance DUA-MQTT eliminates single-broker bottleneck, delivering high throughput & low latency vital for real-time IIoT. MAC-GC automates unstructured-to-structured transformation, enhancing knowledge reuse & decision-making. Together, they form an end-to-end pipeline from raw sensor/log data to interoperable, high-performance message flows. Future Work & Outlook Embed intelligent scheduling/AI-driven load balancers to further optimise agent cooperation. Build large-scale, domain-specific Chinese industrial corpora for stronger generalisation and benchmarking. Enhance security: integrate hybrid intrusion-detection (e.g., KNN-NN) directly into broker mesh. Explore edge-cloud co-training so that MAC-GC fine-tunes on-device to protect IP & reduce bandwidth. Ethical, Practical & Industrial Implications High availability & failover crucial for safety-critical plants; DUA-MQTT’s multi-agent redundancy mitigates downtime, potential production losses. Semantic consistency via OPC UA nodes eases vendor interoperability, promoting open ecosystems. Automation of manual modelling reduces human error, accelerates digital-twin/Industry 4.0 deployments. Numerical / Statistical Summary Latency reduction: 28.6 %. Throughput increase: 22.2 %. MAC-GC metrics: P = 0.9701 , R = 0.9601 , F 1 = 0.9651 P=0.9701, R=0.9601, F_1=0.9651 P = 0.9701 , R = 0.9601 , F 1 = 0.9651 Ability formula: Z = α S + β R Z=\alpha S + \beta R Z = α S + β R RTT formula: RTT = T < e m > r e s p o n s e − T < / e m > r e q u e s t \text{RTT}=T<em>{response}-T</em>{request} RTT = T < e m > r es p o n se − T < / e m > r e q u es t Funding & Acknowledgements (Concise) Supported by multiple Chinese provincial foundations (Liaoning, Fujian) & industry–university projects. Authors declare no conflict of interest. Reference Nuggets (select) Integrating OPC UA with MQTT-SN (Nast et al.). BORDER framework for distributed MQTT benchmarking (Longo et al.). Peer-to-peer privacy-enhanced MQTT-A (Buccafurri et al.). Pre-trained biomedical NLP analogues (BioBERT, MacBERT Chinese).