DT-Create Suite for Digital Twin Maintenance

• DT-Create Suite is a modular architecture for creating and managing Digital Twins tailored to predictive maintenance in Industry 5.0.
• It integrates real-time data collection, automated machine learning for model selection, and ontology-based semantic enrichment to enable agile decision making.
• Empirical validation via Design Science Research highlights its high accuracy and adaptability across diverse industrial contexts.

DT-Create Suite specifies a modular service Architecture for the creation and management of Digital Twins (DTs) tailored to predictive maintenance within Industry 5.0. DT-Create integrates IoT sensor data ingestion, ontological enrichment, automated machine learning model selection, and adaptive feedback mechanisms to support agile, data-driven decision making. The suite was devised and refined through two cycles of Design Science Research (DSR), addressing data collection and storage, semantic data processing, predictive modeling, and self-adaptive decision support as core functions.

1. Architectural Organization

DT-Create consists of loosely coupled microservices forming a pipeline for real-time DT composition and management. The architectural structure is formalized as: $$\mathcal{P} = \bigl\langle C, S, P, \mathrm{IP}, \mathrm{SE}, \mathrm{PS}, \mathrm{DS}, \mathrm{SA} \bigr\rangle$$ where:

• $C$: Data Collect (MQTT subscriber, message buffering, time-series DB interfacing)
• $S$: Storage
• $P$: Preprocessing (missing data cleansing, normalization, outlier removal, optional ontology-based validation)
• $\mathrm{IP}$: Intelligent Processing
• $\mathrm{SE}$: Semantic Enrichment
• $\mathrm{PS}$: Predictive-model Selection (AutoML)
• $\mathrm{DS}$: Decision Support (dashboards, alerting)
• $\mathrm{SA}$: Self-Adaptation

Data proceeds as: $$D_{\mathrm{raw}} \xrightarrow{C} D_{\mathrm{collected}} \xrightarrow{P} D_{\mathrm{clean}} \xrightarrow{\mathrm{IP}}\left\{D_{\mathrm{ML}}, D_{\mathrm{ont}}\right\} \xrightarrow{\mathrm{DS}} \text{actions/visualizations} \xrightarrow{\mathrm{SA}} \text{system reconfiguration}$$

Key pipeline components include:

• Data Collect: Continuous MQTT ingestion and buffering.
• Preprocessing: Standard and ontology-driven cleansing and validation.
• Storage & Intelligent Processing: Use of MySQL (or equivalent) and standardized APIs.
• AutoML: Automatic feature selection, candidate model training, cross-validation, hyper-parameter tuning, with model selection objective

$$m^* = \underset{m\in M}{\arg\max}\; \mathrm{score}_{CV}(m;D)$$

• Ontology Model: OWL2-based.
• Inference Machine: Fact enrichment and relationship inference via SWRL and Description Logic.
• Autonomous Agent: Subscribes to alerts, issues operational commands via MQTT.
• Web Interface: Real-time dashboards and mobile alerts.

2. Semantic Data Processing and Ontology-based Enrichment

DT-Create employs domain-specific ontologies and rule-based inference to supplement numeric sensor data with semantically rich, actionable information. The ontologies, specified in OWL 2, capture entities such as Machine, Sensor, Failure, Alert, and their relationships, for example:

• $\text{Failure} \sqsubseteq \top$
• $$\text{CriticalFailure} \sqsubseteq \text{Failure} \sqcap \exists\text{hasCriticality}.\times$$
• $$\text{hasTemperature} \sqsubseteq \text{Failure} \rightarrow \text{TemperatureReading}$$

SWRL (Semantic Web Rule Language) rules drive inference, generating domain-specific alert codes, e.g.: $$\begin{aligned} &\{\textit{Failure}(f),\;\textit{TemperatureReading}(t),\;\textit{temperatureValue}(t,v),\;v\ge35,\ &\quad \textit{HumidityReading}(h),\;\textit{humidityValue}(h,u),\;u\le25\} \ &\quad \longrightarrow\;\textit{alertCode}(f,100) \end{aligned}$$

In applied settings, such as the heat-treatment furnace scenario, the ontology encapsulated physical equipment, sensor channels, and enabled event labeling, such as inferring maintenance requirements from binary sensor values.

3. Automated Predictive Model Selection via Machine Learning

The AutoML subsystem operationalizes the selection of ML models in response to DT data characteristics. Given $D=\{(x_i, y_i)\}$ and model classes $\mathcal{M} = \{ m_j(\theta) \}$, cross-validation scoring is defined as: $$\mathrm{CVscore}(m_j(\theta); D) = \frac{1}{K} \sum_{k=1}^K \mathrm{Acc}\left(m_j(\theta), D^{\mathrm{train}}_k, D^{\mathrm{test}}_k \right)$$ Model and hyper-parameter selection is then: $$(m^*, \theta^*) = \arg\max_{m_j \in \mathcal{M},\, \theta} \mathrm{CVscore}(m_j(\theta); D)$$ Evaluated models include Decision Trees, Random Forests, LightGBM, XGBoost, KNN, and SVM, with the Decision Tree yielding top metrics for the heat-treatment dataset (Accuracy≈0.9526, AUC≈0.9898, F1≈0.9659, MCC≈0.8902).

Preprocessing for AutoML includes class balancing and confidence-interval-based label assignment: $CI = \bar{x} \pm z_{0.975} \frac{s}{\sqrt{n}}$ and feature aggregation: $E = \sum_i p_i s_i$

4. Self-Adaptation Mechanism

A MAPE-K (Monitor-Analyze-Plan-Execute over Knowledge) feedback loop underlies DT-Create’s capacity for adaptivity. Monitoring involves the assessment of performance metrics ($\mathrm{Perf}(m_t, D_t)$), such as accuracy, AUC, and F1. Anomaly detection hinges on deviations

$$e_t = \mathrm{Perf}_{\mathrm{ref}} - \mathrm{Perf}(m_t)$$

The adaptation policy enforces model re-selection if performance drops below threshold $\delta$: $$m_{t+1} = \begin{cases} \arg\max_{m \in \mathcal{M}} \mathrm{score}_{CV}(m; D_t) & \text{if } \mathrm{Perf}(m_t; D_t) < \delta \ m_t & \text{otherwise} \end{cases}$$ Corrective actions triggered by the Autonomous Agent may include sensing re-configuration, alert threshold adjustment, or process intervention (e.g., machine stop commands).

5. Design Science Research Methodology and Empirical Validation

DT-Create was iteratively refined over two DSR cycles:

Cycle	Artifact Features	Use Case	ML Approach	Ontology	Noted Outcomes
1	Data Collect, Preprocessing, MLP, SmartMaintenance	Textile, 10,000 events	Multilayer Perceptron	SWRL alertCodes	High accuracy (≈0.9976); no auto model selection
2	Added AutoML, tuning, self-adaptation	Heat treatment, 1M records, 7 chans	Diverse (13) models, Decision Tree best	SensorEquipment	ACC≈0.95, real-time dashboard, model diversity

This process highlighted the value of explicit model selection (mitigating overfitting), the impact of ontological enrichment, and the labor intensiveness of preprocessing. Self-adaptive re-training responded effectively to performance degradation, supporting non-functional requirements for adaptability and extensibility.

6. Practical Implications and Limitations

DT-Create's fusion of real-time IoT integration, ontology-driven data enrichment, automated predictive model selection, and a MAPE-K self-adaptive loop enables robust, human-centric decision support for predictive maintenance in Industry 5.0. Ontology design and SWRL rule alignment with physical domain semantics are critical for reliable inference. Preprocessing remains the principal bottleneck, particularly in label generation and class balancing. The suite demonstrates feasibility across textile and heat-treatment contexts but requires iterative tuning to generalize across domains and evolving operational data. A plausible implication is that future DT architectures for Industry 5.0 will increasingly adopt similar modular, self-adaptive, and semantically integrative paradigms.