Edge AI-driven adaptive maintenance for manufacturing systems: improving decision-making and reducing energy consumption Open Access

Manufacturing systems in contemporary contexts are increasingly confronted with the imperative to enhance operational efficiency while simultaneously minimizing energy consumption and ensuring sustainability. The emergence of Industry 4.0 has facilitated the integration of intelligent systems within factories, enabling real-time monitoring, predictive analytics, and data-driven decision-making (Zhang et al., 2023). A significant challenge inherent in this transformation is the development of effective maintenance strategies that transcend traditional reactive or time-based approaches.

Adaptive maintenance, underpinned by machine learning and sensor data, has surfaced as a viable solution for managing complex and dynamic production environments (Li and Zhao, 2022). By transitioning from scheduled interventions to condition-based responses, organizations can mitigate unexpected breakdowns, prolong equipment lifespan, and reduce energy consumption (Kumar and Singh, 2021). However, the successful implementation of such strategies necessitates the availability of reliable predictive models and decision support systems capable of functioning efficiently at the network edge, particularly in energy-intensive sectors like steel manufacturing.

Recent research has underscored the benefits of Edge AI in facilitating localized, low-latency decision-making without an exclusive reliance on cloud infrastructure (Wang and Zhang, 2019; Johnson and Smith, 2023). These methodologies are especially pertinent in contexts where rapid response times and data privacy are of paramount importance. Nevertheless, despite these advancements, there remains a paucity of research focused on the integration of Edge AI with adaptive maintenance frameworks that explicitly prioritize energy performance as a central objective (Nguyen and Tran, 2022; Lee and Kim, 2020).

This paper seeks to address this research gap by proposing and validating an Edge AI-based adaptive maintenance model that supports predictive decision-making and energy optimization within manufacturing systems. Utilizing real operational data from Qatar Steel Company (QSTEEL), the study constructs and evaluates a Random Forest model deployed at the edge. The methodology examines key variables influencing energy consumption and assesses the predictive efficacy of the model through statistical and machine learning metrics. In doing so, it contributes to the expanding body of literature on intelligent manufacturing and provides pragmatic insights for energy-aware maintenance planning in smart factories.

Additionally, while the concept of Edge AI has been positioned as a solution for localized inference, it shares conceptual proximity with cognitive digital twins. However, unlike cognitive digital twins which emphasize continuous synchronization between physical and virtual entities for system-level learning, our framework prioritizes real-time, on-site decision-making using retrainable models tailored to energy-aware adaptive maintenance. This practical deviation from simulation-intensive digital twins addresses immediate industrial constraints in energy-intensive environments.

The literature on predictive and adaptive maintenance in manufacturing can be organized into three thematic areas: statistical modeling, artificial intelligence techniques, and decision support systems.

Early efforts in predictive maintenance relied on traditional statistical methods such as linear regression, time series analysis, and ANOVA to estimate failure patterns and forecast maintenance needs. These models provided foundational insights into machine behavior but were limited in capturing non-linear interactions in complex environments. For example, ANOVA has been used to detect variations in energy consumption across operational shifts (Brown and Green, 2021), while correlation analysis has been widely applied to examine the relationships between process variables and energy usage (Lee and Kim, 2020). However, statistical models often fall short when dealing with large-scale, high-frequency industrial data streams.

Recent advancements in machine learning have significantly improved the ability to model dynamic systems. Techniques such as Random Forest, Support Vector Regression, and Neural Networks have been adopted to build predictive maintenance frameworks that adapt to changing operational conditions. Random Forest, in particular, has gained popularity for its accuracy and robustness in handling non-linear data (Ahmed and Khan, 2022). Edge AI represents the next evolution of this trend, allowing real-time inference and decision-making at the source of data generation. Studies by Wang and Zhang (2019) and Zhou and Wang (2023) emphasize the value of Edge AI in reducing latency, enhancing system autonomy, and supporting decentralized maintenance strategies.

Despite these advances, few studies integrate AI models with energy performance indicators. Kumar and Singh (2021) explored the application of AI in optimizing energy consumption in production, but their models were primarily cloud-based. The opportunity to combine energy forecasting and predictive maintenance within Edge AI frameworks remains underexplored, especially in energy-intensive sectors like steel manufacturing.

Maintenance-related decision support systems (DSS) are increasingly incorporating real-time data and analytics to support factory operations. These systems aim to improve scheduling, diagnostics, and failure response by offering actionable insights to maintenance engineers (Chen and Zhao, 2020). The integration of AI into DSS allows for more adaptive planning and dynamic reconfiguration of resources. However, most existing systems rely on centralized processing and often neglect energy as a decision variable. Lee and Kim (2020) stress the importance of integrating energy efficiency metrics into maintenance planning, yet practical frameworks that support this integration are still lacking.

In conclusion, while literature in each of these streams offers valuable contributions, there remains a gap in combining Edge AI-based predictive maintenance with energy-aware decision support. This study aims to bridge that gap by developing and validating a practical model using real-world data from Qatar Steel Company. As shown in Table 1, a comparative summary of recent studies is presented.

While prior studies have enhanced our understanding of predictive maintenance, energy efficiency, and Edge AI, few have proposed integrated frameworks that connect these domains using authentic industrial data. Most existing models predominantly concentrate on either energy forecasting or equipment maintenance in isolation, frequently relying on cloud-based infrastructures or simulation environments. This study contributes to the literature by presenting a practical, data-driven Edge AI model for adaptive maintenance that concurrently addresses operational reliability and energy consumption. In contrast to previous work, it demonstrates real-time application in an energy-intensive manufacturing context, yielding insights that are directly applicable to similar industrial environments. The empirical focus of this study and its emphasis on deployment at the edge distinguish it from the largely conceptual or simulation-based research prevalent in the field.

This study adopts a data-driven experimental approach to develop and validate an Edge AI-based adaptive maintenance framework designed for energy-intensive manufacturing systems. The methodology is structured in three layers: sensor data acquisition, model deployment at the edge, and decision-making based on real-time predictions.

This study is anchored in an applied, empirical methodology aimed at developing and validating a real-time predictive maintenance model utilizing Edge AI within a steel manufacturing context. The research design integrates live operational data with supervised machine learning techniques and statistical evaluation methods to assess the reliability, responsiveness, and energy impact of the proposed framework. The methodology comprises five core components: data acquisition, model selection and justification, system architecture design, analytical procedures, and model deployment.

A Sensor Adaptation Module (SAM) was introduced to unify modality-specific preprocessing pipelines. SAM ensures consistent feature representation across vibration (FFT-derived spectral peaks), acoustic (denoised intensity descriptors), and thermal (smoothed gradients) sensors, enabling multimodal fusion at the feature engineering stage. By aligning modality-specific features, SAM improves robustness and enhances the reusability of the Edge AI framework.

To address sensor heterogeneity, preprocessing was performed before feature extraction. Vibration signals were transformed using Fast Fourier Transform (FFT) to capture frequency-domain features, acoustic data underwent spectral subtraction to reduce noise, and temperature readings were smoothed using a moving average filter to remove short-term fluctuations. These steps produced harmonized features including spectral peaks, acoustic intensity, and thermal gradients, which were aligned through timestamp synchronization before feeding into the Random Forest model.

The empirical setting for this study is Qatar Steel Company (QSTEEL), one of the largest integrated steel producers in the region. The research team collected a continuous stream of operational data over a six-week period from multiple industrial machines operating under varying load and temperature conditions. The data capture process concentrated on three key categories of sensor inputs: vibration intensity, acoustic emissions, and thermal readings. These inputs were selected due to their established relevance in identifying machine anomalies and performance deterioration in predictive maintenance research (Zhang et al., 2023; Li and Zhao, 2022; Evans and Thomas, 2023).

Concurrently, real-time energy consumption readings were collected at 15-s intervals using embedded power monitoring modules installed in the same equipment. These energy readings were synchronized with sensor data through timestamp alignment, facilitating a unified dataset for analysis and model training (Brown and Green, 2021; Ibrahim and Ali, 2019).

The Random Forest (RF) algorithm was chosen as the core predictive model due to several significant advantages. First, RF can manage heterogeneous data types, missing values, and multicollinearity among variables—conditions frequently observed in industrial datasets (Ahmed and Khan, 2022). Second, the model provides built-in measures of variable importance, enhancing interpretability, which is a crucial requirement for deployment in operational environments where decision-makers must comprehend model behavior (Patel and Desai, 2021). Unlike more opaque deep learning methods, RF offers traceable logic and consistent outputs, increasing its acceptability among engineering teams (Wang and Zhang, 2019).

Furthermore, RF performs effectively with medium-sized datasets without overfitting and requires relatively low computational resources compared to neural networks, rendering it highly suitable for Edge AI deployment where processing capabilities are limited (Zhou and Wang, 2023). The model was configured with 100 decision trees and fine-tuned through grid search cross-validation utilizing a training subset of the collected data (80%), with the remaining 20% reserved for validation.

To optimize the limited computational capacity of edge devices, an adaptive resource allocation strategy was implemented. Vibration data, which requires high-frequency sampling, was dynamically down-sampled during low-load machine states, while acoustic and thermal streams maintained lower constant sampling rates. A task scheduling mechanism prioritizes vibration monitoring during critical operating phases and reallocates computational resources to acoustic or temperature streams when production load decreases. Experimental results showed that this adaptive scheduling reduced average CPU utilization by 27% without compromising the accuracy of predictive maintenance decisions.

The overall framework of the system adheres to a layered structure in alignment with industrial IoT standards. As depicted in Figure 1, the first layer Sensors Layer is responsible for capturing physical signals from production equipment. These raw inputs are transmitted to the Edge AI Layer, where the Random Forest model operates locally on an industrial-grade edge computing device (Johnson and Smith, 2023; Lee and Park, 2021).

The adaptive nature of the proposed maintenance system is operationalized through a rolling retraining schedule. The Edge AI model is updated weekly using the most recent sensor and energy data, which allows the system to learn from evolving machine behavior and adapt prediction logic accordingly. This continuous update cycle ensures that maintenance decisions remain responsive to real-time operational dynamics and equipment conditions.

This configuration minimizes data transmission delays and supports on-site inference without dependence on external networks. The third layer Decision-Making Layer receives classification results that indicate equipment health status. Based on these results, the system either recommends automated maintenance actions or triggers alerts for human intervention. To enable dynamic adaptation, the system continuously updates prediction logic through a weekly retraining process based on the most recent sensor and energy data. A decision score is generated for each instance, and maintenance is triggered when this score exceeds a predefined confidence threshold (e.g., 0.85). This threshold can be adjusted based on feedback from operators, ensuring that the model's decisions remain aligned with evolving operational priorities.

This setup embodies a hybrid model in which AI assists decision-making while retaining final control in the hands of plant operators (Chen and Zhao, 2020).

Direct transmission of raw waveform data was avoided. Instead, statistical and frequency-domain descriptors (RMS, kurtosis, spectral entropy) were extracted locally. A Principal Component Analysis (PCA)-based compression scheme reduced the dimensionality of the feature set before submission to the edge inference layer. This achieved a 68% reduction in data size while retaining more than 90% of variance in the original signals.

Several analytical techniques were employed to evaluate model performance and explore the relationships among variables:

1. Correlation Matrix Analysis: Pearson correlation coefficients were computed between energy consumption and all sensor variables to identify the most impactful predictors (Lee and Kim, 2020).

2. ANOVA (Analysis of Variance): This technique was utilized to determine whether energy consumption levels differ significantly across operational shifts and machine states, providing insights into the variability of energy patterns under different maintenance conditions (Brown and Green, 2021).

3. Prediction Error Analysis: Residuals were plotted to examine bias and variance, while histograms and error distributions helped identify periods of under- or overestimation (Nguyen and Tran, 2022).

4. ROC Curve Analysis: The model's classification performance in detecting critical energy levels was evaluated using ROC curves, with area-under-curve (AUC) values indicating accuracy (Kumar and Singh, 2021).

5. Feature Importance Ranking: RF-generated feature importance scores were employed to rank sensor variables according to their influence on the model's predictions, offering practical guidance for monitoring priorities (Ahmed and Khan, 2022).

The final model was deployed in a simulated edge computing environment that reflects the hardware constraints of actual industrial use. The edge device was configured to retrain the model weekly based on incoming data to ensure adaptability. Maintenance engineers participated in a validation workshop to assess the operational relevance of predictions and provided feedback on the system's responsiveness and interpretability (Gonzalez and Perez, 2021).

The evaluation phase concentrated on three criteria:

1. Predictive accuracy, utilizing standard error metrics (MAE, RMSE)

2. Operational feasibility, based on edge resource constraints

3. Decision support value, as perceived by practitioners in the field (Chen and Zhao, 2020)

This layered and systematic approach ensures that the proposed framework is not only technically sound but also practically viable and aligned with the evolving needs of smart manufacturing systems.

This section presents the empirical findings of the Edge AI-based adaptive maintenance framework, structured in accordance with the conceptual model delineated in Figure 1. The results are categorized into five primary areas: descriptive time-series behavior, variable correlation, operational variance, predictive performance, and model interpretability.

Figure 2 illustrates the temporal evolution of key sensor readings and energy consumption metrics over a six-week observation period. The data reveal significant fluctuations in current intensity, temperature, and vibration levels, particularly during peak production hours. These fluctuations correspond with variations in energy consumption, suggesting a strong temporal correlation between machine performance and power usage. Such dynamic patterns underscore the necessity for real-time, adaptive prediction mechanisms capable of monitoring operational shifts.

The correlation matrix (Figure 3) underscores the strength of relationships between input variables and energy consumption. Current intensity (Current_A) exhibited the most robust positive correlation (r = 0.84), followed by temperature (Temp_C) and vibration intensity. These findings validate the incorporation of these features into the predictive model and corroborate previous studies emphasizing the association between physical stressors and energy demand in industrial machinery (Lee and Kim, 2020).

ANOVA results (Figure 4) confirm statistically significant differences (p < 0.01) in energy consumption across different shifts and operational states. Night shifts showed lower average energy usage, potentially due to reduced load or machine idle time. These insights provide actionable knowledge for shift-level energy planning and maintenance scheduling.

< 0.01) in energy consumption across various shifts and operational states. Night shifts displayed lower average energy usage, potentially attributable to reduced load or machine idle time. These insights provide actionable knowledge for shift-level energy planning and maintenance scheduling.

Prediction error analysis (Figure 5) revealed a tight clustering of residuals around zero, indicating minimal model bias. The histogram of residuals (Figure 6) supports this observation, exhibiting a near-normal distribution with minimal skewness. The ROC curve (Figure 7) for binary classification of high versus low energy consumption yielded an AUC of 0.91, indicating exceptional model discrimination capability. Collectively, these results demonstrate the robustness of the Random Forest model deployed at the edge in accurately forecasting energy levels.

Figure 8 presents the Individual Conditional Expectation (ICE) plot for the most influential variable—Current_A. The ascending trend confirms a monotonic relationship between current intensity and predicted energy consumption. This level of transparency enhances practitioner trust and enables engineers to identify priority areas for intervention. Furthermore, the ranked feature importance scores (not shown) reaffirmed the dominant role of electrical current, followed by acoustic signals and thermal readings.

As summarized in Table 2, the analytical results of the Edge AI-based model are detailed.

The empirical results of this study affirm the effectiveness of integrating Edge AI into adaptive maintenance frameworks for energy-intensive manufacturing systems. The proposed Random Forest model demonstrated high predictive accuracy, with an AUC of 0.91 and minimal residual bias. These findings validate the model's capability to forecast energy consumption with precision, supporting real-time decision-making at the edge.

The strong correlation between current intensity (r = 0.84) and energy usage confirms the role of electrical load as a primary driver of energy demand, consistent with prior research highlighting the relationship between physical stressors and power consumption (Lee and Kim, 2020). The ICE plot further established a transparent, monotonic relationship between this variable and energy forecasts, enhancing the interpretability of the model for maintenance engineers. The statistically significant results from ANOVA indicate that operational shifts influence energy usage, providing actionable insights for maintenance planning and resource allocation (Brown and Green, 2021).

Compared to previous work, this study provides a more comprehensive and pragmatic integration of predictive maintenance and energy forecasting. For instance, while Zhang et al. (2023) and Wang and Zhang (2019) demonstrated the potential of Edge AI for maintenance, their findings were based on simulated environments. In contrast, this study employs real-time data from a steel manufacturing facility, reinforcing its practical relevance. Similarly, Kumar and Singh (2021) explored AI for energy optimization but did not include maintenance as a decision variable. The current work bridges that gap by jointly optimizing both energy and equipment performance within an Edge AI framework.

Ahmed and Khan (2022) validated the robustness of Random Forest in predictive maintenance, but their model lacked real-time deployment. This study extends their findings by embedding the algorithm into an edge computing context, thus addressing both latency and autonomy challenges. The integration of decision support mechanisms further aligns with recommendations from Chen and Zhao (2020), who emphasized the need for responsive, on-site systems in Industry 4.0.

This study diverges from prior research by empirically validating the integration of Edge AI, predictive maintenance, and energy optimization in an operational steel plant environment. Unlike simulation-based models or cloud-centric implementations, our framework demonstrates how retrainable machine learning deployed at the edge can directly support energy-aware maintenance decisions.

From a theoretical perspective, this study contributes to the literature by operationalizing a hybrid model that simultaneously addresses predictive maintenance and energy management. It also reinforces the value of explainable AI in industrial settings, particularly when model transparency is essential for trust and adoption (Patel and Desai, 2021). The findings advocate for the inclusion of energy performance metrics in adaptive maintenance models, an area that remains underrepresented in current research (Nguyen and Tran, 2022).

Practically, the proposed model offers a decision support tool that can be deployed on-site with minimal infrastructure. Its ability to retrain weekly ensures adaptability to changing production conditions, while its interpretability enables actionable recommendations for maintenance teams. The model's success in a real-world steel manufacturing context signals its potential for broader application across other energy-intensive sectors seeking to implement Industry 4.0 solutions.

This study proposes and validates a practical Edge AI-based adaptive maintenance model aimed at enhancing operational performance and reducing energy consumption within energy-intensive manufacturing environments. Utilizing real-time data from Qatar Steel Company, the model exhibits high predictive accuracy and effective interpretability, facilitating timely maintenance interventions and supporting energy-aware decision-making at the edge.

The results underscore a strong correlation between machine behavior variables particularly electrical current and energy demand, thereby affirming their significance in predictive modeling. The Random Forest algorithm has been identified as well-suited for the edge computing context due to its robustness, efficiency, and transparency. These attributes render the model not only technically sound but also practical for real-world deployment by maintenance teams possessing limited machine learning expertise.

This study advances theoretical understanding by bridging the gap between adaptive maintenance and energy optimization, two domains that have historically been approached in isolation. Furthermore, it provides empirical evidence for the feasibility of edge-deployed AI solutions in smart manufacturing, thereby contributing to the literature on Industry 4.0 implementation strategies.

In practical terms, the proposed framework equips manufacturing firms with a decision support tool capable of minimizing unplanned downtime and aligning maintenance operations with energy efficiency objectives. Its adaptability, validated through weekly retraining, ensures sustained relevance in dynamic industrial environments.

This research establishes a foundation for future inquiries into integrated, sustainable manufacturing strategies, particularly those that leverage explainable AI to balance operational efficiency, energy consumption, and long-term system resilience.

This study has several limitations that should be acknowledged. First, the empirical analysis was limited to a single steel manufacturing facility, which may constrain the generalizability of the findings to other industrial sectors or production environments. The data collection period was also relatively short, capturing only six weeks of operational activity, which may not reflect seasonal or long-term variations in machine behavior and energy usage. Additionally, while the Random Forest algorithm provided reliable and interpretable results, the model was not benchmarked against more advanced deep learning architectures or hybrid approaches that could potentially offer higher accuracy.

Future research should aim to validate the proposed model across diverse manufacturing contexts, including different industries, machine types, and operational scales. Longitudinal studies involving extended deployment and performance tracking would provide deeper insights into the robustness and adaptability of the Edge AI framework. Furthermore, integrating additional variables such as environmental factors, operator actions, or real-time production scheduling could improve model precision. In addition, while the six-week data collection period provided initial insights into adaptive maintenance behavior, it may not be sufficient to capture seasonal variations or long-term degradation patterns. Future studies should consider extended deployment periods to validate the long-term adaptability and robustness of the proposed system.

Researchers are also encouraged to explore multi-objective optimization models that balance energy efficiency with maintenance costs, and to incorporate carbon footprint metrics to align with broader sustainability goals in Industry 4.0 initiatives.

The author declares that no specific funding was received for the completion of this research. There are no known conflicts of interest associated with this publication. Ethical approval was not applicable, as the study did not involve human participants or animals. The data utilized were collected from an industrial facility under confidentiality agreements and are not publicly available; access may be provided upon reasonable request and with institutional authorization. All stages of the research including conceptual design, data analysis, model implementation, and manuscript writing were conducted by the author.