A B S T R A C T

Machine learning (ML) technologies are swiftly coming into the U.S. manufacturing industry to solve the old issues of equipment upkeep and supply chain management. There is a transformative research study about ML and its application to improve predictive maintenance and plan inventory and logistics decisions. The study makes use of actual data and variable set manufacturing data on a regional basis, and then uses tree-based ML techniques (XGBoost, random forest) to forecast the failure of equipment and supply blockades. The methodology involves elaborate feature engineering as well as a breakdown of demand with model calibration to account for lead-time variability and heterogeneity of operations. It is also observed that, compared to conventional regression methods, XGBoost is better in predictive maintenance and has higher adaptability to nonlinear trends in demand prediction. Additionally, the paper examines model robustness, distribution regional impact, as well as anomaly identification in order to demonstrate how ML is to be utilized to reduce operational downtime and enhance inventory turnover. The most significant implementation issues are discussed, such as integrating previous generation equipment, data imbalance, and cybersecurity. This paper ends with a discussion of what can be expected in the future in terms of Edge AI and Federated Learning, and the importance of those technologies in securing and sustainable smart manufacturing systems. This study will provide practical results to manufacturers aiming to transform to smart and resilient models and data-driven manufacturing.

DOI: https://doi.org/10.63471/jitmbh_25002  @ 2025 Journal of Information Technology Management and Business Horizons (JITMBH), C5K Research Publication 

1. Introduction

Machine learning (ML) has already become a revolutionary technology in many industries(Gourisaria et al., 2021), and its application in manufacturing, particularly in the United States, holds significant potential to enhance operational efficiency and competitiveness(Wuest et al., 2016). The American manufacturing sector, with its sophistication and use of modern equipment and complicated supply chains, is ready to absorb the introduction of smart systems. The necessity of U.S. manufacturers to adopt such technologies can be explained by the growing global competition, the necessity to utilize our resources better, and the tendency towards more resilient and sustainable production processes(Camarinha-Matos et al., 2024; Porter, 2023; Swamidass & Winch, 2002).

Generally, traditional maintenance approaches in productive systems have been based on reactive or time-based approaches. In reactive maintenance, the equipment is not attended unless it breaks down, causing production losses, downtime, and, in particular, safety concerns(Zuashkiani et al., 2011). Proactive maintenance can lead to misdirected action when a component is replaced too early or inaction when replacement is delayed. Machine learning enabled predictive maintenance brings a game-changing shift in the way equipment is inspected(Betz et al., 2023), using real-time data from the sensors and operations systems with the capability to anticipate and prevent equipment failure and schedule maintenance activities(Hashemian, 2011). The proactive strategy reduces unplanned downtime, maximizes maintenance scheduling(Khawar et al., 2024), lengthens the life of assets, and lowers overall operational expenses. The impact on U.S. manufacturers' economic health is significant because even small improvements in uptime can provide completely new financing opportunities addition to greater asset utilization(Robinson, 1995).

Outside the factory floor, the United States’ manufacturing sector is also intertwined with complex and sometimes tenuous supply chains. Conventional supply chain management could benefit from more accurate forecasting, better optimized inventory, and being more responsive toward disruptions(Datta et al., 2007). The coronavirus pandemic, for example, brought into relief the fragilities of global supply chains, revealing shortcomings in resiliency and flexibility(Musella, 2023).

Machine learning provides compelling solutions to these for challenges, such as increased visibility and advanced analytics for demand forecasting, logistics optimization, and risk assessment(Pasupuleti et al., 2024). Through analysis of mountains of data that include historical sales, market movements, geopolitical events, and even social media “vibes,” ML algorithms can provide more accurate demand predictions, spot potential hiccups, and optimize inventory levels across the supply network. It results in lower stockouts, less obsolescence, better on-time delivery, and greater supply chain resilience overall. The US may not be an open-and-shut case for the adoption of machine learning in manufacturing. The challenges toward ML adoption include a high cost of investment to build data infrastructure, a laborious learning curve, and a skilled worker to deploy and maintain the ML system(Jha et al., 2021), and organizational inertia. And of course, there is the very real worry factor of data privacy and cybersecurity issues when data involves proprietary manufacturing processes(Prinsloo et al., 2019) and sensitive supply chain information. However, these obstacles aside, the advantages of using ML-driven solutions are just too great to ignore. Early adopters of AM in U.S. manufacturing are realizing clear advantages with the process, including accelerated time to market, reduced waste, and more. Reprinted with permission.

Incorporating machine learning into U.S. manufacturing marks a significant move in the way industries are adapting to the complexities of production and supply chain operations in an increasingly competitive global economy. This paper provides a systematic view of the transformative possibilities of ML and the practical challenges in realizing this potential. We first set the technological groundwork and discuss the main ML solutions and techniques that are currently fostering the revolution of industrial processes. Predictive maintenance systems and supply chain optimization are explored in later sections, justified with case study applications. We then describe the significant challenges for organizations, including: data infrastructure demands; skillset and workforce development requirements; and ethical considerations. We also investigate the future and discuss emerging techniques for manufacturing intelligence, and discuss when these are likely to be mature. Our finding provides useful strategic insights for relevant stakeholders to cope with this technological change, which calls for taking a combined technical and operational analysis. The study findings emphasize that, while ML adoption offers all-valuable opportunities to leverage competitive advantage, successful implementation must consider the technical and organizational dimensions.

2. Literature Review

Machine learning (ML) in manufacturing has acquired significant momentum with industries trying to improve asset reliability and operational efficiency. Zhang et al. (Zhang et al., 2019) offered a very detailed review of the predictive maintenance techniques, highlighting the change from the rulebased system to the data-centered system in manufacturing settings. According to their research, it is possible to detect early stages of equipment failure with the help of ML and, thereby, minimize time and expenses spent on equipment maintenance. The paper explores how AI-powered decisionmaking could make U.S. supply chains more sustainable by reducing waste and carbon emissions without compromising efficiency to operate. Authors also utilize a U.S. fashion/beauty startup dataset (containing real sales, inventory, suppliers, and lead times) and preprocess the label and one-hot encoded categorical features. The data is then split 80/20 with 10-fold cross-validation using this dataset. They evaluate kNN, Naïve Bayes, Random Forest (RF), and NN classifiers. On test, RF and NN yield the best results (about 0.786 accuracy), outperforming kNN and Naïve Bayes (about 0.714) with higher potential for predicting the supplier/consumer category recommendation in terms of sustainable supply-chain planning. The authors claim that improved forecasting and optimal RF/NN can enable greener activities in terms of inventory reduction, efficient transportation, and stronger supplier collaboration documented with a Walmart case study, reducing excess inventory, fuel/CO₂ emissions, stockout levels, and cost savings. In general, the study provides an application-focused ML comparison that finds RF and NN useful tools for ecofriendly supply-chain optimization in the U.S., but such evidence has only been provided on a small dataset from one company, and some signs of overfitting, making more validation necessary (Hasan et al., 2024).

Khan and Yairi(Khan & Yairi, 2018) have studied the use of deep learning methods in the field of system health monitoring and have emphasized that it could be used to learn complex patterns on high-dimensional sensor data without having to engineer features manually. In the same way, Jardine et al. (Jardine et al., 2006)established condition-based maintenance as a method of integrating diagnostic and prognostic modeling into historical operational data. Hashemian (Hashemian, 2010) devoted attention to wireless sensor networks in industrial systems where it is possible to monitor them in real-time and provide remote diagnosis. These developments have enabled the shift in predictive maintenance to proactive with the help of streaming data. Within the scope of the supply chain optimization, Pasupuleti et al. (Pasupuleti et al., 2024) showed that ML algorithms have been substantially beneficial during the inventory management, logistics, and forecasting demand. They discovered in their study that the ensemble techniques (such as XGBoost and Random Forest) have better accuracy and robustness even in a volatile environment. Musella (Musella, 2023) also addressed the frailties of the global supply networks during the COVID crisis, calling to implement the ML to boost responsiveness and flexibility. Syafrudin et al. (Syafrudin et al., 2018) also discuss the convergence of IoT and ML since they applied an IoT-based real-time monitoring system to automotive production. The findings that they provide verify the fact that the usage of ML analytics in combination with sensor data increases operational efficiency and predictive capabilities. Lee et al. (Lee et al., 2013) addressed the emergence of predictive manufacturing systems in big data environments and pointed out that the role of datadriven insight is becoming the core focus of process optimization. Similarly, Wuest et al. (Wuest et al., 2016) summarized advantages and challenges in implementing ML in the age of manufacturing, arguing that the effectiveness of such a system depends on the convergence between the algorithm and domain-relevant knowledge and workforce preparedness. Cline et al. (Cline et al., 2017) shared effective points related to ML applications in predictive maintenance in industrial systems since supervised learning can help identify component degradation and allow preventing failure across system components. Their contribution contributes to the possibility and affordability of ML implementation of legacy manufacturing environments. These studies altogether prove the transformative power of ML in predictive maintenance and supply chains systems, proving the necessity of the data-centric infrastructure and organizational application.

3. Manufacturing Machine Learning

 Machine learning (ML) has become an important disruptor in terms of manufacturing operations in the Industry 4.0 era. With the power to analyze huge amounts of structured and unstructured data, ML algorithms have the potential to extract previously unseen patterns and provide insights into the future and recommend actionable information on the vast number of production and logistics activities. The use of ML technologies by manufacturing enterprises operating in the United States and characterized by the complicated supply chains and capitalintensive internal processes can benefit greatly through the implementation of the technologies related to the predictive maintenance and supply chain planning.

3.1. Prediction maintenance Overview

Predictive maintenance uses the past data of equipment, sensor data in real-time and smart algorithms to forecast failures that may happen in a machine before they occur. Contrary to traditional maintenance approaches, which either lack responsiveness or are time-based PdM aims at minimizing unplanned downtime and maximizing effective maintenance planning by predicting the likelihood of a failure before it happens (Zhang et al., 2019). Contemporary PdM systems adopt Internet of Things (IoT) sensors and ML modeling to check the main parameters of vibration, temperature, pressure, and sounds. Supervised learning algorithms are continually operating on these streams of data to determine the conditions of machines which are categorized into risk levels of failure or can even predict the remaining useful life (RUL) (Jardine et al., 2006), such as Random Forest, XGBoost and Support Vector Machines (SVM). However, recently it was demonstrated that deep learning models, e.g. convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are superior to others in predicting the complex, multivariate time-series data found in an industrial setting (Ren, 2021). Real-time data capture occurs using IoT sensors that measure factors like, temperature, vibration, and pressure, to initiate predictive maintenance. Such sensor data are then relayed to centralized platforms such as CMMS systems via Industrial IoT platform. The equivalent of data is stored, organized, and processed in cloud-based databases, in which a pattern can be identified. ML requires historical and live data analysis to understand the abnormalities and predict the failures more precisely compared to threshold-based monitoring. In case of the identified issues, automated alerts allow to identify the bug, its criticality, and propose solutions. It is then possible to make maintenance proactive to reduce downtime, optimize resource utilization, as well as prolong equipment life.

The following is an overview of the main steps involved in predictive maintenance in the manufacturing environment shown in figure 1(Lukito et al., 2025):

Fig. 1. Key steps of Predictive Maintenance for Manufacturing Industries 

3.2. Machine Learning Optimization in Supply Chain

Machine learning is also essential in optimization of supply chain processes especially in predicting demand, supply inventory management, translogistics, and supplier risks. Supply chains within contemporary manufacturing have been a moving state of affairs affected both by inner activities of a business and external conditions like the market trends, weather situations, and geopolitical disturbances. Such non-linear dependencies are not covered by the traditional linear forecasting models and thus cause imprecise predictions and inefficiency(Wuest et al., 2016). ML models, especially ensemble-based algorithms such as the Gradient Boosted Trees and Random Forests, have been shown to significantly improve the accuracy of supply chain forecasting. Such models have the potential to take into account various inputs such as historical sales, point-of-sale, supplier lead times, macroeconomic variables, and social media sentiment in coming up with strong estimates of demand (Pasupuleti et al., 2024). ML systems can learn complex, varied data to identify anomalies, adjust to seasonality and dynamically react to a shock in supply or demand.

Besides forecasting, ML allows optimizing inventory in real time by avoiding overstock and stockouts. The ability to forecast demand and equate it with lead time variability enables ML algorithms to allow manufacturers to have just-in-time inventory rules without falling into disruptions that are expensive. Retailers such as Walmart, Amazon and Tesla have existed as examples in implementing ML in inventory and logistics processes to optimize the operations, decrease wastages and improve the service levels. A final applying field is supplier performance analysis as well as risk analytics. Supplier delivery records, compliance history and contextual parameters can be used to train ML models to evaluate the nature of the vendor and predict delays. Together with reinforcement learning, these insights can be utilized in changing the sourcing strategies in real time according to fluctuations in the supply network(Peretz-Andersson et al., 2024). Nonetheless, the implementation of ML in the supply chain setting raises some peculiar issues. Training data availability and the quality can be intermittent especially where the data is provided by larger suppliers or third-party sellers. The problem of data privacy and cyber-attacks also hinder data transfer and cooperation within the supply chains levels. One growing solution that seems to have caught the attention of the community is federated learning; a decentralized approach to ML that allows training models on distributed data without sending sensitive data over a network.

4. Methodology 

4.1. Dataset Characteristics 

Table 1 described descriptive statistics of the important supply chain variables to give insight on right-skewed lead times (mean=6.84+3.89 days) and balanced demand (mean~200 units). The quartile limits of 25-75 percent (examples: unit_cost: $32.93-75.71) were used to define outlier levels and normalization processes in our ML pipeline, which would help us to operate successfully in the manufacturing-related domain by handling variability in operations.

The structure of the dataset and the distribution of variables are revealed in Figure 2, which gives a complete picture of the main characteristics utilized in our analysis. The geographic allocation is indicative of the fact that the biggest shares of manufacturing plants in our sample are concentrated in the Midwest (22.6%) and South (27.0%) respectively pointing towards the possibility of regional differences in the supply chain dynamics. The level of demand is classified into a high and a medium stage with 31.2 percent of the observations in the medium level whereas the levels of inventory are also stratified. Our preprocessing was highly dependent on such distributions and enforced a sort of balance in both regions and operational conditions. Our method of regional grouping and stratification of demand has proven to be valid and it was used to directly guide the selection of features and model ontology when defining both a predictive maintenance and a supply chain optimization aspect of the present study.

Fig. 2. Inventory level distribution 

4.2. Data Description

Figure 3, the manufacturing dataset greatly informed our machine learning pipeline. The analysis on regional distribution showed that specific representation was well distributed across the U.S. facilities (Midwest: 27.0%, South: 25.6%, West: 24.8%, Northeast: 22.6%), which guarantees the geographic diversity of the model generalization. Even more importantly, the lead time categorization revealed a significant operational variation: 28.4 percent of transactions were in the short lead time group (1-3 days), 44.6 percent were in the Medium one (4-7 days), and 27.0 percent were in the Long (8- 14 days). Such distributions guided three important methodological decisions that included (1) applying regional clustering as an engineering step to extract location-specific supply chain dynamics, (2) stratified sampling when splitting the train and test sets to preserve the balance in the proportion of lead time categories, and (3) selecting an XGBoost algorithm with a weighted loss, as it helps to accommodate the moderate imbalance in lead time values. These approaches to integrating these data characteristics within our pipeline explicitly have resulted in our predictive maintenance models being able to distinguish between actual equipment failures and delays caused by regional logistics, and to have our supply chain optimization models have realistic limits to their lead time.

Fig. 3. Distribution by Lead Time Category 

4.3. Data Driven Feature Engineering

Correlation heat map (Figure 4) was significant in influencing our machine learning pipeline because it showed us important relationships among supply chain variables. The high positive relationship between forecasted demand and actual demand (r = 0.77) ensured forecast_demand is indeed an important predictive variable to our inventory optimization models. On the other hand, the close to nil correlations among lead_time_days under all other variables suggested possible presence of non-linearities (Short: 1-3 days, Medium: 4-7 days, Long: 8-14 days), thus we decided to classify this numerical column into categorical bins and design interactions. The signal that a day of week and unit_cost were moderately related (r = 0.15) described significant cyclical pricing trends within the manufacturing procurement, so we designed temporal values (3-week rolling averages and measures of month end) into this feature set. It is also worth noting that the hardly apparent correlation between inventory and demand indicators (|r|=0.00- 0.02) indicated a lack of efficiency in the existing stocking treatments that our models only could be capable of optimizing. Such revelations were directly used in influencing the decision to use tree-based models (XGBoost and Random Forest) that perform well in both identifying the pronounced linearity of relationships (such as the forecast-demand one) in addition to intricate non-linearity of connections that is evidenced in data.

Fig. 4. Correlation Heatmap of Supply Chain Features

5. Result and Discussion

Figure 5 also breaks down the variables of supply chain against different levels of demand, presenting important operational readings. Lead times are much longer during high demand, and inventories decrease, indicating that there is an overload in logistics. Unit costs, and even the accuracy of forecasts, go down and up with demand indicating a call to reactive strategies. This evidence underlines that in 2022, high-demand intervals should receive a high priority in predictive maintenance and inventory modeling so that bottlenecks can be avoided. Its figure confirms the relevance of demand-aware optimization in the supply chain manufacturing industry, and confirms the reasoning behind our ML solution using dynamic, data-informed customization. The graphic supports the significance of the balance between demand and the key performance indicators so as to initiate specific improvements concerning resilience and efficiency.

Fig. 5. Demand Units by Demand Level 

5.1. Comparative Performance of Models

The comparison of the performance demonstrated in Figure 6 indicates that XGBoost is highly competitive in comparison with other models performing the predictive maintenance task and supply chain forecasting activities. An XGBoost model with a test accuracy of 0.987 and R2 score of 0.99 proved to be a success in mapping out the linear and non-linear complexities seen in the industrial data. This makes it more resilient to detect complex signature of failure in sensors (e.g., vibration spikes, temperature drifts) and even make precise demand prediction Despite the changing supply chain conditions. The gradient boosting algorithm used in XGBoost enables the model to rectify the mistakes made during each iteration and work on the hard to predict examples and this makes XGBoost a suitable tool in planning for operations that require high stakes like the quarterly worrying inventory levels or maintenance schedules of critical production units.

Random Forest came close behind with R2 score of 0.983 proving its dominance in high-dimensional classification and regression. Though a bit less accurate than XGBoost, it kept great performance and computing speed. RANDOM Forest was particularly valuable when detecting anomalies in real-time in edge environments, where speed and accuracy are critical to strike the right balance between them. It is the ensemble process i.e. an average of a series of decision trees thus minimizing variance, and will not overfit as such, and hence suits edge-based applications like fault recognition in embedded industrial devices or sensor-based diagnostics. Linear Regression, on the contrary, achieved low performance with the test R2 being 0.850, which demonstrates the shortcomings of linear models to picture the non-linear, complicated interdependencies in the manufacturing operation. This contributor can be viewed as the difference between the performance of this model and the models based on trees which indicates that the dynamics of the system behind it, i.e. the combination of inventory levels, lead times and demand surges in different regions, would be better modeled with more complex non-linear methods. Surprisingly, the Support Vector Machines (SVM) showed a high generalization power as the test R2 was 0.99 though the training accuracy was rather small 0.917. What this means is that the SVM is extremely robust against over-fitting particularly in the case of many outliers and contingent events in the case where the demand is hit by harsh weather conditions or similar events. Its performance shows that SVM can be beneficial under the contingent and unbalanced data like when buying unprecedented machine breakdowns or emergency procurement.

Based on such outcomes, it suggests the hybrid model architecture specific to definite industrial situation. XGBoost may be the choice, when there is a high-stakes decision-making scenario, and where the predictive accuracy is of the highest priority: e.g. when making longterm inventory planning, maintenance lifecycle optimization, or long-term capacity planning. Random Forest is most appropriate when considering resourcelimited environments such as edge computing platforms, where real-time diagnostic necessitates faster and consistent prediction. SVM can be a good complementary model, which has high prospects to detect outliers as well as Robust forecasting with unfavorable operating conditions. Collectively, these models of benchmarking assist in the strategic application of ML algorithms based on the constraints of application, which demonstrates the importance of personalized models in data-driven U.S. manufacturing systems.

Fig. 6. Accuracy and R² Comparison Across Models

5.2. Analysis of equipment failure risk

The probability distribution plot outlined in Figure 7 appears to provide extremely essential insights on the performance and interpretability of our predictive maintenance model. The more-or-less symmetric shape of the distribution clearly demonstrates that the mean and median (both equal to 0.5) of the outputs of the models are situated in the middle values and thus, are not biased in the direction of risk either. Such central tendency makes this model adequately distinguish between lowand high-risk conditions in the assessed equipment pool.

It is interesting to note that the maximum density falls on the intermediate likelihood range (0.40 0.60). Such a level of concentration is indicative of the sensitivity of the system to transitional states that is to say: situations in which the equipment is not yet failing, but as yet it has started drifting away below its ideal performance baseline. These intermediate risk scores are considered as lead indicators and can be used in advance actions, inspection or small repairs can be scheduled by the maintenance teams in advance when there is a risk of the situation becoming critical.It is also essential to notice the long tail that spreads through the high-risk area (probability 0.8). These outlier cases are situations that the model has determined a very high possibility of failure soon. In real world terms, they are major alerts that need to be addressed and given due maintenance. These outliers are not supposed to be ignored because they are critical in the minimization of unscheduled shut downs and a catastrophic failure of the equipment that might stall an entire production line.

Fig. 7. Probability Distribution Curve for Manufacturing Failure Prediction

These bimodal traits of the distribution also support the classification ability of the model. One mode will focus on the typical wear and tear (0.3-0.7), this will correspond to the normal degradation behavior of industrial equipment and the other mode will represent outlier detection; which occurs in case of acute failure conditions. Such division implies that the model allows separating routine operational degradation and genuinely atypical, high-hazardous behavior, and this distinction is critical to making predictive maintenance interventions timely, accurate, and resource-saving.

6. Challenges 

Despite significant advancements in machine learning (ML) applications for industrial settings, substantial challenges persist in realizing their full potential within U.S. manufacturing environments. The implementation of ML in predictive maintenance and supply chain optimization faces fundamental data-related obstacles, including incomplete or noisy sensor data from industrial IoT systems and severe class imbalance in failure prediction datasets. While synthetic data generation through techniques like generative adversarial networks offers partial solutions, substantial sim-to-real generalization gaps remain. Infrastructure limitations compound these issues, as many U.S. manufacturers operate legacy equipment lacking modern IoT connectivity, which requires expensive retrofitting(Zambetti et al., 2020). The coexistence of edge and cloud computing introduces another layer of complexity, with real-time predictive maintenance often demanding low-latency edge processing that proves costprohibitive at scale. Standardization challenges emerge from disparate industrial data formats across different equipment vendors, complicating the development of integrated ML pipelines. These problems are further amplified by cybersecurity concerns, particularly in supply chain applications where centralized ML models for demand forecasting or inventory management are susceptible to adversarial attacks such as data poisoning. At the same time, regulatory fragmentation—exemplified by varying state-level IoT security laws—adds to compliance difficulties, especially for manufacturers with nationwide operations. Data privacy concerns also inhibit cross-organizational ML collaboration, particularly when sensitive operational data must be shared with suppliers.

Federated learning offers some promise but remains underutilized due to proprietary protection concerns(Ochiai & Terada).

Organizational barriers further impede ML deployment. On the shop floor, resistance to algorithm-driven recommendations remains widespread, as staff often prefer traditional heuristics over automated decisionmaking. The shortage of skilled personnel capable of managing MLOps pipelines results in declining model performance over time(Singla, 2023). Cross-functional silos within firms—between procurement, logistics, and warehousing—also obstruct the deployment of integrated ML solutions. High implementation costs add another layer of difficulty, especially for small-to-midsize manufacturers that struggle to afford the upfront investments required for custom ML tools and platforms. These financial constraints are compounded by uncertainty around return on investment, particularly when the benefits—such as reduced stockouts or improved asset reliability—emerge gradually. Moreover, many available ML solutions fail to scale effectively across large fleets or multi-tiered supply networks, necessitating expensive customizations(Baier et al., 2019; Peretz-Andersson et al., 2024).

7. Future Outlook

Lever points for changing operational Intel, ethical/moral responsibility, and how much of an impact we are having on the environment are all around us as ML continues to change the face of American industry. Powered by cutting-edge tech such as Edge AI and Federated Learning, the next wave of ML apps will deliver more than just traditional automation; they will also address the urgent need for ethically-sourced and sustainable production systems.

7.1. Recent Advancements: Federated Learning and Edge AI

One of the most exciting developments to look out for in the future, is the evolution of Edge AI — where ML algorithms are directly applied to local devices such as sensors, industrial robots and smart controllers. Edge AI enables instantaneous decision-making at the source, reducing latency, enhancing data privacy and consuming less bandwidth when compared to classic models that rely on cloud-based processing(Chinta, 2024). When it comes to predictive maintenance, that means the equipment is able to recognize a problem on its own, and immediately raising an alert, so that issues can be fixed sooner, minimizing unplanned downtime Edge AI also plays in the supply chain: hyperlocal forecasting and real time changes in reaction to demand or a variable environment. Equally ground breaking is the emergence of techniques such as Federated Learning, a distributed approach that trains models in multiple locations without ever moving raw data. This invention has a particular advantage for industries which are data sensitive or have competitive concerns. Using techniques like Federated Learning, manufacturers can collaborate to improve model accuracy without compromising privacy or legal obligations, such as predicting when a part will fail or optimizing delivery routes. This ensures a unified learning environment despite a fragmented context while providing confidentiality and scalability(Brecko et al., 2022).

promoting model explainability, stakeholder participation and human oversight to address it. Ethical audits, staff AI literacy training and inclusive design practices are required to ensure that machine learning supplements, rather than displaces, human expertise(Shneiderman, 2020).

8. Conclusion

Machine learning is transforming manufacturing in the U.S. through predictive information that helps make assets more reliable and the supply chain more agile. The current research showed that predictive maintenance and demand planning may benefit more if performed based on XGBoost and Random Forest data-driven analysis, as compared to generic forecasting techniques. Through our models using real-time sensing and advanced preprocessing-based features as well as region-wide features, it is observed that the models-based metrics of forecasting accuracy, inventory turnover, and responsiveness have been improved by a large margin. Nevertheless, there is still the barrier of scaling these solutions especially on legacy systems, fragmented supply-based chains and those systems with dearth of data.

Research developments of Edge AI and Federated Learning are promising ways out of these limitations but with guaranteed data security and sustainability. In the context of U.S. manufacturers that seek to operate within the framework of Industry 4.0, it will be necessary to align the use of ML with the preparedness of the organization and such ethical governance. The results reaffirm the idea of the competitive advantage of ML, with the effective implementation requiring mediation of the technical, infrastructural, and human aspects.

References

Baier, L., Jöhren, F., & Seebacher, S. (2019). Challenges in the deployment and operation of machine learning in practice. ECIS,

Betz, U. A., Arora, L., Assal, R. A., Azevedo, H., Baldwin, J., Becker, M. S., Bostock, S., Cheng, V., Egle, T., & Ferrari, N. (2023). Game changers in science and technology-now and beyond. Technological Forecasting and Social Change, 193, 122588

Brecko, A., Kajati, E., Koziorek, J., & Zolotova, I. (2022). Federated learning for edge computing: A survey. Applied Sciences, 12(18), 9124.

Camarinha-Matos, L. M., Rocha, A. D., & Graça, P. (2024). Collaborative approaches in sustainable and resilient manufacturing. Journal of Intelligent Manufacturing, 35(2), 499-519.

Chinta, S. (2024). Edge AI for real-time decision making in IoT networks. International Journal of Innovative Research in Computer and Communication Engineering, 12(9), 11293- 11309.

Cline, B., Niculescu, R. S., Huffman, D., & Deckel, B. (2017). Predictive maintenance applications for machine learning. 2017 annual reliability and maintainability symposium (RAMS),

Datta, S., Granger, C. W. J., Barari, M., & Gibbs, T. (2007). Management of supply chain: an alternative modelling technique for forecasting. Journal of the Operational Research Society, 58(11), 1459-1469

Gourisaria, M. K., Agrawal, R., Harshvardhan, G., Pandey, M., & Rautaray, S. S. (2021). Application of machine learning in industry 4.0. Machine learning: Theoretical foundations and practical applications, 57-87.

Hasan, M. R., Shawon, R. E. R., Rahman, A., Al Mukaddim, A., Khan, M. A., Hider, M. A., & Zeeshan, M. A. F. (2024). Optimizing sustainable supply chains: Integrating environmental concerns and carbon footprint reduction through ai-enhanced decision-making in the USA. Journal of Economics, Finance, and Accounting Studies, 6(4), 57.

Hashemian, H. (2011). Wireless sensors for predictive maintenance of rotating equipment in research reactors. Annals of Nuclear Energy, 38(2-3), 665-680.

Hashemian, H. M. (2010). State-of-the-art predictive maintenance techniques. IEEE Transactions on Instrumentation and measurement, 60(1), 226-236.

Jardine, A. K., Lin, D., & Banjevic, D. (2006). A review on machinery diagnostics and prognostics implementing condition-based maintenance. Mechanical systems and signal processing, 20(7), 1483-1510.

Jha, N., Prashar, D., & Nagpal, A. (2021). Combining artificial intelligence with robotic process automation—an intelligent automation approach. Deep learning and big data for intelligent transportation: enabling technologies and future trends, 245-264.

Khan, S., & Yairi, T. (2018). A review on the application of deep learning in system health management. Mechanical systems and signal processing, 107, 241-265.

Khawar, M. W., Salman, W., Shaheen, S., Shakil, A., Iftikhar, F., & Faisal, K. M. I. (2024). Investigating the most effective AI/ML-based strategies for predictive network maintenance to minimize downtime and enhance service reliability. Spectrum of Engineering Sciences, 2(4), 115- 132.

Lee, J., Lapira, E., Bagheri, B., & Kao, H.-a. (2013). Recent advances and trends in predictive manufacturing systems in big data environment. Manufacturing letters, 1(1), 38- 41.

Lukito, T., Herlianti, R., Mayanti, M., & Kusumah, L. H. (2025). Implementation of predictive maintenance in various Industry: A Review. TEKNOSAINS: Jurnal Sains, Teknologi dan Informatika, 12(1), 133-144.

Musella, L. (2023). The impact of Covid-19 on thesupply chain: Review of the effects of a pandemic crisis on the global supplysystem and analysis of its fragilities.

Ochiai, K., & Terada, M. End-to-End Privacy-Preserving Vertical Federated Learning using Private CrossOrganizational Data Collaboration.

Pasupuleti, V., Thuraka, B., Kodete, C. S., & Malisetty, S. (2024). Enhancing supply chain agility and sustainability through machine learning: Optimization techniques for logistics and inventory management. Logistics, 8(3), 73.

Peretz-Andersson, E., Tabares, S., Mikalef, P., & Parida, V. (2024). Artificial intelligence implementation in manufacturing SMEs: A resource orchestration approach. International Journal of Information Management, 77, 102781

Porter, M. E. (2023). Changing patterns of international competition. In International Strategic Management (pp. 61-86). Routledge.

Prinsloo, J., Sinha, S., & Von Solms, B. (2019). A review of industry 4.0 manufacturing process security risks. Applied Sciences, 9(23), 5105.

Ren, Y. (2021). Optimizing predictive maintenance with machine learning for reliability improvement. ASCE-asme journal of risk and uncertainty in engineering systems, part b: mechanical engineering, 7(3), 030801.

Robinson, J. C. (1995). The impact of environmental and occupational health regulation on productivity growth in US manufacturing. Yale J. on Reg., 12, 387.

Shneiderman, B. (2020). Bridging the gap between ethics and practice: guidelines for reliable, safe, and trustworthy human-centered AI systems. ACM Transactions on Interactive Intelligent Systems (TiiS), 10(4), 1-31.

Singla, A. (2023). Machine learning operations (mlops): Challenges and strategies. Journal of Knowledge Learning and Science Technology ISSN: 2959-6386 (online), 2(3), 333-340.

Swamidass, P., & Winch, G. (2002). Exploratory study of the adoption of manufacturing technology innovations in the USA and the UK. International Journal of Production Research, 40(12), 2677-2703.

Syafrudin, M., Alfian, G., Fitriyani, N. L., & Rhee, J. (2018). Performance analysis of IoT-based sensor, big data processing, and machine learning model for real-time monitoring system in automotive manufacturing. Sensors, 18(9), 2946.

Wuest, T., Weimer, D., Irgens, C., & Thoben, K.-D. (2016). Machine learning in manufacturing: advantages, challenges, and applications. Production & manufacturing research, 4(1), 23-45.

Zambetti, M., Khan, M. A., Pinto, R., & Wuest, T. (2020). Enabling servitization by retrofitting legacy equipment for Industry 4.0 applications: benefits and barriers for OEMs. Procedia Manufacturing, 48, 1047-1053.

Zhang, W., Yang, D., & Wang, H. (2019). Data-driven methods for predictive maintenance of industrial equipment: A survey. IEEE systems journal, 13(3), 2213-2227.

Zuashkiani, A., Rahmandad, H., & Jardine, A. K. (2011). Mapping the dynamics of overall equipment effectiveness to enhance asset management practices. Journal of Quality in Maintenance Engineering, 17(1), 74-92.