Leveraging Generative AI for Modelling and Optimization of Maintenance Policies in Industrial Systems

Academic Editor: Gabriel Luque Received: 4 February 2025 Revised: 28 February 2025 Accepted: 9 March 2025 Published: 11 March 2025 Citation: Crespo Márquez, A.; Pérez Oliver, D. Leveraging Generative AI for Modelling and Optimization of Maintenance Policies in Industrial Systems. Information 2025,16, 217. https://doi.org/10.3390/ info16030217 Copyright: © 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/ licenses/by/4.0/). Article Leveraging Generative AI for Modelling and Optimization of Maintenance Policies in Industrial Systems Adolfo Crespo Márquez 1,* and Diego Pérez Oliver 2 1Department of Industrial Management, University of Sevilla, 41004 Sevilla, Spain 2Asociación para el Desarrollo de la Ingeniería de Mantenimiento (INGEMAN), 41092 Sevilla, Spain; perezoliver[email protected] *Correspondence: [email protected] Abstract: This paper explores how generative AI can enhance the modelling and optimization of maintenance policies by incorporating real-time problem-solving techniques into structured maintenance frameworks. Maintenance policies, evolving from simple calendar-dependent or age-dependent preventive maintenance strategies to more complex approaches involving partial system replacement, minimal repairs, or imperfect maintenance, have traditionally been optimized based on minimizing costs, maximizing reliability, and ensuring operational continuity. In this work, we leverage AI models to simulate and analyze the implementation and overlap of different maintenance strategies to an industrial asset, including the combined use of different preventive (total and partial replacement) and corrective actions (minimal repair and normal repairs), with perfect or imperfect maintenance results. Integrating generative AI with well-established maintenance policies and optimization criteria, this paper tries to demonstrate how AI-assisted tools can model maintenance scenarios dynamically, learning from predefined strategies and improving decision-making in real-time. Python-based simulations are employed to validate the approach, showcasing the benefits of using AI to enhance the flexibility and efficiency of maintenance policies. The results highlight the potential for AI to revolutionize maintenance optimization, particularly in single-unit systems, and lay the groundwork for future studies in multi-unit systems. Keywords: maintenance policies; maintenance optimization; generative AI; simulation 1. Introduction Maintenance policies have long been a critical component of industrial asset management, ensuring that systems continue to operate efficiently while minimizing downtime and operational costs. Traditional maintenance approaches—ranging from corrective maintenance (performed after failures) to preventive strategies like ageor condition-based maintenance—have evolved to include more sophisticated models that optimize resource use, asset reliability, and cost efficiency. The optimization of these maintenance policies has typically been driven by well-established mathematical models and simulation techniques. These models are often designed to minimize expected maintenance costs over a finite period while ensuring that reliability and operational and eco-efficiency targets are met [ 1 ]. However, while effective, these traditional approaches may be limited in flexibility and adaptability, often requiring manual intervention and domain expertise for adjustment and re-formulation [2]. The advent of generative artificial intelligence (AI) introduces new possibilities to the optimization and management of maintenance policies. Unlike traditional AI, which often Information 2025,16, 217 https://doi.org/10.3390/info16030217 Information 2025,16, 217 2 of 23 focuses on pattern recognition and predictive analysis, generative AI can interpret, modify, and even generate new models of optimization based on industry-standard frameworks. This ability to understand and build upon existing optimization models allows for dynamic problem-solving, enhancing both the formulation of maintenance strategies and their real-time application [3,4]. Generative AI not only helps optimize maintenance policies but also offers a suite of other capabilities. It can adapt existing models to specific constraints, explore alternative formulations for finite or infinite time periods, and generate insightful visualizations of the output, allowing decision-makers to better understand the implications of various maintenance strategies [ 5 ]. Moreover, AI-assisted models can suggest innovative ways to overlap different policies, combining and creating hybrid strategies tailored to specific operational needs [ 5 ]. Therefore, the novelty of this research lies in the application of generative AI to dynamically model, optimize, and interact with maintenance policies. Specifically, the research demonstrates how generative AI can accomplish the following: 1. Understand and re-formulate traditional maintenance models: instead of simply optimizing predefined models, generative AI can adapt, restructure, and improve models based on real-time data or shifting operational needs [4]. 2. Combine multiple policies dynamically: it offers a unique capability to overlap various maintenance strategies (e.g., preventive overhauls, corrective interventions, minimal repairs, imperfect maintenance, etc.), creating hybrid strategies that are more adaptable to industrial complexities [5]. 3. Offer enhanced interaction with optimization models: by enabling users to interact with the AI, it allows for adjustments not just in policy but also in the visualization of outcomes, the re-formulation of the model in finite time periods, and other exploratory tasks that go beyond traditional optimization techniques [5]. 4. Real-time adaptability: the AI’s ability to learn and refine strategies dynamically allows for real-time problem-solving in maintenance operations, unlike static models that require frequent manual updates [4]. This integration of generative AI with maintenance policy optimization represents a novel approach to industrial asset management, providing greater flexibility, deeper insights, and more comprehensive solutions to complex maintenance challenges. 2. The Maintenance Policy Concept In the context of maintenance and dependability standards, such as IEC Electropedia 192-06-02:2015 [ 6 ], a maintenance policy is a comprehensive framework that delineates maintenance objectives, lines of maintenance, indenture levels, maintenance levels, maintenance support, and their interrelationships. Maintenance objectives articulate the goals and desired outcomes of maintenance activities, such as maximizing uptime or minimizing costs. The line of maintenance specifies the organizational segment responsible for executing maintenance activities. Indenture levels define the hierarchical structure of system components and sub-components, which influences maintenance execution at various levels. Maintenance levels categorize different types of maintenance activities, including routine, intermediate, and depot-level maintenance. Maintenance support encompasses the necessary resources and infrastructure, such as tools, personnel, and spare parts, required to conduct maintenance activities. In the academic literature, however, a maintenance policy is often more narrowly defined as a set of predefined rules and strategies that dictate how maintenance activities are conducted on a system, whether single unit or multi-unit, to ensure its reliability and operational efficiency [7–14]. Some examples of this, for a single unit, are the following: Information 2025,16, 217 3 of 23 • The age-dependent preventive maintenance (PM) policy for a single unit system is a common approach where a unit is replaced at a predetermined age T or upon failure, whichever occurs first [ 7 ]. This policy has evolved to include concepts like minimal, imperfect, or perfect repair for CMs [ 15 ] or imperfect maintenance for PMs [ 15 – 17 ], leading to various extensions and modifications (failure limit policy, failure cost limit policy, etc.). • A very useful extension of previous policies was the partial replacement policy [ 18 ], which is based on the idea that in most cases where a system fails, a full replacement of the system is not necessary to restore it to proper operating conditions. Instead, a partial preventive replacement (PPR) of some components is sufficient. PPRs are interventions, often named overhauls in industry, performed when the system reaches a certain age, which restore the system’s failure rate to its initial level. After a certain number of partial preventive replacements (to be determined), the accumulated costs of these interventions may exceed the cost of performing a total replacement of the system. • A random age-dependent maintenance policy addresses scenarios where it is impractical to maintain a unit on a strictly periodic basis. Instead, maintenance is performed at random intervals, taking advantage of available free time [ 8 , 14 ]. In this context, a unit’s age is measured from the time of the last replacement, and minimal repairs are undertaken upon failure, leading to a periodic replacement with a minimal repair at failure policy. • Recently, ref. [ 2 ] introduced an imperfect maintenance policy that defines a reliability improvement factor according to the intervention level of maintenance actions. This assumes the intervention level to be independent of the time between consecutive maintenance actions, removing the common constraint that longer intervals necessitate higher intervention levels. Additionally, it optimizes not only the number of maintenance activities and their timing but also the intervention levels of maintenance activities. While maintenance policies for single-unit systems focus on the timing and type of maintenance actions to ensure reliability and operational efficiency, multi-unit systems introduce additional complexity due to the interdependence between units. This interdependence necessitates a coordinated approach to maintenance, considering the hierarchical structure of components, redundancy, and the interactions between different units. These policies are designed to address the specific challenges of multi-unit systems as a whole, including efficiently managing redundancy and spare parts. Some examples of this, for a single unit, are the following: • Group maintenance: maintenance is conducted simultaneously on a group of units to reduce costs and downtime [1,9]. • Selective maintenance: maintenance prioritizes specific units based on their criticality or condition to optimize overall system reliability [19]. • Condition-based maintenance (CBM): maintenance schedules are based on the real-time monitoring of unit conditions, adapting to the varying states of different units [ 15 , 20 – 22 ]. • Opportunistic maintenance: maintenance takes advantage of the planned downtimes or failures of one unit to perform maintenance on other units, minimizing overall downtime and costs [14,23]. Single-unit and multi-unit maintenance policies also account for different types of cost structures, which can be dependent on the age and number of PMs and/or number of CMs; can neglect maintenance times, or not; and consider, or not, penalties, the residual value of the unit, etc. Information 2025,16, 217 4 of 23 For the purposes of this paper, we will adopt the more focused definition of a maintenance policy as described in the academic literature. This includes the specific strategies for performing maintenance actions based on the age and condition of the unit, incorporating preventive and corrective maintenance actions and considering total or partial replacements and various degrees of repair and PM. Also, we will focus on single-unit systems, with the intention of further similar investigations into multi-unit ones. The selection of maintenance policies in this study is rooted in their historical significance within the field of maintenance modelling. These classical policies have been extensively studied and applied in both academic research and industrial practice, serving as foundational models for optimizing maintenance strategies. By leveraging these well-established frameworks, we demonstrate how AI can reinterpret, model, and integrate them into more complex, data-driven decision-making processes. The evolution of maintenance practices, from basic preventive approaches to more sophisticated conditionbased and hybrid strategies, reflects the continuous advancements in information systems and asset management. Highlighting these traditional models allows us to showcase the adaptability of AI in understanding, formalizing, and enhancing maintenance policies, while ensuring that our findings remain grounded in widely recognized principles of maintenance optimization. For readers seeking further insights into advanced AI-driven maintenance strategies, we recommend consulting recent studies that explore predictive and AI-enhanced maintenance frameworks. Gawde et al. [ 24 ] present explainable AI techniques for the predictive maintenance of rotating machines, emphasizing interpretability in AI decision-making. Yang et al. [ 25 ] discuss AI-driven operations and maintenance within telecommunication networks, showcasing AI’s role in optimizing large-scale infrastructure. Kirschbaum et al. [ 26 ] provide insights into AI-driven maintenance support for downhole tools, demonstrating applications in oil and gas industries. Additionally, Dhyani [ 27 ] explores AI-driven predictive maintenance strategies for manufacturing plants, offering valuable perspectives on failure prediction. These works complement our study by providing broader context and application-oriented discussions of AI-driven maintenance solutions. 3. Criteria for Maintenance Policies’ Optimization One of the most prevalent optimization criteria in maintenance policies is the minimization of costs. This criterion involves reducing the total expected cost of maintenance over a specified period, considering both preventive and corrective maintenance actions. For example, in the age-dependent preventive maintenance policy discussed by [ 7 ], the goal is to find the optimal replacement age TT that minimizes the combined cost of scheduled replacements and unscheduled failures. Similarly, ref. [ 9 ] focus on minimizing maintenance costs by determining the optimal intervals for preventive maintenance and the degree of repair required after failures. Another important optimization criterion is the maximization of system reliability and availability. This involves ensuring that the system operates without failures for as long as possible, thereby increasing its uptime and reducing downtime. The authors of ref. [ 15 ] emphasize the importance of reliability in their models, incorporating concepts of minimal, imperfect, and perfect maintenance to optimize maintenance schedules and actions. These policies are critical in environments where operational continuity is essential. Policies are optimized to ensure that maintenance activities cause minimal disruption to production processes. This involves the strategic scheduling of maintenance tasks during non-peak hours or integrating predictive maintenance techniques to anticipate and prevent failures before they occur [28,29]. Information 2025,16, 217 5 of 23 In some cases, optimization criteria also include risk and safety considerations. This involves ensuring that the maintenance policy not only minimizes costs and maximizes reliability but also reduces the risk of catastrophic failures and ensures the safety of operations. The authors of ref. [ 14 ] highlight the importance of considering the risk of failure and the potential consequences in their maintenance models. By incorporating risk and safety factors, they aim to develop maintenance policies that are not only cost-effective but also safe and reliable. Maintenance policies are designed to mitigate risks associated with equipment failures and ensure compliance with safety regulations, thereby protecting both personnel and assets [30,31]. Recent studies have highlighted the importance of considering energy consumption and greenhouse gas (GHG) emissions in maintenance policy optimization. By integrating these environmental factors, organizations aim to reduce their carbon footprint and improve sustainability. This approach often balances the trade-off between maintenance costs and environmental impact, promoting more eco-friendly maintenance strategies [1]. Some frameworks assess the long-term financial impact of maintenance policies by calculating their net present value (NPV). This approach evaluates the overall financial benefits of a maintenance policy over its lifecycle, considering both immediate costs and long-term savings. Following this criterion, organizations can justify maintenance investments that offer significant returns over time [32]. Finally, there is always the possibility of using multi-criteria decision-making models, incorporating several criteria simultaneously. These models help in developing maintenance policies that optimize multiple objectives, ensuring a more balanced and comprehensive approach. For instance, the use of MCDM allows maintenance managers to prioritize actions based on the criticality of equipment, potential safety hazards, and operational impact, rather than solely focusing on costs [10,13]. For the purposes of this paper, we will use maintenance policy optimization criteria based on improving the expected total cost of the policy. This approach does not preclude the possibility of conducting the same study using other optimization criteria, ensuring the generality of our findings. 4. The Problem Statement The idea of this section is to investigate some distinct use cases where generative AI can deliver actionable insights, optimize maintenance plans, and facilitate the assessment of performance enhancements. Through a rigorous examination of varied prompts, we aim to demonstrate the capacity of generative AI not only to streamline existing processes but also to contribute to the ongoing refinement of maintenance strategies. For this exercise, we use CHAT GPT-4o [ 33 ] as the chatbot or conversational AI system and Jupyter Notebook (which is a web-based, interactive computing notebook environment using Python 3) as the open-source interactive web application to create live code, equations, visualizations, and narrative texts [34]. We will be presenting chat prompts or statements or questions provided to the chatbot to initiate a conversation or to request specific information or assistance. A prompt serves as an input for the AI model to generate a relevant response or carry out a particular task. Chat prompts are commonly used in natural language processing applications, chat-based interfaces, and virtual assistants to facilitate human–computer interactions. Users can input prompts to engage in a dialogue, ask questions, or seek assistance, and the AI model responds based on its training and understanding of the provided input. After this introduction, let us continue with our prompts. The idea when building the prompts is to arrive at a formulation that generates a suitable answer when prompted. Information 2025,16, 217 6 of 23 This is sometimes not an easy task. The user must make sure that in rephrasing the prompt, containing similar structured information, the answer is similar. In the first prompt, we provide the chatbot with the data needed for a unit reliability analysis for a given failure mode. We assume an existing PM cycle (25 calendar days) before questioning about the adequacy of that policy. Therefore, this first prompt is only meant to explore the failure behavior. Notice that the number of total days of operation is not needed (redundant data). Also, we are not questioning the goodness of fit to save space in this section. The prompt is as follows: •Prompt 1. Fitting the failure data to a Weibull distribution using the MLE. Suppose we are operating a component, under the assumption that this component maintenance program is in place, employing a calendar-based schedule with a preventive maintenance (PM) cycle of 25 calendar days, this program was decided according to initial guesstimates about preventive and failure/corrective operations’ cost. Additionally, we are assuming there is no significant downtime for either preventive or corrective maintenance, or, in other words, we disregard it in comparison to the time between maintenance operations. After 120 days of operation (starting 1 January), a specific number of failure incidents (measured in operating time to failure in days) and preventive replacements (measured in operating time to replacement in days) have been recorded. The total days of operation (125) is the sum of all the times recorded. Prior to questioning the appropriateness of our current PM cycle (25 calendar days), we have opted to investigate the failure behavior. Failure data (time to failure data) is: 10, 12, 18, 13 and 23. Censored data (time to preventive maintenance) is: 15, 13, 7, 12 and 2. With this data we would like to generate a histogram of failure and censored data (in different color) (See Figure 1). Fit the Weibull distribution: Use the maximum likelihood estimator to fit a Weibull distribution to the pooled data. Interpret the results: Once the fit is complete, we print out the parameters of the Weibull fit for review and analysis. After doing the above, present the plot of the reliability and CFD functions, with the time to failure, x of the graph. Information 2025, 16, x FOR PEER REVIEW 6 of 23 After this introduction, let us continue with our prompts. The idea when building the prompts is to arrive at a formulation that generates a suitable answer when prompted. This is sometimes not an easy task. The user must make sure that in rephrasing the prompt, containing similar structured information, the answer is similar. In the first prompt, we provide the chatbot with the data needed for a unit reliability analysis for a given failure mode. We assume an existing PM cycle (25 calendar days) before questioning about the adequacy of that policy. Therefore, this first prompt is only meant to explore the failure behavior. Notice that the number of total days of operation is not needed (redundant data). Also, we are not questioning the goodness of fit to save space in this section. The prompt is as follows: • Prompt 1. Fitting the failure data to a Weibull distribution using the MLE. Suppose we are operating a component, under the assumption that this component maintenance program is in place, employing a calendar-based schedule with a preventive maintenance (PM) cycle of 25 calendar days, this program was decided according to initial guesstimates about preventive and failure/corrective operations’ cost. Additionally, we are assuming there is no significant downtime for either preventive or corrective maintenance, or, in other words, we disregard it in comparison to the time between maintenance operations. After 120 days of operation (starting 1 January), a specific number of failure incidents (measured in operating time to failure in days) and preventive replacements (measured in operating time to replacement in days) have been recorded. The total days of operation (125) is the sum of all the times recorded. Prior to questioning the appropriateness of our current PM cycle (25 calendar days), we have opted to investigate the failure behavior. Failure data (time to failure data) is: 10, 12, 18, 13 and 23. Censored data (time to preventive maintenance) is: 15, 13, 7, 12 and 2. With this data we would like to generate a histogram of failure and censored data (in different color) (See Figure 1). Fit the Weibull distribution: Use the maximum likelihood estimator to fit a Weibull distribution to the pooled data. Interpret the results: Once the fit is complete, we print out the parameters of the Weibull fit for review and analysis. After doing the above, present the plot of the reliability and CFD functions, with the time to failure, x of the graph. Figure 1. Graphical representation of the scenario described for Prompt 1. This is the answer of the chatbot (GPT-4o): Here is the graph displaying the histograms and density of failure and censored data, along with the reliability and cumulative distribution functions fitted with a Weibull distribution. 1. Blue: Histogram and density of failure data. 2. Green: Histogram and density of censored data. 3. Red Line: Cumulative distribution function. 4. Orange Dashed Line: Reliability function. Figure 1. Graphical representation of the scenario described for Prompt 1. This is the answer of the chatbot (GPT-4o): Here is the graph displaying the histograms and density of failure and censored data, along with the reliability and cumulative distribution functions fitted with a Weibull distribution. 1. Blue: Histogram and density of failure data. 2. Green: Histogram and density of censored data. 3. Red Line: Cumulative distribution function. 4. Orange Dashed Line: Reliability function. The Weibull distribution parameters are: Information 2025,16, 217 7 of 23 •Shape (β): 3.38 •Scale (η): 18.50 This visualization helps in understanding the data distribution and the fitted Weibull model for better maintenance scheduling and reliability analysis. (See Figure 2). Information 2025, 16, x FOR PEER REVIEW 7 of 23 The Weibull distribution parameters are: • Shape (𝛽): 3.38 • Scale (𝜂): 18.50 This visualization helps in understanding the data distribution and the fitted Weibull model for better maintenance scheduling and reliability analysis. (See Figure 2). Figure 2. Graph containing the results of Prompt 1. 5. Optimization Using Total Replacement Policies Now, let us scrutinize the appropriateness of our existing PM cycle (25 calendar days), which necessitates a standard risk cost–benefit analysis. To facilitate this analysis, it is imperative to incorporate information pertaining to costs. In addition to this, we need to state the maintenance approach we want to assess and the criteria for optimization we wish to apply. We understand that accurately formulating the optimization problem mathematically is crucial. Here is the prompt: • Prompt 2. Properly formulating the maintenance optimization problem. Solve a maintenance optimization problem that involves determining the best maintenance policy to apply to a failure mode and the optimal preventive maintenance cycle time (tp) that characterizes that policy. Data to consider for the failure mode: - Equipment reliability: Weibull (3.38, 18.5) - Cp: Preventive maintenance cost = 50 - Cc: Corrective maintenance cost = 250 Optimization criterion: Minimize the expected total maintenance cost CTE (tp). Maintenance policies to compare: Calendar-based vs. operating time-based CTE (tp) for the calendar-based policy: CTE (tp) = (Cp + Cc × N(tp))/tp with N(tp) = integral of λ(t) from 0 to tp; λ(t) is the failure rate CTE (tp) for the operating time-based policy: CTE (tp) = (Cp × R(tp) + Cc × F(tp))/(tp × R(tp) + M(tp)) where: F(tp) is the cumulative failure probability at time tp, i.e., the integral of f(t) from 0 to tp; f(t) is the probability density function of failures. Figure 2. Graph containing the results of Prompt 1. 5. Optimization Using Total Replacement Policies Now, let us scrutinize the appropriateness of our existing PM cycle (25 calendar days), which necessitates a standard risk cost–benefit analysis. To facilitate this analysis, it is imperative to incorporate information pertaining to costs. In addition to this, we need to state the maintenance approach we want to assess and the criteria for optimization we wish to apply. We understand that accurately formulating the optimization problem mathematically is crucial. Here is the prompt: •Prompt 2. Properly formulating the maintenance optimization problem. Solve a maintenance optimization problem that involves determining the best maintenance policy to apply to a failure mode and the optimal preventive maintenance cycle time (tp) that characterizes that policy. Data to consider for the failure mode: - Equipment reliability: Weibull (3.38, 18.5) - Cp: Preventive maintenance cost = 50 - Cc: Corrective maintenance cost = 250 Optimization criterion: Minimize the expected total maintenance cost CTE (tp). Maintenance policies to compare: Calendar-based vs. operating time-based CTE (tp) for the calendar-based policy: CTE (tp) = (Cp + Cc ×N(tp))/tp with N(tp) = integral of λ(t) from 0 to tp; λ(t) is the failure rate CTE (tp) for the operating time-based policy: CTE (tp) = (Cp ×R(tp) + Cc ×F(tp))/(tp ×R(tp) + M(tp)) where: F(tp) is the cumulative failure probability at time tp, i.e., the integral of f(t) from 0 to tp; f(t) is the probability density function of failures. Information 2025,16, 217 8 of 23 R(tp) is the survival probability at time tp, i.e., R(tp) = 1 −F(tp). M(tp) = integral of t * f(t) from 0 to tp. The chatbot answer in Figure 3is very interesting. We can check that the model, as used by the AI, is the correct one for each maintenance strategy, for each probability function, and for the optimization criteria. Information 2025, 16, x FOR PEER REVIEW 8 of 23 R(tp) is the survival probability at time tp, i.e., R(tp) = 1 − F(tp). M(tp) = integral of t * f(t) from 0 to tp. The chatbot answer in Figure 3 is very interesting. We can check that the model, as used by the AI, is the correct one for each maintenance strategy, for each probability function, and for the optimization criteria. Figure 3. Chatbot problem formulation (original output format). In the successive Prompt 3, we enable the tool to handle problem resolution, referring to the anticipated outcomes that we have condensed for space in this section. See Figure 4 for the graph generated by the code provided by the chatbot. All this specific information is provided in the prompt, and this is the result: • Prompt 3. Ordering to solve the maintenance optimization problem. I concur with the formulation. Now, compute CTE(tp) for each specified policy, plot them for different tp values, and identify the tp that yields the minimum CTE value for each policy. Mark these values on the graph. GPT-4o’s answer is as follows: Figure 3. Chatbot problem formulation (original output format). In the successive Prompt 3, we enable the tool to handle problem resolution, referring to the anticipated outcomes that we have condensed for space in this section. See Figure 4 for the graph generated by the code provided by the chatbot. All this specific information is provided in the prompt, and this is the result: •Prompt 3. Ordering to solve the maintenance optimization problem. Information 2025,16, 217 9 of 23 Information 2025, 16, x FOR PEER REVIEW 9 of 23 Figure 4. Python generated graph for the optimization problem solution. Once the tool demonstrates a solid grasp of the problem and a preferred (dominant in this case) maintenance strategy is chosen, we can explore additional potential scenarios to enhance decision-making in the event of changes in failure distribution and/or future changes in corrective maintenance costs. Here are the prompts (see graphical results in Figures 5 and 6), now paying more attention to results, given code obtention: • Prompt 4. Gaining knowledge to face potential failure scenarios. Is it possible to have a graph showing the sensitivity of these optimal intervals with respect to variations of beta values? GPT-4o’s answer is as follows: Minimum CTE vs. Beta. Here are the sensitivity analysis graphs: • Optimal tp vs. β: This graph shows how the optimal preventive maintenance interval (tp) changes with different values of β for both calendar-based and operating timebased policies. Both policies show a decrease in optimal tp as β increases, reflecting the higher failure rates associated with higher β values. • Minimum CTE vs. β: This graph illustrates how the minimum expected total maintenance cost (CTE) varies with different values of β. The operating time-based policy consistently results in a lower minimum CTE compared to the calendar-based policy across the range of β values. These graphs in Figure 5 provide a clear visualization of how the optimal maintenance strategy and costs are sensitive to changes in the Weibull shape parameter (β). If you have any further questions or need additional analyses, please let me know! Figure 4. Python generated graph for the optimization problem solution. I concur with the formulation. Now, compute CTE(tp) for each specified policy, plot them for different tp values, and identify the tp that yields the minimum CTE value for each policy. Mark these values on the graph. GPT-4o’s answer is as follows: Once the tool demonstrates a solid grasp of the problem and a preferred (dominant in this case) maintenance strategy is chosen, we can explore additional potential scenarios to enhance decision-making in the event of changes in failure distribution and/or future changes in corrective maintenance costs. Here are the prompts (see graphical results in Figures 5and 6), now paying more attention to results, given code obtention: •Prompt 4. Gaining knowledge to face potential failure scenarios. Information 2025, 16, x FOR PEER REVIEW 10 of 23 Figure 5. Optimal tp when Weibull Beta changes. • Prompt 5. Gaining more knowledge to face potential cost scenarios. Is it possible now to have a graph showing the sensitivity of these optimal intervals with respect to variations in corrective cost from 50 to 500? This is GPT-4o’s answer: Figure 6. Optimal tp when Cc changes. The graphs in Figure 6 provide insights into how the optimal maintenance intervals and costs respond to changes in the corrective maintenance cost. The results in Figures 5 and 6 leave us anything but indifferent. The possibility of optimizing the problem repeatedly under different conditions, obtaining the plot, and visualizing the pattern of solution behavior with the effort invested is simply brilliant. 6. Optimization Using Partial Replacement Policies A dialog to transfer the principles of this maintenance policy model is as follows: • Prompt 6. Transferring principles of partial replacement policy. We will now model partial replacement policies, these models are based in the idea that on many occasions a system fails does not need to be completely replaced (ST) to come back to normal operating conditions, only a partial preventive replacement (SPP) is needed. SPPs are overhauls after a certain time interval (with length Ti), setting the system to as an “as good as new” situation. After a certain number of overhauls, these will be more costly than a total system replacement. The subindex i measures the number of periods before the overhaul, with i = 1 to k, with k − 1 as maximum number of overhauls Figure 5. Optimal tp when Weibull Beta changes. Is it possible to have a graph showing the sensitivity of these optimal intervals with respect to variations of beta values? GPT-4o’s answer is as follows: Information 2025,16, 217 16 of 23 Information 2025, 16, x FOR PEER REVIEW 16 of 23 Figure 8. Sample output of the finite-time problem simulation in Prompt 11. 9. The AI Generated Code per Case Study—Considerations This paper focuses on the use of ChatGPT [33] as an assistant for the development and analysis of maintenance models. As a software, it is relevant to perform a quality analysis of the interaction with it and its capabilities. A critical approach is required to ensure that the defined models are valid, up to date, and applicable in the different contexts that may arise. With this in mind, we have dipped into the ISO 25000 series (SQuaRE—System and Software Quality Requirements and Evaluation) [35], specifically into the 25010 series (Quality Model Division). ISO 25010 proposes the following criteria to be analyzed to evaluate a code, concerning the possible outcomes resulting from the interaction with it, within its “Quality in Use” model (see Figure 9 and some examples of applications in [36] or [37]): • Effectiveness: the accuracy and completeness with which users achieve specified goals. • Efficiency: resources expended in relation to the accuracy and completeness with which users achieve goals. • Satisfaction: the degree to which user needs are satisfied when a product or system is used in a specified context of use. • Freedom from risk: the degree to which a product or system mitigates the potential risk to economic status, human life, health, or the environment. • Context coverage: the degree to which a product or system can be used with effectiveness, efficiency, freedom from risk, and satisfaction in both specified contexts of use and in contexts beyond those initially explicitly identified. Figure 8. Sample output of the finite-time problem simulation in Prompt 11. 9. The AI Generated Code per Case Study—Considerations This paper focuses on the use of ChatGPT [ 33 ] as an assistant for the development and analysis of maintenance models. As a software, it is relevant to perform a quality analysis of the interaction with it and its capabilities. A critical approach is required to ensure that the defined models are valid, up to date, and applicable in the different contexts that may arise. With this in mind, we have dipped into the ISO 25000 series (SQuaRE—System and Software Quality Requirements and Evaluation) [ 35 ], specifically into the 25010 series (Quality Model Division). ISO 25010 proposes the following criteria to be analyzed to evaluate a code, concerning the possible outcomes resulting from the interaction with it, within its “Quality in Use” model (see Figure 9and some examples of applications in [ 36 ] or [37]): • Effectiveness: the accuracy and completeness with which users achieve specified goals. • Efficiency: resources expended in relation to the accuracy and completeness with which users achieve goals. • Satisfaction: the degree to which user needs are satisfied when a product or system is used in a specified context of use. • Freedom from risk: the degree to which a product or system mitigates the potential risk to economic status, human life, health, or the environment. • Context coverage: the degree to which a product or system can be used with effectiveness, efficiency, freedom from risk, and satisfaction in both specified contexts of use and in contexts beyond those initially explicitly identified. Information 2025,16, 217 17 of 23 Information 2025, 16, x FOR PEER REVIEW 17 of 23 Figure 9. SQuaRE “Quality in Use” evaluation criteria. A qualitative assessment of the code generated by ChatGPT has been carried out using this framework. The intention of this chapter is not to perform a measured evaluation of ChatGPT’s code but to give insights into possible problems that may arise and general guidance on where to look, so users are critical when using AI to solve these kinds of problems. As will be explained below, the capabilities of the generated code can vary from case to case, depending on the complexity of the assignment, but will be highly influenced by the specificity of the prompts provided by the user. In the end, the expectations from the user can also affect the evaluation of the obtained code; while some would expect any mathematical function to be defined as a code function with all its parameters defined as function inputs without the need of specifying it (even when there are no indications pointing to variation in some of the parameters), for others, this would be irrelevant. Quality-in-Use Assessment • Effectiveness: Concerning goal’s satisfaction, the chatbot is proven useful when it comes to assessing simple tasks. It has to be noted that the model might usually behave in a narrow-minded way. For example, when defining functions, parameters such as alpha or costs will not be included as function inputs but as simple constants, unless explicitly told to do so. When it comes to more complex tasks, the model may start to provide wrong answers. One example is Prompt 11; when asked to develop a fairly simple simulation model, several attempts are necessary until a decent solution is achieved—the main failures in this case were wrong logical rules when inputting costs, a misunderstanding of concepts clearly stated when deterministic parameters, such as scheduled preventive maintenance time intervals, was randomized, and a not-so-reliable method for generating random time intervals. Similarly, when performing stochastic simulations, the chatbot does not consider the necessity of statistical analysis to obtain reliable results, unless explicitly told to. • Efficiency: The time consumption of the chatbot’s processing is negligible. However, there are some further issues that need to be addressed when it comes to resource consumption. First of all, it is important to consider that many companies will not be able to/want to upload sensitive data to an external cloud-based data system, such as OpenAI’s ChatGPT platform. Running extensive AI models might require high-end hardware, posing an elevated cost that might not be so efficient if this is the only task it will be used for. Also, it is necessary to provide very specific and detailed instructions, and colloquial speech should be avoided at any costs—this usually leads to the need of thoroughly review the written prompts. If personnel with decent programming skills is available, the same task could be achieved as fast as with the chatbot Figure 9. SQuaRE “Quality in Use” evaluation criteria. A qualitative assessment of the code generated by ChatGPT has been carried out using this framework. The intention of this chapter is not to perform a measured evaluation of ChatGPT’s code but to give insights into possible problems that may arise and general guidance on where to look, so users are critical when using AI to solve these kinds of problems. As will be explained below, the capabilities of the generated code can vary from case to case, depending on the complexity of the assignment, but will be highly influenced by the specificity of the prompts provided by the user. In the end, the expectations from the user can also affect the evaluation of the obtained code; while some would expect any mathematical function to be defined as a code function with all its parameters defined as function inputs without the need of specifying it (even when there are no indications pointing to variation in some of the parameters), for others, this would be irrelevant. Quality-in-Use Assessment • Effectiveness: Concerning goal’s satisfaction, the chatbot is proven useful when it comes to assessing simple tasks. It has to be noted that the model might usually behave in a narrow-minded way. For example, when defining functions, parameters such as alpha or costs will not be included as function inputs but as simple constants, unless explicitly told to do so. When it comes to more complex tasks, the model may start to provide wrong answers. One example is Prompt 11; when asked to develop a fairly simple simulation model, several attempts are necessary until a decent solution is achieved—the main failures in this case were wrong logical rules when inputting costs, a misunderstanding of concepts clearly stated when deterministic parameters, such as scheduled preventive maintenance time intervals, was randomized, and a not-so-reliable method for generating random time intervals. Similarly, when performing stochastic simulations, the chatbot does not consider the necessity of statistical analysis to obtain reliable results, unless explicitly told to. • Efficiency: The time consumption of the chatbot’s processing is negligible. However, there are some further issues that need to be addressed when it comes to resource consumption. First of all, it is important to consider that many companies will not be able to/want to upload sensitive data to an external cloud-based data system, such as OpenAI’s ChatGPT platform. Running extensive AI models might require high-end hardware, posing an elevated cost that might not be so efficient if this is the only task it will be used for. Also, it is necessary to provide very specific and detailed instructions, and colloquial speech should be avoided at any costs—this usually leads to the need of thoroughly review the written prompts. If personnel with decent programming skills is available, the same task could be achieved as fast as with the chatbot but Information 2025,16, 217 18 of 23 avoiding any extensive hardware requirements and the use of the AI itself, as the task is fairly simple. • Satisfaction: Personal user satisfaction has been evaluated from two different Points of View: a skilled programmer perspective and a non-skilled programming one. On the one hand, for a person with little knowledge of programming, the use of the chatbot is highly satisfactory, as it leverages programming tools overtaking the lack of knowledge. On the other, a person with enough programming knowledge to develop these models in Python might find uncomfortable the use of the chatbot because of several reasons: these are, for example, the necessity of explicitly mentioning rather trivial things, such as the inclusion of parameters as inputs rather than as constants of functions. It is very important to consider trust implications. Users may trust or distrust the chatbot because of very different reasons. If a person with enough programming knowledge considers the source code provided simple or notices flaws, they might feel discouraged to continue using the tool. However, the chatbot can be asked to provide the mathematical formulation being used, so it can be checked and can improve users trust when simply prompting the chatbot to develop models (not source codes). Nevertheless, this option is not always suitable (it might not be as simple to mathematically formulate and revise a simulation model). • Freedom from risk: As stated before, the chatbot might sometimes commit errors that can only be checked by looking at the source code that the chatbot generated and used to accomplish the task. This means that a user unexperienced in coding might consider these as valid wrong solutions, posing a potential risk. This also leads to an ironic situation where a programmer might be requested to check the source code and carry out necessary corrections but could have been tasked before to develop the model from scratch, making this potentially more efficient. • Context coverage: Chatbots like ChatGPT allow the user to interact using different languages—this entails that language barriers might be easily surpassed when the user’s mother tongue is supported. Also, it is possible to generate models dynamically and update input data through interaction with the chatbot. However, it has been proven that in certain cases, the chatbot might give wrong answers, leading to potential risks. Also, the nuances defining the barrier between a good or a bad answer might be unclear; therefore, the context coverage might be considered as unclear, or null in the worst case, based on the inability of potential users to judge the correctness of answers. 10. Conclusions Regarding the Use of AI for Modelling Policies The use of generative AI like GPT for generating and optimizing maintenance policies offers several advantages (See Table 1). First, GPT can dynamically create tailored maintenance strategies that adjust to varying scenarios, combining different policy types, such as minimal repairs, preventive overhauls, and corrective interventions. This flexibility allows for hybrid strategies that better reflect the complexity of real-world maintenance needs. Additionally, GPT can autonomously re-formulate models based on operational changes, ensuring that maintenance policies remain up to date and are optimized. Another key advantage is the ability to analyze policies over finite periods and explore probabilistic outcomes using Monte Carlo simulations, which can improve decision-making by offering insights into potential risks and costs. AI-driven models also provide improved visualizations and interaction, allowing users to simulate, visualize, and adjust strategies as needed. Ultimately, this results in better cost optimization, risk mitigation, and the ability to implement data-driven maintenance decisions more efficiently. Information 2025,16, 217 19 of 23 Table 1. Advantages and disadvantages found when using AI for modelling maintenance policies. Advantages Disadvantages - Dynamic model generation: GPT adapts to various scenarios. - Data dependency: results depend on the quality of training data. - Combination of maintenance policies into hybrid strategies. - Complexity of real-world scenarios: models may miss key nuances. - Improved decision support through simulations and visualizations. - Computational cost: high cost for generating and updating models. - Automation of model re-formulation based on changing conditions. - Interpretability: AI models may be difficult to interpret. - Handles finite-time and real-time problem-solving. - Risk of overfitting: models may be too specific for general use. - Cost optimization and risk mitigation. - Lack of physical constraints: generated models may be impractical. - Provides visualization and interaction for better decision-making. - Dependence on human input for final decisions and validation. On the other hand, several challenges exist when applying GPT in this context. One key limitation is that the quality of AI-generated models depends heavily on the availability and accuracy of training data. If the data are incomplete or biased, this may lead to unreliable or overly simplified policies. Additionally, complex real-world scenarios may require more sophisticated modelling, which AI systems may oversimplify or fail to account for. While AI models are flexible, they can also be computationally expensive to run, especially when combined with techniques like Monte Carlo simulations. Furthermore, the interpretability of GPT-generated models may pose a challenge for domain experts, as the reasoning behind the AI’s decisions might be difficult to trace. Finally, there is a risk of overfitting to specific scenarios, making the models less effective in broader applications. Despite these challenges, the potential benefits of using generative AI in maintenance policy development make it a powerful tool for industrial optimization. 11. Conclusions Regarding the Code Generated The overall findings indicate that ChatGPT’s code generation capabilities can be highly beneficial for relatively simple programming tasks, especially for users with limited technical knowledge. In such cases, the tool allows for the rapid development of basic models or prototypes, thereby reducing the initial effort needed to begin coding. However, when confronted with more complex or specialized tasks, ChatGPT often omits crucial parameters or treats them inappropriately, for instance, by fixing them as constants rather than incorporating them as variables. It also tends to produce logical inconsistencies, particularly in cost modelling, scheduled maintenance intervals, and stochastic simulations, where it does not account for the necessary statistical analysis unless prompted explicitly. From the perspective of “Quality in Use” as articulated by ISO 25010, ChatGPT is effective at providing guidance for solution development but requires meticulous prompt specifications to ensure completeness and accuracy. Insufficiently detailed prompts can result in code that is incomplete or erroneous, posing risks for non-expert users. While the efficiency of ChatGPT in terms of response time is high, concerns remain regarding the management of sensitive data on external servers and the potential costs of implementing on-premise hardware solutions. User satisfaction, moreover, depends heavily on programming expertise. Novices may find it highly satisfactory, as ChatGPT delivers functional Information 2025,16, 217 20 of 23 code quickly, whereas experienced programmers may find the need to explicitly spell out trivial details and verify or correct the resulting code burdensome. Despite the broad applicability of ChatGPT across different contexts and languages, its reliability varies, particularly when users are not equipped to detect potential errors. This variability increases the risk of integrating unverified code in critical environments. As a result, any code produced by ChatGPT should be perceived as an initial draft rather than a definitive solution. Review and, if necessary, refinement by domain experts or skilled programmers are essential to ensure that the final implementation is robust and aligns with the intended purpose. In sum, ChatGPT offers valuable support for rapid prototyping and instructional activities. However, its limitations underscore the importance of active expert oversight, the rigorous validation of generated outputs, and careful consideration of the context in which the code will be deployed. This balanced approach preserves the tool’s advantages while minimizing the risks associated with incomplete or erroneous solutions (See Table 2). Table 2. Summary of pros and cons found in the quality of ChatGPT code generation. Pros Cons - Rapid development of basic models or prototypes - Omission of crucial parameters - Ease of use for non-experts - Logical inconsistencies - Efficiency in response time - Dependency on prompt specificity - Applicability across contexts and languages - Incomplete or erroneous code - Instructional support - Reliability concerns - User satisfaction for novices - Burden for experts - Cost-effective for prototyping - Security risks - Drafting support - Requires expert oversight 12. Limitations and Future Work This study relies on a hypothetical scenario rather than real industrial data. While this approach allows for controlled experimentation, it does not capture the full complexity of real-world maintenance environments. Future research should focus on validating AI-generated maintenance policies using historical maintenance records. These datasets would provide valuable insights into failure patterns, repair histories, and operational constraints, improving AI-driven optimization techniques. 12.1. Risks of AI-Driven Maintenance Optimization Despite its advantages, AI-driven maintenance modelling presents key risks: • Oversimplification of constraints—AI may overlook factors like workforce availability, spare parts, or regulatory requirements, leading to impractical recommendations. • Limited consideration of critical variables—AI may fail to account for industry-specific factors that influence maintenance decisions. • Bias in decision-making—unbalanced training data may cause AI to favor certain maintenance strategies while neglecting others. • Suboptimal recommendations—AI-generated policies may appear optimal in simulations but fail when applied to real-world conditions. To mitigate these risks, AI-generated policies should be reviewed by human experts, integrated into expert-driven frameworks, and refined using real-world data. Testing AI under different operational conditions would help ensure its reliability. Information 2025,16, 217 21 of 23 12.2. Computational Considerations and Deployment Costs AI-driven maintenance modelling requires significant computational resources. Unlike traditional methods, AI relies on large datasets and iterative processing, which increase infrastructure demands. • Processing power and data requirements—AI models require high-performance computing for training and optimization. Their accuracy depends on access to high-quality maintenance records, which may not always be available. • Cloud vs. on-premise deployment—cloud-based AI offers scalability but raises security concerns, while on-premise solutions provide more control but demand greater infrastructure investment. • Computational cost vs. optimization accuracy—While AI can refine maintenance strategies, running complex models may not always justify the computational expense. Balancing cost and performance is crucial for practical adoption. Despite these challenges, advancements in cloud computing, edge AI, and hybrid optimization approaches make AI-driven maintenance increasingly feasible. Future research should explore cost-effective AI implementations that integrate expert knowledge while ensuring scalability and efficiency. 12.3. Future Directions Key areas for further research may include the following: validation with realworld data—testing AI-generated maintenance policies on industrial datasets; hybrid AI-optimization models—combining AI insights with traditional techniques for better decision-making; and computational efficiency—reducing the computational cost of AI models for real-time applications. If these challenges are addressed, AI-driven maintenance optimization can evolve into a practical tool for industrial asset management, improving reliability and cost efficiency. Author Contributions: Conceptualization, A.C.M.; Software, D.P.O.; Validation, D.P.O.; Formal analysis, D.P.O.; Investigation, A.C.M. and D.P.O.; Resources, A.C.M.; Data curation, D.P.O.; Writing— original draft, A.C.M. and D.P.O.; Writing—review & editing, A.C.M. All authors have read and agreed to the published version of the manuscript. Funding: This work has been developed as part of the AMADIT Project (PID2022-137748OB-C32), funded by MCIN/AEI/10.13039/501100011033/FEDER, EU. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author. Conflicts of Interest: The authors declare no conflict of interest. References 1. Shafiee, M.; Finkelstein, M. An optimal age-based group maintenance policy for multi-unit degrading systems. Reliab. Eng. Syst. Saf. 2015,134, 230–238. [CrossRef] 2. Pereira, F.H.; Melani, A.H.d.A.; Kashiwagi, F.N.; da Rosa, T.G.; dos Santos, U.S.; de Souza, G.F.M. Imperfect Preventive Maintenance Optimization with Variable Age Reduction Factor and Independent Intervention Level. Appl. Sci. 2023,13, 10210. [CrossRef] 3. Haridasan, P.K.; Kumar, V. Generative AI in Manufacturing: A Review of Innovations, Challenges, and Future Prospects. J. Artif. Intell. Mach. Learn. 2024,2, 1–15. [CrossRef] [PubMed] 4. Mohapatra, A. Generative AI for Predictive Maintenance: Predicting Equipment Failures and Optimizing Maintenance Schedules Using AI. Int. J. Sci. Res. Manag. (IJSRM) 2024,12, 1648–1672. [CrossRef] Information 2025,16, 217 22 of 23 5. ´ Celi´c, J.; Bronzin, T.; Horvat, M.; Jovi´c, A.; Stipi´c, A.; Prole, B.; Mariˇcevi´c, M.; Pavlovi´c, I.; Pap, K.; Mikota, M.; et al. Generative AI in E-maintenance: Myth or Reality? Int. J. Adv. Comput. Sci. Appl. 2024,15, 45–52. 6. Electropedia 192-06-02:2015; Maintenance Policy. IEC: Geneva, Switzerland, 2015. Available online: https://www.electropedia. org/iev/iev.nsf/display?openform&ievref=192-06-02 (accessed on 3 February 2025). 7. Barlow, R.; Hunter, L. Optimum Preventive Maintenance Policies. Oper. Res. 1960,8, 90–100. [CrossRef] 8. Chang, C.C. Optimum preventive maintenance policies for systems subject to random working times, replacement, and minimal repair. Comput. Ind. Eng. 2014,67, 185–194. [CrossRef] 9. Cho, D.I.; Parlar, M. A survey of maintenance models for multi-unit systems. Eur. J. Oper. Res. 1991,51, 1–23. [CrossRef] 10. Ding, S.H.; Kamaruddin, S. Maintenance policy optimization—Literature review and directions. Int. J. Adv. Manuf. Technol. 2014, 76, 1263–1283. [CrossRef] 11. Journal, E.; Pham, H.; Wang, H. Imperfect maintenance. Eur. J. Oper. Res. 1996,94, 425–438. 12. Pierskalla, W.P.; Voelker, J.A. A survey of maintenance models: The control and surveillance of deteriorating systems. Nav. Res. Logist. Q. 1976,23, 353–388. [CrossRef] 13. de Jesus Santos, A.C.; Cavalcante, C.A.V.; Wu, S. Maintenance policies and models: A bibliometric and literature review of strategies for reuse and remanufacturing. Reliab. Eng. Syst. Saf. 2023,231, 108983. [CrossRef] 14. Valdez-Flores, C.; Feldman, R.M. A survey of preventive maintenance models for stochastically deteriorating single-unit systems. Nav. Res. Logist. (NRL) 1989,36, 419–446. [CrossRef] 15. Wang, H.; Pham, H. Optimal maintenance policies for several imperfect repair models. Int. J. Syst. Sci. 1996,27, 543–549. [CrossRef] 16. Coster, M.; Liu, Q. An Imperfect Preventive Maintenance Policy Considering Reliability Limit. J. Eng. Res. Appl. 2015,5, 49–56. 17. Nakagawa, T. Optimum policies when preventive maintenance is imperfect. IEEE Trans. Reliab. 1979,R-28, 331–332. [CrossRef] 18. Nguyen, D.G.; Murthy, D.N.P. Optimal Preventive Maintenance Policies for Repairable Systems. Oper. Res. 1981,29, 1181–1194. [CrossRef] 19. Xu, Q.Z.; Guo, L.M.; Wang, N.; Fei, R. Recent advances in selective maintenance from 1998 to 2014. J. Donghua Univ. (Engl. Ed.) 2015,32, 986–994. 20. Jafari, L.; Naderkhani, F.; Makis, V. Joint optimization of maintenance policy and inspection interval for a multi-unit series system using proportional hazards model. J. Oper. Res. Soc. 2017,69, 36–48. [CrossRef] 21. Tian, Z.; Liao, H. Condition based maintenance optimization for multi-component systems using proportional hazards model. Reliab. Eng. Syst. Saf. 2011,96, 581–589. [CrossRef] 22. Wakiru, J.; Pintelon, L.; Muchiri, P.N.; Chemweno, P.K.; Mburu, S. Towards an innovative lubricant condition monitoring strategy for maintenance of ageing multi-unit systems. Reliab. Eng. Syst. Saf. 2020,204, 107200. [CrossRef] 23. Laggoune, R.; Chateauneuf, A.; Aissani, D. Opportunistic policy for optimal preventive maintenance of a multi-component system in continuous operating units. Comput. Chem. Eng. 2009,33, 1499–1510. [CrossRef] 24. Gawde, S.; Patil, S.; Kumar, S.; Kamat, P.; Kotecha, K.; Alfarhood, S. Explainable Predictive Maintenance of Rotating Machines Using LIME, SHAP, PDP, ICE. IEEE Access 2024,12, 29345–29361. [CrossRef] 25. Yang, Y.; Yang, S.; Zhao, C.; Xu, Z. TelOps: AI-Driven Operations and Maintenance for Telecommunication Networks. IEEE Commun. Mag. 2024,62, 104–110. [CrossRef] 26. Kirschbaum, L.; Roman, D.; Singh, G.; Bruns, J.; Robu, V.; Flynn, D. AI-Driven maintenance support for downhole tools and electronics operated in dynamic drilling environments. EEE Access 2020,8, 78683–78701. [CrossRef] 27. Dhyani, B. Predicting Equipment Failure in Manufacturing Plants: An AI-driven Maintenance Strategy. Math. Stat. Eng. Appl. 2021,70, 1326–1334. [CrossRef] 28. Chareonsuk, C.; Nagarur, N.; Tabucanon, M.T. A multicriteria approach to the selection of preventive maintenance intervals. Int. J. Prod. Econ. 1997,49, 55–64. [CrossRef] 29. Sharma, A.; Yadava, G.S.; Deshmukh, S.G. A literature review and future perspectives on maintenance optimization. J. Qual. Maint. Eng. 2011,17, 5–25. [CrossRef] 30. Anders, G.J.; Sugier, J. Risk assessment tool for maintenance selection. In Proceedings of the International Conference on Dependability of Computer Systems, DepCoS-RELCOMEX, Szklarska Poreba, Poland, 25–27 May 2006. [CrossRef] 31. Samrout, M.; Châtelet, E.; Kouta, R.; Chebbo, N. Optimization of maintenance policy using the proportional hazard model. Reliab. Eng. Syst. Saf. 2009,94, 44–52. [CrossRef] 32. Marais, K.B.; Saleh, J.H. System value trajectories, maintenance, and its present value. In Safety, Reliability and Risk Analysis: Theory, Methods and Applications—Proceedings of the Joint ESREL and SRA-Europe Conference, 1st ed.; Taylor & Francis Group: London, UK, 2009. [CrossRef] 33. OpenAI. Hello GPT-4o. 2024. Available online: https://openai.com/index/hello-gpt-4o/ (accessed on 1 December 2024). 34. Pérez, F.; Granger, B.E. IPython: A System for Interactive Scientific Computing. Comput. Sci. Eng. 2007,9, 21–29. [CrossRef] Information 2025,16, 217 23 of 23 35. ISO 25000. SQuaRE—System and Software Quality Requirements and Evaluation. Available online: https://iso25000.com/index. php/en/iso-25000-standards (accessed on 3 February 2025). 36. Pereira, D.S.; Bezerra, L.F.V.; Nunes, J.S.; Barroca Filho, I.M.; Lopes, F.A.S. Performance Efficiency Evaluation based on ISO/IEC 25010:2011 applied to a Case Study on Load Balance and Resilient. In Workshop de Testes e Tolerância a Falhas (WTF); Sociedade Brasileira de Computação: Porto Alegre, Brazil, 2023. [CrossRef] 37. Prabowo, W.A. Developing Compliant Audit Information System for Information Security Index: A Study on Enhancing Institutional and Organizational Audits Using Web-based Technology and ISO 25010:2011 Total Quality of Use Evaluation. Int. J. Inform. Vis. 2024,8, 343. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.