Condition-Based Maintenance of Degrading Assets underIntermittent Accessibility

arXiv:2609.38474 · eess.SY, cs.SY · Submitted 2026-09-29 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Condition-Based Maintenance of Degrading Assets underIntermittent Accessibility".

Dev: Many maintenance models implicitly assume that maintenance can be performed whenever intervention is warranted, but in practice, environmental uncertainty can make maintenance opportunities intermittent and dynamically evolving.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So, we're looking at this paper today about "Condition-Based Maintenance of Degrading Assets under Intermittent Accessibility," which sounds really relevant to how we handle things in the field. I'm curious if these ideas actually translate well outside of a controlled lab setting, and how long these policies can reliably run before they start failing in real-world conditions.

Dev: I think that’s a big question, Rosa; from my side, I'm focused on the operational reality—the loop rates and how latency affects decision-making when accessibility is uncertain. If the system has to wait for an opportunity, that uncertainty directly impacts how fast we can react to failure modes.

Taro: I'm interested in what happens when the world misbehaves, because this paper introduces stochastic evolution for accessibility, which means things aren't just degrading smoothly; they get stuck waiting for good weather or access. It makes the asset management problem much more dynamic than traditional models suggest.

Rosa: Exactly; it moves us away from simple schedules and toward something that reacts to the environment in real-time, whether that's a sudden change in accessibility or a slow drift in condition. This paper tackles the complexity of maintenance opportunities being intermittent, which is a huge practical hurdle for field operations.

Dev: And when you look at the formulation as a finite-state Markov decision process, it sounds like we’re dealing with states that change based on two things: the asset's physical condition and its current accessibility status. That coupling is what makes this model interesting from an engineering standpoint because it introduces dependencies between those two factors.

Taro: That coupling is where the autonomy aspect comes in for me; when you talk about the accessibility state evolving stochastically over time, that means the system needs to anticipate future opportunities, which is a core challenge for any autonomous agent operating in a changing environment. How does it handle that probabilistic leap?

Rosa: The authors show that instead of just having one fixed rule for maintenance, the optimal threshold actually changes depending on whether the asset is currently accessible or not, which means the decision-making logic has to adapt instantly to the current state of access. This adaptability is what makes this paper’s approach potentially very useful in dynamic settings.

Dev: That dependency on accessibility states directly impacts the cost function because you have to account for fixed intervention costs that depend on that accessibility state, which changes with every decision epoch. I'm thinking about how we can implement a system where the decision logic shifts instantly when sensor data indicates a change in access conditions.

Taro: If the asset is in an inaccessible state, the paper suggests you have to account for not just condition but also when another maintenance opportunity might exist later; that implies a strategic patience that goes beyond just fixing things when they hit a certain wear level. It’s about managing the wait time effectively.

Title and authors: Rosa: Right, so it’s not just about whether the asset is broken enough; it's about whether we are allowed to fix it right now, and if not, how much value we place on waiting for a better window. This accessibility-dependent threshold policy is what they arrive at as the optimal structure.

Dev: I see the implication for latency here; if the system relies on predicting future accessibility states to decide whether to maintain now or wait, any lag in that prediction could lead to suboptimal decisions regarding when to intervene or continue operating. We need a very tight control loop for this kind of foresight.

Taro: And when we consider the structural results, they establish that for every feasible state, there's a threshold above which intervention is optimal, but this threshold itself is governed by the accessibility state w, meaning it’s not universal. This suggests a highly tailored approach rather than a one-size-fits-all rule.

Rosa: That tailoring means we can use richer data about the environment to make smarter choices, rather than just relying on the asset's internal degradation metrics alone. It shifts the focus from purely condition monitoring to integrating environmental foresight into maintenance planning.

Dev: The paper also outlines a specific condition for when preventive maintenance is optimal: Q P(w) at most Q zero(w, x) or when inaccessibility occurs, which gives us concrete mathematical criteria to follow during operation. This helps define the boundary conditions for our control logic.

Taro: Those concrete criteria are important because they give us a way to quantify the trade-off between immediate efficiency loss and the risk of missing a future intervention window based on that stochastic evolution of accessibility. It frames the decision mathematically.

Rosa: It seems like this paper offers a very robust mathematical framework for integrating these two uncertainties—degradation and accessibility—into one cohesive policy structure, which is exactly what we need when deploying complex systems outside the lab.

Dev: I'm just wondering about the assumptions they made, because they rely on monotone degradation and cost conditions; if the real world has non-monotone wear or costs that change drastically with environment, does this optimal threshold structure still hold up? That’s a point for me to push on for robustness.

Taro: If we think about misbehaving worlds, we have to consider what happens when those assumptions break; if degradation isn't monotone, the relationship between condition and cost might become much more chaotic, which could invalidate the simple threshold ordering they establish.

Rosa: That’s a fair challenge; the real world is rarely perfectly smooth in its wear or its access patterns, so we have to keep an eye on those boundary conditions where this model might not apply as cleanly.

Dev: So, moving into the quantitative results, the numerical studies motivated by offshore wind turbines show that incorporating condition and accessibility information can reduce long-run average cost by about thirty-three point nine five percent compared to just a condition-based policy. That’s a substantial saving if those assumptions hold true in practice.

Title and authors: Taro: That reduction shows the economic benefit of using this more complex model; it validates the effort to incorporate environmental uncertainty into maintenance planning when dealing with assets that have intermittent access, which is a major factor in offshore environments.

Rosa: And then there's the four point three one percent additional cost reduction they found when adapting that threshold specifically to accessibility, showing that knowing *when* you can fix it is just as valuable as knowing *how degraded* it is. That really highlights the importance of the accessibility state information itself.

Dev: From a loop rate perspective, this suggests that incorporating an accessible-state-dependent threshold into our decision algorithm requires careful engineering to ensure we process that accessibility data fast enough to be relevant before a maintenance window closes or opens. The latency around those external environmental updates matters a lot here.

Taro: Thinking about the broader impact, if we can effectively model and exploit this dynamic relationship between degradation and intermittent access, it could lead to much more efficient long-term management of large, aging infrastructure globally that faces unpredictable environmental challenges.

Rosa: So to wrap up the core idea of this paper, it’s that the optimal threshold isn't a single number but a vector dependent on the accessibility state w, because failing to adapt it can actually increase costs. This moves us toward much smarter, context-aware maintenance scheduling.

Dev: We should keep thinking about how to implement those accessibility state transitions within our real-time control framework so that we can effectively use this policy structure without introducing unacceptable delays in the action itself. That’s my focus for the next iteration of the design.

Taro: For me, the implication is that future research needs to focus on how these dynamic thresholds interact when multiple assets are competing for limited maintenance resources, which is a real-world scenario we haven't explored here.

Rosa: So, we’ve seen that this paper provides a solid mathematical structure for handling intermittent access in condition-based maintenance, showing how the threshold varies based on accessibility. It’s definitely something to keep studying as we think about deploying these kinds of adaptive systems in complex field robotics applications.

Dev: I agree; the quantitative reduction figures are compelling evidence that this level of modeling is justified when dealing with assets where access is a known stochastic variable. We need to ensure our implementation can handle those state transitions reliably.

Taro: And looking ahead, we should definitely follow up on how this framework scales when you move from a single asset to managing a whole fleet where accessibility patterns are correlated across different assets. That’s the next logical step for autonomy research.

Rosa: It’s clear that understanding how environmental uncertainty shapes maintenance opportunities is key to achieving truly optimized long-term operation, and this paper gives us the mathematical tools to get there. We'll be looking for papers that build on this dynamic threshold concept next.

The paper's summary: Rosa: So, to quickly recap, this paper shows that instead of just having one universal rule for maintenance based on how degraded an asset is, the best approach is to have a different rule depending on whether or not we can actually get there to fix it right now.

Dev: That's exactly right; the main finding boils down to this accessibility-state-dependent threshold policy, meaning the optimal point to intervene shifts based on current access conditions. It moves us away from that one static condition metric and makes it much more context-aware.

Taro: I think what's really important is how this handles the uncertainty of when we can act; it doesn't just look at the asset’s wear, but also how likely we are to find a good weather window in the future. That forward-looking aspect is where it gets interesting for autonomy.

Rosa: It's definitely that forward-looking element; they show that by incorporating this accessibility state into the decision, you can actually reduce costs significantly compared to just using a simple condition-based schedule. The numerical results point to substantial savings when you factor in these environmental dynamics.

Dev: I agree with Rosa on the cost reduction; those figures suggest that the information about accessibility is a very strong driver for economic efficiency, but we still need to figure out how fast the AI can process that state change so it doesn't miss a critical window. The authors define specific conditions for when preventive maintenance becomes optimal based on those accessibility states, which gives us some concrete rules to follow.

Taro: That concrete rule definition is key because it tells us exactly when we should be patient versus when we have to act immediately, especially in those situations where the asset might be accessible but the weather outlook is poor. It means the system has a strategy for managing that waiting time effectively based on probability.

Rosa: And that's what makes me wonder about its real-world application; if this works well in a lab setting with clean transition matrices, how long can we trust this policy to hold up when the actual weather patterns or degradation rates in the field are much messier?

Dev: That’s a valid concern for me; they rely on assumptions like monotone degradation, and if those assumptions break down in a real-world scenario—say, sudden extreme events—the whole threshold ordering might get skewed. We need to test how sensitive this policy is to those non-monotone changes.

Taro: If the world misbehaves and those assumptions fail, the system needs a robust fallback mechanism; it can't just rely on a mathematically derived optimal path if the underlying model of reality is wrong. That’s where we need to push for more adaptive behavior when uncertainty spikes.

Rosa: So, it seems like this paper gives us a strong foundation for building maintenance logic that isn't just reactive to failure but anticipates both physical wear and environmental opportunity. It definitely has the potential to make asset management much smarter in harsh conditions.

Dev: I think the real power here is taking that structural result—that threshold changes with accessibility—and making it operational within a fast loop rate, which is still a major engineering challenge for any decision-making system based on stochastic processes.

Taro: And looking at the broader picture, if we can successfully deploy this kind of dynamic planning, it could fundamentally change how we manage large fleets of assets that operate in unpredictable environments like offshore wind or remote infrastructure.

Rosa: Indeed, this paper suggests that future work should really focus on scaling this idea to handle multiple competing assets simultaneously and seeing how those accessibility patterns interact across the fleet.

The paper's improvements: Rosa: This paper doesn't just stop at finding an optimal threshold; it actually proposes a much richer policy structure where you have different thresholds tailored specifically for every possible accessibility state. It’s like having a whole map of rules instead of just one single guideline for fixing things.

Dev: That’s the structural improvement they highlight, Rosa; moving from a universal threshold to a vector of thresholds tau = (tau w) w in W A is what lets the AI make context-aware decisions in real-time based on current environmental feedback. It directly addresses the intermittent nature of maintenance opportunities.

Taro: I think this capability to have state-specific thresholds allows for a much more nuanced handling of uncertainty; it means the policy can explicitly account for not just how degraded an asset is, but also the specific probability and timing associated with future accessibility states. That’s a significant step in dealing with complex, evolving environments.

Rosa: It really shows that we can use the stochastic evolution of accessibility itself to inform when we decide to intervene, which is a key piece of information that was previously overlooked in simpler models. This means the system can be much more strategic about using those intermittent windows.

Dev: From a control engineering standpoint, implementing this requires the AI to constantly monitor and update its estimate of the current accessibility state w and then look up the corresponding optimal threshold tau w; that’s a dynamic decision-making loop that needs to run very efficiently to maintain low latency. The cost function calculation itself is also more complex because it has to incorporate these state-dependent costs.

Taro: If we can build a system that effectively uses those accessibility rules, it could lead to much more resilient autonomous operations in situations where access conditions are constantly fluctuating, which is exactly the kind of messy reality we see in deep-sea or offshore work. It’s about building systems that don't just react, but anticipate the *next* available opportunity.

Rosa: The numerical studies mentioned earlier further back up this; they show that incorporating this accessibility information actually yields an additional cost reduction by adapting the threshold to accessibility, proving that using this environmental foresight is economically sound. It’s not just theoretical; it’s quantifiable savings.

Dev: That quantification is compelling, Rosa, but we have to keep pushing on the robustness issue—if the underlying assumptions about degradation being monotone don't hold up under real-world stress, we could run into unexpected failure modes in our threshold logic. We need to know exactly where this structure breaks down when things get truly chaotic.

Taro: I agree with Dev; the next step for autonomy research has to be testing how well this whole framework performs when those assumptions about environmental smoothness are violated, because in the real world, things rarely follow smooth mathematical curves.

Rosa: So, in summary, the improvement is moving from a single condition threshold to a set of thresholds that change based on accessibility conditions and future expectations. It’s much more powerful for handling dynamic maintenance environments.

Dev: And I think we need to focus our next efforts on designing the control system that can execute those state-dependent actions with the precision required by those varying thresholds, ensuring no unnecessary delays creep into our operational loop.

Conclusion: Rosa: So we've seen how the paper on "Condition-Based Maintenance of Degrading Assets under Intermittent Accessibility" suggests that the optimal maintenance threshold isn't a fixed number, but actually changes depending on whether or not an asset is accessible. It’s a really sophisticated way to handle the uncertainty of when you can actually perform work.

Dev: That structural result is quite powerful, Rosa; it gives us a mathematical framework for making dynamic decisions based on the joint state of condition and access. I'm still focused on how we implement that vector of thresholds within our control loop while keeping latency low enough for real-time operation.

Taro: From an autonomy standpoint, this means we can design systems that are much better at strategic patience; they learn to wait intelligently when conditions are favorable because they understand the stochastic nature of future opportunities. It gives us a way to build more robust behaviors for unpredictable settings.

Rosa: I think the real-world impact is huge because it lets us move beyond simple reactive fixes and toward truly proactive, context-aware maintenance planning in harsh environments where access is sporadic. It really opens up new ways to manage long-term asset health.

Dev: I agree with Rosa about the practical application; those quantified cost reductions show that this modeling isn't just academic, it directly translates to significant economic benefits for operators managing these types of assets. We need to focus on making sure the transition between accessibility states is handled without introducing any jitter or instability in the control action.

Taro: If we can get this kind of dynamic planning working reliably in the field, it could fundamentally change how we manage fleets of robots or infrastructure that have intermittent access—it moves us toward much smarter, anticipatory operational strategies.

Rosa: It sounds like a really solid piece of work for anyone building autonomous systems in complex environments; the paper on "Condition-Based Maintenance of Degrading Assets under Intermittent Accessibility" provides a lot to chew on.

Dev: I think the next thing we should look at is how this framework scales when we start managing multiple assets that all share similar, but slightly different, accessibility patterns. That’s where things get really interesting for large-scale operations.

Taro: Scaling up to multiple interacting assets with shared resources is definitely the logical next frontier; seeing how these dynamic thresholds interact across a whole fleet would give us a much better sense of system-wide optimization.

Rosa: Well, it's been fascinating looking at this paper, and I think we should keep an eye out for follow-up research that builds on this idea of state-dependent policies.

Muhammet Ceyhan Sahin, Murat Yildirim, Onur A. Kilic

Wayne State University · University of Groningen

eess.SY, cs.SY

Submitted: 2026-09-29

Updated: 2026-09-29

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 83/100

The gist: Many maintenance models implicitly assume that maintenance can be performed whenever intervention is warranted, but in practice, environmental uncertainty can make maintenance opportunities

Key concepts

Accessibility State (w): This represents whether a machine or asset can actually be reached for maintenance at any given time. It is a stochastic variable that changes over time based on random processes, meaning maintenance opportunities are not guaranteed and can appear or disappear unpredictably.
Asset Condition State (x): This measures the physical health of the asset, ranging from 'as-good-as-new' (0) to complete failure (K). This state degrades over time due to operation. Maintenance resets this condition back to zero if performed.
Accessibility-State-Dependent Threshold Policy: The optimal strategy is not a single maintenance threshold for all conditions. Instead, the best condition level for performing maintenance changes depending on whether the asset is currently accessible or not, reflecting the dynamic nature of repair opportunities.

Terminology

Benchmark — 8× in this paper · explained in 2 episodes
Accessibility-State-Dependent Threshold Policy — 3× in this paper · explained in 1 episode
Action Space — 2× in this paper · explained in 1 episode
Core — 2× in this paper · explained in 2 episodes
Failure Modes — 2× in this paper · explained in 3 episodes
Robustness — 2× in this paper · explained in 5 episodes
Adaptability — 1× in this paper · explained in 1 episode
Interpretability — 1× in this paper · explained in 15 episodes
Predictive Maintenance — 1× in this paper · explained in 2 episodes

Summary

Many maintenance models implicitly assume that maintenance can be performed whenever intervention is warranted, but in practice, environmental uncertainty can make maintenance opportunities intermittent and dynamically evolving. This paper investigates how asset accessibility stochastically evolves over time and its effect on optimal condition-based maintenance policies.

The gist: An optimal policy retains a threshold form in asset condition, but unlike a single threshold, the optimal threshold varies with the accessibility state.

Problem Formulation as an MDP

The setting is formulated as a finite-state Markov decision process under a long-run average cost criterion for an asset that degrades and whose accessibility evolves stochastically. The system state is defined as s = (w, x), where w represents the accessibility condition and x represents the asset condition state, ranging from as-good-as-new (0) to failure (K). At each decision epoch, an action is selected: either continue without maintenance or perform maintenance.

The transition dynamics are governed by two independent stochastic processes. The accessibility state follows a transition matrix P W, and the asset condition follows a transition matrix P X, which depends on whether maintenance is performed (resetting the condition to 0) or not. The single-period cost c((w, x), a) incorporates efficiency loss g(w, x) from continued operation and fixed intervention costs c PM(w) or c CM(w), which depend on the accessibility state w.

Structural Results of Optimal Policy

The analysis establishes that an optimal policy exists for every combination of asset condition and accessibility. The key structural result is the accessibility-state-dependent threshold policy: for each state in which maintenance is feasible, there exists a condition threshold above which intervention is optimal. This means the maintenance decision depends jointly on (i) the asset’s current condition, (ii) the stochastic evolution of degradation, (iii) the current accessibility state, and (iv) the evolution of future maintenance opportunities.

The optimality equation reveals that for functional states x in 1 to K-1, preventive maintenance is optimal whenever Q P(w) ≤ Q 0(w, x) or when inaccessibility occurs (w ∈/ W A). For the failed state (x=K), corrective maintenance is mandatory if accessible (Q C(w)), otherwise continued operation is the only option.

Monotonicity and Threshold Ordering

The structural analysis relies on several assumptions, including monotone degradation (Assumption 2) and monotone efficiency loss (Assumption 3). Lemma 1 formally establishes the ordering required for threshold determination: there exists a solution (ρ∗, v) to the average-cost optimality equation such that v(w, x + 1) ≥ v(w, x), w ∈ W, x = 1,…, K - 1. This lemma proves that a more degraded asset has a weakly larger relative long-run cost because continued operation combines a larger current efficiency loss with stochastically worse future degradation.

Quantifying the Value of Information

Numerical studies motivated by offshore wind turbine maintenance quantify the economic benefits of incorporating condition and accessibility information. The optimal policy reduces long-run average cost from a purely condition-based policy by 33.95% relative to an optimized age-based policy, demonstrating that condition information determines whether maintenance is warranted by the physical state of the asset.

Furthermore, adapting the threshold to accessibility provides an additional reduction in long-run average cost by 4.31% relative to a constant-threshold policy, showing that accessibility information determines whether a current maintenance opportunity should be used. The sensitivity analysis further shows that stronger weather persistence increases the differentiation between favorable and unfavorable accessibility periods, while faster degradation calls for earlier intervention.

Threshold Policy Representation

The optimal stationary policy can be represented by a collection of condition thresholds associated with each accessible state. A threshold vector τ = (τw)w∈W A is defined, where τw ∈ 1 to K represents the condition level above which maintenance is performed in accessibility state w. The optimal threshold vector satisfies τ∗ ∈ arg min τ∈Θ ητ, where ητ is the long-run average cost induced by policy πτ. This structural result reduces the representation of the optimal stationary maintenance policy from arbitrary state-action decisions to a collection of condition thresholds associated with the accessible states.

Conclusion

The paper concludes that the optimal threshold is not universal: it changes with accessibility, degradation, and economic conditions, and failing to adapt the threshold to these conditions can increase long-run cost. The results show that current accessibility determines whether maintenance can be performed, while the stochastic evolution of accessibility determines the value of postponing an available intervention. Future research directions include studying multiple interacting assets competing for limited maintenance resources.

How it works

Improvements for AI systems

Here are the specific improvements to AI systems derived from this research, and what those improved systems can achieve:

)1. Adaptive Maintenance Thresholding in Stochastic Environments:

The core improvement is moving beyond static or time-based maintenance schedules toward a dynamic, state-dependent maintenance policy that explicitly incorporates both asset condition and the stochastic evolution of future opportunities (accessibility).

Specific Improvements and Capabilities:

The improved AI system will be a sophisticated Markov Decision Process (MDP) solver capable of making optimal maintenance decisions in scenarios characterized by intermittent accessibility (e.g., offshore wind, complex industrial assets).

Specifically, the system can perform the following functions:

a) Adaptive Threshold Determination: Instead of using a single fixed threshold for when to fix, the system will determine a unique optimal condition threshold, denoted as θ(w), for every possible current accessibility state (w). This means an asset in state (Condition=5, Accessible) might require maintenance if its condition is below 4, but if it's in state (Condition=5, Inaccessible), the decision is to continue operating until the next opportunity arises.

b) Forward-Looking Decision Making: The system will utilize the expected evolution of future accessibility (governed by transition probabilities Pw) to inform its current action. If it is currently accessible but future accessibility is predicted to worsen significantly, the system may choose to perform preventive maintenance immediately, even if the current condition is only moderately degraded. Conversely, if favorable conditions are expected soon, it will tolerate a more degraded state now.

c) Benchmark Comparison and Value Quantification: The system can automatically compare its performance against simpler models (constant-threshold policies and age-based policies). It can quantify the economic benefit of incorporating accessibility information—demonstrating quantifiable cost reductions (up to 33.95% improvement in the base case) versus relying solely on condition or time.

d) Sensitivity Analysis for Robustness: The system's decision-making rules can be dynamically adjusted based on uncertainty parameters:

i. If degradation is highly uncertain (high standard deviation), the system will become more sensitive to changes in accessibility outlook, leading to a more responsive threshold policy.

ii. If weather persistence is high, the system will exhibit anticipatory behavior—deferring maintenance when conditions are favorable because they are likely to persist.

e) Optimized Policy Implementation: The final output is not just a decision but an explicit policy map: for any observed state (Condition, Accessibility), the AI provides the optimal action (Continue or Maintain). This allows for real-time deployment in systems where environmental sensors feed into the accessibility state.

)2. What the Improved AI System Can Do (Specific Applications):

This enhanced AI system is ideal for high-stakes, capital-intensive asset management:

Integrated Offshore Wind Farm Management: The system can optimize the maintenance schedule for an entire wind farm by understanding that weather patterns dictate when turbines can be serviced, and that favorable seasons might justify delaying maintenance on slightly degraded turbines. It minimizes total long-run cost by balancing immediate efficiency losses against the risk of missing a future, potentially more favorable, maintenance window.
Predictive Asset Health in Harsh Environments: For assets like deep-sea drilling equipment or remote infrastructure (where access is weather/sea condition dependent), the AI can proactively adjust maintenance schedules based on forecasts. It won't just wait for a pre-set degradation level; it will wait intelligently, using the forecast to determine if waiting is economically rational.
Resource Allocation Under Uncertainty: In scenarios with limited maintenance crews or parts, this system can prioritize which assets to service next based on the joint risk profile of condition (how degraded it is) and accessibility (how likely we are to service it soon).
Cost Optimization for Maintenance Contracts: For asset owners, the system provides a rigorous mathematical justification for their maintenance contracts. It proves that paying for condition-based sensors and real-time weather data leads to significant long-term savings compared to traditional fixed schedules or simple time-based rules.

Abstract

Many maintenance models implicitly assume that maintenance can be performed whenever intervention is warranted. In practice, however, environmental uncertainty, such as weather and sea conditions, can make maintenance opportunities intermittent and dynamically evolving. We consider a degrading asset for which preventive maintenance may be performed before failure, but intervention is possible only when the asset is accessible. Accessibility evolves stochastically over time and therefore affects not only whether maintenance can be performed now, but also the value of waiting for future opportunities. We formulate this setting as a finite-state Markov decision process under a long-run average cost criterion and investigate the structure of the optimal maintenance policy. Under monotone degradation and cost conditions, we show that an optimal policy retains a threshold form in asset condition, but unlike a single threshold, the optimal threshold varies with the accessibility state. Thus, intervention depends jointly on the asset condition and the expected evolution of future maintenance opportunities. Through a numerical study motivated by offshore wind turbine maintenance, we examine how accessibility dynamics, degradation characteristics, and economic parameters shape the optimal thresholds and compare the optimal policy with constant-threshold and age-based maintenance policies. In a representative setting, adapting the condition threshold to accessibility reduces long-run average cost by 4.31% relative to an optimized condition-based maintenance policy. Relative to an optimized age-based policy, the optimal policy reduces long-run average cost by 33.95%. The results show that maintenance thresholds are not universal: they should adapt to the degradation and accessibility characteristics of the system, and failing to account for these conditions can lead to substantial cost increases.