Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory

arXiv:2606.05334 · cs.AI · Submitted 2026-06-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory".

Jane: Here's the gist of the paper. Returned products in circular factories re-enter production with heterogeneous degradation states, usage histories, and remaining capability The proposed framework combines the current tool state with recent force–torque usage windows This paper addresses this gap for an angle grinder by combining uncertainty-aware functional prediction with component-level fatigue assessment in an instance-specific reliability workflow Pruning redundant reasoning tokens The research question asks how the current state, usage history, and future usage scenario of a returned product can be represented in a unified functional–material reliability space to support instance-specific redeployment decisions under uncertainty The framework establishes a unified functional–material assessment space that combines the current system state, recent usage history, and material degradation state into a common representation Redeployment is formulated as an admissibility problem where a returned product is suitable for reuse if its predicted functional behavior and material capability remain within the admissible reliability region under a prescribed future usage scenario How the training signal is built The functional model uses two input types: the current system state and a short history of recent loading The prediction target is the functional behavior vector yt which contains thermal, electrical, rotational, and geometry-related condition variables The learning problem is formulated as pθ(yt st, Ut−L+1∶t) The network predicts increments rather than absolute target values, and the absolute trajectory is reconstructed by adding the predicted changes to the measured initial value Where the gains come from The model architecture consists of a usage-history encoder, a state encoder, a recurrent sequence model, and an uncertainty-aware output head The LSTM is used because the present task involves short sequences of overlapping usage windows whose local structure is already summarized by the convolutional encoder The uncertainty-aware loss function uses a weighted Gaussian negative log-likelihood loss to predict both an expected value and an uncertainty estimate Component-level material assessment The material side lifetime assessment is commonly based on two established approaches: stress-life fatigue models and fracture-mechanics-based crack-growth models This pipeline combines three steps: finite element analysis for local stress reconstruction, a stress–life fatigue characterization of the shaft under rotating bending, and a fracturemechanics-based crack-growth model The Paris-law propagation life is determined by numerical integration of the Paris equation over the crack length interval from a0 to ac Consolidation of functional-material reliability algorithm The final assessment step consolidates the full methodology into a single streaming reliability analysis, as demonstrated in Algorithm 3 This procedure updates functional reliability Rf unc, material reliability Rmat, and system reliability Rsys at each inspection point while accounting for gaps between non-consecutive observations The governing branch identifies which reliability dimension is binding at each inspection point, functional or material Case study results The case study considered in this work is an angle grinder, selected as a representative returned product for instance-specific functional and material assessment in a circular-factory context The component–material assessment indicates a high reuse potential of the output shaft under the investigated nominal service loading Rare high-load events and overload history can dominate crack propagation life and therefore determine whether direct reuse is still justified The sequential replay over 18 ordered inspection files demonstrated that the functional and material branches can be integrated coherently within a single streaming reliability workflow Implications for circular factory decision-making The framework supports a shift from inspection-based classification to prediction-based redeployment A returned product should not only be judged by whether it currently operates, but by whether it is expected to fulfill its function in the next use phase and whether its critical components retain sufficient structural capability This supports more differentiated Circular Factory decisions such as direct reuse, restricted reuse, reconditioning, component replacement, further inspection, or rejection The research contributes the integration of data-driven functional prognosis and material-informed fatigue assessment in a redeployment-oriented reliability workflow Limitations and research outlook Generalization to real Circular Factory conditions still requires validation on multiple returned products, heterogeneous usage histories, variable operating profiles, and reprocessed product states The material model is load-case-specific and the predicted reuse cycles should be interpreted as scenario-specific estimates rather than universal lifetime values for all shafts Future work should focus on validating the functional model under heterogeneous operating conditions, calibrating shaft-specific crack-growth parameters, improving load-to-stress reconstruction, and testing scenario-conditioned predictions against prescribed future load profiles

Title and authors: Tom: So we're looking at this paper today about "Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory." It’s tackling that big problem where you take a used machine, like an angle grinder, back into production, and you can't just look at it once to decide if it’s good to reuse.

Jane: Right. So the core idea here is that when a product comes back into the loop, its condition isn't just one snapshot; it’s a mix of how it looks now, how much it has been used before, and what kind of stresses it might take next.

Lu: Exactly. The paper proposes combining the current state with recent usage history to predict future behavior while also assessing the material degradation separately. It sets up this unified functional material assessment space that brings everything together for reuse decisions under uncertainty.

Meng: From a production standpoint, what I want to know is how much of this prediction actually translates into something we can deploy reliably on the shop floor. Are we talking about a simple pass or fail, or something more nuanced?

Lalam: The paper frames redeployment as an admissibility problem, meaning the tool is suitable for reuse if its predicted function and material capability stay within a certain reliability region under a planned future usage scenario.

Tom: That’s the functional side—predicting what it *does*—and it relies on this conditional sequence learning model that predicts increments instead of absolute values based on the current state and recent loading history. It tries to anchor the forecast to where it is now, which seems smart for short sequences of usage windows.

Jane: So, if we think about how we build that training signal, it uses two things: the current system state and a short history of recent loading. The target prediction is a functional behavior vector containing variables like thermal, electrical, rotational, and geometry related conditions.

Lu: And to handle the uncertainty around those predictions, they use an uncertainty-aware loss function that estimates both an expected value and its uncertainty estimate using a weighted Gaussian negative log-likelihood loss. That helps penalize bad predictions while making sure the variance doesn't explode unnecessarily.

Meng: Okay, so that sounds like a lot of complexity for what we need in a fast production line. How do they actually manage the material part—the physical wear and tear? That’s where things usually get really messy when you have to link force-torque measurements to actual material damage.

Title and authors: Jane: They handle that by combining several steps into a pipeline. First, finite element analysis to reconstruct local stress, then an S-N or Miner damage evaluation with a Haibach extension, and finally a fracture mechanics based crack growth model using the Paris law for the propagation life between initial crack length zero and final crack length.

Tom: That’s how they link those measured force-torque windows into cumulative material damage and a corresponding material reliability estimate. It’s not just looking at the stress once; it’s modeling how that stress accumulates over time under rotating bending.

Lu: The paper consolidates this into a single streaming reliability analysis using Algorithm three which updates the functional reliability, the material reliability, and the system reliability at each inspection point while accounting for gaps between those observations <ref:2606.05334#pg1>. The governing branch tells you whether it’s the functional degradation or the material fatigue that is limiting your decision at that moment.

Meng: A single stream update sounds efficient on paper, but what does that look like when we try to deploy this at scale? Does this streaming analysis run fast enough for a high-throughput sorting line?

Jane: The case study uses an angle grinder as the example, and the component material assessment showed a high reuse potential for the output shaft under nominal service loading. But they pointed out that rare high-load events or overload history can actually dominate how long that crack propagation life lasts, which is important context.

Tom: That’s a key finding—the paper shows that even if the nominal load looks fine, a history of high loads can dictate whether you should still reuse the shaft directly. The sequential replay over eighteen ordered inspection files showed that the functional and material branches can work together coherently in this single workflow.

Lu: What this means for us is a shift away from just classifying a product as 'good' or 'bad' based on current operation, toward a prediction-based redeployment strategy. A tool needs to be judged not just by whether it works now, but if it’s expected to fulfill its function in the next use phase and if the critical components have enough structural capability for that future load.

Jane: So, what is this actually changing for someone who only listens to this show? It means we move from simple inspection-based classification to a system that predicts redeployment based on both expected behavior and material integrity.

Title and authors: Tom: It’s about integrating data-driven functional prognosis with material-informed fatigue assessment in a workflow specifically designed for making reuse decisions. That integration seems like the main contribution of this work, "Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory."

Meng: The limitation they flag is that generalizing this to real circular factory conditions still requires testing it on many different returned products with varied histories and operating profiles. Also, the material model is load-case specific, so the predicted reuse cycles should be treated as scenario-specific estimates, not a universal lifetime value for every shaft.

Lu: That’s fair. The authors themselves say that they need to validate the functional model under heterogeneous operating conditions and also calibrate those crack-growth parameters for different shafts. Future work needs to focus on testing those scenario-conditioned predictions against prescribed future load profiles.

Jane: So, the big picture is that this paper gives us a more detailed way to make informed decisions about what happens next for returned products, moving beyond simple pass or fail checks. We need to be looking at both the expected function and the physical wear on those components when making reuse calls.

Tom: It’s about being able to give more differentiated decisions—direct reuse, restricted reuse, reconditioning, component replacement—instead of just a single category. This framework supports that by showing you whether the decision is limited by predicted behavioral degradation or by accumulated structural fatigue at any given inspection point.

Lalam: To wrap up on the main points: this paper proposes a unified way to assess functional behavior and material fatigue using sequence learning for prediction and physical models for damage, all tied together in a streaming reliability analysis.

Meng: I just want to stress that deploying this requires a lot of clean data inputs—the usage history, the state encoders—and we need to make sure the uncertainty estimates are actually reliable enough when we put them into production systems.

Lu: That’s where the ambition lies, but also where the engineering challenge is. The model needs to handle real-world noise and variability across different tools.

Jane: And that's what we leave you with today: "Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory." It’s a framework that helps us move from just inspecting a tool to predicting its future service life under uncertainty, which is crucial for making smarter decisions in the circular economy.

The paper's summary: Tom: So this paper is basically about how to figure out if a returned product is still good to use again by looking at its past and predicting its future performance under uncertainty.

Jane: Right, so it takes a used machine, like an angle grinder, and instead of just checking if it’s broken now, they want to predict what it will do in the next use phase.

Tom: Exactly. The researchers created this unified space where you combine the current state of the tool with its entire usage history and what they think will happen next.

Jane: And they tie that functional prediction to a separate assessment of how much material wear and tear has happened, which is called material fatigue assessment.

Tom: So it’s not just one thing; it’s combining what the tool *does* with what the tool is *made of*. They formulate reuse as an admissibility problem, meaning it can only be reused if its predicted function and its structural capability stay within a certain reliability region for a planned future use.

Jane: That sounds like a really practical way to make decisions in a real factory setting. How do they actually build that functional prediction part?

Tom: They use two inputs for the functional model, the current system state and this short history of recent loading windows. The network predicts increments rather than absolute values, which keeps its forecast anchored to where it is right now.

Jane: And to handle the uncertainty in that prediction, they use a loss function that estimates both an expected value and an uncertainty estimate. That way, they know not just what the tool will do but how sure they are about it.

Tom: That’s smart because both those pieces—the predicted mean and the uncertainty estimate—determine if you can actually trust whether the tool will perform as expected in the future.

Jane: So, on top of that functional prediction, they have this material side where they translate those measured force-torque windows into actual material damage.

Tom: They use a pipeline there that combines finite element analysis to reconstruct local stress, then an S-N or Miner damage evaluation with a Haibach extension, and finally a crack growth model using the Paris law for the propagation life between initial crack length zero and final crack length.

Jane: That means they are linking those force measurements directly into how much material is degrading over time.

Tom: The whole thing wraps up in a single streaming reliability analysis where they update the functional reliability, the material reliability, and the system reliability at every inspection point while accounting for any gaps between those observations.

Jane: So what does this mean for someone running a circular factory? It shifts decisions from just checking if it’s working today to predicting whether it will fulfill its function in the next use phase and whether its components have enough structural capability.

Tom: It supports more differentiated reuse decisions, like direct reuse, restricted reuse, or even component replacement, because you can see which branch—the functional degradation or the material fatigue—is actually limiting your decision at that moment.

Jane: It’s about integrating data-driven functional prognosis with material-informed fatigue assessment into a workflow specifically designed for making those redeployment calls.

Tom: The limitation they mention is that this still needs to be tested on many different returned products with varied histories and operating profiles, because the material model is load-case specific, so the results are scenario specific estimates rather than universal lifetimes for every shaft.

Jane: So the authors need to validate how this works under really messy, real-world conditions before it can be widely applied. This paper shows a way to combine prediction and physical modeling for reuse decisions.

The paper's improvements: Tom: So, looking at how they improved their original idea, they focused on making the functional prediction better by using a conditional sequence learning problem that predicts increments instead of absolute target values.

Jane: That makes sense, so it’s not trying to guess the exact final state of the tool but predicting how much it will change based on what's already happening and its history.

Tom: Exactly. It keeps the forecast anchored to the known starting condition, which is that future behavior depends jointly on what's currently in there and that usage history.

Jane: And they also enhanced reliability assessment by adding a Gaussian negative log-likelihood loss function that predicts both an expected value and its uncertainty estimate.

Tom: That’s important because it penalizes bad predictions but also discourages excessively large variances, which helps because both the predicted mean and the uncertainty estimate determine if you can actually trust whether the tool will perform as expected.

Jane: So they are making sure that when they make a prediction, they aren't just giving you a number but also telling you how much doubt there is attached to it.

Tom: That uncertainty estimation is crucial because it directly influences the probability of functional unfulfillment, and that’s something we need for reliable deployment.

Jane: And on the material side, they made the fatigue assessment more robust by combining measured force-torque windows into damage using a pipeline that links stress reconstruction to S-N or Miner damage evaluation and then to a Paris law crack growth analysis.

Tom: That’s a big step because it allows the system to link those raw force measurements directly into cumulative material damage and give you a corresponding material reliability estimate.

Jane: So they are making sure that when you look at the tool, you get two separate but connected warnings: one about how well it will function next, and one about its physical structure breaking down.

Tom: The consolidation part is also key; they put everything into a single streaming reliability analysis that updates functional, material, and system reliability at every inspection point while tracking gaps between observations.

Jane: That governing branch identifies which part—the functional behavior or the material fatigue—is actually limiting your reuse decision at each specific moment.

Tom: That means instead of just getting a final pass or fail, you get an answer that tells you exactly what's causing the restriction right now.

Jane: It’s about giving the user a more detailed picture of why they might not be able to use that tool again, whether it’s because it’s going to act weird or because its physical shaft is getting too worn.

Tom: The limitation they state is that generalizing this to real circular factory conditions still requires testing it on many different returned products with varied histories and operating profiles, because the material model is load-case specific.

Jane: So they are saying this isn't a universal solution yet; it needs more testing on really messy real-world scenarios before you can trust it for every single tool.

Conclusion: Tom: So to wrap up, this paper introduces a method for assessing returned products in circular factories by combining uncertainty-aware functional prediction with component-level material fatigue assessment through a single streaming reliability analysis framework.

Jane: It’s like giving a tool a sophisticated health check that looks at both what it does and how physically worn out it is, and then uses that to decide if it can go back into service.

Tom: Exactly. The main implication is moving from just checking if something works now to predicting whether it will fulfill its function in the next use phase while also confirming its structural integrity for that future load.

Jane: It supports making better decisions, like choosing between direct reuse and reconditioning or component replacement, based on a much richer set of data than just a simple inspection result.

Tom: The authors are using this unified approach to link the functional and material branches together so you can see which one is actually limiting your decision at any given moment.

Jane: It’s really about integrating data-driven prognosis with material fatigue assessment into a workflow specifically designed for making those redeployment calls.

Tom: The limitation they flag is that generalizing this to real circular factory conditions still requires testing it on many different returned products with varied histories and operating profiles, because the material model is load-case specific.

Jane: They need more validation under messy real-world scenarios before we can apply this to every single tool out there.

Tom: That’s fair. The authors themselves say they need to calibrate those crack growth parameters for different shafts and test the functional model under heterogeneous operating conditions.

Jane: So the paper shows a solid framework, but it still needs more real-world stress testing before it becomes a universal rule for every returned item.

Tom: That’s the challenge. We need to make sure that when we put this into production, those uncertainty estimates are actually reliable enough when dealing with real-world noise and variability across different tools.

Jane: It’s about getting from a theoretical framework to a reliably deployed system that can handle the messy reality of recycled products.

Tom: So, "Uncertainty Aware Functional Behavior Prediction and Material Fatigue Assessment for Circular Factory" gives us a way to get more nuanced decisions in the circular economy.

Lu: If this functional prediction model works as intended, it could fundamentally change how we manage asset lifecycles in manufacturing by allowing for much more precise reuse decisions based on predicted performance.

Meng: From my side, I need to see if the data pipelines can actually run fast enough to handle the streaming analysis without creating a massive computational overhead that breaks at scale.

Lalam: I'm fairly sure of this answer; based on the literature, this integrated approach is a very sound way to handle these complex multi-variable problems in asset management.

Nehal Afifi, Mehdi Khabou, Victor Mas, Jonas Hemmerich, Patric Grauberger, Stefan Dietrich, Volker Schulze, Sven Matthiesen

IPEK Institute of Product Engineering at Karlsruhe Institute of Technology (KIT) · IAM-WK Institute for Applied Materials – Materials Science and Engineering at Karlsruhe Institute of Technology (KIT) · wbk Institute of Production Science at Karlsruhe Institute of Technology (KIT)

cs.AI

Submitted: 2026-06-03

Updated: 2026-11-06

Comments: 30 pages, submitted to the Journal of Manufacturing Systems' special issue about circular factories, the manuscript is accepted and under publication

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Returned products in circular factories re-enter production with heterogeneous degradation states, usage histories…

Key concepts

Functional–Material Reliability Space
This is a unified representation where the tool's predicted operational capability (functional) and its remaining structural integrity (material) are mapped together. This allows researchers to assess reuse suitability by checking if both aspects remain within acceptable limits under future use scenarios.
Uncertainty-Aware Prediction
The model predicts future functional behavior not just as a single value, but as a distribution of possibilities, including an estimate of how uncertain the prediction is. This uncertainty measure is crucial because it directly informs the decision-making process about the risk associated with reusing a product.
Streaming Reliability Analysis
Instead of making one final judgment at the end, this method updates functional and material reliability continuously as new inspection data arrives. At each step, it determines which factor—functional degradation or material fatigue—is currently limiting the reuse decision.
Component-Level Fatigue Assessment
This involves a detailed analysis of how the physical parts of a tool degrade over time. It combines stress calculations from simulations with models that predict crack growth based on material properties, linking measured forces to cumulative damage.

Terminology

Summary

Returned products in circular factories re-enter production with heterogeneous degradation states, usage histories, and remaining capability The proposed framework combines the current tool state with recent force–torque usage windows This paper addresses this gap for an angle grinder by combining uncertainty-aware functional prediction with component-level fatigue assessment in an instance-specific reliability workflow

Pruning redundant reasoning tokens

The research question asks how the current state, usage history, and future usage scenario of a returned product can be represented in a unified functional–material reliability space to support instance-specific redeployment decisions under uncertainty The framework establishes a unified functional–material assessment space that combines the current system state, recent usage history, and material degradation state into a common representation Redeployment is formulated as an admissibility problem where a returned product is suitable for reuse if its predicted functional behavior and material capability remain within the admissible reliability region under a prescribed future usage scenario

How the training signal is built

The functional model uses two input types: the current system state and a short history of recent loading The prediction target is the functional behavior vector yt which contains thermal, electrical, rotational, and geometry-related condition variables The learning problem is formulated as pθ(yt st, Ut−L+1∶t) The network predicts increments rather than absolute target values, and the absolute trajectory is reconstructed by adding the predicted changes to the measured initial value

Where the gains come from

The model architecture consists of a usage-history encoder, a state encoder, a recurrent sequence model, and an uncertainty-aware output head The LSTM is used because the present task involves short sequences of overlapping usage windows whose local structure is already summarized by the convolutional encoder The uncertainty-aware loss function uses a weighted Gaussian negative log-likelihood loss to predict both an expected value and an uncertainty estimate

Component-level material assessment

The material side lifetime assessment is commonly based on two established approaches: stress-life fatigue models and fracture-mechanics-based crack-growth models This pipeline combines three steps: finite element analysis for local stress reconstruction, a stress–life fatigue characterization of the shaft under rotating bending, and a fracturemechanics-based crack-growth model The Paris-law propagation life is determined by numerical integration of the Paris equation over the crack length interval from a0 to ac

Consolidation of functional-material reliability algorithm

The final assessment step consolidates the full methodology into a single streaming reliability analysis, as demonstrated in Algorithm 3 This procedure updates functional reliability Rf unc, material reliability Rmat, and system reliability Rsys at each inspection point while accounting for gaps between non-consecutive observations The governing branch identifies which reliability dimension is binding at each inspection point, functional or material

Case study results

The case study considered in this work is an angle grinder, selected as a representative returned product for instance-specific functional and material assessment in a circular-factory context The component–material assessment indicates a high reuse potential of the output shaft under the investigated nominal service loading Rare high-load events and overload history can dominate crack propagation life and therefore determine whether direct reuse is still justified The sequential replay over 18 ordered inspection files demonstrated that the functional and material branches can be integrated coherently within a single streaming reliability workflow

Implications for circular factory decision-making

The framework supports a shift from inspection-based classification to prediction-based redeployment A returned product should not only be judged by whether it currently operates, but by whether it is expected to fulfill its function in the next use phase and whether its critical components retain sufficient structural capability This supports more differentiated Circular Factory decisions such as direct reuse, restricted reuse, reconditioning, component replacement, further inspection, or rejection The research contributes the integration of data-driven functional prognosis and material-informed fatigue assessment in a redeployment-oriented reliability workflow

Limitations and research outlook

Generalization to real Circular Factory conditions still requires validation on multiple returned products, heterogeneous usage histories, variable operating profiles, and reprocessed product states The material model is load-case-specific and the predicted reuse cycles should be interpreted as scenario-specific estimates rather than universal lifetime values for all shafts Future work should focus on validating the functional model under heterogeneous operating conditions, calibrating shaft-specific crack-growth parameters, improving load-to-stress reconstruction, and testing scenario-conditioned predictions against prescribed future load profiles

Conclusion

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve functional behavior prediction by using a conditional sequence learning problem that combines current system state with recent usage history to predict increments rather than absolute target values. This allows for local evolution focus and keeps forecasts anchored to the known starting condition, which is what captures the main assumption that future functional behavior depends jointly on the current tool state and recent usage history.

  2. Enhance reliability assessment by incorporating a Gaussian negative log-likelihood loss function that predicts both an expected value and its uncertainty estimate. This penalizes inaccurate predictions while discouraging excessively large variances, which is important because both the predicted mean and uncertainty determine the probability of functional unfulfillment.

  3. Integrate material fatigue assessment by translating measured force-torque windows into cumulative material damage using a pipeline that combines finite element analysis for stress reconstruction, S-N/Miner damage evaluation with Haibach extension, and Paris-law crack growth analysis. This allows the system to link measured force–torque windows into cumulative material damage and a corresponding material reliability estimate.

  4. Enable scenario-conditioned inference by prescribing future usage windows to evaluate functional behavior under explicit operating conditions. This allows the model to estimate the future response under this operating condition as p theta(Y t+1:t+H st, Ut-L+1:t, Ũ t+1:t+H).

  5. Consolidate functional and material reliability into a single streaming reliability analysis that updates functional, material, and system reliability at each inspection point. This procedure allows for the identification of the governing branch, which provides a direct indication of whether the redeployment decision is limited by predicted behavioral degradation or by accumulated structural fatigue.

Abstract

Returned products in circular factories re-enter production with heterogeneous degradation states, usage histories, and remaining capability. Reuse cannot be decided from the current inspection alone, because future function fulfillment and component integrity may evolve differently under the next service scenario. Existing PHM approaches support degradation prediction, but often target fixed operating conditions or isolated component benchmarks, while material-fatigue assessment is rarely linked to system-level functional prognosis. This paper addresses this gap for an angle grinder by combining uncertainty-aware functional prediction with component-level fatigue assessment in an instance-specific reliability workflow. The proposed framework combines the current tool state with recent force--torque usage windows. A convolutional encoder extracts loading patterns from spindle forces and shaft torque, and an LSTM backbone predicts nine functional variables as Gaussian mean and variance estimates. In parallel, the same loading history is translated into output-shaft fatigue information through finite-element-supported stress reconstruction, S--N/Miner damage evaluation with Haibach extension, and Paris-law crack-growth analysis. A streaming replay algorithm consolidates both branches into functional, material, and system reliability trajectories. Held-out tests show mean 2% -tolerance accuracy of 0.9652 across nine outputs. Thermal variables are predicted near-perfectly, while drive motor current and load speed remain the most demanding dynamic outputs, with R squared values of 0.9750 and 0.9924. Torque history is especially important for these variables, and the conventional LSTM outperforms GRU and xLSTM in the short-history setting. Reliability calibration is most informative for drive motor current, where predicted and observed exceedance probabilities...

Related papers