PaperlayerView exact revision on arXiv

Source-linked AI summary

Explainable, Interpretable & Trustworthy AI for Intelligent Digital Twin: Case Study on Remaining Useful Life

Kazuma Kobayashi, Syed Bahauddin Alam

arXiv:2301.06676v2cs.LGstat.APstat.CO

TL;DR

The paper addresses how to make AI/ML-based RUL prediction in intelligent digital twins explainable, interpretable, and trustworthy. It explains XAI and IML across local and global analyses, applies interpretable models and PiML to RUL data, and reports that operating cycle is the dominant predictive factor while identifying continued validation and online-update needs.

  • Problem

    RUL prediction in intelligent digital twins requires explainable, interpretable, and trustworthy AI so users can understand and trust model decisions.

  • Method

    The paper explains XAI and IML for digital-twin components and applies local and global analyses with ReLU-DNN, EBM, FIGS, decision trees, and PiML.

  • Results

    Operating cycle is consistently identified as the most influential factor governing RUL prediction across the paper’s XAI analyses.

  • Takeaways & Limitations

    Explainable RUL forecasts can support trust, transparency, accountability, and data-driven maintenance planning in safety-critical infrastructure.

  • Takeaways & Limitations

    The paper identifies future needs for sensor-related degradation modeling and online learning that continuously updates digital-twin models and parameters.

Abstract

from arXiv · show

Artificial intelligence (AI) and Machine learning (ML) are increasingly used in energy and engineering systems, but these models must be fair, unbiased, and explainable. It is critical to have confidence in AI's trustworthiness. ML techniques have been useful in predicting important parameters and in improving model performance. However, for these AI techniques to be useful for making decisions, they need to be audited, accounted for, and easy to understand. Therefore, the use of explainable AI (XAI) and interpretable machine learning (IML) is crucial for the accurate prediction of prognostics, such as remaining useful life (RUL), in a digital twin system, to make it intelligent while ensuring that the AI model is transparent in its decision-making processes and that the predictions it generates can be understood and trusted by users. By using AI that is explainable, interpretable, and trustworthy, intelligent digital twin systems can make more accurate predictions of RUL, leading to better maintenance and repair planning, and ultimately, improved system performance. The objective of this paper is to explain the ideas of XAI and IML and to justify the important role of AI/ML in the digital twin framework and components, which requires XAI to understand the prediction better. This paper explains the importance of XAI and IML in both local and global aspects to ensure the use of trustworthy AI/ML applications for RUL prediction. We used the RUL prediction for the XAI and IML studies and leveraged the integrated Python toolbox for interpretable machine learning~(PiML).

1 Introduction

Digital twins and PHM use AI/ML to support real-time system assessment and RUL-based maintenance decisions, but trustworthy deployment requires explainable and interpretable models. The paper frames XAI/IML as ways to make RUL predictions transparent, understandable, and useful while noting unresolved data, validation, uncertainty, and trust challenges.

  • 1 Introduction: Digital-twin explainability is complicated by noisy or inconsistent data, uncertainty, complex deterioration, and the difficulty of validating explanations and operator effects.The paper also identifies unresolved safety, uncertainty-propagation, and user-probing issues.
  • 1 Introduction: XAI is presented as improving transparency, trust, ethical accountability, user experience, and debugging of AI-enabled digital twins.For digital twins specifically, explanations can identify improvement opportunities and support more reliable performance.
  • 1 Introduction: XAI and interpretable ML are crucial for transparent, understandable, and trustworthy RUL prediction in intelligent digital twins.They help users understand influential prediction factors and assess whether models rely on relevant, non-misleading features.
  • 1 Introduction: Digital twins support PHM by monitoring system health, anticipating failures, and enabling condition-based maintenance and other RUL-driven decisions.RUL estimates inform predictive maintenance, production-plan modifications, and component management.
  • 1 Introduction: XAI can build confidence in RUL forecasts by revealing operating conditions and failure modes associated with predicted degradation.The paper connects these explanations to actionable maintenance insights and broader adoption in high-reliability domains.

2 Core Components of Digital Twins: ML Algorithms and the XAI Requirement for DTs

The paper describes digital-twin components that use ML for prediction and system updating, then motivates XAI for understanding their prognostic behavior. It also presents Bayesian-filter updates, operator-learning surrogates, and an open, Python-based platform architecture.

  • 2.1 Update Module in a Digital Twin: A digital-twin framework contains prediction, system-update, data-processing, visualization, and decision-making modules, with ML concentrated in prediction and updating.The system-update module combines a Bayesian filter with ML and is characteristic of digital twins rather than ordinary simulations.
  • 2.1 Update Module in a Digital Twin: The update module estimates states and parameters with Bayesian filtering while using temporal evolution and online sequential learning to maintain the model.A Gaussian-process ML component predicts the degradation mechanism as a future system state.
  • 2.2 Operator Learning as a Faster Surrogate for a Digital Twin: Operator learning is presented as a faster, more robust surrogate because conventional supervised ML may be accurate without being fast enough for digital-twin use.The paper illustrates this direction with DeepONet, which uses branch and trunk networks for a 2D diffusion system.
  • 2.2 Operator Learning as a Faster Surrogate for a Digital Twin: The operator-learning example maps an input function u(x) to a target system state s(x), while its lack of explainability remains a stated concern.The input function is sampled from a Gaussian random field and numerical solutions provide training targets.
  • 2.3 Platform-Agnostic and Surrogate-Driven Digital Twin Framework: The proposed platform is platform-agnostic and modification-friendly, integrating Python-based AI/ML code, third-party software, databases, and a browser user interface.Flask provides connectivity among the platform’s four major components and supports a graphical interface.

3 Explainable and Interpretable AI for Digital Twins

The section distinguishes explainable and interpretable AI and situates them within digital-twin frameworks, emphasizing transparent, understandable decision-making. It introduces interpretable-model approaches and an accuracy–explainability trade-off for trustworthy AI applications.

  • Digital-twin components lack clear explainability, motivating interpretable AI methods that justify model judgments and predictions.
  • XAI can use transparent models or post hoc techniques such as feature importance and sensitivity analysis to explain decisions.
  • XAI and interpretable AI overlap but differ in emphasis: XAI explains actions, whereas interpretability concerns understanding model outcomes.
  • Interpretable machine learning explains relationships between model inputs and outputs while supporting human-readable justifications for predictions.
  • Interpretable machine learning is framed around balancing prediction accuracy against model explainability.
  • The study uses inherently interpretable ReLU, EBM, FIGS, and tree models instead of post hoc explainability models.

4 Inherent Interpretable ML Models: Problem Setup, Methodologies, and Data Preparation

The study prepares PHM08 turbofan-engine data for RUL prediction and evaluates four interpretable machine-learning models. RUL is estimated from operational cycles to create a target suitable for XAI analysis.

  • PHM08 uses NASA C-MAPSS simulations to represent turbofan-engine degradation under different operating conditions.
  • The dataset is organized as time-interval rows with 26 feature columns, including engine identifiers, operational cycles, operating settings, and sensor measurements.
  • RUL is estimated by subtracting each engine’s operational cycle from its maximum observed cycle because the original PHM08 data lacks explicit RUL values.
  • The study formulates two XAI test cases for RUL prediction using ReLU-DNN, EBM, FIGS, and decision-tree surrogates.

5 XAI for RUL: Results and Discussion

The paper frames RUL as the target explanation parameter and applies PiML to study XAI and interpretability using sensor and operating-setting features.

  • RUL is treated as the target explanation parameter, with 26 independent features comprising sensor data and operating settings.
  • The study leverages the integrated Python toolbox for interpretable machine learning, PiML, for its XAI and interpretability analyses.

5.1 Feature Selection

Feature selection and complementary dependence analyses identify cycle operation as the dominant factor influencing RUL prediction, while interpretable model training supports transparent model comparison.

  • 5.1 Feature Selection: Cycle operation most strongly influenced RUL prediction across Pearson correlation, distance correlation, and LGBM-based feature importance analyses.These analyses are presented in Figure 7.
  • 5.1 Feature Selection: RCIT analyses with different initializations likewise identified cycle operation as the feature most impacting RUL prediction.The reported RCIT settings use a p-value of 0.01 and kernel size of 100.
  • 5.1 Feature Selection: Exploratory analysis and correlation assessment also confirmed that cycle operation affects RUL more than other operational input parameters.Figure 9 combines univariate, bivariate, and multivariate analyses.
  • 5.1 Feature Selection: The study trains tree, FIGS, EBM, and ReLU-DNN models using interpretable machine-learning practices and feature-selection methods.Feature selection is described as supporting model precision and interpretability.
  • 5.1 Feature Selection: ReLU-DNN exhibited superior performance, followed by EBM, FIGS, and tree models across MAE, MSE, and R2 evaluations.The comparison is reported in Table 2.

5.2 Interpretability of the Model

Local and global interpretability analyses examine how features contribute to RUL predictions across several model classes, with cycle operation repeatedly identified as influential.

  • 5.2.1 Local Interpretability: Local exact interpretability confirms that cycle operation significantly impacts RUL predictions across ReLU-DNN, EBM, FIGS, and tree models.The analysis uses feature weight, effect, and local importance views.
  • 5.2.2 Global Interpretability: LLM coefficient plots visualize each feature’s contribution, effect direction, and relative magnitude for global interpretability.The paper illustrates this analysis for ReLU-DNN and provides other model analyses in an appendix.

5.3 Trustworthy AI: Model Diagnosis and Validation

Model diagnosis compares candidate RUL predictors across accuracy, overfitting, reliability, robustness, and resilience, with ReLU-DNN showing the strongest overall performance.

  • ReLU-DNN exhibited the best overall performance, followed by tree and FIGS, across the reported evaluation matrices.The diagnosis covers accuracy, overfit, reliability, robustness, and resilience.
  • Residual plots compare training and testing behavior for ReLU-DNN, EBM, FIGS, and tree models.
  • ReLU-DNN and FIGS significantly outperformed EBM on accuracy, while ReLU-DNN showed superior predictive performance.Accuracy is evaluated using MSE, MAE, and R2.
  • EBM showed greater overfitting because of a large MSE gap between training and testing datasets.Overfitting is assessed through the difference between true and fitting errors.
  • Under perturbations across all features, EBM had inferior predictive performance, consistent with degradation under worst-case and out-of-distribution testing.The robustness and resilience assessments use covariate perturbation, worst-case subsampling, and out-of-distribution scenarios.

6 Conclusions

The conclusions position explainable and interpretable AI as central to trustworthy RUL prediction in digital twins while identifying several priorities for realistic deployment.

  • Explainable and interpretable AI is presented as indispensable for reliable RUL predictions within intelligent digital twin frameworks.The paper links these mechanisms to trust, transparency, and accountability.
  • The study applies ReLU-DNN, EBM, FIGS, and decision tree models to examine trustworthy RUL prediction locally and globally.The operating cycle feature is reported as consistently most influential across the analyses.
  • Future work targets explainable real-time inference that can handle multi-fidelity and scarce data, missing or approximate physics, and new environments.
  • The proposed future agenda includes sensor-location optimization, sensor degradation, recalibration, and signal reconstruction for system-degradation prediction.
  • Additional priorities are online model updating for temporal synchronization and uncertainty quantification for scarce, noisy, and previously unseen data.

A Global Model Interpretability

Global model interpretability is examined through feature-importance analyses for ReLU-DNN, EBM, and FIGS.

  • ReLU-DNN global interpretability is represented through LLM feature importance.
  • EBM global interpretability is represented through feature and effect importance.
  • FIGS global interpretability is represented through feature importance.

Declaration of Generative AI and AI-assisted technologies in the writing process

The authors disclose using GPT-4 for language editing and refinement during preparation of the work, followed by author review and editing.

  • GPT-4 was used for language editing and refinement during preparation of the work.
  • The authors reviewed and edited the content after using the tool or service.
  • The authors state that they retain full responsibility for the publication’s content.
Loading 2301.06676v2…