Interpretable Machine Learning Approach for Reliability Analysis

More, Kalpesh; Mathpati, Yogesh; Tripura, Tapas; Nayek, Rajdip; Alam, Syed Bahauddin; Chakraborty, Souvik

Permalink

https://hdl.handle.net/2142/121822 Copy

Description

Title

Interpretable Machine Learning Approach for Reliability Analysis

Author(s)

• More, Kalpesh
• Mathpati, Yogesh
• Tripura, Tapas
• Nayek, Rajdip
• Alam, Syed Bahauddin
• Chakraborty, Souvik

Issue Date

2023

Keyword(s)

• Reliability
• Equation discovery
• Bayesian
• Uncertainty

Abstract

The reliability assessment of stochastic dynamic systems is a crucial issue, specifically for complex engineering systems. Mathematically, reliability can be estimated as the probability of not failing while meeting particular objective functions and constraints. It is achieved by shrinking the area under the probability distribution function (PDF) while moving the average value. There are well-established techniques available in the literature for reliability estimation; however, the reliability analysis of existing systems, particularly complex interdependent and integral structures, is often overlooked despite being an equally significant problem. It is widely recognized that the behavior of structures can change over time due to degradation. Understanding, controlling, and mitigating component degradation are key priorities for complex engineering assets. As expensive engineering systems age beyond their design lifetimes, it is important to ensure reliability: detect and track degradation and changes in degradation rates; monitor system components for degradation; classify and characterize their degradation modes; and perform prognosis of their future state. Similarly, industrial systems that undergo multiple maintenance tasks and component replacements can also experience alterations in their governing physics, making it challenging to estimate their reliability using traditional methods that rely on a model based on the design blueprint of the system [1]. To address this issue, We propose and develop an innovative approach named "model-agnostic reliability analysis framework". This development has been published [2], and now we are extending this tool for trustworthy reliability analysis of complex nuclear systems with Missouri S&T. This method integrates Bayesian statistics, interpretable machine learning, and identification of stochastic dynamic equations (SDEs) to estimate the reliability of complex systems with unknown/approximate physics. In the conference presentation, we will demonstrate the effectiveness of our development for complex systems while extending the test cases for nuclear systems and structures through numerical examples, highlighting its potential application in reliability analysis in the domain of nuclear systems.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121822

Owning Collections