Hybrid State-Space Model for Asset End-of-Life Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting the end-of-life (EOL) of industrial assets, such as gas turbines and automobiles, face challenges due to non-linearities and complexities in real-life assets, requiring a known model that captures system dynamics, which is often not met in practice.
Innovation Solution
The development of a hybrid state-space model that combines physics-based and data-driven approaches, allowing for the construction of dynamical models from empirical training data, and incorporating data from a fleet of similar assets to provide a universal model that can be adapted for unit-specific predictions, including the use of Radial Basis Functions (RBFs) and system identification techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional Bayesian filtering techniques are used to predict remaining useful life, then prediction capability is achieved, but the method requires a known model capturing system dynamics which is often unavailable in practice
Solution Approach 1:
The patent transforms the approach from requiring a complete known model to using learnable parameters. The system identifies unknown system dynamics by learning parameters from operational data, converting a model-dependent problem into a data-driven parameter estimation problem that maintains prediction reliability without requiring complex pre-defined models
Solution Approach 2:
The patent replaces the traditional mechanical modeling approach (requiring known physical models and equations) with a data-driven learning approach. Instead of relying on mechanical understanding of system dynamics, the system uses observed data to infer behavioral patterns, substituting complex mechanical model requirements with statistical learning from operational data
2Reliability
If data from a fleet of similar assets is used to build a universal model, then model robustness is improved, but the model must be adapted for unit-specific predictions
Solution Approach 1:
The patent segments the modeling process into two distinct phases: first building a universal model from fleet-wide data that captures general patterns, then adapting this base model to individual units using their specific operational data. This segmentation allows the system to leverage both the robustness of aggregate data and the specificity of unit-level characteristics
Solution Approach 2:
The patent performs preliminary model building using fleet-wide data before adapting to individual units. By pre-establishing a universal model that captures common failure patterns and system behaviors across the fleet, the system reduces the amount of data and computational effort needed for unit-specific adaptation, while maintaining both robustness and specificity
Data Source
AI summary
In an example embodiment, a method of calculating end-of-life (EOL) predictions for a physical asset is provided. A state-space model for the physical asset is obtained, the state-space model being a physics-based model describing a state of the physical asset at a particular time given measurements or observations for the physical asset. Then a current state of the physical asset is inferred. Then a long-term prediction is derived for the physical asset based on the inferred current state of the physical asset and the state-space model for the physical asset. Then an EOL probability distribution function is generated for the physical asset based on the long-term prediction, the EOL probability distribution function describing a range of estimates of EOL for the physical asset and their corresponding confidence intervals.