Conformal Filtering for Asset RUL Prediction Under Sensor Noise
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Reliably predicting the transition from a quasi-steady degradation phase to an accelerated degradation phase in engineering assets is challenging due to noisy and stochastic sensor measurements, making it difficult to accurately estimate remaining useful life (RUL).
Innovation Solution
Applying conformal mathematical filters to preprocess noisy sensor data to conform to known wear curve formulations, followed by using sequence-based machine learning models with asymmetric loss functions to predict RUL, ensuring conservative predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conformal mathematical filters are applied to preprocess sensor data, then noise is reduced and prediction accuracy is improved, but device complexity increases
Solution Approach 1:
A conformal mathematical filter is introduced as an intermediary component between the sensor measurements and the machine learning model. This filter preprocesses the noisy sensor data by conforming it to a wear curve formulation, thereby reducing noise and improving the quality of input data for the RUL prediction model without requiring changes to the core prediction algorithm
Solution Approach 2:
The conformal filter performs preliminary processing of the sensor data before it is fed into the machine learning model. By pre-conforming the measurements to the wear curve formulation and resampling the output, the system prepares cleaner, more structured data in advance, which improves prediction accuracy while keeping the main model relatively simple
2Reliability
If asymmetric loss functions with increased penalty for overprediction are used, then reliability of predictions is improved, but manufacturing precision of the model requires higher control
Solution Approach 1:
An asymmetric loss function is implemented in the machine learning model where the penalty for overprediction is intentionally made heavier than the penalty for underprediction. This asymmetric penalty structure guides the model to produce more conservative and reliable RUL estimates, prioritizing safety and reliability over symmetric accuracy
Solution Approach 2:
The loss function parameters are specifically configured to create an asymmetric penalty structure. By adjusting the weight parameters in the loss function (with higher weight for overprediction errors), the model training process is directed toward producing more reliable predictions that favor underprediction, thereby improving overall system reliability
Data Source
AI summary
A system determines that an asset of an engineering system has transitioned from a quasi-steady degradation stage to an accelerated degradation phase based on sensor measurements received from an asset. During the accelerated degradation phase, features are extracted from the sensor measurements that are indicative of wear of the asset. A conformal mathematical filter is applied to the features that causes the features to conform to a wear curve formulation associated with the asset. An output of the filter is resampled to form a noise-reduced signal. The noise-reduced signal is input into a sequence machine learning model. A loss function of the sequence machine learning model uses an increased penalty to overprediction and a relaxed penalty for underprediction. An output of the sequence machine learning model is used to predict a remaining useful life (RUL) of the asset.