Failure Prediction Model for Non-Linear Degradation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Predicting non-linear performance degradation in mechanical systems, such as wear-induced deterioration in progressing cavity pumps, is challenging due to its gradual nature, imbalanced training datasets, and the difficulty in identifying exclusive failure signatures, which hinders accurate predictive modeling.
Innovation Solution
A computer-implemented method using a failure prediction model that clusters data samples with look-back windows and prediction horizons, calculating normalcy and anomaly scores, and outputting classifications based on these scores to predict performance degradation without relying on over-parameterized neural networks or class balancing techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If highly parameterized nonlinear methods (e.g., ANNs) are used to predict non-linear performance degradation, then the model can capture complex non-linear relationships, but the number of parameters is massive and requires abundant data for fine-tuning and optimization, which is not available in imbalanced datasets
Solution Approach 1:
The patent segments the prediction task into two distinct stages: (1) a feature extraction stage using a lightweight autoencoder to learn compact representations from imbalanced data, and (2) a prediction stage using a simple linear model on the extracted features. This segmentation avoids the need for massive parameters while capturing non-linear relationships through the autoencoder's latent space representation.
Solution Approach 2:
The patent introduces an intermediary component - the autoencoder's latent space representation - that bridges the gap between raw imbalanced data and the final prediction. This intermediary transforms the complex non-linear data into a compressed feature space that can be effectively modeled with simple linear methods, eliminating the need for highly parameterized models.
2Quantity of substance
If over-sampling and/or under-sampling methods are employed to remedy imbalanced data, then the number of examples in the anomaly state increases, but these methods can skew the original distribution in the data and thus bias the solution (e.g., decreased accuracy by virtue of increased false positive indications)
Solution Approach 1:
The patent extracts the essential information from imbalanced data through the autoencoder's feature extraction capability, taking out only the relevant patterns and relationships needed for prediction. This extraction process works directly with the original data distribution without requiring artificial balancing, thereby avoiding the introduction of bias while still capturing sufficient anomaly information for accurate prediction.
3Device complexity
If linear models (e.g., Naïve Bayes, SVMs) are used for prediction, then the model is simple and requires less data, but the relationship between normal and worn states is nonlinear which makes linear models inapplicable
Solution Approach 1:
The patent transforms the problem from a direct linear classification in the original feature space to a linear classification in a transformed latent space. The autoencoder maps the non-linear relationships from the original space into a new dimensionality where linear models become applicable, thus combining the simplicity of linear models with the capability to handle non-linear relationships.
Data Source
AI summary
Described are techniques for predicting gradual performance degradation with non-linear characteristics. The techniques including a method comprising inputting a new data sample to a failure prediction model, wherein the failure prediction model is trained using a labeled historical dataset, wherein respective data points are associated with a look-back window and a prediction horizon to create respective training samples, wherein the respective training samples are clustered in a plurality of clusters, and wherein the plurality of clusters are each associated with a normalcy score and an anomaly score. The method further comprises outputting a classification associated with the new data sample based on comparing a first anomaly score of a first cluster of the plurality of clusters that includes the new data sample to an average anomaly score of clusters of the plurality of clusters having the normalcy score greater than the anomaly score.