Physically Explainable Bearing Fault Detection via Feature Attribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fault detection methods for bearings in rotating machinery, particularly those using machine learning, are complex and lack interpretability, making it difficult for domain experts to understand the reasoning behind the predictions and trust the results, and they struggle with noisy sensor data and confounding factors.
Innovation Solution
A method and apparatus that utilize a fault detection model to map sensor data from an input domain to a selected domain with physical meaning, applying feature attribution to quantify the importance of individual features in this domain, thereby providing interpretable fault information and root-cause analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are used to detect faults in bearings, then detection accuracy and robustness are improved, but the models become complex and not human-understandable (black-box algorithms)
Solution Approach 1:
The patent introduces an explanation layer as an intermediary component between the black-box machine learning model and the end user. This explanation layer processes the model's internal representations and generates human-understandable explanations, thereby mediating between the complex model and the need for interpretability without modifying the core detection model itself.
Solution Approach 2:
The patent segments the fault detection system into distinct functional components: the core machine learning model for detection, an explanation layer for interpretability, and a visualization interface for user interaction. This segmentation allows each component to be optimized independently while maintaining overall system performance.
2Reliability
If complex machine learning models are deployed for fault detection, then detection capability is improved, but domain experts cannot understand or trust the results
Solution Approach 1:
The explanation layer serves as an intermediary that translates the complex internal representations of the machine learning model into domain-expert-friendly explanations. It processes feature importance scores and generates interpretable outputs that align with domain knowledge, enabling experts to understand and trust the model's predictions.
Solution Approach 2:
The patent employs visual encoding techniques where different colors represent different levels of feature importance or different types of fault indicators. This visual representation makes it easier for domain experts to quickly comprehend the model's reasoning and the significance of various features in the fault detection process.
3Loss of information
If feature attribution methods are applied to explain model predictions, then interpretability is improved, but computational overhead increases
Solution Approach 1:
The patent applies feature attribution methods selectively rather than comprehensively to all model operations. The explanation layer focuses on generating explanations only when needed (e.g., when faults are detected or when user requests explanations), rather than continuously computing attributions for all predictions, thereby reducing overall computational overhead.
Solution Approach 2:
The explanation layer is pre-integrated into the model architecture during the training phase, allowing it to learn efficient explanation generation strategies alongside the main detection task. This preliminary integration enables the system to produce explanations with minimal additional computational cost during inference.
Data Source
Figure 1
Figure 1~2
Figure 2A
AI summary
Fault detection apparatus and computer-implemented method for providing physically explain-able fault information of a bearing built in a machine by a fault detection model (11), comprising the steps: - obtaining sensor data measured at the bearing as input data relating to an input data domain and the fault detection model (11) which is trained on sensor data related to said input data domain to output a predicted failure value (12) of the bearing by processing the obtained sensor data (10), - mapping the measured sensor data (10) from the input data domain to a selected data domain resulting in an augmented fault detection model (13) which outputs augmented predicted failure value related to the selected data do-main, wherein the selected data domain has a physical meaning to the fault of the bearing, - performing a feature attribution (14) on the augmented fault detection model (13) quantifying an importance of at least one individual feature to the augmented failure value related to the selected data domain, and - displaying (S4) the individual feature (15) and the respective quantified importance (16) in the selected data domain at a user interface.