Industrial Machine Failure Prediction with Reference Sensor Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex industrial systems present challenges in monitoring due to a low number of malfunctions per component, unique configurations, and uncorrelated sensed information across components, leading to insufficient training data for machine learning models, which decreases their accuracy.
Innovation Solution
A method is introduced that selects reference sensors based on similarities with the target components, uses multiple mappings to associate sensed information with predicted failures, and employs pre-training and fine-tuning of models to enhance monitoring accuracy, allowing for accurate monitoring of industrial systems with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained using only component-specific sensed information, then the model learning process is simple, but the accuracy decreases due to insufficient training data
Solution Approach 1:
The patent combines sensed information from multiple reference components with the target component's sensed information to create an expanded training dataset. This merging allows the machine learning model to learn from a larger volume of data while maintaining relevance to the target component, thereby improving model accuracy without requiring more physical training samples from the target component itself.
Solution Approach 2:
Reference components serve as intermediaries that bridge the gap between limited target component data and the need for comprehensive training data. By selecting reference components with similar characteristics and using their sensed information as a proxy, the system can indirectly augment the training dataset while maintaining the relationship between sensed information and component failures.
2Quantity of substance
If reference sensor data from multiple components is used to augment training data, then the training data volume increases, but the data quality decreases due to uncorrelated sensed information across different components
Solution Approach 1:
The patent applies local quality by selecting reference components based on their similarity to the target component. Not all reference components are treated equally; instead, the system identifies and uses only those reference components whose sensed information characteristics closely match the target component. This selective approach ensures that the augmented training data maintains high quality and relevance, avoiding the inclusion of uncorrelated or irrelevant data from dissimilar components.
Solution Approach 2:
The system transforms and adapts sensed information from reference components to match the characteristics of the target component's sensed information. By applying parameter changes such as normalization, scaling, or feature transformation, the patent ensures that reference data can be effectively integrated with target component data while maintaining consistency in data quality and correlation patterns.
3Ease of operation
If sensed information formats are equalized across all components, then data integration becomes easier, but information loss occurs due to the loss of component-specific characteristics
Solution Approach 1:
The patent segments the data integration process into two distinct stages: format equalization and characteristic preservation. First, sensed information from reference and target components is standardized in format to enable integration. Second, component-specific characteristics are preserved through selective feature extraction and weighting, ensuring that the equalization process does not erase important individual differences. This segmented approach allows the system to achieve both ease of integration and retention of critical information.
Data Source
AI summary
A method for monitoring an industrial machine, the method may include selecting a selected reference component for each component out of one or more components of the industrial system; the selecting provides one more selected reference components; wherein each selected reference component is selected out of multiple reference components; wherein for each component the selecting is based on similarities between the component and reference components of the multiple reference components; determining one or more learnt mappings between (i) sensed information related to the one or more components and (ii) predicted failures of the one or more components ; wherein the learning is based, at least in part, on one or more reference mappings between (i) sensed information related to the one or more selected reference components and (ii) predicted failures of the one or more selected reference components; monitoring the one or more components to receive monitoring results; and evaluating the operation of the one or more components based on the monitoring results and the one or more reference mapping.