Sensor Data Preprocessing for Low-Data Abnormality Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality predictor diagnosis systems require a large amount of sensor data and knowledge to generate a physical model, making it impractical to implement without initial setup and extensive data collection.
Innovation Solution
A system that uses a preprocessor to generate a physical model by combining sensor data with domain knowledge, adding latent variables through nonlinear functions, and performing correlation coefficient calculations to create a mathematical expression for abnormality prediction without needing extensive sensor data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a physical model is generated using traditional methods (JP 2015-179443 A), then diagnosis accuracy is improved, but a large amount of sensor data and initial setup are required
Solution Approach 1:
The patent introduces an intermediary processing step that transforms raw sensor data into a simplified representation suitable for physical model generation. The system extracts essential features and relationships from sensor data, creating an intermediate structured format that enables model generation without requiring large volumes of raw data. This intermediary processing layer bridges the gap between raw data and the physical model, reducing data requirements while maintaining diagnostic accuracy.
2Reliability
If a physical model is generated using traditional methods, then diagnostic capability is improved, but device complexity and initial setup are increased
Solution Approach 1:
The patent divides the physical model generation process into distinct modular segments: sensor data acquisition, feature extraction, relationship identification, and model construction. Each segment performs a specific function and can be independently configured or adjusted. This segmentation reduces overall system complexity by breaking down the complex model generation task into manageable, well-defined steps that are easier to implement and maintain.
Solution Approach 2:
The system performs preliminary processing of sensor data before physical model generation, including data validation, feature extraction, and preliminary relationship analysis. This preliminary action prepares the data in advance, reducing the complexity of the subsequent model generation step and eliminating the need for extensive initial setup. The preliminary processing ensures that the input data is ready for efficient model construction.
3Extent of automation
If machine learning methods are used to generate a physical model, then model generation is automated, but large amounts of data are still required
Solution Approach 1:
The patent extracts essential features and key relationships from sensor data, separating the critical information needed for physical model generation from the voluminous raw data. By taking out only the necessary features and relationships, the system achieves automated model generation without requiring large amounts of raw sensor data. This extraction process identifies and isolates the most relevant data elements for model construction.
Solution Approach 2:
The system transforms the data representation by changing parameters from raw sensor readings to extracted features and relationships. This parameter transformation converts voluminous raw data into a compact feature space that retains the essential information needed for model generation. The parameter change enables automated processing with reduced data volume by operating in the transformed feature domain rather than the original sensor data domain.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A preprocessor into which time-series sensor data is input, the preprocessor including: a correspondence table associating sensor data with a physical unit; a list associating a physical unit with a nonlinear function; a physical unit extraction unit that obtains sensor data, and specifies a physical unit from the obtained sensor data on the basis of the correspondence table; a latent variable generation unit that specifies, on the basis of the list, a nonlinear function from the physical unit specified by the physical unit extraction unit, and generates a latent variable for applying the sensor data to the specified nonlinear function; a latent variable addition unit that obtains time-series sensor data and sets the data as a variable, applies the obtained time-series sensor data to the latent variable generated by the latent variable generation unit to generate time-series data and sets the data as a variable, and adds the generated time-series data to the obtained time-series sensor data to generate a data set of a plurality of variables; and a relational expression estimation processing unit that analyzes a relationship between the variables of the data set generated by the latent variable addition unit, and generates a linear relational expression.