Mechanical Soft Sensors Using SINDy for Low-Overhead Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing soft sensors for mechanical systems are often unreliable, difficult to troubleshoot, and require significant computing resources, making them inefficient and costly for embedded systems.
Innovation Solution
The method involves defining a training dataset of sensor values, training a Sparse Identification of Nonlinear Dynamics (SINDy) model to generate a mathematical function relating sensor values to soft sensor values, and embedding machine-readable instructions based on this function into an embedded system, allowing it to generate soft sensor values efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are used to create soft sensors, then the soft sensor can generate estimates of physical quantities, but the computing overhead (memory and processor power) increases significantly
Solution Approach 1:
The patent replaces complex neural network computations with a simplified mathematical model (polynomial function) that can be evaluated efficiently. Instead of using heavy machine learning infrastructure, the invention uses a lightweight analytical approach where sensor measurements are directly transformed into physical quantity estimates through pre-determined mathematical relationships, dramatically reducing processor power and memory requirements while maintaining estimation capability
Solution Approach 2:
The patent changes the computational parameters from complex neural network weights and activations to simple polynomial coefficients. By transforming the estimation problem into a parameter-based mathematical evaluation rather than iterative neural network computation, the system achieves the same estimation function with significantly reduced computing overhead suitable for embedded systems
2Reliability
If neural networks are used to create soft sensors, then the soft sensor can generate estimates of physical quantities, but the system becomes difficult to troubleshoot and customize
Solution Approach 1:
The patent replaces the black-box neural network architecture with an interpretable mathematical model consisting of polynomial equations with explicit coefficients. This substitution enables engineers to understand, verify, and modify the estimation logic directly through mathematical inspection rather than relying on opaque neural network internals, significantly improving troubleshootability and ease of customization
Solution Approach 2:
The patent enables the soft sensor system to be self-explanatory and self-modifiable through its mathematical formulation. The explicit polynomial relationships allow users to independently verify calculations, adjust parameters, and troubleshoot issues without requiring specialized machine learning expertise or complex debugging tools, making the system more maintainable and adaptable
Data Source
AI summary
A method of generating a soft sensor for a mechanical system includes defining a training dataset of sensor values received from at least one first sensor corresponding a first physical quantity. The method further includes training a Sparse Identification of Nonlinear Dynamics (SINDy) model using the training dataset to generate a mathematical function defining a relationship between the sensor values and soft sensor values corresponding to a second physical quantity. The method further includes embedding machine readable instructions based on the mathematical function into an embedded system of the mechanical system including at least one first sensor. The machine readable instructions, when executed by the embedded system, cause the embedded system to receive sensor values corresponding to the first physical quantity and generate soft sensor values based on the sensor values and the mathematical function.