Virtual Sensor Training for Non-Differentiable Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for optimizing sensor parameters for deep learning-based object identification in automated vehicles and robots are inefficient, leading to suboptimal performance due to reliance on differentiable models that are limited in applicability and accuracy, particularly for non-differentiable sensor parameters and complex imaging systems.
Innovation Solution
A training method that determines and optimizes both sensor parameters and neural network models by generating sensor data sets with correct answer information, training neural network models, calculating identification performance, and selecting the best pair of sensor parameters and models to improve identification performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional differentiable models are used to optimize sensor parameters, then the optimization process can be performed, but the applicability is limited and accuracy is suboptimal for non-differentiable parameters and complex imaging systems
Solution Approach 1:
The patent creates a virtual sensor that replicates the optical characteristics and imaging properties of a real sensor through software simulation. This virtual copy can be configured with any parameter values and used for training neural networks without requiring physical sensor modifications or differentiable model assumptions, thereby extending applicability to non-differentiable parameters while maintaining high identification accuracy through realistic optical modeling
Solution Approach 2:
The patent replaces the need for physical sensor parameter adjustment and complex differentiable optical models with a computational virtual sensor system. By substituting mechanical/optical parameter optimization with software-based virtual imaging and neural network training, the system achieves high accuracy for both differentiable and non-differentiable parameters without relying on traditional differentiable model constraints
2Reliability
If sensor parameters are optimized based on deep learning identification performance, then identification performance can be improved, but the process requires complex formulations and multiple iterations
Solution Approach 1:
By creating a virtual sensor that replicates real sensor behavior, the patent enables direct training of neural networks on synthesized data without requiring complex differentiable formulations of optical systems. This virtual copy approach simplifies the optimization process while maintaining high identification performance, as the virtual sensor can generate diverse training data for various parameter configurations without physical constraints
Solution Approach 2:
The patent performs preliminary generation of virtual sensor data and neural network training before actual sensor deployment or physical parameter adjustment. By pre-training models using virtual sensor simulations with various parameter settings, the system identifies optimal parameter configurations in advance, reducing the need for complex iterative optimization after deployment and simplifying the overall process
Data Source
AI summary
The training system determines a plurality of sensor parameter candidates to be used for an operation of a sensor, generates a plurality of sensor data sets corresponding to each of the plurality of sensor parameter candidates and including sensor data to be obtained by the operation of the sensor and a plurality of pieces of correct answer identification information corresponding to each of the sensor data, generates a plurality of trained neural network model candidates corresponding to the plurality of sensor parameter candidates, calculates identification performance of the plurality of trained neural network model candidates, selects a pair of the trained neural network model candidate with the highest identification performance and the sensor parameter candidate corresponding to the trained neural network model candidate with the highest identification performance, and outputs the selected pair of the sensor parameter candidate and the trained neural network model candidate.