Virtual Sensor Data Generation for Work Machine ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems face challenges in generating sufficient machine learning data for neural network models used in physical work machines, as obtaining appropriate operation commands for varied and amorphous subjects, such as pieces of fried chicken, is difficult with deterministic algorithms, requiring extensive time and cost to prepare real data.
Innovation Solution
A machine learning data generation device that virtually generates sensor inputs and operation commands based on virtual subject models, allowing for simulation of physical work in a virtual space to produce machine learning data efficiently, including a virtual sensor input generator, virtual operation command generator, simulator, and machine learning data generator.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real physical work is performed to collect training data, then data quality and authenticity are improved, but time consumption and costs increase significantly
Solution Approach 1:
The patent creates virtual copies of physical objects (subjects) in a virtual space that replicate their geometric and physical properties. These virtual subjects are generated from real subject data through 3D scanning and modeling, allowing the system to simulate physical work scenarios without requiring actual physical objects. This copying approach maintains data authenticity while eliminating the time and resource costs of physical data collection.
Solution Approach 2:
The patent introduces a virtual space as an intermediary between the physical world and the machine learning model training. This virtual environment acts as a mediator that translates real-world physical properties into simulated scenarios, enabling data generation without direct physical interaction. The virtual space preserves the essential characteristics of physical work while removing the constraints of physical resource limitations.
2Adaptability or versatility
If diverse subjects are used for training, then model adaptability is improved, but data preparation complexity increases
Solution Approach 1:
The patent segments the data preparation process into distinct modules: subject data acquisition, 3D modeling, virtual scene generation, and simulation execution. Each module handles specific aspects of diversity independently, allowing the system to manage complex varied subjects through structured decomposition. This segmentation enables diverse training data generation while keeping the overall system manageable through modular architecture.
Solution Approach 2:
The patent utilizes parameter changes to generate diversity in virtual subjects by modifying geometric parameters, material properties, and environmental conditions in the virtual space. Instead of requiring physically diverse objects, the system achieves variety by systematically varying parameters of virtual models, thereby improving model adaptability while simplifying data preparation through controlled parameter manipulation rather than physical object collection.
3Quantity of substance
If extensive real data is collected, then training comprehensiveness is improved, but costs and resource requirements increase
Solution Approach 1:
The patent generates large quantities of training data by creating and manipulating virtual copies of subjects in a simulated environment. These virtual subjects can be replicated indefinitely without consuming additional physical resources, allowing the system to achieve high data quantities while maintaining constant resource requirements. The virtual copying process eliminates the need for proportional physical resource investment as data volume increases.
Solution Approach 2:
The virtual space system is self-sufficient for data generation, creating its own training datasets without requiring external physical resources proportional to data volume. The system uses pre-acquired subject models to generate unlimited variations through virtual manipulation, making the data generation process independent of ongoing physical resource investment and enabling scalable data production with fixed resource overhead.
Data Source
AI summary
Provided is a machine learning data generation device including: a virtual sensor input generator configured to generate a virtual sensor input, which is obtained by virtually generating a sensor input obtained as a result of performing sensing, by a sensor of a work machine, on a plurality of randomly piled subjects to be subjected to physical work by an operating machine of the work machine; a virtual operation command generator configured to generate a virtual operation command, which is obtained by virtually generating an operation command for the operating machine of the work machine; a virtual operation outcome evaluator configured to evaluate an outcome of the physical work in response to the virtual operation command in a virtual space; and a machine learning data generator configured to generate machine learning data based on the virtual sensor input, the virtual operation command, and the evaluation of the virtual operation outcome evaluator.