Action Analysis Using Pretrained Behavior Recognition for Process Timing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current action analysis methods require significant annotation effort and high computational resources for machine learning, leading to increased operational burdens and introduction costs.
Innovation Solution
An action analysis device that acquires behavior data recognized by a preliminarily learned model and determines the required time for a process based on this data, reducing the need for extensive annotation and machine learning processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are used to recognize operator postures and analyze operations, then action analysis capability is improved, but annotation burden and operational complexity increase
Solution Approach 1:
The system performs preliminary learning in advance to create a general-purpose action recognition model that can be applied across multiple factories without requiring site-specific annotation. The model is pre-trained on diverse operation data to recognize various postures and actions, eliminating the need for each factory to perform time-consuming annotation work while maintaining high action analysis capability
Solution Approach 2:
The invention creates a universal action recognition model that can be applied across multiple factories and operation types without requiring factory-specific customization. The model serves multiple functions including posture recognition, operation analysis, and time measurement across different production sites, reducing the need for separate annotation efforts at each location
2Measurement precision
If machine learning processing is performed at each factory, then localized action analysis is improved, but introduction cost and computational resource requirements increase
Solution Approach 1:
The invention extracts the complex machine learning processing from individual factory systems and consolidates it into a centralized learning server. The learning server performs all heavy computational tasks including model training and inference, while factory terminals only handle lightweight data transmission and result display, dramatically reducing the computational resources needed at each factory location
Solution Approach 2:
The invention introduces a learning server as an intermediary between data collection and action analysis. The learning server acts as a mediator that receives operation data from multiple factories, performs centralized machine learning processing, and returns analysis results to the factories, eliminating the need for each factory to maintain complex computational infrastructure
3Reliability
If sufficient GPUs and learning time are provided for on-site learning, then model accuracy is improved, but introduction cost and implementation time increase
Solution Approach 1:
The system performs all model learning and accuracy optimization in advance at a centralized server before deployment to factories. The model is pre-trained on extensive data to achieve high accuracy, eliminating the need for time-consuming on-site learning processes that would require significant GPU resources and extended learning periods at each factory location
Data Source
AI summary
An action analysis device includes an acquisition unit that acquires behavior data indicating a behavior of an object during an operation process, which has been recognized by a model preliminarily learned for recognizing the object; and a determination unit that determines a required time for a process corresponding to the behavior data based on the behavior data acquired by the acquisition unit. For example, based on time information set as the required time for the process, the determination unit determines a required time for a process corresponding to the behavior data.