Sensor-Based Task Clustering for Annotation-Free Cycle Time Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In smart manufacturing, detecting task and cycle times for workers is challenging due to the diversity of individual behavior, and existing machine learning-based solutions require significant costs for data preparation and model customization.
Innovation Solution
A method involving the extraction of features from time-series sensor data, clustering these features into tasks, and calculating cycle times using a cycle pattern model, which reduces the need for data annotation and customization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning-based task and cycle time detection is implemented using RGB or depth cameras, then detection accuracy is improved, but deployment cost increases due to data preparation and model customization requirements
Solution Approach 1:
The patent creates a universal task and cycle time detection system that works across multiple manufacturing sites without requiring site-specific customization. The system uses pre-trained machine learning models that can generalize to different environments, eliminating the need for each site to prepare their own annotated datasets and customize models, thus reducing deployment costs while maintaining detection accuracy
Solution Approach 2:
The patent performs data preparation and model training in advance at a central location before deployment. By pre-processing the data and pre-training the models beforehand, the system eliminates the need for on-site data annotation and model customization, significantly reducing the time and cost required for deployment at each manufacturing site
2Ease of manufacture
If traditional task and cycle time detection methods are used, then deployment cost is reduced, but detection capability is insufficient due to diversity of individual worker behavior
Solution Approach 1:
The patent introduces machine learning models as intermediaries between the raw sensor data and the task/cycle time detection process. These models automatically learn to handle the diversity of worker behaviors from training data, enabling accurate detection without requiring complex manual configuration or high deployment costs, thus bridging the gap between low cost and high capability
3Measurement precision
If custom machine learning models are developed for each manufacturing site, then detection accuracy is improved, but implementation time increases due to data preparation and model customization
Solution Approach 1:
The patent develops a universal detection system with pre-trained models that can be deployed across multiple manufacturing sites without requiring site-specific customization. This approach maintains high detection accuracy while dramatically reducing implementation time by eliminating the need for each site to go through the lengthy processes of data collection, annotation, and model training
Data Source
AI summary
Example implementations described herein involve systems and methods that can involve extracting features from each of a plurality of time-series sensor data, the plurality of time-series sensor data associated with execution of one or more operations; clustering the extracted features into a plurality of tasks that occur from execution of the one or more operations, each of the plurality of tasks associated with a clustering identifier (ID) from the clustering; and calculating a cycle time of the cycle based on the initiation and end of the cycle recognized by referencing a cycle pattern model, wherein the cycle pattern model comprises configuration information of a cycle including a set from a plurality of the clustering IDs.