Edge Inference Model Management Using Multi-Model Feature Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in efficiently creating and managing models for various edge devices across different production control systems, such as substrate processing, electric apparatuses, and pharmaceuticals, due to the complexity of integrating and optimizing learning models across multiple devices.
Innovation Solution
An information processing method that involves acquiring feature values from multiple learning models, performing learning on a second model to estimate results, and inputting these values to output efficient inference models for edge devices, utilizing a system with an apparatus group server and edge devices connected via a communication network to generate and manage learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple learning models are used for different edge devices, then the adaptability to various production control systems is improved, but the device complexity and difficulty of management increase
Solution Approach 1:
The patent implements a unified learning model management system that can handle multiple types of learning models (first learning models for individual edge devices and second learning models for group-level inference) through a common architecture. The system uses standardized data collection, feature extraction, and model training processes that work across different production control systems including substrate processing, electric apparatuses, electronic apparatuses, automobiles, pharmaceuticals, chemical products, and foods, thereby achieving multi-functionality without proportionally increasing management complexity.
2Measurement precision
If learning models are created for each edge device individually, then the measurement precision and reliability of device-specific inference are improved, but the productivity and efficiency of model creation decrease
Solution Approach 1:
The patent divides the learning model creation process into two segments: first learning models trained individually for each edge device using device-specific data to ensure high precision for device-specific inference, and second learning models trained on aggregated data from multiple edge devices to enable efficient group-level inference. This segmentation allows simultaneous achievement of device-specific precision and overall creation efficiency through parallel processing and data sharing.
Solution Approach 2:
The patent merges data from multiple edge devices to train second learning models, combining datasets from different production control systems to create generalized inference models. This merging approach maintains the precision benefits of individual device modeling while achieving productivity gains through shared training data and consolidated model management, reducing redundant computation and resource consumption.
3Adaptability or versatility
If data from multiple edge devices is aggregated for learning, then the adaptability and generalization capability are improved, but the loss of time and computational resources increase
Solution Approach 1:
The patent performs preliminary data preprocessing and feature extraction at the edge device level before aggregation, preparing data in advance for efficient centralized model training. This preliminary action reduces the time required for data aggregation by ensuring data is already in the appropriate format and contains extracted features, thereby minimizing computational overhead during the model training phase while maintaining generalization capability.
Data Source
AI summary
To provide an information processing method, an information processing apparatus, and an information processing system. Acquiring a feature value of data processed by a plurality of first learning models, performing learning of a second learning model that outputs information relating to an estimation result in a case where the feature value of data processed by the first learning model is input based on the acquired feature value, and inputting the acquired feature value of data into the second learning model after learning to output an estimation result based on information obtained from the second learning model are included.