Split Neural Network Classification for Edge Device Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In industrial control devices like power conversion devices and industrial controllers, the low performance and storage capacity of microcomputers make it difficult to execute neural network learning and improve classification accuracy, especially when the function is fixed and prone to unauthorized duplication.
Innovation Solution
An operating state classification system where a neural network is divided between an edge device and a server, with the edge device storing the pre-stage side and the server storing the post-stage side, allowing for additional learning on the server to improve classification accuracy while reducing the burden on the edge device and preventing unauthorized duplication.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire neural network is installed on the edge device for autonomous operation, then classification accuracy can be maintained, but the device complexity and storage requirements exceed the capabilities of microcomputers in industrial control devices
Solution Approach 1:
The neural network is divided into two segments: the pre-stage (feature extraction layer) is installed on the edge device, while the post-stage (classification layer) is installed on the server. This segmentation allows the edge device to perform initial processing with limited resources while maintaining overall system accuracy through server-based completion of the classification task.
2Ease of operation
If the neural network is divided between edge device and server, then the burden on the edge device is reduced, but the system requires continuous connection to the server which may expose security risks
Solution Approach 1:
Dimensionally compressed intermediate data serves as an intermediary between the edge device and server. This compressed representation makes it difficult to reconstruct the original sensor data or reverse-engineer the neural network parameters, thereby preventing unauthorized duplication while still enabling accurate classification through server processing.
3Measurement precision
If all sensor data is transmitted to the server for processing, then classification accuracy can be maintained, but the communication cost and data transmission volume increase significantly
Solution Approach 1:
The essential features are extracted from the raw sensor data through the pre-stage neural network on the edge device. Only these extracted feature amounts (intermediate data) are transmitted to the server, rather than transmitting the complete high-volume sensor data. This extraction process maintains classification accuracy while dramatically reducing communication overhead.
Data Source
AI summary
The objective of the present invention is to realize an operating state classification system having a classification accuracy that continuously improves through additional learning, and which is appropriately protected from unauthorized duplication of a classification function. In this operating state classification system, in which an edge device and a server are connected by means of a communication network, and which inputs sensor data into a neural network and outputs a state label: the edge device includes a first storage unit which stores an upstream side of the neural network, and a dimensionality reduction unit which inputs the sensor data into the upstream side of the neural network and outputs intermediate data; and the server includes a second storage unit which stores a downstream side of the neural network, an inference executing unit which inputs the intermediate data into the downstream side of the neural network and outputs the state label, and a learning unit which updates the downstream side of the neural network by means of additional learning.