Edge Inference Computing for Manufacturing Data Latency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing user-cloud computing architecture in industrial production faces challenges with high network bandwidth requirements, delayed data processing, and resource burden on cloud devices due to large volumes of production data, which affects the timeliness and efficiency of inference computing.
Innovation Solution
Implementing a user-side-edge-side-cloud computing architecture that utilizes an inference computing apparatus with a GPU cluster to perform inference tasks closer to the user-side device, reducing data upload and leveraging cloud resources for model training without real-time decision-making, and updating inference models based on performance evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If all production data is uploaded to cloud devices for inference computing, then cloud computing power and model training capabilities are improved, but network bandwidth requirements increase and data processing time is delayed
Solution Approach 1:
The system segments the computing architecture into three distinct layers: user-side devices for data collection and preliminary processing, edge computing devices for real-time inference computing with GPU clusters, and cloud devices for model training and management. This segmentation allows different computing tasks to be performed at appropriate locations, reducing the need to upload all production data to the cloud and thereby decreasing network bandwidth requirements and data processing time.
Solution Approach 2:
Edge computing devices serve as intermediaries between user-side production devices and cloud computing resources. These edge devices receive production data, perform inference computing using deployed inference models, and only upload necessary results or updated training samples to the cloud, rather than uploading all raw production data. This intermediary role significantly reduces network bandwidth consumption and accelerates data processing.
2Measurement precision
If large volumes of production data are uploaded to cloud devices, then inference computing accuracy is improved, but resource burden on cloud devices increases
Solution Approach 1:
The system extracts the inference computing function from the cloud and places it at the edge computing layer. Edge computing devices download inference models from the cloud and execute them locally on production data, extracting only the necessary computational logic from the cloud. This leaves the cloud resources free from the burden of processing large volumes of production data, while still maintaining high inference accuracy through locally executed models.
Solution Approach 2:
The architecture transitions from a traditional two-dimensional user-cloud structure to a three-dimensional user-edge-cloud hierarchy. This dimensional change enables inference computing to occur at the edge layer, intermediate between user devices and cloud resources. Production data is processed at the edge dimension rather than being transmitted to the cloud dimension, reducing cloud resource burden while preserving inference computing accuracy.
3Measurement precision
If inference models are continuously updated with all production data, then model accuracy is improved, but training time and computational resources are increased
Solution Approach 1:
The system applies partial action by selectively uploading only certain production data samples to the cloud for model retraining, rather than uploading all production data. Edge computing devices identify and upload only those data samples that are most valuable for model improvement, such as samples with high uncertainty or representative of new production conditions. This partial uploading approach maintains model accuracy while significantly reducing training time and computational resources required.
Data Source
Figure 1~3
Figure 4
Figure 5
AI summary
An inference computing apparatus includes at least one processor and a memory with program instructions stored therein, the program instructions can be executed by the at least one processor to cause the inference computing apparatus to perform following operations: receiving a first inference model from a model training apparatus, the first inference model being obtained through a model training by the model training apparatus based on a first training sample library, the first training sample library comprising training samples from historical data generated in a manufacturing stage, the model training apparatus comprising a cloud device; performing an inference computing on data to be processed generated in the manufacturing stage based on the first inference model to obtain the inference result which is sent to a user-side device; and evaluating performance of the first inference model to determine whether the first inference model needs to be updated, and if yes, updating the first inference model.