Industrial Image Monitoring With Selective OBIA Layer Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial automation control systems face inefficiencies in training and retraining object-based image analysis (OBIA) models due to resource-intensive processes, particularly when unexpected classifications occur, leading to undesirable retraining of the entire model and resource allocation issues.
Innovation Solution
Implementing incremental retraining of specific property layers within the OBIA model based on identified anomalies, allowing for targeted updates without disrupting the entire system, thereby reducing processing power requirements and improving model robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire OBIA model is retrained when unexpected classifications occur, then model accuracy is improved, but resource consumption and processing time increase significantly
Solution Approach 1:
The OBIA model is divided into multiple independent property layers (spectral, texture, spatial, contextual layers). When an unexpected classification occurs, only the specific property layer responsible for the error is retrained, rather than retraining the entire model. This segmentation allows selective updates that maintain accuracy while reducing resource consumption.
Solution Approach 2:
The problematic property layer is extracted and identified as the source of the classification error. Only this specific layer is then retrained using the erroneous classification as training data, while the other property layers remain unchanged. This extraction approach isolates the retraining process to only what is necessary.
2Reliability
If the entire OBIA model is retrained when unexpected classifications occur, then model robustness is improved, but system downtime increases
Solution Approach 1:
By segmenting the model into independent property layers, the retraining process is divided into smaller, parallelizable tasks. Multiple property layers can be retrained simultaneously or in sequence with minimal interruption to the overall system, reducing total downtime compared to retraining the entire model as a single unit.
Solution Approach 2:
Instead of performing the complete retraining action on the entire model, only the necessary partial action (retraining specific property layers) is executed. This partial approach achieves sufficient robustness improvement without the excessive time cost of full model retraining.
3Productivity
If high-capability computing devices are used for OBIA model retraining, then retraining speed is improved, but device complexity and cost increase
Solution Approach 1:
The retraining workload is segmented into smaller property layer tasks that can be distributed across multiple computing devices. This allows the use of simpler, lower-capability devices in a distributed architecture, avoiding the need for single high-capability devices while maintaining overall retraining efficiency.
Solution Approach 2:
Multiple simpler computing devices are merged into a distributed computing system that collectively performs the retraining tasks. The combined computational power of multiple devices achieves the necessary retraining speed without requiring any single device to have high complexity or capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A non-transitory computer-readable medium comprising computer-executable instructions that, when executed, are configured to cause a processor to perform operations that include receiving image data after an operation is performed by an industrial automation device on a product; analyzing the image data based an object-based image analysis (OBIA) model to classify the product as one of a plurality of conditions related to manufacturing quality and the OBIA model includes property layers associated with features related to a manufacturing of the product; determining whether the one of the conditions indicates an anomaly being present in the product; sending a notification indicative of the one of the plurality of conditions is presently associated with the product; identifying a property layer associated with classifying the one of the plurality of conditions; and updating the OBIA model based on the property layer and the input indicative of the anomaly being incorrectly associated with the product.