Battery Image Inspection Using Dual AI for Novel Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In battery manufacturing and assembly processes, appearance defects such as deformation, crushing, or damage to the sealing part of the battery casing can occur, leading to unreliable and potentially unsafe batteries. Current defect detection methods struggle to identify new types of defects, as they rely on pre-learned data and may misclassify novel defects as good products.
Innovation Solution
A battery defect detection apparatus utilizing two artificial intelligence models is employed. The first AI model classifies the battery product based on pre-learned data, while the second AI model determines whether the product corresponds to normal data, re-classifying it as defective if it does not match. This system allows for the detection of new defect types and learns from new defective images to improve precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single AI model is used for defect detection based on pre-learned data, then the detection speed is fast and the system is simple, but new defect types cannot be detected and are misclassified as normal products
Solution Approach 1:
The defect detection system is segmented into two independent AI models: a first AI model for initial classification of products as normal or defective based on pre-learned data, and a second AI model for re-classification of borderline cases. This segmentation allows each model to specialize in specific detection tasks, improving overall reliability while maintaining manageable complexity through modular architecture.
Solution Approach 2:
A determination module acts as an intermediary between the first and second AI models. It receives classification results from the first model, determines whether re-classification is necessary by comparing against learning data, and selectively inputs images to the second model. This intermediary prevents unnecessary processing while ensuring new defect types are caught, balancing reliability and efficiency.
2Measurement precision
If more learning data is collected to improve detection of new defect types, then the detection precision improves, but the time and resources required for data collection and model training increase
Solution Approach 1:
The system performs preliminary classification using the first AI model trained on existing learning data before invoking the more resource-intensive second model. This preliminary action filters out clearly normal or defective products, reserving extensive data processing and model training resources only for borderline cases that may represent new defect types, thereby reducing overall time and resource consumption.
Solution Approach 2:
Instead of processing all products through the comprehensive second AI model, the system applies partial action by using the first model for routine classifications and only invoking the second model with excessive computational resources when the determination module identifies uncertain cases. This selective approach achieves high precision for critical cases while minimizing time loss through efficient resource allocation.
Data Source
AI summary
A battery defect detection apparatus according to an embodiment disclosed herein includes: a communication module; a processor; and a memory configured to store a first artificial intelligence model, a second artificial intelligence model, and instructions, in which the processor is configured to execute the instructions by the battery defect detection apparatus; to perform operations including: obtaining an image of a subject product by using the communication module, inputting the image of the subject product to the first artificial intelligence model to classify the subject product, and inputting the image of the subject product to the second artificial intelligence model to determine whether the image of the subject product corresponds to first data for classifying the subject product as normal among learning data of the first artificial intelligence model, when the subject product is classified as normal.