Deep Anomaly Detector for Manufacturing Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial manufacturing lines face challenges in real-time anomaly detection, particularly in scenarios where product architecture and features change frequently, leading to concept drift and the need for large datasets that are often unavailable at the start of production.
Innovation Solution
A method using a Deep Anomaly Detector (DAD) system that employs meta-learning, few-shot learning, and automatic patch extraction to detect anomalies in manufacturing products, even without a robust base of anomaly images. This system generates synthetic anomalous features and adapts to new products by comparing input images with pairs of reference images, allowing for efficient adaptation to changing inspection scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning vision systems are used for anomaly detection, then detection accuracy is improved, but data acquisition cost and complexity increase significantly
Solution Approach 1:
The patent creates synthetic anomaly copies by generating pseudo-anomalous samples from normal product images through parameter perturbation and transformation. This copying approach eliminates the need for expensive real anomaly data acquisition while maintaining detection accuracy, as the synthetic anomalies preserve the visual characteristics needed for training the deep learning model.
Solution Approach 2:
The patent applies parameter changes by systematically modifying image parameters (color, brightness, geometric transformations) to generate diverse synthetic anomaly samples from a single normal image. This parameter variation technique creates a robust training dataset without requiring additional physical samples or complex data acquisition equipment.
2Reliability
If traditional deep learning models are used for anomaly detection, then detection robustness is improved, but adaptability to new products deteriorates due to concept drift
Solution Approach 1:
The patent implements dynamics by creating a dynamic training process where synthetic anomaly parameters are continuously adjusted and diversified. The system adapts to new product versions by regenerating synthetic anomalies based on the new product's normal images, maintaining robustness while achieving adaptability through parameter dynamic adjustment rather than fixed training data.
Solution Approach 2:
The patent applies preliminary action by pre-generating diverse synthetic anomaly samples across multiple parameter spaces before actual detection begins. This preliminary synthetic data generation prepares the model for various anomaly types and product variations, enabling faster adaptation to new products without requiring extensive retraining on real anomaly data.
3Measurement precision
If human inspectors are used for quality control, then detection accuracy is maintained, but inspection speed and productivity decrease significantly
Solution Approach 1:
The patent replaces the mechanical human inspection system with an automated deep learning vision system trained on synthetic anomalies. This substitution maintains high detection accuracy comparable to human inspectors while achieving inspection speeds thousands of times faster, as the automated system can process images at machine speed without fatigue or attention limits.
4Adaptability or versatility
If large datasets are collected for training, then model generalization is improved, but data collection time and loss of time increase
Solution Approach 1:
The patent uses copying to generate large volumes of training data by creating multiple synthetic anomaly variants from a small set of normal product images. This copying approach achieves the dataset size needed for good model generalization without the time-consuming process of collecting, annotating, and curating real anomaly data from production lines.
Solution Approach 2:
The patent applies preliminary action by pre-computing and storing diverse synthetic anomaly samples across different parameter configurations before deployment. This preliminary generation of comprehensive training data ensures the model learns robust anomaly patterns, reducing the need for additional data collection and model retraining when deploying to new product versions.
Data Source
AI summary
The present invention refers to a method of training a neural network to detect anomalies in a manufacturing product comprising: obtaining a dataset including multiple manufacturing product images and annotation files with coordinates of predetermined anomalous regions corresponding to the manufacturing product images, respectively; selecting a test query set of images and a support set of images from the dataset multiple to train a deep learning model, the support set of images comprises pairs of reference images and each pair comprises at least one anomalous manufacturing product image and at least one non-anomalous manufacturing product image; inputting the test query set into a deep learning model to rate a similarity score based on a similarity distance between the pairs of reference images of the support set based on the annotation files; and adjusting parameters characterizing the deep learning model through a model based on the similarity score.