Autoencoder Code Layer Compression for Unknown Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autoencoder-based abnormality determination methods struggle to accurately identify unknown abnormal images due to the diversity of abnormality patterns in manufacturing, as they are trained only on no-defects images, leading to ineffective estimation and determination.
Innovation Solution
The proposed solution involves an abnormality determination computer that maximizes the dimensional compression rate of the autoencoder's code layer and reduces its expressive power, allowing for accurate determination of unknown abnormal images by comparing inspection images with estimated no-defects images, using a comparative inspection engine that evaluates differences and adjusts processing parameters for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If the autoencoder is trained only on no-defects images to perform abnormality determination, then the inspection cost and dependency on inspector skills are reduced, but the accuracy of determining unknown abnormality patterns deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-training the autoencoder on a large dataset of no-defects images before deployment. This pre-training enables the model to learn normal patterns comprehensively, allowing it to effectively identify deviations (abnormalities) when inspecting new images, even though it hasn't been explicitly trained on abnormal examples.
Solution Approach 2:
The patent introduces a comparative inspection engine as an intermediary component that bridges the gap between the autoencoder's estimation and the actual inspection images. This engine compares the original inspection image with the autoencoder's reconstructed no-defects image, highlighting differences that indicate abnormalities, thereby improving determination accuracy without requiring direct training on abnormal data.
2Ease of manufacture
If the autoencoder learns to estimate the same no-defects image from no-defects images, then the learning process is simplified, but the ability to estimate accurate no-defects images from unknown abnormal images deteriorates
Solution Approach 1:
The patent implements feedback by using the comparative inspection engine to evaluate the difference between inspection images and reconstructed images. This feedback mechanism allows the system to identify when the autoencoder fails to accurately reconstruct abnormal images, enabling continuous improvement of the model's ability to handle unknown abnormality patterns while maintaining simple no-defects-only training.
Solution Approach 2:
The patent applies parameter changes by adjusting the autoencoder's architecture parameters (such as the number of layers, neurons per layer, and activation functions) to optimize its reconstruction capability. These parameter adjustments enhance the model's ability to accurately estimate no-defects images from various inputs, including unknown abnormal images, without complicating the training process.
3Measurement precision
If the dimensional compression rate of the code layer is increased, then the abnormality determination correct answer rate improves, but the information loss in the compressed representation increases
Solution Approach 1:
The patent applies partial action by using a moderate dimensional compression rate that retains sufficient information for accurate abnormality determination. Rather than maximizing compression, the system uses a compression level that is just enough to capture the essential features needed for detecting abnormalities, balancing information retention with determination accuracy.
Solution Approach 2:
The patent addresses information loss by transitioning to another dimension through the comparative inspection approach. Instead of relying solely on the compressed code layer to preserve all information, the system compares the original high-dimensional inspection images with the reconstructed images, using the difference in this alternative dimension (image space vs. code space) to detect abnormalities effectively.
Data Source
AI summary
In anomaly determination using an autoencoder, the present invention enables highly accurate anomaly determination with respect to an unknown anomaly image, even in the case of learning using only a non-defective image. In a learning step of learning a parameter of the autoencoder using a non-defective learning image /{f_i/} captured of a non-defective test target, an anomaly determination accuracy rate Rc of the non-defective learning image /{f_i/} is used as an evaluation value, and a dimension compression rate Rd of a code layer of the autoencoder is maximized or an input signal amount to the code layer or an output signal amount from the code layer is decreased, such that the anomaly determination accuracy rate Rc is maximized or becomes at least a predetermined threshold th.