ML Validation Using Disturbance Images for Quality Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning algorithms used for quality inspection in manufacturing are prone to errors due to disturbance variables like vibrations, humidity, and dust, leading to incorrect component rejection or acceptance, which can cause unnecessary costs and safety risks.
Innovation Solution
A method and system for validating machine learning algorithms using labeled validation data generated by a generative adversarial network to assess robustness against disturbance variables, allowing for reliable selection of robust algorithms for quality inspection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If machine learning algorithms are used for quality inspection, then inspection reliability and automation are improved, but sensitivity to disturbance variables (noise, vibrations, humidity, dust) causes error rates to increase
Solution Approach 1:
The patent applies preliminary action by generating validation data with disturbance variables before deploying the machine learning algorithm for actual quality inspection. This pre-validation process prepares the algorithm to handle real-world disturbances by testing it against synthetically corrupted images containing noise, vibrations, humidity effects, and dust artifacts, thereby improving its robustness before automated inspection begins
Solution Approach 2:
The patent converts the harmful effect of disturbance variables into a beneficial validation mechanism. By intentionally introducing disturbance variables (noise, vibrations, humidity, dust) into validation images, the system transforms these previously harmful factors into useful test conditions that reveal algorithm weaknesses and enable targeted improvements, turning a source of errors into a tool for enhancing reliability
2Reliability
If validation data with disturbance variables is generated, then algorithm robustness is improved, but computational resources and validation time increase
Solution Approach 1:
The patent uses copying by creating synthetic validation images that replicate real-world disturbance conditions through digital manipulation rather than requiring physical reproduction of all possible disturbance scenarios. Generative adversarial networks generate copied representations of disturbed images, allowing comprehensive validation without the time cost of capturing and processing actual disturbed samples for every validation case
Solution Approach 2:
The patent applies parameter changes by systematically varying disturbance parameters (noise levels, vibration intensities, humidity effects, dust concentrations) in validation data generation. This allows efficient exploration of robustness across multiple disturbance conditions by adjusting numerical parameters rather than requiring separate validation processes for each condition, reducing overall validation time while maintaining comprehensive coverage
Data Source
AI summary
A method for validating a machine learning algorithm. The machine learning algorithm is trained to recognize objects in image data. The method includes: providing a machine learning algorithm which is trained to recognize objects in image data; generating labeled validation data for validating the machine learning algorithm, wherein the validation data each contain at least one disturbance variable; and validating the machine learning algorithm on the basis of the generated validation data.