ML Quality Prediction Thresholds for Weakly Correlated Inspections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In production lines, it is challenging to accurately predict defective products in process inspections when inspection data is not strongly correlated with final inspection results, leading to inefficiencies and increased production costs.
Innovation Solution
A prediction score calculation device using a machine learning model to calculate a prediction score from first inspection data, which includes determining a threshold value to maximize cost merit by considering profit and loss, thereby enabling accurate quality determination in downstream inspections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional correlation-based inspection item selection is used, then inspection criteria can be reset for strongly correlated items, but it becomes difficult to predict quality when inspection data is not strongly correlated with final inspection results
Solution Approach 1:
The patent transforms the inspection prediction approach by changing from selecting individual inspection items based on correlation to using machine learning models that process combinations of inspection data. This parameter change enables the system to handle weakly correlated data by considering complex interactions among multiple inspection items simultaneously, thereby improving prediction accuracy for cases where individual item correlation is low.
Solution Approach 2:
The patent applies the composite principle by combining multiple inspection data items into a composite prediction model. Instead of relying on single inspection items, the system integrates results from multiple inspection processes into a unified machine learning model that can predict final inspection outcomes even when individual items show weak correlation, effectively creating a composite indicator of product quality.
2Measurement precision
If machine learning models process combinations of inspection data, then prediction accuracy improves for weakly correlated items, but the complexity of determining how items are combined increases
Solution Approach 1:
The machine learning model performs self-service by automatically learning and determining the optimal combinations of inspection data items through training on historical data. The model autonomously identifies which inspection items and combinations are most predictive of final inspection results, eliminating the need for manual analysis of item combinations and reducing the perceived complexity for users while maintaining high prediction accuracy.
Solution Approach 2:
The system incorporates feedback mechanisms where the machine learning model continuously learns from the outcomes of predictions and actual final inspection results. This feedback loop allows the model to refine its understanding of how inspection items combine to predict quality, automatically adjusting its internal parameters and combinations without increasing operational complexity for end users.
3Productivity
If process inspection excludes defective products early, then production efficiency improves and losses are reduced, but accurate prediction requires strong correlation between inspection data and final results
Solution Approach 1:
The patent enables early defective product exclusion by changing the prediction approach from correlation-based single item selection to machine learning-based combination analysis. This parameter change allows the system to achieve high prediction accuracy even with weakly correlated inspection data, making it feasible to perform accurate quality prediction at upstream inspection stages and thereby improve production efficiency by excluding defective products earlier in the manufacturing process.
Data Source
AI summary
When inspection data of a process inspection in a production line is input, a machine learning unit 420 of a prediction score calculation device 202 performs machine learning so as to output a prediction score of quality determination of a final inspection. In addition, a prediction score calculation unit 410 outputs a prediction score predicting the quality determination result of the final inspection from the inspection data of the process inspection using a machine learning model that has performed the machine learning. In addition, a threshold value determination unit 440 compares the prediction score calculated by the prediction score calculation unit 410 and determines a threshold value for predicting the quality determination from learning data, the prediction score, and cost data.