Manufacturing Anomaly Detection Using Expert-Adjusted ML Probabilities
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In manufacturing processes, especially in large-scale equipment like semiconductor manufacturing, engineers face challenges in real-time anomaly detection due to the complexity of managing multiple sensors and varying expertise levels, leading to inefficiencies and inconsistent anomaly identification.
Innovation Solution
A processor-implemented method that converts manufacturing process data into an image test sample, using a machine learning model to generate a probability value for anomalies, which is adjusted based on expert classification criteria by mapping to an embedded space and adjusting probabilities based on distances to hard samples, allowing for automated anomaly identification without human input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If engineers manually monitor and analyze manufacturing process data, then they can identify anomalies using their expertise, but the efficiency decreases due to the large number of sensors and varying abilities between engineers
Solution Approach 1:
The system enables automated anomaly detection by training a machine learning model to independently identify anomalies in manufacturing process data without requiring continuous human intervention. The model learns from historical data and automatically classifies new data points, making the system self-sufficient in the anomaly detection task while engineers only need to review final results.
Solution Approach 2:
The patent replaces the manual mechanical process of engineer-based anomaly detection with an automated machine learning system. The machine learning model substitutes human engineers' analytical work, processing sensor data and identifying anomalies algorithmically, thereby eliminating the limitations of human capacity and consistency while maintaining detection accuracy.
2Productivity
If a machine learning model is used to detect anomalies automatically, then monitoring efficiency improves, but the accuracy may decrease due to lack of expert judgment
Solution Approach 1:
The patent introduces an intermediary adjustment mechanism that bridges the machine learning model and expert judgment. The system calculates probability values from the machine learning model, then adjusts these probabilities based on characteristics of hard samples (difficult-to-classify cases). This intermediary step incorporates expert knowledge into the automated system, ensuring that final anomaly detections align with expert criteria while maintaining automated efficiency.
Solution Approach 2:
The system implements feedback by using hard samples (cases that are difficult to classify) to adjust and refine the machine learning model's probability outputs. The model learns from these challenging cases and continuously improves its accuracy, with the feedback loop ensuring that automated predictions remain aligned with expert judgment standards while maintaining high monitoring efficiency.
3Measurement precision
If the system adjusts probability values based on expert classification criteria, then anomaly detection accuracy improves, but the system complexity increases
Solution Approach 1:
The patent adjusts the probability parameter output by the machine learning model based on characteristics of hard samples. Instead of changing the overall system architecture, the invention modifies the probability values through mathematical adjustments that reflect expert classification criteria. This parameter-based approach maintains system simplicity while improving accuracy, avoiding the need for complex additional components.
4Reliability
If engineers manually analyze all sensor data, then comprehensive anomaly detection is possible, but the time required increases significantly
Solution Approach 1:
The patent creates a virtual copy of expert judgment capability through the machine learning model. The model is trained on historical data and learns to replicate expert anomaly detection patterns. This copying approach allows the system to process data at machine speed while maintaining the reliability and completeness of expert-level analysis, eliminating the time bottleneck without sacrificing detection quality.
Data Source
AI summary
A processor-implemented method of an apparatus or system includes obtaining an expert classification criterion from a memory of the apparatus or system; converting manufacturing process data associated with a manufacturing process to a test sample in a form of an image; generating, using a machine learning model provided the test sample, a probability value that the test sample corresponds to a target class representing an anomaly occurring in the manufacturing process; adjusting the probability value by reflecting the expert classification criterion for the anomaly; and identifying, by classifying the anomaly based on the adjusted probability value, whether a final abnormality in the manufacturing process has occurred.