Outlier Detection in Substrate Processing Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current outlier detection methods in semiconductor manufacturing are not sensitive enough to detect anomalies in refined sensor data and lack scalability to handle large batches of substrates, necessitating an improved approach for classifying outliers in time series data from substrate processing chambers.
Innovation Solution
A method involving a client device that receives time series data, converts it into a bounded uniform signal, identifies sub-segments not matching expected behavior, and classifies them using a combination of Grubbs' rule for outlier removal, k-nearest neighbor computations, logit functions for probability estimation, and Radial Basis Function networks for categorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional outlier detection methods are used, then the system is simple to implement, but the sensitivity to detect anomalies in refined sensor data is insufficient
Solution Approach 1:
The patent segments the outlier detection process into multiple specialized stages: data normalization, sub-segment identification using Grubbs' rule, pattern matching against expected behavior templates, and classification. Each stage handles a specific aspect of the detection problem, improving sensitivity while keeping individual components manageable in complexity.
Solution Approach 2:
The patent transforms sensor data into a standardized format with bounded uniform signals, converting raw sensor readings into a normalized dimensional space. This dimensional transformation enables more sensitive anomaly detection by comparing data points against standardized expected behavior patterns rather than raw sensor values.
2Productivity
If traditional outlier detection methods are used, then the computational requirements are low, but the scalability to handle large batches of substrates is limited
Solution Approach 1:
The patent performs preliminary actions by pre-defining expected behavior patterns and normalization parameters before processing substrate batches. By establishing these reference templates in advance, the system can rapidly compare incoming sensor data against known good patterns without performing complex computations during actual batch processing, improving scalability.
Solution Approach 2:
The patent changes parameters by normalizing sensor data into bounded uniform signals with standardized ranges. This parameter transformation allows the system to process large batches efficiently by comparing normalized values against fixed thresholds and patterns, reducing computational complexity while maintaining detection accuracy across varying batch sizes.
3Measurement precision
If the detection range is refined to detect more subtle anomalies, then the anomaly detection capability improves, but the computational complexity increases
Solution Approach 1:
The patent segments the detection range into specific sub-segments of sensor data, analyzing each segment independently against expected behavior patterns. This segmentation allows refined detection of subtle anomalies in specific time windows or sensor ranges without requiring complex algorithms to analyze the entire dataset simultaneously, managing computational complexity while improving detection precision.
Solution Approach 2:
The patent transforms sensor data into bounded uniform signals, changing the dimensional representation to a standardized format. This dimensional change enables the system to detect subtle anomalies across a refined range by comparing normalized values against expected patterns, achieving high detection precision with manageable algorithmic complexity through standardized data representation.
Data Source
AI summary
Embodiments disclosed herein generally relate to a method, system, and non-transitory computer readable medium for classifying an outlier in time series data collected by a sensor positioned in a substrate processing chamber. The client device receives time series data from the sensor positioned in the substrate processing chamber. The client device converts the time series data to a bounded uniform signal. The client device identifies signal sub-segments that do not match an expected behavior. The client device classifies the identified sub-segments that do not match the expected behavior.