Neural Network Anomaly Detection from Aggregate Sensor Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sensor technologies in semiconductor device manufacturing struggle to efficiently monitor and control the quality of processing and product yield due to the complexity of processing chambers and the large volume of data generated by multiple sensors, leading to increased costs and potential for human error in anomaly detection.
Innovation Solution
The implementation of a method that aggregates data from multiple sensors and uses neural networks to process and analyze the data, reducing the dimensionality of sensor statistics, applying outlier detection models, normalizing scores, and generating anomaly scores to indicate manufacturing anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple sensors are used to monitor processing chamber environment and wafer transportation, then measurement precision and reliability of monitoring is improved, but device complexity and data processing burden increase
Solution Approach 1:
The patent segments the complex monitoring system into distinct functional modules: sensor data acquisition module, statistical parameter calculation module, anomaly detection model module, and alert generation module. Each module handles specific tasks independently, reducing overall system complexity while maintaining comprehensive monitoring capabilities across multiple sensors.
Solution Approach 2:
The patent introduces statistical parameters (mean, standard deviation, skewness, kurtosis) as intermediary representations that aggregate raw sensor data. These statistical intermediaries simplify the complex multi-sensor data into manageable metrics that feed into anomaly detection models, reducing data processing burden while preserving measurement precision.
2Reliability
If traditional statistical methods are used for anomaly detection, then ease of implementation is maintained, but detection accuracy and reliability deteriorate due to inability to capture complex patterns
Solution Approach 1:
The patent transforms the anomaly detection approach by changing from traditional single-threshold statistical methods to a multi-parameter evaluation system. It calculates multiple statistical parameters (mean, standard deviation, skewness, kurtosis) and applies them through trained anomaly detection models, enabling detection of complex patterns while maintaining implementation feasibility through systematic parameter transformation.
Solution Approach 2:
The patent implements preliminary training of anomaly detection models using historical sensor data before actual monitoring deployment. This preliminary action prepares the models with learned patterns of normal and anomalous behavior, significantly improving detection reliability while the models themselves remain relatively simple to execute during operational monitoring.
3Manufacturing precision
If comprehensive sensor monitoring is implemented across all manufacturing operations, then product quality control is improved, but loss of time for data processing and analysis increases
Solution Approach 1:
The patent extracts only the essential statistical parameters (mean, standard deviation, skewness, kurtosis) from comprehensive sensor data, discarding redundant raw data points. This extraction approach maintains product quality monitoring by capturing critical process variations while dramatically reducing data processing time and computational resources required.
Solution Approach 2:
The patent applies anomaly detection thresholds selectively to different statistical parameters based on their relevance to specific manufacturing operations. Rather than uniformly processing all parameters with equal depth, it applies partial analysis to less critical parameters and more rigorous analysis to critical ones, optimizing the balance between quality control and processing time.
Data Source
AI summary
Implementations disclosed describe systems and techniques to detect anomalies in a manufacturing operation. The techniques include generating, using a plurality of outlier detection models, a plurality of outlier scores. The outlier scores are representative of a degree of presence, in a plurality of sensor statistics, of an anomaly associated with the manufacturing operation. Individual outlier scores are generated using a respective one of the plurality of outlier detection models. The techniques further include determining, using the outlier scores, a likelihood of the anomaly associated with the manufacturing operation.