Deep Auto-Encoder Fault Detection for Semiconductor Tool Sensors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for Equipment Health Monitoring (EHM) and Fault Detection in semiconductor processing chambers are inefficient in detecting short-time signal perturbations and suffer from high false positive rates, requiring significant feature engineering and being restricted to specific tool types or recipe types, making them time-consuming and inaccurate for anomaly detection.
Innovation Solution
A deep auto-encoder type neural network is employed to derive a model of training time-series traces, which minimizes reconstruction error, allowing for the detection of anomalies in sensor data by calculating mean square error between input and output time-series traces, with the server declaring anomalies when the error exceeds a pre-determined value.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If classical machine learning methods with feature engineering are used, then model interpretability is improved, but detection accuracy and automation capability deteriorate
Solution Approach 1:
The patent replaces manual feature engineering (mechanical process) with deep learning automatic feature extraction (intelligent system). The neural network automatically learns relevant features from raw sensor data, eliminating the need for domain expert intervention while improving detection accuracy through sophisticated pattern recognition capabilities.
Solution Approach 2:
The deep learning model performs self-service by automatically extracting features and detecting anomalies without human intervention. The system trains on historical data and independently identifies outliers in real-time sensor streams, reducing operational complexity while maintaining high detection accuracy.
2Device complexity
If statistical methods are used for anomaly detection, then computational simplicity is improved, but detection capability for short-time perturbations deteriorates
Solution Approach 1:
The patent replaces simple statistical thresholding (mechanical method) with deep learning-based anomaly detection (intelligent system). The neural network captures complex temporal patterns and short-time perturbations that statistical methods miss, achieving superior detection capability while the trained model provides efficient inference.
Solution Approach 2:
The deep learning model transforms the detection problem from simple statistical comparison to multi-dimensional feature space analysis. By learning representations in higher-dimensional space, the model can detect subtle short-time perturbations that are invisible to traditional statistical methods operating in raw sensor value space.
3Reliability
If traditional machine learning models are rebuilt for each recipe or chamber change, then model specificity is improved, but productivity deteriorates
Solution Approach 1:
The patent creates a universal deep learning model that can handle multiple recipes and chambers simultaneously. The model learns general anomaly patterns across different processing conditions and adapts to specific scenarios through the input data itself, eliminating the need to rebuild models for each tool type or recipe type while maintaining high detection accuracy.
Solution Approach 2:
The deep learning model provides dynamic adaptability to different recipes and chambers without requiring retraining. The system adjusts to new processing conditions by learning from incoming data streams, enabling the model to maintain specificity for different tool types while avoiding the time-consuming model rebuilding process.
4Measurement precision
If deep auto-encoder neural networks are used, then detection accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary action by training the deep auto-encoder model offline on historical sensor data. This pre-training phase captures normal operational patterns and anomaly characteristics, enabling the model to make rapid real-time predictions during actual manufacturing operations. The computationally intensive work is done in advance, leaving only efficient inference for real-time detection.
Solution Approach 2:
The deep auto-encoder extracts and isolates the essential features of normal and anomalous sensor patterns during training. By separating the complex learning process from real-time operation, the system achieves high detection accuracy while maintaining efficient runtime performance through feature extraction and dimensionality reduction.
Data Source
AI summary
Implementations described herein generally relate to a method for detecting anomalies in time-series traces received from sensors of manufacturing tools. A server feeds a set of training time-series traces to a neural network configured to derive a model of the training time-series traces that minimizes reconstruction error of the training time-series traces. The server extracts a set of input time-series traces from one or more sensors associated with one or more manufacturing tools configured to produce a silicon substrate. The server feeds the set of input time-series traces to the trained neural network to produce a set of output time series traces reconstructed based on the model. The server calculates a mean square error between a first input time series trace of the set of input time series traces and a corresponding first output time series trace of the set of output time-series traces. The server declares the sensor corresponding to the first input time-series trace as having an anomaly when the mean square error exceeds a pre-determined value.