Deep Auto-Encoder Monitoring for Semiconductor Tool Anomaly Detection

The patent transforms the anomaly detection problem from statistical hypothesis testing to neural network reconstruction error analysis. By changing the detection parameter from statistical significance thresholds to reconstruction error magnitudes, the system achieves both high precision in detecting true anomalies and low false positive rates, as the neural network learns the normal operational patterns and only flags deviations that exceed learned thresholds.

The patent replaces traditional statistical mechanical methods (hypothesis testing, p-values, confidence intervals) with a neural network-based information processing system. This substitution allows the system to capture complex non-linear relationships in sensor data while providing more reliable anomaly detection with reduced false positives, as the neural network implicitly models the joint distribution of sensor readings without requiring explicit statistical assumptions.

The patent implements self-service by allowing the neural network to automatically learn relevant features from raw sensor data during training. The autoencoder architecture inherently performs feature extraction and selection by learning to compress and reconstruct the input data, eliminating the need for manual feature engineering by domain experts. This self-organizing approach significantly reduces time consumption while maintaining or improving detection accuracy.

The patent performs preliminary action by pre-training the neural network on extensive normal operational data before deployment. This preliminary training phase allows the system to learn the baseline patterns and relationships in sensor data, so that during actual operation, anomaly detection can proceed rapidly without requiring real-time feature engineering. The model is prepared in advance to handle various operational conditions.

The patent achieves universality by designing a neural network architecture that can handle multiple tool types and recipes within a single model. The autoencoder learns a unified representation of sensor data patterns across different processing conditions, allowing the same model to detect anomalies in various chamber types and recipe configurations without requiring separate models. This multi-functional approach maintains detection accuracy across diverse scenarios while eliminating the need for time-consuming model rebuilding.

The patent introduces dynamics by enabling the neural network to adapt to different operational contexts through continuous learning or fine-tuning mechanisms. Rather than requiring complete model rebuilding when tool types or recipes change, the system can dynamically adjust to new conditions by learning from additional data or adapting existing representations, thereby maintaining high detection accuracy while significantly improving productivity and reducing downtime.