TLDR: This paper introduces two unsupervised deep learning methods, N-BEATS and a Graph Neural Network (GNN), for predicting anomalies in multi-variate time series data from semiconductor manufacturing. The GNN consistently outperforms N-BEATS in forecasting accuracy and computational efficiency, even when largely operating in a univariate manner, making it a promising solution for real-time fault prevention in complex production environments.
Semiconductor manufacturing is an incredibly intricate and precise process, involving thousands of interconnected parameters across various tools, production steps, and timeframes. Monitoring these systems in real-time, detecting faults, and predicting maintenance needs are crucial. However, predicting anomalies in this environment faces significant hurdles: vast amounts of sensor data, the rarity of actual faults, noisy and incomplete measurements, and the constantly changing nature of production systems due to adjustments and equipment aging.
A new research paper introduces two innovative approaches to move beyond just detecting anomalies to actively predicting them. This shift is vital for enabling real-time process corrections and preventing faults before they occur. The proposed framework operates in two main stages: first, a forecasting model is trained on a dataset assumed to be free of anomalies. Second, this trained model then forecasts future behavior on new, unseen time series data. Any deviations beyond a set threshold are flagged as potential anomalies.
Two Approaches to Anomaly Prediction
The paper explores two distinct forecasting models. The first approach, called N-BEATS, treats each signal in the multi-variate time series independently. It uses the N-BEATS model, a deep learning architecture known for univariate time series forecasting. This means a separate model is essentially trained for each sensor or variable.
The second, more advanced approach, utilizes a Graph Neural Network (GNN). This model is designed to capture the complex interdependencies between different process variables, moving beyond the independence assumption of N-BEATS. The GNN architecture, an extension of previous work, focuses on creating a graph structure where variables are nodes, and connections (edges) are learned based on their relationships, enhancing the model’s ability to understand system dynamics.
Key Contributions of the Research
The researchers highlight several key contributions:
- A clear anomaly prediction procedure involving training on normal data and then forecasting new data to identify deviations.
- The implementation of two distinct forecasting models: N-BEATS for independent variable forecasting and a GNN for capturing inter-variable correlations.
- An unsupervised framework that requires minimal preprocessing, making it highly generalizable and efficient.
- A comprehensive performance study demonstrating strong forecasting and anomaly detection capabilities for both models, with the GNN consistently outperforming N-BEATS.
- An ablation study on the GNN’s graph construction, revealing surprising insights into its optimal configuration.
Data and Methodology
The study used real-world data from a Coat/Develop Track tool at the imec fabrication facility, comprising multi-variate time series from 912 process runs across 14 recipes. Since real anomaly labels were unavailable, synthetic anomalies were carefully induced into the test datasets. These included ‘amplitude shift’ (changing signal intensity), ‘time shift’ (lagging or early onset), and ‘step shift’ (altering step duration) anomalies, specifically targeting signals with step-like behavior to ensure realism and diversity.
Minimal preprocessing was applied to keep the models lightweight and adaptable, primarily involving min-max normalization to scale data between 0 and 1. The data was split into training, validation, and test sets, with three consecutive process runs concatenated to increase dataset size.
Model Performance and Efficiency
Both N-BEATS and the GNN demonstrated strong forecasting performance, maintaining an F1-score above 75% for anomaly detection even with prediction horizons up to 100 time points. As expected, performance generally declined with longer prediction horizons due to increased uncertainty.
A crucial finding was the GNN’s superior performance, especially for longer horizons, where it achieved significantly lower test loss values compared to N-BEATS. This indicates the GNN’s robustness in long-term anomaly detection. Interestingly, an ablation study on the GNN’s ‘top-K’ parameter (which limits the number of neighbors a node can have in the graph) showed that the best performance was achieved when top-K was set to 1. This implies that each node primarily connected to itself, suggesting that either the GNN wasn’t effectively leveraging cross-variable dependencies in this specific dataset, or such dependencies were minimal.
Despite this, the GNN proved to be remarkably more computationally efficient. It required 2.5 to 5.3 times fewer trainable parameters than N-BEATS (which trains a separate model for each variable) and completed training in minutes compared to hours for N-BEATS. Both models, however, were capable of real-time inference, making them suitable for deployment in manufacturing environments.
Also Read:
- Unifying Anomaly Detection: How ShortcutBreaker Tackles Complex Data
- Time-Conditioned Contraction Matching: A New Era for Scalable and Explainable Anomaly Detection
Future Directions
The researchers acknowledge several areas for future work, including evaluating the models on a wider variety of anomaly types and real-world faults, developing models that can adapt to varying numbers of input sensors across different tools, integrating causal inference to better understand variable relationships, and predicting the downstream impact of anomalies on device fabrication.
In conclusion, this study successfully extended and applied two forecasting models for online anomaly prediction in semiconductor manufacturing. The Graph Neural Network emerged as a highly promising solution due to its superior performance and computational efficiency, even when operating with minimal explicit cross-variable connections. This work represents a significant step towards proactive fault prevention and enhanced process control in the complex world of chip fabrication. You can read the full research paper here.
