Anomaly Detection for Semiconductor Wafer Multi-wire Sawing Machines Using Statistical and Deep Learning Methods
At a Glance
Section titled âAt a Glanceâ| Metadata | Details |
|---|---|
| Publication Date | 2025-10-27 |
| Journal | Mobile Networks and Applications |
| Authors | Zhen-Yin Annie Chen, ChunâCheng Lin, Hsin-Cheng Huang, WeiâJuin Su, Chun-Yi Cheng |
| Analysis | Full AI Review Included |
Executive Summary
Section titled âExecutive SummaryâThis analysis outlines the development and validation of two anomaly detection models for Diamond Multi-wire Sawing Machines (DMSMs) used in slicing hard, brittle third-generation semiconductors (SiC, GaN).
- Core Challenge: Preventing catastrophic failures, such as wire breakage, in DMSMs, which leads to significant production losses and wafer damage.
- Proposed Solution: Implementation of a Prognostics and Health Management (PHM) system using two tailored anomaly detection models: a Rule-based Statistical Model and a data-driven Univariate Autoencoder (UAE) Model.
- Data Source: Models were trained and validated using confidential, unlabeled industrial sensor data collected at 5-second intervals from actual production DMSMs.
- Superior Performance: The UAE model demonstrated exceptional robustness, achieving a True Positive Rate (TPR) up to 0.9935 while maintaining zero observed False Positives (FPR = 0) across all experiments.
- Rule-Based Utility: The statistical model offers high interpretability and low computational cost, making it suitable for scenarios with limited historical data or where expert-defined thresholds are preferred.
- Diagnostic Capability: The UAE model converts reconstruction errors into anomaly scores, allowing engineers to trace anomalies back to specific sensor features for root cause diagnosis.
- Operational Impact: The framework provides actionable value for factory deployment, enabling early fault warnings and supporting predictive maintenance strategies in semiconductor wafer slicing.
Technical Specifications
Section titled âTechnical Specificationsâ| Parameter | Value | Unit | Context |
|---|---|---|---|
| Target Equipment | Diamond Multi-wire Sawing Machine (DMSM) | N/A | Essential for slicing SiC and GaN ingots. |
| Sensor Sampling Interval | 5 | seconds | Uniform recording interval via industrial DAQ system. |
| Feature Count | Hundreds | N/A | Synchronized sensor readings per record. |
| Data Volume (Normal Sample) | Tens of thousands | records | Per independent cutting run. |
| Data Normalization Range | [0, 1] | N/A | Min-Max scaling applied to all features. |
| Best TPR (UAE Model) | 0.9935 | N/A | Achieved on Test Data 2-8 (long wire breakage anomaly). |
| Best FPR (UAE Model) | 0 | N/A | Zero false positives observed across all UAE tests. |
| Rule-Based Z-score Threshold | ±3 | Standard Deviations | Default threshold based on three-sigma rule. |
| UAE Anomaly Score Method | Gauss-S | N/A | Selected method for reconstruction error transformation. |
| UAE Final Static Threshold | 10-6 | N/A | Threshold applied to the overall anomaly score. |
Key Methodologies
Section titled âKey Methodologiesâ- Data Preprocessing: Raw, unlabeled sensor data was cleaned by removing non-numeric and zero-variance features (filter-based selection). Remaining features were normalized using Min-Max scaling to ensure equal contribution during model training.
- Rule-Based Model Construction: A sliding window technique was applied to historical normal data to extract statistical metrics (mean, standard deviation, min, max) at each time point.
- Dynamic Threshold Establishment: Two dynamic threshold strategies were implemented: Min-Max (based on historical range) and Z-score (based on historical mean and standard deviation, typically ±3).
- Rule-Based Anomaly Flagging: An anomaly alert is triggered if any selected feature continuously exceeds its respective dynamic threshold for a user-defined duration (e.g., 150 consecutive time steps).
- Univariate Autoencoder (UAE) Training: A separate autoencoder was trained for each individual sensor feature using historical normal data in an unsupervised manner, learning to minimize reconstruction error.
- Anomaly Score Transformation: During inference, the reconstruction error from each UAE was transformed into a standardized anomaly score using the Gauss-S method, which utilizes a global Gaussian distribution of errors.
- Overall Anomaly Detection (UAE): The feature-wise anomaly scores were summed to create an overall anomaly score. If this score exceeded a predefined static threshold (10-6), the observation was flagged as anomalous.
- Diagnostic Insight: The UAE framework supports diagnosis by ranking the feature-wise anomaly scores, identifying which specific sensors contributed most significantly to the detected anomaly.
Commercial Applications
Section titled âCommercial ApplicationsâThe developed anomaly detection framework is highly relevant for high-precision manufacturing and industrial monitoring, particularly in sectors dealing with advanced materials:
- Third-Generation Semiconductor Manufacturing: Direct application in slicing hard and brittle materials like SiC and GaN, which are critical for 5G and high-power electronics.
- Predictive Maintenance (PdM) Systems: Provides real-time, data-driven fault prediction for complex mechanical systems, moving maintenance from reactive/preventive to predictive.
- High-Value Component Protection: Essential for machinery where failure (like wire breakage) results in the loss of expensive materials (SiC/GaN ingots) and significant downtime.
- Industrial Internet of Things (IIoT) Deployment: The lightweight, interpretable rule-based model and the robust, accurate UAE model offer flexible options for integrating PHM into existing IIoT monitoring platforms.
- Unsupervised Time Series Anomaly Detection: The UAE methodology is broadly applicable to any industrial setting generating large volumes of unlabeled time-series sensor data, such as monitoring turbines, pumps, or electric motors.
View Original Abstract
Abstract Diamond multi-wire sawing machines are essential in semiconductor manufacturing, especially for slicing hard and brittle third-generation materials such as silicon carbide (SiC) and gallium nitride (GaN). The increased difficulty in processing these materials has highlighted the urgent need for reliable machine health monitoring and anomaly detection systems. While Predictive Maintenance and Prognostics and Health Management (PHM) frameworks have been widely applied across various industries, little research has specifically addressed semiconductor cutting equipment, where operational dynamics and data confidentiality present unique challenges. This study, in collaboration with an industry partner, develops two anomaly detection models tailored for diamond multi-wire sawing machines. The first model is a rule-based approach that utilizes sliding window techniques to extract statistical features and establish dynamic thresholds for anomaly detection. The second model employs a data-driven Univariate Autoencoder (UAE) to perform unsupervised anomaly detection by learning reconstruction errors from normal operating data. Both models are trained and validated using confidential industrial sensor datasets. Experimental results demonstrate that the UAE-based model achieves high detection accuracy with no observed false positives, providing an effective solution for enhancing operational reliability and production efficiency in semiconductor wafer slicing processes.