Optimization of Arc Detection Algorithms for High-Voltage Power Supplies in Electrostatic Chucks

Electrostatic chucks, core components in semiconductor manufacturing and precision instrument assembly, rely on high-voltage power supplies to generate electrostatic fields for non-contact workpiece fixation. However, high-voltage environments can easily trigger arc faults, ranging from workpiece displacement or damage to equipment burnout and production interruptions. Traditional arc detection methods face challenges such as high false alarm rates and response delays due to strong noise interference from high-voltage power supplies, load diversity, and the transient nature of arcs. Recent advances combining deep learning and signal decomposition algorithms have significantly improved detection accuracy and real-time performance, becoming key breakthroughs. 
Core Challenges in Arc Detection
1. High-Noise Background: High-frequency switching noise from high-voltage power supplies overlaps with arc signal frequency bands, making separation difficult with conventional filtering. 
2. Load Variability: Arc characteristics vary significantly across voltage levels (e.g., 220 V, 750 V, 1500 V). For instance, low-voltage arcs concentrate at 70–120 kHz, while high-voltage arcs extend to 100–150 kHz, rendering single-model detection ineffective. 
3. Transient Nature: Arcs last only microseconds and exhibit nonlinear current distortion, demanding ultra-fast algorithms. 
Limitations of Traditional Methods & Optimization Paths
Conventional approaches (e.g., Fourier transforms, threshold comparisons) rely on manual feature extraction and lack adaptability in complex scenarios: 
• Frequency-Domain Analysis: Fast Fourier Transform (FFT) is highly sensitive to non-stationary signals and noise, with false alarm rates exceeding 10% in 750 V systems. 
• Time-Domain Thresholding: Current amplitude or rate-of-change thresholds struggle to distinguish arcs from load variations, leading to high missed detection rates. 
To overcome these limitations, novel algorithms focus on: 
1. Signal Decomposition & Feature Enhancement 
   Improved Complete Ensemble Empirical Mode Decomposition (ICEEMDAN) adaptively injects white noise components to suppress mode mixing and endpoint effects. Key steps include: 
   • Signal preprocessing: DC component removal and normalization eliminate baseline drift. 
   • Intrinsic Mode Function (IMF) screening: IMFs with energy ratios > threshold γ are retained to reconstruct arc-dominant signals. Experiments show this method boosts SNR by 15 dB under strong noise, reducing feature extraction errors to <5%. 
2. Deep Learning Model Fusion 
   • Dual-Channel 1D Convolutional Neural Network (1D-CNN): Processes arc voltage and load voltage in parallel to capture temporal dependencies and transient features. A Softmax classifier outputs arc severity (none/mild/severe) with 98.3% accuracy. 
   • Optimized Transformer Model: Sliding window operations and feature fusion (Patch Merging) reduce computational load while expanding the receptive field. Combined with Focal Loss to address data imbalance, it cuts detection latency to 2 ms at 512 kHz sampling rates, with only 2.47% false alarms. 
3. Multi-Algorithm Collaboration & Clustering 
   Post-feature fusion requires high-robustness classifiers: 
   • DBSCAN Clustering: Groups feature vectors (e.g., kurtosis, frequency variance) without preset cluster counts, offering strong noise immunity. Separation between arc and normal signal clusters exceeds 90%, outperforming K-Means. 
   • Random Forest (RF) Adaptive Optimization: Improved Mel-Frequency Cepstral Coefficients (MFCC) filter designs reduce sensitivity to high-frequency noise. RF uses feature importance-weighted voting, achieving 99.12% classification accuracy. 
Multi-Algorithm Fusion Trends & Practical Efficacy
Combining ICEEMDAN preprocessing with deep learning models into a decomposition-reconstruction-classification three-tier architecture represents the state-of-the-art: 
1. Speed Optimization: 1D-CNN first checks arc existence; 2D-CNN triggers only for mild/severe arcs, slashing response time by 40%. 
2. Accuracy Gains: ICEEMDAN+Transformer achieves 99.1% accuracy in 750 V systems, a 35% improvement over FFT-based methods. 
3. Engineering Adaptability: Voltage-specific feature extraction (e.g., wavelet node 1 for 220 V, node 4 for 1500 V) improves computational efficiency by 50%. 
Conclusion & Future Directions
Algorithm optimization for arc detection in high-voltage electrostatic chucks is evolving from single models to multi-modal fusion. Future work should explore edge-computing architectures to reduce deployment costs for deep learning models, while physical modeling of arc generation mechanisms (e.g., plasma collision dynamics) could guide feature engineering to minimize data dependency. Through coordinated innovation in algorithms and hardware, microsecond-level arc interception is achievable, ensuring yield and safety in semiconductor manufacturing.