Zero Shot Learning for Novel Defect Detection 2026
Zero Shot Learning for Novel Defect Detection
ZSL and low-data paradigms are emerging as critical enablers for industrial defect detection, allowing AI systems to identify previously unseen defect categories without labeled examples at inference time. Patent filings concentrated between 2021 and 2026 signal accelerating innovation across semiconductor, additive manufacturing, and surface inspection domains.
ZSL Defect Detection: From Supervised Baselines to Zero-Shot Architectures
Zero-shot learning for defect detection spans methods that identify defect categories absent from training data — encompassing ZSL, one-shot, few-shot, and incremental cold-start variants — applied to visual quality inspection across manufacturing and semiconductor domains. The core challenge is data scarcity: defective samples are rare by definition and annotation is expensive at scale.
Three overlapping technical families structure this field: semantic attribute and prompt-based ZSL that bridges visual and textual class descriptors; metric and similarity learning that projects images into latent spaces for novel-instance recognition; and anomaly detection under distribution shift, trained exclusively on normal samples to flag any deviation without prior defect exposure.
Publication dates in the retrieved records span 2017–2026, with the most ZSL-specific filings concentrated between 2021 and 2026. The 2017–2019 period established supervised deep learning baselines, particularly KLA-Tencor’s dual-network architecture for semiconductor inspection. The 2020–2021 period saw the first explicit low-shot and generalized ZSL frameworks for industrial defects, including Siamese networks for steel and a generalized ZSL framework for wheel hub defects.
Among 70+ retrieved records in this dataset, 6 patent records explicitly address zero-shot or cold-start defect detection and 8 address few-shot or one-shot detection. KLA Corporation holds the largest filing count in this dataset with 6+ US/WO patent families, while ZSL-native filings are distributed primarily across Chinese academic institutions, indicating an early competitive stage with limited incumbent lock-in.
Patent Activity and Technology Cluster Distribution
Retrieved patent records show a clear shift from supervised baseline architectures (2017–2020) toward low-shot and zero-shot native approaches (2021–2026), with the most ZSL-specific filings concentrated in the most recent period. Application domains span semiconductor wafer inspection, additive manufacturing, steel and textile surfaces, aerospace, and automotive.
Patent Filings by Application Domain — ZSL Defect Detection (Dataset Snapshot)
Semiconductor and wafer inspection accounts for the largest single cluster in this dataset, driven by KLA Corporation’s multiple patent families, followed by additive manufacturing and surface inspection domains.
↗ Click bars to exploreZSL/FSL Patent Filings by Era — Innovation Phase Progression (Dataset Snapshot)
Filing volume in this dataset shows a clear acceleration of ZSL-native and prompt-based architectures in the 2024–2026 window, versus the metric-learning and supervised baseline dominance of the 2017–2021 era.
↗ Click bars to exploreKey Deployment Domains for ZSL Defect Detection Technology
Retrieved records span five primary application verticals — semiconductor wafer inspection, additive manufacturing, steel and textile surfaces, aerospace components, and automotive/electronics — each with distinct technical requirements driving adoption of low-data learning architectures.
Semiconductor Wafer Inspection
The largest single patent cluster in this dataset, anchored by KLA Corporation’s 6+ US/WO filings on learnable defect detection using deep metric learning in latent space for wafer and chip-level inspection. ASML’s charged-particle inspection system (TW, 2025) applies deep learning-based classification to wafer SEM images. A 2022 survey covers convergence of deep learning with nanophotonics at sub-10 nm nodes.
Semiconductor InspectionAdditive Manufacturing / L-PBF
Divergent Technologies holds 3 US filings on ML-based defect identification in additive manufacturing systems (2022–2023), with a Sigma Labs WO counterpart. Boeing’s graph-learning neural network patent (US, 2024) applies GNN architecture to spectral sensor data for defect detection in additive processes. ArianeGroup’s 2023 US patent employs autoencoder-type CNNs on infrared images — an anomaly-detection-compatible architecture for laser additive manufacturing.
Advanced ManufacturingSteel, Textile Surface Inspection
One-shot recognition of manufacturing defects in steel surfaces (2020) demonstrated Siamese networks for steel quality control using a single reference image. A 2022 paper applied randomized networks for fast few-shot classification of steel surface defects. Fabric defect segmentation using a lightweight GAN for industrial IoT (2022) and few-shot contrastive learning for wear debris detection (2022) extend the low-data paradigm to textiles and mechanical wear monitoring.
Surface InspectionAerospace and Defense Inspection
Nanjing University of Aeronautics and Astronautics filed a US patent (2023) on a few-shot defect detection method based on metric learning, explicitly targeting aviation, aerospace, and navigation equipment under high-precision, scarce-label conditions. A 2023 literature record covers deep-learning-based borescope inspection of high-pressure compressor blades using U-Net and GAN architectures under data-scarce conditions.
Aerospace InspectionLeading Patent Assignees in ZSL Defect Detection — Dataset Snapshot
In this dataset, KLA Corporation holds the highest filing count with 6+ US/WO patent families concentrated in semiconductor-domain metric learning and annotation efficiency. The ZSL-native filing space — zero-shot anomaly detection, cold-start, and prompt-based architectures — is distributed across Chinese academic institutions, with no equivalent US or EU industrial filings retrieved in this dataset for those sub-approaches.
Top Assignees by Filing Count — ZSL Defect Detection in Retrieved Records
↗ Click bars to exploreKLA Corporation
KLA Corporation holds 6+ retrieved US/WO patent filings spanning 2019–2026, making it the highest-volume assignee in this dataset. Core technology areas include deep metric learning for learnable defect detection in semiconductor applications (US/WO, 2020–2023), neural network training for low-resolution wafer images (US/WO, 2019–2020), deep learning defect detection (US, 2022), and a signal-to-noise metric for annotation guidance and DL model tunability (US/WO, 2026). Patent families include both active and granted US and WO members.
United StatesShanghai Jiao Tong University
Shanghai Jiao Tong University has 2 retrieved CN patent filings on cold-start defect detection for industrial quality inspection, dated 2022 and 2025. The 2025 filing addresses cross-domain defect dataset transfer with distribution-aware feature migration, directly tackling the deployment cold-start scenario where inspection algorithms must function before defect data is available. These filings are among the most directly ZSL-native patents retrieved, targeting the practical manufacturing deployment gap.
China — CNForward-Looking Signals from 2024–2026 Filings
Four distinct forward-looking directions are identifiable from filings dated 2024–2026 in this dataset, spanning prompt-learning architectures, one-shot wafer meta-learning, cold-start deployment systems, and continuous incremental adaptation with thermal fusion.
Prompt-Learning and Vision-Language Model Adaptation for ZSAD
Xi’an Jiaotong-Liverpool University’s 2024 CN patent is the first explicitly ZSAD-native filing in this dataset, adapting large vision-language models to industrial inspection through dynamic prompt adjustment per input image. It addresses the weakness of large models in local detail understanding and uses publicly available defect datasets to transfer large model capacity into anomaly detection. This direction enables generalization across product types without any target-domain training data.
One-Shot Wafer Inspection with Siamese Meta-Learning (2025)
Two active CN filings from the University of the Chinese Academy of Sciences, Hangzhou Institute (May–June 2025) introduce Siamese dual-branch networks with meta-learning strategies requiring only one reference sample plus limited defect data with augmentation. Dynamic threshold adjustment via image sharpness value and single inner-loop fine-tuning enable rapid adaptation to new defect tasks. These filings directly address the zero/one-shot gap for semiconductor quality control at advanced nodes.
Metric Learning vs. Prompt-Based ZSL for Defect Detection
Click any row to explore further.
| Dimension | Metric / Similarity Learning | Prompt-Based / Semantic ZSL |
|---|---|---|
| Core mechanism | Embedding function trained so novel defects are identified by proximity to 1–5 reference examples in feature space | Bridges visual features of unseen defect classes with textual or attribute-based class descriptions via prompt tokens or semantic vectors |
| Training requirement | Requires meta-training across many defect class episodes; one reference image needed at inference | Leverages pre-trained vision-language model capacity; zero visual training samples required for target defect class |
| Representative patents | KLA Corporation Learnable Defect Detection (US/WO, 2020–2023); University of Chinese Academy of Sciences One-Shot Wafer Detection (CN, 2025) | Xi’an Jiaotong-Liverpool University Zero-Shot Defect Detection via Prompt Learning (CN, 2024) |
| Primary application | Semiconductor wafer inspection; aerospace assembly components; steel surface quality control | Industrial anomaly detection across product types without target-domain training data |
| Key challenge addressed | Data scarcity for novel defect classes; rapid adaptation to new defect tasks via single inner-loop fine-tuning | Weakness of large models in local detail understanding; generalization across unseen product types |
| Filing jurisdiction concentration | US, WO, CN — dominant patent strategy for incumbent players (KLA) and Chinese academic institutions | CN only as of 2024 in this dataset; no equivalent US or EP filings retrieved |
| Competitive landscape | KLA holds strongest patent position in this dataset; metric-learning space has moderate incumbent lock-in | Structurally open in this dataset — only one filing retrieved; narrow window to establish foundational IP in US/EP |
Frequently Asked Questions: Zero Shot Learning for Defect Detection
In this dataset’s framing, ZSL for defect detection refers to methods that detect defect categories absent from training data by using semantic descriptors, attribute vectors, or prompt tokens to generalize to unseen defect classes without any visual training examples for those classes at inference time. It addresses the fundamental data scarcity problem where defective samples are rare and annotation is expensive.
Zero-shot approaches require zero visual training examples for the target defect class, relying on semantic bridging or anomaly scoring. Few-shot and one-shot approaches use a small number of reference examples (typically 1–5) in a metric or Siamese learning framework to identify novel defects by proximity in feature space. Among retrieved records, 6 patents explicitly address zero-shot or cold-start detection and 8 address few-shot or one-shot detection.
In this dataset, KLA Corporation holds the highest filing count with 6+ US/WO patent families spanning 2019–2026, covering deep metric learning for semiconductor defect detection, neural network training for low-resolution wafer images, deep learning defect detection, and SNR-guided annotation efficiency.
As described in retrieved records, cold-start defect detection refers to the inability to quickly deploy inspection algorithms when defect data is unavailable at launch — such as when a new product line starts or a novel defect type first appears. Shanghai Jiao Tong University addresses this with cross-domain defect dataset transfer and distribution-aware feature migration in CN filings dated 2022 and 2025.
ZSAD is framed in the retrieved records as a distinct task class where a model generalizes to unseen industrial defect types without any target-domain training data, using prompt learning to adapt large vision-language models. The first explicitly ZSAD-native patent in this dataset is from Xi’an Jiaotong-Liverpool University (CN, 2024), which uses dynamic prompt adjustment per input image to address local detail understanding weaknesses in large models.
According to retrieved records, the 2026 CN filings from Shaanxi University of Technology use channel-selective elastic weight consolidation (C-EWC) combined with thermal imaging fusion to prevent catastrophic forgetting while continuously recognizing newly emerged defect classes. Beijing Huatai Hengnuo Technology’s 2026 CN system constructs a new-defect sample set upon emergence of a new defect type and updates the recognition model incrementally, outputting defect category, location coordinates, and area.
Data and insights on this page are based on a limited patent and literature dataset and are for reference only. Figures may not represent the complete technology landscape.