Wear Classification of Abrasive Flap Wheels using a Hierarchical Deep Learning Approach

Conference: CVPR 2025
arXiv: 2603.12852
Code: None
Area: Other
Keywords: Grinding wheel wear classification, hierarchical classification, transfer learning, EfficientNetV2, Grad-CAM

TL;DR

This paper proposes a hierarchical visual classification framework based on EfficientNetV2, which decomposes the wear state of abrasive flap wheels into three levels (usage state \(\rightarrow\) wear type \(\rightarrow\) severity), achieving classification accuracy of 93.8% to 99.3% across various subtasks.

Background & Motivation

Background: Abrasive flap wheels are widely used for finishing complex free-form surfaces in aerospace and mold manufacturing, adapting to different surface shapes due to their flexible nature. Tool condition monitoring in automated grinding is key to achieving adaptive process control.

Limitations of Prior Work: Due to their flexible structure, abrasive flap wheels exhibit complex and highly variable wear patterns (e.g., concave/convex contour changes, flap tearing), and traditional rule-based image processing methods are error-prone when dealing with such high-variance wear patterns. Existing CNN-based wear monitoring methods mainly target rigid tools like milling cutters and turning tools, lacking specialized frameworks for the wear characteristics of flexible tools.

Key Challenge: The wear of abrasive flap wheels is multi-dimensional (geometry, flap integrity, severity), whereas a single multi-class classifier requires sufficient training data for each wear combination, leading to data fragmentation issues.

Goal: To design a multi-level hierarchical classification framework that decomposes wear detection into multiple independent sub-decisions, where each sub-network is responsible for only one specific classification task.

Key Insight: Logical dependency relationships exist between different wear features. Hierarchical decomposition not only reduces the classification difficulty of each sub-network but also enables the detection of misclassifications through logical consistency checks.

Core Idea: Decompose the wear analysis of abrasive flap wheels into a three-level hierarchical decision tree (usage state \(\rightarrow\) profile type + flap tearing \(\rightarrow\) severity), using independent EfficientNetV2 classifiers for each level, and utilizing logical constraints between hierarchies to enhance robustness.

Method

Overall Architecture

The input consists of radial and axial perspective images of the abrasive flap wheel. A three-level hierarchical classification is sequentially performed to judge: Level 1 determines the usage state (new/used), Level 2 determines the wear profile type (rectangular/concave/convex) and flap tearing (present/absent), and Level 3 evaluates the severity of the concave or convex profile (partial/complete deformation).

Key Designs

1. Hierarchical Classification Architecture:

   • Function: Decomposes complex multi-class wear classification into multiple binary or ternary sub-classification tasks.
   • Mechanism: Unlike a monolithic classifier, hierarchical classification allows each sub-network to focus on a specific wear dimension. Level 1 uses radial images to determine if the tool is new or used, Level 2 simultaneously performs profile classification and flap tear detection, and Level 3 refines the severity of deformity. Each sub-network can be trained independently, avoiding the data scarcity issues caused by category combination explosion.
   • Design Motivation: The optimal camera angles vary for different wear features—flap tearing is easiest to observe from the axial perspective, whereas contour changes require a radial view. The hierarchical structure allows different sub-networks to utilize different perspective inputs.

2. Logical Consistency Check:

   • Function: Detects contradictory classification results between levels.
   • Mechanism: Defines 11 consistent outcome paths and 3 contradictory paths. For example, a logical contradiction occurs if Level 1 is classified as "new" but Level 2 detects flap tearing or a non-rectangular profile. Contradictory results can trigger re-detection or manual review.
   • Design Motivation: Leverages physical prior knowledge of wear to provide additional validation for classification results, improving reliability in industrial environments.

3. Transfer Learning based on EfficientNetV2:

   • Function: Achieves high-accuracy classification on limited datasets.
   • Mechanism: Uses ImageNet pre-trained EfficientNetV2 as the feature extractor. EfficientNetV2-L is used for flap tear detection, and EfficientNetV2-S is used for the remaining sub-networks. The AdamW optimizer is employed with both the learning rate and weight decay set to \(1 \times 10^{-4}\).
   • Design Motivation: Labeled data in industrial scenarios is limited (approximately 13,240 images). Transfer learning can fully leverage pre-trained features.

Loss & Training

Each sub-network is trained independently using standard cross-entropy loss. During training, data augmentation techniques such as random rotation, brightness, and contrast adjustments are applied to the images. Axial images are converted to grayscale for flap tear detection.

Key Experimental Results

Main Results

Evaluated on a total of 13,240 images from 105 abrasive flap wheels with different wear conditions:

Subtask	Accuracy	F1-score	AUC
Usage State Classification (Level 1)	98.6%	0.983	0.999
Flap Profile Classification (Level 2)	95.4%	0.954	0.99
Flap Tear Detection (Level 2)	93.8%	0.935	0.98
Concave Severity (Level 3)	99.3%	0.993	1.00
Convex Severity (Level 3)	95.0%	0.948	≥0.98

Ablation Study

Configuration	Key Metric	Description
Hierarchical Classification (Ours)	93.8-99.3% across subtasks	Each subtask is optimized independently
Logical Consistency Check	Detects contradictory paths	Enhances industrial reliability
EfficientNetV2-L (Tearing)	93.8%	Larger model enhances accuracy in difficult-to-distinguish tasks
EfficientNetV2-S (Others)	>95%	Lightweight models are sufficient

Key Findings

• Flap tear detection (93.8%) is the most difficult subtask. The average confidence of misclassification (0.76) is significantly lower than the overall average (0.92), which can be further filtered using a confidence threshold.
• The unidirectional misclassification pattern of "complete \(\rightarrow\) partial" in convex profile classification indicates that the visual boundary between partial and complete convex profiles is gradual.
• Grad-CAM visualization validates that the model has learned physically relevant features (flap edge shapes, contour change regions).
• Rotational symmetry causes some concave/convex misclassifications; thus, vertical flip data augmentation is recommended.

Highlights & Insights

• Hierarchical Decomposition Approach: Decomposing complex multi-attribute classification tasks into multiple simple sub-decisions avoids the data demand issue caused by category combination explosion and can be transferred to other industrial inspection scenarios.
• Logical Consistency Check: Establishing constraint relationships between levels using domain priors to detect misclassifications is a lightweight yet effective way to improve robustness.
• Grad-CAM Validating Physical Features: Not only used to interpret the model, but also to verify whether the model has learned physically relevant wear features.

Limitations & Future Work

• The dataset scale is relatively small (approx. 13K images), and the generalization capability on few-shot categories remains to be validated.
• Lack of direct comparison experiments with monolithic multi-class classifiers.
• Flap tear detection has the lowest accuracy, and attention mechanisms or object detection methods have not been attempted.
• The error propagation problem between levels has not been analyzed in depth.

Related Work & Insights

• vs. Direct Multi-Class Classification: Hierarchical classification decomposition reduces the difficulty of sub-tasks and data requirements but may introduce error propagation.
• vs. Indirect Methods (Force/Vibration Signals): Direct visual detection decouples monitoring from workpiece geometric complexity.
• This paper demonstrates a practical case of combining industrial inspection problems with deep learning.

Rating

• Novelty: ⭐⭐⭐⭐ No breakthrough in methodology, mainly an engineering combination of hierarchical classification and transfer learning
• Experimental Thoroughness: ⭐⭐⭐⭐ Lack of comparison with end-to-end multi-class classifiers
• Writing Quality: ⭐⭐⭐⭐⭐ Clear description of industrial context, detailed Grad-CAM analysis
• Value: ⭐⭐⭐⭐ Practical value for wear detection of abrasive flap wheels, limited generality

Related Papers

• [CVPR 2025] Potential Field Based Deep Metric Learning
• [CVPR 2025] Improving Accuracy and Calibration via Differentiated Deep Mutual Learning
• [ICML 2025] Function Encoders: A Principled Approach to Transfer Learning in Hilbert Spaces
• [CVPR 2026] HierUQ: Hierarchical Uncertainty Quantification with Adaptive Granularity Reconciliation for Degraded Image Classification
• [CVPR 2026] Revisiting F-measure Optimization in Multi-Label Classification: A Sampling-based Approach