Technical Summary: Wire Arc Additive Manufacturing: A Progressive Multi-Architecture Computer Vision Framework for Real-Time Defect Detection

1. Problem Statement

Wire Arc Additive Manufacturing (WAAM) is a directed energy deposition technology used for fabricating large-scale metallic components. However, the process is highly sensitive to thermal gradients and arc-wire interactions, leading to defects such as humping, bead waviness, lack of fusion, and excessive spatter. Real-time monitoring is critical for closed-loop control, yet traditional methods relying on electrical signals (current/voltage) are insufficient for detecting complex visual anomalies.

Computer vision approaches face significant challenges due to the extreme luminosity of the welding arc, which causes overexposure in the image center and underexposure in the cooling weld trail. Furthermore, existing studies often rely on "black-box" deep learning models that lack interpretability or focus on single physical phenomena (e.g., only spatter or only melt pool width) rather than a holistic process state. There is a need for a framework that balances computational efficiency, model interpretability, and the ability to detect dynamic anomalies in real-time.

2. Methodology

The paper proposes and evaluates a progressive, multi-tiered computer vision framework. Rather than deploying a single complex model immediately, the authors systematically characterize trade-offs between computational efficiency, interpretability, and accuracy by testing four distinct model families.

2.1 Data Collection and Preprocessing

• Dataset: 11 high-resolution video samples (approx. 6,842 frames) captured at 30 fps using an HDR monochrome camera with narrow-band spectral filtering.
• Annotation: Frames are manually categorized into three classes: Background (idle), Good (stable deposition), and Bad (defects including bead irregularities and spatter).
• Class Imbalance: The dataset is heavily imbalanced (stable frames dominate). The authors employ resampling techniques (oversampling the "Bad" class to 40% of stable frames) and data augmentation to mitigate this.
• Heuristic Filtering: A dual-metric heuristic (mean intensity and variance) filters out idle periods. A "Continuous Segmentation Block" tracks spatial persistence to distinguish the weld pool from transient spatter, improving active frame identification recall to 1.00.
• Preprocessing Pipeline: Includes Contrast Limited Adaptive Histogram Equalization (CLAHE), Sobel edge mapping, logarithmic transforms, and temporal differencing to handle dynamic range and highlight motion.

2.2 Model Architectures Evaluated

The study progresses through four architectural stages:

1. Modular Two-Stage Classical Pipeline:

   • Uses a frozen CNN backbone (MobileNetV2, EfficientNet-B0, or ResNet-50) pre-trained on ImageNet to extract spatial features.
   • Feeds these embeddings into classical classifiers (Logistic Regression, Random Forest).
   • Goal: Establish a baseline and test interpretability.

2. End-to-End Deep Learning Models:

   • Fine-tunes lightweight architectures (MobileNetV2, EfficientNet-B0) with all parameters trainable.
   • Goal: Overcome the "frozen backbone" bottleneck and learn domain-specific features.

3. Hybrid CNN–LSTM Architecture:

   • Combines a spatial CNN backbone (EfficientNet-B0) with a Long Short-Term Memory (LSTM) layer.
   • Processes a temporal window of frames to capture the onset and progression of defects, distinguishing transient noise from sustained defect states.

4. MCFE-Net (Proposed Novelty):

   • Multi-Channel Feature Engineering Network: Replaces fixed, hand-crafted preprocessing with learnable, differentiable modules.
   • Contrast Enhancement Module (CEM): A learnable convolutional network approximating local contrast normalization.
   • Edge Salience Module (ESM): A learnable depthwise convolution initialized with Sobel filters.
   • These modules dynamically synthesize input channels (R, G, B) from the raw grayscale frame, which are then fed into the CNN backbone.
   • MCFE-CNN-LSTM: The final proposed architecture combines the learnable preprocessing (MCFE) with the temporal modeling (LSTM).

3. Key Results

The models were evaluated on a held-out test set of unseen video runs to ensure sequential integrity and prevent data leakage.

• Classical Two-Stage Pipeline: The best configuration (CLAHE + EfficientNet-B0 + Random Forest) achieved an Active class recall of 0.995 and an overall F1-score of 0.997. However, this was limited to binary classification (Active vs. Background) and could not perform fine-grained 3-class quality monitoring.
• End-to-End Deep Learning:
  • MobileNetV2: Achieved 99% accuracy and a "Bad" class F1-score of 0.99, demonstrating that joint optimization overcomes the frozen-backbone bottleneck.
  • EfficientNet-B0: Achieved 93% accuracy but suffered from a significant drop in defect recall (0.53), likely due to overfitting on the dominant classes given the smaller dataset size.
• CNN-LSTM: By adding temporal context, the model improved the "Bad" class F1-score to 0.74 (compared to per-frame models) and reduced false positives caused by transient arc flashes.
• MCFE-Net and MCFE-CNN-LSTM (Proposed):
  • MCFE-Net (EfficientNet backbone): Achieved 98% accuracy with perfect precision (1.00) for the "Bad" class.
  • MCFE-CNN-LSTM: The final proposed architecture achieved an overall F1-score of 0.987, a Defect Recall of 0.980, and a Defect Precision of 1.00.
  • This architecture successfully eliminated false positives from transient noise while maintaining high sensitivity to gradual weld pool geometry drift.

4. Key Contributions

1. Progressive Evaluation Framework: The paper provides a systematic comparison of model complexity, demonstrating that increased complexity is only justified when simpler baselines fail (e.g., classical models suffice for binary detection, but deep learning is required for 3-class quality monitoring).
2. Learnable Preprocessing (MCFE-Net): The introduction of differentiable preprocessing modules (CEM and ESM) that are jointly optimized with the classification objective. This removes the reliance on fixed, hand-crafted heuristics and allows the network to discover optimal feature representations.
3. Temporal Modeling for WAAM: The demonstration that temporal sequence modeling (LSTM) is critical for distinguishing developing defects from transient arc fluctuations, significantly improving defect recall over stateless per-frame models.
4. Interpretability: The framework maintains structural interpretability. The learned CEM and ESM filters can be visualized, allowing engineers to understand which visual features the model prioritizes for defect detection.
5. Real-Time Interface: Development of a web-based dashboard (FastAPI) for real-time model evaluation and visualization of predictions against ground truth, facilitating potential integration into closed-loop control systems.

5. Significance and Claims

The paper claims that the proposed MCFE-CNN-LSTM framework represents a significant advancement in WAAM monitoring by resolving the trade-off between accuracy and interpretability.

• Industrial Relevance: The perfect precision (1.00) for the defect class is highlighted as crucial for industrial applications, as it minimizes costly false alarms that could disrupt production.
• Robustness: The combination of learnable preprocessing and temporal recurrence allows the system to handle the extreme dynamic range of WAAM imagery and the temporal nature of defects better than previous stateless or fixed-preprocessing models.
• Scalability: The framework is designed to be lightweight (using MobileNetV2 or EfficientNet-B0 backbones) and suitable for real-time deployment.

The authors conclude that while their framework achieves near-perfect performance on the specific test dataset, future work should address class imbalance via advanced synthesis (GANs/SMOTE), explore Transformer architectures for longer temporal dependencies, and integrate multi-modal sensor data (current/voltage) to further enhance robustness.