Abstract
<title>Abstract</title> <p> Wire Arc Additive Manufacturing (WAAM) involves complex, multi-parameter interactions in which weld pool stability, arc behavior, and deposition quality must be continuously monitored. To address the limitations of traditional black-box deep learning models, this paper evaluates a progressive, multi-tiered computer vision and machine learning framework. We first establish a highly interpretable two-step architecture utilizing Convolutional Neural Networks (CNNs) for spatial feature extraction from preprocessed video frames, followed by classical machine learning classifiers to establish baseline operational states. Building upon these spatial embeddings, we evaluate end-to-end deep learning models including MobileNetV2 and EfficientNet-B0, and introduce a hybrid Recurrent Convolutional architecture (CNN-LSTM) to capture critical spatial-temporal dynamics. Finally, we propose <bold>MCFE-Net</bold> (Multi-Channel Feature Engineering Network), a novel architecture in which the preprocessing pipeline itself becomes a set of differentiable, learnable modules jointly optimized with the classification objective. This progressive evaluation explicitly characterizes the trade-offs between computational efficiency, interpretability, and the system's ability to accurately detect dynamic anomalies and irregular fusion states in real-time WAAM monitoring. </p>