Laser cladding technology has established itself as a critical surface treatment technique in modern manufacturing due to its ability to enhance surface properties and extend the service life of components. Its wide industrial applications span the aerospace, automotive, and energy sectors. However, the laser cladding process commonly encounters quality issues such as not fused, uneven coating, and crack, which limit its broader adoption. These defects arise from challenges such as fluctuating laser power, inconsistent powder feed rates, and rapid cooling during the cladding process. Traditional defect detection methods, relying on manually designed feature extraction or machine vision systems, often fail to capture the complexity and variability of cladding defects under different manufacturing conditions, leading to suboptimal detection accuracy. To address these limitations, we aim to develop an intelligent and automated defect recognition framework based on deep learning. A dual-channel residual neural network (ResNet) is proposed to automatically extract and classify defect features in cladding images. The model’s performance is further enhanced by integrating an improved loss function that combines the traditional Softmax loss with a center loss function, which improves feature separability.

The proposed framework utilizes a dual-channel ResNet-18 architecture for its efficient feature extraction capabilities and robust performance on moderately sized datasets. The ResNet-18 model is selected due to its lightweight design, which balances computational efficiency and classification accuracy. A novel combination of the Softmax loss and center loss functions is used to enhance the network’s ability to distinguish between defect types and improve inter-class separability. The center loss function computes the Euclidean distance between each feature vector and its corresponding class center, encouraging features of the same class to cluster closely while maintaining separation from other classes. To construct a comprehensive dataset, three types of cladding states—good cladding, uneven cladding, and not fused—are considered. Each cladding state is subjected to 10 experimental trials, with different cladding parameters used for each trial. A total of 7200 images are collected, with 6000 allocated for training and 1200 for validation. Data augmentation techniques, including cropping, flipping, rotation, and random aspect ratio changes, are applied to increase the diversity of the dataset and mitigate overfitting risks (Fig. 6). The images are resized to 256 pixel×256 pixel to standardize input dimensions and reduce computational demands. One of the critical challenges in laser cladding defect detection is uneven illumination and strong reflections caused by the laser process. To address this, the K-SVD algorithm is employed for image enhancement. This algorithm separates illumination and reflection components, effectively improving image contrast and highlighting defect features (Fig. 7). The enhanced images are used to train the network, ensuring accurate defect detection under varying lighting conditions. During training, the Adam optimizer is used with an initial learning rate of 0.001 and a batch size of 20. The learning rate is decayed by a factor of 0.1 every seven epochs to ensure convergence. The training process incorporates Xavier initialization to prevent gradient vanishing or explosion, ensuring stable performance.

1) Model performance: The dual-channel ResNet demonstrates outstanding performance, achieving an overall recognition accuracy of 98% on the validation dataset [Fig. 9(a)]. The accuracy curve shows rapid improvement in the early epochs, stabilizing after the 12st epoch. The model’s loss curve [Fig. 9(b)] indicates effective learning, with minimal overfitting observed due to the data augmentation strategies employed. 2) Comparison with SVM: To evaluate the effectiveness of the proposed model, its performance is compared with that of a traditional support vector machine (SVM) classifier. The SVM model, trained on the same dataset, achieves lower accuracy and precision across all defect categories. For example, the SVM model achieves only 27.5% accuracy in identifying not fused, while the dual-channel ResNet achieves 100% accuracy for the same category (Table 4). The results confirm the superiority of the proposed framework in handling complex defect patterns and varying cladding conditions. 3) Feature visualization: Guided Grad-CAM technology is employed to visualize the regions of interest that the model focuses on during classification (Fig. 11). The visualization highlights critical areas of the cladding surface, such as defect boundaries and morphological features, providing insights into the model’s decision-making process. For instance, in good cladding images, the model primarily focuses on smooth, uniform regions, while for not fused, it highlights areas with irregular textures and unbonded material. This interpretability is crucial for industrial applications, as it enhances trust in the system’s predictions. 4) K-SVD enhancement: The K-SVD algorithm significantly improves image quality by reducing noise and enhancing defect features (Fig. 7). This is particularly important for images with strong reflections or uneven lighting, which are common in laser cladding processes. The enhanced images enable the model to accurately identify subtle defect patterns, contributing to its high classification accuracy. 5) Comparison with traditional ResNet: The proposed dual-channel ResNet with center loss outperforms the traditional ResNet model in most defect categories. For not fused, the dual-channel ResNet achieves perfect precision and accuracy (Table 5), demonstrating its ability to cluster features more effectively. While the traditional ResNet performs slightly better in identifying good cladding images, the dual-channel model exhibits superior performance overall, particularly in distinguishing between defect categories. 6) Interpretability and Generalization: The use of the center loss function improves the model’s ability to generalize to unseen data. The improved inter-class separability ensures that even subtle differences between defect categories are effectively captured. This is particularly important for real-world applications, where defect patterns may vary significantly between production batches or materials.

We present a novel dual-channel residual neural network for defect recognition in laser cladding processes. By combining the ResNet-18 architecture with a center loss function, the proposed framework achieves an overall recognition accuracy of over 98%, significantly outperforming traditional methods such as SVM. The use of advanced image enhancement techniques, including the K-SVD algorithm, further improves the model’s performance by addressing challenges related to uneven illumination and reflections. Grad-CAM visualization with guidance provides valuable insights into the model’s decision-making process, enhancing its interpretability and reliability for industrial applications. The results demonstrate the potential of the proposed approach for intelligent quality control in laser cladding processes. Its ability to accurately identify and classify defects, combined with its interpretability, makes it a promising solution for industrial defect detection. Future work will focus on expanding the dataset to include additional defect categories and further enhancing the model’s robustness under diverse cladding conditions. By addressing these challenges, the proposed framework has the potential to significantly enhance manufacturing quality and efficiency.