Selective laser melting (SLM) is a critical additive manufacturing process. However, defects that occur during the powder spreading can significantly affect the final product quality. Real-time detection of these anomalies is crucial for maintaining high manufacturing standards. This research proposes a deep learning-based method to identify powder spreading anomalies during the SLM process. We develop a dataset containing images under five common powder bed conditions: normal, scratch, uneven coating, insufficient coating, and excessive coating. The purpose is to classify these conditions accurately, thereby enhancing real-time monitoring and quality control in industrial applications.

We construct the dataset using an SLM machine, simulating five distinct powder bed states of normal, scratch, uneven powder coating, insufficient powder coating, and excessive powder coating. A total of 1327 images are captured from various angles for each defect type to enhance dataset diversity. Three deep learning models, namely, VGG-16, ResNet-101, and EfficientNetV2-XL, are trained and evaluated on this dataset. The dataset is divided into training and testing sets, and image preprocessing is applied that includes resizing and normalization. Model performance is assessed based on metrics such as accuracy, precision, recall, and F1-score (Table 1 and Fig. 2).

Experimental results show that all three models perform well in classifying powder bed defects, with VGG-16 achieving the best balance between speed and accuracy. VGG-16 attains a classification accuracy of 99.63% and frame rate of 68.24 frame/s [Fig. 4(a)], making it suitable for real-time applications. ResNet-101 and EfficientNetV2-XL also perform well, particularly in identifying complex defects such as uneven powder distribution (Table 3). However, ResNet-101 exhibits slower inference speeds, restricting its use in scenarios requiring rapid detection. EfficientNetV2-XL demonstrates robustness in detecting larger and more complex defects, achieving an accuracy of 99.19% (Table 4), but its slower processing speed limits its suitability for real-time systems. Heatmap analysis using Grad-CAM (Fig. 7) reveals that VGG-16 focuses more on localized defect regions, which contributes to its enhanced capabilities in detecting minor defects such as scratches. By contrast, ResNet-101 and EfficientNetV2-XL exhibit broader attention, making them more effective in handling images with complex backgrounds. These findings underscore the importance of model selection based on defect type and real-time processing requirements.

We establish a novel multi-angle dataset capturing five powder bed states in the SLM process, significantly enhancing model training diversity and robustness. The VGG-16 model outperforms the others in terms of speed and precision, making it more suitable for real-time monitoring in industrial applications. EfficientNetV2-XL shows better performance in handling complex defects but has a slower inference speed, making it more suitable for less time-sensitive scenarios. Future research will focus on optimizing these models through techniques such as multi-scale feature fusion and integrating multi-modal sensor data to improve defect detection accuracy across a wider range of manufacturing conditions.