Abstract
In selective laser melting (SLM), spattering is an important phenomenon that is highly related to the quality of the manufactured parts. Characterisation and monitoring of spattering behaviours are highly valuable in understanding the manufacturing process and improving the manufacturing quality of SLM. This paper introduces a method of automatic visual classification to distinguish spattering characteristics of SLM processes in different manufacturing conditions. A compact feature descriptor is proposed to represent spattering patterns and its effectiveness is evaluated using real images captured in different conditions. The feature descriptor of this work combines information of spatter trajectory morphology, spatial distributions, and temporal information. The classification is performed using support vector machine (SVM) and random forests for testing and shows highly promising classification accuracy of about 97%. The advantages of this work include compactness for representation and semantic interpretability with the feature description. In addition, the qualities of manufacturing parts are mapped with spattering characteristics under different laser energy densities. Such a map table can be then used to define the desired spatter features, providing a non-contact monitoring solution for online anomaly detection. This work will lead to a further integration of real-time vision monitoring system for an online closed-loop prognostic system for SLM systems, in order to improve the performance in terms of manufacturing quality, power consumption, and fault detection.
Introduction: The Challenge of SLM Quality Control
Selective Laser Melting (SLM) is a powerful additive manufacturing technique, but it is prone to internal defects like pores and cracks that are often only detectable after the build is finished. While "spattering"—the ejection of molten material during the laser-matter interaction—is a known indicator of build quality, monitoring it effectively is difficult.
Existing solutions usually rely on expensive high-speed cameras, X-ray imaging, or thermal sensors that are hard to integrate into industrial workflows. Furthermore, many machine learning approaches (like deep learning) act as "black boxes," making it hard for engineers to understand why a process is failing. This paper seeks a middle ground: high-accuracy monitoring using "consumable" (off-the-shelf) cameras and a feature descriptor that humans can actually interpret.
Methods: A Specialized Vision Pipeline
The authors propose a specialized computer vision pipeline designed to turn raw video frames into a compact "fingerprint" of the spattering process:
- Preprocessing: The system uses morphological operations (binarization, erosion, and distance transforms) to automatically locate the center of the melt pool, ensuring the analysis remains accurate even as the laser moves across the build plate.
- Polar Decomposition: Because spatters radiate outward from the laser point, the image is divided into a grid of cells using polar coordinates (radial distance and angle).
- Line Segment Detection (LSD): Instead of analyzing raw pixels, the system extracts line segments to represent spatter trajectories. For each polar cell, it calculates the total length of segments, their principal orientation, and the standard deviation of these values.
- Temporal Integration: To make the system robust against outliers, the authors "stack" line segments from up to 13 consecutive frames into a single feature vector without increasing the vector's dimensionality.
- Classification: These feature vectors are fed into Support Vector Machines (SVM) and Random Forest classifiers to identify the laser energy density setting.
The study used a Renishaw AM250 SLM machine to produce 316L stainless steel samples. A total of 35,000 images were captured using a Panasonic Lumix DMC-LX100. The data was categorized into 5 distinct classes based on laser energy density ranging from 33 J/mm³ to 166.7 J/mm³.
Key Insights and Ideas
The effectiveness of the proposed descriptor lies in its ability to capture the physical nuances of different spatter types, which are directly linked to the stability of the molten pool.
- The Power of Polar Coordinates: Traditional image grids (Cartesian) struggle with radial movement. By using polar cells, the descriptor naturally captures the physics of ejections flying away from a central point.
- Semantic Interpretability: Unlike deep learning features, these descriptors mean something to an engineer. A "high standard deviation in orientation" in a specific cell translates directly to erratic, unstable spattering.
- Temporal Boosting: Stacking segments from multiple frames significantly improved accuracy (from ~76% to ~97%). This suggests that the consistency of spatter over a few milliseconds is a much stronger signal than any single snapshot.
- Spatter as a Proxy for Density: The paper confirms that spatter morphology changes predictably with energy density. Low energy produces few, short spatters (partial melting), while excessive energy produces long, upright trajectories (keyhole effect).
Key Results
The Random Forest classifier achieved a peak accuracy of 97% when using 13-frame temporal features and a 36x20 polar grid. The feature extraction process is compact, making the feature vector small enough for potential real-time application. Crucially, the authors successfully mapped spatter characteristics to part density, identifying an "optimum" spatter profile that corresponds to a relative part density of over 99%.
Limitations and Future Work
While the results are promising, the authors note a few hurdles for industrial adoption. The current implementation uses Python and Scikit-Learn for offline processing; to achieve true closed-loop control (adjusting the laser mid-print), the code needs to be optimized for GPU hardware or ported to C++. Additionally, the camera was placed at an angle outside the build chamber. Changes in viewpoint or distance could affect perceived trajectory lengths, requiring more robust spatial normalization. Finally, while the model was trained on 316L steel, future work is needed to see how well these descriptors transfer to different alloys or SLM machine architectures.
Summary updated