PaperlayerView exact revision on arXiv

Source-linked AI summary

In-situ process monitoring and adaptive quality enhancement in laser additive manufacturing: a critical review

Lequn Chen, Guijun Bi, Xiling Yao, Jinlong Su, Chaolin Tan, Wenhe Feng, Michalis Benakis, Youxiang Chew, Seung Ki Moon

arXiv:2404.13673v1eess.SPeess.SY

TL;DR

Consistent quality and repeatability remain difficult in LAM because of its stochastic nature and defects that can occur even with optimized parameters. This review critically examines in-situ monitoring and adaptive quality-enhancement approaches, highlighting multisensor and machine-learning methods for improved defect prediction, detection, and correction.

  • Problem

    LAM’s stochastic nature and defects that persist despite optimized process parameters make consistent quality and repeatability difficult.

  • Method

    The paper provides a comprehensive and critical examination of in-situ monitoring, machine-learning prediction, and adaptive defect-management approaches in LAM.

  • Results

    Multisensor approaches can predict defects with superior accuracy than traditional single-sensor monitoring, while in-process defect correction supports improved overall product quality.

  • Takeaways & Limitations

    Future LAM systems should pursue robust predictive models and comprehensive multiscale defect-identification systems to support self-adaptation.

  • Takeaways & Limitations

    The review identifies a lack of standardized methodology for validating machine-learning model accuracy, alongside challenges in real-time image processing and sensor capability and cost.

Abstract

from arXiv · show

Laser Additive Manufacturing (LAM) presents unparalleled opportunities for fabricating complex, high-performance structures and components with unique material properties. Despite these advancements, achieving consistent part quality and process repeatability remains challenging. This paper provides a comprehensive review of various state-of-the-art in-situ process monitoring techniques, including optical-based monitoring, acoustic-based sensing, laser line scanning, and operando X-ray monitoring. These techniques are evaluated for their capabilities and limitations in detecting defects within Laser Powder Bed Fusion (LPBF) and Laser Directed Energy Deposition (LDED) processes. Furthermore, the review discusses emerging multisensor monitoring and machine learning (ML)-assisted defect detection methods, benchmarking ML models tailored for in-situ defect detection. The paper also discusses in-situ adaptive defect remediation strategies that advance LAM towards zero-defect autonomous operations, focusing on real-time closed-loop feedback control and defect correction methods. Research gaps such as the need for standardization, improved reliability and sensitivity, and decision-making strategies beyond early stopping are highlighted. Future directions are proposed, with an emphasis on multimodal sensor fusion for multiscale defect prediction and fault diagnosis, ultimately enabling self-adaptation in LAM processes. This paper aims to equip researchers and industry professionals with a holistic understanding of the current capabilities, limitations, and future directions in in-situ process monitoring and adaptive quality enhancement in LAM.

2.1. Laser Powder Bed Fusion

The supplied material lists terminology associated with Laser Powder Bed Fusion and related laser additive manufacturing monitoring systems, but does not describe LPBF principles or findings.

  • The glossary also includes machine-learning methods such as CNNs, ANNs, SVMs, GANs, and reinforcement learning.
  • The section’s supplied terminology includes LPBF alongside LDED and other laser additive manufacturing process variants.
  • Listed monitoring-related terms span optical imaging, infrared sensing, acoustic analysis, and X-ray computed tomography.

1. Introduction

LAM enables complex, customized manufacturing but remains vulnerable to stochastic defect formation that can undermine part reliability. This review integrates monitoring, machine-learning defect detection, adaptive quality enhancement, and future research directions across LAM techniques.

  • LAM’s stochastic process can produce porosity, cracks, and distortions despite optimized parameters, degrading mechanical properties and threatening reliability.
  • Earlier reviews often narrowed coverage to particular sensing methods, process-control topics, or AM techniques, limiting holistic assessment of LAM.
  • The review addresses earlier gaps by covering optical, acoustic, and infrared thermography monitoring alongside machine-learning-assisted defect detection.
  • The review emphasizes standardization, improved reliability and sensitivity, enhanced data interpretation, and decision-making strategies as continuing research needs.
  • It further examines closed-loop adaptive quality enhancement and proposes a roadmap toward fully autonomous, zero-defect LAM production.

2. Laser Additive Manufacturing and Defects

LAM encompasses LPBF and LDED processes that enable complex or large-scale fabrication but remain sensitive to process conditions and thermal dynamics. These sensitivities produce defects across multiple scales, motivating in-situ monitoring and adaptive quality enhancement.

  • LPBF and LDED processes: LPBF selectively melts successive powder layers with a laser, while LDED deposits powder or wire into a laser-generated melt pool.LPBF repeats recoating, scanning, solidification, and platform movement; LDED melts feedstock as it is deposited onto a substrate.
  • Process challenges: LPBF limitations include porosity, cracks, restricted build capacity, and residual-stress-induced distortion, while LDED can produce coarse microstructures, rougher surfaces, and dimensional inaccuracies.LDED’s high thermal input and large melt pool can weaken mechanical strength, while increased build rates compromise surface roughness and dimensional accuracy.
  • Process challenges: Both processes require careful parameter optimization because laser power, scan speed, layer or hatch dimensions, feedstock, and thermal conditions strongly influence quality.Even trial-and-error, mechanistic modelling, or machine-learning-based optimization may not eliminate variations in part quality.
  • Defects in LAM: LAM defects span micro/meso to macro scales, including porosity, cracks, discontinuities, layer unevenness, distortions, and microstructure inhomogeneity.Defect formation is linked to dynamic heat accumulation, residual stress, and stochastic transitions between conduction, keyhole-porosity, and lack-of-fusion regimes.
  • Defects in LAM: In-situ monitoring enables defect detection during manufacturing and supports immediate detection and in-process remediation as part of quality assurance.The review frames process-specific monitoring as a response to persistent stochastic defects and uncertain melt-pool behavior.

3. In-Situ Process Monitoring And Defect Detection In LAM

The review critically examines in-situ monitoring and defect-detection approaches for LAM, covering multiple sensing modalities and machine-learning methods. It emphasizes process-specific monitoring and evaluates current capabilities while identifying directions for more comprehensive defect detection.

  • Monitoring methods: The review evaluates optical, acoustic, laser-line, operando X-ray, and multisensor monitoring methods for LAM defect detection.These approaches target process signatures including melt-pool dynamics, thermal histories, and acoustic features from laser–material interactions.
  • Monitoring objectives: In-situ monitoring is positioned as a means of detecting defects early enough to help prevent quality deterioration and potential build failure.The review links monitoring with the broader goal of adaptive quality enhancement in LAM.
  • Machine-learning-assisted detection: The review examines machine-learning models specifically designed for in-situ defect detection as part of a multifaceted evaluation of current practice.The evaluation is intended to provide a comprehensive picture of state-of-the-art defect detection and its future directions.

3.1. Optical-Based Monitoring

Optical monitoring captures melt-pool visual and thermal features for process-quality assessment, defect detection, anomaly diagnosis, and localized quality prediction in LAM. Camera configuration, processing requirements, calibration, integration, and sensor trade-offs constrain reliability and real-time deployment.

  • Optical monitoring applications: Optical monitoring extracts melt-pool visual and thermal features that support process-quality assessment, defect detection, anomaly diagnosis, and fault diagnosis.The review organizes optical monitoring around melt-pool dynamics, feature extraction, defect detection, anomaly detection, and infrared thermal imaging.
  • Camera configurations: Coaxial cameras provide direct overhead views that avoid the image transformation and calibration required by off-axis monitoring.Off-axis views can make melt-pool dimensions difficult to measure because of oblique viewing angles, whereas coaxial setups are increasingly favored for direct viewing.
  • Melt-pool features: Melt-pool geometry and visual characteristics correspond to process stability, melting and cooling states, solid-liquid interfaces, and underlying metallurgical phenomena.Reported features include melt-pool width, size, area, grayscale distribution, and other physics-informed indicators.
  • Machine-learning-assisted detection: 91.2% accuracy was achieved when a CNN classified porosity occurrences during DED using high-speed camera data, with micropores below 100 µm also predicted in titanium samples.
  • Infrared thermal imaging: Infrared thermal-history features enabled voxel-level porosity estimation and ML prediction, with models using time above melting threshold and maximum IR radiance achieving F1 scores above 0.96.Maximum IR radiance was identified as the most significant feature for predicting voxel state.
  • Challenges and limitations: Optical monitoring remains limited by real-time data-processing demands, sensor capability–cost trade-offs, emissivity calibration, custom integration, occlusion, and off-axis viewing constraints.Accurate temperature profiles remain difficult because emissivity varies with temperature, wavelength, material phase, surface roughness, and other factors.

3.2. Acoustic-Based Monitoring

Acoustic monitoring detects process anomalies and defects across LPBF and LDED, especially when signal processing and machine learning relate acoustic features to defect states. Its broader deployment remains constrained by noise, resolution, interpretation, and reproducibility challenges.

  • Acoustic monitoring is vulnerable to environmental and process noise, while multi-layer structures generate complex signatures that are difficult to interpret.
  • Acoustic features distinguish process regimes and defect types, including pores, cracks, balling, overheating, and penetration states.
  • 97% accuracy was reported for keyhole-pore prediction using an SVM classifier, with pore locations spatially resolved in LPBF.
  • In LDED, acoustic emissions correlate with machine status, deposition parameters, powder flow, laser power, porosity, spatter, and line-width variation.
  • 89% overall accuracy, 93% keyhole-pore accuracy, and 98% AUC-ROC were achieved by MFCC-CNN, outperforming other models.
  • Further progress requires higher temporal and frequency resolution, standardized sensor setups, improved noise removal, and integration with optical or thermal sensing.

3.3. Laser Line Scanning

Laser line scanning uses triangulation to reconstruct surface topography and identify defects during LDED. Its usefulness is demonstrated for surface morphology, but noise, speed, integration, and limited sensitivity constrain broader deployment.

  • Laser triangulation converts reflected-line displacement into surface topography for in-process defect identification.
  • Raw point clouds contain outliers, substrates, and edge-related inaccuracies that can degrade surface-defect detection.
  • 2D depth images and point-cloud processing support CNN-based classification and pixel-wise identification of over-extrusion, under-extrusion, bulges, and dents.
  • Scanning may miss subtle defects, lag production rates, and produce false or missed readings under dust, smoke, and changing illumination.
  • Integration is easier in robotic LDED than in CNC-based LDED or LPBF because other systems lack accessible motion coordinates and space.
  • Future work targets adaptable, miniaturized scanners, real-time coordinate estimation, noise reduction, and robust point-cloud algorithms.

3.4. Other Emerging Monitoring Methods

Operando X-ray and related emerging methods reveal subsurface dynamics and defects during LAM. They provide high-resolution evidence and labels for monitoring models, but accessibility, safety, computation, integration, and scalability remain limiting factors.

  • Operando X-ray provides rapid, high-resolution keyhole-pore detection and ground-truth labels for annotating in-situ monitoring data.
  • Synchrotron X-ray combined with thermal imaging revealed keyhole oscillations and enabled sub-millisecond deep-learning detection of pore-generation events.
  • High-speed X-ray imaging in LDED tracks powder-particle flow, laser–matter interaction, and porosity formation.
  • ICI monitored surface topology and cracks, while its off-axis integration avoided alterations to laser-delivery optics and correlated above 0.93 with X-ray radiographs.
  • Operando X-ray is difficult to scale because synchrotron access, radiation safety, expense, data volume, and multi-layer monitoring impose substantial burdens.
  • ICI remains vulnerable to integration complexity, environmental conditions, and limited detection of small or deeply located defects.

3.5. Multisensor Monitoring And Data Fusion

Multisensor monitoring combines complementary measurements to improve defect detection across LAM’s spatial and temporal scales. Feature correlation, synchronization, and fusion support more informative quality assessment and process control.

  • Multisensor systems counterbalance single-sensor limitations, improve signal quality, and provide a more comprehensive view of complex LAM processes.
  • Heterogeneous sensor fusion detects LPBF flaws from micro-scale porosity to meso-scale layer inconsistencies and macroscale geometry defects.
  • Representative setups combine optical, thermal, acoustic, emission, and trajectory measurements across robotic LDED and LPBF platforms.
  • Sudden multisensor feature shifts can correspond to cracks and keyhole pores, while shared trends track heat accumulation and quality deterioration.
  • BR and AE sensors carry the most informative content, with AE gaining importance at shorter windows while BR becomes more reliable over longer integration periods.
  • Timestamp synchronization and feature-level fusion support location-specific defect estimation, although missing real-time TCP information makes spatial prediction difficult.

4. Adaptive Quality Enhancement

Adaptive quality enhancement addresses defects caused by fixed process settings through closed-loop control and in-process correction. Reported approaches improve geometric accuracy, porosity suppression, and microstructure homogeneity, but remain limited in validation scope and model fidelity.

  • Open-loop settings can create heat-distribution variability, over-melting, geometric deviations, keyhole pores, and microstructure inhomogeneity.
  • 4.1. Closed-Loop Feedback Control: Closed-loop control adjusts process parameters from feedback to reduce porosity, improve geometric accuracy, and enhance microstructure homogeneity.
  • 4.1. Closed-Loop Feedback Control: Adaptive control has improved geometric accuracy across varying shapes, materials, toolpaths, and process parameters without manual parameter adjustment.
  • Closed-loop methods are mainly validated on single tracks or thin walls, often use linear plant approximations, and underexplore complex geometries, microstructures, and mechanical properties.
  • 4.1. Closed-Loop Feedback Control: Closed-loop laser-power control produces a more consistent and homogeneous cellular grain structure by suppressing heat accumulation.
  • 4.2. In-Process Defect Correction: In-process defect correction uses additive or subtractive methods as a backup mechanism, compensating detected height variations within a few layers.

5. Summary And Future Perspectives

The paper’s concluding section reviews its key findings and identifies major challenges. It then sets a direction for future advancements in in-situ monitoring and control.

  • The section reviews developments from the preceding parts of the paper.
  • The discussion is organized around a summary of findings followed by a future outlook.
  • It identifies major challenges in the field and uses them to set a direction for future advancements.

5.1. Summary

The review compares optical, acoustic, scanning, X-ray, and multisensor monitoring approaches, then connects defect detection with adaptive control and correction. It emphasizes trade-offs among resolution, cost, deployment, and responsiveness while supporting multiscale monitoring and quality enhancement.

  • Optical, acoustic, laser-scanning, X-ray, and other sensing methods provide complementary capabilities for detecting and characterizing LAM defects.
  • Acoustic monitoring offers flexible installation, rapid response, and lower hardware cost than vision-based monitoring, but its noise rejection needs improvement.
  • Multisensor fusion predicts defects with superior accuracy to traditional single-sensor monitoring approaches.
  • Defect detection spans region-based, layer-wise, and real-time levels, trading detection resolution and speed against efficiency, computational demands, and cost.
  • A hierarchical multisensor system spanning these levels may provide desirable multiscale defect detection for application-specific needs.
  • Closed-loop control and in-process correction can reduce porosity, improve geometric accuracy and microstructure homogeneity, and support final product quality.
  • Monitoring and adaptive enhancement enable early nondestructive correction, process adaptation, microstructure control, and reduced waste.

5.2. Research Gaps And Future Perspectives

LAM’s broader deployment is constrained by incomplete sensor and validation standardization, limited spatially specific prediction, and insufficiently adaptive decision-making. The review proposes standardized evaluation, realistic datasets, multimodal sensing, and proactive defect mitigation as future priorities.

  • Standardization And Reproducibility: Standardized sensor setups, validation procedures, reporting practices, and test components are needed to improve reproducibility and industrial translation.Variations in sensor positioning, materials, process parameters, validation structures, and reporting limit comparability and robustness across studies.
  • Standardization And Reproducibility: ML models must be validated across diverse conditions and materials using approaches that assess defect detectability rather than only conventional accuracy.Probability-of-detection and probability-of-false-alarm curves aligned with NDE standards, including the a90/95 metric, are presented as more relevant assessment tools.
  • Location-Specific Quality Prediction: Current models often overlook spatial information, limiting location-specific quality prediction because monitoring data are difficult to correlate with localized quality characteristics.Heat distribution, melt-pool dynamics, residual stresses, and limited access to real-time position data complicate spatial prediction in LPBF and LDED.
  • Robust ML Models: Training datasets should include components produced under nonideal settings because single-track and thin-wall studies do not adequately represent real production complexity.Real production uses predefined parameters across complex geometries, where defects can arise unpredictably at different locations.
  • Multisensor Monitoring: Future monitoring systems should combine noise-resistant processing with tailored sensor fusion while minimizing redundant sensors to improve reliability and cost-effectiveness.Understanding sensor noise is identified as necessary for reducing disturbances and developing efficient multisensor monitoring.
  • Decision-Making Strategies Beyond Early Stopping: Decision-making should progress beyond anomaly-triggered termination toward dynamic parameter adjustments that mitigate and rectify defects during production.The review highlights physics-informed, data-driven, or rule-based strategies for real-time process adjustment and proactive defect prevention.
  • Future Directions: A hierarchical multisensor framework spanning real-time, layer-wise, and region-based monitoring is proposed to support multiscale defect prediction and autonomous self-adaptation.The roadmap links complementary sensing levels with improved process predictability, quality, and reliability.

5.3. Roadmap Towards Fully Autonomous And Self-Adaptation LAM Processes

The roadmap moves LAM from defect detection toward proactive, self-adaptive operation by combining multimodal sensing, predictive mitigation, and hierarchical defect correction. Its intended endpoint is continuous autonomous production with improved quality control despite process anomalies.

  • Roadmap Overview: The roadmap shifts LAM from defect detection toward proactive defect prediction and elimination before defects emerge.It emphasizes anticipating anomalies from multimodal sensor data and adjusting process parameters preventively.
  • Roadmap Overview: Together, these stages outline a path toward fully autonomous, self-adaptive LAM with continuous unsupervised operation and improved product quality.The roadmap combines sensor fusion, preventive process adjustment, and hierarchical correction into a structured research direction.
  • Multisensor Monitoring And Data Fusion Framework: Multisensor data fusion combines complementary sensor strengths to improve process understanding and predictive-model reliability.The framework is intended to reduce sensor redundancies while retaining comprehensive monitoring.
  • Proactive Defect Prediction And Mitigation Models: Predictive ML and deep-learning models would use historical and real-time multimodal data, including spatio-temporal physics knowledge, to trigger automatic parameter changes.Examples include modifying laser power and speed before anticipated defect occurrence.
  • Multi-scale Hierarchical Defect Identification And Rectification Models: Defect identification should operate hierarchically across micro- to macro-scale defects and real-time, layer-wise, and region-based monitoring levels.Different sensors or sensor combinations would specialize in defect types at their respective operational scales.
  • Multi-scale Hierarchical Defect Identification And Rectification Models: When defects occur, autonomous rectification such as robotic machining should replace simple process halting where feasible.The proposed approach targets continuous operation and removal of detected defects within the process chain.

Appendix A. Surveys on Machine Learning-Assisted Defect Detection in Laser Additive Manufacturing

The appendix surveys ML-assisted vision, acoustic, and multisensor methods for defect, quality, process-state, and geometry prediction in LAM. Reported performance is accompanied by recurring limits involving generalizability, spatial or temporal resolution, and restricted materials or geometries.

  • Vision-Based Methods: The appendix surveys ML-assisted vision-based defect and fault detection methods in LAM.Table A1 frames the surveyed methods and their reported applications and limitations.
  • Vision-Based Methods: Vision methods frequently remain limited by single-track or thin-wall studies, restricted materials and settings, and missing location-specific information.These constraints recur across melt-pool, porosity, surface, and layer-wise inspection studies.
  • Vision-Based Methods: Vision-based studies report outcomes including Accuracy: 0.997 for melting-state classification and Accuracy: 98.37% with Latency: 2.9 ms for online monitoring.Other reported results include F1: 86.6% and Accuracy: 87.3% for anomaly classification.
  • Acoustic-Based Methods: Acoustic methods include process-regime, keyhole-pore, anomaly, quality, and melting-state classification using CNN, SVM, VAE/GAN, and DBN models.Reported results include Accuracy: 97% for keyhole-pore detection and Accuracy: 96% - 97% for VAE and GAN anomaly detection.
  • Acoustic And Multisensor Methods: Acoustic and multisensor studies still report limited spatial or temporal resolution, restricted defect coverage, and limited generalization across materials, geometries, or settings.Several studies are single-track or thin-wall investigations and are not location-specific.
  • Multisensor Fusion: Multisensor surveys combine cameras, infrared and visible sensors, microphones, acoustic emission, photodiodes, and powder-bed imaging for broader process characterization.Table A3 identifies multisensor fusion as a distinct survey category.

Declaration of competing interest

The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.

  • The authors declare no known competing financial interests or personal relationships influencing the reported work.
Loading 2404.13673v1…