I. Core Breakthroughs

Breakthrough 1 : The world’s first end-to-end open-source system AFP Online Detection Software-Hardware Integrated Framework

Existing AFP All online inspection devices employ proprietary, closed-source systems from the manufacturers, with opaque algorithms and data-processing logic, making it impossible for researchers to reproduce or benchmark various defect-detection algorithms. This study fully discloses the hardware communication protocol, Python Processing source code and complete experimental raw datasets, the system ensures end-to-end transparency—from raw laser profiles to finished digital twins—while establishing a universally applicable open-source benchmark platform, thereby significantly lowering the R&D barriers for universities and enterprises engaged in online monitoring.

Breakthrough 2 : Achieve real-time, lane-by-lane synchronous monitoring during the layup process, completely eliminating wasted man-hours associated with offline inspection.

Traditional layered offline inspection requires shutdowns and disassembly of tooling, consuming substantial production time. This system integrates a line laser into the layup head, enabling the robot to perform scanning in sync with each layer deposition and extract tow‑edge data in real time. Defects can thus be identified before the next layer is laid, allowing for on‑line adjustments to process parameters, directly reducing downtime for quality inspections and boosting line utilization.

Breakthrough 3 : Constructing Temporal Synchronization + A comprehensive algorithmic framework for multidimensional geometric quantification

Overcome the challenge of temporal misalignment between robotic motion data and high-frequency laser data, and develop innovative adaptive binarization and contextual filament-edge labeling algorithms to distinguish the first layer. / Subsequent layer‑placement recognition logic; meanwhile, lateral path deviation was successfully addressed ( LPD ), a dual-fidelity evaluation metric for longitudinal surface waviness, separates the positioning error of the inspection robot from the geometric distortion caused by interlayer unevenness, thereby accurately distinguishing the two types of defect origins.

Breakthrough 4 : Interactive 3D Measured Digital Twin and Virtual Sectioning Inspection Tool

Unlike traditional 2D defect drawings, the framework generates a complete, measured 3D point cloud model, supporting layered and lane‑specific color‑coded visualization and virtual sectioning at any location. It enables precise measurement of local filament width, gaps, and layer thickness without damaging the component, thereby achieving nondestructive, full‑field, high‑resolution geometric quality inspection.

II. Technological Innovation

Innovation 1 Robot - Laser Synchronous Acquisition Hardware Communication Architecture

• Hardware configuration: KUKA Equipped with a six-axis robot Micro-Epsilon Line laser, 100Hz High-frequency acquisition, lateral resolution 12.9 μm , enabling sub-millimeter defect detection;
• Open-source communication protocol: based on py_openshowvar Implement the robot controller TCP Real-time acquisition of pose and motion parameters; laser communication via Ethernet. SDK Synchronous sampling with millisecond‑level timestamp alignment of two data streams;
• Sequential trigger-based data acquisition: relies on the robot’s digital signals. IO Signal-controlled laser on/off operation collects data only during the paving phase, reducing redundant storage and archiving data in a structured format organized by layer and lane.

Innovation 2 Five-stage fully automatic Python Data Processing Pipeline

• Temporal Synchronization: Using timestamps to retrospectively align laser profiles with their corresponding robot positions. TCP Pose and temporal error control 1–2ms , maximum spatial offset 0.1mm；
• Signal preprocessing: Savitzky–Golay Filtering for noise reduction, linear interpolation to fill in laser‑induced frame gaps, while preserving the steep edges of the filament.
• Adaptive feature extraction: does not use a fixed height threshold, instead employing a profile-based approach. 10%/90% Percentile-based dynamic partitioning distinguishes three types of edges—current filament, adjacent filament, and gap—and employs two separate recognition logics for single-layer and multi-layer placement. IQR Statistically filter out abnormal measurement values;
• 3D Coordinate transformation: One-time calibration of the hand–eye transformation matrix, for serial robots’ base. -  Tool -  Laser-based three coordinate systems enable global 3D reconstruction from 2D laser profiles.

Path fidelity quantification:

• Lateral Path Deviation LPD : Calculate the offset of the tow’s centerline relative to its average trajectory, and quantify the robot’s left–right oscillation during placement.
• Longitudinal profile waviness: characterizes the surface resulting from orthogonal ply orientation. “ Pebble ”  Ripple, quantizing inter-layer convex–concave distortions.

Innovation 3 Multimodal Visual Interactive Detection Tool

• Layered color‑coded view: Quickly identify overall layer misalignment and large‑area placement offsets.
• Single-channel independent view: Analyzes the width and narrowness of the tow and the uniformity of the gaps on a per‑channel basis.
• Local height heatmap: adaptively normalized across cross-sections, enhancing the visualization of subtle wrinkles and surface irregularities.
• Virtual sectioning function: Extract 2D cross-sections at any location, accurately measure filament width, interlayer spacing, and single-layer thickness—no physical sectioning required.

Innovation 4 Standardized benchmark datasets are open-source and shared.

Adopting the standard [0/90] The orthogonal layering calibration test has been completed, and the raw laser profile, robot pose, processed point cloud, and quantitative quality‑inspection report have been fully uploaded to. Zenodo ; all Python Source Code Hosting GitHub It includes a complete set of scripts for data collection, synchronization, reconstruction, and visualization, enabling researchers worldwide to directly reproduce the experiments and compare their own defect‑identification algorithms.

III. Application Scenarios

• AFP Online Quality Control for Automated Composite Manufacturing Lines (Core Use Case)

Aircraft fuselage, wing box, wind turbine blades AFP Real-time quality inspection on the production line detects gaps, overlaps, and tow misalignment during layup, providing early warnings and enabling online path adjustments, thereby significantly reducing scrap rates and eliminating the need for layer-by-layer offline scanning.

• Universities and colleges / Benchmark Platform for Composite Material Processing Research at Scientific Research Institutes

Researchers can directly download the open-source code and benchmark datasets to develop deep-learning-based defect detection, AI Process optimization algorithms establish a unified benchmark, eliminating data biases introduced by varying equipment and testing protocols, thereby enhancing the reproducibility of experimental results in academic papers.

• Robot Path Planning Simulation and Digital Twin Calibration

Import the measured three-dimensional ply‑by‑ply model. CAE Simulation incorporates real-world gaps and tow waviness into the mechanical analysis, thereby replacing the idealized, defect-free model. CAD The model accurately simulates the strength reduction caused by manufacturing defects, thereby enhancing the credibility of structural simulations.

• AFP In-house equipment development and modification of the placement head

Equipment manufacturers can leverage open-source communication and data‑acquisition frameworks to develop low‑cost, in‑house online monitoring modules, eliminating the need to procure expensive, proprietary commercial inspection systems and thereby shortening the development cycle for quality‑inspection capabilities of new layup heads and multi‑axis robots.

• Iterative Optimization Experiment of Composite Manufacturing Process Parameters

Geometric data are collected in bulk under varying fiber placement speeds, compaction pressures, and tow tensions, enabling quantitative characterization of the relationships between process parameters and interlaminar gaps, misalignments, and waviness, thereby facilitating rapid optimization and stabilization of the fiber placement process window.

IV. Summary of the Article

(1) Core Values

This article reports an industry‑first innovation: fully open source and integrated hardware–software. AFP An online geometric quality inspection framework that seamlessly integrates laser data acquisition. -  Temporal Synchronization -  3D Reconstruction -  Quantitative Analysis -  Interactive visualization of the complete technical pipeline, accompanied by a reproducible standard ply‑by‑ply benchmark dataset. System measurement error is controlled within 0.168mm Within this framework, it can simultaneously detect all critical manufacturing defects—such as fiber width, inter‑ply gaps, lateral offsets, and interlayer waviness—thereby addressing two major industry pain points: the closed nature of commercial inspection systems and the low efficiency of offline inspection. This provides a universal, open‑source platform for online monitoring in composite‑material smart manufacturing.

(II) Summary of Technical Advantages

• Fully transparent and open-source: No proprietary black boxes; both the code and datasets are completely open, enabling free secondary development for both research and industry.
• Synchronous online inspection: scan immediately upon placement, enabling proactive defect detection and significantly reducing rework and downtime.
• High-precision 3D reconstruction: unified hand–eye coordinate transformation, generating digital twins of real-world components and enabling non-destructive virtual sectioning and measurement.
• Dual‑dimensional fidelity metrics: separate robot positioning errors from interlayer ripple defects, enabling precise identification of the root cause.
• Lightweight Modularization: Python Modular script design enables rapid adaptation. KUKA Mainstream, etc. AFP Robots and various line laser devices.

(III) Current Limitations and Future Directions for Expansion

Existing limitations

Currently, edge‑detection algorithms are only compatible with simple, adjacent single‑filament layups and do not yet support complex curved surfaces involving filament loss or angular‑change transitions. They rely solely on a single‑line laser and lack multi‑sensor fusion, nor do they incorporate real‑time closed‑loop path correction.

Future Expansion Directions

Optimized the feature recognition algorithm to handle complex operating conditions, including variable‑angle steering, wire breakage, and multi‑wire overlap.

• Integrating infrared and force-sensing multi-channel data to construct a multimodal synchronous monitoring system;
• Develop a real-time closed-loop control module that detects gaps exceeding the threshold. / Automatically corrects the robot’s laying path after offset.
• Coupled with machine learning algorithms, it automatically classifies multiple defect types—such as wrinkles, overlaps, and variations in width—and provides tiered early warnings.
• Expand to thermoplastics AFP , Dry fiber placement ADFP and various composite material forming equipment.

Image and Text Guide

Figure 1.  System architecture and data flow of the data collection and analysis framework.

Figure 2.  Flowchart of the multi-stage feature extraction algorithm. The process comprises initial signal preprocessing, context labeling based on channel numbers, and a final validation stage, ensuring that the detected drag points and gap coordinates are robust.

Figure 3.  Visualization diagram of representative laser profilometry feature extraction algorithms. (a) Signal processing: The originally restored contour is first processed using Savitzky–Golay filtering. - Golay filtering is applied to suppress high-frequency sensor noise while preserving the sharpness of the towed‑array edges. (b) Adaptive binarization: Based on the contour’s own nth… 10 Percentiles and the nth 90 Dynamic thresholds are calculated based on percentile distributions. (c) Final feature labeling: The state-awareness logic analyzes the binarized signal, identifies, and labels the coordinates of the final features: in this example ( Pass > 1 ) In the adjacent drag line, “Current Left” / “Right Drag Edge” and “Gap Edge.”

Figure 4.  A demonstration of the algorithm’s state-aware contextual logic.

Figure 5.  Coordinate systems and transformation chains. Left: Robot base and flange frame; Middle: Flange, TCP The relationship between the point and the sensor; right: BTT With and Text-to-Speech The linking method used to map sensor points to the world coordinate system.

Table 1  Metrological Specifications and Performance of Integrated Inspection Frameworks

Figure 6.  A data processing flowchart that converts 2D sensor frame features into 3D world‑coordinate system data by leveraging synchronized robot pose data and a static hand–eye calibration matrix.

Figure 7.  Laser line scanner and AFP Integration of the end effector.

Figure 8.  Multi-layered [0/90] Schematic diagram of top-down 2D reconstruction of a stacked structure. The interfacial regions between adjacent fiber bundles are color-mapped according to the measured gap widths.

Table 2 [0/90] Summary of quantitative quality control metrics for the layup motion. All values are expressed in millimeters.

Figure 9.  Multi-layered [0/90] Top-view photograph of a composite laminated structure

Table 3 [0/90] Comprehensive quantitative metrics for the layup motion and validation results. Frame‑based measurements are compared with vision‑based reference values. All values are expressed in millimeters.

Figure 10.  Schematic diagram of the virtual inspection workflow, illustrating the …th 1 The analytical process for a single cross-section of a layer. (a) The three-dimensional geometric digital superposition model of this layer, rendered using in‑slice height‑coloring techniques, shows the position of the virtual slicer (red box). (b) The blend mode diagram of this slice; (c) An illustration of the entity model for this slice. While these static panels showcase the tool’s core functionalities, its true diagnostic advantage lies in the ability to interactively navigate slicers within the three-dimensional model. We provide the complete dataset and source code to enable users to engage in direct, hands-on exploration (see the Data Availability Statement).

Figure 11.  Each layer's first 3 Lateral path deviation during the second pass ( LPD A comparative analysis of ) was conducted, with all data plotted using a unified Y-axis scale to facilitate an intuitive comparison of error magnitudes.

Figure 12.  Each layer's first 3 Vertical cross-sectional comparative analysis following the completion of the passage. A unified Y-axis scale enables direct comparison of topographic features.