I. Core Breakthroughs

Breakthrough 1 : Innovative preparation Transformer Spatial perception model solves AFP The challenge of long-range spatial dependency prediction

Traditional LSTM Serial processing of placement sequence data tends to result in the loss of long-range tow‑tow interaction information, making it impossible to fully capture the gap fluctuations arising from tow‑to‑tow coupling across the entire mold domain. Transformer The multi-head self-attention mechanism processes the entire segment’s layup‑space sequence in parallel, autonomously capturing cross‑regional dependencies between the current filament and multiple upstream and downstream tow bundles. It accurately reproduces the complex gap‑evolution patterns on curved surfaces and along long layup paths, achieving prediction accuracy that significantly outperforms conventional time‑series deep‑learning baseline models.

Breakthrough 2 Customized spatial position encoding to distinguish between global mold‑region effects and local tow‑coupling influences.

Targeting AFP Develop a domain-specific positional encoding for 2D layout space, integrating it with the robot. X 、 And Global coordinates and the adjacent filament are relative. And The offset feature enables the model to simultaneously detect two types of defect sources—intrinsic positioning errors in specific regions of the mold and localized slippage of adjacent filaments—thereby addressing the limitations of conventional temporal models, which lack spatial awareness and cannot distinguish between these two kinds of variations.

Breakthrough 3 : Prediction +  Mechanism-based analysis of dual-capability integration, breaking… AI Barriers to the Implementation of Black-Box Engineering

Currently, deep learning approaches for composite material defect detection only output predicted values and cannot pinpoint the root causes of defects; this study introduces SHAP Explainable analytics quantifies the contribution of each sensor feature, clearly distinguishing among three categories of variables—robotic motion error, tow‑material slippage, and compaction width/height—in terms of their hierarchical influence on gap dimensions. The resulting predictions can be directly applied to optimize process parameters, thereby meeting the stringent requirements of aerospace composite certification for process traceability and mechanistic interpretability.

Breakthrough 4 : The industrial open dataset has undergone standardized validation, and the method’s generality is reproducible.

Relying on TU Delft Open source AFP The multi-sensor industrial dataset has undergone end-to-end training and testing, with rigorous time-series–based stratification for training. /  Verification /  The test set eliminates data leakage caused by random splits, and the experimental workflow is standardized, enabling universities, equipment manufacturers, and composite‑material producers to reproduce the model‑training pipeline, thereby facilitating rapid deployment and portability across the industry.

II. Technological Innovation

Innovation 1 Facing AFP of the Transformer Custom Network Architecture

• Feature Embedding Layer: Unified Mapping of Multi-Source Sensing Data 64 The feature space is normalized to unify the scales and enable attention computation.
• Spatial Position Encoding Module: Fusing Absolute X/Y Coordinates, relative offset of adjacent filaments rel_Y Three-dimensional spatial information endows the model with full‑domain spatial perception capabilities.
• Three-layer encoder + 4 Head‑to‑head self‑attention: concurrently captures multi‑scale coupling relationships between local neighboring tokens and long‑range positional patterns, while residual layers and normalization help mitigate overfitting in deep networks.
• A single-layer linear regression output layer directly produces millimeter-level gap predictions, enabling lightweight inference to ensure real-time online computation.

Innovation 2 Multi-source Industrial Sensor Data Standardization and Preprocessing Pipeline

• Temporal Synchronization: With Robots X Coordinate alignment of multi-channel asynchronous data from lasers, cameras, and trackers enables one-to-one correspondence of features across the same layup cross-section.
• Feature Engineering: Split the measurement data into two categories—current thread and previous thread—and add a relative feature. Y Offset feature, which distinguishes between machine positioning error and tow slippage;
• Time-series dataset partitioning: split according to the sequence of tape placement (previous 22 Dao training, intermediate 4  Verification of the Dao, finally 5  (Production line testing), fully simulating the real production line. “ Place first, predict later ”  Operating conditions, preventing future information leaks;
• Z-score Standardization eliminates sensor range discrepancies and facilitates network convergence.

Innovation 3 SHAP Explainable AI Defect Root Cause Analysis System

• Global Feature Weight Ranking: A Quantitative Proof ** Lateral slip on the roller ( camera Deviation) **  It is the top impact factor in the field of gaps, followed by robotics. And To positioning deviation, post-compaction filament width, and original tow width;
• Single-feature dependency curves: elucidating the physical laws governing the positive and negative correlations between each variable and the gap. —— Increased roller slip directly widens the gap; after compaction, the wider the tow, the smaller the gap becomes.
• Feature interaction visualization reveals a coupled amplification effect between tow width and lateral slip: when slip is excessive, fluctuations in tow width further exacerbate gap‑exceedance, providing a quantitative basis for integrated process optimization.

Innovation 4 Millisecond‑level online inference for real‑time quality control on the production line

Model training employs GPU Accelerated, with single‑pass inference latency at the millisecond level, and compatible with existing systems. AFP An online laser inspection system provides real-time, in‑process gap predictions, eliminating the need for post‑layup offline scanning and enabling proactive defect detection, thereby significantly reducing rework and scrap costs.

III. Application Scenarios

• Aviation AFP Automated Online Proactive Quality Control for Production Lines (Core Scenario)

Mass production of large composite preforms for fuselages, wing boxes, inlet ducts, and other components, coupled with laser line‑scanning for online multi‑source data acquisition, enables real‑time prediction of gap dimensions. Areas exceeding tolerance thresholds trigger immediate alerts, allowing proactive adjustments to layup speed, compaction pressure, and tow tension—eliminating the need for costly, large‑scale offline rework. C Scanning and detection.

• AFP Development of Closed-Loop Adaptive Path Control for Equipment

Integrate the prediction model into the robot’s control system, feed real-time gap values back to the laying‑path algorithm, and perform automatic fine‑tuning. Y By adopting a placement coordinate system, the process achieves adaptive gap correction, thereby reducing the time and cost associated with manual iterative debugging.

• Digital Iterative Optimization of Composite Manufacturing Process Parameters

Collect sensor data in batches under varying compaction forces, infrared temperatures, and tow tensions, and combine it with… SHAP Analyze and pinpoint the root cause of critical variations, rapidly converge on the optimal process window, and significantly reduce the number of trial‑run test parts.

• Input Correction for Composite Material Simulation Digital Twin

to put AI The predicted global gap distribution is incorporated into finite element simulations, replacing the idealized defect-free case. CAD The model accurately captures the stress concentrations and stiffness reductions caused by gaps, thereby enhancing the credibility of structural strength simulations.

• Universities and colleges / Research institute AI Benchmark Study of Composite Algorithms

Fully reproducible based on publicly available industrial datasets. Transformer Training process, developed as a comparative baseline. CNN 、 LSTM , novel defect‑prediction models such as graph neural networks, and a unified benchmark for algorithmic comparison across the industry.

IV. Summary of the Article

(1) Core Research Achievements

This paper successfully attempts to incorporate spatial positional encoding into Transformer Attention models are used for AFP Gap prediction for paving, multi-sensor fusion model test set R²=0.9143 、 MAPE Only 1.04% , millimeter-level prediction accuracy meets the tolerance control requirements for aerospace composite materials; relying on SHAP Explainable AI Quantify the mechanisms by which various process variables influence gap dimensions, and address deep learning. “ The black box cannot be implemented. ”  Industry pain points. End-to-end data processing. +  Model Training +  The mechanistic analysis is fully standardized across the entire workflow and can be completely reproduced using publicly available industrial datasets, providing… AFP Digital monitoring for intelligent manufacturing offers a brand-new… AI Tool.

(II) Core Technological Advantages

• Leading spatial modeling capabilities: Multi-head attention captures long-range couplings across the entire fiber bundle, outperforming… LSTM Time-series model;
• Engineering explainability: SHAP Quantify defect root causes to meet the traceability requirements of aerospace component certification.
• Real-time line adaptation: Lightweight inference enables simultaneous online fault detection and alerting.
• General-purpose and portable: a standardized data processing pipeline, compatible with various platforms. KUKA 、 ABB Mainstream AFP Robots and laser inspection equipment.

(3) Current Research Limitations

• Only the unidirectional straight‑layup flat‑plate condition was verified; variable‑angle steering and complex layups using hyperbolic molds were not evaluated.
• Only synchronization gap estimation has been implemented; the multi-step ahead prediction feature has not yet been developed.
• The closed-loop control system for the robot has not yet been fully integrated—both hardware and software—and remains at the prediction and early-warning stage only.
• The dataset is limited in scope and does not encompass multiple categories of tow materials, such as glass fiber and thermoplastic prepregs.

(4) Future Expansion Directions

• Expand the dataset of variable-curvature, variable-angle layup configurations to enhance the generalization capability of predictions for complex composite components.
• Develop a multi-step ahead gap prediction module to enable proactive path correction.
• Integrating multi-dimensional sensing of compaction roller pressure and infrared temperature, a predictive model for all process variables is constructed.
• Model-embedded robot PLC Control system, setup “ Prediction -  Automatic hyperparameter tuning ”  Closed-loop adaptive placement system;
• Extend simultaneous multi-output prediction to multiple defect types, including overlap, wrinkling, and tow shift.

Image and Text Guide

Figure 1. Detailed methodological workflow.
 

Figure 2. AFP Sensor arrangement on the head. The stacking direction is along the positive axis. X Axial direction [9]

Table 1. Description of the variables in the dataset.

Table 2. Transformer model: Performance specifications.

Figure 3. Transformer model: predicts the relationship between the predicted gap width and the actual gap width.

Figure 4. Transformer model: 50 The predicted confidence interval width of a sample point versus the actual confidence interval width.

Figure 5. Transformer Model: Prediction Error Distribution

Figure 6. Transformer model: the relationship between prediction error and predicted values.

Figure 7. Distribution of the global feature importance (average impact on the output) for predicting gap width.

Figure 8. Graph of the shape dependence of the gap width prediction.

Figure 9. Shap Summary plot of predicted gap widths