Defect detection method for optical fiber connector based on deep learning

Abstract

The invention discloses an optical fiber connector defect detection method and system based on deep learning, and belongs to the field of optical communication equipment quality detection. Aiming at the problems of insufficient feature extraction, insufficient utilization of multi-modal data, poor environmental adaptability and the like of a traditional detection method, detection is realized through a multi-modal data fusion-feature enhancement-dynamic calibration closed-loop architecture. An improved Retinex algorithm and the like are adopted for preprocessing data, features are fused through a cross-modal attention mechanism and the like, features are enhanced through YOLOv8, GAN and the like, and a threshold value is dynamically calibrated in combination with Bayesian optimization and the like. Various defects can be recognized, the detection accuracy rate is larger than or equal to 99.5%, the false alarm rate is smaller than or equal to 0.1%, full-process intelligentization of detection is achieved, and the method is suitable for high-precision defect recognition.

Description

Technical Field

[0001] This invention relates to the field of optical communication equipment quality inspection technology, specifically to an intelligent detection method and system for fiber optic connector end face defects through multimodal data fusion, adaptive feature enhancement, and dynamic threshold calibration, which is suitable for high-precision identification of defects such as scratches, contamination, and geometric deformation. Background Technology

[0002] High-precision detection of defects on the end face of fiber optic connectors is crucial to the performance of optical communication systems. Existing detection methods based on traditional image processing suffer from the following technical bottlenecks:

[0003] Insufficient feature extraction capability: Relying on manually designed features (such as edge gradients and gray-level co-occurrence matrices) makes it difficult to capture the complex features of micron-level defects (such as the light scattering characteristics of subsurface cracks), resulting in a false negative rate as high as 10%-15%.

[0004] Insufficient utilization of multimodal data: Optical images, three-dimensional contours, and spectral data are not deeply integrated, making it impossible to form a complete defect feature vector, which affects the identification accuracy of complex defects (such as stress-induced hidden cracks).

[0005] Poor environmental adaptability: The fixed threshold detection model cannot adapt to the manufacturing differences of different batches of connectors, requiring frequent manual parameter adjustment and low degree of automation.

[0006] Existing technologies lack detection schemes that deeply integrate multimodal deep learning, adaptive feature enhancement, and dynamic threshold calibration. There is an urgent need for an intelligent detection method that can automatically extract defect features and adaptively optimize detection parameters. Summary of the Invention

[0007] This invention provides a method and system for detecting defects in fiber optic connectors based on multimodal deep learning. Through a closed-loop architecture of multimodal data fusion, feature enhancement, and dynamic calibration, it solves the problems of insufficient feature extraction, fragmented multimodal data, and poor environmental adaptability in traditional detection, achieving intelligent detection throughout the entire process with a detection accuracy of ≥99.5% and a false alarm rate of ≤0.1%.

[0008] Multimodal data fusion algorithm:

[0009] Data preprocessing module: Optical image enhancement: An improved Retinex algorithm is used to suppress the effects of uneven illumination, and adaptive histogram equalization (CLAHE) is used to improve the contrast of defect areas. I ′ (x,y)=log(I(x,y))-log(F(x,y) * I(x,y)) Where F(x,y) is the Gaussian blur kernel, and I(x,y) is the pixel value of the original image.

[0010] 3D point cloud reconstruction: Based on the Delaunay triangulation algorithm, spectral confocal displacement sensor data is reconstructed into an end-face curved surface model, and the radius of curvature R is calculated. i and surface roughness Ra:

[0011] Multimodal Feature Fusion Network: Cross-modal Attention Mechanism: A dual-channel attention module is designed to focus on the optical image feature map F. img and 3D point cloud feature map F 3D Perform feature weighting based on channel and spatial dimensions: Where σ is the Sigmoid activation function, δ is the ReLU activation function, and W1 and W2 are learnable weight matrices.

[0012] Feature fusion strategy: Employing spatiotemporal alignment technology to fuse optical, 3D, and spectral features into a unified feature vector V = [V img V 3D V spec Input defect identification model.

[0013] Adaptive Feature Enhancement Algorithm: Dynamic Feature Enhancement Module: Defect Sensitive Area Localization: Based on the YOLOv8 target detection model, potential defect areas are located, ROI masks M are generated, and features within the ROI are enhanced a second time.

[0014] Multi-scale feature pyramid: Construct a feature pyramid containing shallow texture features (3×3 convolution), mid-level geometric features (5×5 convolution), and deep semantic features (7×7 convolution), and achieve multi-scale feature fusion through cross-layer skip connections.

[0015] Adversarial Feature Optimization: Generative Adversarial Networks (GANs): Introducing a defect generator G and a discriminator D, and enhancing the model's sensitivity to minor defects through adversarial training. min G max D E x～pdata [logD(x)]+E z～pz [[log(1-D(G(z)))] Where, p data For the actual defect data distribution, p z This represents the noise distribution.

[0016] Dynamic threshold calibration algorithm: Online learning mechanism: Real-time sample acquisition: The system automatically acquires images of 10 standard parts for every 100 products inspected and updates the dynamic threshold database.

[0017] Adaptive threshold adjustment: Based on a Bayesian optimization algorithm, the defect confidence threshold θ is dynamically adjusted to ensure consistency in the testing of different batches of products. Where α is the learning rate and L(θ) is the loss function.

[0018] Defect confidence assessment: Uncertainty quantification: Monte Carlo dropout is used to estimate the prediction uncertainty. Secondary detection is triggered for samples with confidence below 95%. The false detection rate is reduced through a multi-model voting mechanism. Attached Figure Description

[0019] Figure 1 This refers to the process of defect feature classification and model training. Detailed Implementation

[0020] The fiber optic connector end-face defect detection method and system of this invention achieves automation and high precision in the detection process through multi-modal data fusion and intelligent algorithm collaboration. The specific implementation steps are as follows:

[0021] Multimodal data acquisition and preprocessing: Multi-source data acquisition. This step uses a hardware system to acquire multi-dimensional feature data of the fiber optic connector end face, providing a foundation for subsequent defect analysis.

[0022] Laser confocal microscopy (resolution 2048×2048 pixels) was used to acquire optical microscopic images of the end face to capture micron-level surface defect textures;

[0023] Three-dimensional point cloud data were acquired using a spectral confocal displacement sensor (point spacing ≤ 1 μm) to construct an end face curved surface contour model;

[0024] Spectral characteristic data are acquired by using a multi-band ring light source (switched to ultraviolet light mode) to enhance the optical contrast of defects such as contamination and oxidation.

[0025] Data preprocessing: To improve data quality and algorithm processing efficiency, multi-source data undergoes standardization processing.

[0026] Optical image enhancement: The Retinex algorithm is used to eliminate the effects of uneven illumination. Combined with adaptive histogram equalization (CLAHE) technology, the contrast between the defect area and the background is significantly improved, which facilitates subsequent feature extraction.

[0027] 3D point cloud reconstruction: Based on the Delaunay triangulation algorithm, point cloud data is reconstructed into an end-face surface model, and the radius of curvature R is calculated simultaneously. i and surface roughness R a Quantify geometric deformation characteristics;

[0028] Dimensionality reduction of spectral data: Principal component analysis (PCA) is used to reduce the dimensionality of spectral data, compressing the original feature dimensions to 12 dimensions, thus reducing computational complexity while retaining the main information.

[0029] Feature Fusion and Enhancement: Cross-modal feature deep fusion is achieved through a dual-channel attention module to realize optical image feature F. img With 3D point cloud features F 3D Weighted fusion: First, channel-level and spatial-level cross-attention calculations are performed on the two types of features to generate a feature weight matrix; based on the formula...

[0030] The original features are weighted to highlight defect-sensitive areas, forming a unified feature vector V that includes texture, geometry, and spectral information.

[0031] Dynamic enhancement of defect features

[0032] ROI Region Localization: The YOLOv8 target detection model is used to automatically identify potential defect areas and generate a region of interest (ROI) mask M to focus on the detection target and reduce background interference.

[0033] Multi-scale feature pyramid: Construct a pyramid structure containing shallow texture features (3×3 convolution), mid-level geometric features (5×5 convolution), and deep semantic features (7×7 convolution), and achieve feature fusion of defects of different scales (0.5μm-500μm) through cross-layer skip connections;

[0034] Adversarial sample generation: Introducing Generative Adversarial Networks (GANs) and using adversarial training formulas. min G max D E x～pdata [logD(x)]+E z～pz [[log(1-D(G(z)))]

[0035] Generate simulated defect samples to enhance the model's sensitivity to and generalization ability for minor defects.

[0036] Intelligent Defect Identification and Dynamic Calibration: Defect classification and localization will input the fused feature vector V. The improved EfficientNet-B5 deep learning model is implemented through a multi-task output layer:

[0037] Defect Classification: Accurately identifies 5 types of defects, including scratches, contamination, chipping, dents, and oxidation, with a classification accuracy rate of ≥99.5%.

[0038] Geometric measurement: Simultaneous calculation of end face radius of curvature (accuracy ±2μm), tilt angle (accuracy) Parameters such as ±0.1° and contamination area (pixel-level segmentation) are used to quantify the severity of defects.

[0039] Dynamic calibration of detection threshold: using Bayesian optimization algorithm (formula: The defect confidence threshold θ is adjusted in real time to adapt to manufacturing differences between different batches of products; Monte Carlo dropout technology is introduced to assess prediction uncertainty, and secondary detection is triggered for samples with confidence levels below 95%; the false detection rate is further reduced through a multi-model voting mechanism (false alarm rate ≤ 0.1%).

[0040] The inspection results are output and traceable. The inspection results are uploaded to the factory manufacturing execution system (MES) in real time via the OPC UA protocol. The results include information such as defect type, location, geometric parameters and confidence level. The system automatically generates a quality file for each product and supports historical data traceability and statistical process control (SPC) analysis.

[0041] Model training methods: Sample augmentation and dataset construction. To improve the model's generalization ability, a multi-dimensional data augmentation strategy is used to expand the training set.

[0042] The original samples were rotated (±15°), scaled (0.8-1.2 times), and subjected to additive Gaussian noise (σ=0.05) to expand the sample size from 50,000 to 200,000.

[0043] The annotations include defect categories, pixel-level masks, and geometric parameters, constructing an annotation dataset that covers a variety of defect types and severity levels.

[0044] Loss function and optimization strategy: To address the class imbalance problem, a combined loss function is designed. Here, α = 0.25 and γ = 2 are balancing parameters, and A and B are the predicted defect region and the actual defect region, respectively. Focal Loss is used to increase the weight of hard samples, and Dice loss is combined to enhance pixel-level segmentation accuracy, ensuring the model's ability to detect minor defects (such as microcracks).

[0045] The training process employs a transfer learning strategy, initializing the backbone network based on the ImageNet pre-trained model and using the Adam optimizer for end-to-end training. The training cycle is controlled within 48 hours, and the final performance index of detection accuracy ≥99.5% is achieved on the validation set.

Claims

1. A method for detecting defects on the end face of an optical fiber connector, characterized in that, Includes the following steps: Multimodal data preprocessing: Retinex enhancement and CLAHE processing are performed on optical images, surface reconstruction is performed on 3D point cloud data, and PCA dimensionality reduction is performed on spectral data; Multimodal feature fusion: A weighted fusion of optical, 3D, and spectral features is achieved through a dual-channel attention module to generate a unified feature vector; dynamic Feature enhancement: Based on YOLOv8, potential defect regions are located, and defect-sensitive features are enhanced through multi-scale feature pyramids and generative adversarial networks; Dynamic threshold calibration: The defect confidence threshold is dynamically adjusted using a Bayesian optimization algorithm to trigger secondary detection for uncertain samples.

2. The detection method according to claim 1, characterized in that, The dual-channel attention module uses a cross-modal feature weighting formula. F att =σ(W2δ(W1(F img ⊕F 3D ))) Achieve deep integration of optics and three-dimensional features.

3. The detection method according to claim 1, characterized in that, The dynamic feature enhancement module adopts a multi-scale feature pyramid structure, which includes cross-layer fusion of shallow texture features, mid-level geometric features, and deep semantic features.

4. The detection method according to claim 1, characterized in that, The generative adversarial network is trained using an adversarial training formula. min G max D E x～pdata [logD(x)]+E z～pz [log(1-D(G(z)))] Optimize the model's sensitivity to minute defects.

5. The detection method according to claim 1, characterized in that, The dynamic threshold calibration employs a Bayesian optimization algorithm, using the formula... Achieve adaptive threshold adjustment.

Images