Optical fiber fusion quality detection and optimization method, device, equipment and storage medium

Abstract

The present application relates to the technical field of optical fiber fusion, and particularly relates to an optical fiber fusion quality detection and optimization method, device, equipment and storage medium, the method first collects a multi-dimensional fusion point data set, carries out data preprocessing on the multi-dimensional fusion point data set, obtains a to-be-tested fusion point data set, then carries out multi-dimensional fusion feature extraction on the to-be-tested fusion point data set, obtains a late fusion feature set, then calls a pre-trained target quality detection model to carry out multi-dimensional quality evaluation on the late fusion feature set, obtains a quality evaluation result, finally optimizes parameters of the to-be-tested fusion point data set based on the quality evaluation result, obtains an optimized fusion point data set, aims to realize accurate quantitative detection of optical fiber fusion quality and intelligent optimization of process parameters, effectively improves detection precision and reduces rework rate, and guarantees optical fiber network transmission performance and long-term reliability.

Description

Technical Field

[0001] This invention relates to the field of optical fiber fusion splicing technology, and in particular to a method, apparatus, equipment, and storage medium for optical fiber fusion splicing quality testing and optimization. Background Technology

[0002] Fiber optic fusion splicing is a critical process in the construction and maintenance of fiber optic communication networks, and its quality directly affects the transmission performance of optical signals and the reliability of network operation. Currently, fiber optic fusion splice quality inspection in the industry mainly relies on two methods: the loss estimation function built into the fusion splicer and manual microscopic observation. However, the overall technical solution has many prominent problems that urgently need to be addressed.

[0003] Traditional fusion splicers rely on two-dimensional image matching to estimate splice loss, neglecting key factors such as the three-dimensional morphology of the splice point and material properties. This results in low estimation accuracy and significant deviations from actual loss, with errors becoming more pronounced in special optical fibers or harsh environments. Quality assessment depends on operator experience, lacking quantitative standards for qualitative observation, leading to significant differences in evaluation among different personnel. Furthermore, the disconnect between inspection and repair means the system can only provide a pass / fail judgment, failing to offer parameter adjustment suggestions, resulting in high rework rates in fiber-to-the-home deployments. In addition, traditional inspection is post-inspection, unable to monitor and optimize in real time. Traditional microscopes struggle to identify microscopic defects such as nanoscale cracks, which threaten long-term network reliability. Moreover, related data is scattered, failing to form a knowledge base to support process improvements. While improvements have emerged in recent years, such as image processing and infrared thermal imaging, the former is susceptible to interference, and the latter cannot correlate with optical performance. Current technologies have not achieved multi-source information fusion and intelligent decision-making. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, apparatus, equipment and storage medium for optical fiber fusion splice quality detection and optimization, which aims to achieve accurate quantitative detection of optical fiber fusion splice quality and intelligent optimization of process parameters, effectively improve detection accuracy and reduce rework rate, and ensure the transmission performance and long-term reliability of optical fiber networks.

[0005] The first aspect of this invention provides a method for detecting and optimizing fiber optic fusion splice quality, comprising: collecting a multi-dimensional fusion splice dataset; preprocessing the multi-dimensional fusion splice dataset to obtain a fusion splice dataset to be tested; extracting multi-dimensional fusion features from the fusion splice dataset to be tested to obtain a late fusion feature set; calling a pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late fusion feature set to obtain a quality evaluation result; and optimizing the parameters of the fusion splice dataset to be tested based on the quality evaluation result to obtain an optimized fusion splice dataset.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the multi-dimensional weld point dataset includes an initial image data subset, an initial spectral data subset, and an initial process data subset; the step of preprocessing the multi-dimensional weld point dataset to obtain a weld point dataset to be tested includes: performing image enhancement processing on the initial image data subset using an image enhancement algorithm to obtain a target image data subset; performing wavelength calibration processing on the initial spectral data subset using a polynomial fitting algorithm to obtain a calibration spectral data subset; and performing normalization processing on the calibration spectral data subset. A normalized spectral data subset is obtained by performing numerical scaling uniformity processing. A smoothing algorithm is then used to remove background noise from the normalized spectral data subset to obtain a target spectral data subset. A dynamic time warping algorithm is used to align the initial process data subset temporally to obtain an aligned process data subset. A statistical filtering algorithm is then used to identify and remove outliers from the aligned process data subset to obtain the target process data subset. Finally, the target image data subset, the target spectral data subset, and the target process data subset are integrated to obtain the dataset of the weld joint to be tested.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the step of extracting multi-dimensional fusion features from the dataset of weld points to be tested to obtain a late fusion feature set includes: constructing a deep learning network, the deep learning network including an image feature extraction subnetwork and a spectral feature extraction subnetwork, the image feature extraction subnetwork including a spatial attention module, a multi-scale feature pyramid module, and a three-dimensional convolutional layer module, the spatial attention module, the multi-scale feature pyramid module, and the three-dimensional convolutional layer module being connected sequentially; performing key region enhancement processing on the dataset of weld points to be tested based on the spatial attention module to obtain an initial image feature set; performing multi-scale fusion processing on the initial image feature set based on the multi-scale feature pyramid module to obtain a fused image feature set; performing three-dimensional morphology modeling processing on the fused image feature set based on the three-dimensional convolutional layer module to obtain a target image feature set; performing spectral feature extraction on the dataset of weld points to be tested based on the spectral feature extraction subnetwork to obtain a target spectral feature set; and performing three-level correlation fusion processing on the target image feature set and the target spectral feature set to obtain the late fusion feature set.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the spectral feature extraction sub-network includes a one-dimensional convolutional layer module, an LSTM (Long Short-Term Memory) layer module, and a spectral feature attention module, wherein the one-dimensional convolutional layer module, the LSTM layer module, and the spectral feature attention module are connected sequentially; the step of extracting spectral features from the test fusion point dataset based on the spectral feature extraction sub-network to obtain a target spectral feature set includes: performing local feature extraction processing on the test fusion point dataset based on the one-dimensional convolutional layer module to obtain an initial spectral feature set; performing time-dependent capture processing on the initial spectral feature set based on the LSTM layer module to obtain a time-series fused spectral feature set; and performing sensitive band filtering processing on the time-series fused spectral feature set based on the spectral feature attention module to obtain the target spectral feature set.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the step of performing a three-level correlation fusion process on the target image feature set and the target spectral feature set to obtain the late-stage fusion feature set includes: using a channel stitching algorithm to perform data-level integration on the target image feature set and the target spectral feature set to obtain an early-stage fusion feature set; using a cross-attention mechanism to construct morphological optical correlation on the early-stage fusion feature set to obtain a mid-stage fusion feature set; and using a weighted fusion algorithm to perform decision-level score integration on the mid-stage fusion feature set to obtain the late-stage fusion feature set.

[0010] Optionally, in the fifth implementation of the first aspect of the present invention, before calling the pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late-stage fusion feature set and obtaining the quality evaluation result, the method further includes: constructing an initial quality detection model based on a gradient boosting tree network and a deep neural network; configuring the input feature dimension of the initial quality detection model based on a preset multi-dimensional feature system; obtaining a preset fiber wave equation and a preset heat conduction equation, and constructing a physical constraint system based on the fiber wave equation and the heat conduction equation; determining and setting the output parameter constraints of the initial quality detection model based on the physical constraint system; constructing a training sample set, and performing multiple rounds of iterative training on the initial quality detection model based on the input feature dimension, the output parameter constraints, and the training sample set to obtain the target quality detection model.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the quality assessment result includes a set of quality indicators corresponding to the weld point dataset to be tested; the step of optimizing the parameters of the weld point dataset to be tested based on the quality assessment result to obtain an optimized weld point dataset includes: using the SHAP algorithm (SHapley Additive ex Planations) to prioritize the parameter adjustment of the weld point dataset to be tested based on the set of quality indicators to obtain a priority ranking table; using a counterfactual reasoning algorithm and a response surface algorithm to construct an optimization path based on the priority ranking table; and using a non-dominated sorting genetic algorithm to optimize the parameters of the weld point dataset to be tested based on the optimization path to obtain the optimized weld point dataset.

[0012] A second aspect of the present invention provides an optical fiber fusion splice quality detection and optimization device, comprising: a data processing module for collecting a multi-dimensional fusion splice dataset and performing data preprocessing on the multi-dimensional fusion splice dataset to obtain a fusion splice dataset to be tested; a feature extraction module for extracting multi-dimensional fusion features from the fusion splice dataset to be tested to obtain a late fusion feature set; a quality detection module for calling a pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late fusion feature set to obtain a quality evaluation result; and a parameter optimization module for optimizing the parameters of the fusion splice dataset to be tested based on the quality evaluation result to obtain an optimized fusion splice dataset.

[0013] A third aspect of the present invention provides an optical fiber fusion splice quality inspection and optimization device, the optical fiber fusion splice quality inspection and optimization device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the optical fiber fusion splice quality inspection and optimization device to perform the various steps of the optical fiber fusion splice quality inspection and optimization method described in any of the preceding claims.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the fiber optic splice quality detection and optimization method described in any of the preceding claims.

[0015] In the technical solution of this invention, a multi-dimensional fusion splice dataset is first collected, and the dataset is preprocessed to obtain a dataset of fusion splices to be tested. Then, multi-dimensional fusion features are extracted from the dataset to be tested to obtain a late fusion feature set. Next, a pre-trained target quality detection model is called to perform multi-dimensional quality evaluation on the late fusion feature set to obtain a quality evaluation result. Finally, based on the quality evaluation result, the parameters of the dataset to be tested are optimized to obtain an optimized fusion splice dataset. This aims to achieve accurate quantitative detection of fiber optic fusion splice quality and intelligent optimization of process parameters, effectively improving detection accuracy and reducing rework rate, while ensuring the transmission performance and long-term reliability of fiber optic networks. Attached Figure Description

[0016] Figure 1 A logic flowchart of the optical fiber fusion splice quality detection and optimization method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the optical fiber fusion splice quality detection and optimization device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the optical fiber fusion splice quality inspection and optimization equipment provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a method, apparatus, device, and storage medium for inspecting and optimizing fiber optic fusion splice quality. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the fiber optic fusion splice quality detection and optimization method of the present invention includes: 101. Collect a multi-dimensional weld point dataset, perform data preprocessing on the multi-dimensional weld point dataset, and obtain the weld point dataset to be tested; In this embodiment, a multi-dimensional fusion splice dataset is acquired. Multiple macro cameras arranged in a ring simultaneously capture images of the fusion splice and surrounding areas from multiple angles, including 0°, 45°, 90°, and 135°, ensuring a resolution of at least 5 megapixels and a frame rate of 30fps. A hybrid illumination scheme combining coaxial incident illumination and dark-field illumination is used to enhance the contrast between surface morphology and internal defects. An integrated miniature fiber optic spectrometer is used to acquire the transmission, reflection, and scattering spectra of the fusion splice area within a wavelength range of 400nm to 1700nm with a resolution not exceeding 0.5nm. The spectral probe is integrated with the imaging optical path via a beam splitter, enabling simultaneous imaging and spectral measurement of the same area. Furthermore, key process parameters such as arc current and voltage waveforms, discharge time, fiber advance speed curves, ambient temperature and humidity, and fiber end-face cleanliness scores are recorded in real time. Simultaneously, the backscattered light time-domain reflection curves or optical time-domain analysis data of the fiber before and after fusion splicing are acquired as a benchmark for fusion splice loss. The collected raw multi-dimensional data undergoes systematic preprocessing. Specifically, image data is processed for denoising, contrast enhancement, distortion correction, and multi-view registration; spectral data is processed for wavelength calibration, intensity normalization, and baseline correction; and process data is processed for temporal alignment and outlier filtering. Then, by unifying the timestamps of all data, a spatiotemporally aligned multi-dimensional fusion point dataset, i.e., the fusion point dataset to be tested, is constructed. This approach overcomes the limitations of traditional single data acquisition, achieving comprehensive coverage and accurate capture of multi-dimensional information on the fusion state. The preprocessing stage effectively removes data noise, corrects deviations, and standardizes data formats, ensuring the accuracy, consistency, and completeness of the dataset to be tested, providing high-quality data support for subsequent multi-dimensional fusion feature extraction and quality assessment.

[0019] 102. Perform multi-dimensional fusion feature extraction on the dataset of weld joints to be tested to obtain a late-stage fusion feature set; In this embodiment, multi-dimensional fusion feature extraction is performed on the dataset of weld joints to be tested. The core is to first mine and refine features based on the characteristics of different dimensions of data, and then systematically fuse them to form a unified feature space, ultimately obtaining a late-stage fusion feature set. Through deep fusion of multi-dimensional features, the inherent correlation and complementary relationship between information in each dimension is accurately captured, overcoming the limitations of single-dimensional features in representing weld joint quality. The fused feature set has a more comprehensive and accurate quality representation capability, providing strong discriminative feature support for the multi-dimensional quality evaluation of subsequent pre-trained models, and significantly improving the model's accuracy in identifying weld joint quality and potential defects.

[0020] 103. Call the pre-trained target quality detection model to perform multi-dimensional quality assessment on the late fusion feature set, and obtain the quality assessment results; In this embodiment, the late-stage fusion feature set is input into a pre-trained target quality detection model to conduct multi-dimensional quality assessment. This model integrates a hybrid prediction model based on gradient boosting trees and deep neural networks, a multi-label classifier, and a reliability analysis module, possessing comprehensive welding quality characterization and assessment capabilities. For the real-time loss prediction dimension, a complex nonlinear mapping relationship is established through extensive experimental data training, outputting welding loss prediction values ​​in the range of 0dB to 3dB and loss prediction confidence levels in the range of 0 to 1. For the defect classification and severity assessment dimension, the model incorporates a multi-label classifier capable of accurately identifying various defects such as axial misalignment, radial misalignment, end face tilting, bubble clusters, microcracks, surface contamination, overheating deformation, insufficient welding, and material degradation. Based on the defect size, quantity, and location, each type of defect is comprehensively classified into three levels: minor, moderate, and severe. Simultaneously, the interaction effect analysis of multiple coexisting defects is completed, clarifying the cumulative impact of coupled defects on welding quality. For the long-term reliability prediction dimension… The model combines material fatigue modeling with accelerated aging test data, taking current defect status, material properties, and expected environmental conditions as inputs, to analyze the performance evolution trend of the weld joint within a preset lifespan. It outputs reliability index, expected lifespan, and failure probability curves, achieving a full life-cycle assessment of weld quality. For the comprehensive quality scoring dimension, the model normalizes real-time loss prediction values, defect severity assessment results, and long-term reliability prediction data to a scoring system of 0 to 100. Differentiated weights are set according to different application scenarios such as backbone networks, access networks, and data centers to generate the final comprehensive quality score and quality levels of excellent, qualified, and unqualified, forming a comprehensive and complete quality assessment result. This multi-dimensional quality assessment process breaks through the limitations of traditional weld quality inspection, which only focuses on loss values. It effectively solves the problem of large deviations between traditional two-dimensional image estimation methods and actual loss. The application of multi-label classifiers enables accurate quantitative judgment of defect type and severity, completely eliminating reliance on human experience and eliminating evaluation differences among different operators. The long-term reliability prediction dimension makes up for the shortcomings of traditional inspection in assessing the long-term performance of weld joints, identifying potential hidden dangers such as nanoscale cracks and internal stress concentration areas in advance, and preventing such defects from threatening the stability of network operation over time.

[0021] 104. Based on the quality assessment results, optimize the parameters of the weld point dataset to be tested to obtain an optimized weld point dataset.

[0022] In this embodiment, the quality assessment results include a set of quality indicators. Parameter optimization of the weld point dataset to be tested specifically involves: based on this set of quality indicators, quantitatively analyzing the contribution of each process parameter in the weld point dataset to the core quality indicators, prioritizing parameter adjustments, and then deducing the potential impact of parameter adjustments on the quality indicators to eliminate invalid directions and construct an operable adjustment path. Guided by this path, under constraints such as equipment operating limits and process ranges, and taking into account objectives such as minimizing weld loss and maximizing long-term reliability, iteratively searching for the Pareto optimal parameter combination, gradient adjusting key parameters, and updating information such as process parameters and quality prediction indicators in the weld point dataset to be tested, ultimately obtaining an optimized weld point dataset.

[0023] In this embodiment, the optimized fusion splice point dataset is incorporated as a core sample into the system's knowledge base construction and self-learning evolution system. Specifically, the system integrates the complete information of each fusion splicing operation into standardized cases and stores them in the fusion splice case library. This includes the input conditions before optimization, the multimodal monitoring data sequence during the process, the result data before and after optimization, and complete optimization records. A graph database storage architecture is adopted, and by establishing a similarity relationship index between cases, rapid retrieval and analogical reasoning of cases in similar scenarios are achieved. This provides reusable experience support for parameter optimization of similar fusion splices in subsequent scenarios, significantly shortening the decision cycle for parameter optimization in similar scenarios and avoiding repeated trial and error. At the same time, based on optimized cases and newly added operation data, the system initiates an online model update and incremental learning mechanism. For cases with low prediction confidence during parameter optimization, manual review is prompted, and the review results are used as new samples to supplement training, continuously improving the model's prediction accuracy. Real-time monitoring of data distribution changes is conducted. If new fiber types or new environmental conditions are encountered, model retraining is automatically triggered to adapt to the new working conditions. Distributed federated learning across multiple devices is supported, enabling cross-device model knowledge sharing and collaborative upgrades while ensuring data privacy and not sharing original data. Furthermore, by using techniques such as association rule mining and cluster analysis, we can deeply mine the optimal combination of process parameters for specific fiber types, effective prevention measures for common defects, and the influence of environmental factors on fusion splicing quality from a large number of optimization cases and historical operation data. Based on these patterns, we can construct a fusion splicing process knowledge graph, transforming implicit optimization experience and process rules into explicit, explainable, and reasonable knowledge, thereby achieving standardized inheritance and efficient reuse of process knowledge.

[0024] In this embodiment of the invention, the multi-dimensional weld point dataset includes an initial image data subset, an initial spectral data subset, and an initial process data subset. The data preprocessing of the multi-dimensional weld point dataset to obtain the weld point dataset to be tested includes: performing image enhancement processing on the initial image data subset using an image enhancement algorithm to obtain a target image data subset; performing wavelength calibration processing on the initial spectral data subset using a polynomial fitting algorithm to obtain a calibration spectral data subset; performing numerical scale unification processing on the calibration spectral data subset using a normalization algorithm to obtain a normalized spectral data subset; performing background noise removal processing on the normalized spectral data subset using a smoothing algorithm to obtain a target spectral data subset; performing time-series alignment processing on the initial process data subset using a dynamic time warping algorithm to obtain an aligned process data subset; performing outlier identification and removal processing on the aligned process data subset using a statistical filtering algorithm to obtain a target process data subset; and integrating the target image data subset, the target spectral data subset, and the target process data subset to obtain the weld point dataset to be tested.

[0025] In this embodiment, the multi-dimensional fusion splice dataset includes an initial image data subset, an initial spectral data subset, and an initial process data subset. The initial image data subset comes from image data of the fusion splice and its surrounding area simultaneously acquired from multiple angles by multiple macro cameras arranged in a ring. The acquisition process adopts a hybrid illumination scheme combining coaxial incident illumination and dark field illumination. The initial spectral data subset comes from the transmission spectrum, reflection spectrum, and scattering spectrum data of the fusion splice area acquired by an integrated miniature fiber optic spectrometer. The spectrometer has a wavelength range covering 400nm to 1700nm and a resolution not exceeding 0.5nm. The spectral probe is integrated with the imaging optical path through a beam splitter to achieve simultaneous imaging and spectral measurement of the same area. The initial process data subset covers key process parameters acquired in real time during the fusion splicing process, such as arc current and voltage waveforms, discharge time, fiber advance speed curves, ambient temperature and humidity, and fiber end-face cleanliness scores, as well as the backscattered light time-domain reflection curves or optical time-domain analysis data of the fiber before and after fusion splicing.

[0026] In this embodiment, an image enhancement algorithm is used to perform a full-process processing of denoising, contrast enhancement, distortion correction, and multi-view registration on the initial image data subset. In the denoising stage, Gaussian filtering or bilateral filtering algorithms can be used to remove image noise. In the contrast enhancement stage, histogram equalization or adaptive histogram equalization algorithms can be used to improve image contrast to highlight the surface morphology and internal defect features of the weld point. In the distortion correction stage, camera calibration combined with perspective transformation algorithms can be used to correct image geometric distortion. In the multi-view registration stage, scale-invariant feature transformation or accelerated robust feature transformation algorithms can be used to achieve accurate registration of images from different angles. After the above processing, the target image data subset is obtained.

[0027] In this embodiment, a polynomial fitting algorithm is used to perform wavelength calibration on the initial spectral data subset. Based on a standard spectral reference source with known characteristic peaks, a deviation model between the measured wavelength and the standard wavelength is constructed. A precise wavelength calibration curve is obtained by fitting the model using a quadratic or cubic polynomial fitting algorithm, and the initial spectral data is corrected point-by-point to eliminate wavelength drift caused by spectrometer system errors. After calibration, a calibrated spectral data subset is obtained. Then, a normalization algorithm is used to unify the numerical scale of the calibrated spectral data subset. The min-max normalization algorithm is selected to map the spectral intensity to the interval between 0 and 1, thereby eliminating the scale difference of spectral intensity under different acquisition conditions, resulting in a normalized spectral data subset. Next, a smoothing algorithm is used to remove background noise from the normalized spectral data subset. The Savitzky-Golay smoothing algorithm, through polynomial fitting within a local window, removes background noise while preserving the spectral characteristic peak shape, further suppressing random noise interference, resulting in the target spectral data subset.

[0028] In this embodiment, a dynamic time warping algorithm is first used to perform time-series alignment processing on the initial process data subset. Addressing the issues of inconsistent time-series lengths and asynchronous sampling times of process parameters collected from different sensors, the optimal time-matching path is found to align the time-series curves of all process parameters, such as arc current and voltage waveforms and fiber propulsion speed curves, to a unified time axis, ensuring consistency of process parameters from different sources in the time dimension, thus obtaining an aligned process data subset. Then, a statistical filtering algorithm is used to identify and remove outliers from the aligned process data subset. For example, the Grubbs criterion algorithm can be used to calculate the deviation of data points from the mean, combined with a preset significance level to determine and remove outliers, identifying abnormal data points caused by instantaneous equipment failures or sudden environmental changes, thus obtaining the target process data subset. Finally, the target image data subset, target spectral data subset, and target process data subset are integrated, and a unified timestamp is applied to all data subsets to establish a spatiotemporal alignment relationship, ultimately obtaining the dataset of the weld joint to be tested. By selectively choosing appropriate algorithms to refine the processing of different types of data, noise interference in various types of data is effectively eliminated, systematic errors and distortions generated during data collection are corrected, and spatiotemporal alignment and scale uniformity of data in different dimensions are achieved, significantly improving the accuracy, consistency and completeness of the dataset.

[0029] In this embodiment of the invention, the step of extracting multi-dimensional fusion features from the test fusion point dataset to obtain a late fusion feature set includes: constructing a deep learning network, the deep learning network including an image feature extraction subnetwork and a spectral feature extraction subnetwork, the image feature extraction subnetwork including a spatial attention module, a multi-scale feature pyramid module, and a three-dimensional convolutional layer module, the spatial attention module, the multi-scale feature pyramid module, and the three-dimensional convolutional layer module being connected sequentially; performing key region enhancement processing on the test fusion point dataset based on the spatial attention module to obtain an initial image feature set; performing multi-scale fusion processing on the initial image feature set based on the multi-scale feature pyramid module to obtain a fused image feature set; performing three-dimensional topography modeling processing on the fused image feature set based on the three-dimensional convolutional layer module to obtain a target image feature set; performing spectral feature extraction on the test fusion point dataset based on the spectral feature extraction subnetwork to obtain a target spectral feature set; and performing a three-level correlation fusion processing on the target image feature set and the target spectral feature set to obtain the late fusion feature set.

[0030] In this embodiment, a multi-input multi-task deep learning network is constructed. The network adopts an improved architecture design to achieve deep extraction and fusion representation of multimodal features. The image feature extraction sub-network uses an improved ResNet-50 as the backbone network, and the core processing link is composed of a spatial attention module, a multi-scale feature pyramid module, and a three-dimensional convolutional layer module connected in sequence.

[0031] In this embodiment, firstly, based on the spatial attention module of the image feature extraction subnetwork, key region enhancement processing is performed on the registered multi-view image stack within the preprocessed test weld point dataset. By adaptively learning the importance difference between the weld core region and the surrounding background region, interference from irrelevant background information is suppressed, focusing on the weld point and potential defect regions, enhancing effective feature signals, and obtaining an initial image feature set that highlights key region information. Then, relying on the multi-scale feature pyramid module, multi-scale fusion processing is performed on the initial image feature set. Through hierarchical feature extraction and integration, macroscopic morphological features and microscopic defect features of the weld point are captured simultaneously, making up for the problem that single-scale features are insufficient to represent defects of different sizes, forming a fused image feature set with both completeness and precision. Finally, based on the three-dimensional convolutional layer module, three-dimensional morphological modeling processing is performed on the fused image feature set. Spatial correlation information is deeply mined in multiple perspective dimensions to construct a three-dimensional point cloud model of the weld point, accurately extracting multiple geometric morphological parameters such as the center position of the weld point, axial and radial misalignment, end face tilt angle, bubble size distribution, and crack length direction, ultimately obtaining the target image feature set. For spectral feature extraction, a spectral feature extraction sub-network is used to extract spectral features from the dataset of weld points under test to obtain the target spectral feature set. Finally, a three-level correlation fusion process is performed on the target image feature set and the target spectral feature set to obtain the late-stage fused feature set. Targeted module design achieves accurate extraction of image and spectral features. Spatial attention and multi-scale fusion mechanisms significantly improve the ability of image features to represent defects. The three-dimensional model constructed by the three-dimensional convolutional layer overcomes the information limitations of two-dimensional images. The hybrid architecture of the spectral sub-network takes into account both local features and temporal evolution patterns. The three-level fusion strategy effectively avoids the information loss problem of single fusion methods, deeply explores the correlation between morphological and optical features, and the final late-stage fused feature set has comprehensive and accurate quality representation capabilities.

[0032] In this embodiment of the invention, the spectral feature extraction subnetwork includes a one-dimensional convolutional layer module, an LSTM layer module, and a spectral feature attention module, which are sequentially connected. The step of extracting spectral features from the weld point dataset to be tested using the spectral feature extraction subnetwork to obtain a target spectral feature set includes: performing local feature extraction processing on the weld point dataset to be tested using the one-dimensional convolutional layer module to obtain an initial spectral feature set; performing time-dependent capture processing on the initial spectral feature set using the LSTM layer module to obtain a time-series fused spectral feature set; and performing sensitive band filtering processing on the time-series fused spectral feature set using the spectral feature attention module to obtain the target spectral feature set.

[0033] In this embodiment, the spectral feature extraction subnetwork consists of a one-dimensional convolutional layer module, an LSTM layer module, and a spectral feature attention module connected in series. Based on this hierarchical architecture, refined spectral feature extraction is performed on the spectral sequence data within the weld point dataset to be tested, ultimately obtaining the target spectral feature set. Specifically, the one-dimensional convolutional layer module performs local feature extraction processing on the weld point dataset to be tested. This module sets the convolutional kernel size and stride to adapt to the spectral data, performs local sliding convolution operations on the transmission, reflection, and scattering spectral curves, and combines activation and pooling operations to accurately mine local feature information in the spectral data, including absorption peaks, reflection valleys, peak values, full width at half maximum (FWHM), and peak spacing of scattering characteristic peaks. These local features are directly related to the composition and microstructure changes of the material in the welded area, and can effectively remove local random noise in the spectral data, retaining effective local features with discriminative value to form an initial spectral feature set. Then, the initial spectral feature set is processed by time-dependent capture based on the LSTM layer module. This module is based on a gating mechanism, which can effectively avoid the gradient vanishing problem of traditional recurrent neural networks. It accurately learns and captures the time-series correlation of spectral features with the evolution of processes such as welding arc discharge, temperature rise and fall, and material melting and solidification. It integrates discrete local spectral features into a feature expression with temporal coherence, fully restores the dynamic evolution trajectory of the spectral signal with the entire welding process, highlights the correlation and change law of spectral features at different welding stages, and forms a time-series fused spectral feature set. Next, the temporal fusion spectral feature set is processed by a sensitive band screening module based on spectral feature attention. This module adaptively learns the correlation between welding quality indicators and different spectral bands, assigning higher weights to specific wavelength ranges that are sensitive to core quality indicators such as welding loss and material defects. At the same time, it suppresses the interference of irrelevant bands and redundant information, and selects spectral bands and corresponding features that have core value for welding quality judgment. Finally, it extracts multiple optical characteristic parameters such as transmittance, reflectance, scattering intensity distribution, spectral slope change, and characteristic peak shift at specific wavelengths. The resulting target spectral feature set has the integrity of local details, the coherence of temporal evolution, and the saliency of core features. It can accurately characterize the material properties, microstructure changes, and process status of the welding area, significantly improving the ability of subsequent welding quality assessment to identify micro-defects such as material degradation and internal stress concentration, while reducing the computational complexity of the overall system.

[0034] In this embodiment of the invention, the three-level correlation fusion processing of the target image feature set and the target spectral feature set to obtain the late-stage fusion feature set includes: using a channel stitching algorithm to perform data-level integration of the target image feature set and the target spectral feature set to obtain an early-stage fusion feature set; using a cross-attention mechanism to construct morphological optical correlation of the early-stage fusion feature set to obtain a mid-stage fusion feature set; and using a weighted fusion algorithm to perform decision-level score integration of the mid-stage fusion feature set to obtain the late-stage fusion feature set.

[0035] In this embodiment, for the data layer fusion stage, a channel stitching algorithm is used to integrate two types of feature sets. The target image feature set, which represents the three-dimensional morphology and geometric defects of the weld point, and the target spectral feature set, which reflects the optical properties and microstructural changes of the material, are expanded and stitched together in the feature channel dimension. This allows the two types of original feature information to be completely preserved and initially fused, forming an early fused feature set. This lays the data foundation for subsequent cross-modal correlation mining, and is simultaneously input into the fusion encoder to complete the initial feature encoding. For the feature layer fusion stage, a morphological optical correlation is constructed on the early fused feature set based on a cross-attention mechanism. The cross-attention mechanism guides the image feature branches to adaptively focus on regions strongly correlated with spectral features, such as the shift of feature peaks and changes in absorption intensity in the spectrum corresponding to the location of weld point defects. At the same time, the spectral feature branches focus on band information that matches the image morphological features, realizing bidirectional interaction and precise correlation between morphological and optical features. This effectively mines the inherent coupling law between the two types of features, strengthens feature complementarity, eliminates cross-modal redundant information, and improves the feature's ability to represent the weld quality, ultimately obtaining the mid-term fused feature set. For the decision-level fusion stage, a weighted fusion algorithm is used to integrate the scores of the mid-stage fusion feature set. Learnable weight parameters are used to adaptively weight the quality prediction scores output by the image feature path and the spectral feature path, respectively. These weight parameters are dynamically adjusted based on the quality assessment contribution of the two types of features under different welding scenarios and defect types, ensuring that features more valuable for quality assessment receive higher weights and maximizing the retention of the discriminative value of both types of features. After weighted fusion and normalization, the late-stage fusion feature set is obtained. This three-level fusion strategy effectively avoids the information loss or insufficient correlation problems of single fusion methods. The channel stitching algorithm preserves the integrity of the original information of the two types of features in the early stage, avoiding distortion of basic features. The cross-attention mechanism builds a precise correlation between morphology and optical properties in the mid-stage, breaking through the limitations of the two types of features representing themselves independently and strengthening the correspondence between defects and material properties. The learnable weighted fusion algorithm achieves optimized integration at the decision level in the late stage, adapting to the differences in feature value under different scenarios. The final late-stage fusion feature set possesses spatial integrity of morphological information, microscopic sensitivity of optical information, and intrinsic correlation of cross-modal features, comprehensively covering the core dimensions of welding quality assessment.

[0036] In this embodiment of the invention, before calling the pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late-stage fusion feature set and obtaining the quality evaluation result, the method further includes: constructing an initial quality detection model based on a gradient boosting tree network and a deep neural network; configuring the input feature dimensions of the initial quality detection model based on a preset multi-dimensional feature system; obtaining a preset fiber wave equation and a preset heat conduction equation, and constructing a physical constraint system based on the fiber wave equation and the heat conduction equation; determining and setting the output parameter constraints of the initial quality detection model based on the physical constraint system; constructing a training sample set, and performing multiple rounds of iterative training on the initial quality detection model based on the input feature dimensions, the output parameter constraints, and the training sample set to obtain the target quality detection model.

[0037] In this embodiment, an initial quality detection model is constructed based on a gradient boosting tree network and a deep neural network. This fully combines the gradient boosting tree network's ability to accurately mine feature correlations with the deep neural network's ability to fit complex nonlinear relationships, laying the foundation for the model's efficient quality assessment performance. The input feature dimensions of the initial quality detection model are configured based on a preset multi-dimensional feature system. This system corresponds to a 38-dimensional feature vector covering 18 geometric morphology parameters, 12 optical characteristic parameters, and 8 process parameters, ensuring that the model input comprehensively covers the core feature information related to fusion splicing quality. Preset fiber wave equations and heat conduction equations are obtained, and a complete physical constraint system is constructed based on these two types of physical equations. This system fully conforms to the light transmission law and thermal melting and solidification law in the fiber optic fusion splicing process. Then, based on the physical constraint system, the output parameter constraints of the initial quality detection model are determined and set, clarifying the reasonable range and physical boundaries of the model's output results. For example, the predicted fusion loss value is constrained to be within a continuous range of 0 to 3 dB, ensuring that the subsequent output results of the model do not violate objective physical laws. The constructed training sample set covers fusion splicing data under different fiber types, environmental conditions, and process parameters. It also includes complete annotation information such as the corresponding measured values ​​of fusion splicing loss, defect type and severity labeling, and long-term reliability verification data. Subsequently, the initial quality detection model is trained in multiple rounds based on the input feature dimension, output parameter constraints, and training sample set. During the training process, the network weights are continuously adjusted, the model structure is optimized, and the prediction bias is corrected, so that the model gradually has stable feature mapping capabilities and accurate multi-dimensional evaluation capabilities, and finally the pre-trained target quality detection model is obtained.

[0038] In this embodiment of the invention, the quality assessment result includes a set of quality indicators corresponding to the weld point dataset to be tested; the step of optimizing the parameters of the weld point dataset to be tested based on the quality assessment result to obtain an optimized weld point dataset includes: using the SHAP algorithm to prioritize the parameter adjustment of the weld point dataset to be tested based on the set of quality indicators to obtain a priority ranking table; using a counterfactual reasoning algorithm and a response surface algorithm to construct an optimization path based on the priority ranking table; and using a non-dominated sorting genetic algorithm to optimize the parameters of the weld point dataset to be tested based on the optimization path to obtain the optimized weld point dataset.

[0039] In this embodiment, the quality assessment result includes a set of quality indicators corresponding to the weld point dataset to be tested. The quality indicator set focuses on the quantitative characterization of the weld point defect status, covering core indicators such as defect type, defect severity level, and defect interaction coupling effect. Based on this quality assessment result, parameter optimization of the weld point dataset to be tested is essentially to conduct precise root cause diagnosis, path planning, and parameter calibration for unqualified or suboptimal weld points, ultimately obtaining an optimized weld point dataset. Specifically, the SHAP algorithm is used based on the quality indicator set to quantitatively analyze various process parameters in the weld point dataset to be tested, accurately calculating the induction probability of different process parameters, the degree of aggravation of defect severity, and the influence weight of multi-defect coupling effect. Based on this, the parameter adjustment priority is sorted and a priority ranking table is generated, clarifying the hierarchical relationship between the core process parameters with the most significant effect on defect improvement and the secondary parameters with a lower degree of influence. Key control factors under dimensions such as arc parameters, alignment parameters, and environmental parameters are prioritized to anchor the direction of subsequent optimization work, avoiding the blindness of traditional parameter adjustment. Priority ranking allows core parameters to be optimized in a focused manner, greatly improving the efficiency and accuracy of adjustment. Then, based on this priority ranking table, a scientifically feasible optimization path is constructed by jointly employing counterfactual reasoning and response surface methodology. The counterfactual reasoning algorithm generates virtual scenarios of parameter adjustments, deduces the changing trends of quality indicators under different parameter adjustment ranges and combinations, simulates the potential impact of parameter changes on welding quality, effectively eliminates ineffective adjustment directions, and locates the optimal adjustment dimension. The response surface methodology constructs a nonlinear correlation model between the process parameter space and the quality indicator space, fits the mapping relationship between parameter combinations and quality results, and outlines the adjustment trajectory from the current parameter state to the optimal quality state. At the same time, combined with parameter sensitivity analysis, the adjustment range and gradient of key parameters are clarified, ensuring that the optimization path has both theoretical support and practical operability.Finally, guided by the constructed optimization path, a non-dominated sorting genetic algorithm was used to optimize the parameters of the weld point dataset under test. The non-dominated sorting genetic algorithm takes minimizing welding loss as its primary objective and maximizing long-term reliability as its secondary objective. It strictly follows the constraints of feasible process parameters, equipment operating limits, and operation time limitations, and iteratively searches for Pareto optimal parameter combinations. For core parameters in the priority sorting table, such as arc current intensity, discharge time, XY axis compensation, and environmental temperature and humidity control thresholds, gradient adjustment and combination optimization are performed. During the optimization process, the effective adjustment directions obtained from counterfactual reasoning and the parameter correlation rules fed back by the response surface model are integrated simultaneously to ensure that the optimized parameters not only meet the equipment capabilities but also accurately improve quality defects. After parameter calibration, the optimized process parameters, expected quality improvement effects, parameter adjustment logic, and other information are synchronously updated to the weld point dataset under test, ultimately forming a complete optimized weld point dataset. This achieves a balance between multi-objective requirements and actual constraints, ensuring that the optimized parameters reduce welding loss while taking into account long-term operational reliability, thus breaking away from the traditional mode of relying on repeated manual debugging.

[0040] The fiber optic splice quality detection and optimization method in the embodiments of the present invention has been described above. The fiber optic splice quality detection and optimization device in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 2 One embodiment of the optical fiber fusion splice quality detection and optimization device of the present invention includes: Data processing module 201: used to collect multi-dimensional weld point datasets, perform data preprocessing on the multi-dimensional weld point datasets, and obtain the weld point dataset to be tested; Feature extraction module 202: used to perform multi-dimensional fusion feature extraction on the test weld point dataset to obtain a late fusion feature set; Quality detection module 203: used to call a pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late fusion feature set and obtain quality evaluation results; Parameter optimization module 204: used to optimize the parameters of the weld point dataset to be tested based on the quality assessment results, so as to obtain an optimized weld point dataset.

[0041] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0042] above Figure 2 The fiber optic splice quality detection and optimization device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The fiber optic splice quality detection and optimization device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0043] Figure 3This is a schematic diagram of the structure of an optical fiber fusion splice quality inspection and optimization device 300 provided in an embodiment of the present invention. The optical fiber fusion splice quality inspection and optimization device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the optical fiber fusion splice quality inspection and optimization device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the optical fiber fusion splice quality inspection and optimization device 300 to implement the steps of the optical fiber fusion splice quality inspection and optimization methods provided in the above-described method embodiments.

[0044] The fiber optic fusion splice quality inspection and optimization equipment 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated fiber optic fusion splice quality inspection and optimization equipment structure does not constitute a limitation on the fiber optic fusion splice quality inspection and optimization equipment. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0045] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the fiber optic splice quality detection and optimization method.

[0046] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0047] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0048] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting and optimizing the quality of optical fiber fusion splices, characterized in that, include: Collect a multi-dimensional weld point dataset, perform data preprocessing on the multi-dimensional weld point dataset, and obtain the weld point dataset to be tested; Multi-dimensional fusion feature extraction is performed on the dataset of weld joints to be tested to obtain a late-stage fusion feature set; The pre-trained target quality detection model is invoked to perform multi-dimensional quality assessment on the late fusion feature set, and the quality assessment results are obtained. Based on the quality assessment results, the parameters of the weld point dataset to be tested are optimized to obtain an optimized weld point dataset.

2. The method for detecting and optimizing fiber optic fusion splice quality according to claim 1, characterized in that, The multi-dimensional weld point dataset includes a subset of initial image data, a subset of initial spectral data, and a subset of initial process data; The process of preprocessing the multi-dimensional weld point dataset to obtain the weld point dataset to be tested includes: An image enhancement algorithm is used to perform image enhancement processing on the initial image data subset to obtain the target image data subset; A polynomial fitting algorithm is used to perform wavelength calibration on the initial spectral data subset to obtain a calibrated spectral data subset. The calibration spectral data subset is subjected to numerical scaling uniformity processing using a normalization algorithm to obtain a normalized spectral data subset; A smoothing algorithm is used to remove background noise from the normalized spectral data subset to obtain the target spectral data subset. The initial process data subset is time-aligned using a dynamic time warping algorithm to obtain an aligned process data subset. A statistical filtering algorithm is used to identify and remove outliers from the aligned process data subset to obtain the target process data subset. The target image data subset, the target spectral data subset, and the target process data subset are integrated to obtain the test weld point dataset.

3. The method for detecting and optimizing fiber optic fusion splice quality according to claim 1, characterized in that, The step of extracting multi-dimensional fusion features from the dataset of weld points to be tested to obtain a late-stage fusion feature set includes: A deep learning network is constructed, which includes an image feature extraction subnetwork and a spectral feature extraction subnetwork. The image feature extraction subnetwork includes a spatial attention module, a multi-scale feature pyramid module, and a three-dimensional convolutional layer module, which are connected sequentially. Based on the spatial attention module, the dataset of weld points to be tested is subjected to key region enhancement processing to obtain an initial image feature set; Based on the multi-scale feature pyramid module, the initial image feature set is subjected to multi-scale fusion processing to obtain a fused image feature set; Based on the three-dimensional convolutional layer module, the fused image feature set is subjected to three-dimensional morphological modeling to obtain the target image feature set; Based on the spectral feature extraction subnetwork, spectral features are extracted from the dataset of weld points to be tested to obtain the target spectral feature set; The target image feature set and the target spectral feature set are subjected to a three-level correlation fusion process to obtain the late fusion feature set.

4. The method for detecting and optimizing fiber optic splice quality according to claim 3, characterized in that, The spectral feature extraction subnetwork includes a one-dimensional convolutional layer module, an LSTM layer module, and a spectral feature attention module, which are sequentially connected. The step of extracting spectral features from the weld point dataset to be tested based on the spectral feature extraction subnetwork to obtain a target spectral feature set includes: Based on the one-dimensional convolutional layer module, local feature extraction processing is performed on the dataset of weld points to be tested to obtain an initial spectral feature set; Based on the LSTM layer module, the initial spectral feature set is subjected to time-dependent capture processing to obtain a time-fused spectral feature set; The target spectral feature set is obtained by performing sensitive band filtering on the time-series fused spectral feature set based on the spectral feature attention module.

5. The method for detecting and optimizing fiber optic fusion splice quality according to claim 3, characterized in that, The three-level correlation fusion processing of the target image feature set and the target spectral feature set to obtain the late fusion feature set includes: A channel stitching algorithm is used to integrate the target image feature set and the target spectral feature set at the data layer to obtain an early fusion feature set; A cross-attention mechanism is used to construct morphological optical correlations on the early fusion feature set to obtain the mid-term fusion feature set; The decision-level scores of the intermediate fusion feature set are integrated using a weighted fusion algorithm to obtain the late fusion feature set.

6. The method for detecting and optimizing fiber optic fusion splice quality according to claim 1, characterized in that, Before the step of calling the pre-trained target quality detection model to perform multi-dimensional quality assessment on the late-stage fusion feature set and obtaining the quality assessment results, the method further includes: An initial quality detection model was constructed based on gradient boosting tree network and deep neural network; The input feature dimensions of the initial quality detection model are configured based on a preset multidimensional feature system; Obtain the preset fiber wave equation and the preset heat conduction equation, and construct a physical constraint system based on the fiber wave equation and the heat conduction equation; The output parameter constraints of the initial quality inspection model are determined and set based on the physical constraint system. A training sample set is constructed, and the initial quality detection model is trained iteratively in multiple rounds based on the input feature dimension, the output parameter constraints, and the training sample set to obtain the target quality detection model.

7. The method for detecting and optimizing fiber optic fusion splice quality according to claim 1, characterized in that, The quality assessment results include a set of quality indicators corresponding to the weld point dataset to be tested; the step of optimizing the parameters of the weld point dataset to be tested based on the quality assessment results to obtain an optimized weld point dataset includes: The SHAP algorithm is used to prioritize and sort the parameters of the weld joint dataset to be tested based on the quality index set, resulting in a priority sorting table. An optimized path is constructed based on the priority ranking table using a counterfactual reasoning algorithm and a response surface algorithm. A non-dominated sorting genetic algorithm is used to optimize the parameters of the test weld point dataset based on the optimized path, resulting in the optimized weld point dataset.

8. A device for detecting and optimizing fiber optic fusion splice quality, characterized in that, include: Data processing module: used to collect multi-dimensional weld point datasets, perform data preprocessing on the multi-dimensional weld point datasets, and obtain the weld point dataset to be tested; Feature extraction module: used to perform multi-dimensional fusion feature extraction on the dataset of weld joints to be tested, to obtain a late fusion feature set; Quality detection module: used to call the pre-trained target quality detection model to perform multi-dimensional quality evaluation on the late fusion feature set and obtain the quality evaluation results; Parameter optimization module: used to optimize the parameters of the weld point dataset to be tested based on the quality assessment results, so as to obtain an optimized weld point dataset.

9. A fiber optic fusion splice quality inspection and optimization device, characterized in that, The fiber optic splice quality detection and optimization device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the fiber optic splice quality inspection and optimization device to perform the various steps of the fiber optic splice quality inspection and optimization method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the fiber optic splice quality detection and optimization method as described in any one of claims 1-7.

Images