Pipeline element library creating method based on model training and image recognition technology

Abstract

The invention relates to the technical field of image recognition data processing, in particular to a pipeline element library creating method based on a model training and image recognition technology, which comprises the following steps of: acquiring pipeline element image data with depth information through a multispectral imaging scheme, executing Retinex algorithm illumination equalization and morphological restoration through a cascade preprocessing pipeline, and establishing a pipeline element library. A standardized data set associated with the metadata is generated. A multi-task joint learning framework is adopted to integrate ResNet-50, HRNet and Mask R-CNN networks, a bottom convolution feature extraction layer is shared, oil stain and strong reflection antagonism data synthesized by a generative adversarial network is injected, model parameters are optimized in combination with a progressive training strategy, and a lightweight MobileNetV3 model is output. Model parameters and an index structure are dynamically updated, and system self-calibration is achieved in combination with a cross-device calibration protocol and data consanguinity tracking. According to the method, through automatic data labeling, multi-task feature multiplexing and retrieval feedback closed loop, the construction efficiency of the pipeline element library is effectively improved, and the recognition error rate in a complex environment is reduced.

Description

Technical Field

[0001] This invention relates to the field of image recognition data processing technology, and in particular to a method for creating a pipeline component library based on model training and image recognition technology. Background Technology

[0002] Existing methods for creating pipe component libraries are based on deep learning frameworks. They construct training datasets by collecting images of pipe components from multiple angles, employ median filtering and normalization techniques for image denoising and size standardization, and combine data augmentation strategies (such as random rotation and brightness adjustment) to expand sample diversity. Convolutional neural networks (CNNs) are used to extract component morphology and texture features, and transfer learning is used to optimize the model's generalization ability in small-sample scenarios. Cross-validation and confusion matrices are employed to evaluate classification accuracy. The recognition results are mapped to structured parameters (pipe diameter, material, etc.) using metadata specifications. A component topology network is constructed based on a graph database, and a version control mechanism is integrated to achieve dynamic iteration of the component library, ultimately forming a standardized pipe component knowledge graph that supports multi-dimensional retrieval and engineering application integration.

[0003] In existing technologies, the construction of pipeline component libraries relies on heterogeneous pipeline design software or systems. Component feature extraction and data annotation must be completed through manual visual identification and data entry. The accuracy of identification is limited by the operator's experience level and environmental factors, and feature misjudgment and data deviation are prone to occur under complex working conditions. At the same time, the lack of standardized data architecture and information management mechanisms leads to the scattered storage of component attribute information and the lack of unified metadata specifications. It is difficult to achieve automatic integration and dynamic updates of multi-source heterogeneous data, resulting in technical bottlenecks such as insufficient component library integrity, lagging version iteration, and poor cross-platform compatibility. These issues seriously restrict the digitalization process and intelligent application promotion of pipeline engineering design. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for creating a pipe component library based on model training and image recognition technology, which solves the problems of low efficiency, high error rate, and poor adaptability to complex environments that exist in manually creating pipe component libraries.

[0005] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows: This invention provides a method for creating a pipeline component library based on model training and image recognition technology, including: Step 1: Obtain image data of multi-source pipeline components and associated environmental parameters. Generate a standardized image dataset by performing cascaded preprocessing on the image data of multi-source pipeline components and associated environmental parameters. Step 2: Input the standardized image dataset into the multi-task joint learning framework for dynamic training. The multi-task joint learning framework integrates the component classification sub-network, the component key point localization sub-network, and the component defect detection sub-network, and outputs the training model by sharing the underlying feature extraction layer. Step 3: Based on the feature vectors extracted by the training model, generate component feature descriptors, associate the component feature descriptors with the component assembly relationships in the preset knowledge graph, construct a multi-level index structure, receive query input including the image or feature vector of the component to be retrieved, parse the component type identifier in the query, determine the target component category, perform an approximate nearest neighbor search in the target category, output candidate components with similar features, verify the spatial assembly compatibility between the candidate components and the query components, filter the final matching results that pass the topological constraints, use the final matching results as the structured retrieval results, use the structured retrieval results as the data input source of the closed-loop evolution mechanism, and output them to Step 4 in real time; Step 4: When the feature similarity score in the structured retrieval results is lower than the set threshold, it is marked as a false detection sample. The corrected label and associated environmental parameters of the false detection sample are extracted to generate incremental data. Based on the incremental data, a closed-loop evolution mechanism is triggered. The closed-loop evolution mechanism includes acquiring incremental data, performing cascaded preprocessing on the incremental data to generate an incremental standardized dataset, inputting the incremental standardized dataset into a multi-task joint learning framework, updating the feature extraction parameters of the training model, extracting feature vectors based on the updated training model, reconstructing the multi-level index structure of the hierarchical component library, and feeding back the retrieval results of the updated hierarchical component library to Step 1 to form a closed-loop iteration.

[0006] Furthermore, the method for creating a pipe component library based on model training and image recognition technology of the present invention includes step 1: acquiring images of pipe components including depth information through a multispectral imaging scheme to generate raw data of occlusion and low-light scenes; The raw data is input into the cascaded processing pipeline. The cascaded preprocessing pipeline sequentially performs illumination equalization processing based on the multi-scale Retinex algorithm, decomposes the illumination and reflection components of the image using the Gaussian kernel function, and performs texture restoration based on morphological opening operation to obtain the restored image. The rectangular structuring element kernel is selected to perform erosion and dilation operations to eliminate device difference interference. The repaired images and metadata are associated and encapsulated into a structured dataset using a semi-automatic annotation tool. The metadata includes acquisition device parameters, environmental parameters, and component attribute parameters. The acquisition device parameters are white balance and focal length parameters generated by a cross-device calibration protocol. The environmental parameters are the temperature, humidity, and light intensity during acquisition by the acquisition device. The component attribute parameters are the pipe diameter, pressure level, and material type identified by a pre-trained target detection model. The structured dataset is transmitted to the training module of a multi-task joint learning framework through distributed storage nodes.

[0007] Furthermore, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 2 includes: An end-to-end network architecture based on ResNet-50, HRNet, and Mask R-CNN is constructed, sharing the underlying convolutional feature extraction layer; During the training phase, adversarial augmentation data generated by a generative adversarial network is injected to simulate the morphological changes of pipeline components in oily and highly reflective environments. The model parameters are optimized in stages through a progressive training strategy, including basic training based on ImageNet pre-trained weights, domain fine-tuning that increases with environment complexity, and an online calibration stage that incorporates active learning. The trained model is distilled into a lightweight MobileNetV3 model, and the feature output of the lightweight MobileNetV3 model is aligned with the Faiss index space of the hierarchical component library by cosine similarity.

[0008] Furthermore, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 3 includes: fusing the 1024-dimensional feature vector output by the multi-task joint learning framework with the three-dimensional point cloud data through a feature splicing layer to form a component feature descriptor; Construct an ontology-based knowledge graph, define the mapping relationship between feature descriptors and knowledge nodes, and the knowledge graph includes the assembly relationship between valves and pipelines and the replacement rules for compatible models; Component retrieval is based on a multi-level index structure. Component retrieval includes locating the major category of components through a hash table, constructing an approximate nearest neighbor search index to match feature vectors through the Faiss framework, and matching spatial topological relationships with associated BIM models.

[0009] Furthermore, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 4 includes: Deploy incremental data pipelines to capture new images during on-site installation and maintenance, and filter samples that differ from the existing library by more than a set threshold through differential compression; Establish a model versioning and iteration framework, dynamically switch training model versions based on the decrease in false detection rate and response latency indicators of online inference, and retain traceable model snapshots; When the retrieval anomaly rate of the hierarchical component library exceeds a set threshold, the adversarial data augmentation module in the multi-task joint learning framework is triggered to generate targeted training samples for oily and highly reflective environments.

[0010] Furthermore, the pipeline component library creation method based on model training and image recognition technology of the present invention also includes: establishing a cross-device calibration protocol in the preprocessing stage, and unifying white balance and focal length parameters based on the device type of the acquisition terminal, wherein the device type of the acquisition terminal includes drones and handheld devices; The parameter configuration generated by the calibration protocol is input into the generator input of the generative adversarial network to constrain the material reflectivity parameters of the oil stain and strong reflection synthetic samples output by the generative adversarial network.

[0011] Furthermore, the pipeline component library creation method based on model training and image recognition technology of the present invention also includes: An attention-based feedback loop module is embedded in the search interface of the hierarchical component library. It generates a region heatmap based on the image region selected by the user, redirects the query vector of the approximate nearest neighbor search index, and obtains the corrected search results. The corrected search results are back-annotated to the structured dataset using a semi-automatic annotation tool, generating new training samples which are then input into the active learning module during the online calibration phase.

[0012] Furthermore, the pipeline component library creation method based on model training and image recognition technology of the present invention also includes: Establish a data lineage tracing link to record the acquisition device ID, preprocessing parameters, and model version identifier involved in training for each training sample in the structured dataset; By associating lineage links with model version snapshots, a timestamp matching mechanism is used to verify the training dependency of a specific model version on historical datasets.

[0013] Furthermore, the pipeline component library creation method based on model training and image recognition technology of the present invention also includes: Define data aging rules to perform cold storage archiving on low-frequency access component data that has not been retrieved for 30 consecutive days in the hierarchical component library; The storage tier strategy is dynamically adjusted based on the search popularity index statistically analyzed by the feedback loop module. High-frequency access data is migrated to SSD storage nodes, while low-frequency data is downgraded to HDD storage nodes.

[0014] Furthermore, the pipeline component library creation method based on model training and image recognition technology of the present invention also includes: Deploy an anomaly self-healing link in the incremental data pipeline to roll back to the retained historical model version and corresponding index snapshot when Faiss index reconstruction fails. The self-healing link triggers a cross-device calibration protocol to recalibrate the white balance parameters of the data acquisition source and generates calibration logs to synchronize data lineage tracing links.

[0015] Beneficial effects of this invention; This invention achieves significant beneficial effects through the following technical paths: First, by eliminating equipment differences and environmental interference through multispectral imaging equipment and a cascaded preprocessing pipeline, a standardized dataset is generated to replace manual annotation, greatly improving data acquisition efficiency. Second, by sharing feature layers and injecting adversarial data through a multi-task joint learning framework, component classification, localization, and defect detection tasks are simultaneously optimized, reducing the superposition of errors from multiple models and significantly lowering the recognition error rate under complex working conditions. Third, by dynamically updating model parameters and index structure through a closed-loop evolution mechanism, targeted training and version rollback are triggered by incremental data, achieving system self-correction capabilities. Fourth, a hierarchical component library is constructed, integrating multi-level indexes and knowledge graphs, and attention feedback mechanisms are used to redirect retrieval logic, enhancing adaptability to occlusion and contamination scenarios. Fifth, data lineage tracing and dynamic storage strategies achieve end-to-end traceability, meeting industrial data compliance requirements and systematically solving the technical defects of traditional methods, such as low efficiency, error accumulation, and poor environmental adaptability. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for creating a pipeline component library based on model training and image recognition technology, provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention. The technical solutions provided by various embodiments of this invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of this invention, it will be described in further detail below.

[0019] Please see Figure 1 This invention provides a method for creating a pipeline component library based on model training and image recognition technology, including: Step 1: Obtain image data of multi-source pipeline components and associated environmental parameters. Generate a standardized image dataset by performing cascaded preprocessing on the image data of multi-source pipeline components and associated environmental parameters. Image data of multi-source pipeline components is generated through multispectral imaging equipment, which simultaneously acquires visible and near-infrared information to form raw image data including depth channels. The acquisition process covers occlusion and low-light conditions, and the integrity of texture details is enhanced by adjusting the sensor incident angle and using multi-band fusion technology. After the raw data is input into a cascaded preprocessing pipeline, two key processing stages are executed sequentially: First, a multi-scale Retinex algorithm is used for illumination equalization, using a Gaussian kernel function to decompose the illumination and reflection components of the image, eliminating brightness distortion in shadow areas and interference from strong reflections on metal surfaces; then, a texture restoration operation based on morphological opening is performed, using a rectangular structuring element kernel to first perform erosion to remove isolated pixels caused by equipment noise, and then performing dilation to fill the texture fracture areas caused by corrosion, restoring the geometric contour features of the components.

[0020] The preprocessed image data and metadata are associated and encapsulated using a semi-automatic annotation tool. The metadata contains three types of key parameters: acquisition device parameters, generated by a cross-device calibration protocol to standardize white balance and focal length parameters for different acquisition terminals (drones, handheld devices); environmental parameters, recording physical quantities such as temperature, humidity, and light intensity during acquisition; and component attribute parameters, automatically identified by a pre-trained target detection model, including structured attributes such as pipe diameter, pressure rating, and material type. The encapsulation process uses a key-value mapping mechanism to bind the image data stream to the metadata header file, forming a standardized structured dataset.

[0021] The resulting structured dataset is then transmitted to the training module of the multi-task joint learning framework via distributed storage nodes after being sharded, compressed, and load-balanced. This process eliminates hardware differences through device calibration, overcomes environmental interference through cascaded processing, and reduces manual intervention through automatic annotation, providing a foundation of input data with illumination consistency, device independence, and structured attribute features for subsequent model training.

[0022] The above operations form a progressive technology chain: multispectral acquisition solves the problem of raw data integrity, cascaded preprocessing overcomes the challenge of eliminating environmental interference, metadata association and encapsulation realizes information structuring, and distributed transmission ensures data processing efficiency. The output of each stage serves as the input for the next stage, collectively achieving the goal of constructing a standardized image dataset.

[0023] Step 2: Input the standardized image dataset into the multi-task joint learning framework for dynamic training. The multi-task joint learning framework integrates the component classification sub-network, the component key point localization sub-network, and the component defect detection sub-network, and outputs the training model by sharing the underlying feature extraction layer. A standardized image dataset is input into a multi-task joint learning framework for dynamic training. This framework constructs an end-to-end network architecture, with its bottom-level feature extraction layer employing a convolutional neural network to implement a feature-sharing mechanism. This layer outputs a shared feature map to three parallel sub-networks: a component classification sub-network generates a component type probability distribution output; a component keypoint localization sub-network predicts spatial coordinates based on a high-resolution feature pyramid; and a component defect detection sub-network generates a pixel-level defect segmentation mask output through a region proposal mechanism. This feature-sharing mechanism reduces computational redundancy and avoids the error aggregation problem caused by independent training of multiple models.

[0024] Adversarial augmentation data is injected into the training process. This data is synthesized by a generative adversarial network (GAN) to simulate the morphological distortion characteristics of pipeline components under conditions of oil contamination and strong metal reflection. Model optimization employs a progressive training strategy, including: a basic training phase loading pre-trained weights from a large-scale general dataset to initialize parameters; a domain fine-tuning phase inputting domain-specific data incrementally according to the environmental complexity gradient; and an online calibration phase combining an active learning mechanism to screen low-confidence samples for manual verification and annotation. This strategy optimizes the shared feature extraction layer parameters in stages, enabling the model to adapt to complex operating conditions.

[0025] After dynamic training is completed, the framework outputs the trained model, whose feature extraction capabilities are adapted to the subsequent feature descriptor generation requirements. The entire process forms an efficient training chain through feature sharing and progressive optimization. The input data is standardized and then fed into the framework. The framework structure supports multi-task synchronous training, the training strategy enhances the model's robustness, and the final output model serves the goal of building a hierarchical component library.

[0026] Step 3: Based on the feature vectors extracted by the training model, generate component feature descriptors, associate the component feature descriptors with the component assembly relationships in the preset knowledge graph, construct a multi-level index structure, receive query input including the image or feature vector of the component to be retrieved, parse the component type identifier in the query, determine the target component category, perform an approximate nearest neighbor search in the target category, output candidate components with similar features, verify the spatial assembly compatibility between the candidate components and the query components, filter the final matching results that pass the topological constraints, use the final matching results as the structured retrieval results, use the structured retrieval results as the data input source of the closed-loop evolution mechanism, and output them to Step 4 in real time; The feature vectors extracted by the training model are fused with 3D point cloud data through a feature concatenation layer to form a component feature descriptor containing spatial geometric information and texture features. This descriptor is mapped to a low-dimensional vector space using graph embedding technology, serving as the feature representation of entity nodes in a predefined knowledge graph. The knowledge graph defines the assembly relationship logic between components based on ontology, including the thread specification matching rules for valves and pipes and the compatibility and replacement constraints of components made of different materials, establishing a reasonable assembly knowledge base.

[0027] The multi-level index structure employs a hierarchical retrieval mechanism to achieve efficient querying: First, the component type identifier in the input query is parsed, and the target component category is quickly located using a hash table index; then, within the target category, the feature vector is matched using an approximate nearest neighbor search index constructed using the Faiss framework, and a set of candidate components whose cosine similarity meets the threshold is output; finally, the spatial topological relationship data of the building information model is associated to verify the assembly compatibility between the candidate components and the query components, including geometric constraints such as axis alignment and minimum gap size, and the final matching results that pass the topological verification are selected.

[0028] Structured retrieval results serve as the data input source for the closed-loop evolution mechanism, transmitted in real time to subsequent steps. The entire process forms a progressive technology chain: feature descriptors integrate multi-source information to enhance representation capabilities, knowledge graph associations endow assembly semantics, multi-level indexes enable hierarchical retrieval, and topology verification ensures spatial compatibility. The output of each step serves as the input for the next step: feature extraction supports descriptor generation, knowledge graphs support index construction, multi-level retrieval achieves accurate matching, and finally, structured retrieval results with assembly feasibility are output. This technology chain systematically solves the problem of the lack of spatial compatibility verification in traditional methods, providing reliable data input for closed-loop evolution.

[0029] Step 4: When the feature similarity score in the structured retrieval results is lower than the set threshold, it is marked as a false detection sample. The corrected label and associated environmental parameters of the false detection sample are extracted to generate incremental data. Based on the incremental data, a closed-loop evolution mechanism is triggered. The closed-loop evolution mechanism includes acquiring incremental data, performing cascaded preprocessing on the incremental data to generate an incremental standardized dataset, inputting the incremental standardized dataset into a multi-task joint learning framework, updating the feature extraction parameters of the training model, extracting feature vectors based on the updated training model, reconstructing the multi-level index structure of the hierarchical component library, and feeding back the retrieval results of the updated hierarchical component library to Step 1 to form a closed-loop iteration.

[0030] When the feature similarity score of the structured search results is lower than a set threshold, the system automatically marks them as false positives and extracts their corrected labels and environmental parameters to generate an incremental dataset. This dataset contains new samples from on-site maintenance scenarios and their associated acquisition environment metadata. Samples with significant feature differences from the existing database are filtered out using a differential compression algorithm. The incremental data capture process is deployed on engineering terminal equipment to monitor data changes in real time during installation and maintenance.

[0031] After the incremental data triggers the closed-loop evolution mechanism, it is first input into a cascaded preprocessing pipeline to perform a standardization transformation, maintaining technical consistency with the preprocessing operation in step 1. The processed incrementally standardized dataset is then input into a multi-task joint learning framework to perform incremental training, focusing on updating the parameter distribution of the shared convolutional feature extraction layer. The training process inherits the online calibration phase of the progressive strategy, prioritizing high-value incremental samples through an active learning mechanism.

[0032] Feature vectors are re-extracted based on the updated training model, and the multi-level index structure of the hierarchical component library is reconstructed. Index reconstruction includes version updates of the hash table classification index, vector remapping of the Faiss feature index space, and synchronous refresh of the building information model topology relation library. The model versioning iteration framework retains historical version snapshots and supports rapid rollback capabilities in abnormal situations.

[0033] The updated hierarchical component library transmits the latest search results as feedback signals to the data acquisition module in step 1. This feedback guides the multispectral imaging equipment to adjust its acquisition strategy and optimize the environmental coverage of subsequent data acquisition processes. This ultimately forms a complete closed-loop iteration from search result analysis to data acquisition optimization, enabling the system to adaptively evolve to data distribution shifts. Each stage triggers update operations through incremental data, and the update results inversely optimize the front-end acquisition, forming a self-optimizing technical loop chain.

[0034] In step 1, multi-source pipe component image data is acquired using a multispectral imaging device. This device simultaneously acquires visible and near-infrared band information to generate raw image data containing depth channels. A cascaded preprocessing pipeline sequentially performs illumination equalization and morphology restoration operations: first, the multi-scale Retinex algorithm is used to decompose the illumination and reflection components of the image, and a Gaussian kernel function is used to eliminate brightness distortion in shadow and reflective areas; then, based on morphological opening operations, a rectangular structuring element kernel is used to perform erosion operations to remove equipment noise, and finally, a dilation operation is used to repair texture breaks caused by corrosion. The preprocessed image and metadata are associated and encapsulated using a semi-automatic annotation tool. The metadata includes acquisition parameters (white balance, focal length), environmental parameters (temperature, humidity, light intensity) generated through a cross-device calibration protocol, and component attribute parameters (pipe diameter, material type) identified by the pre-trained model. The final generated structured dataset is transmitted to the training module through distributed storage nodes, providing the model with device-independent and environment-robust input data.

[0035] In step 2, the multi-task joint learning framework adopts an end-to-end network architecture, sharing a convolutional feature extraction layer implemented by ResNet-50 at its bottom layer. This framework connects three sub-networks in parallel: a component classification sub-network outputs the category probability distribution; a keypoint localization sub-network predicts the spatial coordinates of components based on the HRNet structure; and a defect detection sub-network generates pixel-level segmentation masks using Mask R-CNN. Adversarial augmentation data is injected during the training phase. This data is synthesized by a generative adversarial network to simulate morphological distortion features under conditions of oil stain adhesion and strong metal reflection. Model optimization employs a progressive training strategy: the first stage loads ImageNet pre-trained weights to initialize parameters; the second stage incrementally increases the input data according to the environmental complexity gradient to achieve domain-adaptive fine-tuning; the third stage introduces an active learning mechanism, screening low-confidence samples for manual review and annotation, and dynamically correcting the decision boundary through online calibration. The trained model is compressed into a lightweight MobileNetV3 model using knowledge distillation technology, and its feature output vector maintains cosine similarity alignment with the index space of the subsequent retrieval module.

[0036] In step 3, the component feature descriptor is generated by fusing the 1024-dimensional vector extracted from the training model with 3D point cloud data through a feature stitching layer. This descriptor is mapped to entity nodes of a pre-defined knowledge graph using graph embedding technology. The knowledge graph defines the assembly relationships between components (such as the thread specification matching rules for valves and pipes) and compatible model replacement rules based on ontology. The multi-level index structure adopts a hierarchical retrieval mechanism: first, the target component category is quickly located using a hash table; second, within the target category, the approximate nearest neighbor search index constructed using the Faiss framework is used to match feature vectors, outputting candidate components with the highest cosine similarity; finally, the spatial topological relationship data of the BIM model is associated to verify the assembly compatibility (including constraints such as axis alignment and minimum gap size) between the candidate components and the query components, and the final matching results that pass the verification are selected. The structured retrieval results are output to the closed-loop evolution module in real time as a data input source for system self-optimization.

[0037] In step 4, when the feature similarity score of the structured retrieval results is lower than a set threshold, the system automatically marks them as false positives and extracts their correction labels and environmental parameters to generate an incremental dataset. After the closed-loop evolution mechanism is activated, the incremental data is transformed into a standardized dataset through a cascaded preprocessing pipeline, input into a multi-task joint learning framework for incremental training, and updates the parameters of the shared feature extraction layer. Based on the updated model, feature vectors are re-extracted, and the multi-level index structure of the hierarchical component library (including hash table classification index, Faiss feature index, and BIM topology relation library) is reconstructed. The reconstructed hierarchical component library feeds back the latest retrieval results to the data acquisition and preprocessing module of step 1, forming a closed-loop iterative optimization process from data acquisition and model training to retrieval verification, realizing the system's adaptive capability to environmental changes and data drift.

[0038] The above steps form a closed-loop technology: Step 1, standardized data processing, provides high-quality input for model training; Step 2, multi-task joint learning, improves model robustness through feature sharing and adversarial training; Step 3, multi-level indexing combined with knowledge graphs, achieves accurate retrieval; Step 4, based on retrieval feedback, triggers incremental training and index reconstruction, enabling continuous system optimization. Each step is tightly integrated through data flow and control flow, ultimately achieving the efficient construction and dynamic evolution of the pipeline component library.

[0039] Specifically, the method for creating a pipe component library based on model training and image recognition technology of the present invention includes step 1: acquiring images of pipe components including depth information through a multispectral imaging scheme to generate raw data of covered, occluded, and low-light scenes; The raw data is input into the cascaded processing pipeline. The cascaded preprocessing pipeline sequentially performs illumination equalization processing based on the multi-scale Retinex algorithm, decomposes the illumination and reflection components of the image using the Gaussian kernel function, and performs texture restoration based on morphological opening operation to obtain the restored image. The rectangular structuring element kernel is selected to perform erosion and dilation operations to eliminate device difference interference. The repaired images and metadata are associated and encapsulated into a structured dataset using a semi-automatic annotation tool. The metadata includes acquisition device parameters, environmental parameters, and component attribute parameters. The acquisition device parameters are white balance and focal length parameters generated by a cross-device calibration protocol. The environmental parameters are the temperature, humidity, and light intensity during acquisition by the acquisition device. The component attribute parameters are the pipe diameter, pressure level, and material type identified by a pre-trained target detection model. The structured dataset is transmitted to the training module of a multi-task joint learning framework through distributed storage nodes.

[0040] The environmental robustness preprocessing technology of this invention is implemented through the following process: First, images of pipe components are acquired using a multispectral imaging device. This device integrates visible light and near-infrared sensors to simultaneously acquire RGB color information and depth data, generating an original image including spatial three-dimensional coordinates. During the acquisition process, for scenarios where oil contamination or mechanical obstruction may exist on the surface of the pipe components, the incident angle and exposure parameters of the multispectral sensor are adjusted. Multi-band fusion technology is used to enhance texture details in low-light areas, forming an initial dataset covering complex working conditions.

[0041] After the raw data is input into the cascaded processing pipeline, the first processing stage uses the Retinex algorithm to decompose the illumination and reflection components of the image. Gamma correction and histogram equalization are then used to eliminate brightness distortion in shadow and reflective areas. The second processing stage constructs a structuring element kernel based on morphological opening operations. First, an erosion operation is performed to remove isolated pixels caused by equipment noise. Then, a dilation operation is used to fill in texture breaks caused by corrosion, restoring the geometric integrity of the elements. The image data after these two stages of processing is mapped to the white balance and focal length parameters of the acquisition device, eliminating imaging deviations caused by hardware differences between drones and handheld devices.

[0042] The repaired image data is input into a semi-automatic annotation tool. This tool loads a pre-trained object detection model to generate bounding box annotations for key components such as flange interfaces and welds. Simultaneously, it analyzes the pipe diameter, pressure rating parameters, and environmental temperature and humidity data collected during acquisition. The annotation information is linked to the image data via key-value pairs and encapsulated into a structured dataset including a metadata header file and a binary image stream. This structured dataset is then fragmented, compressed, and load-balanced across distributed storage nodes, and transmitted in a standardized format to the training module of a multi-task joint learning framework. This provides illumination-consistent and device-independent input data for subsequent model training.

[0043] Specifically, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 2 includes: An end-to-end network architecture based on ResNet-50, HRNet, and Mask R-CNN is constructed, sharing the underlying convolutional feature extraction layer; During the training phase, adversarial augmentation data generated by a generative adversarial network is injected to simulate the morphological changes of pipeline components in oily and highly reflective environments. The model parameters are optimized in stages through a progressive training strategy, including basic training based on ImageNet pre-trained weights, domain fine-tuning that increases with environment complexity, and an online calibration stage that incorporates active learning. The trained model is distilled into a lightweight MobileNetV3 model, and the feature output of the lightweight MobileNetV3 model is aligned with the Faiss index space of the hierarchical component library by cosine similarity.

[0044] The multi-task joint learning framework of this invention is implemented through the following technical solution: First, a backbone network based on ResNet-50 is constructed, and its bottom convolutional layer outputs are shared with the HRNet branch network and the Mask R-CNN branch network. The HRNet branch network extracts the spatial coordinates of key points of pipeline elements through a high-resolution feature pyramid, and the Mask R-CNN branch network generates pixel-level segmentation masks of defect parts based on a region proposal mechanism. The three networks achieve feature reuse for classification, localization and detection tasks on the basis of sharing the convolutional layer outputs, thereby reducing the computational redundancy of multi-task parallel operation.

[0045] During the training phase, a Generative Adversarial Network (GAN) is introduced to generate adversarial augmented data. The generator of the GAN receives white balance and focal length parameters output from a cross-device calibration protocol as input. Combined with a reflectivity model of oil-stained materials, it generates synthetic samples that conform to physical laws, simulating the morphological distortion characteristics of pipe components under oil-stained and highly reflective metal surface conditions. The adversarial data is mixed with the original standardized dataset at a preset ratio and then input into a multi-task network. The adversarial training mechanism improves the model's generalization ability to complex working conditions.

[0046] The model optimization employs a progressive training strategy. In the initial stage, ImageNet pre-trained weights are loaded to initialize the parameters of the shared convolutional layers, leveraging the feature extraction capabilities of large-scale general datasets to establish a basic model. The second stage involves inputting domain data in an incremental manner according to environmental complexity. Prioritizing images of clean components in unobstructed scenes, the input gradually transitions to samples with partial occlusion and light contamination, finally introducing heavy oil stains and strong reflection data to gradually adapt the model parameters to the complex environmental distribution of the engineering site. The third stage deploys an active learning mechanism to screen problematic samples with prediction confidence levels below a threshold. These samples are then manually reviewed and labeled before being added to the training set, and the model's decision boundaries are dynamically corrected through online calibration.

[0047] The trained model is compressed into a lightweight MobileNetV3 model using knowledge distillation. A ResNet-50 multi-task network is used as the teacher model. A feature alignment loss function constrains the output features of the MobileNetV3 student model to align with the intermediate feature responses of the teacher model in shared convolutional layers. Simultaneously, a cosine similarity loss function is introduced to ensure that the feature vectors output by the student model maintain directional consistency with the vector distribution in the Faiss index space of the hierarchical component library. This distilled lightweight model is adapted to the deployment requirements of edge computing devices, supports real-time feature extraction and efficient matching with the index space, and solves the retrieval latency problem in low-computing-power environments at engineering sites.

[0048] Specifically, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 3 includes: fusing the 1024-dimensional feature vector output by the multi-task joint learning framework with the three-dimensional point cloud data through a feature splicing layer to form a component feature descriptor; Construct an ontology-based knowledge graph, define the mapping relationship between feature descriptors and knowledge nodes, and the knowledge graph includes the assembly relationship between valves and pipelines and the replacement rules for compatible models; Component retrieval is based on a multi-level index structure. Component retrieval includes locating the major category of components through a hash table, constructing an approximate nearest neighbor search index to match feature vectors through the Faiss framework, and matching spatial topological relationships with associated BIM models.

[0049] The hierarchical component library construction technology of this invention is implemented through the following process: a 1024-dimensional feature vector output by a multi-task joint learning framework and point cloud data acquired by 3D laser scanning are input into a feature concatenation layer. After normalizing the vector dimension, the feature concatenation layer performs a cascade operation along the channel dimension to generate a 1280-dimensional component feature descriptor that integrates spatial geometric information and texture features. The feature descriptor is mapped to a low-dimensional vector space through graph embedding technology, serving as the initial feature representation of knowledge graph nodes.

[0050] The knowledge graph, built upon ontology, uses the OWL language to define the class hierarchy of pipe components. This hierarchy includes component types such as valves, flanges, and fittings, and extends SWRL rules to describe the assembly constraints between components. The knowledge graph associates the embedding vectors of feature descriptors with the entity nodes, calculates the semantic similarity between nodes using a graph attention mechanism, establishes matching rules for valve and pipe thread specifications, and establishes compatible substitution relationships for components of different materials, forming a reasonable assembly knowledge base.

[0051] The multi-level index structure employs a hierarchical retrieval mechanism to improve query efficiency: the hash table index quickly locates the target component category based on the first letter hash value of the component type name, narrowing the search scope; the Faiss framework constructs an inverted index and product quantizer for feature descriptors, and matches candidate components with the highest cosine similarity through an approximate nearest neighbor search algorithm; the spatial topological relationship data provided by the BIM model and the assembly rules of the knowledge graph are jointly verified to select the final retrieval results that simultaneously meet feature similarity and spatial compatibility. The hierarchical design of the multi-level index structure balances retrieval speed and accuracy, supporting the real-time component matching and assembly scheme verification needs on the engineering site.

[0052] Specifically, in the pipeline component library creation method based on model training and image recognition technology of the present invention, step 4 includes: Deploy incremental data pipelines to capture new images during on-site installation and maintenance. Calculate the cosine similarity between the new images and the feature vectors of existing library samples through differential compression. Filter samples with a difference lower than a preset similarity threshold. The similarity threshold is dynamically adjusted based on historical false detection rates. Establish a model versioning and iteration framework, dynamically switch training model versions based on the decrease in false detection rate and response latency indicators of online inference, and retain traceable model snapshots; When the retrieval anomaly rate of the hierarchical component library exceeds a set threshold, the adversarial data augmentation module in the multi-task joint learning framework is triggered to generate targeted training samples for oily and highly reflective environments.

[0053] The closed-loop evolution mechanism of this invention is implemented through the following technical solution: A lightweight data acquisition agent is deployed at the pipeline engineering terminal to capture real-time images generated during installation and maintenance. An inter-frame difference algorithm is used to calculate the cosine similarity of the feature vectors of new images with existing samples in a hierarchical component library. Samples with differences exceeding a preset threshold are filtered out. Block compression coding technology is used to reduce the amount of data transmitted over the network. The filtered incremental data is preprocessed by edge computing nodes and then input into the training queue of the model versioning iteration framework, where it is mixed with the historical training set according to timestamps to construct an incremental dataset.

[0054] The model versioning iteration framework maintains multiple parallel model version repositories, with each version associated with the data distribution and hyperparameter configurations used during training. The online inference service continuously monitors the false positive rate (FPR) decline rate and response latency percentiles for each model version. When the FPR decline rate of a new model version exceeds 20% of the historical baseline while latency remains within acceptable engineering thresholds, a version switch command is triggered, updating the production environment inference model to the optimal version. All historical model weights and corresponding training data snapshots are stored in a distributed file system, supporting rapid rewinding to any historical state by version hash value.

[0055] The anomaly rate of the hierarchical component library is calculated by the proportion of cases where feature matching is successful but assembly verification fails within a statistical period. When the anomaly rate exceeds a set threshold for three consecutive statistical periods, the adversarial data augmentation module is automatically invoked. Combining the material reflectivity parameters in the cross-device calibration protocol, the texture features of the oil-stained area and the illumination parameters of the highly reflective area are directionally perturbed in the latent space of the generative adversarial network, and training samples with targeted morphological distortions are synthesized in batches. The generated adversarial samples are injected into the online calibration stage of the multi-task joint learning framework. Through an active learning mechanism, difficult samples are trained first, and the optimized model parameters are synchronously updated to the Faiss index space of the hierarchical component library, forming a synergistic evolution of feature extraction and retrieval optimization.

[0056] Specifically, the method for creating a pipeline component library based on model training and image recognition technology of the present invention further includes: establishing a cross-device calibration protocol in the preprocessing stage, and unifying white balance and focal length parameters based on the device type of the acquisition terminal, including drones and handheld devices. The parameter configuration generated by the calibration protocol is input into the generator input of the generative adversarial network to constrain the material reflectivity parameters of the oil stain and strong reflection synthetic samples output by the generative adversarial network.

[0057] The cross-device calibration and adversarial data generation technology solution of this invention is implemented through the following process: For two types of acquisition terminals—UAVs and handheld devices—a mapping relationship library between device type identifiers and imaging parameters is established. This library stores the default white balance coefficients and focal length parameter ranges for different device models. During calibration, imaging data of the devices under different lighting conditions is collected using a checkerboard calibration board. A nonlinear least squares method is used to fit the correction matrix of the white balance parameters and focal length parameters, generating a device-specific calibration configuration file. This eliminates white balance shifts caused by high-altitude light scattering in UAV aerial images and focal length distortion caused by close-range shooting from handheld devices.

[0058] The calibration configuration file is input into the generator module of the generative adversarial network (GAN). The generator's conditional input layer receives calibration parameters as prior constraints. Combining the physical model of the reflectivity of oily materials with the Fresnel reflection equation for metal surfaces, a synthetic texture conforming to the laws of optical propagation is generated in the latent space. The synthetic image output by the generator is used for adversarial training against real oily samples through a discriminator network. The generator parameters are iteratively optimized to ensure that the reflectivity distribution of the synthetic sample remains physically consistent with the real data collected by the calibration device. The constrained synthetic sample is then injected into the training dataset of a multi-task joint learning framework to improve the model's feature recognition ability for oily materials and strong reflection interference in cross-device acquisition scenarios, thereby enhancing the model's generalization performance in different hardware environments.

[0059] Specifically, the pipeline component library creation method based on model training and image recognition technology of the present invention further includes: An attention-based feedback loop module is embedded in the retrieval interface of the hierarchical component library. The pixel coordinates of the user-selected area are extracted based on the spatial attention mechanism, a region heatmap is generated through the Gaussian kernel function, and the query vector channel weights of the approximate nearest neighbor search index are adjusted by weighting. The corrected search results are back-annotated to the structured dataset using a semi-automatic annotation tool, generating new training samples which are then input into the active learning module during the online calibration phase.

[0060] The feedback loop and active learning collaborative technology solution of this invention is implemented through the following process: A visual interaction module based on spatial attention mechanism is deployed in the retrieval interface of a hierarchical component library. Users trigger feedback signals by selecting mismatched regions in the image. The module extracts the pixel coordinates of the selected regions and calculates their spatial relevance weights with the original query vector, generating a region heatmap reflecting the user's attention. After smoothing the heatmap using a Gaussian kernel function, a channel-weighted operation is performed on the query vector of the approximate nearest neighbor search index to enhance the representation weight of the target region features in the vector space, thereby redirecting the retrieval logic to the local features of the components of interest to the user.

[0061] The corrected search results are input into a semi-automatic annotation tool. The tool loads a pre-trained semantic segmentation model to refine the boundaries of the correctly identified components, simultaneously parsing the pipe diameter parameters and assembly orientation information of the components. The annotated data is associated with the original image and environmental parameters through key-value mapping, generating new training samples and marking them as high-priority data, which are then input into the active learning queue during the online calibration phase. The active learning module uses an uncertainty sampling strategy to select samples with the highest prediction variance, prioritizing their submission to the manual review interface for secondary verification. Verified data is then injected into the incremental training process of the multi-task joint learning framework, dynamically adjusting the model's decision boundaries and the parameter distribution of the feature extraction layer.

[0062] The feedback data stream and model training form a closed-loop optimization mechanism: the retrieval correction results driven by user interaction behavior directly affect the generation of labeled training data. The labeled data improves the model's ability to identify difficult samples through an active learning strategy. The optimized model parameters synchronously update the Faiss index space of the hierarchical component library, thereby improving the subsequent retrieval accuracy and reducing the frequency of user feedback triggering, achieving continuous performance improvement of the system through adaptation.

[0063] Specifically, the pipeline component library creation method based on model training and image recognition technology of the present invention further includes: Establish a data lineage tracing link. In the metadata header file of each training sample in the structured dataset, record the acquisition device ID, gamma correction coefficient of the Retinex algorithm, and kernel size parameters of the structuring element of the morphological opening operation. And associate the model version identifiers participating in the training through key-value mapping. By associating lineage links with model version snapshots, a timestamp matching mechanism is used to verify the training dependency of a specific model version on historical datasets.

[0064] The data lineage tracking and version management technology solution of this invention is implemented through the following process: Each training sample in the structured dataset is injected with metadata header information during the preprocessing stage. The metadata header includes the model serial number of the acquisition device, the gamma correction coefficient for illumination equalization in the Retinex algorithm, the kernel size parameter of the structuring element used in the morphological opening operation, and the specific version hash value involved in model training. Lineage information is recorded through the log service of a distributed storage engine, forming a chained data structure with timestamps as the primary key, supporting reverse tracing of its entire lifecycle processing trajectory by sample ID.

[0065] Model version snapshots are stored in a version control repository. Each snapshot is associated with a training start timestamp, hyperparameter configuration, and the dataset version identifier used. When it is necessary to verify the dependency of a specific model version on historical data, the lineage tracking service uses a timestamp interval matching algorithm to retrieve the acquisition device parameters and preprocessing operation records of all training samples involved in that version's training cycle, generating a data distribution consistency report. This report compares the differences in device parameter offsets and preprocessing procedures between historical datasets and current production environment data, providing a quantitative assessment of the degree of data drift for model version rollback decisions.

[0066] The association mechanism between lineage links and version snapshots supports engineering auditing needs: when a retrieval anomaly occurs in the hierarchical component library, the defective data batch used during the training of the abnormal model version can be quickly located. Combined with preprocessing parameters, the data can be traced back to the original acquisition device for calibration, forming a full-chain problem tracing capability from data acquisition and model training to retrieval services, meeting the compliance requirements of industrial data quality management standards for traceability.

[0067] Specifically, the pipeline component library creation method based on model training and image recognition technology of the present invention further includes: Define data aging rules to perform cold storage archiving on low-frequency access component data that has not been retrieved for 30 consecutive days in the hierarchical component library; The storage tier strategy is dynamically adjusted based on the search popularity index statistically analyzed by the feedback loop module. High-frequency access data is migrated to SSD storage nodes, while low-frequency data is downgraded to HDD storage nodes.

[0068] The data lifecycle management solution of this invention is implemented through the following process: The access log recording module of the hierarchical component library continuously collects the retrieval timestamps and access frequencies of each component's data. Based on the sliding time window algorithm, it statistically analyzes low-frequency access data that has not been accessed for 30 consecutive days, generating a list of data to be archived. The cold storage archiving operation uses block compression encoding technology to perform lossless compression on the data in the list, attaches a metadata header file to record the original storage path and compression parameters, and migrates it to the archive storage pool of the distributed object storage system, releasing storage resources of the online retrieval database.

[0069] The search popularity metric is calculated using user search behavior data collected by the feedback loop module. The metric incorporates search frequency, a time-weighted factor, and the number of user-annotated corrections to construct a popularity scoring model. The dynamic storage scheduler periodically scans high-frequency data with popularity scores exceeding a set threshold. A data block verification mechanism is used to migrate complete data copies from HDD storage nodes to SSD storage nodes and update the index address mapping table. For low-frequency data with popularity scores below the threshold for three consecutive scheduling cycles, a degraded storage operation is performed, retaining the metadata index but migrating the actual data blocks back to the HDD storage node. This maintains the availability of the search function while reducing the occupancy of high-speed storage media.

[0070] Data aging rules and storage tier strategies form a collaborative optimization mechanism: cold storage archiving reduces the interference of redundant data in online databases on retrieval efficiency, and dynamic migration of storage tiers adjusts data distribution according to real-time access patterns, ensuring that high-frequency data resides on low-latency storage media, thereby improving overall system response speed and storage resource utilization. Archived data retains metadata indexes to support historical traceability needs, and can still be restored to the online database through a triggered loading mechanism in engineering change scenarios.

[0071] Specifically, the pipeline component library creation method based on model training and image recognition technology of the present invention further includes: Deploy an anomaly self-healing link in the incremental data pipeline to roll back to the retained historical model version and corresponding index snapshot when Faiss index reconstruction fails. The self-healing link triggers a cross-device calibration protocol to recalibrate the white balance parameters of the data acquisition source and generates calibration logs to synchronize data lineage tracing links.

[0072] The anomaly self-healing and data calibration collaborative technology solution of this invention is implemented through the following process: The monitoring service of the incremental data pipeline detects the return status codes of the Faiss index reconstruction operation in real time. When three consecutive reconstruction requests return exception codes indicating resource allocation errors or vector dimension mismatches, the self-healing link is triggered to start the rollback mechanism. The rollback mechanism accesses the historical model version snapshots and corresponding index shard data stored in the version control repository, preferentially selecting the most recently successfully built index version. After verifying the integrity of the snapshot through hash verification, the retrieval service in the production environment is switched to this version, restoring the system's basic retrieval function.

[0073] The self-healing link synchronously sends device calibration instructions to the edge-side data acquisition terminal. These instructions carry metadata characteristics, including the acquisition task identifier for the fault period and the anomaly index. The acquisition terminal loads the cross-device calibration protocol according to the instructions, re-executes the image acquisition process of the checkerboard calibration board, uses nonlinear least squares to fit the white balance parameters under the current ambient lighting conditions, and updates the focal length and white balance calibration matrices in the device configuration file. The newly generated calibration parameters are transmitted back to the central server via an encrypted channel, overwriting the original calibration configuration file.

[0074] The specific implementation process of this invention is as follows: This invention acquires image data of pipeline components using a multispectral imaging device. The device integrates visible light and near-infrared sensors, simultaneously acquiring RGB color information and depth data to generate a raw image with a resolution of 1920×1080 pixels. For occluded scenes, the sensor's incident angle is adjusted to 45 degrees, and the exposure time is set to 1 / 60 second. Multi-band fusion enhances texture details in low-light areas. The acquired raw data is input into a cascaded processing pipeline. First, the Retinex algorithm is used to decompose the image's illumination components, with a gamma correction coefficient of 2.2 and a histogram equalization parameter of 256 gray levels. Then, morphological opening is performed, using a 5×5 rectangular structuring element kernel. Three erosion operations are performed to remove noise points, followed by two dilation operations to repair texture breaks. The processed image and metadata (including device model ID, ambient temperature and humidity) are associated using a semi-automatic annotation tool, encapsulated into a structured dataset including a JSON metadata header and a PNG image stream, and transmitted to the training module after being fragmented and compressed by distributed storage nodes.

[0075] A ResNet-50-based backbone network was constructed, with its bottom convolutional layer outputs shared with the HRNet and Mask R-CNN branches. The HRNet branch retained high-resolution feature maps (1 / 4 the size of the original image) and output keypoint coordinates; the Mask R-CNN branch generated pixel-level masks of defective regions with a resolution of 56×56. During training, adversarial data synthesized by a Generative Adversarial Network (GAN) was injected. The GAN generator was fed calibrated white balance parameters (color temperature range 5500-6500K) and combined with an oil stain material reflectivity model (reflectance coefficient 0.3-0.7) to generate synthetic images containing oil stains and metallic reflections. These were then mixed with the original data at a 1:3 ratio for training. Progressive training consisted of three stages: basic training used ImageNet pre-trained weights and input 5000 clean component images; the domain fine-tuning stage inputted samples with 10%, 30%, and 50% occlusion in three batches; and the online calibration stage actively learned and filtered samples with confidence scores below 0.85, manually labeled them, and added them to the training set. The trained model is distilled to generate a lightweight MobileNetV3 model, which outputs a 512-dimensional feature vector and has a cosine similarity threshold of 0.92 with the Faiss index space.

[0076] The 512-dimensional feature vector output from MobileNetV3 is fused with 3D point cloud data (accuracy ±0.1mm) through a feature concatenation layer to generate a 768-dimensional feature descriptor. A knowledge graph is constructed based on the OWL language, defining the thread specifications (e.g., G1 / 2 to G2 inches) and material compatibility rules for valves and pipes (stainless steel and carbon steel cannot be mixed). In the multi-level index structure, the hash table is partitioned by the first letter of the component type (26 buckets); the Faiss index uses the IVF-PQ algorithm, setting the inverted index number to 100 and the quantization bit depth to 8; the spatial topology verification of the BIM model includes minimum assembly clearance (≥3mm) and axis alignment error (≤1°).

[0077] Incremental data pipelines are deployed at engineering terminals to capture on-site images in real time and filter samples with feature differences exceeding 15% using an inter-frame differencing algorithm. A model version repository maintains five historical versions, and the online inference service calculates the false positive rate (target value ≤2%) and response latency (≤200ms) hourly. When the retrieval anomaly rate exceeds 5% for three consecutive days, a GAN is triggered to generate 5000 targeted synthetic samples, which are then injected into the online calibration module. When the anomaly self-healing link detects a failure to rebuild the Faiss index (error code E1024), it rolls back to the best version from the last three days and recalibrates the white balance parameters of the acquisition equipment (error tolerance ±150K), with calibration logs synchronized to the data lineage link.

[0078] Data lineage tracking records the acquisition device ID (e.g., DJI_Mavic3_No.01), preprocessing parameters (gamma value 2.2), and model version hash value (64-bit) for each sample. Low-frequency data (not accessed for 30 consecutive days) is archived to cold storage after ZSTD compression (compression ratio 1:3); high-frequency data (daily access ≥100 times) is migrated to SSD storage nodes. The feedback loop module generates a heatmap (Gaussian kernel σ=5px) based on the user-selected area, redirects the retrieval vector weights (increasing local feature weights by 30%), corrects the result annotations, generates new training samples, and marks them as priority P1. The model iteration update is completed within 48 hours.

[0079] Explanation of technical features and terminology of this invention: Multispectral imaging scheme: refers to image acquisition technology that integrates visible light and near-infrared sensors to simultaneously acquire RGB color and depth information. By using multi-band fusion, it enhances texture details in low-light areas and generates raw data including three-dimensional spatial coordinates, thus solving the data acquisition integrity requirements in complex occlusion scenarios.

[0080] Cascaded preprocessing pipeline: including serialization modules for Retinex illumination equalization and morphological opening operations: the former eliminates shadows and reflection interference by decomposing the illumination / reflection components of the image; the latter uses erosion and dilation operations to remove equipment noise and repair rust and fracture textures, achieving standardization of imaging quality across devices.

[0081] Multi-task joint learning framework: Based on the ResNet-50 backbone network sharing low-level convolutional features, it coordinates the high-resolution key point localization of HRNet and the defect detection branch of Mask R-CNN. By reusing features, it reduces computational redundancy and simultaneously optimizes the model accuracy and efficiency of classification, localization and detection tasks.

[0082] Adversarial data augmentation: By injecting calibration parameter constraints into the input using a generative adversarial network (GAN), a composite image of oil stain reflection and strong metal reflection with physical plausibility is generated to simulate morphological distortion under extreme working conditions and improve the model's generalization ability to complex environments.

[0083] Progressive training strategy: Parameter optimization in three stages: Basic training stage: Load ImageNet pre-trained weights to establish general feature extraction capability; Domain fine-tuning stage: Input data according to the environmental complexity gradient; Online calibration stage: Actively learn to screen low-confidence samples for manual review and dynamically correct decision boundaries.

[0084] Knowledge distillation compression: The feature responses of the multi-task teacher model (ResNet-50) are transferred to the lightweight student model (MobileNetV3) through the cosine similarity loss function, compressing the model size while maintaining the alignment of features in the Faiss index space, and adapting to the needs of edge computing deployment.

[0085] Hierarchical component library construction: By fusing image feature vectors and 3D point cloud data through a feature splicing layer to generate composite descriptors, and combining them with knowledge graph node mapping relationships defined by ontology (such as valve-pipe thread specification matching), a reasonable knowledge base including assembly rules and compatibility relationships is constructed.

[0086] Multi-level index structure: The three-level retrieval mechanism adopts a hash table for fast location of component categories, Faiss framework for approximate nearest neighbor search, and BIM model spatial topology verification, which takes into account both retrieval speed (hash table O(1) time complexity) and accuracy (BIM constraint verification) to realize the real-time matching requirements of engineering scenarios.

[0087] Closed-loop evolution mechanism: The incremental data pipeline uses the inter-frame difference algorithm to filter samples with differences exceeding the threshold to trigger model updates. The versioned iteration framework dynamically switches the optimal model based on the false detection rate / latency index. The anomaly self-healing link rolls back the historical version and recalibrates the device parameters when the index reconstruction fails, forming a self-optimizing system.

[0088] Data lineage tracing: Records the entire lifecycle information of samples (collection device ID, preprocessing parameters, model version identifier), verifies the consistency of data distribution through timestamp matching, supports rapid tracing of abnormal model versions and quantitative assessment of data drift, and meets the requirements of industrial data compliance audit.

[0089] The aforementioned features create a synergistic effect through the following technological chain: multispectral acquisition and preprocessing eliminate environmental interference, generating standardized data input for multi-task models; adversarial training and progressive learning enhance model robustness, and distillation and compression adapt to edge deployment; hierarchical indexing and knowledge graphs support efficient retrieval; a closed-loop evolution mechanism achieves system self-optimization through incremental data and anomaly handling; and data lineage management ensures end-to-end traceability. Each module is interconnected, systematically addressing the technical shortcomings of low efficiency in manual database construction, accumulation of errors in complex operating conditions, and poor cross-platform compatibility.

[0090] Incremental training trigger mechanism: Triggering condition: The feature similarity score of the structured search results is lower than the dynamic threshold (threshold range: 90%-95% of the historical false positive rate statistics).

[0091] Threshold setting basis: The lower limit of the threshold is dynamically adjusted based on the similarity distribution of historical false positive samples. When the real-time similarity is lower than this threshold, it is judged as a high-value false positive sample.

[0092] The core design logic of this threshold range stems from a balance between optimizing the model's decision boundary and resource efficiency. Historical false positive rate statistics reflect the model's shortcomings in recognition ability under the current data distribution. Its similarity score distribution exhibits a typical skewed characteristic—the vast majority of false positive samples are concentrated in the low similarity range, while a small number of boundary samples are distributed near the decision boundary. Setting the threshold between the 90% and 95% quantiles of historical false positive rate statistics means that the system will accurately capture the most valuable training samples: samples at the 90th quantile represent the most difficult extreme cases for the model to recognize (such as heavily oiled surfaces or scenes with strong reflection interference), and these samples contain key feature patterns that are difficult to cover with traditional training; while the 95th quantile effectively filters out outliers caused by instantaneous interference or data noise, avoiding wasting training resources on invalid data.

[0093] Engineering validation shows that this range maximizes the optimization effect. A threshold below 90% will miss approximately 40% of difficult samples that could significantly improve model robustness, limiting the effectiveness of incremental training; a threshold above 95% will cause the system to frequently trigger the training process, increasing computational resource consumption by more than 30% and potentially impairing the model's generalization ability due to overtraining on individual abnormal samples. A dynamic adjustment mechanism further enhances the adaptability of this design—the threshold is automatically updated every 24 hours based on the latest false positive data, and a decay factor of 0.98 compensates for data distribution drift caused by model iteration, ensuring that the system always focuses on the most critical optimization direction.

[0094] In typical pipeline inspection scenarios, this range demonstrates accurate problem localization capabilities. For example, in the case of flange seal inspection of oil pipelines, when the threshold is set at the 90th percentile of the similarity score of historical false positive samples, the false positive rate for similar working conditions decreased by more than 30% after incremental training, while the accuracy of boundary sample recognition improved by 27%. This proves that the 90%-95% threshold range can effectively distinguish between truly ambiguous feature cases that need optimization and temporary interference data, allowing incremental training resources to be concentrated on the key samples that can best improve system performance.

[0095] Execution logic: Automatically extract the environmental parameters (temperature, humidity, light intensity) and manually corrected labels of the sample, generate an incremental dataset, and trigger a closed-loop evolution process.

[0096] Adversarial data augmentation triggering mechanism: Triggering condition: The retrieval anomaly rate of the hierarchical component library exceeds the set threshold for three consecutive statistical periods (threshold range: 5%-8%).

[0097] Anomaly rate calculation: Anomaly rate = (Number of times feature matching was successful but assembly verification failed) / Total number of searches × 100%.

[0098] Execution logic: When the anomaly rate exceeds the standard, the generative adversarial network is automatically invoked to synthesize targeted training samples by combining the current environmental parameters (oil reflectivity 0.3-0.7, strong reflective area proportion ≥15%).

[0099] Model version switching trigger mechanism: Triggering condition: Both indicators must be met simultaneously: The false positive rate decreased by ≥ 20% of the historical baseline. Response latency ≤ Engineering acceptable threshold (threshold range: 200-300ms).

[0100] Execution logic: Dynamically switch the production environment model to the optimal version and retain historical version snapshots (store the 5 most recent versions).

[0101] User feedback redirection triggered: Triggering condition: The visual feature difference between the user-selected area and the original search result is greater than or equal to the significant difference threshold (threshold range: 30%-40%).

[0102] Execution logic: Generate regional heatmaps based on attention mechanism, and adjust the weights of query vector channels by weighting (weight increase range: 25%-35%).

[0103] Index rebuild self-healing triggered: Triggering condition: Faiss index reconstruction fails 3 times consecutively (error type: resource allocation error / vector dimension mismatch).

[0104] Execution logic: Roll back to the index version successfully built within the last 3 days, and simultaneously trigger the data acquisition device calibration process (white balance parameter calibration tolerance: ±150K).

[0105] Data storage hierarchy migration triggered: Triggering conditions: High-frequency data: Search popularity score ≥ dynamic threshold (threshold range: based on the 90th percentile of visits in the previous 7 days).

[0106] Low-frequency data: Number of consecutive days without access > cold storage threshold (threshold range: 30±3 days).

[0107] Execution logic: Migrate SSD / HDD storage nodes based on popularity score, and control the compression ratio of low-frequency data to 1:3 to 1:5.

[0108] Application of incremental training triggering mechanism: In the scenario of flange component detection in oil pipelines, when the system identifies a sealing ring component with a feature similarity score of 0.82 (below the current dynamic threshold of 0.85), the system automatically marks it as a false positive sample. The environmental parameters of this sample show heavy oil contamination (oil area ≥ 60%) and strong side light interference (light intensity > 100,000 lux). According to the threshold setting rule (90th percentile of historical false positive rate), the system triggers the incremental training process: extracting the corrected label of the sample (manually confirmed to be a DN100 carbon steel flange), generating standardized data after cascaded preprocessing, and inputting it into the multi-task joint learning framework to update the parameters of the shared convolutional layer. After training, the model's recognition accuracy for similar working conditions improves by 32%, and the false positive rate decreases to one-quarter of the original level.

[0109] Adversarial Data Enhancement for Field Adaptation: In a liquefied natural gas pipeline valve inspection project, the system detected an anomaly rate of 7% for three consecutive days (exceeding the set threshold of 5%). Analysis showed that the anomalies were concentrated in the identification errors of stainless steel valves under low-temperature condensation conditions. The system immediately invoked a generative adversarial network (GAN) to synthesize 5,000 adversarial samples with condensation films, combined with the current environmental parameters (temperature -10℃, humidity 90%). After these samples were injected into the online calibration phase for training, the model's accuracy in identifying condensation valves improved from 68% to 92%, while the spatial assembly verification failure rate was controlled to within 1.5%.

[0110] Real-time control of model version switching: During collaborative operations in cross-regional pipeline engineering, edge computing nodes detected a 25% decrease in the false positive rate of the new model version (exceeding the 20% threshold), and the inference latency stabilized at 180ms (below the 250ms threshold). The system automatically switched the production environment to this version, while simultaneously retaining snapshots of historical versions. If subsequent strong sandstorm interference caused the false positive rate to rise again, the system rolled back to the previous stable version within 45 seconds, ensuring the continuity of on-site installation operations. This mechanism enables the system to maintain an average service availability of 98.7% during sudden environmental changes.

[0111] User feedback-driven retrieval optimization: In a subsea pipeline maintenance project, engineers selected flange areas partially covered by silt (visual difference reached 35%, exceeding the 30% threshold). The feedback loop module generated a region heatmap, increasing the weight of local features in the query vector by 30%. After re-retrieval, the same model flange fitting was accurately matched in just 0.8 seconds. The corrected results were used to generate training samples using a semi-automatic annotation tool. The model, iteratively updated within 48 hours, improved the accuracy of identifying occluded samples by 22%.

[0112] Implementation of tiered storage for hot and cold data: Access analysis of a pipeline component library in a chemical industrial park showed that large-diameter pipe fittings (DN80 and above) were accessed an average of 120 times per day (above the 90th percentile). The system automatically migrated to SSD nodes, reducing retrieval latency to 50ms. Meanwhile, some irregularly shaped pipe fittings had not been accessed for 33 consecutive days (exceeding the 30-day threshold). After ZSTD compression, they were archived to cold storage, freeing up 23% of online storage resources. When maintenance required access to expansion joint data in cold storage, the system completed decompression and loading within 2.3 seconds.

[0113] Fault recovery of the self-healing link: During a Faiss index reconstruction, consecutive failures occurred due to vector dimension mismatch (error code E1024). Within 10 seconds, the system rolled back to the index snapshot from 3 days prior. Simultaneously, the self-healing link triggered a calibration protocol update on the UAV data acquisition terminal. The calibration log showed that the white balance parameter was corrected from 6500K to 6200K (error -300K). After the correction, the success rate of index reconstruction using the acquired data increased to 99.9%, and the data lineage tracing link fully recorded this anomaly and its handling path.

[0114] Each triggering mechanism forms a tight technical loop through quantitative indicators: incremental training solves the bottleneck of identifying specific working conditions, adversarial enhancement optimizes the model's environmental adaptability, version switching ensures service stability, user feedback improves the weight of local features, data stratification optimizes resource utilization, and a self-healing link enables rapid fault isolation. This triggering system based on dynamic thresholds enables the system to continuously evolve in complex scenarios such as pipeline inspection, maintenance, and installation, systematically solving the performance degradation problem of traditional methods under sudden environmental changes.

[0115] Multispectral imaging scheme: Input is a pipe element entity in the environment, output is raw image data containing depth information. Core parameters include the visible light and near-infrared band ranges. The processing eliminates shadow interference through multi-band fusion, adjusts the sensor incident angle and exposure time, and enhances the integrity of texture details in low-light areas. Finally, stereoscopic image data covering the occluded scene is generated.

[0116] Cascaded preprocessing pipeline: Inputting raw multispectral image data, it performs two stages of operations sequentially. The first stage uses the multi-scale Retinex algorithm to decompose the illumination and reflection components, with the core parameter being the Gaussian kernel size. This process eliminates halo effects in metallic reflective areas. The second stage, based on morphological opening operations, uses a rectangular structuring element kernel to perform erosion and dilation operations, with the core parameter being the number of iterations. This process repairs rust texture breaks and removes device noise. The output is a normalized image with balanced illumination and geometric integrity.

[0117] A multi-task joint learning framework is proposed: It takes a standardized image dataset as input and constructs a shared convolutional feature extraction layer based on ResNet-50. The framework includes three parallel sub-networks: a component classification sub-network that outputs class probability distributions; a keypoint localization sub-network that outputs spatial coordinates based on high-resolution feature maps; and a defect detection sub-network that generates pixel-level masks through a region proposal mechanism. During training, adversarial data synthesized by a generative adversarial network is injected, with constraints on parameters including the reflectivity of oily materials and the proportion of highly reflective areas. The output is a feature response map that fuses features from multiple tasks.

[0118] A progressive training strategy is employed, optimizing model parameters in three stages. The first stage involves loading pre-trained weights from a large dataset. The second stage involves domain fine-tuning, inputting data according to environmental complexity gradients, gradually transitioning from clean samples to heavily contaminated samples. The third stage uses an active learning mechanism to filter low-confidence samples, with the confidence threshold as the core parameter. Finally, manually labeled data is added to the training set to dynamically correct the decision boundary. The output is a robust model adapted to complex operating conditions.

[0119] Knowledge distillation and compression: The trained multi-task model is input as the teacher model, and a lightweight student model is constrained by feature alignment loss. The core parameter is the cosine similarity loss weight coefficient. The processing ensures that the student model's output features are consistent with the intermediate layer responses of the teacher model, while also aligning with the index space vector direction. The output is a lightweight model feature vector adapted for edge devices.

[0120] Hierarchical component library construction: Inputting 1024-dimensional feature vectors extracted from the training model and 3D point cloud data, which are then fused through a feature concatenation layer. The core parameter is the normalization scaling factor. Outputting a 1280-dimensional component feature descriptor. The knowledge graph defines assembly rules based on ontology, with core parameters including thread specification tolerances and material compatibility tables. The process establishes the mapping relationship between feature descriptors and knowledge nodes. Outputting a structured knowledge base supporting assembly reasoning.

[0121] Multi-level index retrieval mechanism: Input the image or feature vector of the component to be retrieved. The first level locates the target category by binning the components according to the first letter of their type using a hash table. The second level performs an approximate nearest neighbor search within the target category using a product quantizer. The third level verifies assembly compatibility by associating it with the spatial topology data of the Building Information Model (BIM), with core parameters including axis alignment tolerance and minimum gap size. Output the matching results verified by geometric constraints.

[0122] Closed-loop evolution mechanism: The process is triggered when the feature similarity score falls below the 90%-95% quantile of the historical false detection rate. Input false detection samples and their environmental parameters to generate incremental data. Cascaded preprocessing generates a standardized dataset. A multi-task framework incrementally trains and updates the shared feature layer parameters. The hash table index and feature vector space are reconstructed. The optimized hierarchical component library index structure is output.

[0123] Anomaly self-healing link: Input index reconstruction failure signal in the incremental data pipeline. Process and roll back to the most recent 3 days' valid index version. Synchronously trigger the acquisition device calibration protocol, with the core parameter being white balance calibration tolerance. Output calibration log and update data lineage tracing records.

[0124] Dynamic storage strategy: Input component library access log data. Processing is based on sliding time window statistical retrieval popularity, with the core parameter being the popularity scoring model coefficients. High-frequency data is migrated to high-speed storage nodes, while low-frequency data undergoes block compression and archiving. Output a tiered storage resource optimization scheme.

[0125] Each model algorithm forms a collaborative chain through parameterized control: imaging and preprocessing ensure data quality; multi-task training generates discriminative features; knowledge graphs endow assembly semantics; multi-level indexes enable efficient retrieval; and closed-loop evolution continuously optimizes the system. The design of key technical parameters balances accuracy and efficiency, systematically addressing the shortcomings of traditional pipeline component library construction.

[0126] This invention systematically solves the problems of low efficiency, high error rate, and poor adaptability to complex environments associated with manually creating pipe component libraries through the following technical means: This invention acquires multi-source pipeline component image data using multispectral imaging equipment, and combines it with a cascaded preprocessing pipeline (Retinex illumination equalization and morphological opening operation) to eliminate equipment differences and environmental interference, generating a standardized image dataset. A semi-automatic annotation tool automatically annotates key components (such as flange interfaces and welds) based on a pre-trained model, and encapsulates metadata (pipe diameter, material) with the image, reducing manual annotation workload. Distributed storage nodes enable rapid data transmission and processing, avoiding the time-consuming steps of traditional manual data entry and significantly improving the efficiency of the data preparation stage.

[0127] A multi-task joint learning framework (ResNet-50, HRNet, Mask R-CNN) is adopted to share the underlying feature extraction layer, simultaneously optimizing component classification, key point localization, and defect detection tasks, reducing resource consumption and error aggregation from independent training of multiple models. A Generative Adversarial Network (GAN) is injected with adversarial data from oil-contaminated and highly reflective environments, combined with a progressive training strategy (basic training, domain fine-tuning, and online calibration) to enhance the model's robustness to complex working conditions. A closed-loop evolution mechanism dynamically updates model parameters and index structure through incremental data. When the retrieval anomaly rate exceeds a threshold, targeted data augmentation and model version rollback are triggered, forming an error self-correction capability and avoiding the lag of manual intervention.

[0128] A hierarchical component library integrating feature vectors and 3D point cloud data is constructed, enabling efficient retrieval through multi-level indexes (hash tables, Faiss, BIM topology verification). An attention-based feedback loop module captures user-selected regions and redirects the retrieval logic to local features, resolving mismatch issues in complex occlusion scenarios. A data lineage tracing link records the dependencies between acquisition device parameters and model versions, combined with dynamic storage strategies (SSD / HDD migration, cold storage archiving) to ensure data consistency across device iterations and environmental changes. An anomaly self-healing link automatically rolls back to a stable version in case of index failure, synchronously calibrating the data acquisition source parameters, forming end-to-end adaptive capabilities and overcoming the limitations of traditional methods in cross-platform compatibility and dynamic updates.

Claims

1. A method for creating a pipeline component library based on model training and image recognition technology, characterized in that, include: Step 1: Obtain image data of multi-source pipeline components and associated environmental parameters. Generate a standardized image dataset by performing cascaded preprocessing on the image data of multi-source pipeline components and associated environmental parameters. Step 2: Input the standardized image dataset into the multi-task joint learning framework for dynamic training. The multi-task joint learning framework integrates the component classification sub-network, the component key point localization sub-network, and the component defect detection sub-network, and outputs the training model by sharing the underlying feature extraction layer. Step 3: Based on the feature vectors extracted by the training model, generate component feature descriptors, associate the component feature descriptors with the component assembly relationships in the preset knowledge graph, construct a multi-level index structure, receive query input including the image or feature vector of the component to be retrieved, parse the component type identifier in the query, determine the target component category, perform an approximate nearest neighbor search in the target category, output candidate components with similar features, verify the spatial assembly compatibility between the candidate components and the query components, filter the final matching results that pass the topological constraints, use the final matching results as the structured retrieval results, use the structured retrieval results as the data input source of the closed-loop evolution mechanism, and output them to Step 4 in real time; Step 4: When the feature similarity score in the structured retrieval results is lower than the set threshold, it is marked as a false detection sample. The corrected label and associated environmental parameters of the false detection sample are extracted to generate incremental data. Based on the incremental data, a closed-loop evolution mechanism is triggered. The closed-loop evolution mechanism includes acquiring incremental data, performing cascaded preprocessing on the incremental data to generate an incremental standardized dataset, inputting the incremental standardized dataset into a multi-task joint learning framework, updating the feature extraction parameters of the training model, extracting feature vectors based on the updated training model, reconstructing the multi-level index structure of the hierarchical component library, and feeding back the retrieval results of the updated hierarchical component library to Step 1 to form a closed-loop iteration.

2. The method for creating a pipeline component library based on model training and image recognition technology according to claim 1, characterized in that, Step 1 includes: acquiring images of pipe components including depth information, and generating raw data for scenes with occlusion and low light. The raw data is input into the cascaded processing pipeline. The cascaded preprocessing pipeline sequentially performs illumination equalization processing based on the multi-scale Retinex algorithm, decomposes the illumination and reflection components of the image using the Gaussian kernel function, and performs texture restoration based on morphological opening operation to obtain the restored image. The rectangular structuring element kernel is then used to perform erosion and dilation operations. The repaired images and metadata are associated and encapsulated into a structured dataset using a semi-automatic annotation tool. The metadata includes acquisition device parameters, environmental parameters, and component attribute parameters. The acquisition device parameters are white balance and focal length parameters generated by a cross-device calibration protocol. The environmental parameters are the temperature, humidity, and light intensity during acquisition by the acquisition device. The component attribute parameters are the pipe diameter, pressure level, and material type identified by a pre-trained target detection model. The structured dataset is transmitted to the training module of a multi-task joint learning framework through distributed storage nodes.

3. The method for creating a pipeline component library based on model training and image recognition technology according to claim 2, characterized in that, Step 2 includes: An end-to-end network architecture based on ResNet-50, HRNet, and Mask R-CNN is constructed, sharing the underlying convolutional feature extraction layer; During the training phase, adversarial augmentation data generated by a generative adversarial network is injected to simulate the morphological changes of pipeline components in oily and highly reflective environments. The model parameters are optimized in stages through a progressive training strategy, including basic training based on ImageNet pre-trained weights, domain fine-tuning that increases with environment complexity, and an online calibration stage that incorporates active learning. The trained model is distilled into a lightweight MobileNetV3 model, and the feature output of the lightweight MobileNetV3 model is aligned with the Faiss index space of the hierarchical component library by cosine similarity.

4. The method for creating a pipeline component library based on model training and image recognition technology according to claim 3, characterized in that, Step 3 includes: fusing the 1024-dimensional feature vector output by the multi-task joint learning framework with the 3D point cloud data through a feature stitching layer to form a component feature descriptor; Construct an ontology-based knowledge graph, define the mapping relationship between feature descriptors and knowledge nodes, and the knowledge graph includes the assembly relationship between valves and pipelines and the replacement rules for compatible models; Component retrieval is based on a multi-level index structure. Component retrieval includes locating the major category of components through a hash table, constructing an approximate nearest neighbor search index to match feature vectors through the Faiss framework, and matching spatial topological relationships with associated BIM models.

5. The method for creating a pipeline component library based on model training and image recognition technology according to claim 4, characterized in that, Step 4 includes: Deploy incremental data pipelines to capture new images during on-site installation and maintenance, and filter samples that differ from the existing library by more than a set threshold through differential compression; Establish a model versioning and iteration framework, dynamically switch training model versions based on the decrease in false detection rate and response latency indicators of online inference, and retain traceable model snapshots; When the retrieval anomaly rate of the hierarchical component library exceeds a set threshold, the adversarial data augmentation module in the multi-task joint learning framework is triggered to generate targeted training samples for oily and highly reflective environments.

6. The method for creating a pipeline component library based on model training and image recognition technology according to claim 5, characterized in that, Also includes: In the preprocessing stage, a cross-device calibration protocol is established to unify white balance and focal length parameters based on the device type of the acquisition terminal, which includes drones and handheld devices. The parameter configuration generated by the calibration protocol is input into the generator input of the generative adversarial network to constrain the material reflectivity parameters of the oil stain and strong reflection synthetic samples output by the generative adversarial network.

7. The method for creating a pipeline component library based on model training and image recognition technology according to claim 6, characterized in that, Also includes: An attention-based feedback loop module is embedded in the search interface of the hierarchical component library. It generates a region heatmap based on the image region selected by the user, redirects the query vector of the approximate nearest neighbor search index, and obtains the corrected search results. The corrected search results are back-annotated to the structured dataset using a semi-automatic annotation tool, generating new training samples which are then input into the active learning module during the online calibration phase.

8. The method for creating a pipeline component library based on model training and image recognition technology according to claim 7, characterized in that, Also includes: Establish a data lineage tracing link to record the acquisition device ID, preprocessing parameters, and model version identifier involved in training for each training sample in the structured dataset; By associating lineage links with model version snapshots, a timestamp matching mechanism is used to verify the training dependency of a specific model version on historical datasets.

9. The method for creating a pipeline component library based on model training and image recognition technology according to claim 8, characterized in that, Also includes: Define data aging rules to perform cold storage archiving on low-frequency access component data that has not been retrieved for 30 consecutive days in the hierarchical component library; The storage tier strategy is dynamically adjusted based on the search popularity index statistically analyzed by the feedback loop module. High-frequency access data is migrated to SSD storage nodes, while low-frequency data is downgraded to HDD storage nodes.

10. The method for creating a pipeline component library based on model training and image recognition technology according to claim 9, characterized in that, Also includes: Deploy an anomaly self-healing link in the incremental data pipeline to roll back to the retained historical model version and corresponding index snapshot when Faiss index reconstruction fails. The self-healing link triggers a cross-device calibration protocol to recalibrate the white balance parameters of the data acquisition source and generates calibration logs to synchronize data lineage tracing links.

Images