Ship pipe fitting identification system and method based on multi-source data enhancement and intelligent recommendation

Abstract

A ship pipe fitting recognition system based on multi-source data enhancement and intelligent recommendation comprises a multi-source data enhancement module, a model training module, a recommendation decision module, a federated learning platform and a block chain evidence storage module, the system is of a'cloud + side 'double-layer architecture, the cloud is used for intelligent evolution, distributed knowledge is integrated through federated learning, and the block chain evidence storage module is used for storing the distributed knowledge. The global model is continuously updated and optimized, the edge end is used for edge real-time response, and efficient pipe fitting feature detection and pipe fitting number recommendation are carried out at the equipment end. According to the method, the detection precision and the non-standard pipe fitting recognition capability in complex illumination and shielding scenes can be remarkably improved, and the misjudgment rate is reduced; the decision-making problem of similar part hybrid scenes is solved, and the assembly efficiency and accuracy are improved; meanwhile, the data privacy security is ensured; the technical blank in the aspects of data diversity, algorithm robustness and industrial adaptability in the field of ship pipe fitting recognition is filled, the efficient requirement of a ship assembly scene is met, and technical support is provided for ship manufacturing intelligence.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and intelligent ship manufacturing technology, and in particular to a ship pipe identification system based on multi-source data enhancement and intelligent recommendation, and also to an identification method of the ship pipe identification system based on multi-source data enhancement and intelligent recommendation. Background Art

[0002] Ship pipe fittings are characterized by large size, numerous non-standard parts, and a high degree of similarity in shape. Currently, the manufacturing of ships and their products requires scheduling, issuing, and managing production tasks through a workshop production management platform. On-site workers refer to product and process design drawings to manufacture ship products. However, traditional inspection methods (such as binocular vision and lidar) have problems such as high cost, limited accuracy, and poor adaptability.

[0003] Existing deep learning-based object detection methods (such as YOLO and Faster R-CNN) utilize end-to-end convolutional neural networks to achieve high-precision detection, real-time processing, and multi-scale feature fusion. However, they still face significant challenges in complex industrial scenarios, including insufficient data, weak model generalization, insufficient edge real-time performance, and insufficient data security. Furthermore, the large number of non-standard pipe fittings and the dynamic updates in information databases lead to low recognition efficiency with traditional single models.

[0004] Therefore, there is an urgent need for a high-precision, lightweight, and scalable intelligent recognition system to improve the level of automation in shipbuilding. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a ship pipe identification system based on multi-source data enhancement and intelligent recommendation that is highly precise, lightweight, and scalable and can solve the problems of detection accuracy and process reliability caused by insufficient data diversity, high misjudgment rate of non-standard parts, and low cross-organizational collaboration efficiency in ship pipe identification.

[0006] Another technical problem to be solved by the present invention is to provide an identification method for the ship pipe identification system based on multi-source data enhancement and intelligent recommendation.

[0007] The technical problem to be solved by the present invention is achieved through the following technical solutions. The present invention is a ship pipe identification system based on multi-source data enhancement and intelligent recommendation, which includes a multi-source data enhancement module, a model training module, a recommendation decision module, a federated learning platform and a blockchain evidence storage module.

[0008] The multi-source data enhancement module includes an industrial camera unit (resolution ≥1920×1080, frame rate ≥30FPS), a CAD drawing parsing unit, a virtual rendering unit, a labeling unit and a storage unit, wherein the industrial camera unit is used to deploy a high-resolution industrial camera to support multi-angle shooting of pipe fitting images; the CAD drawing parsing unit is used to receive pipe fitting CAD design files, parse the geometric parameters of common feature points of ship pipe fittings, and convert them into a Blender-compatible format; the virtual rendering unit is used to configure a multi-light source environment in Blender, adjust model pose parameters, simulate stacking and occlusion scenes, and superimpose Gaussian noise and motion blur to generate a highly simulated interference image; the labeling unit is used to use the Labelme tool to manually / automatically label the six types of feature points in the mixed data set and generate an XML label file; the storage unit is used to export and store the generated standardized ship pipe fitting common feature point data set into a distributed database to support data management and rapid retrieval;

[0009] The model training module is used to be deployed on edge intelligent devices to realize data collection and pipe feature detection at the assembly site through an external camera. At the same time, the model training module is equipped with an optimized YOLOv5s model, supports mixed data set training, and detects common feature points of multiple types of ship pipe fittings including but not limited to elbows, tees, crosses, tapered pipes, flanges and sleeves, identifies and counts the number of various feature points, and forms a multi-dimensional feature vector: V 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common characteristic points of other ship fittings;

[0010] The recommendation decision module includes a feature extraction unit, a K-nearest neighbor (KNN) algorithm number matching unit and a human-computer interaction interface. The feature extraction unit is used to extract the multidimensional feature vector V from the cascade detection module. 检测 =[v1,v2,v3,v4,v5,v6,…,v n ]; The K-nearest neighbor (KNN) algorithm number matching unit is used to perform global similarity matching with the pipe fitting information database through the KNN algorithm, and the Euclidean distance is used to calculate the similarity between the detection vector and the feature vector of each pipe fitting model in the pipe fitting information database; the human-computer interaction interface is used to build a graphical interface based on PyQt5, which displays the detection box, the recommended pipe fitting number list and the process guidance animation of the relevant pipe fittings in real time, and supports touch screen or external devices. Based on the recommended pipe fitting number list, the on-site staff conducts secondary confirmation and final selection of the pipe fitting number, and the key data (including but not limited to the recommended model, manual selection record and related process parameters) is synchronized to the blockchain;

[0011] The federated learning platform includes an edge training unit, a gradient encryption aggregation unit, and a global model delivery unit. The edge training unit is used to deploy a cluster of edge detection devices, fine-tune the detection model locally, and compress the model through channel pruning and knowledge distillation methods to make the model better adapted to edge devices. The gradient encryption aggregation unit is used to encrypt the model gradient data packet using the Paillier homomorphic encryption algorithm and transmit it to the cloud for aggregation. The global model is updated based on the node data volume ratio and dynamic weight allocation. The global model delivery unit is used to securely deliver the updated model parameters to the edge nodes through the HTTPS protocol, and the version number verification ensures consistency.

[0012] The blockchain evidence storage module is used to achieve information sharing based on the distributed ledger structure of blockchain technology. That is, all authorized personnel can view the pipe assembly status and data through the chain, and then use the SHA-256 algorithm to generate the detection image hash value, associate the timestamp with the equipment ID, and store the detection results, process parameters, and timestamp of the pipe information on the blockchain, supporting rapid data verification, traceability and anti-tampering query.

[0013] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions. For the above-mentioned ship pipe identification system based on multi-source data enhancement and intelligent recommendation, the system is a two-layer architecture of "cloud + edge". The cloud is used for intelligent evolution, integrating distributed knowledge through federated learning, and continuously updating and optimizing the global model. The edge is used for real-time edge response, and performs efficient pipe feature detection and pipe number recommendation on the device side.

[0014] The technical problem to be solved by the present invention can be further achieved through the following technical solutions: for the above-mentioned ship pipe identification system based on multi-source data enhancement and intelligent recommendation, a ship pipe identification method based on multi-source data enhancement and intelligent recommendation is provided, which includes hybrid data set construction, pipe feature detection model, pipe number recommendation mechanism, data privacy protection mechanism based on federated learning, and detection traceability mechanism based on blockchain.

[0015] The hybrid dataset is constructed by rendering the CAD model of ship pipe fittings from multiple angles using Blender software to generate a virtual image, which is then combined with the real-shot image to form a hybrid dataset. Labelme is then used to manually or automatically label common feature points of various types of ship pipe fittings in the image, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves. The hybrid dataset contains common feature points of ship pipe fittings.

[0016] The pipe feature detection model is constructed by optimizing the YOLOv5s network architecture, replacing the loss function with WIoUv3, and optimizing the detection accuracy by dynamically adjusting the geometric penalty weight and the distance attention mechanism. The model is then trained using the mixed dataset in the mixed dataset construction. The model then detects common feature points of multiple types of ship pipe fittings, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and casings, identifies and counts the number of various feature points, and outputs a multidimensional feature vector.

[0017] The pipe fitting number recommendation mechanism is based on the multi-dimensional feature vector of common feature points of ship pipe fittings output by detection. It uses the KNN algorithm to perform global similarity matching with the pipe fitting information database, sorts the distance from small to large, and outputs the four most likely recommended ship pipe fitting numbers. It also supports manual selection.

[0018] The privacy protection mechanism based on federated learning trains the model locally on the edge device, encrypts and uploads the model gradient to the cloud for aggregation, and dynamically allocates node weights to optimize the global model. Multiple edge devices train the detection model locally, encrypt the model gradient using the Paillier homomorphic encryption algorithm, and upload it to the cloud for aggregation. The cloud updates the global model using a dynamic weight allocation strategy and sends it to each edge node. The knowledge distillation technology is used to migrate the knowledge of the large cloud model to the small edge model, and the feature distribution is aligned using the KL divergence loss.

[0019] The blockchain-based detection traceability mechanism is used to support key data traceability, quickly locate error nodes, and realize distributed storage of detection logs based on blockchain technology. Each evidence record includes but is not limited to the detection timestamp, SHA-256 image hash, pipe number, and complete metadata of process parameters. Based on Merkle tree index and chain tracking technology, the block where the abnormal record is located is located, and related data is extracted through a multi-dimensional query interface to automatically generate an error analysis report.

[0020] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions. For the above-mentioned ship pipe fitting identification method based on multi-source data enhancement and intelligent recommendation, the hybrid data set is constructed to collect multi-angle image data of pipe fittings at the assembly site through an industrial camera, obtain real-shot images of ship pipe fittings, and import the CAD design drawings of ship pipe fittings into Blender software. A multi-light source environment is set to simulate the lighting conditions of the ship assembly site, and the model pose and background parameters are adjusted to generate a virtual rendered image containing occlusion and rust interference. The real-shot image and the virtual rendered image are then mixed in a 6:4 ratio to form a standardized data set. The Labelme tool is used to manually / automatically annotate multiple types of feature points in the image, including but not limited to elbows, tees, crosses, tapered pipes, flanges and sleeves, to generate an XML label file containing bounding box coordinates and feature categories.

[0021] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions. For the above-mentioned ship pipe identification method based on multi-source data enhancement and intelligent recommendation, the pipe feature detection model is used to adopt the network architecture of the optimized YOLOv5 detection model, replace the CIoU in the border loss function with WIoUv3, and dynamically adjust the parameters. The aspect ratio difference and edge feature error during the training process are monitored in real time, and the geometric penalty weight coefficient α of WIoUv3 is dynamically updated through the distance attention mechanism; the border loss function is specifically as follows:

[0022] L WIoUv3 =rL WIoUv1

[0023] The specific formula is:

[0024]

[0025] L WIoUv1 =R WIoU L IoU

[0026]

[0027] L IoU =1-IoU

[0028] Among them, IoU is the interaction ratio, xgt, ygt are the coordinates of the center point of the real box, x, y are the coordinates of the center point of the predicted box, Wg, Hg are the sizes of the smallest enclosing boxes; the mapping of outlier β and gradient gain r is controlled by hyperparameters α and δ.

[0029] The technical problem to be solved by the present invention can be further achieved by the following technical solutions. For the above-mentioned ship pipe identification method based on multi-source data enhancement and intelligent recommendation, the pipe number recommendation mechanism is used to extract the multi-dimensional feature vector V from the detection model. 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common characteristic points of other ship fittings;

[0030] The pipe fitting number recommendation mechanism is used to perform global similarity matching with the pipe fitting information database through the KNN (K nearest neighbor) algorithm, and the Euclidean distance is used to calculate the similarity between the detection vector and the feature vector of each pipe fitting model in the pipe fitting information database. The formula is:

[0031] v数据库，i is the i-th eigenvector in the pipe fitting information database;

[0032] The pipe fitting number recommendation mechanism is used to sort the distances from smallest to largest to obtain the four most likely ship pipe fitting numbers, and push them as recommendation results to the manual interaction interface for on-site staff to view and select.

[0033] The technical problem to be solved by the present invention can be further achieved through the following technical solutions. For the ship pipe identification method based on multi-source data enhancement and intelligent recommendation described above, the data protection mechanism based on the federated learning framework is used for local training of edge devices and cloud model aggregation, specifically:

[0034] Step 1: Local model adaptation on edge devices

[0035] (1) Model initialization: Each edge device obtains the latest version of the global model from the cloud and performs lightweight adaptation based on the computing power of the edge device at the production site. It removes redundant structures through pruning technology, retains the key feature extraction layer, and deploys a lightweight detection model locally.

[0036] (2) Local training: Input the local dataset; set the training rounds to 50, the batch size to 8, and the learning rate to 0.001×0.95 t , t is the current training round, and the SAM (Sharpness Perceptual Minimization) optimizer is used to improve the generalization of the model. During the training process, the model simultaneously learns the detection task and imitates the knowledge features of the complex cloud model, and optimizes its own performance by comparing the output of the intermediate layer.

[0037] (3) Gradient calculation: After training is completed, the edge device calculates the adjustment direction (gradient) of the model parameters and normalizes the gradient to limit its numerical range to prevent information leakage. Subsequently, random noise is added to the processed gradient to ensure that the training details of a single device cannot be reversely inferred.

[0038] Step 2: Gradient encrypted transmission and dynamic aggregation in the cloud

[0039] (1) Gradient encryption protection: Paillier homomorphic encryption technology is used to encrypt the gradient, allowing the cloud to perform aggregate calculations without decryption. The encrypted data packets are transmitted to the cloud via a secure communication protocol. Even if they are intercepted during the process, the original content cannot be decrypted.

[0040] (2) Dynamic weight allocation: The cloud allocates aggregate weights based on the data contribution and quality of each device. Devices with larger data volumes and more stable data quality receive higher weights. An exponential decay strategy is used to prevent a small number of devices from dominating the global model, ensuring fairness among multiple participants.

[0041] Step 3: Global model update and distribution

[0042] (1) Joint parameter update: The cloud merges the encrypted gradients of each device according to the weights to generate an updated global model. The update process uses a noise cancellation mechanism to ensure that the added random noise does not affect the model validity after aggregation;

[0043] (2) Knowledge distillation and model compression: Through knowledge distillation technology, the key recognition capabilities of complex cloud models are refined into lightweight models. Specifically, the edge model (student model) is made to imitate the characteristic response pattern of the large cloud model (teacher model), reducing the number of parameters while maintaining high accuracy.

[0044] (3) Lightweight model delivery: Only the changed parts of the model are delivered to the edge device instead of the complete parameters, which greatly reduces the amount of transmitted data. The update package uses digital signature verification to prevent tampering. After the device receives it, it is automatically integrated into the local model.

[0045] The technical problem to be solved by the present invention can also be further achieved through the following technical solutions. For the above-mentioned ship pipe identification method based on multi-source data enhancement and intelligent recommendation, the blockchain-based detection traceability mechanism is used to support the trusted storage of key data and rapid error location. The detection results and process parameters are written into a distributed ledger in the form of structured records. Each record contains a timestamp, SHA-256 image hash, pipe number, manual operation log and process parameter fields. The ledger adopts a block storage structure, each block has a capacity of 1MB, and can accommodate 50-100 records; the blocks are connected in series through cryptographic hashes to ensure that the global state cannot be tampered with. At the same time, the record hash level is aggregated internally through the Merkle tree, and the root hash is written to the block header to achieve second-level verification of a single piece of data. The ledger provides a query interface, supports multi-dimensional retrieval by timestamp range, pipe model code, and process parameters, locates abnormal nodes in the assembly process, and generates error reports in PDF / CSV format.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. This invention constructs a hybrid dataset by fusing real-life images with virtual rendered images. Combined with Blender's multi-light source simulation technology, it generates highly realistic images of scenes with complex lighting, occlusion, and interference. This effectively addresses the difficulty in collecting data on ship pipe fittings (especially the scarcity of non-standard parts). Virtual data augmentation significantly expands the coverage of training samples, improves the model's adaptability to dynamic changes in industrial scenarios, and reduces the error rate caused by insufficient data.

[0048] 2. This paper adopts an improved YOLOv5s network architecture, replaces the traditional CIoU loss function with WIoUv3, and enhances the model's sensitivity to the edge features and subtle differences of non-standard pipe fittings through dynamic geometric penalty weights and distance attention mechanism. Combined with mixed data set training, it significantly improves the detection robustness in complex lighting and occlusion scenarios, and reduces the misjudgment rate of non-standard pipe fittings.

[0049] 3. Based on the federated learning framework, this invention supports local training of lightweight models on edge devices, uploads gradients to the cloud for aggregation through Paillier homomorphic encryption, and dynamically allocates node weights to optimize the global model. This not only addresses the problem of limited computing power at the edge, but also protects the privacy of sensitive data in multi-organizational collaboration, breaks down data silos, and enables cross-device knowledge sharing and continuous model evolution.

[0050] 4. The present invention designs a multi-dimensional feature vector (such as the number statistics of elbows, tees, etc.) and a KNN global similarity matching algorithm to quickly screen out the four pipe fitting number recommendation results with the highest probability. Combined with the human-computer interaction interface, it supports manual secondary confirmation, which not only improves the recommendation efficiency, but also reduces the decision-making error rate in scenarios with mixed similar parts, shortening the assembly process time. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Constructing a flow chart for the hybrid dataset of the present invention;

[0052] Figure 2 A flow chart of the pipe numbering recommendation mechanism of the present invention;

[0053] Figure 3 This is a diagram of the edge-cloud collaborative pipe detection architecture based on a federated platform of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0055] Reference Figure 1-3 A ship pipe identification system based on multi-source data enhancement and intelligent recommendation, including a multi-source data enhancement module, a model training module, a recommendation decision module, a federated learning platform, and a blockchain evidence storage module;

[0056] The multi-source data enhancement module includes an industrial camera unit, a CAD drawing parsing unit, a virtual rendering unit, a labeling unit, and a storage unit. The industrial camera unit is used to deploy high-resolution industrial cameras (resolution ≥1920×1080, frame rate ≥30FPS) to support multi-angle capture of pipe fitting images. The CAD drawing parsing unit receives pipe fitting CAD design files, analyzes the geometric parameters of common feature points of ship pipe fittings, and converts them into a Blender-compatible format. Through the virtual rendering unit, a multi-light source environment (including but not limited to parallel light, point light, and ambient light) is configured in Blender. Model pose parameters are adjusted to simulate stacking and occlusion scenarios. Gaussian noise and motion blur are superimposed to generate highly simulated interference images. The labeling unit uses the Labelme tool to manually / automatically label the six types of feature points in the mixed dataset and generate an XML label file. The storage unit exports and stores the generated standardized dataset of common feature points of ship pipe fittings in a distributed database, supporting data management and rapid retrieval.

[0057] The model training module is deployed on edge intelligent devices and uses external cameras to collect data from the assembly site and detect pipe features. The model training module is equipped with an optimized YOLOv5s model, supports mixed data set training, and detects common feature points of various types of ship pipes, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves. It identifies and counts the number of various feature points to form a multi-dimensional feature vector V 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common feature points of other ship fittings.

[0058] The recommendation decision module includes a feature extraction unit, a K-nearest neighbor (KNN) algorithm number matching unit, and a human-computer interaction interface. The feature extraction unit extracts six multi-dimensional feature vectors V from the cascade detection module. 检测 =[v1,v2,v3,v4,v5,v6,…,v n The K-nearest neighbor (KNN) algorithm number matching unit uses the KNN algorithm to perform global similarity matching with the pipe fitting information database, using Euclidean distance to calculate the similarity between the detection vector and the feature vectors of each pipe fitting model in the pipe fitting information database. The human-computer interaction interface is built on a PyQt5 graphical interface, displaying the detection box, a list of recommended pipe fitting numbers, and process guidance animations for related pipe fittings in real time. Supporting touch screens or external devices, on-site staff can re-confirm and finally select the pipe fitting number based on the recommended pipe fitting number list. Key data (including but not limited to recommended models, manual selection records, and associated process parameters) is synchronized to the blockchain.

[0059] The federated learning platform includes an edge training unit, a gradient encryption aggregation unit, and a global model delivery unit. The edge training unit deploys a cluster of edge detection devices, fine-tunes the detection model locally, and compresses the model through methods such as channel pruning and knowledge distillation to make the model better adapted to edge devices. The gradient encryption aggregation unit uses the Paillier homomorphic encryption algorithm to encrypt the model gradient and the gradient data packet, and transmits them to the cloud for aggregation; the global model is updated according to the proportion of node data volume and dynamic weight allocation. The global model delivery unit is used to securely deliver the updated model parameters to the edge nodes through the HTTPS protocol, and version number verification ensures consistency.

[0060] The blockchain evidence storage module realizes information sharing based on the distributed ledger structure of blockchain technology, that is, all authorized personnel can view the assembly status and data of pipe fittings through the chain; the SHA-256 algorithm is used to generate the hash value of the inspection image, and the timestamp and equipment ID are associated; the inspection results, process parameters, timestamp and other pipe fitting information are stored in the blockchain, supporting rapid data verification, traceability and tamper-proof query.

[0061] A ship pipe identification method based on multi-source data enhancement and intelligent recommendation, mainly including hybrid dataset construction, pipe feature detection model, pipe number recommendation mechanism, data privacy protection based on federated learning, and detection traceability mechanism based on blockchain;

[0062] The hybrid dataset is constructed by using Blender software to perform multi-angle light source rendering on the CAD model of ship pipe fittings to generate a virtual image, which is combined with the real-shot image to form a hybrid dataset. The Labelme tool is used to automatically annotate the common feature points of various types of ship pipe fittings in the image, including but not limited to elbows, tees, crosses, tapered pipes, flanges and sleeves. The hybrid dataset contains common feature points of ship pipe fittings.

[0063] The pipe feature detection model optimizes the YOLOv5s network architecture and replaces the loss function with WIoUv3. The detection accuracy is optimized by dynamically adjusting the geometric penalty weight and the distance attention mechanism. The model is trained using the aforementioned mixed dataset. The model detects common feature points of various types of ship pipe fittings, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and casings, identifies and counts the number of various feature points, and outputs a multidimensional feature vector.

[0064] The pipe fitting number recommendation mechanism is based on the multidimensional feature vector of common feature points of ship pipe fittings output by detection. It uses the KNN (K nearest neighbor) algorithm to perform global similarity matching with the pipe fitting information database, and outputs the four recommended ship pipe fitting numbers with the highest probability in ascending order of distance, and supports manual selection.

[0065] The privacy protection mechanism based on federated learning trains the model locally on the edge device, encrypts and uploads the model gradient to the cloud for aggregation, and dynamically allocates node weights to optimize the global model. Multiple edge devices train the detection model locally, encrypt the model gradient through the Paillier homomorphic encryption algorithm, and upload it to the cloud for aggregation; the cloud updates the global model through a dynamic weight allocation strategy and sends it to each edge node; the knowledge distillation technology is used to migrate the knowledge of the large cloud model to the small edge model, and the feature distribution is aligned through the KL divergence loss.

[0066] The blockchain-based inspection traceability mechanism supports the traceability of key data, enabling rapid location of faulty nodes. Blockchain technology enables distributed storage of inspection logs, with each record containing complete metadata, including but not limited to the inspection timestamp, SHA-256 image hash, pipe number, and process parameters. Using Merkle tree indexing and chain tracing technology, the system locates the block containing the abnormal record, extracts related data through a multi-dimensional query interface, and automatically generates an error analysis report.

[0067] The ship pipe identification system, based on multi-source data enhancement and intelligent recommendation, consists of a two-tier "cloud + edge" architecture. The edge provides real-time response, enabling efficient pipe feature detection and number recommendation on the device side. The cloud-side intelligent evolution integrates distributed knowledge through federated learning, continuously updating and optimizing the global model.

[0068] Specifically:

[0069] A hybrid dataset was constructed by using industrial cameras to capture multi-angle image data of pipe fittings at the assembly site to obtain real-life images of ship pipe fittings. CAD design drawings of ship pipe fittings were imported into Blender software, and a multi-light environment was set up to simulate the lighting conditions at the ship assembly site. The model pose and background parameters were adjusted to generate virtual rendered images that included occlusion and corrosion interference. The real-life images and virtual rendered images were mixed in a 6:4 ratio to form a standardized dataset. The Labelme tool was used to automatically annotate multiple feature points in the images, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves, generating an XML label file containing bounding box coordinates and feature categories.

[0070] The pipe feature detection model uses the optimized YOLOv5 detection model network architecture, replacing CIoU in the bounding box loss function with WIoUv3. Dynamic parameter adjustment monitors the aspect ratio difference and edge feature error in real time during training, and dynamically updates the geometric penalty weight coefficient α of WIoUv3 through the distance attention mechanism. The bounding box loss function is as follows:

[0071] L WIoUv3 =rL WIoUv1

[0072] The specific formula is:

[0073]

[0074] L WIoUv1 =R WIoU L IoU

[0075]

[0076] L IoU =1-IoU

[0077] Among them, IoU is the interaction ratio, x gt 、y gt is the center point coordinate of the real box, x and y are the center point coordinates of the predicted box, and W g 、H g is the size of the smallest enclosing box; the mapping between outlier degree β and gradient gain r is controlled by hyperparameters α and β.

[0078] Pipe fitting number recommendation mechanism, multi-dimensional feature vector V extracted from the detection model 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common feature points of other ship fittings.

[0079] The pipe fitting number recommendation mechanism uses the KNN (K nearest neighbor) algorithm to perform global similarity matching with the pipe fitting information database. The Euclidean distance is used to calculate the similarity between the detection vector and the feature vector of each pipe fitting model in the pipe fitting information database. The formula is:

[0080] v 数据库，i is the i-th feature vector in the pipe fitting information database.

[0081] The pipe fitting number recommendation mechanism sorts the distance from smallest to largest to obtain the four most likely ship pipe fitting numbers, and pushes them to the manual interaction interface as the recommendation results for on-site staff to view and select;

[0082] The data protection mechanism based on the federated learning framework enables local training on edge devices and aggregation of cloud models:

[0083] Step 1: Local model adaptation on edge devices

[0084] (1) Model initialization: Each edge device obtains the latest version of the global model from the cloud and performs lightweight adaptation based on the computing power of the edge device at the production site. It removes redundant structures through pruning technology, retains the key feature extraction layer, and deploys a lightweight detection model locally.

[0085] (2) Local training: Input the local dataset; set the training rounds to 50, the batch size to 8, and the learning rate to 0.001×0.95 t (t is the current training round) and uses the SAM (Sharpness Perceived Minimization) optimizer to improve model generalization. During training, the model simultaneously learns the detection task and imitates the knowledge features of complex cloud models, optimizing its own performance by comparing the output of intermediate layers.

[0086] (3) Gradient calculation: After training is completed, the edge device calculates the adjustment direction (gradient) of the model parameters and normalizes the gradient to limit its numerical range to prevent information leakage. Subsequently, random noise is added to the processed gradient to ensure that the training details of a single device cannot be reversely inferred.

[0087] Step 2: Gradient encrypted transmission and dynamic aggregation in the cloud

[0088] (1) Gradient encryption protection: Paillier homomorphic encryption technology is used to encrypt the gradient, allowing the cloud to perform aggregate calculations without decryption. The encrypted data packets are transmitted to the cloud via a secure communication protocol. Even if they are intercepted during the process, the original content cannot be decrypted.

[0089] (2) Dynamic weight allocation: The cloud allocates aggregate weights based on the data contribution and quality of each device. Devices with larger data volumes and more stable data quality receive higher weights. An exponential decay strategy is used to prevent a small number of devices from dominating the global model, ensuring fairness among multiple participants.

[0090] Step 3: Global model update and distribution

[0091] (1) Joint parameter update: The cloud merges the encrypted gradients of each device according to the weights to generate an updated global model. The update process uses a noise cancellation mechanism to ensure that the added random noise does not affect the model validity after aggregation;

[0092] (2) Knowledge distillation and model compression: Through knowledge distillation technology, the key recognition capabilities of complex cloud models are refined into lightweight models. Specifically, the edge model (student model) is made to imitate the characteristic response pattern of the large cloud model (teacher model), reducing the number of parameters while maintaining high accuracy.

[0093] (3) Lightweight model delivery: Only the changed parts of the model are delivered to the edge device instead of the complete parameters, which greatly reduces the amount of transmitted data. The update package uses digital signature verification to prevent tampering. After the device receives it, it is automatically integrated into the local model.

[0094] The blockchain-based inspection and traceability mechanism supports the trusted storage of key data and rapid error location. Inspection results and process parameters are written into the distributed ledger in the form of structured records. Each record contains fields such as timestamp, SHA-256 image hash, pipe number, manual operation log and process parameters. The ledger adopts a block storage structure, with each block capacity of 1MB, accommodating 50-100 records; blocks are connected in series through cryptographic hashes to ensure that the global state cannot be tampered with. At the same time, the record hash level is aggregated internally through the Merkle tree, and the root hash is written to the block header to achieve second-level verification of a single data piece. The ledger provides a query interface that supports multi-dimensional retrieval by timestamp range, pipe model code, process parameters, etc., locates abnormal nodes in the assembly process, and generates error reports in PDF / CSV format.

[0095] The present invention provides a ship pipe identification system and method based on multi-source data enhancement and intelligent recommendation, aiming to solve the problems of detection accuracy and process reliability caused by insufficient data diversity, high misjudgment rate of non-standard parts and low cross-organizational collaboration efficiency in ship pipe identification. By constructing a hybrid data set that integrates real-shot images and virtual rendered images, combining Blender software with multi-angle light source simulation to generate highly realistic images, and using automated annotation tools to achieve standardized feature point annotation, the data diversity and model generalization ability are effectively improved; further optimizing the target detection model based on YOLOv5, and converting the traditional CIoU loss function into a real-world image. The algorithm replaces the traditional WIoUv3 with WIoUv3, introduces dynamic geometric penalty weights and distance attention mechanism, and significantly enhances the detection sensitivity of edge features of non-standard pipe fittings. At the same time, a multi-dimensional model recommendation system based on the KNN algorithm is designed. By matching the six-dimensional feature vector with the global similarity of the engineering database, it outputs probability-sorted recommendation results to support rapid decision-making and manual correction. The local training model of the edge device is realized through the federated learning framework, and the model gradient is encrypted and uploaded to the cloud for aggregation, and dynamic weight allocation is performed to protect data privacy. The detection results and process parameters are stored on the chain with the combination of blockchain technology to ensure that key data cannot be tampered with and support rapid backtracking.

[0096] The implementation of the present invention can significantly improve the detection accuracy and non-standard pipe identification capabilities in complex lighting and occlusion scenarios, and reduce the misjudgment rate; solve the decision-making problem in scenarios with mixed similar parts, and improve assembly efficiency and accuracy; while ensuring data privacy and security; fill the technical gaps in the field of ship pipe identification in terms of data diversity, algorithm robustness and industrial adaptability, adapt to the high-efficiency requirements of ship assembly scenarios, and provide technical support for intelligent ship manufacturing.

Claims

1. A ship pipe identification system based on multi-source data enhancement and intelligent recommendation, characterized by: It includes multi-source data enhancement module, model training module, recommendation decision module, federated learning platform and blockchain evidence storage module. The multi-source data enhancement module includes an industrial camera unit, a CAD drawing parsing unit, a virtual rendering unit, a labeling unit, and a storage unit. The industrial camera unit is used to deploy a high-resolution industrial camera to support multi-angle shooting of pipe fitting images; the CAD drawing parsing unit is used to receive pipe fitting CAD design files, parse the geometric parameters of common feature points of ship pipe fittings, and convert them into a Blender-compatible format; the virtual rendering unit is used to configure a multi-light source environment in Blender, adjust model pose parameters, simulate stacking and occlusion scenes, and superimpose Gaussian noise and motion blur to generate a highly simulated interference image; the labeling unit is used to use the Labelme tool to manually / automatically label the six types of feature points in the mixed data set and generate an XML label file; the storage unit is used to export and store the generated standardized ship pipe fitting common feature point data set into a distributed database to support data management and rapid retrieval; The model training module is deployed on edge intelligent devices to collect data from the assembly site and detect pipe features through external cameras. The module is equipped with an optimized YOLOv5s model, supports mixed dataset training, and detects common feature points of various types of ship pipe fittings, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves. It identifies and counts the number of various feature points to form a multidimensional feature vector: V 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common characteristic points of other ship fittings; The recommendation decision module includes a feature extraction unit, a K-nearest neighbor algorithm number matching unit and a human-computer interaction interface. The feature extraction unit is used to extract the multidimensional feature vector V from the cascade detection module. 检测 =[v1,v2,v3,v4,v5,v6,…,v n The K-nearest neighbor algorithm number matching unit is used to perform global similarity matching with the pipe fitting information database using the KNN algorithm, and uses the Euclidean distance to calculate the similarity between the detection vector and the feature vectors of each pipe fitting model in the pipe fitting information database. The human-computer interaction interface is used to build a graphical interface based on PyQt5, which displays the detection box, the recommended pipe fitting number list, and the process guidance animation of the relevant pipe fittings in real time. It also supports touch screens or external devices. On-site staff conduct secondary confirmation and final selection of the pipe fitting number based on the recommended pipe fitting number list, and key data is synchronized to the blockchain. The federated learning platform includes an edge training unit, a gradient encryption aggregation unit, and a global model delivery unit. The edge training unit is used to deploy a cluster of edge detection devices, fine-tune the detection model locally, and compress the model through channel pruning and knowledge distillation methods to make the model better adapted to edge devices. The gradient encryption aggregation unit is used to encrypt the model gradient data packets using the Paillier homomorphic encryption algorithm, transmit them to the cloud for aggregation, dynamically allocate weights according to the node data volume ratio, and update the global model. The global model delivery unit is used to securely deliver the updated model parameters to the edge nodes through the HTTPS protocol, and verify the version number to ensure consistency. The blockchain evidence storage module is used to achieve information sharing based on the distributed ledger structure of blockchain technology. That is, all authorized personnel can view the pipe assembly status and data through the chain, and then use the SHA-256 algorithm to generate the detection image hash value, associate the timestamp with the equipment ID, and store the detection results, process parameters, and timestamp of the pipe information on the blockchain, supporting rapid data verification, traceability and anti-tampering query.

2. The ship pipe identification system based on multi-source data enhancement and intelligent recommendation according to claim 1 is characterized by: The system has a two-layer architecture of "cloud + edge". The cloud is used for intelligent evolution, integrating distributed knowledge through federated learning, and continuously updating and optimizing the global model. The edge is used for real-time edge response and efficient pipe feature detection and pipe number recommendation on the device side.

3. A ship pipe identification method based on multi-source data enhancement and intelligent recommendation, characterized by: The method uses the ship pipe identification system based on multi-source data enhancement and intelligent recommendation according to any one of claims 1-2, and includes hybrid data set construction, pipe feature detection model, pipe number recommendation mechanism, data privacy protection mechanism based on federated learning, and detection traceability mechanism based on blockchain. The hybrid dataset is constructed by rendering the CAD model of ship pipe fittings from multiple angles using Blender software to generate a virtual image, which is then combined with the real-shot image to form a hybrid dataset. Labelme is then used to manually or automatically label common feature points of various types of ship pipe fittings in the image, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves. The hybrid dataset contains common feature points of ship pipe fittings. The pipe feature detection model is constructed by optimizing the YOLOv5s network architecture, replacing the loss function with WIoUv3, and optimizing the detection accuracy by dynamically adjusting the geometric penalty weight and the distance attention mechanism. The model is then trained using the mixed dataset in the mixed dataset construction. The model then detects common feature points of multiple types of ship pipe fittings, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and casings, identifies and counts the number of various feature points, and outputs a multidimensional feature vector. The pipe fitting number recommendation mechanism is based on the multi-dimensional feature vector of common feature points of ship pipe fittings output by detection. It uses the KNN algorithm to perform global similarity matching with the pipe fitting information database, sorts the distance from small to large, and outputs the four most likely recommended ship pipe fitting numbers. It also supports manual selection. The privacy protection mechanism based on federated learning trains the model locally on the edge device, encrypts and uploads the model gradient to the cloud for aggregation, and dynamically allocates node weights to optimize the global model. Multiple edge devices train the detection model locally, encrypt the model gradient using the Paillier homomorphic encryption algorithm, and upload it to the cloud for aggregation. The cloud updates the global model using a dynamic weight allocation strategy and sends it to each edge node. The knowledge distillation technology is used to migrate the knowledge of the large cloud model to the small edge model, and the feature distribution is aligned using the KL divergence loss. The blockchain-based detection traceability mechanism is used to support key data traceability, quickly locate error nodes, and realize distributed storage of detection logs based on blockchain technology. Each evidence record includes but is not limited to the detection timestamp, SHA-256 image hash, pipe number, and complete metadata of process parameters. Based on Merkle tree index and chain tracking technology, the block where the abnormal record is located is located, and related data is extracted through a multi-dimensional query interface to automatically generate an error analysis report.

4. The ship pipe identification method based on multi-source data enhancement and intelligent recommendation according to claim 3 is characterized by: The hybrid dataset is constructed for collecting multi-angle image data of pipe fittings at the assembly site using an industrial camera, obtaining real-life images of ship pipe fittings, importing CAD design drawings of ship pipe fittings into Blender software, setting a multi-light source environment to simulate the lighting conditions at the ship assembly site, adjusting the model pose and background parameters, generating a virtual rendered image containing occlusion and corrosion interference, and then mixing the real-life image and the virtual rendered image in a 6:4 ratio to form a standardized dataset. The Labelme tool is used to manually / automatically annotate multiple feature points in the image, including but not limited to elbows, tees, crosses, tapered pipes, flanges, and sleeves, to generate an XML label file containing bounding box coordinates and feature categories.

5. The ship pipe identification method based on multi-source data enhancement and intelligent recommendation according to claim 3 is characterized by: The pipe feature detection model adopts the network architecture of the optimized YOLOv5 detection model, replaces CIoU in the bounding box loss function with WIoUv3, and dynamically adjusts parameters. It monitors the aspect ratio difference and edge feature error in real time during training, and dynamically updates the geometric penalty weight coefficient α of WIoUv3 through the distance attention mechanism. The border loss function is as follows: L WIoUv3 =rL WIoUv1 The specific formula is: L WIoUv1 =R WIoU L IoU L IoU =1-IoU Among them, IoU is the interaction ratio, xgt, ygt are the coordinates of the center point of the real box, x, y are the coordinates of the center point of the predicted box, Wg, Hg are the sizes of the smallest enclosing boxes; the mapping of outlier β and gradient gain r is controlled by hyperparameters α and δ.

6. The ship pipe identification method based on multi-source data enhancement and intelligent recommendation according to claim 3 is characterized by: The pipe number recommendation mechanism is used to extract the multidimensional feature vector V from the detection model. 检测 =[v1,v2,v3,v4,v5,v6,…,v n ], where v1 is the number of elbows, v2 is the number of tees, v3 is the number of crosses, v4 is the number of tapered pipes, v5 is the number of flanges, v6 is the number of sleeves, and v n The number of common characteristic points of other ship fittings; The pipe fitting number recommendation mechanism is used to perform global similarity matching with the pipe fitting information database through the KNN algorithm, and the Euclidean distance is used to calculate the similarity between the detection vector and the feature vector of each pipe fitting model in the pipe fitting information database. The formula is: v 数据库，i is the i-th eigenvector in the pipe fitting information database; The pipe fitting number recommendation mechanism is used to sort the distances from smallest to largest to obtain the four most likely ship pipe fitting numbers, and push them as recommendation results to the manual interaction interface for on-site staff to view and select.

7. The ship pipe identification method based on multi-source data enhancement and intelligent recommendation according to claim 3 is characterized by: The data protection mechanism based on the federated learning framework is used for local training on edge devices and aggregation of cloud models. Specifically: Step 1: Local model adaptation on edge devices (1) Model initialization: Each edge device obtains the latest version of the global model from the cloud and performs lightweight adaptation based on the computing power of the edge device at the production site. It removes redundant structures through pruning technology, retains the key feature extraction layer, and deploys a lightweight detection model locally. (2) Local training: Input the local dataset; set the training rounds to 50, the batch size to 8, and the learning rate to 0.001×0.95 t , t is the current training round, using SAM optimizer to improve model generalization; During training, the model simultaneously learns the detection task and imitates the knowledge features of complex cloud models, optimizing its own performance by comparing the output of intermediate layers. (3) Gradient calculation: After training is completed, the edge device calculates the adjustment direction of the model parameters and normalizes the gradient to limit its numerical range to prevent information leakage. Subsequently, random noise is added to the processed gradient to ensure that the training details of a single device cannot be reversely inferred. Step 2: Gradient encrypted transmission and dynamic aggregation in the cloud (1) Gradient encryption protection: Paillier homomorphic encryption technology is used to encrypt the gradient, allowing the cloud to perform aggregate calculations without decryption. The encrypted data packets are transmitted to the cloud via a secure communication protocol. Even if they are intercepted during the process, the original content cannot be decrypted. (2) Dynamic weight allocation: The cloud allocates aggregate weights based on the data contribution and quality of each device. Devices with larger data volumes and more stable data quality receive higher weights. An exponential decay strategy is used to prevent a small number of devices from dominating the global model, ensuring fairness among multiple parties. Step 3: Global model update and distribution (1) Joint parameter update: The cloud merges the encrypted gradients of each device according to the weights to generate an updated global model. The update process uses a noise cancellation mechanism to ensure that the added random noise does not affect the model validity after aggregation; (2) Knowledge distillation and model compression: Through knowledge distillation technology, the key recognition capabilities of complex cloud models are refined into lightweight models. Specifically, the edge model is made to mimic the characteristic response patterns of the large cloud model, reducing the number of parameters while maintaining high accuracy. (3) Lightweight model delivery: Only the changed parts of the model are delivered to the edge device instead of the complete parameters, which greatly reduces the amount of transmitted data. The update package uses digital signature verification to prevent tampering. After the device receives it, it is automatically integrated into the local model.

8. The ship pipe identification method based on multi-source data enhancement and intelligent recommendation according to claim 3 is characterized by: The blockchain-based inspection and traceability mechanism is used to support the trusted storage of key data and rapid error location. Inspection results and process parameters are written into a distributed ledger in the form of structured records. Each record contains a timestamp, SHA-256 image hash, pipe number, manual operation log, and process parameter fields. The ledger adopts a block storage structure, with each block having a capacity of 1MB and accommodating 50-100 records. Blocks are linked together into a chain through cryptographic hashing to ensure that they cannot be tampered with globally. At the same time, Merkle trees are used to aggregate record hash levels internally, and the root hash is written into the block header to achieve second-level verification of single data. The ledger provides a query interface that supports multi-dimensional retrieval by timestamp range, pipe model code, and process parameters, locates abnormal nodes in the assembly process, and generates error reports in PDF / CSV format.

Images