Federal learning-based ship inspection method and system, and computer program product

Abstract

The invention discloses a federated learning-based ship inspection method and system and a computer program product, and relates to the field of ship inspection, and the method comprises the steps: deploying nodes at each inspection point, and deploying a federated learning center server at the center; an initial face recognition model is constructed, and initial parameters are configured and distributed; each node uses the preprocessed image to update and generate a model parameter update quantity and uploads the model parameter update quantity; aggregating all model parameter update quantities to obtain new model parameters, distributing the new model parameters to update the model parameters of the nodes, and updating the federal learning center server and the nodes to obtain a new model; repeatedly updating until a final model of which the identification accuracy of each node reaches a preset threshold value is obtained, distributing final model parameters to all nodes, and deploying the final models of all nodes; and calling the deployed final model by each node, executing face recognition to generate a polling recognition result, and uploading the polling recognition result. The method has the effect of high inspection face recognition precision.

Description

Technical Field

[0001] This application relates to the field of ship inspection, and in particular to a ship inspection method, system and computer program product based on federated learning. Background Technology

[0002] Ship equipment inspection is crucial for navigational safety, and currently relies primarily on crew members manually clocking in and out to record inspection routes. This method has significant drawbacks: crew members must actively operate the clocking-in device, which is prone to missed clocking-ins or unauthorized clocking-ins, leading to insufficient record accuracy; the lack of automated monitoring during inspections may result in incomplete checks; the additional clocking-in hardware increases costs and has a high failure rate in high-temperature, high-humidity, and vibrating ship environments. While facial recognition technology can replace manual clocking-in, the complex lighting and vibration environments on ships reduce recognition accuracy, and centrally uploading facial data to a central server for processing poses a risk of crew privacy data leakage. Therefore, a ship inspection method that combines high-precision automated recognition is urgently needed. Summary of the Invention

[0003] To address the issues of low authenticity in existing manual inspection check-in systems and low accuracy in automated identification, this application provides a ship inspection method, system, and computer program product based on federated learning.

[0004] Firstly, the ship inspection method based on federated learning provided in this application adopts the following technical solution: S1, deploy local nodes at each inspection point and deploy a federated learning center server at the central location to establish a federated learning network; S2. The Federated Learning Center server builds the initial face recognition model and configures the initial face recognition model parameters, and distributes the initial face recognition model parameters to each local node; Each local node is equipped with an independent edge computing unit, which is used to preprocess local personnel image samples and update face recognition model parameters. S3. Each local node independently updates the face recognition model parameters using the preprocessed local personnel image samples to generate its own independent model parameter update amount and upload it to the Federated Learning Center server. S4. The Federated Learning Center server aggregates all uploaded model parameter updates to obtain new face recognition model parameters, and distributes them to each local node to synchronously update the face recognition model parameters of each local node. At the same time, the Federated Learning Center server and each local node update the new face recognition model based on the new face recognition model parameters. S5. Repeat steps S3-S4 until the final face recognition model is updated and the face recognition accuracy of each local node reaches the preset threshold. Stop iterating and synchronously distribute the final face recognition model parameters corresponding to the final face recognition model to all local nodes to complete the update and deployment of the final face recognition model of all local nodes. S6. Each local node calls the deployed final face recognition model, performs face recognition to generate inspection recognition results, and uploads the inspection recognition results to the Federated Learning Center server.

[0005] By adopting the above technical solution, the Federated Learning Center server achieves distributed optimization and synchronous updates of the face recognition model by distributing initial face recognition model parameters and aggregating the model parameter updates uploaded by each local node. This avoids uploading original local personnel image samples, effectively protecting the security of user privacy data. Each local node independently performs local sample preprocessing and parameter updates using edge computing units, reducing data transmission volume and central processing burden, improving overall computing efficiency, and reducing network bandwidth consumption. Through iterative optimization until all local nodes reach the preset face recognition accuracy threshold, the global robustness and high-precision generalization ability of the face recognition model are ensured. Finally, the synchronous deployment of the face recognition model and the uploading of face recognition inspection results simplify the system maintenance process, support efficient distributed intelligent inspection, reduce centralized storage and hardware deployment costs, and enhance the system's real-time response capability.

[0006] Optionally, the face recognition model includes a face detection sub-model for quickly detecting face regions from local personnel image samples; a face alignment sub-model for aligning the detected face regions to obtain a frontal face image; and a face recognition sub-model for performing face recognition on the frontal face image to generate inspection recognition results.

[0007] By adopting the above technical solutions, the face detection sub-model significantly shortens the preprocessing time of local personnel image samples and reduces processing latency by efficiently locating face regions, thereby improving the overall response speed. The face alignment sub-model performs standardization processing on face regions, effectively eliminating interference from pose deviation and illumination differences, ensuring the consistency and stability of frontal face images, thereby enhancing the accuracy and reliability of subsequent recognition. The face recognition sub-model performs deep feature matching based on the aligned images, quickly generating accurate inspection and recognition results, optimizing decision-making efficiency and system automation level. Through phased processing, the adaptability and robustness of the face recognition model in complex ship inspection environments are strengthened.

[0008] Optionally, the face detection sub-model uses the YOLOv8 lightweight network to quickly detect face regions; the face alignment sub-model uses a key point detection algorithm to adjust the face region to a frontal pose; and the face recognition sub-model uses the MobileNetV3 network combined with the ArcFace loss function to perform face recognition on frontal images.

[0009] By adopting the above technical solutions, the face detection sub-model uses the lightweight YOLOv8 network, significantly improving the real-time performance and efficiency of face region detection, reducing computational latency and edge device resource consumption, and is suitable for rapid response scenarios in ship inspection. The face alignment sub-model accurately adjusts face pose through key point detection algorithms, effectively overcoming interference from lighting changes and motion blur in the ship environment, ensuring the standardization and consistency of frontal face images, and greatly enhancing the accuracy of subsequent face recognition. The face recognition sub-model combines the lightweight characteristics of the MobileNetV3 network with the strong discriminative power of the ArcFace loss function to efficiently perform feature extraction and matching, generating high-precision inspection and recognition results, and optimizing decision-making speed and system automation. This targeted combination of technologies not only enhances the efficiency of local edge computing and reduces reliance on federated learning center servers and network bandwidth consumption, but also improves the adaptability and reliability of the face recognition model under complex ship inspection conditions by optimizing the lightweightness and robustness of the face recognition model, realizing low-cost, high-efficiency distributed face recognition applications.

[0010] Optionally, preprocessing of personnel image samples may include image denoising, brightness equalization, normalization, and image enhancement.

[0011] By adopting the above technical solutions, image denoising effectively eliminates image noise interference caused by humidity and vibration in the ship environment, improving data quality; brightness equalization technology adaptively adjusts the differences in cabin lighting to ensure that facial features are clearly distinguishable; normalization processing unifies image scale and color space, reducing model learning complexity; and image enhancement strengthens key texture details, providing high-value image input for subsequent face recognition.

[0012] Optionally, the Federated Learning Center server aggregates all uploaded model parameter updates, including the model parameter updates received by the Federated Learning Center server from each local node, and uses a federated averaging algorithm to weight and aggregate all model parameter updates.

[0013] By adopting the above technical solution, the federated learning center server uses a federated averaging algorithm to weight and aggregate the model parameter updates uploaded by each local node, significantly improving the efficiency and fairness of face recognition model optimization. This algorithm dynamically allocates weights based on the distribution characteristics of local sample data from each node, avoiding the negative impact of single-node data bias on the face recognition model and ensuring that the aggregation result fully integrates ship inspection features from multiple scenarios. The weighting mechanism effectively balances the differences in data volume among different nodes, enhancing the model's generalization ability and robustness, enabling it to adapt to complex and ever-changing ship environments. Simultaneously, this lightweight aggregation strategy significantly reduces the computational load and communication overhead of the federated learning center server, avoiding network congestion caused by frequent interactions.

[0014] Optionally, the inspection identification results may include crew member identity, equipment point number, inspection time, or unidentified information.

[0015] By adopting the above technical solutions, the structured output of key fields of the inspection and identification results accurately binds the local personnel's operations, inspection point locations and execution time, providing a complete and traceable digital record for ship inspection management. At the same time, it automatically reports abnormal situations through unidentified information, significantly improving the efficiency of safety supervision and the speed of decision response.

[0016] Secondly, the ship inspection system based on federated learning provided in this application includes: a central service module, multiple local node modules, an edge computing unit, and a face recognition unit; The central service module is deployed in the ship's central control room. It is used to build and distribute the initial face recognition model parameters to each local node module; receive and aggregate the model parameter updates uploaded by all local node modules, generate new face recognition model parameters and distribute them to all local node modules; and receive and store the inspection recognition results uploaded by the face recognition unit. Multiple local node modules are distributed and deployed at various inspection points, and are connected to the central service module via a local area network. They include edge computing units and face recognition units. Edge computing units are deployed within each local node module to preprocess local personnel image samples, independently update face recognition model parameters based on the preprocessed local personnel image samples, generate model parameter update quantities, and upload them to the central service module. The face recognition unit is deployed in each local node module; it is used to receive face recognition model parameters; call the face recognition model to perform face recognition to generate inspection recognition results, and upload the inspection recognition results to the central service module.

[0017] Thirdly, this application provides a computer program product, which includes a computer program or instructions that enable the computer program or instructions to implement the method described in any of the first aspects.

[0018] Understandably, the federated learning-based ship inspection system provided in the second aspect and the computer program product provided in the third aspect are both used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0019] In summary, this application includes at least one of the following beneficial technical effects: 1. Due to the mechanism of independently updating face recognition model parameters by distributed local nodes and only uploading the parameter update, local personnel image samples do not need to leave the local node device, effectively protecting user privacy data security; at the same time, the edge computing unit performs local preprocessing and face recognition model parameter updates, greatly reducing the load and network transmission pressure on the federated learning center server, significantly reducing system latency and bandwidth consumption, and achieving efficient and low-cost model training.

[0020] 2. Because the Federated Learning Center server dynamically aggregates the model parameter updates of each local node and updates the face recognition model synchronously, combined with the iterative termination condition of the preset face recognition accuracy threshold, it ensures that the final face recognition model fully learns the complex environmental features of multiple inspection points, such as changes in lighting and posture differences in ship cabins, which significantly enhances the cross-scene adaptability and recognition accuracy of face recognition and avoids model degradation caused by local data deviations.

[0021] 3. After the final face recognition model is deployed synchronously on each local node, it generates structured inspection recognition results containing crew member identity, equipment point number, and inspection time or unrecognized information by performing face recognition, and transmits them back to the Federated Learning Center server in real time, forming a closed-loop inspection management process. This significantly improves the real-time performance and traceability of ship safety supervision, while reducing manual verification costs and providing a highly reliable solution for distributed inspection scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the ship inspection method based on federated learning provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the ship inspection system based on federated learning provided in the embodiments of this application. Detailed Implementation

[0024] This application discloses a ship inspection method, system, and computer program product based on federated learning, as illustrated below. Figure 1-2 This application will be described in further detail.

[0025] Reference Figure 1 This is a flowchart illustrating a ship inspection method based on federated learning in an embodiment of this application.

[0026] S1. Deploy local nodes at each inspection point and deploy a federated learning center server at the central location to establish a federated learning network.

[0027] In this context, "inspection point" refers to the physical location within the ship's engine room where equipment inspections are required. "Local node" refers to a computing unit deployed at the inspection point, possessing independent data processing capabilities, including image acquisition equipment and edge computing units. "Federated learning center server" refers to the central computing equipment deployed in the ship's control room, used to coordinate distributed training. "Federated learning network" refers to a distributed computing architecture composed of the federated learning center server and multiple local nodes connected via communication links, supporting the interaction of facial recognition model parameters without requiring centralized transmission of original local personnel image samples.

[0028] In practical implementation, to achieve automated inspection, a federated learning infrastructure needs to be built within the ship's engine room. First, local nodes are fixedly installed in the vicinity of each inspection point in the engine room. Each local node integrates high-definition image acquisition equipment and an edge computing unit. The image acquisition equipment is responsible for capturing facial image samples of the inspection personnel, and the edge computing unit performs localized facial recognition model calculations. Simultaneously, a federated learning center server is deployed in the ship's control room as the central coordination unit. After physical deployment, all local nodes are interconnected with the federated learning center server via the ship's internal wired local area network, forming a star-topology federated learning network architecture. In this architecture, the federated learning center server acts as a parameter aggregation node, and each local node acts as a distributed training node. The network communication protocol uses an encrypted transmission mechanism to ensure the security of parameter exchange, ensuring that the edge computing units of all local nodes can directly synchronize model parameters with the federated learning center server, but prohibiting direct data transmission across nodes, thus constructing a distributed computing environment that conforms to the federated learning paradigm.

[0029] S2, the Federated Learning Center server builds the initial face recognition model and configures the initial face recognition model parameters, and distributes the initial face recognition model parameters to each local node.

[0030] Each local node is equipped with an independent edge computing unit, which is used to preprocess local personnel image samples and update face recognition model parameters.

[0031] Here, the face recognition model refers to a composite neural network structure that includes a face detection sub-model, a face alignment sub-model, and a face recognition sub-model. The face recognition model parameters refer to the set of weight values ​​in the face recognition model. Distribution refers to the transmission of face recognition model parameter data to each local node through a federated learning network. Edge computing units perform local computing tasks on each local node.

[0032] Specifically, the parameters of a face recognition model refer to the set of weight values ​​within the model. These weights, as adjustable variables within the neural network, learn the mapping from input data to output during training, enabling the model to extract abstract features from original local human image samples and complete the recognition task. This set of weight values ​​represents the connection strength between neurons in each layer of the neural network and has different data structures in the three face recognition sub-models. In the face detection sub-model, the weight values ​​of the convolutional layers are stored as four-dimensional tensors, controlling the extraction of spatial features from the image. In the face alignment sub-model, the weight values ​​of the fully connected layers are stored as two-dimensional matrices, responsible for the geometric transformation mapping of facial keypoint coordinates. In the face recognition sub-model, the weight values ​​of the depthwise separable convolutional layers are stored as composite tensors, encoding identity-discriminating features. All sub-model weight values ​​are organized and stored in multi-dimensional tensor form, with the tensor dimension determined by the network structure.

[0033] The weights serve two main purposes: firstly, they act as the smallest synchronization unit in distributed training, with each local node iteratively adjusting weights based on local personnel image samples using backpropagation to optimize the model's adaptability to specific environments, such as cabin lighting and vibration. Secondly, they serve as a privacy protection medium, transmitting weight updates only between each local node and the federated learning center server, preventing the original personnel image sample data from leaving the local node. Throughout the entire lifecycle of the face recognition model, the federated learning center server uses mathematical algorithms to generate an initial set of weights; the edge computing units calculate the gradient of the loss function based on preprocessed local image samples and update their local weights; finally, the federated learning center server uses a weighted averaging algorithm to fuse the weight updates from each local node to generate a globally optimized new set of weights, ultimately forming a distributed face recognition model with high generalization capabilities.

[0034] The Federated Learning Center server initializes the face recognition model in the ship's control room. The face detection sub-model uses a lightweight YOLOv8 network architecture to achieve fast face region localization, the face alignment sub-model uses a key point detection algorithm to standardize face pose, and the face recognition sub-model uses a MobileNetV3 network architecture combined with the ArcFace loss function to achieve feature extraction. After initialization, the central server generates model parameter data packets, which are transmitted and distributed to each local node through an encrypted network. The edge computing units of each local node receive and store these initial parameters, laying the foundation for subsequent local preprocessing and parameter update calculations.

[0035] Specifically, the face detection sub-model uses a lightweight YOLOv8 network architecture to achieve fast face region localization. Its weight values ​​are stored in a four-dimensional tensor structure of depthwise separable convolutional layers. Through a single-stage end-to-end detection mechanism, the image is divided into uniform grid cells. Each cell optimizes the bounding box coordinate prediction by updating its weight values. The face region coordinates are output in real time in the edge computing unit, enabling the face detection sub-model to output face region coordinates in real time even under the limited computing power of the edge computing unit, overcoming the image sample blurring interference caused by ship engine room vibration. The face alignment sub-model relies on a key point detection algorithm to standardize face pose. Its weight values ​​are represented as a two-dimensional matrix of fully connected layers. Through convolutional neural network regression, the weight values ​​are updated to accurately calibrate the coordinates of key facial points such as the outer canthi of the eyes and the tip of the nose, constructing a facial set structure. Using these key point coordinates, an affine transformation matrix is ​​generated to perform rotation, scaling, and translation operations on the original face region, eliminating the perspective deviation caused by changes in the posture of the inspection personnel, and generating a standard face image with a unified frontal orientation. The face recognition sub-model integrates MobileNetV3 and ArcFace loss functions to achieve high discriminative feature extraction. Its weight values ​​are stored as composite tensors in the depthwise separable convolution module. The optimized weight values ​​are searched through the neural architecture, which enhances illumination robustness while maintaining low computational complexity. At the same time, the weight value update driven by the ArcFace loss function compresses the intra-class distribution and expands the inter-class distance in the hyperspherical feature space, so that the extracted face feature vector has strong identity discrimination ability and provides high-precision input for face feature matching.

[0036] S3. Each local node independently updates the face recognition model parameters using the preprocessed local personnel image samples to generate its own independent model parameter update and upload it to the Federated Learning Center server.

[0037] Among them, the preprocessed local personnel image sample refers to the standardized personnel face image sample data that has undergone image denoising, brightness equalization, normalization, and image enhancement in the local node edge computing unit; the model parameter update refers to the set of difference data between the face recognition model parameters trained by the local node based on the preprocessed image and the initial model parameters; and uploading refers to the operation of encrypting and transmitting the model parameter update to the federated learning center server through the federated learning network.

[0038] In implementing this step, each local node first loads the initial face recognition model parameters distributed by the federated learning center server into the edge computing unit, and calls the preprocessed crew face image samples stored locally as training input. Subsequently, the edge computing unit executes the local training process of the face recognition model, iteratively optimizing the model parameters through the backpropagation algorithm to adapt the face recognition model to the specific environmental conditions of the inspection point. After training, the difference between the current face recognition model parameters and the initial face recognition model parameters is calculated and combined into the model parameter update amount. Finally, the model parameter update amount is encrypted and packaged and transmitted to the federated learning center server through the ship's internal LAN, ensuring that the original personnel image sample data is always retained on the local node, realizing distributed knowledge sharing under privacy protection.

[0039] S4. The Federated Learning Center server aggregates all uploaded model parameter updates to obtain new face recognition model parameters, and distributes them to each local node to synchronously update the face recognition model parameters of each local node. At the same time, the Federated Learning Center server and each local node update to obtain a new face recognition model based on the new face recognition model parameters.

[0040] Weighted aggregation refers to the operation performed by the federated learning center server to calculate a weighted average of the difference vectors based on preset weight coefficients. The new face recognition model parameters refer to the set of globally optimized face recognition model parameters generated after weighted aggregation.

[0041] The Federated Learning Center server first receives the encrypted model parameter updates transmitted from each local node, decrypts and stores the data after verifying its integrity. When performing aggregation operations, the Federated Learning Center server calls the Federated Averaging algorithm to process the model parameter updates. The formula is... ,in This is the parameter update amount for the face recognition model, calculated and output by the Federated Learning Center server. The total number of all local nodes participating in the aggregation is equal to the number of inspection points; For local node index identifiers (all of which appear below) (These all have this meaning, which will not be elaborated further below.) They are generally numbered according to the physical deployment order of local nodes and are unique numbers for a specific inspection point, for example... It can represent the host monitoring point; For the first The weight coefficient of each local node reflects the weight of the contribution of the local personnel image sample data of that node; For the first The amount of model parameter updates uploaded by each local node , For the first The parameters of the face recognition model trained on the local node. These are the initial face recognition model parameters for initial distribution.

[0042] The specific formula for the weighting coefficient is as follows: ,in, The weighting factor is the number of local personnel image samples. , For the first The number of local crew member face image samples stored on the node. This represents the maximum number of samples across all nodes. For the quality assessment factor of local personnel image samples, among which, , Image sharpness is scored to indicate its resistance to blurring; The attitude diversity score covers yaw or pitch angle scenarios. The lighting balance score is represented by the proportion of overexposure and underexposure. , is the quality dimension adjustment coefficient. The federated averaging algorithm assigns weight coefficients to each local node. These weight coefficients are dynamically calculated from two dimensions of weighting factors. The first dimension's weighting factor is the number of crew face image samples for each local node; a larger number of image samples results in a higher weight. The second dimension's weighting factor is the image sample data quality, quantified by evaluating indicators such as image sharpness, pose diversity, and illumination evenness; a higher quality level results in a higher weight. Based on these weight coefficients, all model parameter updates are weighted and summed to generate new face recognition model parameter updates. Then, the initial face recognition model parameters are superimposed with the new face recognition model parameter updates to output a new set of face recognition model parameters. The formula is... ,in This is a new set of face recognition model parameters to be distributed to each local node. , The definition is consistent with the above. The Federated Learning Center server immediately distributes the new set of face recognition model parameters to all local nodes through an encrypted channel. Upon receiving the parameters, the edge computing units of each local node overwrite the old face recognition model parameters, completing the synchronous update of the face recognition model parameters. At the same time, the Federated Learning Center server itself also loads the new face recognition model parameters into the face recognition model copy, achieving version unification of the face recognition model between the Federated Learning Center server and the edge computing units.

[0043] S5. Repeat steps S3-S4 until the final face recognition model is updated and the face recognition accuracy of each local node reaches the preset threshold. Stop the iteration and synchronously distribute the final face recognition model parameters corresponding to the final face recognition model to all local nodes to complete the update and deployment of the final face recognition model of all local nodes.

[0044] Here, the preset threshold refers to the minimum performance standard that the face recognition model must meet, pre-set by the federated learning center server. The final face recognition model refers to the globally optimized model that meets the preset accuracy requirements after multiple rounds of federated learning iterations. The final face recognition model parameters refer to the set of weight values ​​corresponding to this model. Synchronous distribution refers to the operation of ensuring that all local nodes simultaneously receive the face recognition model parameters through an atomic broadcast protocol. Update deployment refers to the operation of loading the final face recognition model parameters and activating the final face recognition model on local nodes.

[0045] In this step, the Federated Learning Center server initiates a distributed evaluation process after each round of facial recognition model parameter aggregation. The edge computing units of each local node use their local validation sets to perform facial recognition model inference, calculate the recognition accuracy of the current facial recognition model in the cabin environment, and encrypt and report the results. The Federated Learning Center server verifies the accuracy data reported by each node, and determines that the facial recognition model has converged only if the recognition accuracy of all local nodes consistently meets the preset threshold requirements. At this point, the Federated Learning Center server terminates the federated learning iteration process, encrypts and packages the final facial recognition model parameters generated in the last round of aggregation, and sends an atomic broadcast to all local nodes via the ship's internal LAN. Each local node synchronously receives the parameter data packet, decrypts it, overwrites the old facial recognition model parameters, and reinitializes the facial recognition model. After deployment, each local node automatically executes the local verification process, sending a model readiness confirmation signal to the Federated Learning Center server. The Federated Learning Center server aggregates the confirmation signals to complete the final global state synchronization, marking the final facial recognition model as effective across the entire federated learning network.

[0046] S6. Each local node calls the deployed final face recognition model, performs face recognition to generate inspection recognition results, and uploads the inspection recognition results to the Federated Learning Center server.

[0047] The final face recognition model refers to the face recognition model deployed on the local node after iterative optimization in step S5 of federated learning. Performing face recognition refers to the process by which the local node calls this model to perform face detection, alignment, and recognition calculations on real-time collected local personnel image samples. Inspection recognition results refer to structured data records containing crew member identities, equipment point numbers, inspection times, or unidentified information; uploading refers to the operation of transmitting the inspection recognition results to the federated learning center server via an encrypted communication protocol.

[0048] In practice, when inspection personnel enter the monitoring range of the local node image acquisition device, the device automatically triggers the acquisition of a high-definition image sequence. The edge computing unit immediately loads the final face recognition model. The face detection sub-model first locates the bounding box coordinates of the face region in the image. After successful detection, the face alignment sub-model generates a standardized frontal face image through key point spatial transformation. The face recognition sub-model then extracts the feature vector of the standardized frontal face image and performs similarity matching calculation with the locally encrypted crew feature library. The cosine similarity between the feature vector to be identified and the feature vectors of each crew member in the crew feature library is calculated. If the maximum cosine similarity is greater than the preset recognition threshold, the recognition is considered successful, and the corresponding crew member's identity identifier is output. Combined with the current inspection point location code and the coordinated world time timestamp accurate to the second, a valid inspection recognition result is generated. If matching fails for multiple consecutive frames, an unrecognized status code is generated, and the inspection recognition result "No valid crew member identified" is output. Finally, the edge computing unit encapsulates and encrypts the inspection and identification results, and transmits them to the Federated Learning Center server through the ship's internal local area network. After the transmission is completed, the original image data cached locally is automatically cleared, and only the structured records are retained to ensure that the crew's privacy data does not leave the local node.

[0049] In practical applications, the Federated Learning Center server continuously monitors the latest inspection record timestamps of each inspection point in the ship's inspection database. For each inspection point, it independently calculates the time interval between the current moment and the last valid inspection. When this time interval exceeds a preset period, such as 24 hours, an early warning response process is triggered. First, a standardized early warning data packet is generated, encapsulating the early warning level code, the physical location code of the inspection point, the precise duration of the uninspected period, and a suggested re-inspection route plan. Second, the early warning data packet is synchronously pushed to the alarm devices in the ship's control room and the crew's mobile application via the ship's internal LAN. Upon receiving the packet, the control room alarm device activates an audible and visual alarm signal and prominently displays the 3D location of the early warning inspection point on the monitoring screen. The crew's mobile application parses the data packet and generates a pop-up alarm with a navigation path. Finally, a closed-loop tracking mechanism is initiated. When the crew arrives at the early warning inspection point and completes a compliant inspection, the inspection identification results uploaded by the local node automatically trigger the early warning status to be lifted, and the Federated Learning Center server updates the device status and stops the alarm signal. This ensures that no inspection is missed.

[0050] Reference Figure 2 This is a schematic diagram of a ship inspection system based on federated learning in an embodiment of this application.

[0051] The ship inspection system based on federated learning includes a central service module 1 and multiple local node modules 2. The local node modules 2 include an edge computing unit 21, a face recognition unit 22, and an image acquisition device 23.

[0052] The central service module 1 is deployed in the ship's central control room. It is used to build and distribute the initial face recognition model parameters to each local node module 2; receive and aggregate the model parameter updates uploaded by all local node modules 2, generate new face recognition model parameters and distribute them to all local node modules 2; receive the model parameter updates generated by the edge computing unit 21; and store the inspection recognition results uploaded by the face recognition unit 22.

[0053] The central service module is deployed in the ship's central control room and consists of a central processing server 11 and a distributed storage device 12. The central processing server is equipped with a multi-core processor and high-speed memory for model parameter aggregation calculation, and the distributed storage device uses a redundant disk array to store model parameters and inspection records.

[0054] Multiple local node modules 2 are distributed and deployed at various inspection points, and are connected to the central service module 1 via a local area network. Edge computing units 21 are deployed within each local node module 2, and are used to preprocess local personnel image samples, independently update the face recognition model parameters based on the preprocessed local personnel image samples, generate model parameter update quantities, and upload them to the central service module 1.

[0055] The face recognition unit 22 is deployed in each local node module 2; it is used to receive face recognition model parameters; call the face recognition model to perform face recognition to generate inspection recognition results, and upload the inspection recognition results to the central service module 1.

[0056] Image acquisition device 23 is used to capture images of inspection personnel and provide on-site image data of people to be identified for edge computing module 21 and face recognition unit 22.

[0057] Multiple local node modules are distributed across various inspection points. Each local node module includes an image acquisition device 23, an edge computing unit hardware processor (i.e., edge computing unit 21), and a dedicated face recognition acceleration chip (i.e., face recognition unit 22), all interconnected via a local data bus. The system establishes a communication connection through the ship's internal local area network switch, and the central service module transmits data with all local node modules using an encrypted communication protocol.

[0058] Specifically, image acquisition device 23 is a high-definition industrial camera with a wide dynamic range, fixedly installed near the inspection points in the cabin to capture images of inspection personnel. The edge computing unit hardware processor is an embedded artificial intelligence computing chip, integrating an independent memory module to perform local personnel image sample preprocessing and local face recognition model training. The dedicated face recognition acceleration chip is a neural network processor that loads the face recognition model and performs real-time inference calculations. During system operation, each local node module acquires raw personnel image sample data through image acquisition device 23. After preprocessing, the edge computing unit hardware processor updates the face recognition model parameters and generates model parameter update values, which are then encrypted and transmitted to the central service module. The central processing server of the central service module aggregates global parameters and distributes new face recognition model parameters to each node. The dedicated face recognition acceleration chip loads the latest face recognition model, performs the recognition task, and sends the encrypted inspection recognition results back to the central service module for storage, forming a federated learning closed loop.

[0059] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0060] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0061] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0062] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. "Above," "below," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A ship inspection method based on federated learning, characterized in that: This includes S1, deploying local nodes at each inspection point, and deploying a federated learning center server in a central location to establish a federated learning network; S2. The Federated Learning Center server builds the initial face recognition model and configures the initial face recognition model parameters, and distributes the initial face recognition model parameters to each local node; Each local node is equipped with an independent edge computing unit, which is used to preprocess local personnel image samples and update face recognition model parameters. S3. Each local node independently updates the face recognition model parameters using the preprocessed local personnel image samples to generate its own independent model parameter update amount and upload it to the Federated Learning Center server. S4. The Federated Learning Center server aggregates all uploaded model parameter updates to obtain new face recognition model parameters, and distributes them to each local node to synchronously update the face recognition model parameters of each local node. At the same time, the Federated Learning Center server and each local node update the new face recognition model based on the new face recognition model parameters. S5. Repeat steps S3-S4 until the final face recognition model is updated and the face recognition accuracy of each local node reaches the preset threshold. Stop iterating and synchronously distribute the final face recognition model parameters corresponding to the final face recognition model to all local nodes to complete the update and deployment of the final face recognition model of all local nodes. S6. Each local node calls the deployed final face recognition model, performs face recognition to generate inspection recognition results, and uploads the inspection recognition results to the Federated Learning Center server.

2. The ship inspection method based on federated learning according to claim 1, characterized in that: The face recognition model includes a face detection sub-model, which is used to quickly detect face regions from local personnel image samples; The face alignment sub-model is used to align detected face regions to obtain a frontal face image; The face recognition sub-model is used to perform face recognition on frontal face images and generate inspection recognition results.

3. The ship inspection method based on federated learning according to claim 2, characterized in that: The face detection sub-model uses the YOLOv8 lightweight network to quickly detect face regions; the face alignment sub-model uses a key point detection algorithm to adjust the face region to a frontal pose; and the face recognition sub-model uses the MobileNetV3 network combined with the ArcFace loss function to perform face recognition on frontal images.

4. The ship inspection method based on federated learning according to claim 1, characterized in that: The preprocessing of personnel image samples includes image denoising, brightness equalization, normalization, and image enhancement.

5. The ship inspection method based on federated learning according to claim 1, characterized in that: The federated learning center server aggregates all uploaded model parameter updates, including the model parameter updates received by the federated learning center server from each local node, and uses a federated average algorithm to weight and aggregate all model parameter updates.

6. The ship inspection method based on federated learning according to claim 1, characterized in that: The inspection and identification results include crew member identity, equipment point number, inspection time, or unidentified information.

7. A ship inspection system based on federated learning, characterized in that, It includes a central service module, multiple local node modules, an edge computing unit, and a face recognition unit; The central service module is deployed in the ship's central control room. It is used to build and distribute the initial face recognition model parameters to each local node module; receive and aggregate the model parameter updates uploaded by all local node modules, generate new face recognition model parameters and distribute them to all local node modules; and receive and store the inspection recognition results uploaded by the face recognition unit. Multiple local node modules are distributed and deployed at various inspection points, and are connected to the central service module via a local area network. They include edge computing units and face recognition units. Edge computing units are deployed within each local node module to preprocess local personnel image samples, independently update face recognition model parameters based on the preprocessed local personnel image samples, generate model parameter update quantities, and upload them to the central service module. The face recognition unit is deployed in each local node module; it is used to receive face recognition model parameters; call the face recognition model to perform face recognition to generate inspection recognition results, and upload the inspection recognition results to the central service module.

8. A computer program product, characterized in that, The computer program product includes a computer program or instructions that enable the computer program or instructions to perform the steps in the ship inspection method based on federated learning as described in any one of claims 1 to 6.

Images