Art examination evaluation method based on cloud edge-end collaboration

Abstract

The invention discloses an art examination evaluation method based on cloud edge-end collaboration, and belongs to the technical field of art examination evaluation. According to the method, a cloud-edge node-terminal collaborative architecture is adopted, and full-process management and control are realized through four steps of collaborative configuration before examination, task allocation and encrypted communication link establishment, examination process management and control, data aggregation and return and result generation, wherein the cloud issues evaluation parameters, encrypts resources and lightens a model; the edge node completes resource verification, task allocation and state monitoring, and the terminal collects answer data and extracts features, and supports temporary storage and continuous transmission after network disconnection. According to the scheme, the defect of a single framework is avoided, the resource transmission and data processing efficiency is improved, the examination process stability and data safety are guaranteed, large-scale and standardized evaluation of art examinations is achieved, and the requirements of large-capacity resource processing and special scenes of the art examinations are accurately met.

Description

Technical Field

[0001] This invention relates to the field of art examination and assessment technology, and in particular to an art examination and assessment method based on cloud-edge-device collaboration. Background Technology

[0002] With the popularization of arts education and the advancement of digital transformation in education, arts examinations and assessments have gradually moved away from the traditional offline single model and are developing towards large-scale, standardized, and digitalization. Arts examinations have significant unique characteristics, requiring the processing of large amounts of multimedia resources such as images, audio, and video, placing high demands on the stability of resource transmission, the efficiency of data processing, and the standardization of the examination process. Currently, various digital examination architectures have emerged in the industry, mainly including pure cloud architecture, pure local area network architecture, and a hybrid model combining traditional offline and digital approaches. Pure cloud architecture relies on cloud servers for centralized resource management and remote access, while pure local area network architecture focuses on rapid data interaction between local devices. Both architectures are applied in arts examinations of different scales and scenarios. Simultaneously, cloud-edge-device collaborative technology is increasingly widely used in digital systems across various industries. By combining macro-management in the cloud with localized processing at edge nodes and terminals, new technical approaches are provided to solve resource allocation and data processing problems in complex scenarios, laying a technical foundation for the optimization and upgrading of arts examination and assessment systems.

[0003] However, existing technologies related to art exam assessment still have many problems that urgently need to be solved, which directly affect the efficiency, stability, and adaptability of the exams. First, existing technologies lack a cloud-edge-device collaborative architecture design for the entire art exam process. Pure cloud architectures rely too heavily on the stability of the public network, and are prone to process interruptions due to resource transmission pressure during large-scale exams. Pure local area network architectures, on the other hand, struggle to achieve unified resource distribution and efficient data aggregation, failing to meet the needs of both scalability and localization. Second, there is a lack of deep adaptation to the special scenarios of art exams, and a lack of a full-process collaborative mechanism of "pre-exam resource preparation - in-exam process control - post-exam data feedback" has been formed. This results in low efficiency of pre-exam resource allocation, lack of in-exam status monitoring, and easy data loss in case of network outages or other emergencies. Finally, the technology connections between various links are not smooth. Functions such as resource verification, identity verification, and data encryption are scattered and not deeply integrated with the collaborative architecture, resulting in insufficient controllability of the exam process, high data security risks, and an inability to meet the core requirements of standardization, efficiency, and security for art exam assessments. Furthermore, they are not adaptable to large-scale assessment applications of different scales and scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an art examination and evaluation method based on cloud-edge-device collaboration.

[0005] The objective of this invention is achieved through the following technical solution: A cloud-edge-device collaborative method for art examination assessment is provided, which includes the following steps: S1. Perform pre-exam collaborative configuration, distribute evaluation activity parameters, encrypted exam resources and lightweight feature extraction models to the cloud, and verify the integrity of resources and model adaptability after receiving them at the edge nodes, while simultaneously detecting the hardware performance and network connectivity of the candidate's terminal. S2. Based on the terminal status and resource adaptation results, the edge node evaluates the task allocation scheme through a collaborative decision-making formula, pushes exam notifications to ready terminals and allocates exclusive test papers, and establishes an encrypted communication link between the edge and the terminal. S3. During the exam, the terminal collects artistic answer data and completes local feature extraction through a lightweight model. Edge nodes receive feature data and answer status in real time. When the network is normal, the data is synchronized to the cloud. When the network is disconnected, the data is temporarily stored locally and the network recovery status is continuously monitored. S4. Edge nodes aggregate the answer data, feature results, and exam logs from all terminals, encrypt and package them, and then send them back to the cloud in batches according to the collaboration strategy. The cloud combines the data with the global model to perform comprehensive processing and generate art exam evaluation results.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Configure the exam duration, scoring dimensions, and answer format parameters of the evaluation activity in the cloud, upload the exam resources containing test questions and materials and encrypt and encapsulate them, and distribute the lightweight feature extraction model adapted to edge nodes at the same time. S1.2. Edge nodes download encrypted exam resources and lightweight models from the cloud, and complete integrity verification by comparing resource file identifiers and check codes, and test the compatibility between the model and local hardware. S1.3. The edge node sends a detection command to the candidate's terminal, collects the terminal's storage space, computing performance, network bandwidth and client version information, filters ready terminals that meet the examination requirements, generates a detection report containing terminal identifier and status parameters and synchronizes it to the cloud.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. The edge node calls the collaborative decision-making module, inputting terminal hardware performance, network bandwidth, resource size, and model complexity parameters, and calculates using the formula: ; Calculate the task unloading priority, where These are the weighting coefficients. For network bandwidth, For edge computing capabilities, For cloud computing capabilities, For data volume, For terminal load; S2.2. Determine the task allocation between local processing and cloud collaboration based on priority results, push exam start notifications to ready terminals, and assign unique test papers with no duplicates according to preset rules; S2.3. Establish a dedicated communication link between the edge node and the candidate's terminal through an asymmetric encryption algorithm, verify the terminal's access permissions, and ensure data transmission security.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Candidates browse the decrypted exam resources through the terminal and complete artistic responses such as drawing, singing, and video recording. The terminal collects data on the response process and the final result in real time. S3.2. The terminal calls the built-in lightweight feature extraction model to extract feature points and standardize the format of the answer data, generate feature vectors and upload them to the edge nodes in real time. S3.3. When the network is normal, the edge nodes will synchronize the feature vector and the response status to the cloud in real time. When the network is interrupted, the terminal will temporarily store the response data and feature results in the local encrypted storage area. The edge nodes will continuously monitor the network status and trigger the terminal to automatically resume transmission after the network is restored.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Edge nodes collect complete answer data, feature vectors, answer status records, and examination process logs from all candidate terminals, classify and organize them according to candidate identifiers and examination subjects, and package and aggregate them into data packets of a unified format; S4.2. Through calculation formula: ; Calculate the cloud collaboration weight, where, For activation function, For network latency, For delay weighting, For edge load, For load weight, As a bias term, the timing of batch data return is selected based on the weighting results; S4.3. After receiving the data packet in the cloud, integrity verification is completed by comparing the check code. Combined with the global evaluation model, the feature vector and the answer data are comprehensively analyzed to generate standardized results that include the candidate's score, feature attainment status and overall evaluation trend.

[0010] Furthermore, in step S1, the edge node performs adaptation optimization on the downloaded lightweight model, adjusts the model parameters to match the local hardware computing power, and compares the terminal detection results with the cloud-based preset standards, marks terminals that do not meet the requirements and provides optimization suggestions to ensure that the examination resources and models can be stably accessed.

[0011] Furthermore, in step S3, the edge node collects the login status, answering progress, model running status, and network transmission status of the candidate's terminal in real time, integrates and processes the collected multi-dimensional status information, presents the overall examination process through a visual interface, and identifies and logs network outages, timeouts, and abnormal model running situations in real time.

[0012] Furthermore, in step S4, the edge node encrypts the aggregated data packets using a symmetric encryption algorithm to generate a unique checksum corresponding to the data packets. If a network interruption occurs during the backhaul process, the identifiers of the transmitted and untransmitted data segments are recorded. After the network is restored, the optimal timing for resuming transmission is determined based on the collaborative weight formula, and the data backhaul is resumed from the point of interruption. After the backhaul is completed, the cloud verifies the data integrity by comparing the checksums.

[0013] An art examination and evaluation system based on cloud-edge-device collaboration is provided. The system includes a cloud collaboration module, an edge processing module, and a terminal execution module. The cloud-based collaboration module is used to configure evaluation activity parameters, store encrypted exam resources, and deploy a global evaluation model. It receives data from the edge processing module and generates evaluation results. The edge processing module downloads exam resources and a lightweight model from the cloud-based collaboration module and performs verification and adaptation. It evaluates task allocation schemes through collaborative decision-making formulas, establishes an encrypted communication link between the edge and the terminal, receives answer data and feature results from the terminal execution module, aggregates them, and sends them back to the cloud according to the collaboration strategy. The terminal execution module receives exam resources, supports candidates in completing artistic answers, extracts answer features through a built-in lightweight model, temporarily stores relevant data during network interruptions, and retransmits it after network recovery.

[0014] Furthermore, the cloud-based collaborative module includes a parameter configuration unit, a resource storage unit, a global model unit, and a data processing unit; the parameter configuration unit configures evaluation-related parameters, the resource storage unit encrypts and stores examination resources and model files, the global model unit provides comprehensive evaluation algorithm support, and the data processing unit verifies the returned data and generates evaluation results; The edge processing module includes a resource verification unit, a collaborative decision-making unit, a communication encryption unit, a data aggregation unit, and a backhaul control unit. The resource verification unit verifies the integrity and compatibility of resources and models. The collaborative decision-making unit calculates task allocation and backhaul strategies using formulas. The communication encryption unit establishes a secure edge link. The data aggregation unit organizes and categorizes relevant answer data. The backhaul control unit executes data backhaul according to the collaborative strategy. The terminal execution module includes a resource receiving unit, a response acquisition unit, a feature extraction unit, and a local temporary storage unit. The resource receiving unit receives and decrypts exam resources, the response acquisition unit acquires artistic response data, the feature extraction unit extracts response features through a lightweight model, and the local temporary storage unit stores relevant data when the network is disconnected and supports resume transmission.

[0015] The beneficial effects of this invention are: (1) Relying on the cloud-edge-device collaborative architecture and the full-process collaborative operation of "pre-exam configuration-task allocation-process control-data feedback", the dependence on a single architecture is reduced, the examination process is ensured to proceed continuously, and the adaptability of large-scale art examination assessment is improved. (2) By using lightweight feature extraction and data standardization processing technology, we can achieve unified and efficient processing of different types of artistic answer data, reduce manual intervention, and enhance the standardization of the evaluation process and the accuracy of data processing. (3) Combining multiple encryption mechanisms with fault-tolerant designs such as resume transmission after network outage and local temporary storage, the risk of data leakage, tampering and loss is avoided, the impact of the external environment on the examination is reduced, and the stability of the examination process and data security are guaranteed. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of an art examination and evaluation method based on cloud-edge-device collaboration; Figure 2 A schematic diagram of the structure of an art examination and evaluation system based on cloud-edge-device collaboration is provided for an embodiment; Figure 3 The following is a flowchart illustrating the specific steps of an art examination and evaluation method based on cloud-edge-device collaboration, provided as an example. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 See Figure 1 This paper presents an art examination assessment method based on cloud-edge-device collaboration, which includes the following steps: S1. Perform pre-exam collaborative configuration, distribute evaluation activity parameters, encrypted exam resources and lightweight feature extraction models to the cloud, and verify the integrity of resources and model adaptability after receiving them at the edge nodes, while simultaneously detecting the hardware performance and network connectivity of the candidate's terminal. S2. Based on the terminal status and resource adaptation results, the edge node evaluates the task allocation scheme through a collaborative decision-making formula, pushes exam notifications to ready terminals and allocates exclusive test papers, and establishes an encrypted communication link between the edge and the terminal. S3. During the exam, the terminal collects artistic answer data and completes local feature extraction through a lightweight model. Edge nodes receive feature data and answer status in real time. When the network is normal, the data is synchronized to the cloud. When the network is disconnected, the data is temporarily stored locally and the network recovery status is continuously monitored. S4. Edge nodes aggregate the answer data, feature results, and exam logs from all terminals, encrypt and package them, and then send them back to the cloud in batches according to the collaboration strategy. The cloud combines the data with the global model to perform comprehensive processing and generate art exam evaluation results.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Configure the exam duration, scoring dimensions, and answer format parameters of the evaluation activity in the cloud, upload the exam resources containing test questions and materials and encrypt and encapsulate them, and distribute the lightweight feature extraction model adapted to edge nodes at the same time. S1.2. Edge nodes download encrypted exam resources and lightweight models from the cloud, and complete integrity verification by comparing resource file identifiers and check codes, and test the compatibility between the model and local hardware. S1.3. The edge node sends a detection command to the candidate's terminal, collects the terminal's storage space, computing performance, network bandwidth and client version information, filters ready terminals that meet the examination requirements, generates a detection report containing terminal identifier and status parameters and synchronizes it to the cloud.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. The edge node calls the collaborative decision-making module, inputting terminal hardware performance, network bandwidth, resource size, and model complexity parameters, and calculates using the formula: ; Calculate the task unloading priority, where These are the weighting coefficients. For network bandwidth, For edge computing capabilities, For cloud computing capabilities, For data volume, For terminal load; S2.2. Determine the task allocation between local processing and cloud collaboration based on priority results, push exam start notifications to ready terminals, and assign unique test papers with no duplicates according to preset rules; S2.3. Establish a dedicated communication link between the edge node and the candidate's terminal through an asymmetric encryption algorithm, verify the terminal's access permissions, and ensure data transmission security.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Candidates browse the decrypted exam resources through the terminal and complete artistic responses such as drawing, singing, and video recording. The terminal collects data on the response process and the final result in real time. S3.2. The terminal calls the built-in lightweight feature extraction model to extract feature points and standardize the format of the answer data, generate feature vectors and upload them to the edge nodes in real time. S3.3. When the network is normal, the edge nodes will synchronize the feature vector and the response status to the cloud in real time. When the network is interrupted, the terminal will temporarily store the response data and feature results in the local encrypted storage area. The edge nodes will continuously monitor the network status and trigger the terminal to automatically resume transmission after the network is restored.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Edge nodes collect complete answer data, feature vectors, answer status records, and examination process logs from all candidate terminals, classify and organize them according to candidate identifiers and examination subjects, and package and aggregate them into data packets of a unified format; S4.2. Through calculation formula: ; Calculate the cloud collaboration weight, where, For activation function, For network latency, For delay weighting, For edge load, For load weight, As a bias term, the timing of batch data return is selected based on the weighting results; S4.3. After receiving the data packet in the cloud, integrity verification is completed by comparing the check code. Combined with the global evaluation model, the feature vector and the answer data are comprehensively analyzed to generate standardized results that include the candidate's score, feature attainment status and overall evaluation trend.

[0023] In some embodiments, in step S1, the edge node performs adaptation optimization on the downloaded lightweight model, adjusts the model parameters to match the local hardware computing power, and compares the terminal detection results with the cloud-preset standards, marks terminals that do not meet the requirements and provides optimization suggestions to ensure that the examination resources and models can be stably accessed.

[0024] In some embodiments, in step S3, the edge node collects the login status, answering progress, model running status and network transmission status of the candidate's terminal in real time, integrates the collected multi-dimensional status information, presents the overall examination process through a visual interface, and identifies and logs network outages, timeouts, and abnormal model running situations in real time.

[0025] In some embodiments, in step S4, the edge node encrypts the aggregated data packets using a symmetric encryption algorithm to generate a checksum uniquely corresponding to the data packets. If a network interruption occurs during the backhaul process, the identifiers of the transmitted data segments and the untransmitted data segments are recorded. After the network is restored, the optimal timing for resuming the transmission is determined based on the collaborative weight formula, and the data backhaul is resumed from the point of interruption. After the backhaul is completed, the cloud verifies the data integrity by comparing the checksums.

[0026] See Figure 2 This paper presents an art examination and evaluation system based on cloud-edge-device collaboration, which includes a cloud collaboration module, an edge processing module, and a terminal execution module. The cloud-based collaboration module is used to configure evaluation activity parameters, store encrypted exam resources, and deploy a global evaluation model. It receives data from the edge processing module and generates evaluation results. The edge processing module downloads exam resources and a lightweight model from the cloud-based collaboration module and performs verification and adaptation. It evaluates task allocation schemes through collaborative decision-making formulas, establishes an encrypted communication link between the edge and the terminal, receives answer data and feature results from the terminal execution module, aggregates them, and sends them back to the cloud according to the collaboration strategy. The terminal execution module receives exam resources, supports candidates in completing artistic answers, extracts answer features through a built-in lightweight model, temporarily stores relevant data during network interruptions, and retransmits it after network recovery.

[0027] In some embodiments, the cloud collaboration module includes a parameter configuration unit, a resource storage unit, a global model unit, and a data processing unit; the parameter configuration unit configures evaluation-related parameters, the resource storage unit encrypts and stores examination resources and model files, the global model unit provides comprehensive evaluation algorithm support, and the data processing unit verifies the returned data and generates evaluation results. The edge processing module includes a resource verification unit, a collaborative decision-making unit, a communication encryption unit, a data aggregation unit, and a backhaul control unit. The resource verification unit verifies the integrity and compatibility of resources and models. The collaborative decision-making unit calculates task allocation and backhaul strategies using formulas. The communication encryption unit establishes a secure edge link. The data aggregation unit organizes and categorizes relevant answer data. The backhaul control unit executes data backhaul according to the collaborative strategy. The terminal execution module includes a resource receiving unit, a response acquisition unit, a feature extraction unit, and a local temporary storage unit. The resource receiving unit receives and decrypts exam resources, the response acquisition unit acquires artistic response data, the feature extraction unit extracts response features through a lightweight model, and the local temporary storage unit stores relevant data when the network is disconnected and supports resume transmission.

[0028] In some embodiments, the method adopts a four-layer collaborative architecture of cloud-edge server-teacher-student, where the cloud is responsible for global resource configuration and result aggregation, the edge server (IES.ExamServer.exe) is deployed in the local area network and is responsible for exam process control, data processing and AI inference, the teacher's end (the edge server process embedded in the Electron application) provides the exam management interface, and the student's end (the front-end SPA remotely loaded by the Electron application) provides the answering interface. Multi-layer collaboration is achieved through real-time communication via HTTPS API and SignalR.

[0029] Example 2 This embodiment provides a specific implementation process for an art examination assessment method based on cloud-edge-device collaboration. Through the collaborative cooperation of the cloud, edge nodes, and terminals, it achieves standardized and efficient operation of the art examination assessment. The overall process includes pre-examination collaborative configuration, task allocation and link establishment, examination process control, data aggregation and feedback, and result generation. Figure 3 As shown, the specific steps are as follows: S1. Pre-exam collaborative configuration: Pre-exam collaborative configuration is fundamental to the smooth conduct of art exams. Through collaborative operations between cloud and edge nodes, it completes the preparation of exam-related parameters and resources, as well as terminal status detection, ensuring the availability of exam resources and compatibility of terminal devices. Specifically, it includes the following sub-steps: S1.1. Configure the exam duration, scoring dimensions, and answer format parameters for the evaluation activity in the cloud, upload exam resources containing questions and materials, encrypt and encapsulate them, and simultaneously distribute a lightweight feature extraction model adapted to edge nodes: The cloud serves as the configuration and distribution center for examination resources. First, the administrator completes the basic parameter settings for the evaluation activities through the operation interface. The examination duration is configured according to the characteristics of the art examination subjects (such as painting requiring a longer creation time and music performance requiring a fixed performance time) and the answering requirements. The scoring dimensions cover core evaluation aspects such as the technical performance of the artwork (such as the smoothness of lines in painting and the pitch control in music), creative conception (such as the expression of the theme of the work and the innovation of the form of expression), and completion (such as the completeness of the work and the absence of obvious flaws). The answer format parameters clearly specify the storage format of the candidate's answer results (such as PNG format for images, WAV format for audio, and MP4 format for video) to ensure the consistency of subsequent data processing.

[0030] Exam resources consisting of test questions and materials must be encrypted and packaged using the AES encryption algorithm. During the packaging process, the unique identifier of the resource, the associated subject information, and the validity period are integrated to prevent the resource from being illegally accessed or tampered with during transmission and storage.

[0031] Meanwhile, a lightweight feature extraction model is distributed from the cloud to edge nodes. This model adopts a simplified CNN architecture, which includes an input layer, convolutional layer, pooling layer, fully connected layer, and output layer. The input layer receives standardized answer data, the convolutional layer uses 3×3 convolutional kernels for feature extraction, the pooling layer uses max pooling to reduce dimensionality, the fully connected layer compresses the feature dimension to generate feature vectors, and the output layer outputs standardized feature results. The number of model parameters is controlled within a preset range to reduce redundant calculations, reduce computational complexity while ensuring feature extraction accuracy, and avoid affecting the examination process due to hardware performance limitations.

[0032] In some embodiments, the teacher's client adopts a hybrid architecture of an Electron framework with an embedded ASP.NET Core edge server. The edge server process is automatically created and a health check is performed at startup. Exclusive use of port 8326 is ensured through single instance locking. The teacher's client interface communicates with the embedded server via HTTPS. When the server starts, it automatically loads a self-signed certificate (cert.pem / key.pem) and provides a health check interface (GET / index / health) and a graceful shutdown interface (GET / index / shutdown).

[0033] In some embodiments, a speech recognition model (SenseVoiceSmall, belonging to the FunASR series) and a punctuation prediction model adapted to the edge server are distributed from the cloud. The edge node provides speech-to-text services for voice-based question answering. Music evaluation adopts a third-party platform iframe embedding method, and the edge node only provides API proxy, AES encryption and decryption, and OSS relay services. Artistic answer data such as paintings and videos are stored directly without local feature extraction.

[0034] S1.2. Edge nodes download encrypted exam resources and lightweight models from the cloud, and complete integrity verification by comparing resource file identifiers and checksums, and check the compatibility between the model and local hardware. After the edge node starts up, it establishes a secure communication connection with the cloud via HTTPS and initiates a download request for encrypted exam resources and a lightweight model. The download request carries the edge node's unique identifier and authentication information. After successful verification by the cloud, data transmission begins. During the download process, the edge node receives data in blocks and writes it to a temporary storage area, recording the number and size of the received data blocks in real time. After the download is complete, an integrity verification process is initiated.

[0035] Resource file identification verification involves comparing the unique file identifiers (filename, size, modification time, etc.) of downloaded resources with the identifiers in the resource list provided by the cloud to confirm that the resource files have not been replaced or tampered with. Checksum verification involves calculating the MD5 checksum of the downloaded resource (obtained by concatenating block hashes of the resource file) and verifying its consistency with the original MD5 checksum provided by the cloud to ensure that no data loss or corruption occurred during transmission. After completing resource integrity verification, the edge node further tests the compatibility between the lightweight model and local hardware. It obtains parameters such as the number of CPU cores, memory capacity, and GPU computing power by calling the local hardware information query interface. It compares the minimum hardware configuration required for model operation (preset in the model configuration file) with the local hardware parameters and runs model test cases to check if the model startup time and single inference time are within reasonable ranges. If incompatibility issues are detected (such as insufficient memory causing model startup failure or inference time exceeding a threshold), the edge node records the relevant information and reports it to the cloud.

[0036] In some embodiments, edge nodes can use a dual checksum comparison method to improve the reliability of resource integrity verification. That is, the MD5 checksum and SHA-256 checksum of the resource are calculated simultaneously and compared with the corresponding checksum provided by the cloud. Only when both checksums match are the resource integrity verification considered successful.

[0037] S1.3. The edge node sends a detection command to the candidate's terminal, collects the terminal's storage space, computing performance, network bandwidth, and client version information, filters ready terminals that meet the examination requirements, generates a detection report containing terminal identifiers and status parameters, and synchronizes it to the cloud: Edge nodes send UDP testing commands to all connected candidate terminals via the local area network. These commands include a list of testing items and data return format requirements. Upon receiving the commands, the candidate terminals automatically start their local testing programs: storage space information is obtained by calling the system file management interface to query the remaining capacity of the terminal's storage partition; computational performance information is obtained by running preset lightweight computational test programs (such as matrix multiplication) to calculate computation time, peak CPU usage, and peak memory consumption; network bandwidth information is obtained by sending fixed-size test data packets to the edge nodes, recording the time difference between packet transmission and reception, and calculating real-time upload and download speeds. Client version information is obtained by reading the version configuration file in the client's installation directory. The terminal encapsulates all collected information into JSON data according to a specified format and sends it back to the edge nodes via the TCP protocol.

[0038] After receiving the information from each candidate's terminal, the edge node compares it with the preset examination requirements (such as remaining storage space not less than the preset value, peak CPU usage not exceeding the preset percentage, network download speed not less than the preset value, and client version being the latest compatible version). It then selects all ready terminals that meet the requirements, compiles the terminal identifier, various status parameters, and the judgment results of whether the terminals are ready into a test report, and synchronizes the test report to the cloud through the communication link, so that the cloud can monitor the terminal's readiness status in real time.

[0039] S1.4. Edge nodes perform adaptability optimization on the downloaded lightweight model, adjusting model parameters to match local hardware computing capabilities. Simultaneously, they compare terminal detection results with cloud-based preset standards, marking terminals that do not meet the requirements and providing optimization suggestions to ensure stable access to examination resources and models. Edge nodes optimize the lightweight model based on local hardware computing power data (such as the number of CPU cores and GPU memory size): if the local GPU computing power is strong, the number of convolutional kernels in the convolutional layers is appropriately increased (not exceeding the maximum number supported by the model) to improve feature extraction accuracy; if the local hardware computing power is limited, the number of neurons in the fully connected layers is reduced to reduce the model's computational load, and the batch size of the model is adjusted to a value that fits the local memory, so that the model can run efficiently on local hardware and avoid stuttering or crashing.

[0040] For terminal detection results, edge nodes compare each terminal's storage space, computing performance, network bandwidth, and client version with the minimum standards preset in the cloud: terminals with insufficient storage space are marked, and optimization suggestions such as "clean up redundant local files to free up storage space" are provided; terminals with insufficient computing performance are marked, and optimization suggestions such as "close unnecessary background programs" are provided; terminals with insufficient network bandwidth are marked, and optimization suggestions such as "check network connection or change network access point" are provided; terminals with outdated client versions are marked, and optimization suggestions such as "update client to the latest compatible version" are provided. Through model optimization and terminal problem marking, it is ensured that examination resources and models can be stably used in subsequent examinations, reducing examination anomalies caused by device or model compatibility issues.

[0041] S2. Task allocation and communication link establishment: Based on the pre-exam collaborative configuration results, edge nodes complete task allocation scheme evaluation, exam notification push, exam paper distribution, and secure communication link establishment, ensuring the orderly conduct of the exam process. This includes the following sub-steps: S2.1. The edge node calls the collaborative decision-making module, inputting parameters such as terminal hardware performance, network bandwidth, resource size, and model complexity, and determines the task offloading priority through calculation: After the edge node starts, it invokes the built-in collaborative decision-making module. This module adopts a modular design, including a parameter input unit, a calculation unit, and a result output unit. The parameters of the input module are all derived from the detection results of the pre-exam collaborative configuration phase. Among them, the terminal hardware performance parameters are obtained by extracting standardized values ​​of core indicators such as CPU computing speed (number of instructions completed per unit time), memory capacity, and GPU floating-point computing power; the network bandwidth parameter is the average of the real-time download and upload speeds between the terminal and the edge node; the resource size parameter is the total number of bytes of resources such as the test question package and material package for this exam; and the model complexity parameter is a standardized value calculated from the total number of model parameters, the number of convolutional layers, and the computational cost (FLOPs).

[0042] The calculation formula is: ; in, This is a weighting coefficient used to adjust the influence of each parameter on the task unloading priority. Its value ranges from 0 to 1 and satisfies the following conditions: ; Standardized values ​​for network bandwidth; The standardized numerical value for edge computing capability is obtained by weighting parameters such as CPU processing speed, GPU floating-point processing power, and memory bandwidth of edge nodes; The standardized values ​​for cloud computing capabilities are calculated from the hardware performance parameters of the cloud server (preset in the edge node configuration file); Standardize the data volume to a numerical value; The standardized value for terminal load is calculated by weighting the terminal's current CPU utilization and memory usage. The collaborative decision-making module outputs the task offloading priority for each terminal using this calculation formula (with a value ranging from 0 to 1). The higher the priority value, the more suitable the terminal is to offload some tasks to edge nodes or the cloud for processing, and vice versa.

[0043] In some embodiments, the collaborative decision-making module may employ a dynamic weight adjustment mechanism to adjust the weights based on the real-time status during the examination process. The value of can be adjusted, for example, when network bandwidth is generally low, by increasing . The weighting of network bandwidth makes its impact on priority more significant, ensuring that the task allocation scheme can adapt to real-time changes in system status.

[0044] In some embodiments, the edge node extracts the corresponding test paper from the pre-downloaded assessment package based on the student's login information and exam configuration, and pushes it to the terminal via an HTTPS communication link. The unique identifier of the test paper is bound to the candidate's identity and stored in the LiteDB database to ensure the accuracy and uniqueness of the test paper allocation. The speech recognition model supports two deployment modes: embedded mode (default) loads the SenseVoice model within the edge server process and uses a locking mechanism to ensure thread safety; standalone service mode deploys the ASR function as a standalone process (port 5000) and supports concurrent recognition through an object pool.

[0045] S2.2. Based on the priority results, determine the task allocation between local processing and cloud collaboration, push exam start notifications to ready terminals, and assign unique exam papers according to preset rules: Based on the task offloading priority of each terminal, the edge nodes clearly define the task division between local processing and cloud collaboration: For terminals with priority values ​​higher than a preset threshold, computationally intensive tasks such as complex feature extraction (e.g., multi-dimensional feature fusion calculation) and deep data analysis (e.g., preliminary determination of feature compliance) are offloaded to the edge nodes for processing, with the terminals only responsible for data collection and basic preprocessing (e.g., data format conversion); for terminals with priority values ​​lower than or equal to the preset threshold, the terminals themselves complete data collection, complete feature extraction, and data format standardization, while the edge nodes are only responsible for data reception and status monitoring. After the task division is determined, the edge nodes push exam start notifications to all ready terminals via the local area network. The notification content is displayed in a pop-up window, including key information such as the exam start time, answer instructions, submission deadline, and exception handling methods, ensuring that candidates receive exam-related requirements in a timely manner. Meanwhile, edge nodes allocate a unique exam paper to each candidate from the locally stored exam paper library according to a preset random allocation rule (based on the modulo operation of the hash value of the candidate's identifier). During the allocation process, the unique identifier of the exam paper is used for deduplication verification to ensure that the exam paper is not duplicated and that the exam paper obtained by each candidate is unique, thus avoiding the leakage of exam questions or cheating by candidates. After the allocation is completed, the unique identifier of the exam paper is associated with the candidate's identity information and stored. After the candidate is identified, he / she can accurately obtain his / her own unique exam paper.

[0046] S2.3. Establish a dedicated communication link between the edge node and the candidate's terminal using an asymmetric encryption algorithm to verify the terminal's access permissions and ensure data transmission security: Edge nodes use the RSA asymmetric encryption algorithm to generate a 2048-bit public and private key pair. The public key is broadcast and distributed to all ready terminals via the local area network, while the private key is stored in the edge node's encrypted secure area (a hardware encrypted storage module independent of the system disk), accessible only to the collaborative decision-making module and the data transmission module. After receiving the public key, the candidate terminal uses it to encrypt its own terminal physical address, candidate identity identifier, client verification code, and other information, generating an encrypted data packet which is then sent to the edge node.

[0047] After receiving the encrypted data packet, the edge node decrypts it using its private key to obtain information such as the terminal's physical address, the examinee's identity identifier, and the client's verification code. It then compares the terminal's physical address with the physical addresses in the pre-exam ready terminal list, compares the client's verification code with the preset valid client verification code, and confirms whether the examinee's identity identifier is on the examinee list for this exam. Only terminals that pass all three comparisons are deemed to have passed access permission verification and establish a dedicated communication link with the edge node. Terminals that fail verification will be denied access, and the edge node will send them a permission verification failure message and will not provide any access to exam resources.

[0048] Once the communication link is established, all data transmitted between the terminal and the edge node (such as answer data, status information, control commands, etc.) is transmitted through this encrypted link. The data sender uses the public key to encrypt the data, and the receiver uses the private key to decrypt the data, ensuring that the data is not stolen, tampered with, or forged during transmission, thus guaranteeing the security of the examination data transmission.

[0049] In some embodiments, the edge server uses a self-signed SSL certificate to establish an HTTPS connection, and the certificate is automatically installed on a locally trusted root certificate authority. The Electron client verifies the server certificate through a certificate whitelist mechanism, supporting all versions of TLS 1.0-1.3 to ensure compatibility with older systems such as Windows 7. The teacher client verifies whitelisted domains through the app.on('certificate-error') event, and the student client ignores non-whitelisted certificate errors through the --ignore-certificate-errors parameter.

[0050] In some embodiments, both the teacher and student clients adopt the Electron cross-platform desktop application framework. The teacher client embeds an edge server process and provides a management interface, while the student client remotely loads a Vue single-page application as the answering interface. Both clients support features such as self-signed certificate verification, automatic version updates, and offline / online dual-mode switching, and are optimized for x86 architecture for older systems such as Windows 7.

[0051] S3. Examination Process Control: Examination process control is a crucial step in ensuring the fairness and orderliness of the examination. Comprehensive control over the examination process is achieved through terminal data collection, feature extraction, status monitoring, and network outage fault tolerance handling. This includes the following sub-steps: S3.1. Candidates use the terminal to browse the decrypted exam resources and complete artistic responses including drawing, singing, and video recording. The terminal collects data on the response process and the final result in real time. Candidates enter their username and password on the terminal's login screen. The terminal encrypts the identity information and sends it to the edge node. After successful authorization verification at the edge node, the terminal obtains a unique decryption key from the edge node (transmitted via an encrypted communication link). This key is then used to decrypt the locally stored encrypted exam resources using AES. After decryption, candidates can browse the exam content (such as descriptions of painting themes, musical pieces, and video recording requirements), material information (such as reference images and accompaniment audio), and answer requirements (such as artwork size, audio duration, and video resolution) through the terminal's visual interface (developed based on a lightweight UI framework).

[0052] According to the exam requirements, candidates will complete the corresponding art-related tasks: drawing will be completed using the terminal's built-in drawing tools, supporting functions such as brush thickness adjustment, color selection, and layer management; singing will be completed using the terminal's audio capture device (microphone), recording the candidate's audio in real time; video recording will be completed using the terminal's camera, simultaneously capturing video and audio signals. The terminal will collect two types of data in real time during the candidate's performance: One type is the answer process data. Drawing answer process data records the coordinate trajectory, color value, brush thickness, and drawing timestamp of each stroke through the operation log of the drawing tool; music performance process data records the volume changes, frequency fluctuations, and rhythm during the performance through real-time sampling of audio acquisition devices; video recording process data records the action changes, facial expressions, scene stability, and corresponding timestamps of the frame through frame sampling of the camera device, which can fully reflect the candidate's answer process. Another type is the final result data, which includes static or dynamic result data such as artwork files, music audio files, and performance video files generated after the examinee completes the test. The terminal stores the collected data in real time, and manages it according to the naming rules of "examinee identifier-subject name-data type-timestamp" and the preset hierarchical storage path to ensure the orderliness and traceability of the data. At the same time, a data verification mechanism is adopted, and a verification value is calculated after storing a certain number of data blocks to avoid data loss or confusion.

[0053] In some embodiments, the terminal can support the expansion of multiple art forms of answering questions. In addition to painting, singing, and video recording, it can also support other art forms such as calligraphy creation and instrumental performance. The data types collected are adjusted accordingly based on the form of answering questions. For example, calligraphy creation collects data on changes in ink density and stroke continuity, while instrumental performance collects data on timbre characteristics and rhythmic stability, ensuring that it can fully cover the examination needs of different art subjects.

[0054] In some embodiments, the music evaluation function uses an independent front-end project (IES.Music) deployed on the / musicplay path of the edge server, supporting functions such as device debugging, pitch adjustment (pitch rise and fall), speed adjustment (original speed / 0.8x / 0.6x), real-time waveform display, and sheet music following; the recording files support two modes: local server priority upload and Alibaba Cloud OSS backup upload, and the recording plugin uses the self-developed js-recorder, which supports the sheet music display component (OpenSheetMusicDisplay).

[0055] S3.2. The terminal calls the built-in lightweight feature extraction model to extract feature points and standardize the format of the response data, generate feature vectors, and upload them to the edge nodes in real time: While collecting the answer data, the terminal calls the built-in lightweight feature extraction model to process the data at preset time intervals (or data volume thresholds). For drawing-related answer data, the model first converts the image data into a grayscale image, extracts line contour features through edge detection algorithms, calculates color distribution features through color histogram statistics, extracts composition proportion features through image segmentation algorithms, and obtains brushstroke texture features through texture feature extraction algorithms. Each feature is quantized into a value between 0 and 1. For music performance-related response data, the model uses Fourier transform to convert the audio signal into a frequency spectrum, extracting feature points such as fundamental frequency (pitch), frequency variation amplitude (pitch stability), rhythm period (beat accuracy), and spectral energy distribution (timbre), which are also quantified into standardized values. For video recording-related response data, the model extracts action continuity features using frame difference, facial expression features using facial recognition algorithms, and scene suitability features using scene matching algorithms, all quantified into standardized values. After feature point extraction, the model performs format standardization on all feature point data, arranging feature point data of different types and dimensions in a preset order, converting it into a fixed-dimensional numerical matrix (such as a 128-dimensional vector), and then generating standardized feature vectors. This ensures that edge nodes and the cloud can uniformly process and analyze different types of response data.

[0056] After the feature vector is generated, the terminal uploads it to the edge node in real time through the established encrypted communication link. During the upload process, a fragmented upload mechanism is adopted to divide the feature vector into small data blocks. After each block is uploaded, the terminal receives confirmation feedback from the edge node to ensure that the data can be transmitted to the edge node completely and in a timely manner. If no confirmation feedback is received, the data block is automatically retransmitted.

[0057] S3.3. When the network is normal, the edge nodes synchronize the feature vectors and response status to the cloud in real time. When the network is interrupted, the terminal temporarily stores the response data and feature results in the local encrypted storage area. The edge nodes continuously monitor the network status and trigger the terminal to automatically resume transmission after the network is restored. Edge nodes monitor the network connection status with terminals and the cloud in real time through a network status detection module (based on ICMP protocol heartbeat packet detection). When the network status is normal, the edge node receives the feature vector uploaded by the terminal, extracts the terminal identifier, subject, current answering progress (such as the number of questions completed and the remaining time), and packages it with the feature vector. This information is then synchronized to the cloud in real time via an encrypted communication link. The cloud receives the data and stores it in the corresponding data partition, enabling the cloud to monitor the exam progress and the examinee's answering status in real time.

[0058] If a network outage occurs (heartbeat detection timeout), the terminal immediately initiates a local temporary storage mechanism, storing any unuploaded response data (process data and result data) and feature results in a local encrypted storage area. This storage area uses AES encryption algorithm for partitioned encryption, and the encryption key is bound to the terminal identifier, allowing only the current terminal to decrypt and access the data, preventing unauthorized access. Simultaneously, the network status detection module at the edge node continuously and periodically checks the network status (the detection period is a preset fixed value) to ensure timely detection of network recovery.

[0059] Once the network is restored (two consecutive heartbeat packet checks are normal), the edge node sends a resume transmission command to the terminal (containing the latest timestamp and data block identifier of the received data). Upon receiving the command, the terminal automatically reads the data temporarily stored in its local encrypted storage area and resumes transmission to the edge node through the encrypted communication link. During the resume transmission process, the timestamp and data block identifier of the local data are first compared with the information sent by the edge node. Only data that has not been completely uploaded (timestamp later than the latest timestamp or data block identifier not in the received list) is transmitted to avoid duplicate transmission and ensure the efficiency and integrity of data transmission.

[0060] In some embodiments, after receiving the answer data submitted by the terminal, the edge node first persists it to a local embedded NoSQL database (LiteDB), and then adds it to an asynchronous push queue based on a channel. The background push service uses CPU adaptive concurrency control (1-8 concurrent requests) and an exponential backoff retry mechanism (1s→2s→4s) to send the data back to the cloud in batches. The edge node maintains three independent push queues, implemented using BoundedChannel: subject answer queue (DataQueue, capacity 1000 entries), music recording queue (MusicQueue, capacity 500 entries), and activity result queue (ActivityQueue, capacity 500 entries). Each queue is processed independently and does not affect the others.

[0061] S3.4. Edge nodes collect real-time data on the login status, answering progress, model running status, and network transmission status of examinee terminals. This multi-dimensional status information is then fused and processed to present the overall examination process through a visual interface. Real-time identification and logging of network outages, timeouts, and model running anomalies are also implemented. While receiving data uploaded by the terminals, the edge nodes collect multi-dimensional status information of each terminal in real time through encrypted communication links: Login status is determined by the terminal's login session identifier and heartbeat feedback, including logged-in (valid session and normal heartbeat), not logged-in (no valid session), login timeout (session timeout and not renewed), and login failure (identity verification error); Answering progress is determined by the completed answering steps, the current amount of data processed, and the remaining answering time (calculated based on the total exam time and the time already spent answering), as reported by the terminal, including the number of questions completed, the current question number, and the remaining answering time. Information such as the percentage of remaining time; the model running status is determined by the model startup status, inference time, and CPU / GPU utilization reported by the terminal, including normal model startup (successful startup and inference time within the threshold), running lag (inference time exceeds the threshold), and running crash (model process terminates); the network transmission status is determined by the success rate of data transmission, latency, and number of retransmissions, including normal transmission (success rate higher than the threshold, latency lower than the threshold, no retransmission), slow transmission (latency higher than the threshold but no retransmission), and interrupted transmission (number of retransmissions exceeds the threshold or heartbeat packet interruption).

[0062] Edge nodes fuse the collected multi-dimensional status information, using a weighted average method to convert each status parameter into a unified standardized status index (value range 0-1). The weights are preset based on the importance of the status information (e.g., network transmission status has a higher weight than login status). A data fusion algorithm eliminates redundant information and abnormal data interference, resulting in a comprehensive status index for each terminal. The fused status information is presented through a visual monitoring interface on the edge nodes. The interface combines lists and charts. The list displays each terminal's identifier, examinee's name, login status, answering progress, and comprehensive status index. The charts display the overall examination progress (e.g., percentage of logged-in terminals, percentage of terminals that have completed the answering process). For detected network outages, timeouts (terminals without data transmission and status updates exceeding a preset threshold), and model malfunctions, these are marked on the visualization interface with a preset prominent color (e.g., red), and anomaly logs are automatically recorded. The logs include the time of the anomaly, terminal identifier, examinee's name, anomaly type, anomaly parameter value, and current network status, providing a basis for subsequent anomaly handling and traceability.

[0063] S4. Data aggregation and result generation: After the exam, the edge nodes complete data aggregation, encrypted packaging, and back transmission. The cloud verifies and processes the data and generates evaluation results to ensure the security of the exam data and the accuracy of the evaluation results. This process includes the following sub-steps: S4.1. Edge nodes collect complete answer data, feature vectors, answer status records, and examination process logs from all candidate terminals, classify and organize them according to candidate identifier and examination subject, and package and aggregate them into a data package with a unified format: Once the exam ends, the edge node sends a deadline instruction to all terminals. Upon receiving the instruction, the terminals automatically stop collecting answer data, complete the final upload of any unuploaded data, and then lock the answering function. The edge node then initiates the data collection process, sending data collection requests to each terminal via an encrypted communication link. The terminals return complete answer data (process data and result data), feature vectors, answer status records (the trajectory of answer status changes throughout the exam), and exam process logs (terminal operation logs and anomaly records) as requested.

[0064] After receiving all data, the edge nodes group the data according to the examinee's identifier, grouping all subject data for the same examinee into one group. Within each group, they further subdivide by exam subject, grouping answer data, feature vectors, status records, and logs for the same subject into a subgroup, ensuring clear and orderly data classification. After classification and organization, the edge nodes package and aggregate the data, using a ZIP compression algorithm to compress all classified data into a unified format data packet. An encryption password (associated with the edge node identifier and exam batch) is set during compression, and the data packet is named in the format "exam batch-edge node identifier-data aggregation timestamp.zip" for easy identification and management in the cloud. During packaging, data integrity is verified by calculating a checksum for each subgroup and storing it in the data packet's verification information area, ensuring that the data packet does not contain missing or erroneous data, laying the foundation for subsequent data transmission and processing.

[0065] In some embodiments, edge nodes use the LiteDB embedded NoSQL database to store exam data. The database file is located on the local file system (%LocalAppData% / ExamServer / LiteDB / data.db), supporting offline operation and ACID transactions. The core data set includes the evaluation activity table (EvaluationClient), the student result table (EvaluationStudentResult), the student exam paper table (EvaluationStudentPaper), and the push queue list (SubjectPushData / ActivityPushData). All data is persisted locally and then pushed to the cloud through an asynchronous queue.

[0066] S4.2. Calculate the cloud collaboration weight using a computational formula, and select the timing for batch data return based on the weight results: While packaging and aggregating data packets, edge nodes initiate a cloud-based collaborative weight calculation process to determine the timing of batch data packet return. The parameters required for this calculation are obtained through real-time data acquisition: network latency is calculated by sending fixed-size test data packets to the cloud, recording the round-trip time from transmission to reception, and taking the average of multiple tests as the network latency. Edge load balancing is achieved by querying system resource monitoring data from edge nodes to obtain CPU utilization, memory usage, and disk I / O utilization, and then calculating a standardized edge load using weighted averages. (Value range 0-1); Delay weight Load weight and bias terms These are preset parameters, and their values ​​are determined based on historical data transmission and system performance test results to ensure that the weight calculation results accurately reflect the quality of the transmission conditions.

[0067] The calculation formula is: ; in, The Sigmoid activation function maps the calculation results to a weight range of 0-1. Its function expression is: To ensure that the collaborative weight results are within a unified range for easy threshold determination, edge nodes collect network latency and edge load data in real time, and substitute them into the calculation formula to obtain the cloud collaborative weight. When the weight result is greater than a preset threshold (e.g., 0.7), it indicates that the current network status (low latency) and edge load (low resource consumption) are at a relatively good level, suitable for data backhaul, and the edge node starts the batch backhaul process. If the weight result does not reach the threshold, the parameters are continuously collected and the weight is recalculated at preset intervals until the preset threshold is reached before starting backhaul, ensuring the stability and efficiency of the data backhaul process.

[0068] In some embodiments, the preset threshold can be adjusted according to the size of the exam and the amount of data. For large-scale exams with a large amount of data, the threshold can be appropriately lowered (e.g., 0.6) to ensure that data transmission can be completed within a reasonable time. For small-scale exams with a small amount of data, the threshold can be appropriately increased (e.g., 0.8) to prioritize transmission when the network status and edge load are optimal, ensuring the reliability of data transmission.

[0069] In some embodiments, edge nodes employ a dynamic batch processing strategy, automatically adjusting the batch size (5-20 messages / batch) based on push time, with a push interval of 500ms to avoid network congestion; adaptively adjusting the concurrency based on the number of CPU cores (1 core → 1 concurrent, 2 cores → 1 concurrent, 3-4 cores → 2 concurrent, 5-8 cores → number of cores / 2, >8 cores → min(number of cores * 3 / 4, 8)) to ensure network stability and transmission efficiency; and employing an exponential backoff retry mechanism (1s → 2s → 4s, up to 3 times) when push fails.

[0070] S4.3. Edge nodes encrypt the aggregated data packets using a symmetric encryption algorithm, generating a unique checksum corresponding to each data packet. If a network interruption occurs during the backhaul process, the identifiers of the transmitted and untransmitted data segments are recorded. After the network is restored, the optimal resumption time is determined based on the collaborative weight formula, and the data backhaul continues from the point of interruption. After the edge node determines the time for backhaul, it uses the AES-256 symmetric encryption algorithm to perform secondary encryption on the aggregated data packets. The encryption key is determined through a preset key negotiation mechanism between the edge node and the cloud (based on the exchange of keys using an asymmetric encryption algorithm) to ensure that the cloud can decrypt correctly.

[0071] After encryption, the edge node calculates a unique checksum based on the encrypted data packet content using the SHA-256 hash algorithm. This checksum is stored in association with the data packet and used by the cloud to verify the packet's integrity. During data transmission, the edge node uses the HTTP resumable transmission protocol to divide the data packet into several fixed-size segments (e.g., 10MB / segment). Each segment is assigned a unique identifier (e.g., segment sequence number + segment checksum). These segments are then sent to the cloud via an encrypted communication link, while simultaneously recording the transmission status of each segment in real time (not transmitted, transmitting, transmission completed).

[0072] If a network interruption occurs during the backhaul process (heartbeat packet detection timeout or data segment transmission confirmation timeout), the edge node immediately stops the backhaul, records the sequence number list of transmitted data segments and the sequence number list of untransmitted data segments, to avoid confusion of data segment identifiers.

[0073] After network recovery, the edge node restarts the cloud-based collaborative weight calculation process. Based on the weight result, it determines the optimal time to resume transmission (when the weight result is greater than a preset threshold), ensuring that the network status and edge load are at an optimal level during transmission resumption. During transmission resumption, the edge node sends a resumption request to the cloud, containing a list of sequence numbers of the transmitted data segments. Upon receiving this request, the cloud sends a confirmation message. The edge node then transmits only the untransmitted data segments to the cloud. The cloud receives these segments and concatenates them with the received data segments according to their sequence numbers, forming a complete data packet, ensuring the integrity and efficiency of data return.

[0074] S4.4. After receiving the data packet, the cloud verifies its integrity by comparing the checksum. It then combines this with a global evaluation model to perform a comprehensive analysis of the feature vectors and response data, generating standardized results that include the candidate's score, feature attainment status, and overall evaluation trend. After receiving the data packets transmitted back from the edge nodes, the cloud first extracts the SHA-256 checksum associated with the data packets. By recalculating the SHA-256 checksum of the received data packets and comparing it with the extracted checksum, the integrity of the data packets is verified.

[0075] If the checksums match, it indicates that the data packet was not lost or tampered with during transmission, and the verification is successful; if the checksums do not match, the cloud sends a retransmission request to the edge node, requesting the edge node to retransmit the data packet.

[0076] After integrity verification is successful, the cloud uses the negotiated AES-256 key to decrypt the data packet, extracts the feature vectors and response data, and calls the global evaluation model for comprehensive analysis. The global evaluation model adopts a deep learning architecture, including an input layer, hidden layers (multiple fully connected layers), and an output layer. The input layer receives standardized feature vectors, the hidden layers perform non-linear transformations and fusions on the features through activation functions, and the output layer outputs multi-dimensional evaluation results.

[0077] The model assigns preset weights to each feature point corresponding to each feature vector according to the preset scoring criteria (stored in the cloud scoring configuration library), calculates the individual feature score (feature point quantization value × feature weight) based on the quantized value (0-1) of the feature point, and sums up all individual feature scores to obtain the candidate's final score; Simultaneously, the quantified values ​​of each feature point are compared with preset thresholds (e.g., 0.6). If the value is higher than the threshold, the feature is considered to meet the standard; if it is lower than the threshold, it is considered to fail to meet the standard, thus clarifying the feature compliance status of the candidates' answers. In addition, the scores of all candidates are statistically grouped (e.g., 0-60 points, 61-80 points, 81-100 points), and the proportion of candidates in each score range is calculated. The feature compliance status of all candidates is statistically analyzed, the overall compliance rate of each feature is calculated, and the overall evaluation trend is summarized, such as which feature points high-scoring candidates excel in and which feature points low-scoring candidates generally fail to meet.

[0078] After the analysis is completed, a standardized evaluation result report is generated in the cloud. The report adopts a unified XML format and includes the candidate's personal information (candidate identifier, subject), final score, details of compliance with various characteristics (characteristic name, quantitative value, whether compliance is met) and overall evaluation trend (score range distribution ratio, characteristic compliance rate statistics). This allows administrators and invigilators to view, download and export the report through the cloud management platform.

[0079] This cloud-edge-device collaborative art examination and evaluation method achieves standardized operation of the entire art examination and evaluation process through a hierarchical collaborative architecture of cloud, edge nodes and terminals. The collaborative approach between cloud, edge, and endpoint reduces reliance on a single architecture, enabling the core examination process to be conducted efficiently within a local area network. This minimizes the impact of public network fluctuations on the examination and ensures its stability. Multiple security mechanisms, including encrypted encapsulation, asymmetric encrypted communication links, and local encrypted storage, ensure the security of examination resources and answer data during storage, transmission, and processing, effectively mitigating risks such as data leakage and tampering. The lightweight feature extraction model's adaptation and optimization to the terminal, along with standardized data format processing, achieves unified processing of different types of artistic answer data, improving data processing efficiency and accuracy. Fault-tolerant mechanisms such as network interruption resumption, local temporary storage, and breakpoint resumption ensure the continuity of the examination process and prevent data loss due to network outages or other unforeseen circumstances. The global evaluation model, through quantitative analysis and weighted scoring of feature vectors, makes the evaluation results more standardized and objective, comprehensively reflecting the candidates' artistic level and overall examination trends, providing reliable technical support for art examination evaluation. Meanwhile, the entire method reduces manual intervention by automating resource allocation, data collection, transmission and processing, thereby improving the overall efficiency of art examination and assessment and meeting the application needs of large-scale art examinations.

[0080] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An art examination assessment method based on cloud-edge-device collaboration, characterized in that, Includes the following steps: S1. Perform pre-exam collaborative configuration, distribute evaluation activity parameters, encrypted exam resources and lightweight feature extraction models to the cloud, and verify the integrity of resources and model adaptability after receiving them at the edge nodes, while simultaneously detecting the hardware performance and network connectivity of the candidate's terminal. S2. Based on the terminal status and resource adaptation results, the edge node evaluates the task allocation scheme through a collaborative decision-making formula, pushes exam notifications to ready terminals and allocates exclusive test papers, and establishes an encrypted communication link between the edge and the terminal. S3. During the exam, the terminal collects artistic answer data and completes local feature extraction through a lightweight model. Edge nodes receive feature data and answer status in real time. When the network is normal, the data is synchronized to the cloud. When the network is disconnected, the data is temporarily stored locally and the network recovery status is continuously monitored. S4. Edge nodes aggregate the answer data, feature results, and exam logs from all terminals, encrypt and package them, and then send them back to the cloud in batches according to the collaboration strategy. The cloud combines the data with the global model to perform comprehensive processing and generate art exam evaluation results.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Configure the exam duration, scoring dimensions, and answer format parameters of the evaluation activity in the cloud, upload the exam resources containing test questions and materials and encrypt and encapsulate them, and distribute the lightweight feature extraction model adapted to edge nodes at the same time. S1.

2. Edge nodes download encrypted exam resources and lightweight models from the cloud, and complete integrity verification by comparing resource file identifiers and check codes, and test the compatibility between the model and local hardware. S1.

3. The edge node sends a detection command to the candidate's terminal, collects the terminal's storage space, computing performance, network bandwidth and client version information, filters ready terminals that meet the examination requirements, generates a detection report containing terminal identifier and status parameters and synchronizes it to the cloud.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. The edge node calls the collaborative decision-making module, inputting terminal hardware performance, network bandwidth, resource size, and model complexity parameters, and calculates using the formula: ; Calculate the task unloading priority, where These are the weighting coefficients. For network bandwidth, For edge computing capabilities, For cloud computing capabilities, For data volume, For terminal load; S2.

2. Determine the task allocation between local processing and cloud collaboration based on priority results, push exam start notifications to ready terminals, and assign unique test papers with no duplicates according to preset rules; S2.

3. Establish a dedicated communication link between the edge node and the candidate's terminal through an asymmetric encryption algorithm, verify the terminal's access permissions, and ensure data transmission security.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Candidates browse the decrypted exam resources through the terminal and complete artistic responses such as drawing, singing, and video recording. The terminal collects data on the response process and the final result in real time. S3.

2. The terminal calls the built-in lightweight feature extraction model to extract feature points and standardize the format of the answer data, generate feature vectors and upload them to the edge nodes in real time. S3.

3. When the network is normal, the edge nodes will synchronize the feature vector and the response status to the cloud in real time. When the network is interrupted, the terminal will temporarily store the response data and feature results in the local encrypted storage area. The edge nodes will continuously monitor the network status and trigger the terminal to automatically resume transmission after the network is restored.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Edge nodes collect complete answer data, feature vectors, answer status records, and examination process logs from all candidate terminals, classify and organize them according to candidate identifiers and examination subjects, and package and aggregate them into data packets of a unified format; S4.

2. Through calculation formula: ； Calculate the cloud collaboration weight, where, For activation function, For network latency, For delay weighting, For edge load, For load weight, As a bias term, the timing of batch data return is selected based on the weighting results; S4.

3. After receiving the data packet in the cloud, the integrity is verified by comparing the check code. The feature vector and the answer data are comprehensively analyzed by combining the global evaluation model to generate standardized results that include the candidate's score, feature attainment status and overall evaluation trend.

6. The method according to claim 1, characterized in that, In step S1, the edge node performs adaptation optimization on the downloaded lightweight model, adjusts the model parameters to match the local hardware computing power, and compares the terminal detection results with the cloud-based preset standards, marks terminals that do not meet the requirements and provides optimization suggestions to ensure that the examination resources and models can be stably accessed.

7. The method according to claim 1, characterized in that, In step S3, the edge node collects the login status, answering progress, model running status and network transmission status of the candidate's terminal in real time. The collected multi-dimensional status information is fused and processed, and the overall examination process is presented through a visual interface. Network outages, timeouts, and abnormal model running situations are identified and logged in real time.

8. The method according to claim 1, characterized in that, In step S4, the edge node encrypts the aggregated data packets using a symmetric encryption algorithm to generate a unique checksum corresponding to each data packet. If a network interruption occurs during the backhaul process, the identifiers of the transmitted and untransmitted data segments are recorded. After the network is restored, the optimal timing for resuming transmission is determined based on the collaborative weight formula, and the data backhaul is resumed from the point of interruption. After the backhaul is completed, the cloud verifies the data integrity by comparing the checksums.

9. An art examination and evaluation system based on cloud-edge-device collaboration, characterized in that, It includes a cloud collaboration module, an edge processing module, and a terminal execution module; The cloud-based collaboration module is used to configure evaluation activity parameters, store encrypted exam resources, and deploy a global evaluation model. It receives data from the edge processing module and generates evaluation results. The edge processing module downloads exam resources and a lightweight model from the cloud-based collaboration module and performs verification and adaptation. It evaluates task allocation schemes through collaborative decision-making formulas, establishes an encrypted communication link between the edge and the terminal, receives answer data and feature results from the terminal execution module, aggregates them, and sends them back to the cloud according to the collaboration strategy. The terminal execution module receives exam resources, supports candidates in completing artistic answers, extracts answer features through a built-in lightweight model, temporarily stores relevant data during network interruptions, and retransmits it after network recovery.

10. The system according to claim 9, characterized in that, The cloud-based collaborative module includes a parameter configuration unit, a resource storage unit, a global model unit, and a data processing unit. The parameter configuration unit configures evaluation-related parameters, the resource storage unit encrypts and stores examination resources and model files, the global model unit provides comprehensive evaluation algorithm support, and the data processing unit verifies the returned data and generates evaluation results. The edge processing module includes a resource verification unit, a collaborative decision-making unit, a communication encryption unit, a data aggregation unit, and a backhaul control unit. The resource verification unit verifies the integrity and compatibility of resources and models. The collaborative decision-making unit calculates task allocation and backhaul strategies using formulas. The communication encryption unit establishes a secure edge link. The data aggregation unit organizes and categorizes relevant answer data. The backhaul control unit executes data backhaul according to the collaborative strategy. The terminal execution module includes a resource receiving unit, a response acquisition unit, a feature extraction unit, and a local temporary storage unit. The resource receiving unit receives and decrypts exam resources, the response acquisition unit acquires artistic response data, the feature extraction unit extracts response features through a lightweight model, and the local temporary storage unit stores relevant data when the network is disconnected and supports resume transmission.

Images