Development and training methods and systems based on artificial intelligence in the logistics field

Abstract

This invention discloses a development and training method and system based on artificial intelligence in the logistics field, including the following steps: receiving input of logistics professional training requirements, preprocessing and standardizing the input content, completing the training environment configuration and resource initialization, and obtaining clear training objectives and resource configuration parameters; based on the determined training objectives, collecting relevant data of logistics scenarios. This invention deploys full-process logistics professional training by constructing an edge-cloud collaborative training architecture, breaking the limitations of traditional training equipment and scenarios, improving the flexibility and practicality of training, processing training data through a combination of automated annotation and manual verification to improve data annotation efficiency and accuracy, quickly constructing high-quality standardized logistics training datasets, and building AI models for logistics scenarios through hyperparameter optimization and incremental training mechanisms to improve model adaptability and recognition capabilities, and lower the model development threshold.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a development and training method and system based on artificial intelligence in the logistics field. Background Technology

[0002] Logistics is an interdisciplinary field that integrates management, economics, information technology, and engineering technology. It focuses on the entire process of the physical flow of goods from the supply location to the receiving location, covering key links such as transportation, warehousing, loading and unloading, packaging, distribution processing, delivery, and information processing. It aims to cultivate compound talents with the ability to plan and design logistics systems, operate and manage them, and apply intelligent technologies, so as to meet the needs of the modern logistics industry to transform towards digitalization and intelligence.

[0003] Generally, traditional logistics training methods mainly consist of theoretical lectures combined with simple simulation operations. These methods are primarily conducted through offline sand table simulations, hands-on operation of single logistics equipment, and static case analysis. Although some institutions have introduced basic computer simulation software, they lack a practical environment that is deeply integrated with actual logistics scenarios.

[0004] However, traditional training methods have significant drawbacks: First, the training content is disconnected from industry realities, mostly focusing on basic process simulations and failing to touch upon the core technological applications of intelligent logistics. Second, training equipment is fixed and outdated, unable to adapt to the training needs of new technologies such as artificial intelligence and edge computing, making it difficult for students to access the entire industrial-grade development and deployment process. Third, training data collection is difficult, annotation efficiency is low, and there is a risk of data privacy leaks, making it difficult to form high-quality, secure, and controllable training datasets. Fourth, the training model lacks collaboration, failing to support concurrent training in multiple groups and efficient interaction between teachers and students, and also making it difficult to achieve seamless integration of model training and edge deployment. Fifth, the evaluation of training effectiveness is singular, focusing only on the completion of operational steps, lacking a comprehensive consideration of model performance, resource utilization, and training efficiency, and failing to achieve dynamic optimization of the training system based on evaluation results.

[0005] In summary, it is necessary to propose development and training methods and systems based on artificial intelligence in the field of logistics to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a development and training method and system based on artificial intelligence in the field of logistics, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes a development and training method based on artificial intelligence in the field of logistics. The method is implemented based on an artificial intelligence logistics development and training system and includes the following steps: S1. Receive the input of logistics professional training requirements, preprocess and standardize the input content, complete the training environment configuration and resource initialization, and obtain clear training objectives and resource configuration parameters; S2. Based on the training objectives determined in step S1, collect relevant data on logistics scenarios, complete data annotation through a combination of automated annotation and manual verification, and perform data augmentation processing to generate a standardized training dataset; S3. Based on the standardized training dataset generated in step S2, construct a model training process specifically for the logistics field, perform model training, hyperparameter optimization and incremental training, and output an optimized AI model for the logistics scenario. S4. Based on the AI ​​model output in step S3, perform model format conversion and compression quantization, and distribute the model to edge devices through a security authentication mechanism to achieve end-to-cloud collaborative training deployment. The end-to-cloud collaboration adopts a closed-loop architecture of server model training and edge device inference feedback. The server receives inference latency, computing power consumption, accuracy fluctuation and inference performance data from edge devices in real time, and dynamically adjusts the model optimization strategy. Edge devices only transmit key inference results and status data to reduce network bandwidth consumption. S5. Based on the edge-cloud deployment results of step S4, conduct logistics professional training operations, and monitor and evaluate the data flow, model performance and resource utilization in real time and in multiple dimensions during the training process. S6. Record the key data and evaluation results of the entire process from S1 to S5, and optimize the training dataset, model parameters and training process based on user feedback to achieve dynamic iteration of the training system.

[0008] Preferably, the implementation process of step S1 is as follows: S1.1. Provides a multimodal input interface for text and document uploads and visual interaction to receive logistics training requirements, including logistics scenario type, training task type, number of participants and equipment requirements; S1.2. The input logistics training requirements are cleaned and preprocessed by word segmentation. Key training requirements are selected by the core element weight calculation algorithm, as shown in equation (1): (1); In the formula, The weight of the j-th core element in the i-th training requirement text is set to a threshold of 0.65. The threshold adopts a dynamic adaptation mechanism and is automatically adjusted according to the logistics scenario type: the threshold is set to 0.62 for warehousing scenarios, 0.65 for transportation scenarios, and 0.68 for delivery scenarios. Elements exceeding the threshold are judged as key elements. The word frequency of the j-th core element in the i-th training requirement text is the number of times the element appears in the text divided by the total number of elements in the text. The inverse document frequency of the j-th core element is given by equation (2): (2); In the formula, This represents the total number of documents in the corpus for logistics training. Let j be the number of documents containing the j-th core element; This is the scenario matching coefficient, ranging from 0.85 to 1.25. If the element belongs to the core logistics scenario, then... ,otherwise ; This is the training correlation coefficient, ranging from 0.9 to 1.1. If an element has a direct correlation with the identified training objective, then... ,otherwise For the missing key elements, supplementary information can be obtained through an interactive completion mechanism; S1.3. Complete the environment configuration based on the logistics training resource library, including server hardware initialization, edge device network configuration, and database and training framework deployment.

[0009] Preferably, the implementation process of step S2 is as follows: S2.1. Based on the logistics scenario type determined in step S1, collect the corresponding logistics data, including cargo images, transportation videos, warehousing time-series data and sensor signals. The data format supports JPEG image format, 720P to 8K resolution video format, and time-series data text format. S2.2. Start the built-in logistics scenario pre-labeling model and perform automatic labeling on the collected data. The labeling types include target detection box labeling, time series feature labeling and category labeling. Among them, target detection labeling supports rectangle drawing, tracking labeling and interpolation labeling functions, and category labels support custom addition of Chinese and English. S2.3. Ensure labeling accuracy through a label consistency mapping algorithm, as shown in equation (3): (3); In the formula, Let be the matching degree between the k-th automatically labeled result and the l-th manually verified result, with a value ranging from 0 to 1. Once the annotations are deemed consistent, the matching threshold is dynamically adjusted based on the annotation type: the threshold for object detection annotations is set to 0.85, the threshold for temporal feature annotations is set to 0.82, and the threshold for category label annotations is set to 0.90. A consistent annotation is determined when the corresponding threshold is reached. It has been determined that manual correction is required. The annotation is deemed invalid. Let be the overlap between the k-th automatically labeled result and the l-th manually verified result. Let k be the total number of annotations for the k-th automatic annotation result. This represents the total number of annotations for the l-th manual verification result; The semantic similarity between the k-th automatically labeled result and the l-th manually verified result is calculated based on a dictionary in the logistics domain. S2.4. Perform data augmentation on the labeled data, using a combination of flipping, shifting, scaling, cropping, adding noise, and adjusting brightness. Each original data point automatically generates 5 augmented samples. The augmented data is divided into training, validation, and test sets in a 7:2:1 ratio to form a standardized training dataset, which supports searching, editing, deleting, and downloading operations.

[0010] Preferably, the implementation process of step S3 is as follows: S3.1. Based on the standardized training dataset in step S2, load pre-built models in the logistics field, including object detection models, image classification models, and time series analysis models, supporting the import of models from TensorFlow and PyTorch frameworks; configure model training parameters, including training epochs, batch size, and initial learning rate. S3.2. Start the hyperparameter optimization process and iteratively adjust the hyperparameters using the Bayesian optimization algorithm. The default number of optimization attempts is 30, as shown in equation (4): (4); In the formula, For the optimal combination of hyperparameters, The set of hyperparameters includes learning rate, batch size, and weight decay coefficient. For hyperparameters In the training set The posterior probability on, For hyperparameters The corresponding model performance score; S3.3. Execute the model training process, use the gradient descent optimization algorithm to adjust the model parameters, record the model loss value and accuracy every 20 rounds during training, and automatically stop training when the loss function does not improve for 50 consecutive rounds; S3.4. Start the incremental training mechanism. On the basis of retaining the original training dataset, add new logistics scenario samples or negative sample samples, and readjust the model parameters, as shown in equation (5): (5); In the formula, These are the model parameters after incremental training. For the original model parameters, The learning rate decay coefficient ranges from 0.001 to 0.01. For the loss function in the merged dataset gradient on, For the original training dataset, To add a new training dataset; S3.5. After training is completed, the accuracy, precision, recall, F1-score and mAP of the model are calculated using the validation set, and the optimized logistics scenario AI model that meets the training requirements is output.

[0011] Preferably, the implementation process of step S4 is as follows: S4.1. Convert the AI ​​model output from step S3. First, convert the TensorFlow and PyTorch framework models to ONNX format. During the conversion, perform operator compatibility optimization, remove redundant high-order operators in logistics scenarios, and retain core operators supported by RK series chips (such as RK3588 and RV1106) such as Conv2d, BatchNorm, and Linear. Then, call the RKNN Toolkit to convert the ONNX format to RKNN format. During the conversion, take into account the computing power characteristics of ARM architecture TPUs, use layer fusion technology to merge continuous convolutional layers and activation layers, and replace the SiLU activation function with the ReLU activation function. Because the SiLU function has instruction execution latency on ARM architecture TPUs, the replacement can reduce the hardware computing power consumption during inference. In addition, through data augmentation and hyperparameter optimization in the early stage, the accuracy loss is compensated to ensure that the model accuracy decreases by no more than 2%. S4.2. A three-level optimization process of "quantization-pruning-distillation" is adopted: First, 8-bit integer quantization is performed based on the logistics scenario dataset. The quantization calibration adopts the KL divergence calibration algorithm to preserve the quantization accuracy of key information such as cargo features and warehouse shelf edges. Second, structured pruning is performed on the model's convolutional layers. The pruning ratio is dynamically adjusted according to the computing power of the target edge device (RK3588 / RV1106). The pruning ratio of RK3588 is controlled at 40%-50%, and the pruning ratio of RV1106 is controlled at 30%-40%. During pruning, the core channels related to logistics target detection and classification are preserved. Finally, through knowledge distillation technology, the original high-precision model is used as the teacher model, and the quantized and pruned model is used as the student model. The distillation loss function adopts cross-entropy loss weighted by logistics scenario features to ensure that the inference speed of the model on the edge device is improved by more than 30%, while the model size is compressed to less than 1 / 5 of the original size. S4.3. Establish an edge device security authentication mechanism. The edge device actively creates an authentication file to initiate a network access application. The server decrypts and verifies the device using a symmetric encryption method with a public key medium. The authentication process is shown in equation (6): (6); In the formula, For the authentication results, For decryption function, For encryption functions, For the server's private key, For edge device public key, The edge device information includes the device model, hardware parameters, and network address. During the authentication process, an additional hardware fingerprint verification of the device (extracting the unique identifier ID of the RK series chip) is added to prevent unauthorized devices from accessing the network. S4.4. After successful authentication, based on the resource configuration parameters in step S1, the converted model is distributed to edge devices within the specified IP range via the MQTT protocol. The MQTT protocol adopts a dedicated configuration for logistics training: the QoS level is set to 2 (to ensure reliable transmission of model files), message topics are divided into levels according to "logistics scenario - device type - training group", and the payload uses the LZ77 compression algorithm to reduce transmission volume. After deployment, the model inference test is automatically executed, a performance optimization report adapted to the current hardware is generated, and the device network configuration and model deployment are completed, supporting the training requirements of 3-channel 1920×1080@30FPS input and 1280×720@30FPS output.

[0012] Preferably, the implementation process of step S5 is as follows: S5.1. Based on the edge-cloud deployment results of step S4, it supports the development and learning of 8 groups of 4 people in concurrent models, and provides training tasks for logistics scenario training operations including target detection, image classification, target tracking, monocular ranging, time series analysis, and unmanned system control. S5.2. Real-time monitoring of key data during the training process includes data transmission rate, model inference latency, device CPU / GPU utilization, and memory usage. The monitoring frequency is set to once per second, and real-time monitoring logs are generated. S5.3. Construct a multi-dimensional practical training evaluation system: Quality assessment: Calculate the model's accuracy, recall, F1-score, mAP50, and mAP50-95 metrics on the logistics test set; Computing power evaluation: Compare the model training time and inference speed under different hardware configurations. The evaluation metrics are the time spent per training round and the number of frames per second for inference. Efficiency assessment: Computational resource utilization includes CPU utilization, memory utilization, and network bandwidth utilization. The assessment is shown in equation (7): (7); In the formula, To score overall efficiency, For task completion rate, For average resource utilization rate, To delay satisfaction, , , These are weighting coefficients, with values ​​of 0.4, 0.3, and 0.3 respectively. S5.4 generates training evaluation reports, clearly defining training results and improvement directions, and supports the visualization and export of evaluation data.

[0013] Preferably, the implementation process of step S6 is as follows: S6.1. Record all process data from steps S1 to S5, including requirement parameters, dataset information, model parameters, training logs, deployment configuration, and evaluation results, and store them in PostgreSQL and Mongo databases to ensure data traceability; S6.2. Based on the evaluation results of step S5, establish a problem diagnosis mechanism: if the overall mAP decreases, it is determined that the quality of the newly added dataset is not up to standard; if the mAP amplitude is not large but the mAP50-95 amplitude is large, it is determined that there is a labeling quality problem; if the resource utilization rate is less than 30%, it is determined that the resource allocation is unreasonable. S6.3. Based on feedback from logistics major teaching, optimize the training dataset by supplementing high-quality samples and correcting low-quality annotations, adjust model parameters including hyperparameter range and incremental training sample ratio, and optimize the training process by adjusting task order and resource allocation strategy. S6.4. After each iteration, re-execute steps S2 to S5 to verify the optimization effect. The overall efficiency score is then calculated. When the improvement exceeds 15% or the model accuracy improves by more than 8%, the current optimization results are solidified to form a new training benchmark configuration.

[0014] Based on the above methods, this invention also proposes a development and training system for the field of artificial intelligence logistics, the system comprising a front-end architecture and a back-end architecture; The front-end architecture includes: The requirement interaction module provides a multimodal requirement input interface, supporting text and document uploads as well as visual interaction. The annotation operation module provides an interface for label configuration, manual annotation, and automatic annotation result review, and supports label addition, modification, and consistency verification. The training operation module provides a visual operation interface for logistics training tasks, supporting real-time viewing of training progress and equipment status. The evaluation and display module provides visual charts of evaluation indicators and an interface for exporting evaluation reports. The backend architecture includes: The requirement processing module is used to preprocess the input training requirements, extract core elements, and calculate weights. The environment configuration module is used to complete the hardware initialization of the server and edge devices, network configuration, and deployment of the database and training framework. The data management module is used for logistics data collection, storage, data augmentation, and dataset partitioning. The model training module is used to load a pre-set model, perform hyperparameter optimization, model training, and incremental training. The model evaluation module is used to calculate model performance indicators and output the optimized AI model. A model conversion module, which performs model format conversion and compression quantization processing; A device authentication module, which is used to implement secure authentication and authorization management for edge devices; The model delivery module delivers the model to the edge device via the MQTT protocol to complete the deployment. The monitoring module is used to collect key data during the training process in real time and generate monitoring logs. The evaluation module is used to conduct quality evaluation, computing power evaluation, and efficiency evaluation, and generate an evaluation report. The iterative optimization module is used to record full-process data, diagnose problems, optimize training configurations, and verify effects.

[0015] Preferably, the front-end architecture further includes: The resource management interface module provides interfaces for querying, editing, deleting, and downloading datasets and model files, and supports viewing resource metadata including name, tags, and addition time. The parameter configuration interface module provides a custom configuration interface for training parameters and hyperparameter optimization parameters. The device management interface module provides an interface for viewing edge device status, configuring the network, and managing authorization. The backend architecture also includes: The federated learning module is used to implement base model distribution, tenant local incremental training, and model parameter isolation. The computing network scheduling module is used to intelligently schedule computing resources and network bandwidth, and optimize the allocation of computing power and data flow during task execution. The log storage module is used to centrally store demand processing logs, training logs, monitoring logs, and evaluation logs, and supports log retrieval and export.

[0016] Preferably, the front-end architecture further includes: The collaborative interaction module is used to support teachers to batch distribute labeled items and models to students and provides permission management functions. The help and guidance module provides practical training operation guides, answers to frequently asked questions, and access to technical support. The backend architecture also includes: A multi-tenant management module is used to allocate independent computing, storage, and network resources to different training groups. A compatibility adaptation module is used to support adaptation to different hardware architectures (CPU / GPU / TPU), operating systems, and training frameworks. The feedback processing module is used to collect user feedback on the training process, model performance, and interface operation, classify and store the feedback, and provide feedback analysis results.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention deploys logistics professional training across the entire process by constructing an edge-cloud collaborative training architecture, breaking the limitations of traditional training equipment and scenarios, and improving the flexibility and practicality of training. It processes training data through a combination of automated annotation and manual verification, improving data annotation efficiency and accuracy, and rapidly constructing high-quality, standardized logistics training datasets. It builds AI models for logistics scenarios through hyperparameter optimization and incremental training mechanisms, improving model adaptability and recognition capabilities, and lowering the model development threshold. Finally, it adapts the model to edge devices through model format conversion and compression quantization techniques, enabling the model to be deployed on different hardware architectures. The system ensures efficient operation and fills the gap between model training and actual deployment. It evaluates the training process and results through a multi-dimensional assessment system, comprehensively grasps the training effect, and accurately identifies training problems. The system upgrades through full-process data recording and iterative optimization mechanisms, continuously improving training quality and adapting to the technological development and changing teaching needs of the logistics industry. At the same time, it manages concurrent training in multiple groups through federated learning and multi-tenant management modules, ensuring data privacy and security, and achieving efficient collaboration and resource isolation between teachers and students. The modular design of front-end interaction and back-end services separates training operations from core processing, reducing the difficulty of training operations and improving system stability and scalability. Attached Figure Description

[0018] Figure 1 The flowchart of the development and training method of the present invention based on the field of artificial intelligence logistics is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0020] Please see Figure 1 This invention proposes a development and training method based on the field of artificial intelligence logistics, which is implemented based on an artificial intelligence logistics development and training system. Based on this, the development and training system of the present invention based on the field of artificial intelligence logistics includes a front-end architecture and a back-end architecture. The server configuration meets the following requirements: main processor with no less than 6 cores and 12 threads, main frequency no less than 2.9GHz, maximum turbo frequency no less than 4.3GHz, accelerated processor video memory no less than 12GB, memory no less than 64GB DDR4, storage no less than 2TB solid-state drive plus 8TB mechanical drive, and network support for 10 Gigabit Ethernet. Edge devices must be configured to have a main processor of at least four 64-bit Cortex-A55 cores with a maximum clock speed of 2.0GHz, at least 4GB of DDR4 memory, at least 16GB of GFlash storage, and support for Gigabit Ethernet and Wi-Fi / Bluetooth modules. The system operating environment must meet the following requirements: operating system Ubuntu 20.04, Python version 3.8 or above, and database PostgreSQL 13 or above. In practical applications, specifically, the development and training method of this invention based on the field of artificial intelligence logistics includes the following steps: S1. Receive the input of logistics professional training requirements, preprocess and standardize the input content, complete the training environment configuration and resource initialization, and obtain clear training objectives and resource configuration parameters; It should also be noted that the implementation process of step S1 is as follows: S1.1. Provides a multimodal input interface for text and document uploads and visual interaction, and receives logistics training requirements, including logistics scenario type, training task type, number of participants and equipment requirements; S1.2. The input logistics training requirements are cleaned and preprocessed by word segmentation. Key training requirements are selected by the core element weight calculation algorithm, as shown in equation (1): (1); In the formula, The weight of the j-th core element in the i-th training requirement text is set to a threshold of 0.65. The threshold adopts a dynamic adaptation mechanism and is automatically adjusted according to the logistics scenario type: the threshold is set to 0.62 for warehousing scenarios, 0.65 for transportation scenarios, and 0.68 for delivery scenarios. Elements exceeding the threshold are judged as key elements. The word frequency of the j-th core element in the i-th training requirement text is the number of times the element appears in the text divided by the total number of elements in the text. The inverse document frequency of the j-th core element is given by equation (2): (2); In the formula, This represents the total number of documents in the corpus for logistics training. Let j be the number of documents containing the j-th core element; This is the scenario matching coefficient, ranging from 0.85 to 1.25. If the element belongs to the core logistics scenario, then... ,otherwise ; This is the training correlation coefficient, ranging from 0.9 to 1.1. If an element has a direct correlation with the identified training objective, then... ,otherwise For the missing key elements, supplementary information can be obtained through an interactive completion mechanism; S1.3. Complete the environment configuration based on the logistics training resource library, including server hardware initialization, edge device network configuration, and database and training framework deployment.

[0021] S2. Based on the training objectives determined in step S1, collect relevant data on logistics scenarios, complete data annotation through a combination of automated annotation and manual verification, and perform data augmentation processing to generate a standardized training dataset; It should also be noted that the implementation process of step S2 is as follows: S2.1. Based on the logistics scenario type determined in step S1, collect the corresponding logistics data, including cargo images, transportation videos, warehousing time-series data and sensor signals. The data format supports JPEG image format, 720P to 8K resolution video format, and time-series data text format. S2.2. Start the built-in logistics scenario pre-labeling model and perform automatic labeling on the collected data. The labeling types include target detection box labeling, time series feature labeling and category labeling. Among them, target detection labeling supports rectangle drawing, tracking labeling and interpolation labeling functions, and category labels support custom addition of Chinese and English. S2.3. Ensure labeling accuracy through a label consistency mapping algorithm, as shown in equation (3): (3); In the formula, Let be the matching degree between the k-th automatically labeled result and the l-th manually verified result, with a value ranging from 0 to 1. Once the annotations are deemed consistent, the matching threshold is dynamically adjusted based on the annotation type: the threshold for object detection annotations is set to 0.85, the threshold for temporal feature annotations is set to 0.82, and the threshold for category label annotations is set to 0.90. A consistent annotation is determined when the corresponding threshold is reached. It has been determined that manual correction is required. The annotation is deemed invalid. Let be the overlap between the k-th automatically labeled result and the l-th manually verified result. Let k be the total number of annotations for the k-th automatic annotation result. This represents the total number of annotations for the l-th manual verification result; The semantic similarity between the k-th automatically labeled result and the l-th manually verified result is calculated based on a dictionary in the logistics domain. S2.4. Perform data augmentation on the labeled data, using a combination of flipping, shifting, scaling, cropping, adding noise, and adjusting brightness. Each original data point automatically generates 5 augmented samples. The augmented data is divided into training, validation, and test sets in a 7:2:1 ratio to form a standardized training dataset, which supports searching, editing, deleting, and downloading operations.

[0022] S3. Based on the standardized training dataset generated in step S2, construct a model training process specifically for the logistics field, perform model training, hyperparameter optimization and incremental training, and output an optimized AI model for the logistics scenario. It should also be noted that the implementation process of step S3 is as follows: S3.1. Based on the standardized training dataset in step S2, load pre-built models in the logistics field, including object detection models, image classification models, and time series analysis models, supporting the import of models from TensorFlow and PyTorch frameworks; configure model training parameters, including training epochs, batch size, and initial learning rate. S3.2. Start the hyperparameter optimization process and iteratively adjust the hyperparameters using the Bayesian optimization algorithm. The default number of optimization attempts is 30, as shown in equation (4): (4); In the formula, For the optimal combination of hyperparameters, The set of hyperparameters includes learning rate, batch size, and weight decay coefficient. For hyperparameters In the training set The posterior probability on, For hyperparameters The corresponding model performance score; S3.3. Execute the model training process, use the gradient descent optimization algorithm to adjust the model parameters, record the model loss value and accuracy every 20 rounds during training, and automatically stop training when the loss function does not improve for 50 consecutive rounds; S3.4. Start the incremental training mechanism. On the basis of retaining the original training dataset, add new logistics scenario samples or negative sample samples, and readjust the model parameters, as shown in equation (5): (5); In the formula, These are the model parameters after incremental training. For the original model parameters, The learning rate decay coefficient ranges from 0.001 to 0.01. For the loss function in the merged dataset gradient on, For the original training dataset, To add a new training dataset; S3.5. After training is completed, the accuracy, precision, recall, F1-score and mAP of the model are calculated using the validation set, and the optimized logistics scenario AI model that meets the training requirements is output.

[0023] S4. Based on the AI ​​model output in step S3, perform model format conversion and compression quantization, and distribute the model to edge devices through a security authentication mechanism to realize the practical deployment of edge-cloud collaboration. Edge-cloud collaboration adopts a closed-loop architecture of server model training and edge device inference feedback. The server receives inference latency, computing power consumption, accuracy fluctuation and inference performance data of edge devices in real time, and dynamically adjusts the model optimization strategy. Edge devices only transmit key inference results and status data to reduce network bandwidth consumption. It should also be noted that the implementation process of step S4 is as follows: S4.1. Convert the AI ​​model output from step S3. First, convert the TensorFlow and PyTorch framework models to ONNX format. During the conversion, perform operator compatibility optimization, remove redundant high-order operators in logistics scenarios, and retain core operators supported by RK series chips (such as RK3588 and RV1106) such as Conv2d, BatchNorm, and Linear. Then, call the RKNN Toolkit to convert the ONNX format to RKNN format. During the conversion, take into account the computing power characteristics of ARM architecture TPUs, use layer fusion technology to merge continuous convolutional layers and activation layers, and replace the SiLU activation function with the ReLU activation function. Because the SiLU function has instruction execution latency on ARM architecture TPUs, the replacement can reduce the hardware computing power consumption during inference. In addition, through data augmentation and hyperparameter optimization in the early stage, the accuracy loss is compensated to ensure that the model accuracy decreases by no more than 2%. S4.2. A three-level optimization process of "quantization-pruning-distillation" is adopted: First, 8-bit integer quantization is performed based on the logistics scenario dataset. The quantization calibration adopts the KL divergence calibration algorithm to preserve the quantization accuracy of key information such as cargo features and warehouse shelf edges. Second, structured pruning is performed on the model's convolutional layers. The pruning ratio is dynamically adjusted according to the computing power of the target edge device (RK3588 / RV1106). The pruning ratio of RK3588 is controlled at 40%-50%, and the pruning ratio of RV1106 is controlled at 30%-40%. During pruning, the core channels related to logistics target detection and classification are preserved. Finally, through knowledge distillation technology, the original high-precision model is used as the teacher model, and the quantized and pruned model is used as the student model. The distillation loss function adopts cross-entropy loss weighted by logistics scenario features to ensure that the inference speed of the model on the edge device is improved by more than 30%, while the model size is compressed to less than 1 / 5 of the original size. S4.3. Establish an edge device security authentication mechanism. The edge device actively creates an authentication file to initiate a network access application. The server decrypts and verifies the device using symmetric encryption with a public key medium. The authentication process is shown in equation (6): (6); In the formula, For the authentication results, For decryption function, For encryption functions, For the server's private key, For edge device public key, The edge device information includes the device model, hardware parameters, and network address. During the authentication process, an additional hardware fingerprint verification of the device (extracting the unique identifier ID of the RK series chip) is added to prevent unauthorized devices from accessing the network. S4.4. After successful authentication, based on the resource configuration parameters in step S1, the converted model is distributed to edge devices within the specified IP range via the MQTT protocol. The MQTT protocol adopts a dedicated configuration for logistics training: the QoS level is set to 2 (to ensure reliable transmission of model files), message topics are divided into levels according to "logistics scenario - device type - training group", and the payload uses the LZ77 compression algorithm to reduce transmission volume. After deployment, the model inference test is automatically executed, a performance optimization report adapted to the current hardware is generated, and the device network configuration and model deployment are completed, supporting the training requirements of 3-channel 1920×1080@30FPS input and 1280×720@30FPS output.

[0024] S5. Based on the edge-cloud deployment results of step S4, conduct logistics professional training operations, and monitor and evaluate the data flow, model performance and resource utilization in real time and in multiple dimensions during the training process. It should also be noted that the implementation process of step S5 is as follows: S5.1. Based on the edge-cloud deployment results of step S4, it supports the development and learning of 8 groups of 4 people in concurrent models, and provides training tasks for logistics scenario training operations including target detection, image classification, target tracking, monocular ranging, time series analysis, and unmanned system control. S5.2. Real-time monitoring of key data during the training process includes data transmission rate, model inference latency, device CPU / GPU utilization, and memory usage. The monitoring frequency is set to once per second, and real-time monitoring logs are generated. S5.3. Construct a multi-dimensional practical training evaluation system: Quality assessment: Calculate the model's accuracy, recall, F1-score, mAP50, and mAP50-95 metrics on the logistics test set; Computing power evaluation: Compare the model training time and inference speed under different hardware configurations. The evaluation metrics are the time spent per training round and the number of frames per second for inference. Efficiency assessment: Computational resource utilization includes CPU utilization, memory utilization, and network bandwidth utilization. The assessment is shown in equation (7): (7); In the formula, To score overall efficiency, For task completion rate, For average resource utilization rate, To delay satisfaction, , , These are weighting coefficients, with values ​​of 0.4, 0.3, and 0.3 respectively. S5.4 generates training evaluation reports, clearly defining training results and improvement directions, and supports the visualization and export of evaluation data.

[0025] S6. Record the key data and evaluation results of the entire process from S1 to S5, and optimize the training dataset, model parameters and training process based on user feedback to achieve dynamic iteration of the training system.

[0026] It should also be noted that the implementation process of step S6 is as follows: S6.1. Record all process data from steps S1 to S5, including requirement parameters, dataset information, model parameters, training logs, deployment configuration, and evaluation results, and store them in PostgreSQL and Mongo databases to ensure data traceability; S6.2. Based on the evaluation results of step S5, establish a problem diagnosis mechanism: if the overall mAP decreases, it is determined that the quality of the newly added dataset is not up to standard; if the mAP amplitude is not large but the mAP50-95 amplitude is large, it is determined that there is a labeling quality problem; if the resource utilization rate is less than 30%, it is determined that the resource allocation is unreasonable. S6.3. Based on feedback from logistics major teaching, optimize the training dataset by supplementing high-quality samples and correcting low-quality annotations, adjust model parameters including hyperparameter range and incremental training sample ratio, and optimize the training process by adjusting task order and resource allocation strategy. S6.4. After each iteration, re-execute steps S2 to S5 to verify the optimization effect. The overall efficiency score is then calculated. When the improvement exceeds 15% or the model accuracy improves by more than 8%, the current optimization results are solidified to form a new training benchmark configuration.

[0027] It should also be noted that, specifically, the front-end architecture of the AI-based logistics professional development and training system includes: The requirement interaction module provides a multimodal requirement input interface, supports text and document uploads and visual interaction, and is used to execute the requirement receiving operation in step S1.1. The annotation operation module provides an interface for label configuration, manual annotation, and automatic annotation result review. It supports adding, modifying, and verifying labels and is used to perform the annotation operations in steps S2.2 and S2.3. The training operation module provides a visual operation interface for logistics training tasks, supports real-time viewing of training progress and equipment status, and is used to execute the training operations in step S5.1. The evaluation and display module provides visual charts of evaluation indicators and an interface for exporting evaluation reports. It is used to display and export the evaluation results in step S5.4. The backend architecture includes: The requirement processing module is used to preprocess the input training requirements, extract core elements and calculate weights, and to execute the requirement processing operation in step S1.2. The environment configuration module is used to complete the hardware initialization, network configuration, database and training framework deployment of the server and edge devices, and to execute the environment configuration operation in step S1.3. The data management module is used for logistics data collection, storage, data augmentation and dataset partitioning, and is used to perform the data processing operations in steps S2.1 and S2.4. The model training module is used to load the pre-set model, perform hyperparameter optimization, model training and incremental training, and to perform the model training operations in steps S3.1 to S3.4. The model evaluation module is used to calculate model performance indicators and output the optimized AI model for executing the model evaluation operation in step S3.5. The model conversion module is used to perform model format conversion and compression quantization processing, and to perform the model optimization operations in steps S4.1 and S4.2. The device authentication module is used to implement secure authentication and authorization management for edge devices and to perform the device authentication operation in step S4.3. The model deployment module distributes the model to the edge device via the MQTT protocol to complete the deployment and is used to execute the model deployment operation in step S4.4. The monitoring module is used to collect key data in the training process in real time, generate monitoring logs, and execute the monitoring operations in step S5.2. The evaluation module is used to conduct quality evaluation, computing power evaluation and efficiency evaluation, generate evaluation reports, and perform the evaluation operations in steps S5.3 and S5.4. The iterative optimization module is used to record full-process data, diagnose problems, optimize training configurations and verify effects, and execute the iterative optimization operations in steps S6.1 to S6.4.

[0028] The front-end architecture also includes: The resource management interface module provides a query, editing, deletion, and download interface for datasets and model files. It supports viewing resource metadata, including name, tags, and addition time, and is used in conjunction with the backend data management module and model training module to perform resource operations. The parameter configuration interface module provides a custom configuration interface for training parameters and hyperparameter optimization parameters. It supports setting and saving parameter ranges and is used to perform parameter configuration operations in conjunction with the backend model training module. The device management interface module provides an interface for viewing the status of edge devices, configuring networks, and managing authorization. It works in conjunction with the backend device authentication module and model distribution module to perform device management operations. The backend architecture also includes: The federated learning module is used to implement base model distribution, tenant local incremental training, and model parameter isolation, ensuring that low-quality data will not have a negative impact on the global model, and to enhance the security of incremental training in step S3.4. The computing network scheduling module is used to intelligently schedule computing resources and network bandwidth, optimize the allocation of computing power and data flow during task execution, and improve the training execution efficiency in steps S3.3 and S5.1. The log storage module is used to centrally store requirement processing logs, training logs, monitoring logs, and evaluation logs. It supports log retrieval and export and is used to perform data traceability operations in conjunction with the backend iterative optimization module.

[0029] It should also be noted that the front-end architecture includes: The collaborative interaction module supports teachers to batch distribute labeled projects and models to students, provides permission management functions, and is used to adapt to the concurrent training needs of 8 groups of 4 people each. The Help & Guidance module provides practical training operation guides, answers to frequently asked questions, and access to technical support to lower the threshold for practical training. The backend architecture also includes: The multi-tenant management module is used to allocate independent computing, storage, and network resources to different training groups, ensuring data isolation from computing tasks and supporting multiple groups of concurrent training. The compatibility adaptation module is used to support the adaptation of different hardware architectures (CPU / GPU / TPU), operating systems and training frameworks, ensuring that the model runs normally on edge devices and enhancing the deployment compatibility in steps S4.1 and S4.4. The feedback processing module collects user feedback on the training process, model performance, and interface operation, categorizes and stores the feedback, and provides feedback analysis results to cooperate with the backend iterative optimization module to make optimization decisions.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A development and training method based on artificial intelligence in the field of logistics, wherein the method is implemented based on an artificial intelligence logistics development and training system, characterized in that, Includes the following steps: S1. Receive the input of logistics professional training requirements, preprocess and standardize the input content, complete the training environment configuration and resource initialization, and obtain clear training objectives and resource configuration parameters; S2. Based on the training objectives determined in step S1, collect relevant data on logistics scenarios, complete data annotation through a combination of automated annotation and manual verification, and perform data augmentation processing to generate a standardized training dataset; S3. Based on the standardized training dataset generated in step S2, construct a model training process specifically for the logistics field, perform model training, hyperparameter optimization and incremental training, and output an optimized AI model for the logistics scenario. S4. Based on the AI ​​model output in step S3, perform model format conversion and compression quantization, and distribute the model to edge devices through a security authentication mechanism to realize the practical deployment of edge-cloud collaboration. Edge-cloud collaboration adopts a closed-loop architecture of server model training and edge device inference feedback. The server receives inference latency, computing power usage, accuracy fluctuation and inference performance data of edge devices in real time, and dynamically adjusts the model optimization strategy. S5. Based on the edge-cloud deployment results of step S4, conduct logistics professional training operations, and monitor and evaluate the data flow, model performance and resource utilization in real time and in multiple dimensions during the training process. S6. Record the key data and evaluation results of the entire process from steps S1 to S5, and optimize the training dataset, model parameters and training process based on user feedback.

2. The development and training method for the logistics professional field based on artificial intelligence according to claim 1, characterized in that, The implementation process of step S1 is as follows: S1.

1. Provides a multimodal input interface for text and document uploads and visual interaction to receive logistics training requirements, including logistics scenario type, training task type, number of participants and equipment requirements; S1.

2. The input logistics training requirements are cleaned and preprocessed by word segmentation. Key training requirements are selected by the core element weight calculation algorithm, as shown in equation (1): （1）； In the formula, The weight of the j-th core element in the i-th training requirement text is set to a threshold of 0.

65. The threshold adopts a dynamic adaptation mechanism and is automatically adjusted according to the logistics scenario type: the threshold is set to 0.62 for warehousing scenarios, 0.65 for transportation scenarios, and 0.68 for delivery scenarios. Elements exceeding the threshold are judged as key elements. The word frequency of the j-th core element in the i-th training requirement text is the number of times the element appears in the text divided by the total number of elements in the text. The inverse document frequency of the j-th core element is given by equation (2): （2）； In the formula, This represents the total number of documents in the corpus for logistics training. Let j be the number of documents containing the j-th core element; This is the scenario matching coefficient, ranging from 0.85 to 1.

25. If the element belongs to the core logistics scenario, then... ,otherwise ; This is the training correlation coefficient, ranging from 0.9 to 1.

1. If an element has a direct correlation with the identified training objective, then... ,otherwise For the missing key elements, supplementary information can be obtained through an interactive completion mechanism; S1.

3. Complete the environment configuration based on the logistics training resource library, including server hardware initialization, edge device network configuration, and database and training framework deployment.

3. The development and training method based on artificial intelligence in the logistics field according to claim 2, characterized in that, The implementation process of step S2 is as follows: S2.

1. Based on the logistics scenario type determined in step S1, collect the corresponding logistics data, including cargo images, transportation videos, warehousing time-series data and sensor signals. The data format supports JPEG image format, 720P to 8K resolution video format, and time-series data text format. S2.

2. Start the built-in logistics scenario pre-labeling model and perform automatic labeling on the collected data. The labeling types include target detection box labeling, time series feature labeling and category labeling. Among them, target detection labeling supports rectangle drawing, tracking labeling and interpolation labeling functions, and category labels support custom addition of Chinese and English. S2.

3. Ensure labeling accuracy through a label consistency mapping algorithm, as shown in equation (3): （3）； In the formula, Let be the matching degree between the k-th automatically labeled result and the l-th manually verified result, with a value ranging from 0 to 1. Once the annotations are deemed consistent, the matching threshold is dynamically adjusted based on the annotation type: the threshold for object detection annotations is set to 0.85, the threshold for temporal feature annotations is set to 0.82, and the threshold for category label annotations is set to 0.

90. A consistent annotation is determined when the corresponding threshold is reached. It has been determined that manual correction is required. The annotation is deemed invalid. Let be the overlap between the k-th automatically labeled result and the l-th manually verified result. Let k be the total number of annotations for the k-th automatic annotation result. This represents the total number of annotations for the l-th manual verification result; The semantic similarity between the k-th automatically labeled result and the l-th manually verified result is calculated based on a dictionary in the logistics domain. S2.

4. Perform data augmentation on the labeled data, using a combination of flipping, shifting, scaling, cropping, adding noise, and adjusting brightness. Each original data point automatically generates 5 augmented samples. The augmented data is divided into training, validation, and test sets in a 7:2:1 ratio to form a standardized training dataset, which supports searching, editing, deleting, and downloading operations.

4. The development and training method based on artificial intelligence in the logistics field according to claim 3, characterized in that, The implementation process of step S3 is as follows: S3.

1. Based on the standardized training dataset in step S2, load pre-built models in the logistics field, including object detection models, image classification models, and time series analysis models, supporting the import of models from TensorFlow and PyTorch frameworks; configure model training parameters, including training epochs, batch size, and initial learning rate. S3.

2. Start the hyperparameter optimization process and iteratively adjust the hyperparameters using the Bayesian optimization algorithm. The default number of optimization attempts is 30, as shown in equation (4): （4）； In the formula, For the optimal combination of hyperparameters, The set of hyperparameters includes learning rate, batch size, and weight decay coefficient. For hyperparameters In the training set The posterior probability on, For hyperparameters The corresponding model performance score; S3.

3. Execute the model training process, use the gradient descent optimization algorithm to adjust the model parameters, record the model loss value and accuracy every 20 rounds during training, and automatically stop training when the loss function does not improve for 50 consecutive rounds; S3.

4. Start the incremental training mechanism. On the basis of retaining the original training dataset, add new logistics scenario samples or negative sample samples, and readjust the model parameters, as shown in equation (5): （5）； In the formula, These are the model parameters after incremental training. For the original model parameters, The learning rate decay coefficient ranges from 0.001 to 0.

01. For the loss function in the merged dataset gradient on, For the original training dataset, To add a new training dataset; S3.

5. After training is completed, the accuracy, precision, recall, F1-score and mAP of the model are calculated using the validation set, and the optimized logistics scenario AI model that meets the training requirements is output.

5. The development and training method for the logistics professional field based on artificial intelligence according to claim 4, characterized in that, The implementation process of step S4 is as follows: S4.

1. Convert the AI ​​model output from step S3. First, convert the TensorFlow and PyTorch framework models to ONNX format. During the conversion process, perform operator compatibility optimization, remove redundant high-order operators in the logistics scenario, and retain the core operators supported by Conv2d, BatchNorm, and LinearRK series chips. Then, call the RKNNToolkit tool to convert the ONNX format to RKNN format. During the conversion, take into account the computing power characteristics of the ARM architecture TPU, use layer fusion technology to merge the continuous convolutional layers and activation layers, and replace the SiLU activation function with the ReLU activation function. S4.

2. A three-stage optimization process of quantization, pruning, and distillation is adopted: First, 8-bit integer quantization is performed based on the logistics scenario dataset, and the quantization calibration adopts the KL divergence calibration algorithm to preserve the quantization accuracy of cargo features and key information of warehouse shelf edges; Second, structured pruning is performed on the model convolutional layers, and the pruning ratio is dynamically adjusted according to the computing power of the target edge device. During pruning, the core channels related to logistics target detection and classification are preserved; Finally, through knowledge distillation technology, the original high-precision model is used as the teacher model, and the quantized and pruned model is used as the student model. The distillation loss function adopts the cross-entropy loss weighted by logistics scenario features. S4.

3. Establish an edge device security authentication mechanism. The edge device actively creates an authentication file to initiate a network access application. The server decrypts and verifies the device using a symmetric encryption method with a public key medium. The authentication process is shown in equation (6): （6）； In the formula, For the authentication results, For decryption function, For encryption functions, For the server's private key, For edge device public key, The edge device information includes the device model, hardware parameters, and network address, and an additional device hardware fingerprint verification is added during the authentication process; S4.

4. After successful authentication, based on the resource configuration parameters in step S1, the converted model is distributed to edge devices within the specified IP range via the MQTT protocol. The MQTT protocol adopts a dedicated configuration for logistics training: the QoS level is set to 2, message topics are divided according to logistics scenarios, device types, and training groups, and the payload uses the LZ77 compression algorithm to reduce transmission volume. After deployment, the model inference test is automatically executed, a performance optimization report adapted to the current hardware is generated, and the device network configuration and model deployment are completed, supporting the training requirements of 3-channel 1920×1080@30FPS input and 1280×720@30FPS output.

6. The development and training method for the logistics professional field based on artificial intelligence according to claim 5, characterized in that, The implementation process of step S5 is as follows: S5.

1. Based on the edge-cloud deployment results of step S4, it supports the development and learning of 8 groups of 4 people in concurrent models, and provides training tasks for logistics scenario training operations including target detection, image classification, target tracking, monocular ranging, time series analysis, and unmanned system control. S5.

2. Real-time monitoring of key data during the training process includes data transmission rate, model inference latency, device CPU / GPU utilization, and memory usage. The monitoring frequency is set to once per second, and real-time monitoring logs are generated. S5.

3. Construct a multi-dimensional practical training evaluation system: Quality assessment: Calculate the model's accuracy, recall, F1-score, mAP50, and mAP50-95 metrics on the logistics test set; Computing power evaluation: Compare the model training time and inference speed under different hardware configurations. The evaluation metrics are the time spent per training round and the number of frames per second for inference. Efficiency assessment: Computational resource utilization includes CPU utilization, memory utilization, and network bandwidth utilization. The assessment is shown in equation (7): （7）； In the formula, To score overall efficiency, For task completion rate, For average resource utilization rate, To delay satisfaction, , , These are weighting coefficients, with values ​​of 0.4, 0.3, and 0.3 respectively. S5.4 generates training evaluation reports, clearly defining training results and improvement directions, and supports the visualization and export of evaluation data.

7. The development and training method for the logistics professional field based on artificial intelligence according to claim 6, characterized in that, The implementation process of step S6 is as follows: S6.

1. Record all process data from steps S1 to S5, including requirement parameters, dataset information, model parameters, training logs, deployment configuration, and evaluation results, and store them in PostgreSQL and Mongo databases to ensure data traceability; S6.

2. Based on the evaluation results of step S5, establish a problem diagnosis mechanism: if the overall mAP decreases, it is determined that the quality of the newly added dataset is not up to standard; if the mAP amplitude is not large but the mAP50-95 amplitude is large, it is determined that there is a labeling quality problem; if the resource utilization rate is less than 30%, it is determined that the resource allocation is unreasonable. S6.

3. Based on feedback from logistics major teaching, optimize the training dataset by supplementing high-quality samples and correcting low-quality annotations, adjust model parameters including hyperparameter range and incremental training sample ratio, and optimize the training process by adjusting task order and resource allocation strategy. S6.

4. After each iteration, re-execute steps S2 to S5 to verify the optimization effect. The overall efficiency score is then calculated. When the improvement exceeds 15% or the model accuracy improves by more than 8%, the current optimization results are solidified to form a new training benchmark configuration.

8. A development and training system based on artificial intelligence in the field of logistics, applied to any one of the development and training methods based on artificial intelligence in the field of logistics as described in claims 1-7, characterized in that, The system includes a front-end architecture and a back-end architecture; The front-end architecture includes: The requirement interaction module provides a multimodal requirement input interface, supporting text and document uploads as well as visual interaction. The annotation operation module provides an interface for label configuration, manual annotation, and automatic annotation result review, and supports label addition, modification, and consistency verification. The training operation module provides a visual operation interface for logistics training tasks, supporting real-time viewing of training progress and equipment status. The evaluation and display module provides visual charts of evaluation indicators and an interface for exporting evaluation reports. The backend architecture includes: The requirement processing module is used to preprocess the input training requirements, extract core elements, and calculate weights. The environment configuration module is used to complete the hardware initialization of the server and edge devices, network configuration, and deployment of the database and training framework. The data management module is used for logistics data collection, storage, data augmentation, and dataset partitioning. The model training module is used to load a pre-set model, perform hyperparameter optimization, model training, and incremental training. The model evaluation module is used to calculate model performance indicators and output the optimized AI model. A model conversion module, which performs model format conversion and compression quantization processing; A device authentication module, which is used to implement secure authentication and authorization management for edge devices; The model delivery module delivers the model to the edge device via the MQTT protocol to complete the deployment. The monitoring module is used to collect key data during the training process in real time and generate monitoring logs. The evaluation module is used to conduct quality evaluation, computing power evaluation, and efficiency evaluation, and generate an evaluation report. The iterative optimization module is used to record full-process data, diagnose problems, optimize training configurations, and verify effects.

9. The system according to claim 8, characterized in that, The front-end architecture also includes: The resource management interface module provides interfaces for querying, editing, deleting, and downloading datasets and model files, and supports viewing resource metadata including name, tags, and addition time. The parameter configuration interface module provides a custom configuration interface for training parameters and hyperparameter optimization parameters. The device management interface module provides an interface for viewing edge device status, configuring the network, and managing authorization. The backend architecture also includes: The federated learning module is used to implement base model distribution, tenant local incremental training, and model parameter isolation. The computing network scheduling module is used to intelligently schedule computing resources and network bandwidth, and optimize the allocation of computing power and data flow during task execution. The log storage module is used to centrally store demand processing logs, training logs, monitoring logs, and evaluation logs, and supports log retrieval and export.

10. The system according to claim 8, characterized in that, The front-end architecture also includes: The collaborative interaction module is used to support teachers to batch distribute labeled items and models to students and provides permission management functions. The help and guidance module provides practical training operation guides, answers to frequently asked questions, and access to technical support. The backend architecture also includes: A multi-tenant management module is used to allocate independent computing, storage, and network resources to different training groups. A compatibility adaptation module is used to support adaptation to different hardware architectures, operating systems and training frameworks. The feedback processing module is used to collect user feedback on the training process, model performance, and interface operation, classify and store the feedback, and provide feedback analysis results.

Images