Multi-terminal collaborative AI model dynamic deployment method, device and medium

Abstract

The invention provides a multi-terminal collaborative AI model dynamic deployment method, which comprises the following steps that: a cloud terminal splits a complete AI model into lightweight sub-models according to a real-time state reported by edge equipment and AI model structure characteristics; calculating the sequence of the sub-models distributed to the edge devices through a priority queue algorithm; differential increment transmission is adopted to push difference parameters of the existing model of the edge device; the edge device triggers incremental learning based on an end side abnormal sample, and local model parameters are updated by adopting lightweight transfer learning; the cloud receives the encrypted local model parameters and the associated sample feature values, node parameters of the multiple edge devices are aggregated through federal learning, and an optimized global model is generated. According to the multi-terminal collaborative AI model dynamic deployment method, through a model fragmentation pushing technology and a side incremental learning mechanism, model dynamic adaptation and closed-loop optimization of resource awareness are realized, and the AI model deployment efficiency and scene adaptability in a complex industrial scene are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and edge computing technology, and in particular relates to a multi-terminal collaborative AI model dynamic deployment method, device and medium. Background Art

[0002] In existing technologies, edge deployment of AI models in industrial IoT scenarios (such as construction sites and smart parks) faces the following technical bottlenecks:

[0003] 1. Model update lags and poor resource adaptability: Traditional edge computing relies on full model updates in the cloud. When edge devices (such as the N150 / HS110 gateway) experience network fluctuations, computing power fluctuations, or insufficient storage capacity, they are unable to receive the complete model in real time, resulting in model update cycles lasting hours or even days. For example, construction site safety detection models (such as PPE wear detection) are prone to inference delays or task interruptions due to the inability to dynamically adapt to local resources when lighting changes or equipment experiences high temperatures and frequency drops.

[0004] 2. Challenges of edge device heterogeneity: Different edge devices have significantly different hardware configurations (CPU / GPU computing power, memory) and operating environments (temperature, power consumption). The existing fixed model deployment model makes it difficult to dynamically adjust the model structure based on device status (such as frequency reduction triggered by excessive temperature), resulting in low resource utilization and even device overload.

[0005] 3. Insufficient scene generalization: Abnormal samples collected by edge sensors (such as cameras) (e.g., missing helmets, damaged protective clothing) cannot be promptly fed back into model optimization due to data silos. This results in reduced recognition accuracy in complex scenarios (such as occlusion and low light). Existing solutions rely on periodic full data transmission and training, which is time-consuming and cannot meet real-time requirements. Summary of the Invention

[0006] The purpose of this invention is to solve the problems in the existing technology and propose a multi-terminal collaborative AI model dynamic deployment method. Through model sharding push technology and side incremental learning mechanism, resource-aware model dynamic adaptation and closed-loop optimization are realized, significantly improving the AI model deployment efficiency and scenario adaptability in complex industrial scenarios.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0008] The multi-terminal collaborative AI model dynamic deployment method includes the following steps:

[0009] The cloud breaks down the complete AI model into lightweight sub-models based on the real-time status reported by the edge device and the structural characteristics of the AI model;

[0010] The order in which sub-models are assigned to edge devices is calculated using a priority queue algorithm;

[0011] Adopt differential incremental transmission to push the difference parameters with the existing model of edge devices;

[0012] Edge devices trigger incremental learning based on abnormal samples on the client side and use lightweight transfer learning to update local model parameters.

[0013] The cloud receives the encrypted local model parameters and associated sample feature values, aggregates the node parameters of multiple edge devices through federated learning, and generates an optimized global model.

[0014] Furthermore, the cloud breaks down the complete AI model into lightweight sub-models based on the real-time status reported by the edge device and the AI model structure characteristics, including:

[0015] The AI model is dynamically sharded into a feature extraction sub-model and a classifier sub-model;

[0016] The feature extraction sub-model is used to extract the spatial features of the image as the first priority.

[0017] The classifier sub-model is used to classify features as a second priority.

[0018] Furthermore, it also includes dynamically adjusting the AI model sharding and sub-model delivery strategy according to the status of the edge device, and giving priority to delivering high-priority sub-models to the edge device.

[0019] Furthermore, the order of calculating the sub-models and assigning them to the edge devices by using a priority queue algorithm includes:

[0020] A weighted evaluation formula is used to calculate the sub-model delivery priority for priority scheduling:

[0021] Priority = α*(computing power utilization) + β*(storage margin ratio) - γ*(device temperature threshold deviation)

[0022] Among them, the coefficients α, β, and γ are dynamically optimized and adjusted through historical deployment data and reinforcement learning to ensure that high-load devices receive key sub-models first; the computing power utilization rate is used to measure whether the device is currently in a high-load state; the storage margin ratio is used to evaluate whether the device has sufficient storage space to receive the new model; and the device temperature threshold deviation is used to avoid device performance degradation or failure due to high temperature.

[0023] Furthermore, the difference parameters between the differential incremental transmission push and the existing model of edge devices include:

[0024] The cloud maintains the current model version of a single edge device. When a new version of the model is generated, the parameter matrices of the two versions are compared layer by layer, the location and value of the parameter changes are recorded, and the differences are represented as a sparse matrix.

[0025] Convert the floating-point difference values into a low-precision format, use a lossless compression algorithm to further compress the sparse matrix index and difference values, and divide the difference data into blocks according to model layers or sub-modules.

[0026] Furthermore, it also includes using a lightweight encryption algorithm to encrypt difference data packets, recording transmission progress when the network is interrupted, retransmitting only unfinished blocks after recovery, and prioritizing the transmission of key sub-module differences based on device status.

[0027] Furthermore, the edge device triggers incremental learning based on abnormal samples on the terminal side, and uses lightweight transfer learning to update local model parameters, including:

[0028] The end-side device filters low-confidence samples using the confidence threshold, triggering incremental learning tasks on the edge side.

[0029] Edge nodes use online fine-tuning to update only local model parameters and use abnormal samples on the end side for adaptive training.

[0030] Furthermore, the cloud receives encrypted local model parameters and associated sample feature values, aggregates node parameters of multiple edge devices through federated learning, and generates an optimized global model including:

[0031] The edge side encrypts the fine-tuned model parameters and transmits them back to the cloud to participate in federated learning optimization.

[0032] Integrate the parameter updates of multiple edge nodes, assign weights according to the data volume or contribution of each node, and generate a global optimization model.

[0033] In order to achieve the above-mentioned objectives, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a program running on the processor, and when the processor runs the program, it executes the steps of the multi-terminal collaborative AI model dynamic deployment method as described above.

[0034] In order to achieve the above-mentioned objectives, the present invention also provides a computer-readable storage medium on which computer instructions are stored. When the computer instructions are executed, the steps of the multi-terminal collaborative AI model dynamic deployment method as described above are executed.

[0035] The present invention proposes a multi-terminal collaborative AI model dynamic deployment method, device, and medium, which has the following beneficial effects:

[0036] 1. Improved real-time performance: Shorten model update delays to meet the real-time response requirements of highly dynamic scenarios such as construction sites and industrial parks.

[0037] 2. Resource utilization optimization: Through dynamic sharding and priority scheduling, the peak resource consumption of edge devices is reduced, avoiding frequency reduction and downtime caused by excessive temperature.

[0038] 3. Enhanced scene generalization: The closed-loop feedback mechanism improves the algorithm's detection accuracy under complex conditions and significantly reduces the false negative rate.

[0039] This invention overcomes the difficulties of resource rigidity and update lag in traditional edge computing through the collaborative design of AI model sharding, edge incremental learning and closed-loop feedback, and has significant technological advancement and industrial application value.

[0040] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 This is a flowchart of a multi-terminal collaborative AI model dynamic deployment method of the present invention;

[0043] Figure 2 A schematic diagram of the structure of an electronic device 300 for implementing an embodiment of the present application. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0045] Example 1

[0046] Figure 1 The following is a flow chart of the dynamic deployment method of the multi-terminal collaborative AI model according to the present invention. Figure 1 , the multi-terminal collaborative AI model dynamic deployment method of the present invention is described in detail.

[0047] In step 101, the cloud breaks down the complete AI model (such as YOLOv5-PPE) into lightweight sub-models based on the real-time status (computing power, storage, temperature) reported by the edge device and the structural characteristics of the AI model (such as the dependencies between neural network layers).

[0048] Optionally, the dynamic sharding engine dynamically shards the AI model into two sub-modules:

[0049] Feature extraction sub-model (priority 1): responsible for extracting spatial features of the image.

[0050] Classifier sub-model (priority 2): responsible for classifying features.

[0051] The design of this sharding strategy needs to consider the following factors:

[0052] Computational complexity: The feature extraction sub-model is usually heavier (such as the convolution layer of ResNet-18) and is prioritized for delivery and execution. The classifier sub-model is lighter and can be quickly loaded on edge devices.

[0053] Priority management: Prioritize the delivery of high-priority sub-models to ensure that edge devices can complete preliminary reasoning with limited resources.

[0054] Dynamic adjustment: Dynamically adjust sharding and delivery strategies based on the status of edge devices (such as CPU utilization, storage capacity, and temperature).

[0055] In this embodiment, the cloud is the core of global resource management and model optimization, responsible for global model training, sharding strategy generation, resource scheduling, and multi-node data aggregation. This includes cloud computing servers (such as GPU clusters and distributed storage systems) and / or central management platforms (such as model training engines and federated learning aggregation servers).

[0056] In this embodiment, the feature extraction layer is separated from the classification decision layer, and sub-models with low computing power requirements are preferentially distributed to resource-constrained devices (such as low-power cameras or embedded devices). This sharding strategy can adapt to the hardware limitations of different devices and ensure that AI models can run efficiently on heterogeneous devices.

[0057] In this embodiment, sub-model deployment and inference are executed on the edge, including receiving model shards (such as feature extraction sub-model) sent from the cloud and executing local AI tasks (such as helmet detection).

[0058] For example, in a construction site scenario, the cloud might be a central server deployed on a public cloud (such as Huawei Cloud), responsible for managing model updates and optimization tasks for all edge gateways and cameras on site. If the edge device's storage capacity falls below 200MB, only the feature extraction sub-model is deployed, and the classification task is transferred back to the cloud. If the device temperature approaches a threshold (such as 45°C), model updates are suspended to prioritize normal device operation.

[0059] In step 102, the order in which the sub-models are sent is calculated using a priority queue algorithm.

[0060] Optionally, the cloud dynamically calculates the sub-model delivery priority through a weighted formula and performs priority scheduling to ensure that high-load devices receive key modules first.

[0061] Optionally, a priority queue algorithm includes using a weighted evaluation formula to calculate the priority of the sub-model delivery:

[0062] Priority = α*(computing power utilization) + β*(storage margin ratio) - γ*(device temperature threshold deviation)

[0063] The coefficients α, β, and γ are dynamically optimized and adjusted using historical deployment data and reinforcement learning to ensure that high-load devices prioritize receiving key sub-models to meet the needs of different scenarios. Computing power utilization measures whether the device is currently under high load; storage headroom ratio assesses whether the device has sufficient storage space to receive new models; and device temperature threshold deviation prevents device performance degradation or failure due to high temperatures.

[0064] The priority queue algorithm ensures that critical sub-models are preferentially assigned to high-load devices while avoiding inefficient or unsafe deployment.

[0065] In step 103, the model difference parameters are pushed using differential incremental transmission technology.

[0066] Optionally, perform difference detection to identify parameter differences between the old and new models. This involves maintaining the current model version (e.g., v1.0) of each edge device in the cloud. When a new version of the model (e.g., v2.0) is generated, compare the parameter matrices (e.g., neural network weight matrices) of the two versions layer by layer. Record the location and numerical change of the parameter changes. For example, only mark the parameters in the output layer weight matrix whose change exceeds a threshold (e.g., ±0.01). Represent the differences as a sparse matrix (storing only non-zero difference values and their position indices) to reduce the amount of redundant data.

[0067] Optionally, it also includes difference encoding and compression to optimize the storage and transmission efficiency of difference data, including quantization compression: converting floating-point difference values into a low-precision format (such as 16-bit floating-point numbers or 8-bit integers) to reduce data volume. Entropy coding: using lossless compression algorithms (such as Huffman coding) to further compress sparse matrix indices and difference values. Block packaging: dividing difference data into blocks by model layer or sub-module (such as feature extraction layer differences, classifier differences), supporting on-demand batch transmission.

[0068] Optionally, encrypted transmission is also included to ensure the security and reliability of differential data during transmission. This includes encrypting differential data packets using lightweight encryption algorithms (such as AES-128) to prevent man-in-the-middle attacks. Resumable transmission: During network interruptions, the transmission progress is recorded and only unfinished blocks are retransmitted upon recovery. Priority scheduling: Based on device status (such as remaining bandwidth and task urgency), critical submodule differences (such as classifier updates) are prioritized for transmission.

[0069] Optionally, it also includes edge-side differential merging, completing the differential merging of the old and new models on the edge device. This includes data verification: the receiving end calculates the hash value (such as SHA-256) of the differential packet and compares it with the hash value sent by the cloud to ensure data integrity. Parameter update: Based on the index position in the differential packet, the old model parameters are superimposed with the differential value (for example, new weight = old weight + Δ weight). Model hot loading: Loading updated parameters without shutting down the system to avoid service interruption.

[0070] Optionally, rollback and fault-tolerance mechanisms are also included to ensure rapid recovery in the event of an update failure, including version snapshots: the edge device backs up the current model parameters before the update and stores them as a snapshot file (such as v1.0_backup). Health monitoring: After the update, diagnostic tests (such as inference speed and accuracy verification) are run. If an anomaly is detected (such as a drop in accuracy of more than 5%), the system automatically rolls back to the previous version. Log feedback: Update results (success / failure) and device status logs are reported to the cloud to optimize subsequent transmission strategies.

[0071] In this embodiment, differential incremental transmission only pushes the difference parameters from the existing model on the edge side, rather than the entire model, reducing bandwidth consumption by more than 40% and significantly improving model update efficiency. It is suitable for large-scale distributed AI systems, especially when the number of edge computing nodes is large. Real-time performance improvement: Block transmission and breakpoint resumption support efficient updates in weak network environments. Security assurance: Encryption and verification mechanisms prevent data tampering and leakage. Lightweight resources: The computational overhead of merging differences on the edge is extremely low (only matrix addition operations are required), which is suitable for low-computing-power devices.

[0072] In this embodiment, the edge is the middle layer close to the terminal device, including edge nodes with certain computing capabilities (such as the HS110 edge gateway and edge server of the industrial gateway), which are responsible for local model inference, incremental learning and real-time feedback.

[0073] As an example, in a construction site safety inspection scenario, the old model is YOLOv5-PPE v1.0 (the feature extraction layer weights remain unchanged, but the classifier layer is updated). Difference generation: The cloud marks 200 parameter changes in the classifier layer (the sum of the delta values is +0.15). Transmission content: Only the classifier layer difference package (compressed size 50KB, full model size 200MB) is sent. Edge update: The HS110 gateway receives the difference package, completes the classifier parameter update within 5 seconds, and upgrades the model version to v1.1.

[0074] In step 104, the edge triggers incremental learning based on abnormal samples on the device side and updates local model parameters;

[0075] Optionally, the edge-side abnormal sample screening of the edge-side incremental learning mechanism includes: the terminal device (such as a camera) automatically screens low-confidence samples (such as blurred / occluded helmet images) through a confidence threshold (such as <85%), triggering the edge-side incremental learning task.

[0076] In this embodiment, incremental learning fine-tuning is performed on the edge, including updating local model parameters based on abnormal samples (such as low-confidence images) screened on the end side using lightweight transfer learning (such as Fast Adaptive Meta-Learning).

[0077] In this embodiment, the end-side is the terminal device that directly interacts with the physical environment. It is typically resource-constrained (low computing power and limited storage) and is primarily responsible for data collection, preliminary filtering, and triggering feedback. Typical devices / roles on the end-side include industrial cameras, sensors (such as temperature and humidity sensors and vibration sensors), and embedded devices (such as ARM-based industrial computers).

[0078] Optionally, the end-side performs data collection and anomaly screening, including automatically filtering low-confidence samples (such as blurry images of hard hats) based on a confidence threshold (e.g., <85%). The end-side triggers the edge task by sending an incremental learning request to the edge when the accumulation of abnormal sample data reaches a threshold. The trigger condition is that when a certain number of abnormal samples accumulate, the edge-side incremental learning task is triggered.

[0079] In this embodiment, for the helmet detection task, if the model's prediction confidence for certain images (such as blurred or occluded helmets) is lower than 85%, they are marked as abnormal samples.

[0080] As an example, smart cameras on a construction site continuously collect images of workers. If they detect that a worker is not wearing a hard hat and the confidence level is less than 85%, it is marked as an abnormal sample and triggers the model optimization process of the edge gateway.

[0081] In this embodiment, lightweight computing on the end side includes performing simple preprocessing (such as image cropping and noise reduction) to reduce computing pressure on the edge side.

[0082] Alternatively, lightweight transfer learning can be used: edge nodes use online fine-tuning techniques (such as Fast Adaptive Meta-Learning) to update only the last layer or local parameters of the model, and use abnormal samples on the end side for adaptive training, avoiding the resource overhead of full training.

[0083] In this embodiment, lightweight transfer learning edge nodes use online fine-tuning techniques (such as Fast Adaptive Meta-Learning (FAML)) to update only the last layer or local parameters of the model. This reduces computational overhead and is suitable for resource-limited edge devices. It also allows for rapid adaptation to new scenarios and improves model generalization.

[0084] In this embodiment, training is performed using abnormal samples uploaded from the client side, forming a closed-loop feedback mechanism. The significance of closed-loop feedback optimization lies in real-time and robustness. Through collaboration between the client and the edge, the model is dynamically optimized without relying on cloud intervention. New samples are continuously incorporated for fine-tuning, allowing the model to gradually adapt to complex and changing real-world environments.

[0085] As an example, the HS110 edge gateway deployed at a construction site receives the YOLOv5-PPE sub-model split from the cloud, analyzes the camera video stream in real time, and triggers local model fine-tuning when abnormal samples are detected.

[0086] In step 105, the encrypted feedback fine-tuning parameters are sent to the cloud, and an optimized global model is generated through federated learning.

[0087] In this embodiment, the edge side performs encrypted feedback, including encrypting the fine-tuned parameters and transmitting them back to the cloud to participate in federated learning optimization.

[0088] Optionally, the fine-tuned model parameters and associated sample feature values are encrypted and transmitted back to the cloud. The cloud aggregates the encrypted model parameters of multiple edge nodes through federated learning to generate a globally optimized model version, forming an "end-edge-cloud" closed-loop optimization chain.

[0089] In this embodiment, the associated sample feature value refers to the key feature information directly related to model fine-tuning extracted from the raw data collected by the end-side device (such as cameras and sensors) (such as images and temperature readings). The raw data collected is an abstract representation generated by lightweight feature extraction methods (such as PCA dimensionality reduction and CNN shallow features).

[0090] For example, in helmet detection, it is possible to extract image features such as color histogram, edge density, and target object outline. In equipment failure prediction, it is possible to extract time series features such as frequency domain energy distribution and peak interval of vibration sensors.

[0091] Optionally, the association of sample feature values with model updates: These feature values directly reflect the characteristics of abnormal samples that trigger model fine-tuning (such as the characteristic patterns of low-confidence images). Association with parameter optimization: The cloud analyzes the distribution of feature values to determine which data features lead to parameter changes (for example, the classifier's sensitivity to blurred images increases).

[0092] Optionally, the steps of generating an optimized global model through federated learning may be specifically performed as follows:

[0093] (1) Each edge node updates the local model parameters based on abnormal samples on the end side and performs local model fine-tuning.

[0094] Input: Low-confidence samples (such as blurred hard hat images) filtered by the edge device (such as the HS110 gateway) receiving the end side.

[0095] Incremental learning: Use lightweight transfer learning (such as Fast Adaptive Meta-Learning) to fine-tune only the last layer or specific layer parameters of the model (such as classifier weights).

[0096] Output: Updated local model parameters (such as the weight matrix ΔW of the classifier layer).

[0097] As an example, a construction site edge node changes the weight of the helmet classifier from W to 旧 Updated to W 新 =W 旧 +ΔW.

[0098] (2) Parameter encryption and upload, protecting the privacy of model parameters uploaded by edge nodes.

[0099] Homomorphic encryption: Encrypt the fine-tuned parameter ΔW (such as Paillier homomorphic encryption) to ensure that the cloud cannot directly parse the original parameter.

[0100] Metadata addition: Associates non-sensitive information such as device ID, model version, and data volume for aggregation weight allocation.

[0101] Transmission protocol: Upload encrypted parameters to the cloud via HTTPS or a dedicated secure channel.

[0102] As an example, an edge node associates the encrypted classifier parameter Encrypt(ΔW) with the model version v1.2 and uploads it.

[0103] (3) Cloud-side parameter aggregation integrates parameter updates of multiple edge nodes to generate a global optimization model.

[0104] Weighted average strategy: assign weights based on the amount of data or contribution of each node.

[0105]

[0106] Among them, n i is the number of local samples of the i-th edge node; N 总 is the total number of samples of all nodes.

[0107] Decryption and aggregation: If homomorphic encryption is used, weighted calculations are performed directly in the encrypted state, and then the results are decrypted.

[0108] Outlier filtering: Eliminate parameters that deviate too much from the mean (such as Z-Score>3) to prevent interference from malicious nodes.

[0109] As an example, the classifier parameters of 10 construction sites are aggregated in the cloud. A node gets a higher weight because the sample size accounts for 20%, and the global classifier weight W is finally generated. 全局 .

[0110] (4) Global model update: integrate the aggregated parameters into the basic model to generate a new version of the global model.

[0111] Parameter replacement: Replace global aggregation parameters (such as W 全局 ) replaces the corresponding layer parameters of the cloud-based basic model.

[0112] Model Validation: Test the performance of new models on a cloud-based validation set (e.g., a mixed dataset across construction sites) to ensure improved accuracy.

[0113] Version iteration: Marked as v2.1, record the update log (such as "Optimizing the classifier's recognition of blurred images").

[0114] As an example, the classifier of the global model v2.1 improved its accuracy on the test set from 82% to 87%, and the false negative rate decreased by 5%.

[0115] (5) The model is distributed in shards, and the optimized global model is dynamically deployed to edge devices.

[0116] Dynamic sharding: Re-split the model based on the device status (for example, the feature extraction layer remains unchanged, and the updated classifier sub-model is sent down).

[0117] Differential incremental transmission: Only the difference parameters of the classifier are pushed to reduce the amount of transmitted data.

[0118] Hot update: Edge devices do not need to be restarted after merging different parameters, and continue to provide services.

[0119] As an example, an edge gateway receives a classifier difference package (50KB), completes model hot loading within 5 seconds, and upgrades the version to v2.1.

[0120] As an example, in construction site safety inspection, multiple nodes collaborate:

[0121] Edge gateway at construction site A: Updates the classifier parameter ΔW1 based on 100 fuzzy images.

[0122] Edge gateway at construction site B: Updates the classifier parameter ΔW2 based on 80 occlusion images.

[0123] Cloud aggregation:

[0124] W is generated by weighting the sample size (site A weight 55%, site B weight 45%) 全局 =0.55ΔW1+0.45ΔW2.

[0125] In this embodiment, federated learning enables the following closed-loop optimization process: Localized learning: Edge devices fine-tune models using device-side data to prevent raw data leakage. Secure aggregation: Encrypted parameter upload and a weighted averaging strategy ensure global model quality. Dynamic deployment: Differential incremental technology efficiently synchronizes optimization results to the edge. This approach improves model generalization while strictly meeting the data privacy and real-time requirements of industrial scenarios.

[0126] As an example, taking PPE wear detection at a construction site as an example, the three-party collaborative process includes the following steps:

[0127] Cloud initialization: Deploy the ResNet-18 basic model, set the temperature threshold to 45°C, and set the minimum storage capacity to 200MB.

[0128] Edge device registration: The HS110 edge gateway reports real-time status (CPU utilization 62%, storage remaining 1.2GB, temperature 38°C);

[0129] Dynamic model sharding: The cloud splits the detection model into a feature extraction sub-model (priority 1) and a classifier sub-model (priority 2), and distributes them in order.

[0130] Device-side data trigger: The camera captures five frames of images without a helmet (confidence <85%), triggering an edge incremental learning request.

[0131] On the edge: The HS110 gateway receives incremental learning requests, uses Fast Adaptive Meta-Learning to fine-tune classifier parameters, and sends encrypted feedback of the updated parameters to the cloud.

[0132] Cloud: Aggregate model parameters from multiple construction sites to generate a global optimization model v2.1, dynamically shard it, and send it to edge devices to complete closed-loop optimization.

[0133] This paper proposes a multi-device collaborative dynamic deployment method for AI models. Through dynamic sharding and priority scheduling, it fully utilizes device resources and avoids resource waste. Differential incremental transmission and lightweight transfer learning significantly reduce computing, storage, and communication overhead. A closed-loop feedback mechanism enables the model to self-optimize and adapt to new scenarios and requirements. It is suitable for large-scale distributed systems and supports the collaborative operation of heterogeneous devices.

[0134] Example 2

[0135] Reference below Figure 2 , which shows a block diagram of an electronic device 300 suitable for implementing an embodiment of the present application. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0136] like Figure 2 As shown, the electronic device 300 may include a processing device 310, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 320 or a program loaded from a storage device 380 into a random access memory (RAM) 330. The processing device 310 can be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processing device 310 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processing device 310 performs the various methods and processes described above.

[0137] Various programs and data required for the operation of the electronic device 300 are also stored in the RAM 330. The processing device 310, the ROM 320, and the RAM 330 are connected to each other via a bus 340. An input / output (I / O) interface 350 is also connected to the bus 340.

[0138] Typically, the following devices may be connected to the I / O interface 350: an input device 360 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 370 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 380 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 390. The communication device 390 may allow the electronic device 300 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 2 The electronic device 300 is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead.

[0139] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication device 390, or installed from the storage device 380, or installed from the ROM 320. When the computer program is executed by the processing device 310, the above-mentioned functions defined in the method of the embodiment of the present application are performed. Alternatively, in other embodiments, the processing device 310 may be configured to execute the method in any other appropriate manner (for example, by means of firmware): the cloud decomposes the complete AI model into lightweight sub-models based on the real-time status and AI model structural characteristics reported by the edge device; calculates the order in which the sub-models are assigned to the edge device through a priority queue algorithm; uses differential incremental transmission to push the difference parameters from the existing model of the edge device; the edge device triggers incremental learning based on abnormal samples on the end side, and uses lightweight transfer learning to update local model parameters; the cloud receives encrypted local model parameters and associated sample feature values, aggregates node parameters of multiple edge devices through federated learning, and generates an optimized global model.

[0140] Example 3

[0141] The computer-readable medium mentioned above in this application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this application, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, device or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. This propagated data signal can take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0142] In some embodiments, the client and server can communicate using any currently known or future developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0143] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0144] The above-mentioned computer-readable medium carries one or more programs. When the above-mentioned one or more programs are executed by the electronic device, the electronic device: the cloud breaks down the complete AI model into lightweight sub-models according to the real-time status and AI model structure characteristics reported by the edge device; calculates the order in which the sub-models are assigned to the edge device through the priority queue algorithm; uses differential incremental transmission to push the difference parameters of the existing model of the edge device; the edge device triggers incremental learning based on abnormal samples on the terminal side, and uses lightweight transfer learning to update the local model parameters; the cloud receives the encrypted local model parameters and associated sample feature values, aggregates the node parameters of multiple edge devices through federated learning, and generates an optimized global model.

[0145] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including, but not limited to, object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0146] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0147] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the module itself.

[0148] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof.

[0149] The program code for implementing the methods of the present application can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flow charts and / or block diagrams are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0150] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0151] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0152] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0153] A computer system may include a client and a server. The client and server are typically remote from each other and typically interact via a communication network. This client-server relationship arises through computer programs running on the respective computers, creating a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosting and VPS services. The server may also be a server in a distributed system or a server integrated with blockchain.

[0154] Artificial intelligence (AI) is the study of how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily encompass computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graphs.

[0155] Cloud computing refers to a technology system that provides network access to elastically scalable shared pools of physical or virtual resources. These resources can include servers, operating systems, networks, software, applications, and storage devices, and can be deployed and managed on-demand in a self-service manner. Cloud computing technology provides efficient and powerful data processing capabilities for the application of technologies such as artificial intelligence and blockchain, as well as for model training.

[0156] The embodiments described above are only part of the embodiments of the present application, not all of them. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0157] In the description of the embodiments of the present application, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, or are the orientation or position relationship in which the product of the application is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the electric vehicle or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", "third", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0158] In the description of the embodiments of the present application, it should also be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0159] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the scope of protection of the claims.

Claims

1. A multi-terminal collaborative AI model dynamic deployment method, characterized by: include: The cloud breaks down the complete AI model into lightweight sub-models based on the real-time status reported by the edge device and the structural characteristics of the AI model; The order in which sub-models are assigned to edge devices is calculated using a priority queue algorithm; Adopt differential incremental transmission to push the difference parameters with the existing model of edge devices; Edge devices trigger incremental learning based on abnormal samples on the client side and use lightweight transfer learning to update local model parameters. The cloud receives the encrypted local model parameters and associated sample feature values, aggregates the node parameters of multiple edge devices through federated learning, and generates an optimized global model.

2. The multi-terminal collaborative AI model dynamic deployment method according to claim 1 is characterized in that: The cloud breaks down the complete AI model into lightweight sub-models based on the real-time status reported by the edge device and the AI model structure characteristics, including: The AI model is dynamically sharded into a feature extraction sub-model and a classifier sub-model; The feature extraction sub-model is used to extract the spatial features of the image as the first priority. The classifier sub-model is used to classify features as a second priority.

3. The multi-terminal collaborative AI model dynamic deployment method according to claim 2 is characterized in that: It also includes dynamically adjusting the AI model sharding and sub-model delivery strategies based on the status of edge devices, giving priority to delivering high-priority sub-models to edge devices.

4. The multi-terminal collaborative AI model dynamic deployment method according to claim 1 is characterized in that: The order of calculating the sub-models and assigning them to the edge devices by the priority queue algorithm includes: A weighted evaluation formula is used to calculate the sub-model delivery priority for priority scheduling: Priority = α*(computing power utilization) + β*(storage margin ratio) - γ*(device temperature threshold deviation) Among them, the coefficients α, β, and γ are dynamically optimized and adjusted through historical deployment data and reinforcement learning to ensure that high-load devices receive key sub-models first; the computing power utilization rate is used to measure whether the device is currently in a high-load state; the storage margin ratio is used to evaluate whether the device has sufficient storage space to receive the new model; and the device temperature threshold deviation is used to avoid device performance degradation or failure due to high temperature.

5. The multi-terminal collaborative AI model dynamic deployment method according to claim 1 is characterized in that: The difference parameters between the differential incremental transmission push and the existing model of edge devices include: The cloud maintains the current model version of a single edge device. When a new version of the model is generated, the parameter matrices of the two versions are compared layer by layer, the location and value of the parameter changes are recorded, and the differences are represented as a sparse matrix. Convert the floating-point difference values into a low-precision format, use a lossless compression algorithm to further compress the sparse matrix index and difference values, and divide the difference data into blocks according to model layers or sub-modules.

6. The multi-terminal collaborative AI model dynamic deployment method according to claim 5 is characterized in that: It also includes encrypting difference data packets using a lightweight encryption algorithm, recording transmission progress when the network is interrupted, retransmitting only unfinished blocks after recovery, and prioritizing the transmission of key sub-module differences based on device status.

7. The multi-terminal collaborative AI model dynamic deployment method according to claim 1 is characterized in that: The edge device triggers incremental learning based on abnormal samples on the terminal side and uses lightweight transfer learning to update local model parameters, including: The end-side device filters low-confidence samples using the confidence threshold, triggering incremental learning tasks on the edge side. Edge nodes use online fine-tuning to update only local model parameters and use abnormal samples on the end side for adaptive training.

8. The multi-terminal collaborative AI model dynamic deployment method according to claim 1 is characterized in that: The cloud receives encrypted local model parameters and associated sample feature values, aggregates node parameters of multiple edge devices through federated learning, and generates an optimized global model including: The edge side encrypts the fine-tuned model parameters and transmits them back to the cloud to participate in federated learning optimization. Integrate the parameter updates of multiple edge nodes, assign weights according to the data volume or contribution of each node, and generate a global optimization model.

9. An electronic device, characterized in that: It includes a memory and a processor, the memory stores a program running on the processor, and the processor executes a multi-terminal collaborative AI model dynamic deployment method as described in any one of claims 1-8 when running the program.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed, a multi-terminal collaborative AI model dynamic deployment method as described in any one of claims 1-8 is executed.

Images