End-side cloud collaborative AI process dynamic arrangement method based on optimization graph and related equipment

Abstract

The invention discloses an end-side cloud collaborative AI process dynamic arrangement method based on an optimization graph and related equipment, and can be applied to the technical field of artificial intelligence. According to the method, the resource configuration demand information of the corresponding AI task unit is borne by the node table in the optimization graph, the constraint information of the data transmission between the two connected nodes is borne by the edge in the optimization graph, and then the real-time resource state of the resource pool is combined; a target mapping scheme for scheduling the optimization graph nodes to the target resources in the resource pool is generated by using an intelligent decision algorithm, so that the resource scheduling flexibility and the resource utilization efficiency in the heterogeneous environment can be effectively improved; moreover, at least one intelligent cooperative node can be set in the optimization graph, and related operation configuration of incremental learning and model updating can be carried out in the intelligent cooperative node, so that the stiffness degree of the AI model can be effectively reduced.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and related equipment for dynamic orchestration of edge-cloud collaborative AI processes based on optimization graphs. Background Technology

[0002] In related technologies, with the popularization of artificial intelligence applications, a single cloud computing model can no longer meet the needs of all scenarios, and edge-cloud collaborative computing has become a new application model. However, existing technologies face the following problems in deploying and applying AI processes in heterogeneous environments: Resource heterogeneity: End devices, edge devices, and cloud devices differ greatly in computing architecture (CPU / GPU / NPU), capabilities, and network conditions, making it difficult for traditional static scheduling strategies to achieve efficient utilization; Task complexity: A complete AI application (such as smart power grid inspection) usually includes multiple stages such as data preprocessing, model inference, and model updating. These stages have different resource requirements and complex dependencies. Dynamics and Adaptability: The dynamic distribution of edge data requires AI models to be able to adapt and update quickly, while the traditional "cloud training, edge deployment" model has high latency and is not flexible enough. Low collaboration efficiency: Most existing workflow engines (such as Airflow) only focus on task dependencies and lack awareness and optimization of underlying heterogeneous resources, thus failing to achieve efficient cross-level collaboration. In summary, the technical problems existing in related technologies need to be improved. Summary of the Invention

[0003] The main objective of this application is to propose a dynamic orchestration method and related equipment for edge-cloud collaborative AI processes based on optimization graphs, which can improve the flexibility of resource scheduling and the efficiency of resource utilization in heterogeneous environments, and reduce the rigidity of AI models.

[0004] To achieve the above objectives, one aspect of this application proposes a dynamic orchestration method for edge-cloud collaborative AI processes based on an optimized graph, the method comprising the following steps: Obtain the AI ​​workflow to be processed; The AI ​​workflow to be processed is parsed into an optimization graph, which includes multiple nodes and multiple edges. Each node carries the resource configuration requirements of the corresponding AI task unit; each edge carries the constraint information of data transmission between two connected nodes; and the multiple nodes include at least one intelligent collaborative node. Obtain the real-time resource status of a resource pool composed of heterogeneous resources from endpoints, edges, and cloud; based on the resource configuration requirements of nodes in the optimization graph, the data transmission constraints of edges, and the real-time resource status of the resource pool, use an intelligent decision-making algorithm to generate a target mapping scheme that schedules the nodes in the optimization graph to target resources in the resource pool, wherein the intelligent decision-making algorithm includes an online reinforcement learning algorithm and / or an offline metaheuristic optimization algorithm; The AI ​​workflow to be processed is executed according to the target mapping scheme; The intelligent collaborative node is configured to perform at least one of the following operations: It receives model optimization seed parameters from the cloud and performs automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment. The mixing ratio of new and old training data is dynamically determined based on a genetic algorithm, and the AI ​​model embedded in the node is incrementally learned and updated.

[0005] In some embodiments, the resource pool includes computing resources, storage resources, network resources, AI-specific resources, device status and load information, and software and runtime resources.

[0006] In some embodiments, generating a target mapping scheme by using an intelligent decision-making algorithm to schedule the optimized graph node to the target resource in the resource pool includes: If the AI ​​workflow to be processed is an online task that arrives in real time, then the target mapping scheme is generated using an online decision model based on deep reinforcement learning. If the AI ​​workflow to be processed contains a batch of tasks formed by caching, a fused recurrent neural network is used to divide the tasks into batches, and a metaheuristic optimization algorithm is used to schedule and optimize the tasks within the batches to generate the target mapping scheme.

[0007] In some embodiments, generating the target mapping scheme using an online decision model based on deep reinforcement learning includes: The resource allocation requirement information is encoded into a first state vector; The real-time resource status is encoded into a second state vector; Based on the first state vector and the second state vector, a deep neural network is used to predict the probability distribution of actions. Based on the action probability distribution, determine the candidate mapping decisions from the nodes in the optimization graph to the target physical resources in the global resource view; The simulation operation of the AI ​​workflow to be processed is performed based on the candidate mapping decision; Based on the simulation results, a target mapping decision is determined from among the multiple candidate mapping decisions as the target mapping scheme.

[0008] In some embodiments, after determining the target mapping decision as the target mapping scheme from a plurality of candidate mapping decisions based on simulation operation results, the method further includes the following steps: Obtain the actual execution effect corresponding to the target mapping scheme; Calculate the reward value based on the actual performance results; The network weights of the reinforcement learning algorithm are updated based on the reward value.

[0009] In some embodiments, the use of a fused recurrent neural network for task batch partitioning includes: The tasks corresponding to all the AI ​​workflows to be processed are cached to form a task set; The time-series patterns of all task arrivals within the task set are obtained by fusing recurrent neural networks. All tasks are batched based on the time series pattern.

[0010] In some embodiments, the step of combining metaheuristic optimization algorithms to optimize the scheduling of tasks within a batch to generate the target mapping scheme includes: Determine the execution method and execution order for each task within each batch; Calculate the overall execution cost of the corresponding batch based on the target mapping scheme, the execution method, and the execution order; The particle position is updated based on the comprehensive execution cost, whereby the particle represents the smallest resource unit. Once the particle position update process is determined to meet the convergence condition, the target mapping scheme is generated.

[0011] To achieve the above objectives, another aspect of this application proposes a dynamic orchestration device for edge-cloud collaborative AI processes based on an optimized graph, the device comprising: The first module is used to acquire the AI ​​workflow to be processed; The second module is used to parse the AI ​​workflow to be processed into an optimization graph, which includes multiple nodes and multiple edges. Each node carries the resource configuration requirements of the corresponding AI task unit; each edge carries the constraint information of data transmission between two connected nodes; and the multiple nodes include at least one intelligent collaborative node. The third module is used to obtain the real-time resource status of the resource pool, which consists of heterogeneous resources from the terminal, edge, and cloud. The fourth module is used to generate a target mapping scheme for scheduling the nodes of the optimization graph to target resources in the resource pool based on the resource configuration requirement information of the nodes in the optimization graph, the data transmission constraint information of the edges, and the real-time resource status of the resource pool, using an intelligent decision algorithm. The intelligent decision algorithm includes an online reinforcement learning algorithm and / or an offline metaheuristic optimization algorithm. The fifth module is used to execute the AI ​​workflow to be processed according to the target mapping scheme; The intelligent collaborative node is configured to perform at least one of the following operations: It receives model optimization seed parameters from the cloud and performs automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment. The mixing ratio of new and old training data is dynamically determined based on a genetic algorithm, and the AI ​​model embedded in the node is incrementally learned and updated.

[0012] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0013] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0014] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0015] The embodiments of this application include at least the following beneficial effects: This application provides a method and related equipment for dynamic orchestration of edge-cloud collaborative AI processes based on optimization graphs. This solution parses the AI ​​workflow to be processed into an optimization graph. The node table in the optimization graph carries the resource configuration requirements of the corresponding AI task units, and the edges in the optimization graph carry the constraint information of data transmission between two connected nodes. This enables a paradigm shift from static dependency to dynamic resource optimization. Then, combined with the real-time resource status of the resource pool composed of heterogeneous resources from the edge, cloud, and endpoints, an intelligent decision-making algorithm, including online reinforcement learning algorithms and / or offline metaheuristic optimization algorithms, is used to generate a target mapping scheme that schedules the optimization graph nodes to target resources in the resource pool. This enables closed-loop optimization from heuristic rules to perception-decision-learning, and then executes the AI ​​workflow to be processed according to the target mapping scheme, thereby effectively improving the flexibility of resource scheduling and resource utilization efficiency in heterogeneous environments. Furthermore, this embodiment can also set at least one intelligent collaborative node in the optimization graph, which is configured to perform at least one of the following operations: receive model optimization seed parameters from the cloud and perform automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment; or dynamically determine the mixing ratio of new and old training data based on a genetic algorithm to incrementally learn and update the AI ​​model embedded in the node, enabling the process to have the ability to dynamically evolve and continuously optimize. Therefore, this application solves the core problems of low resource utilization efficiency, rigid scheduling, and poor model adaptability in AI process orchestration in heterogeneous environments by defining an optimization graph carrying resource and constraint information, constructing a hybrid intelligent decision engine with adaptive task types, and introducing intelligent collaborative nodes with embedded automated learning capabilities in a three-pronged approach. Attached Figure Description

[0016] Figure 1 This is a flowchart of the edge-cloud collaborative AI process dynamic orchestration method based on optimization graph provided in the embodiments of this application; Figure 2 This is a flowchart of dynamic orchestration decision-making based on reinforcement learning provided in an embodiment of this application; Figure 3 This is a flowchart of the hybrid scheduling method that integrates RNN and particle swarm optimization provided in the embodiments of this application; Figure 4 This is a schematic diagram of the framework corresponding to the edge-cloud collaborative AI process dynamic orchestration method based on optimization graph provided in the embodiments of this application; Figure 5 This is a schematic diagram of the optimized process corresponding to the smart reading flow of the power grid meter provided in the embodiments of this application; Figure 6 This is a schematic diagram of the internal operation of the intelligent collaborative node provided in the embodiments of this application; Figure 7This is a schematic diagram of the structure of the edge-cloud collaborative AI process dynamic orchestration device based on optimization graph provided in the embodiments of this application; Figure 8 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0018] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0019] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms used in the embodiments of this application will be explained first. The nouns and terms used in the embodiments of this application shall be interpreted as follows: Containers are a lightweight virtualization technology that packages applications and their dependencies together to form an independent runtime environment through operating system-level isolation mechanisms (such as Linux namespaces and cgroups). Containers can share host resources (such as CPU, memory, and disk), but process isolation is achieved through resource restrictions.

[0022] Resource allocation requirements information is a comprehensive description of the historical usage patterns and future needs of various resources (such as CPU, memory, disk I / O, network bandwidth, etc.) of an application, service, task, or project during its operation, which is digitized, tagged, and modeled.

[0023] Recurrent Neural Networks (RNNs) are a type of neural network used to process sequential data. They possess the ability to "memorize" historical information, and the current output depends not only on the current input but also on past information stored in the hidden state. In resource scheduling, system states (such as load, queues, and utilization) naturally evolve over time. RNNs can model this time-varying dependency, thereby enabling better prediction and decision-making. A typical structure includes input, hidden state, and output, forming a "recurrent" computational path through time unfolding.

[0024] In related technologies, with the popularization of artificial intelligence applications, a single cloud computing model can no longer meet the needs of all scenarios, and edge-cloud collaborative computing has become a new application model. However, due to the heterogeneity of computing architecture (CPU / GPU / NPU), capabilities, and network resources among edge-cloud devices, existing static scheduling strategies are difficult to achieve efficient resource utilization. Moreover, a complete AI application (such as smart power grid inspection) typically involves multiple stages, including data preprocessing, model inference, and model updates, making AI application tasks very complex. The model relying on "cloud training and edge deployment" has high latency, thus failing to meet the needs of current data change scenarios. In addition, most existing workflow engines (such as Airflow) only focus on task dependencies and lack awareness and optimization of underlying heterogeneous resources, failing to achieve efficient cross-level collaboration.

[0025] In view of this, this application provides a method and related equipment for dynamic orchestration of edge-cloud collaborative AI processes based on optimization graphs, which can improve the flexibility of resource scheduling and the efficiency of resource utilization in heterogeneous environments, and reduce the rigidity of AI models.

[0026] The edge-cloud collaborative AI process dynamic orchestration method based on optimized graphs provided in this application relates to the field of artificial intelligence technology. This method can be applied to terminals, servers, or software running on either terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the edge-cloud collaborative AI process dynamic orchestration method based on optimized graphs, but is not limited to the above forms.

[0027] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0028] The embodiments of this application will be described in detail below with reference to the accompanying drawings: Figure 1 This is an optional flowchart of the edge-cloud collaborative AI process dynamic orchestration method based on optimization graph provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S150: Step S110: Obtain the AI ​​workflow to be processed; Step S120: Parse the AI ​​workflow to be processed into an optimization graph, wherein the optimization graph includes multiple nodes and multiple edges. Each node represents an AI task unit, and each AI task unit contains AI task processing logic and resource configuration requirements information corresponding to the AI ​​task processing logic. Each edge carries constraint information for data transmission between two connected nodes. Among the multiple nodes, there is at least one intelligent collaborative node. Step S130: Obtain the real-time resource status of the resource pool composed of heterogeneous resources from the terminal, edge, and cloud. Step S140: Based on the resource configuration requirements of nodes in the optimization graph, the data transmission constraints of edges, and the real-time resource status of the resource pool, generate a target mapping scheme to schedule the nodes of the optimization graph to the target resources in the resource pool using an intelligent decision algorithm. The intelligent decision algorithm includes an online reinforcement learning algorithm and / or an offline metaheuristic optimization algorithm. Step S150: Execute the AI ​​workflow to be processed according to the target mapping scheme; The intelligent collaboration node is configured to perform at least one of the following operations: It receives model optimization seed parameters from the cloud and performs automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment. The mixing ratio of new and old training data is dynamically determined based on a genetic algorithm, and the AI ​​model embedded in the node is incrementally learned and updated.

[0029] It is understood that the nodes in the optimization graph in this embodiment are not only execution units, but also agents carrying multi-dimensional attribute information. This attribute information may include computation type (AI-intensive, memory-intensive), desired execution device (CPU / GPU / NPU), model freshness requirements, and whether it is an optimizable intelligent node (such as an incremental learning node or an AutoML node).

[0030] It is understood that this embodiment will also continuously monitor the resource status in the resource pool to form a global resource view. Specifically, the resource pool includes, but is not limited to, the following resources: (1) Computing resources: CPU: General-purpose computing power, suitable for tasks such as logic control and lightweight computing.

[0031] GPUs are suitable for graphics rendering and massively parallel computing (such as deep learning training / inference).

[0032] NPU: Neural Network Processor, optimized for AI inference or training, with high energy efficiency.

[0033] Other accelerators, such as FPGAs and DSPs, are used for specific computationally intensive tasks.

[0034] (2) Storage resources: Memory: Temporary data storage during task execution.

[0035] Local disk / SSD: Persistent storage on edge devices or servers.

[0036] Distributed storage: such as cloud storage and edge storage clusters, supports large-capacity data access.

[0037] Caching resources, such as Redis and Memcached, are used to improve data access speed.

[0038] (3) Online resources: Bandwidth: The data transmission capability between the device, edge, and cloud.

[0039] Delay: Communication latency between task execution nodes.

[0040] Connection status: Whether the device is online, whether the network is stable, etc.

[0041] Protocol support: Capabilities for communication protocols such as MQTT, HTTP / 2, and gRPC.

[0042] (4) AI-specific resources: Model repository: pre-trained models and model version management.

[0043] AutoML Resources: Computational and storage resources used for model structure search and hyperparameter tuning.

[0044] Incremental learning resources: Data caching and computing resources required for online model updates.

[0045] (5) Equipment status and load information: Device utilization: CPU / GPU / NPU utilization, memory usage.

[0046] Task queue status: The number of tasks currently being executed or waiting.

[0047] Power consumption and temperature: Especially for edge devices, power consumption and heat dissipation capabilities affect scheduling decisions.

[0048] Availability and health status: Whether the device is available and requires maintenance.

[0049] (6) Software and runtime resources: Container runtime: such as Docker, containerd.

[0050] AI frameworks supported: such as TensorFlow, PyTorch, MindSpore, etc.

[0051] Orchestration tools: Status information of underlying scheduling platforms such as Kubernetes and KubeEdge It is understood that, in this embodiment, after obtaining the real-time resource status of the optimization graph and the resource pool composed of heterogeneous resources from the edge, cloud, and endpoints, the intelligent decision-making algorithm generates a target mapping scheme to schedule nodes in the optimization graph to target resources in the resource pool. The resource pool is the collection of all available physical resources and serves as the basis and objective for the decision. The global resource view is a snapshot of the real-time resource status of the resource pool and is the input information for the decision. The dynamic orchestration decision-making process in this embodiment utilizes the intelligent decision-making algorithm to optimally match task requirements (nodes in the optimization graph) with resource supply (global resource view), and its output is a series of specific "mapping" instructions. Therefore, the "most suitable physical resource" "mapped" in this embodiment is a specific entity monitored and managed in the resource pool, such as a specific CPU core, a GPU card, an NPU on an edge server, or a network link. The processing in this embodiment not only considers current resources but also predicts task execution time, network transmission overhead, etc., making decisions with the goal of minimizing total completion time or maximizing throughput, thereby obtaining the target mapping scheme. The intelligent decision-making algorithm can be an online reinforcement learning algorithm or an offline metaheuristic optimization algorithm.

[0052] In this embodiment, the process of generating a target mapping scheme for scheduling optimized graph nodes to target resources in the resource pool using an intelligent decision-making algorithm includes, but is not limited to, the following steps: If the AI ​​workflow to be processed is an online task that arrives in real time, then an online decision-making model based on deep reinforcement learning is used to generate a target mapping scheme. If the AI ​​workflow to be processed contains a batch of tasks formed by caching, a fused recurrent neural network is used to divide the tasks into batches, and a metaheuristic optimization algorithm is used to schedule and optimize the tasks within the batch to generate a target mapping scheme.

[0053] Specifically, such as Figure 2As shown, this embodiment obtains resource configuration requirement information from the optimization graph in the input layer and obtains real-time resource status from the resource pool. The real-time resource status characterizes whether the corresponding resource is occupied, the duration of occupation, and the proportion of occupation. Then, the obtained resource configuration requirement information and real-time resource status are input into the reinforcement learning decision core. In the reinforcement learning decision core, the resource configuration requirement information is encoded into a first state vector, and the real-time resource status is encoded into a second state vector. Based on the first and second state vectors, a deep neural network predicts the action probability distribution. Then, based on the action probability distribution, candidate mapping decisions are determined from nodes in the optimization graph to target physical resources in the global resource view. After simulating the AI ​​workflow to be processed based on the candidate mapping decisions, the target mapping decision is determined from multiple candidate mapping decisions based on the simulation results.

[0054] Specifically, from Figure 2 As can be seen, in this embodiment, after obtaining the target mapping scheme, the target mapping scheme is deployed, and then the actual execution effect corresponding to the target mapping scheme is obtained. The reward value is calculated based on the actual execution effect, and then the network weight of the reinforcement learning algorithm is updated based on the reward value, so that the network weight of the reinforcement learning algorithm is continuously updated, thereby improving the accuracy of the reinforcement learning algorithm.

[0055] It is understood that after obtaining the target mapping scheme, this embodiment can also launch containers on the target node through the process orchestration engine (based on K8s / KubeEdge) to execute the tasks of the AI ​​workflow to be processed, and can also manage the lifecycle of the entire process.

[0056] Understandably, if the AI ​​workflow to be processed contains a cached batch of tasks, a fused recurrent neural network is used to divide the tasks into batches, and a metaheuristic optimization algorithm is combined to optimize the scheduling of tasks within each batch, in order to generate the target mapping scheme. Specifically, such as Figure 3As shown, when batch processing of tasks is required, this embodiment can cache batch tasks to a task set online, divide the tasks within each task set into batches, determine the execution method and order of each task within each batch, and then calculate the comprehensive execution cost of the corresponding batch based on the target mapping scheme, execution method, and execution order. Next, based on the comprehensive execution cost, the particle position representing the smallest resource unit is updated, and a target mapping scheme is generated after the particle position update process meets the convergence condition. Specifically, the execution order of each task within each task set can be analyzed and determined using a recurrent neural network (RNN) time series analysis module. This time series analysis module first analyzes the arrival sequences of all tasks within the task set to determine the time series of each task. Based on the identified time series, all tasks within the task set are then divided into batches to form batch scheduling tasks, improving resource scheduling efficiency.

[0057] It is understandable that, such as Figure 3 As shown, for batch scheduling tasks, this embodiment can determine the optimal scheduling scheme through an improved particle swarm optimization method. Specifically, the improved particle swarm optimization method includes particle encoding, fitness evaluation, particle position update, and global / local optimal update. Particle encoding involves encoding the target mapping scheme containing the task-node mapping relationship, the execution method of each task, and the execution order of each task within each batch. Based on the encoded information, the comprehensive execution cost corresponding to each batch is calculated. Then, the nonlinear inertia weight of each particle (i.e., the smallest resource unit) is adjusted to update the position of each particle, that is, to update the relationship between the smallest resource unit and the task. Then, the global or local particles are updated. When each batch meets the convergence condition, such as the number of updates reaching a threshold or the comprehensive execution cost reaching the optimal level, it is determined that the batch meets the convergence condition, and then the current optimal resource scheduling scheme is taken as the target scheduling scheme.

[0058] It is understandable that, such as Figure 4 As shown, after obtaining the target scheduling scheme, this embodiment performs resource scheduling on each end of the edge-cloud infrastructure based on the target scheduling scheme to execute the tasks of the AI ​​workflow to be processed through the scheduled resources, thereby improving the flexibility of resource scheduling and the efficiency of resource utilization in a heterogeneous environment.

[0059] For example, taking the smart meter reading scenario of the power grid as an example, the method of the embodiment of this application will be specifically described: Step 1: Construct the optimization diagram corresponding to the smart meter reading workflow of the power grid. For example... Figure 5As shown, the optimization graph includes nodes A, B, C, and D. Node A, located at the edge, is used for image acquisition and forward rotation preprocessing. Node B, located on the side, is used for segmenting the preprocessed image from node A into prefix regions; its resource requirement is low-latency CPU. Node C, located in the cloud or on the side, is an intelligent collaborative node. It receives the prefix region image segmented by node B, performs automated machine learning (AutoML) optimization based on a general model seed provided by the cloud model repository, and sends the optimized recognition results to node D. Simultaneously, when the performance of the general model is insufficient, it triggers local AutoML optimization to generate an adapted optimized model. Node D, located on the side, is a data collaborative node used for reporting based on the recognition results and updating the local model. It can also use a genetic algorithm to determine the ratio of new to old samples for incremental learning.

[0060] It is understandable that if the nodes in the optimization graph are intelligent collaborative nodes, such as... Figure 6 As shown, the internal data processing of the intelligent collaborative node includes the following steps: Step 1.1: Load the initial general model; Step 1.2: Execute the model inference process; Step 1.3: Determine if the performance of the current general model meets the standard. If AutoML optimization is triggered, proceed to step 1.6; if it meets the standard, continue to execute the current general model and proceed to step 1.4. Step 1.4: After collecting new data, perform incremental learning closed-loop update on the current general model based on the new data, store the updated model, and then execute step 1.2; Step 1.5: After sending the stored model to the cloud model repository, proceed to step 1.6; Step 1.6: Perform AutoML optimization on the general model; Step 1.7: Deploy the optimization model.

[0061] Step 2: Real-time monitoring and scheduling of resources. For example, when the system detects that the current edge node (Atlas200 DK) has a low load but idle NPU resources, and node C is identified as an intelligent collaboration node, the scheduler decides to schedule nodes B and C to this edge node and inject AutoML seed parameters into node C. The seed parameters are essentially an "optimization starting point" or "configuration blueprint" generated by a pre-performed coarse-grained search in the cloud, which is then distributed to node C on the edge side. This aims to significantly reduce the search space, lower computational overhead, and accelerate the optimization process of the edge model, avoiding the resource-intensive operation of searching for a model from scratch on the edge device. Specifically, the AutoML seed parameters can be summarized as a structured data package, including, but not limited to, the following: 1. Architecture search space: pre-selected backbone network, searchable operation pool, depth / width range.

[0062] 2. Training hyperparameters: configuration of optimizer, learning rate, batch size, etc.

[0063] 3. Hardware constraints: latency, model size, and accuracy requirements.

[0064] 4. Prior knowledge: pre-trained weights, performance predictor.

[0065] 5. Search strategy: Search algorithm configuration and budget evaluation.

[0066] Step 3: Perform process execution and optimization based on scheduled resources. For example, using... Figure 5 Taking the optimized graph shown as an example, nodes B and C can be executed sequentially on the edge nodes. After the initial model inference speed is not ideal, node C automatically starts a local NAS search to quickly train a lightweight model that is more suitable for the Ascend 310 chip, improving the subsequent inference speed by 261%. At the same time, node D, based on the optimal ratio calculated by the genetic algorithm, integrates newly acquired serial number images to incrementally update the model, maintaining the model's high accuracy.

[0067] As can be seen from the above, the method of this application embodiment has the following beneficial effects: (1) This embodiment achieves a paradigm shift from static dependencies to dynamic resource optimization by constructing an optimization graph that integrates multi-dimensional attributes. Traditional workflow orchestration techniques can only describe the static logical dependencies between tasks in their graph models, completely ignoring the dynamic resource environment required for task execution. This embodiment redefines the graph model of the AI ​​process and proposes the concept of an optimization graph. This makes each node in the optimization graph a rich "semantic entity," which not only contains the processing logic of the task, but more importantly, it embeds its resource configuration requirements, including but not limited to task type (such as AI-intensive, memory-intensive), affinity to specific computing units (such as CPU, GPU, NPU), expected computing time and memory usage, etc. At the same time, the edges in the optimization graph are no longer just channels for data flow, but also carry the cost and constraints of data transmission between tasks. This design makes the optimization graph itself evolve from a simple "execution flowchart" into a digital expression of the "optimization problem itself." Compared with traditional methods, this embodiment enables the scheduling system to "understand" the essential requirements of the task, not just its execution order. For example, it can identify a node as having an NPU-intensive task and proactively schedule it to an edge device equipped with the corresponding accelerator, instead of simply balancing the load based on the remaining amount of CPU and memory, as is the case with traditional schedulers.

[0068] (2) In this embodiment, Yongge designed a dynamic orchestration engine based on a hybrid intelligent algorithm, realizing closed-loop optimization from heuristic rules to perception-decision-learning. Traditional schedulers are mostly based on fixed heuristic rules (such as shortest queue first, round-robin scheduling), which appear rigid and inefficient when facing extreme heterogeneity and task burstiness in edge-cloud environments. This embodiment constructs a hybrid intelligent decision engine that integrates reinforcement learning and metaheuristic algorithms. This engine is not a single algorithm, but an adaptive system that can use a deep reinforcement learning model for online tasks that arrive in real time. Through continuous interaction with the environment (trying different scheduling actions, observing task completion, and obtaining rewards), it autonomously learns the optimal scheduling strategy. This strategy can dynamically balance task completion time, waiting time, and overall system load, and its performance continuously surpasses static rules as the learning process progresses. For batch tasks or task sets formed by online task caching, a hybrid scheduler that integrates RNN and particle swarm optimization is enabled. The RNN is responsible for understanding the time-series patterns of task arrivals and intelligently performing batch partitioning, transforming the online problem into a more easily optimized offline problem. Subsequently, the improved particle swarm optimization algorithm innovatively considers the task-node mapping, task execution method, and task execution order simultaneously in its encoding. By adjusting nonlinear inertial weights to balance global search and local convergence capabilities, it seeks the scheduling scheme with the lowest overall cost. This "dual-core" driven hybrid mode achieves a fundamental leap from experience-based, passive rule-based scheduling to data-based, proactive perception-decision-learning.

[0069] (3) This embodiment introduces an endogenous "intelligent collaborative node," which enables the static process to evolve dynamically and continuously optimize. Existing workflow systems treat each processing node as a static, functionally fixed black box. Once deployed, its internal logic (especially the AI ​​model) remains unchanged. This is seriously inconsistent with the dynamic changes in data distribution in edge computing scenarios. Therefore, this embodiment creatively introduces an "intelligent collaborative node" into the optimization graph. This intelligent collaborative node is a "living" intelligent agent with embedded automated machine learning and incremental learning capabilities, enabling a closed loop of model optimization in cloud-edge collaboration. The node receives AutoML "seeds" (containing model architecture and hyperparameters obtained through coarse-grained search) from the cloud and performs fine-grained, resource-aware retraining on the edge using local data to generate an extremely optimized model that is fully adapted to the local environment, thereby solving the adaptation problem between the model and hardware and data at the source. In the data-driven model iteration closed loop, the intelligent collaborative node uses a built-in incremental learning mechanism combined with a genetic algorithm to dynamically analyze the value of new and old samples, scientifically determine the optimal training ratio, and thus achieve rapid adaptation of the model to new data patterns with the fewest iterations and computational overhead, effectively resisting model performance degradation.

[0070] Please see Figure 7 This application also provides a device for dynamic orchestration of edge-cloud collaborative AI processes based on an optimized graph. The device includes: The first module is used to acquire the AI ​​workflow to be processed; The second module is used to parse the AI ​​workflow to be processed into an optimization graph. The optimization graph includes multiple nodes and multiple edges. Each node represents an AI task unit, and each AI task unit contains AI task processing logic and corresponding resource configuration requirements. Each edge carries constraint information for data transmission between two connected nodes. Among the multiple nodes, there is at least one intelligent collaborative node, which has an embedded intelligent agent with machine learning and incremental learning capabilities. The third module is used to build a global resource view based on the resource status within the resource pool under real-time monitoring; The fourth module is used to map nodes in the optimization graph to target physical resources in the global resource view based on reinforcement learning algorithms, forming a target mapping scheme. The fifth module is used to schedule resources for the AI ​​workflow to be processed according to the target mapping scheme, so as to execute the tasks of the AI ​​workflow to be processed through the scheduled resources.

[0071] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0072] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0073] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0074] Please see Figure 8 , Figure 8 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 810 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 820 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 820 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 820 and is called and executed by the processor 810 using the methods described in the embodiments of this application. The input / output interface 830 is used to implement information input and output; The communication interface 840 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 850 transmits information between various components of the device (e.g., processor 810, memory 820, input / output interface 830, and communication interface 840); The processor 810, memory 820, input / output interface 830 and communication interface 840 are connected to each other within the device via bus 850.

[0075] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0076] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0077] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0078] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0079] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0080] This application provides a method and related equipment for dynamic orchestration of edge-cloud collaborative AI workflows based on optimization graphs. By parsing the AI ​​workflow to be processed into an optimization graph, and using the node table in the optimization graph to carry the resource configuration requirements of the corresponding AI task units, and the edges in the optimization graph to carry the data transmission constraints between connected nodes, a paradigm shift from static dependency to dynamic resource optimization can be achieved. Then, combined with the real-time resource status of a resource pool composed of heterogeneous resources from the edge, cloud, and endpoints, an intelligent decision-making algorithm, including online reinforcement learning algorithms and / or offline metaheuristic optimization algorithms, is used to generate a target mapping scheme to schedule optimization graph nodes to target resources in the resource pool. This enables a dynamic orchestration process from initialization to dynamic resource optimization. The closed-loop optimization from pattern rules to perception-decision-learning, and then the execution of the AI ​​workflow to be processed according to the target mapping scheme, can effectively improve the flexibility of resource scheduling and the efficiency of resource utilization in heterogeneous environments. In addition, this embodiment can also set at least one intelligent collaborative node in the optimization graph. The intelligent collaborative node is configured to perform at least one of the following operations: receive model optimization seed parameters from the cloud and perform automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment; or dynamically determine the mixing ratio of new and old training data based on genetic algorithms, and incrementally learn and update the AI ​​model embedded in the node, so that the process has the ability to dynamically evolve and continuously optimize.

[0081] In summary, this application addresses the core issues of low resource utilization efficiency, rigid scheduling, and poor model adaptability in heterogeneous AI process orchestration by defining an optimization graph that carries resource and constraint information, constructing a hybrid intelligent decision engine that adapts to task types, and introducing intelligent collaborative nodes with embedded automated learning capabilities.

[0082] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0083] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0084] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0086] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0087] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0089] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for dynamic orchestration of edge-cloud collaborative AI processes based on optimization graphs, characterized in that, The method includes the following steps: Obtain the AI ​​workflow to be processed; The AI ​​workflow to be processed is parsed into an optimization graph, which includes multiple nodes and multiple edges. Each node carries the resource configuration requirements of the corresponding AI task unit; each edge carries the constraint information of data transmission between two connected nodes; and the multiple nodes include at least one intelligent collaborative node. Obtain the real-time resource status of the resource pool, which consists of heterogeneous resources from the terminal, edge, and cloud. Based on the resource configuration requirements of the nodes in the optimization graph, the data transmission constraints of the edges, and the real-time resource status of the resource pool, an intelligent decision-making algorithm is used to generate a target mapping scheme that schedules the nodes in the optimization graph to the target resources in the resource pool. The intelligent decision-making algorithm includes an online reinforcement learning algorithm and / or an offline metaheuristic optimization algorithm. The AI ​​workflow to be processed is executed according to the target mapping scheme; The intelligent collaborative node is configured to perform at least one of the following operations: It receives model optimization seed parameters from the cloud and performs automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment. The mixing ratio of new and old training data is dynamically determined based on a genetic algorithm, and the AI ​​model embedded in the node is incrementally learned and updated.

2. The method according to claim 1, characterized in that, The resource pool includes computing resources, storage resources, network resources, AI-specific resources, device status and load information, and software and runtime resources.

3. The method according to claim 1, characterized in that, The step of generating a target mapping scheme by using an intelligent decision-making algorithm to schedule the optimized graph nodes to target resources in the resource pool includes: If the AI ​​workflow to be processed is an online task that arrives in real time, then the target mapping scheme is generated using an online decision model based on deep reinforcement learning. If the AI ​​workflow to be processed contains a batch of tasks formed by caching, a fused recurrent neural network is used to divide the tasks into batches, and a metaheuristic optimization algorithm is used to schedule and optimize the tasks within the batches to generate the target mapping scheme.

4. The method according to claim 3, characterized in that, The step of generating the target mapping scheme using an online decision model based on deep reinforcement learning includes: The resource allocation requirement information is encoded into a first state vector; The real-time resource status is encoded into a second state vector; Based on the first state vector and the second state vector, a deep neural network is used to predict the probability distribution of actions. Based on the action probability distribution, determine the candidate mapping decisions from the nodes in the optimization graph to the target physical resources in the global resource view; The simulation operation of the AI ​​workflow to be processed is performed based on the candidate mapping decision; Based on the simulation results, a target mapping decision is determined from among the multiple candidate mapping decisions as the target mapping scheme.

5. The method according to claim 4, characterized in that, After determining the target mapping decision as the target mapping scheme from a plurality of candidate mapping decisions based on the simulation operation results, the method further includes the following steps: Obtain the actual execution effect corresponding to the target mapping scheme; Calculate the reward value based on the actual performance results; The network weights of the reinforcement learning algorithm are updated based on the reward value.

6. The method according to claim 3, characterized in that, The method of using a fused recurrent neural network for task batch partitioning includes: The tasks corresponding to all the AI ​​workflows to be processed are cached to form a task set; The time-series patterns of all task arrivals within the task set are obtained by fusing recurrent neural networks. All tasks are batched based on the time series pattern.

7. The method according to claim 6, characterized in that, The step of combining metaheuristic optimization algorithms to optimize the scheduling of tasks within a batch to generate the target mapping scheme includes: Determine the execution method and execution order for each task within each batch; Calculate the overall execution cost of the corresponding batch based on the target mapping scheme, the execution method, and the execution order; The particle position is updated based on the comprehensive execution cost, whereby the particle represents the smallest resource unit. Once the particle position update process is determined to meet the convergence condition, the target mapping scheme is generated.

8. A dynamic orchestration device for edge-cloud collaborative AI processes based on an optimization graph, characterized in that, The device includes: The first module is used to acquire the AI ​​workflow to be processed; The second module is used to parse the AI ​​workflow to be processed into an optimization graph, which includes multiple nodes and multiple edges. Each node carries the resource configuration requirements of the corresponding AI task unit; each edge carries the constraint information of data transmission between two connected nodes; and the multiple nodes include at least one intelligent collaborative node. The third module is used to obtain the real-time resource status of the resource pool, which consists of heterogeneous resources from the terminal, edge, and cloud. The fourth module is used to generate a target mapping scheme for scheduling the nodes of the optimization graph to target resources in the resource pool based on the resource configuration requirement information of the nodes in the optimization graph, the data transmission constraint information of the edges, and the real-time resource status of the resource pool, using an intelligent decision algorithm. The intelligent decision algorithm includes an online reinforcement learning algorithm and / or an offline metaheuristic optimization algorithm. The fifth module is used to execute the AI ​​workflow to be processed according to the target mapping scheme; The intelligent collaborative node is configured to perform at least one of the following operations: It receives model optimization seed parameters from the cloud and performs automated machine learning search based on local data and resource conditions to generate an optimized model adapted to the local environment. The mixing ratio of new and old training data is dynamically determined based on a genetic algorithm, and the AI ​​model embedded in the node is incrementally learned and updated.

9. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images