Artificial intelligence edge cloud cooperative training method and related device for electric power full scene

Abstract

The invention belongs to the technical field of electric power artificial intelligence application, and discloses an artificial intelligence edge cloud cooperative training method and a related device for an electric power full scene. The electric power full scene-oriented artificial intelligence edge cloud cooperative training method comprises the following steps of: training a cloud sub-model by utilizing a cloud sample to obtain a trained cloud sub-model; based on the trained cloud sub-model, obtaining a lightweight cloud student model through knowledge distillation training; and training the professional application sub-model by using the local sample at the side end, calculating a training loss value and a gradient according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model, and obtaining the trained professional application sub-model. According to the technical scheme disclosed by the invention, the generalization and robustness of the model in diversified scenes can be improved, and the resource utilization efficiency can be optimized.

Description

Technical Field

[0001] The present invention belongs to the field of electric power artificial intelligence application technology, and in particular relates to an artificial intelligence edge-cloud collaborative training method and related devices for the entire electric power scenario. Background Art

[0002] With the increasing complexity of the power system and the acceleration of digital transformation, the application of artificial intelligence technology in the power industry is becoming more and more extensive and in-depth. Intelligent management of the entire process from power generation, transmission, distribution to power consumption has become a key means to improve the efficiency of the power system, ensure the stability of power supply and enhance safety.

[0003] Due to the particularity of the power industry, power scenarios often have the characteristics of high real-time, complex and diverse scenarios, huge data volume, and strict security and privacy protection requirements. These characteristics have brought new challenges to the application of artificial intelligence technology in the entire power scenario. The existing traditional centralized computing methods can no longer meet the diverse needs of power scenarios. Specifically, many application scenarios in the operation of the power system (for example, power transmission and distribution equipment inspection, load forecasting, smart grid scheduling, equipment fault diagnosis, and safe production monitoring) often require real-time, efficient, and low-latency intelligent decision support. Although the existing artificial intelligence models based on cloud-based centralized computing have powerful computing power and global optimization capabilities, due to the wide distribution of power systems, large differences in network conditions, and limitations in data transmission bandwidth and latency, it is difficult to meet the requirements of real-time and security by relying solely on cloud-based centralized computing.

[0004] During the operation of the power system, the introduction of edge computing can solve the above-mentioned problems of relying solely on centralized cloud computing. Edge computing can effectively reduce data transmission delays and improve the response speed of the system by sinking computing resources to edge devices close to the data generation end. However, the computing and storage resources of edge devices are limited, and it is difficult to carry complex large-scale deep learning models. This requires the application of cloud-edge collaborative technology, that is, allocating complex model computing tasks to the cloud, and allocating lightweight tasks with high real-time requirements to edge devices for processing, so as to achieve collaborative work between the cloud and the edge.

[0005] In the above context, how to modularly design and decompose large-scale cloud models so that they can adapt to the needs of diversified power scenarios, and realize the efficient deployment and operation of various intelligent applications in the power industry through edge-cloud collaborative architecture has become a key challenge in power artificial intelligence applications. Summary of the invention

[0006] The purpose of the present invention is to provide an artificial intelligence edge-cloud collaborative training method and related devices for the entire power scenario to solve one or more of the above-mentioned technical problems. The artificial intelligence edge-cloud collaborative mechanism for the entire power scenario provided by the present invention can not only improve the generalization and robustness of the model in diversified scenarios, but also optimize resource utilization efficiency, and provide strong technical support for the intelligent transformation of the power industry.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides an artificial intelligence edge-cloud collaborative training method for the entire power scenario, comprising the following steps:

[0009] The power cloud model is split into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output;

[0010] The cloud sub-model is deployed on the cloud, and the cloud sub-model is trained using cloud samples to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, a lightweight cloud student model is obtained through knowledge distillation training; the cloud samples are stored on the edge to obtain cloud local samples; the lightweight cloud student model and the professional application sub-model are deployed on the edge, and the professional application sub-model is trained based on cloud local samples and edge local samples to obtain a trained professional application sub-model; wherein, when the professional application sub-model is trained based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model and the lightweight cloud student model share a low-dimensional feature encoder.

[0011] A further improvement of the present invention is that

[0012] The step of splitting the power cloud large model into a cloud sub-model and a professional application sub-model comprises:

[0013] The basic model part of the power cloud big model is used as a cloud sub-model, and the remaining part of the power cloud big model after splitting the basic model part is used as a professional application sub-model;

[0014] Among them, the sum of the parameter ratio of the cloud sub-model to the power cloud large model and the parameter ratio of the professional application sub-model to the power cloud large model is 1; the low-dimensional feature encoder of the power cloud large model is shared at the bottom layer of the cloud sub-model and the professional application sub-model, and the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model and the network layer adapted to different downstream tasks are completely separated; the network architecture of the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model follows the same backbone network structure as the power cloud large model, and a different number of filters are used in each network layer to amplify and reduce the amount of parameters.

[0015] A further improvement of the present invention is that

[0016] In the step of obtaining a lightweight cloud-based student model through knowledge distillation training based on the obtained trained cloud-based sub-model,

[0017] The lightweight cloud-based student model is isomorphic to the cloud-based sub-model, shares the underlying low-dimensional feature encoder, and uses the same backbone network as the cloud-based sub-model;

[0018] The lightweight cloud-based student model is provided with a certain number of filters to reduce the number of model parameters and is suitable for edge-side deployment;

[0019] The loss function during knowledge distillation training is set to:

[0020]

[0021] Where, L KD is the loss function in the knowledge distillation training process; D c is a cloud sample dataset; (x i ,y i ) is the i-th sample in the cloud sample data set; l kd is the mean square error distance metric; They are the outputs of the cloud-based sub-model and the lightweight cloud-based student model, respectively.

[0022] A further improvement of the present invention is that

[0023] In the step of deploying the lightweight cloud-based student model and the professional application sub-model on the edge side, training the professional application sub-model based on the cloud-based local samples and the edge-side local samples, and obtaining the trained professional application sub-model,

[0024] Setting the weights of the lightweight cloud-based student model to be frozen and setting the shared low-dimensional feature encoder to be unchangeable;

[0025] The calculation expression for calculating the training loss value is:

[0026]

[0027] Where L is the training loss calculation value; (x j ,y j ) is a sample of the side data batch B; l CE is the cross entropy loss function; Output of the sub-model for side-by-side professional applications; Output of the lightweight cloud-based student model.

[0028] A further improvement of the present invention is that

[0029] The method further includes: deploying the lightweight cloud student model and the professional application sub-model on the edge end, training the professional application sub-model based on the cloud local sample and the edge end local sample, and obtaining the trained professional application sub-model;

[0030] Based on the trained professional application sub-model, the cloud is used as a parameter server to aggregate and process the model parameters of the professional application sub-models updated from different sides.

[0031] The second aspect of the present invention provides an artificial intelligence edge-cloud collaborative training system for all power scenarios, including:

[0032] A model splitting module is used to split the power cloud large model into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output;

[0033] A training module is used to deploy the cloud sub-model on the cloud, and train the cloud sub-model with cloud samples to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, a lightweight cloud student model is obtained through knowledge distillation training; the cloud samples are stored on the edge to obtain cloud local samples; the lightweight cloud student model and the professional application sub-model are deployed on the edge, and the professional application sub-model is trained based on cloud local samples and edge local samples to obtain the trained professional application sub-model; wherein, when the professional application sub-model is trained based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model, and the lightweight cloud student model share a low-dimensional feature encoder.

[0034] A further improvement of the present invention is that

[0035] In the model splitting module, the step of splitting the power cloud large model into a cloud sub-model and a professional application sub-model includes:

[0036] The basic model part of the power cloud big model is used as a cloud sub-model, and the remaining part of the power cloud big model after splitting the basic model part is used as a professional application sub-model;

[0037] Among them, the sum of the parameter ratio of the cloud sub-model to the power cloud large model and the parameter ratio of the professional application sub-model to the power cloud large model is 1; the low-dimensional feature encoder of the power cloud large model is shared at the bottom layer of the cloud sub-model and the professional application sub-model, and the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model and the network layer adapted to different downstream tasks are completely separated; the network architecture of the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model follows the same backbone network structure as the power cloud large model, and a different number of filters are used in each network layer to amplify and reduce the amount of parameters.

[0038] A further improvement of the present invention is that

[0039] In the training module, the step of obtaining a lightweight cloud student model through knowledge distillation training based on the obtained trained cloud sub-model is performed.

[0040] The lightweight cloud-based student model is isomorphic to the cloud-based sub-model, shares the underlying low-dimensional feature encoder, and uses the same backbone network as the cloud-based sub-model;

[0041] The lightweight cloud-based student model is provided with a certain number of filters to reduce the number of model parameters and is suitable for edge-side deployment;

[0042] The loss function during knowledge distillation training is set to:

[0043]

[0044] Where, L KD is the loss function in the knowledge distillation training process; D c is a cloud sample dataset; (x i ,y i ) is the i-th sample in the cloud sample data set; l kd is the mean square error distance metric; They are the outputs of the cloud-based sub-model and the lightweight cloud-based student model, respectively.

[0045] A further improvement of the present invention is that

[0046] In the training module, a step is executed in which a lightweight cloud student model and a professional application sub-model are deployed on the edge side, and the professional application sub-model is trained based on the cloud local sample and the edge side local sample to obtain the trained professional application sub-model.

[0047] Setting the weights of the lightweight cloud-based student model to be frozen and setting the shared low-dimensional feature encoder to be unchangeable;

[0048] The calculation expression for calculating the training loss value is:

[0049]

[0050] Where L is the training loss calculation value; (x j ,y j ) is a sample of the side data batch B; l CE is the cross entropy loss function; Output of the sub-model for side-by-side professional applications; Output of the lightweight cloud-based student model.

[0051] A further improvement of the present invention is that it also includes:

[0052] The aggregation update module is used to aggregate the model parameters of professional application sub-models updated from different sides based on the trained professional application sub-models, using the cloud as a parameter server.

[0053] A third aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, an artificial intelligence edge-cloud collaborative training method for the entire power scenario as described in any one of the first aspect of the present invention is implemented.

[0054] A fourth aspect of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the artificial intelligence edge-cloud collaborative training method for the entire power scenario as described in any one of the first aspects of the present invention is implemented.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] In the technical solution disclosed in the present invention, by splitting the large power cloud model into cloud sub-models and professional application sub-models, a modular design of the model is achieved, which is convenient for flexible deployment and updating in different scenarios. The cloud sub-model and the professional application sub-model can be trained using the resources of the cloud and the edge respectively, giving full play to the advantages of cloud computing and edge computing, and improving training efficiency and resource utilization. The lightweight cloud student model is obtained through knowledge distillation training. This process enables the lightweight cloud student model to learn the "knowledge" of the cloud sub-model while maintaining a smaller model size and a faster reasoning speed. This technology reduces the computing and storage requirements of the model when it is deployed on the edge, making the edge more user-friendly. The device can run artificial intelligence applications more efficiently; the cloud sub-model, professional application sub-model and lightweight cloud student model share a low-dimensional feature encoder. This design promotes information sharing and collaborative work between models. The shared encoder helps to extract common feature representations and improves the generalization and adaptability of the model in different scenarios. During the training process, the training loss value and gradient are calculated based on the sum of the output of the lightweight cloud student model and the output of the professional application sub-model. This method combines information from the cloud and the edge, making the training process more comprehensive and accurate. In this way, it can ensure that the performance of the model on the edge is consistent with that on the cloud, thereby improving the stability and reliability of the model.

[0057] To sum up, the artificial intelligence edge-cloud collaborative solution for the entire power scenario disclosed in the present invention has achieved efficient, stable and reliable artificial intelligence application deployment and update through designs such as model splitting and collaborative training, application of knowledge distillation technology, shared low-dimensional feature encoders, and flexible training loss values ​​and gradient calculation methods, providing strong support for the intelligent development of the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below; obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0059] Figure 1 It is a flow chart of an artificial intelligence edge-cloud collaborative training method for the entire power scenario in an embodiment of the present invention.

[0060] Figure 2 It is a schematic diagram of a modular decomposition process in an embodiment of the present invention.

[0061] Figure 3 It is a flowchart of the collaboration between the professional application sub-model and the lightweight cloud-based student model in an embodiment of the present invention.

[0062] Figure 4 It is a flow chart of edge-cloud collaboration in an embodiment of the present invention.

[0063] Figure 5 This is a schematic diagram of an artificial intelligence edge-cloud collaborative training system for the entire power scenario in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention; it is obvious that the described embodiments and technical solutions are only part of the embodiments of the present invention, not all of the embodiments.

[0065] All other embodiments obtained by those of ordinary skill in the art without creative work based on the technical solutions disclosed in the embodiments of the present invention belong to the scope of protection of the present invention. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device including a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0066] See also Figure 1 , an embodiment of the present invention provides an artificial intelligence edge-cloud collaborative training method for the entire power scenario, comprising the following steps:

[0067] Step 1: split the power cloud model into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output;

[0068] Step 2, deploy the cloud sub-model on the cloud, and use cloud samples to train the cloud sub-model to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, obtain a lightweight cloud student model through knowledge distillation training; store the cloud samples on the edge to obtain cloud local samples; deploy the lightweight cloud student model and the professional application sub-model on the edge, and train the professional application sub-model based on cloud local samples and edge local samples to obtain the trained professional application sub-model; wherein, when training the professional application sub-model based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model, and the lightweight cloud student model share a low-dimensional feature encoder.

[0069] In the method for collaborative training of artificial intelligence edge-cloud for the entire power scene disclosed in the embodiment of the present invention, a novel modular decomposition method of a large model on the power cloud is adopted to split the original large power model into two types of sub-models (i.e., a large model for learning on the cloud side and a professional application sub-model on the edge side), realizing modular design, enabling the model to be flexibly combined, and selecting appropriate professional application sub-models from the modules according to specific task requirements, so as to realize efficient and lightweight deployment of small models on the edge side. In the technical solution of the embodiment of the present invention, the edge device and the cloud side each use local and global data for model learning, and through the knowledge fusion mechanism, the edge and the cloud side jointly participate in model optimization, realizing edge-cloud knowledge fusion, and improving the adaptability and accuracy of the overall intelligent system. In the technical solution of the embodiment of the present invention, the distributed computing method is used to train local samples on the edge side, which reduces the dependence on the cloud side, realizes efficient edge-cloud collaboration through data parallelism, reduces delays, and saves bandwidth. In addition, in the preferred technical solution, the cloud side acts as a parameter server, regularly aggregates model parameters updated from different edge ends, realizes continuous synchronization and optimization of the model, and ensures the collaborative evolution of the edge and cloud intelligent systems.

[0070] See also Figure 2 In one embodiment of the present invention, in the step of splitting the power cloud model into a cloud sub-model and a professional application sub-model,

[0071] Based on the existing original power cloud model architecture design, the model modular decomposition strategy is used to split it, and the basic model is split into cloud sub-models, such as the power visual model and the power semantic model. The remaining parts after splitting the basic model are used as professional application sub-models.

[0072] Furthermore, large-scale cloud samples are used to train and fine-tune the cloud sub-model, and the professional application sub-model learns from the edge sample data, and then the new edge knowledge is integrated with the cloud knowledge.

[0073] Explanatoryally, based on the model modular decomposition strategy disclosed above in the embodiment of the present invention, the artificial intelligence edge-cloud collaborative learning task can be decomposed into two subtasks, one subtask is to train the cloud sub-model on the cloud samples, and the other subtask is for the edge and cloud to use their respective samples to collaboratively train the joint sub-model in a data parallel manner.

[0074] In an embodiment of the present invention, in order to further decouple the two sub-models and reduce the dependence of the end-cloud collaborative sub-model training process (for example, forward propagation and back propagation, etc.) on the cloud sub-model, a vertical decomposition method is adopted to cut off the connections such as neurons between the two sub-models, and the final output adopts the sum of the outputs of the two sub-models; wherein, in the process of model reasoning, for specific tasks in different power scenarios, the edge-cloud collaborative system respectively calls the corresponding cloud sub-model and professional application sub-model, and completes the specified tasks through integrated learning and comprehensive judgment.

[0075] See also Figure 3 In one embodiment of the present invention, the cloud sub-model, the lightweight cloud student model and the professional application sub-model share a low-dimensional feature encoder; wherein, the cloud first uses cloud samples to train the shared low-dimensional feature encoder and the cloud sub-model, and then trains the lightweight cloud student model through knowledge distillation to imitate the cloud sub-model, and the professional application sub-model and the lightweight cloud student model collaborate to realize the reasoning and training process at the edge.

[0076] See also Figure 4 In one embodiment of the present invention, the professional application sub-model is trained using the local samples on the edge end, and the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model to obtain the trained professional application sub-model.

[0077] In a data-parallel manner, professional application sub-models are collaboratively trained under the correction of lightweight cloud-based student models. After multiple rounds of local iterative training, the cloud acts as a parameter server to aggregate model parameters updated from different edges to achieve collaborative evolution of the cloud and edge.

[0078] A specific embodiment of the present invention provides an artificial intelligence edge-cloud collaborative training method for the entire power scenario, comprising the following steps:

[0079] Step 1: split the power cloud model into a cloud sub-model and a professional application sub-model;

[0080] Step 2: deploy the cloud sub-model on the cloud, train the cloud sub-model using cloud samples, and obtain a trained cloud sub-model and a low-dimensional feature encoder;

[0081] Step 3: Based on the trained cloud sub-model obtained in step 2, a lightweight cloud student model is obtained through knowledge distillation training;

[0082] Step 4: deploy the professional application sub-model on the edge end, train the professional application sub-model using the local sample on the edge end, calculate the training loss value and gradient according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model, and obtain the trained professional application sub-model;

[0083] Among them, the cloud sub-model, the professional application sub-model, and the lightweight cloud-based student model share a low-dimensional feature encoder.

[0084] In the technical solution disclosed in the embodiment of the present invention, when the edge optimizes the professional application sub-model on its local sample, the training loss value and its gradient should be calculated based on the sum of the cloud sub-model output and the professional application sub-model output; however, the cloud sub-model parameter scale is large and cannot be fully applied to the edge device, and its output cannot be obtained in real time; therefore, the embodiment of the present invention designs a lightweight cloud student model to simulate the output of the cloud sub-model through knowledge distillation. The model deployed on the edge includes a lightweight cloud student model and a professional application sub-model, which jointly approximate the gradient required for local optimization.

[0085] Explanatory, in the technical solution of the embodiment of the present invention, the large power cloud model is split into two sub-models with the same backbone network, by cutting off the neuron connection between the two sub-models and introducing a lightweight cloud student model (used to simulate the cloud sub-model deployed on the edge end). In the present invention, by isolating the two sub-models, the system can support the cloud sub-model and the professional application sub-model to adopt different backbone network structures, and select different backbone networks to build different types of sub-models according to business needs, so as to adapt to different power business scenarios.

[0086] In the specific exemplary technical solution, the power cloud model is a power vision model, which usually includes a visual encoder and a text encoder. The low-dimensional feature encoder of the power cloud model is shared at the bottom layer of the cloud sub-model and the professional application sub-model to extract low-dimensional features of images or texts. At the same time, the high-dimensional feature encoders of the two sub-models and the network layers adapted to different downstream tasks are completely separated. Specifically, the two sub-models use the same low-dimensional features as input, but the feature mapping of the two sub-models and the forward reasoning and reverse transfer processes are functionally decoupled, and the final result is obtained by integrating the comprehensive output of the two.

[0087] In the embodiment of the present invention, the network architecture of the high-dimensional encoders of the two sub-models can follow the same backbone network structure as the large power cloud model, and different numbers of filters are used in each network layer to amplify and reduce the amount of parameters. Furthermore, when the sum of the ratios of the parameters of the two sub-models to the parameters of the cloud basic model is 1, it means that the cloud basic model is vertically split and the connection of the neural network is cut off, completing the splitting and decomposition of the edge cloud model.

[0088] The embodiment of the present invention proposes an artificial intelligence edge-cloud collaborative training method for the entire power scenario. It is based on the architectural design of the cloud-based large model and proposes a method for modular decomposition of the model. The original large model of the power industry is split into two types of sub-models. One type of sub-model, such as the large power vision model, learns from cloud samples, and the other type of professional application sub-model is used to learn from edge samples and perform edge-cloud knowledge fusion. Professional application modules are selected according to specific task requirements, and models adapted to different power scenarios can be generated, thereby realizing lightweight deployment of small models on the edge. Furthermore, a data parallel architecture is adopted, and each sub-model is trained at the edge using local sample data based on distributed computing. After multiple rounds of local iterations, the cloud acts as a parameter server to aggregate updated model parameters from different edge devices to achieve collaborative evolution of the cloud and the edge. In summary, the technical solution of the embodiment of the present invention is aimed at application scenarios in the power field such as defect detection of power transmission and distribution drone inspection equipment, intelligent customer service, and identification of operational safety violations. It provides an intelligent collaboration and scheduling mechanism from large cloud models to lightweight edge models, giving full play to the comprehensive advantages of cloud-edge collaboration. It not only improves the adaptability and flexibility of the model, but also optimizes resource utilization efficiency, and effectively promotes the engineering implementation and application promotion of power artificial intelligence technology, which can provide strong technical support for the intelligent transformation of the power industry.

[0089] In a specific embodiment of the present invention, in the step of obtaining a lightweight cloud student model through knowledge distillation training based on the trained cloud sub-model obtained in step 2,

[0090] The cloud uses knowledge distillation to train a lightweight cloud student model to imitate the cloud sub-model. The lightweight cloud student model is isomorphic to the cloud sub-model, sharing the underlying low-dimensional feature encoder and using the same backbone network as the cloud sub-model, but with a small number of filters to reduce the number of model parameters for edge deployment. The loss function L in the knowledge distillation process is KD Set to:

[0091]

[0092] Where D c is the cloud sample dataset; l kd is the mean square error distance metric; are the outputs of the cloud sub-model and the lightweight cloud student model respectively; (x i ,y i ) represents the i-th sample in the cloud sample data set;

[0093] After iterative knowledge distillation, the output of the lightweight cloud-based student model can approach the output of the cloud-based sub-model.

[0094] In a specific embodiment of the present invention, a professional application sub-model is deployed on an edge end, the professional application sub-model is trained using a local sample on the edge end, and a training loss value and a gradient are calculated based on the sum of the output of the lightweight cloud student model and the output of the professional application sub-model to obtain the trained professional application sub-model.

[0095] The centralized training of the two types of sub-models is implemented in a distributed manner. Based on the data parallel framework, the edge and cloud save their respective data samples locally and train the joint sub-model on their respective samples. During the local training process, in addition to freezing the weight setting of the lightweight cloud student model, the shared low-dimensional feature encoder of the lightweight cloud student model should also be set to be unchangeable. Explanatory, the above technical means are to ensure that the lightweight cloud student model can accurately imitate the behavior of the cloud sub-model, thereby maintaining the consistency and effectiveness of the model performance. Through this strategy, it can be ensured that the local model will not deviate from the expected behavior of the cloud sub-model during training, thereby ensuring the stability and reliability of the model in different deployment environments. In addition, the training loss value of the edge is obtained based on the sum of the logical values ​​(Logits) of the lightweight cloud student model and the sub-model, specifically:

[0096]

[0097] In the formula, (x i ,y i ) is a sample of the side data batch B; l CE is the cross entropy loss function, Output of the sub-model for side-by-side professional applications.

[0098] In an embodiment of the present invention, after multiple local training iterations, the cloud acts as a parameter server, aggregates its updated sub-model with the updated sub-model on the edge, and generates a new joint sub-model as the starting point for the next round of edge-cloud collaboration training. Explanatory, when there are significant differences in the distribution of samples on the edge and cloud, and when observed from a global perspective, they present long-tail distribution characteristics, local updates of the edge-cloud collaborative sub-model may produce deviations, which may weaken the overall performance of collaborative training; therefore, after executing the standard edge-cloud collaborative training process, the classifier of the joint sub-model is fine-tuned on the global dataset according to actual needs. This process mainly focuses on optimizing the output performance of the joint sub-model at the high-dimensional encoder level for samples on the edge and cloud. Through this targeted fine-tuning strategy, the performance degradation problem caused by differences in sample distribution can be effectively alleviated, and the generalization and robustness of the model in diverse scenarios can be further improved.

[0099] In the technical solution disclosed in the embodiment of the present invention, the limitations of the existing edge-cloud collaborative application of power artificial intelligence are overcome through the edge-cloud collaborative architecture design; among them, the present invention adopts a modular decomposition method of the cloud-side large model to split the large model into two types of sub-models: one type learns global power data samples through the cloud, and the other type learns local professional data through the edge to achieve edge-cloud knowledge fusion. This design can flexibly combine professional modules to generate adaptive lightweight small models for specific power tasks, thereby achieving efficient deployment on the edge side; utilizing distributed computing and data parallel architecture, local iterative training on the edge side, and synchronous optimization of the model through cloud parameter aggregation, so that the cloud and edge devices can evolve collaboratively in real time; in this way, the present invention reduces data transmission delay, saves bandwidth and ensures data privacy, while improving the system response speed and processing performance, and is particularly suitable for diversified power application scenarios such as power transmission and distribution inspection, intelligent customer service and operation safety monitoring, thereby effectively promoting the engineering application of power artificial intelligence technology.

[0100] Specific exemplary technical solutions of the embodiments of the present invention include:

[0101] 1) Initialization phase, including: before training starts, the cloud server initializes a global model and distributes it to all participating edge devices. This global model can serve as the starting point for local training of all edge devices;

[0102] 2) Local training iterations, including: After each edge device receives the global model, it uses its local data to train the model. This step usually involves multiple iterations, that is, multiple forward propagations and backpropagations, to optimize the model parameters. When training locally, the edge device can flexibly adjust the training parameters, such as learning rate, batch size, etc., according to the capabilities of the device and the characteristics of the data;

[0103] 3) Sub-model update upload, including: after local training is completed, the edge device uploads the updated sub-model (or the updated amount of model parameters) to the cloud server;

[0104] 4) Cloud aggregation, including: After the cloud server receives the sub-models uploaded by all edge devices, it will aggregate these models. The aggregation method is usually to average the parameters of each sub-model (or other forms of weighted average) to generate a new global model; during the aggregation process, the cloud can also further optimize the model as needed, such as regularization, pruning, etc., to improve the generalization ability of the model and reduce the amount of calculation.

[0105] 5) New joint sub-model distribution, including: the cloud distributes the aggregated new global model to all participating edge devices as the starting point for the next round of local training;

[0106] The above process will be repeated many times until the model reaches the predetermined training target (such as accuracy, loss function value, etc.) or reaches the preset number of training rounds. Through this edge-cloud collaborative learning method, the local data and computing power of the edge device can be effectively utilized, while the powerful processing power of the cloud can be used to achieve more efficient and accurate model training; in addition, this method can also improve data privacy, because sensitive data can be processed locally without uploading to the cloud.

[0107] At present, the existing technology faces technical problems such as diversified sample characteristics, large differences in scenario requirements, and insufficient adaptability of algorithms to actual applications in the application of power artificial intelligence, which makes it difficult to directly apply a unified large model to a variety of power scenarios, restricting the effect of engineering applications. Especially in scenarios with high real-time requirements and sensitive data privacy, the cloud-based centralized processing model not only has a large delay, but also has bandwidth and data security issues. It is impossible to fully combine the real-time processing capabilities of edge computing and the global optimization advantages of the cloud, resulting in low efficiency, untimely response, and poor scalability of intelligent applications in power scenarios. In view of the above existing problems, the embodiment of the present invention discloses an artificial intelligence edge-cloud collaboration solution for all power scenarios, which realizes the collaborative evolution of large cloud models and lightweight edge models. In scenarios such as power transmission and distribution drone inspections, intelligent customer service, and operational safety behavior monitoring, it can not only improve the real-time response and processing efficiency of the system, but also reduce the pressure of data transmission and the risk of privacy leakage. It has important engineering application value and broad application prospects.

[0108] In the technical solution disclosed in the embodiment of the present invention, the large model on the cloud is divided into two types of sub-models, one of which focuses on the learning of global power data samples and uses the powerful computing power of the cloud to process large-scale and diverse data sets. The other type is for learning local professional data of edge devices to achieve effective integration of edge-cloud knowledge. This modular design enables the system to flexibly combine different professional modules and generate adaptive lightweight small models according to specific power task requirements, which is convenient for efficient deployment on the edge side. The present invention adopts a distributed computing and data parallel architecture, allowing edge devices to perform iterative training locally, reducing dependence on the cloud and improving training efficiency. Through the cloud parameter aggregation mechanism, the synchronous optimization of the model between the cloud and the edge device is realized, ensuring the consistency and accuracy of the model. This architecture enables the cloud and edge devices to co-evolve in real time and jointly improve the performance of the model. In the technical solution of the embodiment of the present invention, since the edge device can be trained locally, the data transmission with the cloud is reduced, thereby reducing the delay and bandwidth consumption. Sensitive data can be processed locally on the edge device without uploading to the cloud, effectively protecting data privacy; in addition, through edge-cloud collaboration and modular design, the system can respond to power task requirements more quickly and improve processing performance. The architecture of the present invention is particularly suitable for power application scenarios such as power transmission and distribution inspection, intelligent customer service, and operation safety monitoring. It demonstrates wide applicability and flexibility, effectively overcomes the limitations of existing power artificial intelligence edge-cloud collaborative applications, and provides strong technical support for the intelligent development of the power industry.

[0109] The following are device embodiments of the present invention, which can be used to implement the method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.

[0110] See also Figure 5 In an embodiment of the present invention, an artificial intelligence edge-cloud collaborative training system for all power scenarios is provided, including:

[0111] A model splitting module is used to split the power cloud large model into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output;

[0112] A training module is used to deploy the cloud sub-model on the cloud, and train the cloud sub-model with cloud samples to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, a lightweight cloud student model is obtained through knowledge distillation training; the cloud samples are stored on the edge to obtain cloud local samples; the lightweight cloud student model and the professional application sub-model are deployed on the edge, and the professional application sub-model is trained based on cloud local samples and edge local samples to obtain the trained professional application sub-model; wherein, when the professional application sub-model is trained based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model, and the lightweight cloud student model share a low-dimensional feature encoder.

[0113] In one embodiment of the present invention, a computer device is provided, the computer device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions in a computer storage medium to implement a corresponding method flow or a corresponding function; the processor described in the embodiment of the present invention can be used to perform the operation of an artificial intelligence edge-cloud collaborative training method for the entire power scenario.

[0114] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in a computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space that stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM (Random Access Memory) memory, or a non-volatile memory (non-volatile memory), such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the artificial intelligence edge-cloud collaborative training method for the entire power scene in the above embodiment.

[0115] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) that contain computer-usable program code.

[0116] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An artificial intelligence edge-cloud collaborative training method for the entire power scenario, characterized in that: The following steps are involved: The power cloud model is split into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output; The cloud sub-model is deployed on the cloud, and the cloud sub-model is trained using cloud samples to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, a lightweight cloud student model is obtained through knowledge distillation training; the cloud samples are stored on the edge to obtain cloud local samples; the lightweight cloud student model and the professional application sub-model are deployed on the edge, and the professional application sub-model is trained based on cloud local samples and edge local samples to obtain a trained professional application sub-model; wherein, when the professional application sub-model is trained based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model and the lightweight cloud student model share a low-dimensional feature encoder.

2. According to claim 1, an artificial intelligence edge-cloud collaborative training method for the entire power scenario is characterized in that: The step of splitting the power cloud large model into a cloud sub-model and a professional application sub-model comprises: The basic model part of the power cloud big model is used as a cloud sub-model, and the remaining part of the power cloud big model after splitting the basic model part is used as a professional application sub-model; Among them, the sum of the parameter ratio of the cloud sub-model to the power cloud large model and the parameter ratio of the professional application sub-model to the power cloud large model is 1; the low-dimensional feature encoder of the power cloud large model is shared at the bottom layer of the cloud sub-model and the professional application sub-model, and the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model and the network layer adapted to different downstream tasks are completely separated; the network architecture of the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model follows the same backbone network structure as the power cloud large model, and a different number of filters are used in each network layer to amplify and reduce the amount of parameters.

3. According to claim 1, an artificial intelligence edge-cloud collaborative training method for the entire power scenario is characterized in that: In the step of obtaining a lightweight cloud-based student model through knowledge distillation training based on the obtained trained cloud-based sub-model, The lightweight cloud-based student model is isomorphic to the cloud-based sub-model, shares the underlying low-dimensional feature encoder, and uses the same backbone network as the cloud-based sub-model; The lightweight cloud-based student model is provided with a certain number of filters to reduce the number of model parameters and is suitable for edge-side deployment; The loss function during knowledge distillation training is set to: Where, L KD is the loss function in the knowledge distillation training process; D c is a cloud sample dataset; (x i ,y i ) is the i-th sample in the cloud sample data set; l kd is the mean square error distance metric; They are the outputs of the cloud-based sub-model and the lightweight cloud-based student model, respectively.

4. According to claim 1, an artificial intelligence edge-cloud collaborative training method for the entire power scenario is characterized in that: In the step of deploying the lightweight cloud-based student model and the professional application sub-model on the edge side, training the professional application sub-model based on the cloud-based local samples and the edge-side local samples, and obtaining the trained professional application sub-model, Setting the weights of the lightweight cloud-based student model to be frozen and setting the shared low-dimensional feature encoder to be unchangeable; The calculation expression for calculating the training loss value is: Where L is the training loss calculation value; (x j ,y j ) is a sample of the side data batch B; l CE is the cross entropy loss function; Output of the sub-model for side-by-side professional applications; Output of the lightweight cloud-based student model.

5. According to claim 1, an artificial intelligence edge-cloud collaborative training method for the entire power scenario is characterized in that: The method further includes: deploying the lightweight cloud student model and the professional application sub-model on the edge end, training the professional application sub-model based on the cloud local sample and the edge end local sample, and obtaining the trained professional application sub-model; Based on the trained professional application sub-model, the cloud is used as a parameter server to aggregate and process the model parameters of the professional application sub-models updated from different sides.

6. An artificial intelligence edge-cloud collaborative training system for all power scenarios, characterized in that: include: A model splitting module is used to split the power cloud large model into a cloud sub-model and a professional application sub-model; wherein the cloud sub-model is used for data storage, computing support and data pre-processing, and the professional application sub-model is used for business logic processing and result output; A training module is used to deploy the cloud sub-model on the cloud, and train the cloud sub-model with cloud samples to obtain a trained cloud sub-model and a low-dimensional feature encoder; based on the obtained trained cloud sub-model, a lightweight cloud student model is obtained through knowledge distillation training; the cloud samples are stored on the edge to obtain cloud local samples; the lightweight cloud student model and the professional application sub-model are deployed on the edge, and the professional application sub-model is trained based on cloud local samples and edge local samples to obtain the trained professional application sub-model; wherein, when the professional application sub-model is trained based on cloud local samples and edge local samples, the training loss value and gradient are calculated according to the sum of the output of the lightweight cloud student model and the output of the professional application sub-model; the cloud sub-model, the professional application sub-model, and the lightweight cloud student model share a low-dimensional feature encoder.

7. The artificial intelligence edge-cloud collaborative training system for the entire power scenario according to claim 6 is characterized in that: In the model splitting module, the step of splitting the power cloud large model into a cloud sub-model and a professional application sub-model includes: The basic model part of the power cloud big model is used as a cloud sub-model, and the remaining part of the power cloud big model after splitting the basic model part is used as a professional application sub-model; Among them, the sum of the parameter ratio of the cloud sub-model to the power cloud large model and the parameter ratio of the professional application sub-model to the power cloud large model is 1; the low-dimensional feature encoder of the power cloud large model is shared at the bottom layer of the cloud sub-model and the professional application sub-model, and the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model and the network layer adapted to different downstream tasks are completely separated; the network architecture of the high-dimensional feature encoders of the cloud sub-model and the professional application sub-model follows the same backbone network structure as the power cloud large model, and a different number of filters are used in each network layer to amplify and reduce the amount of parameters.

8. The artificial intelligence edge-cloud collaborative training system for the entire power scenario according to claim 6 is characterized in that: In the training module, the step of obtaining a lightweight cloud student model through knowledge distillation training based on the obtained trained cloud sub-model is performed. The lightweight cloud-based student model is isomorphic to the cloud-based sub-model, shares the underlying low-dimensional feature encoder, and uses the same backbone network as the cloud-based sub-model; The lightweight cloud-based student model is provided with a certain number of filters to reduce the number of model parameters and is suitable for edge-side deployment; The loss function during knowledge distillation training is set to: Where, L KD is the loss function in the knowledge distillation training process; D c is a cloud sample dataset; (x i ,y i ) is the i-th sample in the cloud sample data set; l kd is the mean square error distance metric; They are the outputs of the cloud-based sub-model and the lightweight cloud-based student model, respectively.

9. The artificial intelligence edge-cloud collaborative training system for the entire power scenario according to claim 6 is characterized in that: In the training module, a step is executed in which a lightweight cloud student model and a professional application sub-model are deployed on the edge side, and the professional application sub-model is trained based on the cloud local sample and the edge side local sample to obtain the trained professional application sub-model. Setting the weights of the lightweight cloud-based student model to be frozen and setting the shared low-dimensional feature encoder to be unchangeable; The calculation expression for calculating the training loss value is: Where L is the training loss calculation value; (x j ,y j ) is a sample of the side data batch B; l CE is the cross entropy loss function; Output of the sub-model for side-by-side professional applications; Output of the lightweight cloud-based student model.

10. The artificial intelligence edge-cloud collaborative training system for the entire power scenario according to claim 6 is characterized in that: Also includes: The aggregation update module is used to aggregate the model parameters of professional application sub-models updated from different sides based on the trained professional application sub-models, using the cloud as a parameter server.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the artificial intelligence edge-cloud collaborative training method for the entire power scenario as described in any one of claims 1 to 5.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the artificial intelligence edge-cloud collaborative training method for the entire power scenario as described in any one of claims 1 to 5.

Images