Cloud-edge collaborative power defect identification method based on federated learning and related device

Abstract

The application discloses a cloud-edge collaborative power defect identification method based on federal learning and a related device. The method adopts the cloud-edge collaborative idea, performs coarse identification, and uploads only suspected defect data to the cloud side for fine identification, thereby effectively reducing the communication and calculation burden of the cloud side server. The federal learning idea is adopted to train and update the cloud side model and the edge side model, and the training process is performed with other parties to train the model under the condition that the sample data of each party does not cross the local, so that the data of each party can be safely and reliably used, and the problem of insufficient single training sample is indirectly overcome, thereby realizing the common optimization effect of joint multi-party on the cloud side and edge side models, and finally training an ideal cloud side model and edge side model. It can be seen that the application can effectively improve the collaboration between the cloud server and the edge side device, and enhance the identification accuracy and generalization ability of the cloud side model to the defects of the respective edge side device.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a cloud-edge collaborative power defect identification method and related apparatus based on federated learning. Background Technology

[0002] In the operation and maintenance of power grid systems, it is often necessary to analyze and interpret data from multiple participating devices in order to prevent and resolve equipment malfunctions in advance and ensure the normal and safe operation of the system.

[0003] For example, current equipment defect identification in power grid systems primarily employs a centralized processing model. This involves uploading inspection data from various locations to a central server, where artificial intelligence algorithms process the data uniformly. However, this model suffers from a heavy computational and communication burden on the central node. Furthermore, because the power grid industry is vital to the national economy and people's livelihoods, the confidentiality requirements for data generated by the power grid system are high, making cross-entity data exchange and aggregation difficult. This results in the inability to directly exchange and aggregate data from different power grid companies, and the limited defect data samples from each party lead to poor identification performance of models trained solely on that party.

[0004] In this regard, the relevant technologies do not provide an effective solution. Summary of the Invention

[0005] This application provides a cloud-edge collaborative power defect identification method and related apparatus based on federated learning, which can improve the application level of power data in the model and the model identification effect.

[0006] The first aspect of this application provides a cloud-edge collaborative power defect identification method based on federated learning, applied to participating cloud servers, including:

[0007] The cloud-side or edge-side models stored on their respective cloud servers are trained using power sample data from each party in several rounds of local model training. The initial parameters of the cloud-side or edge-side models are obtained based on the preset model parameter configuration sent by the parameter server. The sample data used by the edge-side model and the cloud-side model are the initial sample data collected by the end-side device and the defect data initially screened out by the edge-side model from the initial sample data. The cloud-side model is used to perform fine screening on the initially screened defect data.

[0008] The updated parameters of each cloud-side or edge-side model obtained from training are transmitted to the parameter server so that the parameter server can aggregate the updated parameters of each cloud-side or edge-side model to obtain the updated parameters of the aggregated model of the cloud-side or edge-side model. The aggregated model of the cloud-side or edge-side model refers to the global model obtained by integrating the cloud-side or edge-side models of each participant.

[0009] The updated parameters of the cloud-side or edge-side aggregated model sent by the parameter server are used as new preset model parameters to update the parameters of each cloud-side or edge-side model. Then, the step of using sample data from each party to perform several rounds of local model training on the cloud-side or edge-side models stored on their respective cloud servers is returned and executed. The updating of each model is stopped until the updated cloud-side or edge-side model or each aggregated model reaches the convergence condition.

[0010] The final updated edge-side models and / or cloud-side models are used as target models. The target models are used to identify the defects of the edge-side devices under test from coarse to fine based on the data collected by the edge-side devices under test. The recipients of the final updated edge-side models include the edge-side computing platforms of each cloud server.

[0011] A second aspect of this application provides a cloud-edge collaborative power defect identification method based on federated learning, applied to a parameter server, including:

[0012] Send preset model parameters to each participating cloud server. The preset model parameters are used to configure the cloud-side or edge-side models stored on each cloud server.

[0013] The update parameters of the cloud-side or edge-side models trained based on power sample data are aggregated to obtain the update parameters of the cloud-side or edge-side aggregated model as the new preset model parameters, which are then sent to the cloud servers. This allows each cloud server to perform the next round of model training on its latest cloud-side model based on the sample data and the new preset model parameters. The cloud-side aggregated model refers to the global model obtained by integrating the various cloud-side models.

[0014] For the cloud-side or edge-side model update parameters obtained from the next round of model training, return to the step of executing the cloud-side or edge-side model update parameters obtained by aggregating each cloud-side model based on the sample data, and stop updating each model until the updated cloud-side or edge-side model, or each aggregated model reaches the convergence condition, so as to use the finally updated cloud-side and / or edge-side models as target models for identifying defects in their respective end-side devices.

[0015] A third aspect of this application provides a cloud-edge collaborative power defect identification method based on federated learning, including:

[0016] The data collected by the end-side device is input into the corresponding target model to output the defect information contained in the end-side device;

[0017] The target model is obtained by the method described in the first or second aspect, and the defect information includes at least one of the following: whether the equipment status is abnormal, the level of abnormality, the type of defect, and the location of the defect.

[0018] A fourth aspect of this application provides a cloud server, including: a first processing unit and a first sending unit;

[0019] The first processing unit is used to perform several rounds of local model training on the cloud-side or edge-side model stored on their respective cloud servers using power sample data from each party. The initial parameters of the cloud-side or edge-side model are obtained according to the preset model parameter configuration sent by the parameter server. The sample data used by the edge-side model and the cloud-side model are the initial sample data collected by the end-side device and the defect data initially screened out by the edge-side model from the initial sample data. The cloud-side model is used to perform fine screening on the initially screened defect data.

[0020] The first sending unit is used to transmit the updated parameters of each cloud-side or edge-side model obtained from the training to the parameter server, so that the parameter server aggregates the updated parameters of each cloud-side or edge-side model to obtain the updated parameters of the cloud-side or edge-side aggregated model, wherein the cloud-side or edge-side aggregated model refers to the global model obtained by integrating each cloud-side or edge-side model.

[0021] The first processing unit is further configured to take the updated parameters of the cloud-side or edge-side aggregated model sent by the parameter server as new preset model parameters to update the parameters of each cloud-side or edge-side model, and return to execute the step of using the sample data of each party to perform several rounds of local model training on the cloud-side or edge-side models stored in their respective cloud servers, and stop updating each model until the updated cloud-side or edge-side model or each aggregated model reaches the convergence condition.

[0022] The first processing unit is further configured to use the finally updated edge-side models and / or cloud-side models as target models, wherein the target models are used to identify the defects of the edge-side devices under test from coarse to fine based on the data collected by the edge-side devices under test; the recipients of the finally updated edge-side models include the edge-side computing platforms of each cloud server.

[0023] A fifth aspect of this application provides a parameter server, including: a second processing unit and a second sending unit;

[0024] The second sending unit is used to send preset model parameters to each participating cloud server, wherein the preset model parameters are used to configure the cloud-side or edge-side model stored in each cloud server.

[0025] The second processing unit is used to aggregate the update parameters of each cloud-side or edge-side model trained based on power sample data, and send the updated parameters of the aggregated cloud-side or edge-side model as new preset model parameters to each cloud server, so that each cloud server can perform the next round of model training on its latest cloud-side or edge-side model based on the sample data and the new preset model parameters; the aggregated cloud-side or edge-side model refers to the global model obtained by integrating each cloud-side or edge-side model;

[0026] The second processing unit is further configured to, for the cloud-side or edge-side model update parameters obtained in the next round of model training, return to the step of executing the cloud-side or edge-side model update parameters obtained by aggregating each cloud-side or edge-side model based on sample data, and stop updating each model until the updated cloud-side or edge-side model, or each aggregated model reaches the convergence condition, so as to use the finally updated cloud-side and / or edge-side models as target models for identifying defects in their respective end-side devices.

[0027] A sixth aspect of this application provides an identification device, including: an input unit;

[0028] The input unit is used to input the data collected by the end-side device into the corresponding target model to obtain the equipment defect information contained in the power data collected by the end-side device.

[0029] The target model is obtained by the method described in the first or second aspect, and the defect information includes at least one of the following: whether the equipment status is abnormal, the level of abnormality, the type of defect, and the location of the defect.

[0030] In specific implementation, the methods described in the first, second, and third aspects of this application may be implemented using the content described in the fourth, fifth, and sixth aspects of this application.

[0031] A seventh aspect of this application provides an electronic device, including:

[0032] Central processing unit, memory, and input / output interfaces;

[0033] The memory is either a short-term storage memory or a persistent storage memory;

[0034] The central processing unit is configured to communicate with the memory and execute instructions in the memory to perform the method described in any specific implementation of the first to third aspects of the embodiments of this application.

[0035] An eighth aspect of this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform a method as described in any specific implementation of the first to third aspects of this application.

[0036] A ninth aspect of this application provides a computer program product comprising instructions or a computer program, which, when run on a computer, causes the computer to perform a method as described in any specific implementation of the first to third aspects of this application.

[0037] As can be seen from the above technical solutions, the embodiments of this application have at least the following advantages:

[0038] By adopting a cloud-edge collaborative approach, a coarse identification (initial screening) process is used, uploading only available defect data to the cloud for fine identification (refined screening), effectively reducing the communication and computational burden on the cloud server. Employing federated learning, both the cloud-side and edge-side models are trained and updated. The training process ensures that each participant's sample data remains local, allowing for collaborative model training with other parties. This ensures the secure and reliable use of data from all parties and indirectly overcomes the problem of insufficient training samples from a single party. This collaborative approach achieves joint optimization of the cloud-side and edge-side models, ultimately training ideal cloud-side and edge-side models. Therefore, this application effectively improves the collaborative effect between the cloud server and edge devices, enhancing the accuracy and generalization ability of the cloud-side model in identifying defects in their respective edge devices. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0040] It should be noted that although the steps in the flowcharts (if any) involved in the embodiments are drawn sequentially according to the arrows, unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts involved in the embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0041] Figure 1 This is a schematic flowchart of a method according to an embodiment of this application;

[0042] Figure 2 This is another schematic flowchart of the method in an embodiment of this application;

[0043] Figure 3This is another schematic flowchart of the method in an embodiment of this application;

[0044] Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments 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 a 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.

[0047] In the following description, expressions such as "one specific implementation" or "one specific example" describe a subset of all possible embodiments. However, it is understood that "one specific implementation" or "one specific example" can be the same or different subset of all possible embodiments and can be combined with each other without conflict. In the following description, the term "multiple" means at least two. When a certain value mentioned in this application reaches a threshold (if it exists), in some specific examples, it may include the former being greater than the latter. When "any" or "at least one" or similar expressions are mentioned, it specifically refers to any one of the listed examples or any combination of these examples.

[0048] 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.

[0049] In recent years, the use of cloud-edge collaborative mode for inspection data processing has been widely recognized in the industry. That is, a certain amount of computing power and artificial intelligence algorithms are configured near the inspection equipment or on the inspection equipment (i.e., the end-side equipment), and the data is initially processed on the edge platform (edge ​​side), and then the data that needs further processing is uploaded to the central cloud server (cloud side) for processing. This mode can largely solve the problem of heavy communication and computing burden on the central server, but some key problems have not been solved: (1) Deep learning technology is data-driven and requires a large amount of high-quality data for training to obtain high-quality models. However, many defects in power grid equipment are sporadic phenomena with a small number of samples, and it is difficult to obtain high-quality samples. Commonly used deep neural networks often produce serious overfitting phenomena for small sample datasets, resulting in poor detection performance. (2) Since the power grid industry is related to the national economy and people's livelihood, the data generated by the power grid system has high confidentiality requirements or is difficult to desensitize and declassify, making it difficult to carry out cross-entity data exchange, sharing and aggregation, resulting in the inability to use data from different power grid companies to jointly update the defect identification model, and there is a problem of difficulty in identifying small sample defects.

[0050] To address the aforementioned issues, this application provides a cloud-edge collaborative power defect identification method and related apparatus based on federated learning, which solves the problem of poor sample data quality and improves the application level of sample data in the model and the model's identification effect.

[0051] The following section uses a power grid system as an example to further illustrate the method of this application. However, similarly, this method can also be applied to scenarios other than power grid systems where model prediction based on federated learning can be achieved, depending on the specific implementation in the actual scenario. It should be noted that the edge-side model operation in this application is similar to the cloud-side model operation. That is, the training and update process of the edge-side model and the edge-side aggregated model can be similarly referenced to the training and update process of the cloud-side model and the cloud-side aggregated model, without further elaboration. Furthermore, in some specific examples, the edge-side model can be trained on the cloud server side and then distributed to the edge computing platform, or it can be trained by the edge computing platform itself. In other words, as long as the edge-side model can be trained, it doesn't matter who performs the operation.

[0052] Please see Figures 1 to 3 This application provides a specific embodiment of power defect identification based on federated learning and cloud-edge collaboration, which includes the following operational steps:

[0053] 11. The parameter server sends preset model parameters to each participating cloud server.

[0054] The preset model parameters are used to configure the cloud-side or edge-side models stored on each cloud server. The method in this embodiment can be applied to... Figure 2The power grid system scenario shown can be specifically assumed to be that each power grid company operates a system corresponding to the method of this application embodiment. This system can be called a device defect identification system based on federated learning and multi-side collaboration. Each system includes a cloud server (cloud side), an edge computing platform (edge ​​side), and end-side devices (end-side). The cloud server and the latter two are connected in communication. The edge computing platform is connected to end-side devices such as handheld inspection terminals, inspection drones, inspection robots, and high-definition monitoring to obtain data such as images, videos, or text collected by them.

[0055] 12. The cloud server trains the cloud-side model or edge-side model based on preset model parameters and sample data from various parties.

[0056] The cloud server uses power sample data from various parties to train several rounds of models on the cloud-side or edge-side models stored on their respective cloud servers. The initial parameters of the cloud-side or edge-side models are obtained based on preset model parameter configurations sent by the parameter server. Therefore, it can be understood that the initial model parameters of the cloud-side or edge-side models are based on the preset model parameter configurations initially issued by the parameter server (global side). Initially, the updated parameters of the cloud-side model refer to the model parameters obtained after training the cloud-side model with sample data, based on the preset model parameter configuration. Here, the sample data used by the edge-side model and the cloud-side model are the initial sample data collected by the edge-side device and the defect data (suspected) initially screened out by the edge-side model from the initial sample data, respectively. The cloud-side model is used to further refine the initially screened defect data.

[0057] As one possible implementation, edge-side models can be used to perform initial screening of the initial sample data collected by edge devices, filtering out defective data with confidence levels higher than a preset confidence level for further screening (i.e., fine-tuning) by cloud-side models. The initial screening process can be implemented by edge-side models stored on an edge computing platform, which communicates with the cloud server and edge devices. In some specific examples, sample data is collected by the corresponding edge devices of each cloud server. The fine-tuning process includes: using each cloud-side model to further screen the sample data initially screened by its respective edge-side model, and using the finely identified defective data with confidence levels lower than a preset confidence level for model training.

[0058] like Figure 2As shown, in order to reduce redundant data, the edge computing platform samples and stores the acquired data, and uses a recognition algorithm model (i.e., edge model) downloaded from the cloud server in advance to perform preliminary screening of the data. That is, the data is input into the model, and the model inference is performed to identify the equipment in the inspection image or video and determine whether the equipment in the picture is in a normal or abnormal state. If the abnormal confidence level is higher than a certain threshold, the data (such as the defect data determined by the preliminary screening) is uploaded to the cloud server for further fine recognition to avoid mistaking non-defect data for defect data.

[0059] On the cloud server, the data uploaded by the edge computing platform is stored. On one hand, this data is input into the pre-stored recognition algorithm model (i.e., the cloud-side model) on the cloud server, performing model inference to perform refined identification (i.e., fine screening) of equipment defects in inspection images or videos. This determines the type, severity, and / or location of defects, and provides corresponding prediction confidence levels and other model inference results. Based on the model inference results, inspection reports are generated, summarizing the defect identification results in text, tables, images, etc., providing a basis for equipment maintenance and repair. On the other hand, defect data with low confidence levels in the model inference results, defect data with small sample sizes, and defect data with identification errors discovered during manual review or spot checks are manually labeled for retraining the intelligent recognition model used on the cloud server and edge computing platform. Model retraining can be performed independently on a local dataset using conventional training, or in collaboration with other power grid companies using federated learning, utilizing data from all parties to jointly train and update the model, addressing the problem of insufficient samples from a single party. After conducting local training or federated learning training, the new model obtained after training is saved and replaced with the existing models on cloud servers and edge computing platforms as needed.

[0060] In some specific examples, the above model training uses labeled sample data for supervised learning, or uses partially labeled sample data for semi-supervised learning; labeled sample data refers to sample data belonging to the defect class or non-defect class, where the type of sample data can be identified by the model or manually labeled.

[0061] 13. The cloud server transmits the updated parameters of each cloud-side or edge-side model obtained from training to the parameter server.

[0062] The cloud server transmits the updated parameters of each cloud-side or edge-side model obtained from training to the parameter server, so that the parameter server can aggregate the updated parameters of each cloud-side or edge-side model to obtain the updated parameters of the cloud-side or edge-side aggregated model. The cloud-side aggregated model refers to the global model obtained by integrating each cloud-side model, and the edge-side aggregated model refers to the global model obtained by integrating each edge-side model.

[0063] 14. The parameter server aggregates the update parameters of each cloud-side or edge-side model, obtains the update parameters of the aggregated cloud-side or edge-side model, and sends them to each cloud server.

[0064] After transmitting the update parameters of each cloud-side or edge-side model to the parameter server, the parameter server can aggregate these parameters to update the stored global model (such as the cloud-side aggregated model and / or edge-side aggregated model), obtaining the update parameters of the global model (which can be called the update gradient). This allows the cloud-side model (and the same applies to the edge-side model) to perform another round of model training based on the global model update gradient (which can be called model update). For example, the parameter configuration of the cloud-side model can be updated based on the global model update gradient, and the updated model can be trained using sample data. In other words, the specific operation can be as follows: the parameter server aggregates the update parameters of each cloud-side or edge-side model trained based on sample data, obtains the update parameters of the cloud-side or edge-side aggregated model (which can be called the global model update gradient) as the new preset model parameters, and sends them to each cloud server. This allows each cloud server to perform the next round of model training on its latest cloud-side or edge model based on the sample data and the new preset model parameters. The cloud-side aggregated model refers to the global model obtained by integrating the various cloud-side models.

[0065] 15. The cloud server performs the next round of model training for each cloud-side model or edge-side model based on the updated parameters of the cloud-side or edge-side aggregated model.

[0066] The cloud server uses the updated parameters of the cloud-side or edge-side aggregated model sent by the parameter server as the new preset model parameters to update the parameters of each cloud-side or edge-side model. It then returns to step 12, which involves performing one round of model training on the cloud-side or edge-side models stored on their respective cloud servers using electricity sample data from each party. This process continues until the updated cloud-side or edge-side models, or each aggregated model, reaches convergence, at which point the updating of each model stops. Here, "preset number" can refer to the total number of models participating in federated learning.

[0067] 16. The parameter server aggregates the updated parameters of the cloud-side or edge-side model obtained from the next round of model training, and sends the aggregated results to the cloud-side or edge-side model to continue training until the model converges.

[0068] The parameter server updates the parameters of the cloud-side or edge-side models obtained from the next round of model training, and returns to step 14, which aggregates the cloud-side or edge-side models trained based on the sample data to update the parameters of the cloud-side or edge-side models. The update of each model stops when the updated cloud-side or edge-side model or each aggregated model reaches the convergence condition.

[0069] 17. The cloud server uses the latest cloud-side and / or edge-side model as the target model.

[0070] The final updated cloud-side and / or edge-side models are used as target models to identify defects in their respective end-side devices. Specifically, compared to edge-side models, cloud-side models are used to more accurately identify defect information in end-side devices based on the data collected (after initial screening) from the end-side devices. Alternatively, a parameter server, rather than a cloud server, can use the final updated edge-side and cloud-side models as target models to accurately identify defects in their respective end-side devices from a coarse to fine perspective. In other words, the entity responsible for determining the target model can be either a parameter server or a cloud server, depending on the requirements.

[0071] In some specific examples, to optimize the initial screening and recognition effect of the edge-side model, similar operations to training the cloud-side model (which can achieve refined recognition) and updating the cloud-side aggregated model can be performed. Specifically, if the edge computing platform trains the edge-side model itself, rather than receiving a pre-trained and distributed edge-side model from the cloud server, the parameter server sends preset model parameters to the corresponding edge computing platform of each cloud server. These preset model parameters configure the edge-side models stored on each edge computing platform. The edge-side models are used for initial screening and recognition of the initial sample data collected by their respective edge devices. The edge computing platform then transmits the initially screened sample data to the corresponding cloud-side model for refined recognition and model training. The parameter server aggregates the edge-side models trained based on the initial sample data and updates the parameters accordingly. The updated parameters of the edge aggregation model, which serves as the parameters of the new preset model, are sent to each edge computing platform. This allows each edge computing platform to perform the next round of model training based on the initial sample data and the new preset model parameters. The edge aggregation model refers to the global model obtained by integrating the edge models. For the edge model update parameters obtained in the next round of model training, the process returns to the step of aggregating the edge model update parameters obtained by training each edge model based on the sample data. The process continues until the updated preset edge models and / or the edge aggregation model reach the convergence condition, at which point the updating of each model stops.

[0072] It should be noted that the execution order of training the cloud-side model and updating the cloud-side aggregated model is not limited and can be performed simultaneously, depending on the actual situation.

[0073] This application can combine federated learning, cloud-edge collaboration, and power inspection scenarios to jointly update the model without leaving the local data of each party. This effectively solves problems such as insufficient power grid equipment defect samples, difficulty in sharing confidential data from all parties, and heavy computing and communication burden on the central server. In turn, it can effectively improve the collaboration between cloud servers and edge devices, and enhance the accuracy and generalization ability of the cloud-side model in identifying defects of their respective edge devices.

[0074] As can be seen, due to the low probability of defects in some power grid equipment and the difficulty in obtaining samples, single-party data samples are insufficient to train large-scale intelligent defect identification algorithms. Therefore, in this case, multiple power grid companies can jointly conduct federated learning to train the intelligent identification algorithm. When conducting federated learning, in addition to the cloud servers of each power grid company, a shared parameter server is also needed. This parameter server can be provided by one of the participating power grid companies or by a trusted third party. For ease of understanding, the operations in steps 11 to 17 above can be summarized as follows:

[0075] (1) Data preparation. Each party prepares its own dataset according to the unified specifications (such as rules for initial screening and / or detailed identification) and places it in the data storage path of its own cloud server.

[0076] (2) The parameter server sends the initial preset model parameter information to each participating cloud server; of course, the parameter server can also choose to send the initial global model (such as cloud-side aggregation model and / or edge-side aggregation model) to each participating cloud server; in some examples, initially, the cloud-side aggregation model and the edge-side aggregation model can be used to differentiate multiple cloud-side models and multiple edge-side models and send them to the cloud server and the edge computing platform.

[0077] (3) The user enables the federated learning application and begins the federated training process. Taking the cloud-side model and the cloud-side aggregated model as the updated objects as an example, the specific process is as follows: Figure 3 As shown:

[0078] ① Each participating party's cloud server performs one round of model training on its local data according to the preset model parameter settings. This training can be supervised learning using fully labeled data of all data types, or semi-supervised learning using partially labeled sample data.

[0079] ② Each participating party's cloud server encrypts and uploads the updated gradient (i.e., model update parameters) of the local model (cloud-side model) for one round of training to the parameter server.

[0080] ③ The parameter server collects and aggregates the update gradients uploaded by each participating cloud server to obtain the global model update gradient. Typical aggregation methods include algorithms such as FedAvg or Scaffold. The FedAvg algorithm aggregates model parameters through weighted averaging. Its basic idea is to upload the parameters of the local model to the server, the server calculates the average of all model parameters, and then broadcasts this average back to all local devices. This process can be iterated multiple times until the model converges.

[0081] ④ The parameter server sends the updated gradients of the global model to the cloud servers of each participant in an encrypted manner. Typical encryption algorithms include homomorphic encryption and differential privacy.

[0082] ⑤ Each participating party's cloud server decrypts the update gradient of the global model and updates its local model, which can be regarded as another round of model training.

[0083] Repeat steps ① to ⑤ until each individual model or the global model reaches convergence, then stop federated training. As a possible implementation, finally, local data can be used to evaluate the accuracy of each model. If its accuracy is higher than existing models or meets other evaluation criteria, the model is saved for later use, ultimately leading to the ideal target model for accurately identifying equipment defects. Similarly, side-side models and side-side aggregated models can act as update targets, performing steps ① to ⑤ as described above to complete the training and updating of various side-side models, thereby improving the initial screening accuracy of initial power data and reducing data processing pressure.

[0084] This application, in another aspect, provides a specific embodiment of a model usage method, which includes the following steps: inputting data collected by an end-side device into a corresponding target model to output defect information contained in the end-side device; wherein, the target model is obtained through... Figures 1 to 3 The defect information obtained by any illustrative method includes at least one of the following: whether the equipment status is abnormal, the level of abnormality, the type of defect, and the location of the defect.

[0085] As explained above, this application's embodiments not only utilize federated learning for model training, but also feature a key characteristic: achieving coarse-to-fine identification of power equipment defects through cloud-edge collaboration (coordination between the cloud and edge sides). This reduces the significant communication overhead and computational burden on central nodes associated with traditional centralized identification. Federated learning addresses the problem of insufficient defect sample data from a single source; cloud-edge collaboration addresses the heavy communication and computational burden on central nodes in traditional centralized processing models. For example, all data collected by edge devices must be uploaded to a central server (which can be considered the cloud side) for processing, inevitably leading to heavy communication and computational burdens. This application's embodiments, however, use an edge-side model for initial coarse screening, uploading only suspected defect data to the cloud side, thus eliminating most non-defective data and significantly reducing the communication and computational burden on the central server. Then, the cloud-side model on the central server performs refined identification.

[0086] Furthermore, it is understandable that both cloud-side and edge-side models can be trained and updated using federated learning, with similar processes. As one possible implementation, the cloud-side model can be trained using pre- or post-screening power data (to update the cloud-side model), and can also train the edge-side model using federated learning, then distribute the trained edge-side model to the edge for updates. This is because the edge-side model can be stored on the cloud side, allowing the edge-side to perform model inference without training.

[0087] In summary, the embodiments of this application can fully utilize edge computing power for data processing through multi-party collaboration, and combine federated learning technology to conduct joint model training. While ensuring that data from each party does not leave their local machine, it securely utilizes data from all parties to jointly train the model, solving the problem of insufficient single-party samples and overcoming the shortcomings of traditional centralized server-side computing and communication burdens, insufficient small-sample defect data, and inability to share data among multiple parties. Therefore, the method of using this model is beneficial to improving the accuracy and generalization ability of power grid equipment defect identification algorithms, and enhancing the level of intelligent operation and maintenance of power grid equipment.

[0088] This application also provides a cloud server, including: a first processing unit and a first sending unit;

[0089] The first processing unit is used to perform several rounds of local model training on the cloud-side or edge-side model stored on their respective cloud servers using power sample data from each party. The initial parameters of the cloud-side or edge-side model are obtained according to the preset model parameter configuration sent by the parameter server. The sample data used by the edge-side model and the cloud-side model are the initial sample data collected by the end-side device and the defect data initially screened out by the edge-side model from the initial sample data. The cloud-side model is used to perform fine screening on the initially screened defect data.

[0090] The first sending unit is used to transmit the updated parameters of each cloud-side or edge-side model obtained from the training to the parameter server, so that the parameter server aggregates the updated parameters of each cloud-side or edge-side model to obtain the updated parameters of the cloud-side or edge-side aggregated model, wherein the cloud-side or edge-side aggregated model refers to the global model obtained by integrating each cloud-side or edge-side model.

[0091] The first processing unit is further configured to take the updated parameters of the cloud-side or edge-side aggregated model sent by the parameter server as new preset model parameters to update the parameters of each cloud-side or edge-side model, and return to execute the step of using the sample data of each party to perform several rounds of local model training on the cloud-side or edge-side models stored in their respective cloud servers, and stop updating each model until the updated cloud-side or edge-side model or each aggregated model reaches the convergence condition.

[0092] The first processing unit is further configured to use the finally updated edge-side models and / or cloud-side models as target models, wherein the target models are used to identify the defects of the edge-side devices under test from coarse to fine based on the data collected by the edge-side devices under test; the recipients of the finally updated edge-side models include the edge-side computing platforms of each cloud server.

[0093] Optionally, the sample data is collected by the corresponding edge devices of each cloud server, and the first processing unit is specifically used for:

[0094] The edge model is used to perform initial screening and identification on the initial sample data collected by the end device, so as to filter out defective data with a confidence level higher than the preset confidence level for fine screening by the cloud model; wherein, the initial screening and identification process is implemented by the edge model stored in the edge computing platform, and the edge computing platform is communicatively connected to the cloud server and the end device.

[0095] Optionally, the first processing unit is specifically used for:

[0096] Each cloud-side model is used to refine the sample data initially screened by its respective side-side model, and the defective data with confidence levels lower than the preset confidence level that are carefully identified are used for model training.

[0097] This application also provides a parameter server, including: a second processing unit and a second sending unit;

[0098] The second sending unit is used to send preset model parameters to each participating cloud server, wherein the preset model parameters are used to configure the cloud-side or edge-side model stored in each cloud server.

[0099] The second processing unit is used to aggregate the update parameters of each cloud-side or edge-side model trained based on power sample data, and send the updated parameters of the aggregated cloud-side or edge-side model as new preset model parameters to each cloud server, so that each cloud server can perform the next round of model training on its latest cloud-side or edge-side model based on the sample data and the new preset model parameters; the aggregated cloud-side or edge-side model refers to the global model obtained by integrating each cloud-side or edge-side model;

[0100] The second processing unit is further configured to, for the cloud-side or edge-side model update parameters obtained in the next round of model training, return to the step of executing the cloud-side or edge-side model update parameters obtained by aggregating each cloud-side or edge-side model based on sample data, and stop updating each model until the updated cloud-side or edge-side model, or each aggregated model reaches the convergence condition, so as to use the finally updated cloud-side and / or edge-side models as target models for identifying defects in their respective end-side devices.

[0101] Optionally, if the edge computing platform trains the edge model itself, rather than receiving a pre-trained edge model from the cloud server, the second processing unit is further used for:

[0102] Preset model parameters are sent to the edge computing platforms corresponding to each cloud server. The preset model parameters are used to configure the edge models stored on each edge computing platform. The edge models are used to perform initial screening and identification on the initial sample data collected by their respective end devices, so that the initially screened sample data can be transmitted to the corresponding cloud-side models for fine identification and model training.

[0103] The update parameters of each edge model trained based on the initial sample data are aggregated to obtain the update parameters of the edge aggregate model, which are then sent to each edge computing platform. This allows each edge computing platform to perform the next round of model training based on the initial sample data and the new preset model parameters. The edge aggregate model refers to the global model obtained by integrating each edge model.

[0104] For the edge model update parameters obtained from the next round of model training, return to the step of aggregating the edge model update parameters trained based on the sample data, and stop updating each model when the preset number of updated edge models and / or the edge aggregate model reach the convergence condition.

[0105] In this application, the operations performed by each unit of the cloud server and parameter server are the same as those described above. Figures 1 to 3 The operations described in the illustrated method embodiments are similar, and will not be repeated here. Of course, this application... Figures 1 to 3 For details on the implementation process of each operation, please refer to the relevant descriptions of the cloud server and parameter server units.

[0106] Another aspect of this application provides an identification device, including: an input unit;

[0107] The input unit is used to input the data collected by the end-side device into the corresponding target model in order to obtain the equipment defect information contained in the power data collected by the end-side device.

[0108] Among them, the target model is through the aforementioned Figures 1 to 3 The method described herein provides defect information that includes at least one of the following: whether the equipment status is abnormal, the level of abnormality, the type of defect, and the location of the defect.

[0109] Please see Figure 4 The electronic device 400 in this application embodiment may include one or more central processing units (CPUs) 401 and a memory 405, wherein the memory 405 stores one or more applications or data.

[0110] The memory 405 can be volatile or persistent storage. The program stored in the memory 405 can include one or more modules, each module including a series of instruction operations on the electronic device. Furthermore, the central processing unit 401 can be configured to communicate with the memory 405 and execute the series of instruction operations stored in the memory 405 on the electronic device 400.

[0111] Electronic device 400 may also include one or more power supplies 402, one or more wired or wireless network interfaces 403, one or more input / output interfaces 404, and / or one or more operating systems, such as Windows Server™, MacOSX™, Unix™, Linux™, FreeBSD™, etc.

[0112] The central processing unit 401 can perform the operations performed by the first aspect or any specific method embodiment of the first aspect, which will not be described in detail here.

[0113] This application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the aforementioned actions. Figures 1 to 3 The method described by any specific implementation thereof.

[0114] This application provides a computer program product containing instructions or a computer program, which, when run on a computer, causes the computer to perform the aforementioned... Figures 1 to 3 The method described by any specific implementation thereof.

[0115] It is understood that, in the various embodiments of this application, the sequence number of each step does not imply the order of execution. The execution order of each step should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0116] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system (if it exists) and device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0117] 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 units 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 apparatus, 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.

[0118] The units described 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.

[0119] 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.

[0120] 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 (computer program product) is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a business server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A federated learning based cloud-edge collaborative power defect identification method, characterized in that, The application is applied to each cloud server participating in, comprising: using the power sample data of each party to respectively perform several rounds of local model training on the cloud side or edge side model stored in the cloud server of each party; the initial parameters of the cloud side or edge side model are obtained according to the preset model parameter configuration sent by the parameter server, and the sample data used by the edge side model and the cloud side model respectively is the initial sample data collected by the terminal side device and the defect data preliminarily screened out from the initial sample data by the edge side model, and the cloud side model is used for fine screening of the preliminarily screened out defect data; the updated parameters of each cloud side or edge side model obtained by training are transmitted to the parameter server, so that the parameter server aggregates the updated parameters of the cloud side or edge side model to obtain the updated parameters of the aggregated model of the cloud side or edge side model, wherein the aggregated model of the cloud side or edge side model refers to a global model obtained by integrating the cloud side or edge side model of each participant; the updated parameters of the cloud side or edge side aggregated model sent by the parameter server are used as new preset model parameters to update the parameters of each cloud side or edge side model, and the step of using the sample data of each party to respectively perform several rounds of local model training on the cloud side or edge side model stored in the cloud server of each party is returned, and the update of each model is stopped until the updated cloud side or edge side model or each aggregated model reaches the convergence condition; each edge side model and / or cloud side model finally updated is used as a target model, and the target model is used to identify the defects of the terminal side device to be tested from coarse to fine through the data collected by the terminal side device to be tested; the receiving end of each edge side model finally updated includes the edge side computing platform of each cloud server party; the preliminary screening process of the sample data comprises: the initial sample data collected by the terminal side device is preliminarily screened and identified by the edge side model stored in the edge computing platform, so that the defect data with confidence higher than the preset confidence is screened out for fine screening by the cloud side model; wherein the preliminary screening process is implemented by the edge side model stored in the edge computing platform, and the edge computing platform is in communication connection with the cloud server and the terminal side device.

2. The method of claim 1, wherein, the fine screening process of the sample data comprises: each cloud side model is used to fine screen the sample data preliminarily screened out by the edge side model, and the defect data with confidence lower than the preset confidence is used for model training.

3. A cloud-edge collaborative power defect identification method based on federated learning, characterized in that, The application is applied to the parameter server, comprising: sending preset model parameters to each cloud server participating in, the preset model parameters being used to configure the cloud side or edge side model stored in each cloud server; aggregating the cloud side or edge side model update parameters obtained by each cloud side or edge side model according to the power sample data to obtain the update parameters of the cloud side or edge side aggregated model as new preset model parameters and send them to each cloud server, so that each cloud server performs the next round of model training on the latest cloud side model of each party according to the sample data of each party and the new preset model parameters; the cloud side aggregated model refers to a global model obtained by integrating each cloud side model; For the cloud side or edge side model update parameters obtained by the next round of model training, the step of aggregating the cloud side or edge side model update parameters obtained by each cloud side model according to the sample data is executed again, and the update of each model is stopped until the updated cloud side or edge side model or each aggregated model reaches the convergence condition, so that the final updated cloud side and / or edge side model is used as the target model for identifying the defects of the respective end side device. The cloud server executes the method of claim 1.

4. The method of claim 3, wherein, If the edge side model is trained by the edge side computing platform itself instead of receiving the edge side model trained by the cloud server and delivered, before the final updated cloud side and / or edge side model is used as the target model for identifying the defects of the respective end side device, the method further comprises: sending preset model parameters to each cloud server corresponding edge side computing platform, the preset model parameters are used to configure the edge side model stored in each edge side computing platform; the edge side model is used for preliminary screening and identification of the initial sample data collected by each end side device, so as to obtain the preliminary screened sample data and transmit it to the corresponding cloud side model for fine identification and model training; aggregating the edge side model update parameters obtained by each edge side model according to the initial sample data to obtain the update parameters of the edge side aggregated model as new preset model parameters and send them to each edge side computing platform, so that each edge side computing platform performs the next round of model training on the latest edge side computing platform according to the initial sample data of each party and the new preset model parameters; the edge side aggregated model refers to the global model obtained by integrating each edge side model; For the edge side model update parameters obtained by the next round of model training, the step of aggregating the edge side model update parameters obtained by each edge side model according to the sample data is executed again, and the update of each model is stopped until the updated preset edge side model and / or edge side aggregated model reaches the convergence condition.

5. A cloud-edge collaborative power defect identification method based on federated learning, characterized in that, It comprises: inputting the data collected by the end side device into the corresponding target model to output the defect information contained in the end side device; wherein the target model is obtained by the method of any one of claims 1 to 4, and the defect information includes at least one of device state abnormality, abnormality level, defect type and defect location.

6. A cloud server, characterized by, It comprises: a first processing unit and a first sending unit; The first processing unit is used for performing several rounds of local model training on the cloud side or edge side model stored in each cloud server using the power sample data of each party, and the initial parameters of the cloud side or edge side model are configured according to the preset model parameters sent by the parameter server; the sample data used by the edge side model and the cloud side model respectively is the initial sample data collected by the end side device and the defect data preliminarily screened out from the initial sample data by the edge side model, and the cloud side model is used for fine screening of the preliminarily screened defect data; The first sending unit is configured to transmit the cloud-side or edge-side model update parameters obtained through training to the parameter server, so that the parameter server aggregates the cloud-side or edge-side model update parameters to obtain update parameters of a cloud-side or edge-side aggregated model, wherein the cloud-side or edge-side aggregated model refers to a global model obtained by integrating each cloud-side or edge-side model. The first processing unit is further configured to use the update parameters of the cloud-side or edge-side aggregated model sent by the parameter server as new preset model parameters to update each cloud-side or edge-side model parameter, and return to execute the step of performing several rounds of local model training on the cloud-side or edge-side model stored in each cloud server using the sample data of each party, and stop updating each model when the updated cloud-side or edge-side model or each aggregated model reaches a convergence condition. The first processing unit is further configured to use the final updated cloud-side or edge-side model as a target model for identifying defects of the to-be-tested end-side device through data collected by the to-be-tested end-side device from coarse to fine, and the receiver of the final updated cloud-side or edge-side model includes an edge computing platform of each cloud server. The first processing unit is specifically configured to: perform preliminary screening and identification on the initial sample data collected by the end-side device using the edge-side model to filter out defect data with a confidence higher than a preset confidence for fine screening by the cloud-side model; wherein the preliminary screening and identification is performed by the edge-side model stored in the edge computing platform, and the edge computing platform is in communication connection with the cloud server and the end-side device.

7. A parameter server, characterized in that, comprises: a second processing unit and a second sending unit; The second sending unit is configured to send preset model parameters to each participating cloud server, and the preset model parameters are used to configure the cloud-side or edge-side model stored in each cloud server. The second processing unit is configured to aggregate cloud-side or edge-side model update parameters obtained by each cloud-side or edge-side model according to power sample data to obtain update parameters of a cloud-side or edge-side aggregated model as new preset model parameters, and send the update parameters to each cloud server, so that each cloud server performs the next round of model training on the latest cloud-side or edge-side model according to the sample data and the new preset model parameters; the cloud-side or edge-side aggregated model refers to a global model obtained by integrating each cloud-side or edge-side model. The second processing unit is further configured to return to execute the step of aggregating the cloud-side or edge-side model update parameters obtained by each cloud-side or edge-side model according to sample data, and stop updating each model when the updated cloud-side or edge-side model or each aggregated model reaches a convergence condition, so as to use the final updated cloud-side or edge-side model as a target model for identifying defects of the corresponding end-side device. The cloud server executes the method of claim 1.

8. An identification device, characterized in that comprises: an input unit; The input unit is configured to input data collected by an end-side device into a corresponding target model to obtain device defect information contained in power data collected by the end-side device; The target model is obtained by the method of any one of claims 1 to 4, and the defect information includes at least one of information about whether a device state is abnormal, an abnormality level, a defect type, and a defect location.

9. An electronic device, comprising: Comprising: a central processing unit, a memory, and an input / output interface; the memory is a volatile memory or a persistent memory; the central processing unit is configured to communicate with the memory and execute instruction operations in the memory to perform the method of any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that, comprising instructions that, when executed on a computer, cause the computer to perform the method of any one of claims 1 to 5.

Images