Unmanned aerial vehicle nest cooperative operation method and system based on personalized federated learning
By using personalized federated learning for scene clustering and feature weight calculation, the problems of declining model generalization ability and data privacy in UAV nest collaborative operation and maintenance are solved, achieving efficient and accurate UAV nest collaborative operation and maintenance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA TOWER CO LTD
- Filing Date
- 2026-03-03
- Publication Date
- 2026-07-21
AI Technical Summary
In the scenario of large-scale networking of drone nests, traditional centralized machine learning leads to the global model overfitting to high-frequency distributions, resulting in a decline in generalization ability. Existing methods suffer from high communication overhead, high risk of data leakage, poor compliance, and a lack of systematic solutions for drone nest scenarios.
By employing a personalized federated learning approach, scene clustering and feature weight calculation are performed to generate weighted training data. Scene-specific basic parameters are trained on the edge server side and federated aggregation is performed in the cloud. Finally, personalized parameter updates are performed on the client side to achieve a dynamic balance between global and local data.
It improves the intelligence level and system robustness of drone nest collaborative operation and maintenance, solves the model adaptability problem under Non-IID data distribution, and ensures data privacy and model accuracy.
Smart Images
Figure CN122434480A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of edge computing and cloud computing collaborative technology, and specifically relates to a method and system for collaborative operation and maintenance of unmanned aerial vehicle (UAV) nests based on personalized federated learning. Background Technology
[0002] Currently, in large-scale drone nest networking scenarios, the data from each nest exhibits typical non-independent and identically distributed (Non-IID) characteristics. Traditional centralized machine learning aggregates heterogeneous data to a central cloud platform for unified modeling, which easily leads to the global model overfitting to high-frequency distributions. In long-tail or edge scenarios, the generalization ability is severely reduced, affecting the accuracy of operation and maintenance tasks.
[0003] Furthermore, existing methods such as data augmentation and transfer learning rely on cross-domain transmission of raw data, which suffers from drawbacks such as high communication overhead, high risk of data leakage, and poor compliance. While standard federated learning can achieve "model moving without data," it suffers from poor local adaptability of the global model in highly heterogeneous scenarios.
[0004] Existing personalized federated learning methods mostly focus on general algorithm optimization and lack systematic solutions for nested scenarios, especially feature selection mechanisms for scenario heterogeneity and a two-layer parameter update architecture that takes into account both global collaboration and local adaptation. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a method and system for collaborative operation and maintenance of unmanned aerial vehicle (UAV) nests based on personalized federated learning. This method can solve the problems of insufficient coverage of outdoor governance measures in specific areas such as high-rise buildings and the lack of targeted end-of-pipe remediation solutions, thereby achieving precise governance from source control to end-of-pipe closed-loop management. The method includes: Collect raw operation and maintenance data from multiple drone nests. The raw operation and maintenance data includes at least one sensor reading and corresponding scene label. Perform scene clustering processing on the raw operation and maintenance data to generate at least one scene cluster. Based on scenario clusters, calculate the feature weights of each operation and maintenance feature; use the feature weights to perform weighted fusion of client data within the scenario clusters to generate weighted training data; Based on weighted training data, scenario-specific basic parameters are generated by training on the edge server side; The scenario-specific basic parameters are uploaded to the cloud server for federated aggregation to generate global model parameters. The global model parameters are then distributed to the client, and the client updates the global model parameters based on local data to generate personalized parameters. The operation and maintenance prediction task of the drone nest is then executed based on the personalized parameters.
[0006] In this embodiment of the application, the original operation and maintenance data is subjected to scene clustering processing to generate at least one scene cluster, including: Normalize and preprocess the raw operation and maintenance data; calculate the mutual information value between each operation and maintenance feature and the scenario label, and generate a set of mutual information indicators; Based on the mutual information index set, using Clustering algorithms divide multiple drone nests into at least one scene cluster; The formula for calculating mutual information value is:
[0007] in, As an operational feature, Let x be the scene label, s be the current observation value of the operation and maintenance feature X, and s be the current scene value of the scene label S. For joint probability distribution, and This represents a marginal probability distribution.
[0008] In this embodiment of the application, based on scene clustering, the feature weights of each operation and maintenance feature are calculated, including: For each scenario cluster, calculate the information gain value of each operation and maintenance feature; The information gain values are normalized to obtain the normalized feature weights. The formula for calculating information gain is:
[0009] in, The information entropy of scene labels. Given operational characteristics Conditional entropy of scene labels under given conditions.
[0010] In this embodiment of the application, based on the weighted training data, scene-specific basic parameters are trained and generated on the edge server side, including: Weighted training data from multiple drone nests belonging to the same scene cluster are aggregated to the corresponding edge server; The deep learning model is trained on the edge server, and the model weights are updated by minimizing the cross-entropy loss function. The weights of the model after training convergence are determined as the scene-specific basic parameters of the scene cluster.
[0011] In this embodiment of the application, the cross-entropy loss function is:
[0012] in, For real labels, Predict labels for the model, For scene categories, The indicator variable for the true label. Predict probabilities for the model.
[0013] In this embodiment of the application, the scene-specific basic parameters are uploaded to a cloud server for federated aggregation to generate global model parameters, including: The cloud server receives scene-specific basic parameters from multiple edge servers; The received scene-specific basic parameters are weighted and aggregated according to the sample quantity ratio of the scene clusters corresponding to each edge server. Update global model parameters based on aggregated gradients; The update formula for the global model parameters is as follows:
[0014] in, Let be the weight matrix of the global model at layer l. The global learning rate, This represents the update amount of the global model weights. For scene categories.
[0015] In this embodiment, global model parameters are sent to the client, and the client updates the global model parameters based on local data to generate personalized parameters, including: The client receives global model parameters from the cloud server; During local training, a regularized loss function is introduced to constrain the global model parameters; The client's personalized parameters are updated by minimizing the enhanced loss function.
[0016] The enhanced loss function is:
[0017] in, For the original monitoring loss, This is the regularization balance coefficient. Personalize parameters for the client. These are global model parameters distributed from the cloud.
[0018] This application also provides a drone nest collaborative operation and maintenance system based on personalized federated learning, the system comprising: Multiple drone nest clients are used to collect raw operation and maintenance data from multiple drone nests. The raw operation and maintenance data includes at least one sensor reading and corresponding scene labels. Scene clustering processing is performed on the raw operation and maintenance data to generate at least one scene cluster. At least one edge server is used to calculate the feature weights of each operation and maintenance feature based on the scene cluster; use the feature weights to perform weighted fusion of client data within the scene cluster to generate weighted training data; and train and generate scene-specific basic parameters on the edge server side based on the weighted training data. A cloud server is used to upload the scenario-specific basic parameters to the cloud server for federated aggregation to generate global model parameters; the global model parameters are then distributed to the client, and the client updates the global model parameters based on local data to generate personalized parameters; and the operation and maintenance prediction task of the UAV nest is executed based on the personalized parameters.
[0019] In this embodiment of the application, the operation and maintenance prediction task includes at least one of equipment failure early warning, flight path optimization, and battery health assessment.
[0020] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method provided in the above embodiments.
[0021] After initial model training is completed on the client side, the system uploads the model parameters from each client to the cloud server for further optimization and integration. The cloud server coordinates and manages the training process across all clients to ensure model consistency and stability.
[0022] Meanwhile, the cloud server is also responsible for calculating the local loss function, evaluating the model's performance on various clients, and adjusting the model parameters based on the evaluation results to further optimize model performance.
[0023] This application introduces a local loss function, enabling the system to monitor and adjust the model training process in real time, thereby ensuring the stability and accuracy of the model in different scenarios.
[0024] Finally, the system returns the optimized model parameters to each client, which then uses these parameters for final predictions and decisions. The system also provides an output module that not only displays the model's prediction results and performance metrics but also helps users understand the model's performance and effectiveness.
[0025] In addition, this application also supports feedback and adjustment of personalized parameters, allowing users to fine-tune the model parameters according to actual needs and application scenarios, thereby further improving the applicability and flexibility of the model.
[0026] The UAV nest collaborative operation and maintenance method based on personalized federated learning provided in this application has a feature selection mechanism oriented towards scenario heterogeneity and a two-layer parameter update architecture that takes into account both global collaboration and local adaptation.
[0027] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a flowchart illustrating a method for collaborative operation and maintenance of unmanned aerial vehicle (UAV) nests based on personalized federated learning, provided in an embodiment of this application.
[0030] Figure 2 This is a schematic diagram illustrating the working process of a drone nest collaborative operation and maintenance system based on personalized federated learning, provided as an embodiment of this application. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0032] Currently, with the miniaturization and cost reduction of onboard sensors and the increasing maturity of autonomous flight control technology, drones have been widely used in urban security patrols, precision agriculture monitoring, power line inspections, and emergency logistics delivery. As ground infrastructure supporting the routine and swarm operations of drones, drone nests undertake core functions such as drone take-off and landing, automatic charging and battery swapping, task scheduling, and data transmission.
[0033] Currently, drone nests are evolving from single-point independent deployment to large-scale, distributed networking, forming a "nest cluster" operation mode covering a wide area.
[0034] However, in real-world large-scale deployment scenarios, nests in different geographical locations generally exhibit significant differences in environmental conditions, task types, and the distribution of collected data, which is to say, they show typical non-independent identically distributed (Non-IID) characteristics.
[0035] For example, the data collected by satellites deployed in urban CBD areas mainly consists of high-density population flow, high-rise building obstruction, and complex electromagnetic environment; while the data collected by satellites deployed in remote suburbs or forest fire prevention areas mainly consists of natural terrain, sparse targets, and weather changes.
[0036] If the traditional centralized machine learning paradigm is adopted, the heterogeneous data from drone nests in various locations will be forcibly aggregated into a central cloud platform for unified modeling. This can easily lead to the "global model" overfitting to high-frequency data distributions, resulting in a severe decrease in generalization ability in long-tail or edge scenarios. This, in turn, affects the safety of drone take-off and landing, the accuracy of fault warnings, and the efficiency of task scheduling.
[0037] To address the model bias issues caused by the aforementioned Non-IID data distribution, existing technical solutions mostly focus on data augmentation or transfer learning. For example, some methods augment sparse scene data using Generative Adversarial Networks (GANs) to balance the data distribution; or they transfer knowledge from the source domain to the target domain using domain adaptation techniques.
[0038] However, such methods often rely on centralized data processing, which not only places extremely high demands on communication bandwidth and storage resources, but more importantly, the cross-domain transmission of raw data poses serious risks of data privacy breaches and compliance issues. In scenarios involving sensitive data such as urban security facial recognition, precision agriculture plot information, or power grid facility coordinates, directly transmitting raw video streams or point cloud data not only fails to comply with the requirements of the Data Security Law and the Personal Information Protection Law, but also cannot meet the data sovereignty control needs of the respective data center owners.
[0039] Furthermore, federated learning, as an emerging distributed machine learning paradigm, allows each participant to retain the original data locally and only interact with model parameters or gradients, based on the principle of "data does not move, model moves," thus providing a privacy-preserving theoretical basis for multi-nest collaborative modeling.
[0040] However, standard federated learning algorithms (such as Federated Averaging, Fed Avg) assume that the data distribution of each client is approximately independent and identically distributed. When directly applied to highly heterogeneous nesting scenarios, the global model is poorly adapted to local clients, and may even experience problems such as local performance degradation and unstable convergence.
[0041] To address this issue, Personalized Federated Learning has become a research hotspot in recent years. Its core idea is to retain personalized model components for each client while sharing global knowledge. However, existing Personalized Federated Learning methods mostly focus on algorithm-level optimization and lack systematic solutions for UAV nest maintenance scenarios, especially feature selection mechanisms for scenario heterogeneity and a two-layer parameter update architecture that balances global collaboration and local adaptation.
[0042] In summary, the following technical challenges urgently need to be addressed in the field of collaborative operation and maintenance of drone nests: first, how to build prediction and decision-making models with scene-adaptive capabilities under highly heterogeneous and non-IID data distribution; and second, how to achieve efficient cross-domain knowledge fusion and model performance improvement while ensuring that data from each nest does not leave its local area.
[0043] To address this, this application proposes a UAV nest collaborative operation and maintenance method based on personalized federated learning. This method structurally divides heterogeneous scenarios through scenario clustering and feature variable weight distribution, and combines the personalized parameter update mechanism of deep learning models to achieve a dynamic balance between global information sharing and local task adaptation within the federated learning framework. This significantly improves the intelligence level and system robustness of multi-nest collaborative operation and maintenance.
[0044] Figure 1 A flowchart illustrating a UAV nest collaborative operation and maintenance method based on personalized federated learning, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes: S1. Collect raw operation and maintenance data from multiple drone nests. The raw operation and maintenance data includes at least one sensor reading and corresponding scene label. Perform scene clustering processing on the raw operation and maintenance data to generate at least one scene cluster.
[0045] Specifically, this application collects raw input data from different clients (e.g., client 1 to client m). This data may come from various application scenarios, such as sensor data, user behavior records, etc.
[0046] To improve the accuracy and efficiency of subsequent analysis, this application performs scene clustering on the collected data, grouping similar data into one category to form different scene datasets (e.g., scene 1 data to scene k data). This processing method helps reduce data complexity and provides a clearer data structure for subsequent feature extraction and model training.
[0047] S2. Based on scene clusters, calculate the feature weights of each operation and maintenance feature; use the feature weights to perform weighted fusion of client data within the scene clusters to generate weighted training data; based on the weighted training data, train and generate scene-specific basic parameters on the edge server side.
[0048] In this embodiment of the application, the data after scene clustering is further used for the calculation of feature variable weight distribution and the application of information gain algorithm.
[0049] It should be noted that feature variable weight distribution refers to assigning corresponding weights to each feature based on the importance and relevance of the data, in order to highlight the impact of key features on model training.
[0050] Information gain algorithm is used to evaluate the importance of features and select the optimal subset of features, thereby improving the predictive and generalization abilities of the model.
[0051] After feature engineering is completed, the system enters the deep learning model training phase. Multiple clients (client 1 to client m) train their models using local data, generating their respective model parameters.
[0052] The aforementioned model parameters include basic layer parameters and personalized parameters. Basic layer parameters are common parameters shared by all clients, while personalized parameters are optimized for each client's specific scenario and data characteristics. This configuration allows the system to maintain overall model consistency while fully considering the personalized needs and differences of each client, thereby improving the model's adaptability and accuracy.
[0053] S3. Upload the scenario-specific basic parameters to the cloud server for federated aggregation to generate global model parameters; distribute the global model parameters to the client, and update the global model parameters on the client based on local data to generate personalized parameters; execute the operation and maintenance prediction task of the drone nest based on the personalized parameters.
[0054] After the initial model training is completed on the client side, the system uploads the model parameters from each client to the cloud server for further optimization and integration.
[0055] The cloud server is responsible for coordinating and managing the training process of each client to ensure the consistency and stability of the model.
[0056] Meanwhile, the cloud server is also responsible for calculating the local loss function, evaluating the model's performance on various clients, and adjusting the model parameters based on the evaluation results to further optimize model performance.
[0057] This application introduces a local loss function, enabling the system to monitor and adjust the model training process in real time, thereby ensuring the stability and accuracy of the model in different scenarios.
[0058] Finally, the system returns the optimized model parameters to each client, which then uses these parameters for final predictions and decisions. The system also provides an output module that not only displays the model's prediction results and performance metrics but also helps users understand the model's performance and effectiveness.
[0059] In addition, this application also supports feedback and adjustment of personalized parameters, allowing users to fine-tune the model parameters according to actual needs and application scenarios, thereby further improving the applicability and flexibility of the model.
[0060] Based on the above embodiments, this application provides a complete distributed machine learning system architecture, including multiple stages such as data collection, feature engineering, model training, cloud optimization, and output feedback, aiming to achieve efficient, accurate, and flexible model training and prediction through deep learning and personalized parameter optimization.
[0061] The UAV nest collaborative operation and maintenance method based on personalized federated learning provided in this application has a feature selection mechanism oriented towards scenario heterogeneity and a two-layer parameter update architecture that takes into account both global collaboration and local adaptation.
[0062] In this embodiment of the application, the original operation and maintenance data is subjected to scene clustering processing to generate at least one scene cluster, including: S10. Perform normalization preprocessing on the original operation and maintenance data; calculate the mutual information value between each operation and maintenance feature and the scenario label, and generate a set of mutual information indicators.
[0063] S20. Based on the mutual information index set, the K-means clustering algorithm is used to divide multiple UAV nests into at least one scene cluster.
[0064] Furthermore, the formula for calculating the above mutual information value is as follows:
[0065] Where X is the operation and maintenance feature, S is the scene label, x is the current observation value of the operation and maintenance feature X, s is the current scene value of the scene label S, p(x, s) is the joint probability distribution, and p(x) and p(s) are the marginal probability distributions.
[0066] In this embodiment of the application, based on scene clustering, the feature weights of each operation and maintenance feature are calculated, including: S30. For each scenario cluster, calculate the information gain value of each operation and maintenance feature.
[0067] S40. Normalize the information gain value to obtain the normalized feature weights.
[0068] Furthermore, the formula for calculating the information gain value is as follows:
[0069] in, The information entropy of scene labels. Given operational characteristics Conditional entropy of scene labels under given conditions.
[0070] In this embodiment of the application, client data within scene clusters is weighted and fused using feature weights to generate weighted training data, including: The feature weights are combined with the original operation and maintenance data of the corresponding client to perform a Hadamard product operation, generating weighted and fused training data.
[0071] In this embodiment of the application, scene-specific basic parameters are generated on the edge server side based on weighted training data, including: S50: The weighted training data of multiple drone nests belonging to the same scene cluster are aggregated to the corresponding edge server.
[0072] S60. Train a deep learning model on an edge server and update the model weights by minimizing the cross-entropy loss function.
[0073] S70. Determine the model weights after training convergence as the scene-specific basic parameters for the scene cluster.
[0074] In this embodiment of the application, the cross-entropy loss function is:
[0075] in, For real labels, Predict labels for the model, For scene categories, The indicator variable for the true label. Predict probabilities for the model.
[0076] In this embodiment of the application, scene-specific basic parameters are uploaded to a cloud server for federated aggregation to generate global model parameters, including the following steps S80-S100: The S80 cloud server receives scenario-specific basic parameters from multiple edge servers.
[0077] S90. Based on the sample quantity ratio of the scene-specific basic parameters received, perform weighted aggregation on the received parameters.
[0078] S100: Update global model parameters based on aggregated gradients.
[0079] The update formula for the global model parameters is as follows:
[0080] in, Let be the weight matrix of the global model at layer l. The global learning rate, This represents the update amount of the global model weights. For scene categories.
[0081] In this embodiment, global model parameters are sent to the client, and the client updates the global model parameters based on local data to generate personalized parameters, including: The client receives global model parameters from the cloud server; During local training, a regularized loss function is introduced to constrain the global model parameters; The client's personalized parameters are updated by minimizing the enhanced loss function.
[0082] The enhanced loss function is:
[0083] in, For the original monitoring loss, This is the regularization balance coefficient. Personalize parameters for the client. These are global model parameters distributed from the cloud.
[0084] Figure 2 A schematic diagram illustrating the working process of a drone nest collaborative operation and maintenance system based on personalized federated learning, as provided in this application embodiment, is shown below. Figure 2 As shown, the system includes: multiple drone nest clients, at least one edge server, and a cloud server.
[0085] Specifically, multiple drone nest clients are used to collect raw operation and maintenance data from multiple drone nests. The raw operation and maintenance data includes at least one sensor reading and corresponding scene labels. Scene clustering processing is performed on the raw operation and maintenance data to generate at least one scene cluster.
[0086] At least one edge server is used to calculate the feature weights of each operation and maintenance feature based on scene clusters; to use the feature weights to perform weighted fusion of client data within the scene clusters to generate weighted training data; and to train and generate scene-specific basic parameters on the edge server side based on the weighted training data.
[0087] The cloud server is used to upload scenario-specific basic parameters to the cloud server for federated aggregation to generate global model parameters; the global model parameters are then distributed to the client, and the client updates the global model parameters based on local data to generate personalized parameters; and the operation and maintenance prediction tasks of the drone nest are executed based on the personalized parameters.
[0088] Furthermore, in this application, the operation and maintenance prediction task includes at least one of equipment failure early warning, flight path optimization, and battery health assessment.
[0089] Specifically, in this embodiment of the application, the collection of raw operation and maintenance data includes: from each drone nest (client). Data table collection The raw data for the scenario. This raw data includes sensor readings, flight logs, environmental parameters, and other data, covering a wealth of information for different scenarios.
[0090] It should be noted that the data from each nest is labeled as coming from a specific scenario (such as city, mountainous area, suburbs, etc.) to form multiple independent datasets to represent different application scenarios.
[0091] The following is a specific example to illustrate this application: Table 1. Data Types of Raw Operation and Maintenance Data Collected
[0092] Based on the raw operation and maintenance data collected from the above different scenarios, this application performs clustering processing on each scenario according to the dependency of features.
[0093] First, the raw operation and maintenance data is preprocessed to ensure that data of different magnitudes do not affect the clustering results. For example, the numerical ranges of temperature (°C) and humidity (%) vary greatly, so normalization is required.
[0094] In some embodiments, the normalization expression is: (Equation 1) Where x represents the original data; x' represents the normalized value.
[0095] This application evaluates the ability of each sensor feature to depend on the "scene category" based on mutual information.
[0096] Calculate the mutual information I(X;S) between each feature X and the scene label S. The larger the value, the more important the feature is in distinguishing different scenes. The expression for the mutual information I(X;S) is: (Equation 2) Based on the above embodiments, this application provides three scenario examples to explain the solution process of Formula 2. Taking temperature calculation as an example: (Equation 3) Feature X=TMP, after discretization, temperature is set as (10 temperature ranges). Table 2 shows a cross-tabulation of the frequency of temperature characteristics (TMP) and scene categories (S1-S5), as shown in Table 2 below: Table 2. Cross-tabulation of frequency of temperature characteristics (TMP) and scene categories (S1-S5)
[0097] Based on the above embodiments, the joint probability in the temperature range T1 (-20℃ to -10℃) in the first scenario S1 is: (Equation 4) Based on the above embodiments, the formulas for calculating the marginal probability are as follows: (Equation 5) (Equation 6) Based on the above calculations, the mutual information of the first scenario in the first temperature range is represented as follows: (Equation 7) Similarly, mutual information is calculated for each indicator (collected data such as temperature, humidity, and air pressure) in each scenario's numerical range, and the calculated mutual information of each indicator is input into the k-means clustering algorithm.
[0098] The algorithm flow of k-means clustering includes: initializing the number of clusters k and randomly selecting k data points as centroids; assigning points to the nearest centroids; for each data point, calculating its distance to each centroid (e.g., Euclidean distance); and assigning each data point to the cluster corresponding to the nearest centroid.
[0099] The Euclidean distance is calculated as follows: (Equation 8) Where I represents the processed mutual information index, This represents the value of a data point in the i-th feature dimension. Let j represent the j-th centroid. Let represent the value of the j-th centroid on the i-th feature dimension, and n represent the number of feature dimensions.
[0100] Next, update the centroid position. For each cluster, recalculate the centroid position as the average of all data points within that cluster.
[0101] (Equation 9) in, This represents the set of all data points belonging to the j-th cluster.
[0102] Finally, iterate until convergence, repeating steps 2 and 3 until the centroid no longer changes or the maximum number of iterations is reached.
[0103] After completing the scene clustering process, the information gain is calculated for the data of each cluster scene. The purpose of information gain is to enhance important information features, thereby increasing the decisive role of important features in the results.
[0104] This application performs mutual information calculation on the above-mentioned collected data, adopts cluster analysis, and then calculates the information gain for each clustered data, thereby identifying which features play a decisive role in the scene classification results in a certain type of clustered data.
[0105] For example, in one example of this application, we assume three scenarios (S1, S2, S3) and three temperature ranges (T1, T2, T3). Based on the above embodiments, mutual information is calculated on the collected data, and the calculated mutual information data is shown in Table 3 below: Table 3 Mutual Information Data
[0106] Hypothetical clusters The CCP There are m samples from m original scenes (S1~S5), with the number of samples in each scene being... .
[0107] (Equation 10) Based on the above embodiments, the conditional entropy H(S|I) is calculated for each feature I.
[0108] Taking temperature as an example, within the range (T1~T10), the number of samples in each interval is... Then the number of elements belonging to scene s is: .
[0109] (Equation 11) in, , ,
[0110] Next, calculate the information gain: (Equation 12) Sort all features in descending order according to IG(S, I) to obtain the feature in the current cluster. The most important features are listed in Table 4, which is a ranking table of the importance of all features in the above embodiments. Table 4. Ranking of Feature Importance
[0111] Therefore, we can normalize the information gain G: (Equation 13) Similarly, other indicators are calculated using the same method. Therefore, the weight distribution of the variables can be expressed as: , , wait.
[0112] The above data is then weighted and fused with the client data to obtain the training set: (Equation 14) in, express In the first scenario of each cluster, the information gain G is normalized, and weighted fusion of data between clients is achieved through Equation 12. This represents the Hadamard product operator.
[0113] This application uses weighted and fused data to perform distributed deep learning according to scenarios to construct basic layer parameters: Assuming each drone nest (client) locally stores the data of a scene, the data obtained after processing by Equation 14 is as follows: , , (Equation 15) in , is the value of the k-th scene in the i-th cluster in Formula 14. All drone nest data categorized into the k-th scene are trained on an edge server. The edge server then updates the data by weighting all drone nest data categorized into the k-th scene, using the cross-entropy loss function through deep learning. (Equation 16) Where y represents the real label; The label represents the model's predicted label, and k represents the scene category. The indicator variable for the true label. Predict probabilities for the model.
[0114] It should be noted that Equation 16 is the server-weighted data of all scenarios classified into k categories, so the predicted values of all k categories of scenarios can be taken.
[0115] Based on Equation 16, SGD (Stochastic Gradient Descent) is used to obtain the basic parameters classified into the k-th scene: (Equation 17) in, For learning rate, This represents the connection weight matrix of the l-th layer. This represents the gradient of the loss function with respect to the weights.
[0116] By up to Equation 17, we can obtain the basic parameters for k types of scenarios. These basic parameters are used to interact with the cloud server to achieve federated optimization.
[0117] In this embodiment of the application, the federated optimization process includes: Will Uploaded to the cloud server, aggregated on the cloud server, and the gradient of the loss function for each edge server is calculated for the k-th type of scene data: (Equation 18) in, is the number of samples of class k on the p-th edge server; n is the total number of samples. The parameters are allocated according to the proportion of data volume to ensure fairness in the aggregation of parameters on the cloud server.
[0118] After aggregating the gradients of participating clients, the cloud server uses the aggregated gradients to update the global model parameters. (Equation 19) in, This represents the connection weight matrix of the l-th layer, representing the global model parameters. This represents the global learning rate, which is the learning rate of the global model parameters. This indicates the amount of global model weight updates. Indicates the scene category.
[0119] Based on the above embodiments, the updated global model parameters are sent to the client for the next round of training.
[0120] Local training is performed based on global model parameters combined with the local samples to obtain personalized parameters. Regularization is used during local training to constrain the global parameters, thereby enhancing the loss function. (Equation 20) in, Let be the regularization loss function. In Equation 20, is used to represent the regularization loss of the k-th layer, which considers not only the original loss. Regularization terms were also added. To prevent overfitting.
[0121] This is called the original loss, also known as the standard supervised loss (such as cross-entropy CE or mean squared error MSE). The balancing coefficient is a constant greater than zero used to control the strength of the regularization term. This is the connection weight matrix of the k-th class and l-th layer of a certain edge server, also known as the personalized parameter.
[0122] In summary, personalized parameters can be calculated. This parameter can be fine-tuned according to the scenario, thereby enabling personalized prediction in the field of drone nest maintenance. It can solve the problems of data silos and privacy protection, address the challenges of non-independent identically distributed (Non-IID) data, and achieve efficient collaboration and resource optimization through cloud-edge collaboration, thus promoting the transformation of drone maintenance from "passive maintenance" to "proactive prediction".
[0123] This application introduces a scene clustering and feature weight adaptive mechanism, uses mutual information to evaluate the scene discrimination ability of each sensor feature, and combines K-means clustering to group non-independent identically distributed (Non-IID) data into scenes, thereby effectively reducing the impact of data heterogeneity.
[0124] This application also constructs a feature selection and weighted fusion strategy based on information gain, calculates the information gain of each feature within each cluster and normalizes it into weights, thereby enhancing key operation and maintenance features and weighted fusion of client data, improving the quality of model input and generalization ability.
[0125] In addition, this application designs a two-layer parameter update architecture that is collaborative between cloud, edge, and client. The edge side trains scenario-specific basic parameters, the cloud aggregates the global model, and the client then fine-tunes personalized parameters based on local data. Regularization constraints are introduced to prevent overfitting, thus balancing the model's generality and local adaptability.
[0126] Based on the above embodiments, this application comprehensively protects data privacy and system scalability, adopts a federated learning framework to avoid raw data transmission, achieves multi-nest joint optimization without sharing sensitive information, supports dynamic access to new nests, is applicable to large-scale city-level drone operation and maintenance networks, and promotes the evolution of operation and maintenance mode from "passive response" to "proactive prediction".
[0127] Based on the above embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described in the above embodiments.
[0128] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0129] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for cooperative operation and maintenance of a UAV nest based on personalized federated learning, characterized in that, The method includes: Collect raw operation and maintenance data from multiple drone nests, the raw operation and maintenance data including at least one sensor reading and corresponding scene labels; perform scene clustering processing on the raw operation and maintenance data to generate at least one scene cluster; Based on the scenario clusters, calculate the feature weights of each operation and maintenance feature; use the feature weights to perform weighted fusion of client data within the scenario clusters to generate weighted training data; Based on the weighted training data, scene-specific basic parameters are trained and generated on the edge server side; The scenario-specific basic parameters are uploaded to the cloud server for federated aggregation to generate global model parameters; the global model parameters are then distributed to the client, and the client updates the global model parameters based on local data to generate personalized parameters; and the operation and maintenance prediction task of the UAV nest is executed based on the personalized parameters.
2. The method according to claim 1, characterized in that, The step of performing scenario clustering processing on the original operation and maintenance data to generate at least one scenario cluster includes: The original operation and maintenance data is preprocessed by normalization; the mutual information value between each operation and maintenance feature and the scene label is calculated to generate a set of mutual information indicators; Based on the mutual information index set, the K-means clustering algorithm is used to divide multiple UAV nests into at least one scene cluster; The formula for calculating the mutual information value is: in, As an operational feature, Let x be the scene label, s be the current observation value of the operation and maintenance feature X, and s be the current scene value of the scene label S. For joint probability distribution, and This represents a marginal probability distribution.
3. The method according to claim 1, characterized in that, The calculation of feature weights for each operation and maintenance feature based on the scenario clustering includes: For each scenario cluster, calculate the information gain value of each operation and maintenance feature; The information gain value is normalized to obtain the normalized feature weights; The formula for calculating the information gain value is: in, The information entropy of scene labels. Given operational characteristics Conditional entropy of scene labels under given conditions.
4. The method according to claim 1, characterized in that, Based on the weighted training data, scene-specific basic parameters are generated on the edge server side, including: Weighted training data from multiple drone nests belonging to the same scene cluster are aggregated to the corresponding edge server; A deep learning model is trained on the edge server, and the model weights are updated by minimizing the cross-entropy loss function. The weights of the model after training convergence are determined as the scene-specific basic parameters of the scene cluster.
5. The method according to claim 4, characterized in that, The cross-entropy loss function is: Where y represents the real label. Predict labels for the model, For scene categories, The indicator variable for the true label. Predict probabilities for the model.
6. The method according to claim 1, characterized in that, The step of uploading the scene-specific basic parameters to the cloud server for federated aggregation to generate global model parameters includes: The cloud server receives scene-specific basic parameters from multiple edge servers; The received scene-specific basic parameters are weighted and aggregated according to the sample quantity ratio of the scene clusters corresponding to each edge server. Update global model parameters based on aggregated gradients; The update formula for the global model parameters is as follows: in, Let be the weight matrix of the global model at layer l. The global learning rate, This represents the update amount of the global model weights. For scene categories.
7. The method according to claim 1, characterized in that, The step of sending the global model parameters to the client and updating the global model parameters on the client based on local data to generate personalized parameters includes: The client receives global model parameters from the cloud server; During local training, a regularized loss function is introduced to constrain the global model parameters; The client's personalized parameters are updated by minimizing the enhanced loss function; The enhanced loss function is as follows: in, The original monitoring loss; This is the regularization balance coefficient; Customize parameters for the client; These are global model parameters distributed from the cloud.
8. A drone nest collaborative operation and maintenance system based on personalized federated learning, characterized in that, The system includes: Multiple drone nest clients are used to collect raw operation and maintenance data from multiple drone nests. The raw operation and maintenance data includes at least one sensor reading and corresponding scene labels. Scene clustering processing is performed on the raw operation and maintenance data to generate at least one scene cluster. At least one edge server is used to calculate the feature weights of each operation and maintenance feature based on the scene cluster; use the feature weights to perform weighted fusion of client data within the scene cluster to generate weighted training data; and train and generate scene-specific basic parameters on the edge server side based on the weighted training data. A cloud server is used to upload the scenario-specific basic parameters to the cloud server for federated aggregation to generate global model parameters; the global model parameters are then distributed to the client, and the client updates the global model parameters based on local data to generate personalized parameters; and the operation and maintenance prediction task of the UAV nest is executed based on the personalized parameters.
9. The system according to claim 8, characterized in that, The operation and maintenance prediction tasks include at least one of equipment failure early warning, flight path optimization, and battery health assessment.
10. 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 steps of the method as described in any one of claims 1 to 7.