Federal learning-based distribution network unmanned aerial vehicle distributed fault identification model collaborative training and security updating method and system

Abstract

The application discloses a kind of based on federal learning's distribution of network unmanned vehicle distributed fault identification model collaborative training and safe updating method and system.The method includes: multiple edge nodes utilize local multi-modal inspection data, train local model by based on positive and negative sample cross-modal contrast distillation network and calculate local class feature prototype;Node is handled by layered processing and is combined with adaptive compression after uploading model information using homomorphic encryption and differential privacy noise;Center server decrypts correction and calculates global mean and variance prototype, combined with dynamic client selection and proximal term constraint, realize fault aggregation by hierarchical prototype alignment to generate global model;Node receives model, under the double restraint of global prototype, combined with time series attention mechanism to carry out individualized fine-tuning, obtain the identification model adapted to local environment and fault evolution trend;For fault identification and trend assessment.The application realizes efficient, robust, individualized cross-domain fault diagnosis under strict privacy protection.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and power system technology, specifically relating to a collaborative training and security update method and system for a distributed fault identification model of a distribution network drone based on federated learning. Background Technology

[0002] With the rapid development of smart grids, the scale and complexity of distribution network equipment continue to increase, posing severe challenges to its safe and stable operation. Traditional centralized fault diagnosis methods rely on transmitting massive amounts of inspection data back to a data center for unified processing and analysis. However, this method has significant drawbacks in practical applications: First, the volume of multimodal data collected by distribution network drones, such as visible light images, infrared thermal images, and partial discharge signals, is enormous, and long-distance transmission faces bandwidth bottlenecks and high latency issues, making it difficult to meet real-time requirements; second, the raw inspection data contains sensitive grid topology and operating status information, and direct transmission and centralized storage pose serious privacy and security risks; finally, the aging patterns and fault characteristics of equipment vary significantly in different regions (such as urban and rural areas) and environments, resulting in insufficient generalization ability of models trained based on globally unified data, and decreased diagnostic accuracy in specific local scenarios.

[0003] Federated learning, as an emerging distributed machine learning paradigm, offers a potential path to solving the aforementioned problems with its core idea of ​​"data remains still, model moves." This paradigm allows each edge node to train a model locally using private data, exchanging only model parameters rather than the original data, thus theoretically balancing data privacy and collaborative efficiency. However, directly applying federated learning to multimodal fault diagnosis in distribution networks still faces a series of core technical challenges: First, the challenge of modal differences and fusion: inspection data with different physical characteristics are inconsistently distributed in the feature space, and simple parameter averaging cannot achieve effective cross-modal knowledge fusion and alignment; Second, the challenge of balancing privacy and communication efficiency: while protecting the privacy of model gradients, the communication overhead problem under unstable field networks must be addressed; Third, the challenge of data heterogeneity and malicious attacks: the data of each node is not independent and identically distributed and may contain outliers, which can affect traditional federated averaging algorithms, leading to decreased model performance or convergence to a suboptimal solution; Fourth, the challenge of the contradiction between personalization and generalization: a single global model is difficult to adapt to the local features of all nodes, while excessive personalization weakens the transferability between models.

[0004] Therefore, there is an urgent need for a new federated learning method that can deeply integrate multimodal information, ensure strict privacy and security, resist heterogeneous data and malicious interference, and achieve personalized and accurate diagnosis, so as to support the evolution of power distribution network drone inspection systems towards intelligence, autonomy, and security. Summary of the Invention

[0005] This invention aims to overcome the problems faced by existing technologies in multimodal fault diagnosis of distribution networks, such as privacy leakage risks, low communication efficiency, difficulties in cross-modal fusion, the impact of data heterogeneity, and insufficient model personalization. To this end, this invention provides a method and system for collaborative training and secure updating of a distributed fault identification model for distribution networks using unmanned aerial vehicles (UAVs) based on federated learning.

[0006] In a first aspect, embodiments of this application provide a method for collaborative training and secure updating of a distributed fault identification model for distribution network drones based on federated learning, the method comprising: Multiple edge nodes utilize locally collected multimodal inspection data of power equipment to train local fault identification models through a cross-modal comparative distillation network based on positive and negative samples, and calculate local category feature prototypes based on the distilled feature space. The edge nodes will upload the trained model parameters, model update information and local category feature prototypes to the central server after using homomorphic encryption and differential privacy noise injection for layered processing, and adaptive compression combined with communication bandwidth. After performing homomorphic decryption and differential privacy noise correction on the encrypted information, the central server calculates the global mean prototype and global variance prototype based on the local category feature prototype. Combined with dynamic client selection and near-term constraints, fine-grained fault-tolerant aggregation is achieved through hierarchical prototype alignment to generate a global model. The global model and global mean prototype are distributed to each edge node. Each node, based on local multimodal inspection data, performs personalized fine-tuning of the global model under the dual constraints of the global mean prototype and the global variance prototype, combined with the temporal attention mechanism, to obtain a fault identification model that adapts to the local environment and fault evolution trend. The fine-tuned model is used to identify faults and assess trends in power equipment, and the identification results and fault evolution trajectories are synchronized to the power equipment health management system.

[0007] Secondly, embodiments of this application provide a collaborative training and security update system for a distributed fault identification model of a distribution network drone based on federated learning, applied to the method described in the first aspect, the system comprising: Multiple edge node devices are deployed in various inspection drones or local stations to collect multimodal inspection data of power equipment and perform local model training, feature prototype calculation, privacy protection processing, and data compression and uploading. The central server device is communicatively connected to the multiple edge node devices and is used to decrypt and correct noise for received encrypted information, perform global prototype calculation, dynamic client selection, model aggregation, and generate and distribute global models. A secure communication module is integrated into the edge node device and the central server device to realize layered secure communication based on homomorphic encryption and differential privacy, as well as adaptive data compression transmission based on communication bandwidth. The personalized fine-tuning module is deployed in each of the aforementioned edge node devices and is used to perform personalized fine-tuning of the model after receiving the global model, combining local time series data and global prototype constraints. The health management system interface module, deployed in the edge node device or central server device, is used to encapsulate and synchronize the fault identification and trend assessment results of the fine-tuned model to the power equipment health management system.

[0008] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is configured to implement, when executing the instructions, the collaborative training and security update method for the distributed fault identification model of the distribution network drone based on federated learning as described in the first aspect.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the collaborative training and security update method for a distributed fault identification model for distribution network drones based on federated learning as described in the first aspect.

[0010] The beneficial effects of this invention are mainly reflected in the following aspects: By using a cross-modal distillation network based on contrastive learning, it effectively aligns features of different modalities, extracts more discriminative cross-modal invariant features, and improves the model's ability to identify complex fault modes. A layered security framework integrating homomorphic encryption and differential privacy achieves strong privacy protection at the algorithm level; combined with bandwidth-aware adaptive compression, it significantly reduces the communication overhead of federated learning in constrained network environments. The introduction of a hierarchical alignment mechanism based on mean and variance prototypes, combined with dynamic client selection and proximate term constraints, effectively suppresses the negative impacts of data heterogeneity, node failures, or malicious attacks, ensuring the convergence stability and performance of the global model. The introduction of dual prototype constraints and temporal attention during the fine-tuning stage enables the model to absorb global knowledge, finely adapt to local data distribution, and possess the ability to model the temporal evolution of faults, achieving accurate personalized diagnosis and trend prediction. Seamless integration of diagnostic results into the health management system enables complete empowerment from intelligent perception to intelligent decision-making, improving the intelligence level and preventative maintenance capabilities of distribution network operation and maintenance. Attached Figure Description

[0011] Figure 1This is a schematic diagram of a collaborative training and security update method for a distributed fault identification model for power distribution network drones based on federated learning, provided as an embodiment of this application.

[0012] Figure 2 The system architecture diagram for collaborative training and secure update of the distributed fault identification model for distribution network drones based on federated learning provided for this application.

[0013] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] 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 a part of the embodiments of this application, and not all of them.

[0015] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. 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 in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0016] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] Example 1

[0018] Figure 1 This is a schematic flowchart illustrating a method for collaborative training and security update of a distributed fault identification model for power distribution network drones based on federated learning, provided as an embodiment of this application. Figure 1 As shown, a collaborative training and security update method for a distributed fault identification model for distribution network drones based on federated learning includes: S1. Multiple edge nodes utilize locally collected multimodal inspection data of power equipment to train local fault identification models through a cross-modal comparative distillation network based on positive and negative samples, and calculate local category feature prototypes based on the distilled feature space. This step is the starting point of federated learning and includes two core tasks: model training and feature prototyping. Each node first trains its local fault identification model based on local multimodal data through a cross-modal comparative distillation network; then, it abstracts the knowledge learned by the model into local category feature prototypes for various faults, providing comparable and aggregateable knowledge units for subsequent federated collaboration.

[0019] Specifically, in this embodiment, step S1 involves training a local fault identification model using a cross-modal contrastive distillation network based on positive and negative samples, specifically including: (S1-1) Construct positive sample pairs from inspection data of at least two different modes collected from the same power equipment at the same time.

[0020] During implementation, the system controls a multi-sensor-equipped inspection drone to perform targeted synchronous data collection on designated power equipment (such as the circuit breaker numbered DL-1107). Specifically, the visible light camera and infrared thermal imager on the drone are triggered by a hardware synchronization signal to acquire a high-resolution visible light image and an infrared thermal image respectively within the same millisecond-level timestamp, forming a one-to-one corresponding data pair. The system then labels this set of data with the same timestamp, device ID, and health status tag, and packages and registers it as a positive sample pair. The principle behind this is that although the visible light image reflects the appearance and texture, and the infrared image reflects the temperature distribution, they together characterize the true physical state of the device at this specific moment, providing a benchmark for subsequent model learning that describes the same fact through different senses.

[0021] (S1-2) Construct negative sample pairs from multimodal inspection data of different power equipment or different fault categories.

[0022] In its implementation, this step involves automatically cross-combining multimodal data in the database according to a preset strategy. The first combination method is heterogeneous pairing of equipment: for example, randomly selecting a visible light image of circuit breaker DL-1107 and pairing it with an infrared thermal image of disconnector GL-2043. The second combination method is heterogeneous pairing of faults: for example, pairing a visible light image marked as a broken conductor strand with an infrared thermal image marked as surface discharge on an insulator. All of these pairings are automatically labeled as negative sample pairs by the system. The core principle is that by actively constructing these samples with mismatched sensory information, the model is forced to learn to distinguish the semantic consistency of cross-modal information, thereby enhancing its ability to reject erroneous associations.

[0023] (S1-3) Using the positive and negative sample pairs, the cross-modal contrastive distillation network is constrained by the contrastive learning loss function.

[0024] During model training, the system inputs batches of positive and negative sample pairs into a cross-modal contrastive distillation network with a dual-branch structure. This network contains two parallel deep feature extractors, one for visible light and one for infrared modalities. The core of the training process is optimizing an objective called the contrastive learning loss function. Contrastive learning is a training paradigm that teaches the model feature similarity. Its basic idea is that in the feature representation space within the model, features of similar samples should be close to each other, while features of dissimilar samples should be far apart. Specifically, in this embodiment, the loss function continuously calculates and feeds back two types of information: for each positive sample pair, it calculates the distance between the feature vectors output by the two branches and continuously shortens this distance through backpropagation; for each negative sample pair, it calculates the distance between their corresponding feature vectors and continuously increases this distance. After numerous iterations, the network gradually learns to map the same device states (regardless of modality) to neighboring regions in the feature space, while mapping different states to far apart regions, thus achieving deep fusion and discriminative alignment of cross-modal features.

[0025] The contrastive learning loss function of the cross-modal contrastive distillation network is used to train the local model, promoting the convergence of multimodal features with the same device state in the feature space. The contrastive learning loss function of the cross-modal contrastive distillation network is defined as follows: , in, This represents the number of positive sample pairs. The number of negative sample pairs and These are the similarity scores for positive and negative sample pairs, respectively. Temperature is a parameter that controls the smoothness of the similarity distribution.

[0026] Furthermore, the calculation of the local category feature prototype based on the distilled feature space in S1 specifically includes: (S1-4) Using the feature extractor of the trained cross-modal contrastive distillation network, forward propagation is performed on all local labeled samples to extract their output feature vectors in the fusion feature layer.

[0027] After the local model training converges, the edge nodes sequentially input all the labeled fault samples stored in the node into the cross-modal contrastive distillation network for processing. The network's feature extractor transforms each sample (e.g., a pair consisting of a visible light image and an infrared thermogram) into a fused feature vector, located at a specific position within the network called the fused feature layer. This vector is a unified digital representation formed by the network's high-level abstraction and fusion of the original multimodal data.

[0028] (S1-5) Based on the fault category label of the sample, perform cluster analysis on the feature vectors of all samples belonging to the same category.

[0029] The system groups all extracted feature vectors based on pre-labeled fault categories for each sample. For example, it groups the feature vectors of all samples labeled "insulator contamination" into one group. Then, it performs an operation called cluster analysis on each group of feature vectors. Cluster analysis is an unsupervised machine learning method that aims to discover the natural grouping structure of data in space, with the goal of finding the most core and representative positions within each group.

[0030] (S1-6) Calculate the feature vector of each cluster center and use the cluster center vector as the local feature prototype of the category.

[0031] For each fault category data group obtained through cluster analysis, the system calculates the mathematical center point of all feature vectors within it. The feature vector corresponding to this center point is called the cluster center. This cluster center vector essentially represents the most typical and prevalent feature pattern of a certain type of fault in the local data of that node. The system formally stores this vector and names it the local feature prototype for that type of fault. For example, the cluster center of all sample feature vectors of the above insulator pollution category will be recorded as the local feature prototype for insulator pollution of that node.

[0032] Specifically, local feature prototypes are computed to extract typical feature vectors for each fault category. For each fault category... Local feature prototype The calculation is as follows: , in, For category The sample set, For the sample The feature vector obtained by the feature extractor.

[0033] S2. The edge nodes upload the trained model parameters, model update information, and local category feature prototypes to the central server after layered processing using homomorphic encryption and differential privacy noise injection, combined with adaptive compression based on communication bandwidth. Secure compression and privacy-preserving upload are core steps in privacy protection and communication optimization, achieving federated communication where data is usable but invisible. The results from the previous step (model parameters, prototypes) undergo three layers of processing: strong encryption ensures that third parties cannot decrypt the content; differential privacy adds noise to obscure individual contributions; and adaptive compression compresses the data volume according to network conditions, ensuring transmission efficiency under unstable networks. Layered processing integrates comprehensive protection at the algorithm level (differential privacy) and the communication level (encryption, compression), securely extracting knowledge from nodes without leaking the original data.

[0034] Specifically, in this embodiment, step S2, which employs homomorphic encryption and differential privacy noise injection for layered processing, and then performs adaptive compression based on communication bandwidth before uploading to the central server, specifically includes: (S2-1) First Processing Layer: The model parameters, model update information, and local category feature prototypes are encrypted using a homomorphic encryption algorithm. Before the edge node prepares to upload data, the system first activates the security module and uses a cryptographic technique called homomorphic encryption to encrypt the data. Homomorphic encryption is a special encryption method that allows specific mathematical operations (such as addition and multiplication) to be performed on the ciphertext without decryption. The result of the operation, after decryption, is consistent with the result of performing the same operation on the plaintext. In specific implementation, the node uses the model parameters, model update information (such as gradients), and calculated local category feature prototypes obtained from this round of training as plaintext input and encrypts them using a preset homomorphic encryption public key. After this layer of processing, the original data is converted into unreadable ciphertext. Even if the ciphertext is intercepted during transmission, attackers cannot directly obtain useful information, but the central server can still perform secure aggregation operations without decryption.

[0035] (S2-2) Second Processing Layer: Injecting random noise that meets differential privacy requirements into the encrypted data. After homomorphic encryption, the system enters the second processing layer. This layer superimposes a carefully generated layer of special random noise onto the ciphertext data. This noise addition must comply with differential privacy requirements. Differential privacy is a strict privacy protection framework whose core idea is to add controllable randomness to data or computation results, making it impossible for attackers to reliably infer information about any specific individual based on the algorithm output. In this embodiment, the system generates noise conforming to a specific mathematical distribution (such as a Laplace distribution or a Gaussian distribution) according to preset privacy budget parameters and securely superimposes it onto the encrypted ciphertext data. This operation further protects the original data statistically, making it difficult to infer from it using reverse engineering attacks.

[0036] Specifically, noise injection is implemented using the Laplace mechanism: In the formula, This is the prototype of the local feature after adding noise; For Laplace noise generator; For function sensitivity (maximum norm of prototype vector change); To control noise levels for privacy budget parameters.

[0037] (S2-3) Third processing layer: Based on the real-time communication bandwidth of the current edge node, dynamically select gradient sparsification or quantization compression strategy to perform lossy compression on the data after injecting noise.

[0038] The third layer of processing focuses on optimizing communication efficiency. The system first monitors the network connection between the current node and the central server in real time to obtain its real-time communication bandwidth data. Based on the bandwidth conditions, the system dynamically selects a compression strategy. If the bandwidth is extremely low, a gradient sparsity strategy may be chosen. This strategy identifies and retains only the most important gradient values ​​in the model update information (e.g., the top 10% with the largest absolute values), while setting the remaining gradients to zero, thus significantly reducing the amount of data. If the bandwidth is acceptable but compression is still necessary, a quantization compression strategy may be chosen. This strategy reduces the number of bits occupied by each value by decreasing the precision of the data representation (e.g., converting 32-bit floating-point numbers to 8-bit integers). After selecting a strategy, the system performs appropriate lossy compression operations on the encrypted data processed by the first two layers to further reduce the size of the data packets to be transmitted.

[0039] (S2-4) Uploading the layered and compressed data to the central server: Finally, the node transmits the final data packet, processed in the above three layers (encryption, noise addition, and compression), to the central server via the wireless network channel. The uploaded data at this point is unreadable ciphertext in terms of content, protected by differential privacy in terms of statistical characteristics, and has been adaptively compressed according to network conditions in terms of size, thus achieving a comprehensive balance between privacy and security, communication efficiency, and subsequent processability.

[0040] S3. After homomorphic decryption and differential privacy noise correction of the encrypted information, the central server calculates the global mean prototype and global variance prototype based on the local category feature prototype. Combined with dynamic client selection and near-end constraint, it achieves fine-grained fault-tolerant aggregation through hierarchical prototype alignment to generate a global model. This central aggregation and fault-tolerant optimization form the brain of federated learning, responsible for filtering out false information and integrating intelligence. The central server executes: information restoration (decryption and noise correction), knowledge extraction (calculating the global mean prototype and global variance prototype from all node prototypes), and intelligent aggregation (hierarchical alignment and fault-tolerant filtering based on prototypes, combined with dynamic selection and near-end constraint to generate a global model). The global mean prototype represents the consensus center of knowledge across all nodes, while the global variance prototype reflects the differences in knowledge distribution among nodes. Hierarchical prototype alignment achieved using these two prototypes is the core of robust aggregation, enabling the robust fusion of high-quality consensus knowledge even in the presence of heterogeneous data or malicious nodes.

[0041] Specifically, in this embodiment, after the central server in S3 performs homomorphic decryption and differential privacy noise correction on the encrypted information, it calculates the global mean prototype and global variance prototype based on the local category feature prototype. Combined with dynamic client selection and near-end constraint, it achieves fine-grained fault-tolerant aggregation through hierarchical prototype alignment to generate a global model, specifically including: (S3-1) Homomorphically decrypt the received encrypted information and perform statistical correction on the noise introduced by differential privacy in the decrypted data.

[0042] After receiving encrypted data packets uploaded by each edge node, the central server first performs homomorphic decryption using the private key corresponding to the encryption, restoring the ciphertext to plaintext data containing model updates, parameters, and prototypes. Since random noise conforming to differential privacy requirements is added to this data at the edge nodes, denoising processing is necessary. The server performs statistical correction, which utilizes the known mathematical and statistical properties of differential privacy noise (e.g., the added Laplace noise has a mean of zero) to mathematically adjust the decrypted data. This aims to estimate and eliminate as much as possible the impact of noise on the expected value (mean) of the data, thereby restoring effective data information that is close to the original, noise-free state.

[0043] (S3-2) Based on the local category feature prototypes uploaded by each edge node, calculate the prototype mean of all nodes by category to obtain the global mean prototype; at the same time, calculate the variance of the prototype distribution of each category to obtain the global variance prototype.

[0044] The server aggregates all local feature prototypes uploaded by each node for the same fault category. For each fault category (such as insulator contamination), the server first calculates the mathematical mean of all relevant prototypes, resulting in a vector. This vector represents the average state or consensus center of this type of fault feature across the entire federated system, called the global mean prototype. Next, the server calculates the dispersion of these prototype data around this mean prototype, i.e., the statistical variance, resulting in a metric measuring the dispersion of the prototype distribution, called the global variance prototype. The global mean prototype indicates where knowledge should converge, while the global variance prototype characterizes the extent of the divergence of knowledge before convergence.

[0045] Specifically, the global mean prototype and variance prototype are calculated to aggregate the feature prototypes of all nodes on the server side. For fault categories... Global mean prototype With global variance prototype The calculation is as follows: , , in, For the number of nodes, For nodes Uploaded categories The local feature prototype.

[0046] (S3-3) Based on the computing resources, network status and historical contribution of each edge node, dynamically select the set of effective clients to participate in this round of aggregation.

[0047] Before each round of global model training begins, the server performs dynamic client selection. First, it evaluates the real-time status of each registered edge node: checking its remaining computing power (e.g., CPU, memory usage), current network connection quality (e.g., latency, packet loss rate), and considering the quality and stability of the model uploaded by that node in previous training rounds (historical contribution). Then, based on a comprehensive evaluation set, the server selects a subset of nodes that are currently best positioned to provide high-quality updates as the effective client set for this round of model aggregation. This prevents nodes with insufficient resources or poor network connections from slowing down the overall training process.

[0048] (S3-4) Introduce a proximal term into the objective function of model aggregation to constrain the distance between the local model update of each node and the previous round of global model.

[0049] When performing model parameter aggregation optimization, the server's optimization objective (objective function) includes a proximal term in addition to the usual loss term aimed at improving model performance. The proximal term is an extra mathematical constraint that measures the difference (distance) between the local model parameters obtained by each node after this round of training and the global model parameters distributed in the previous round. Including this term in the overall objective and minimizing it means that while seeking optimal performance for the new model, the server also aims to ensure that the update magnitude of each node does not deviate too far from the consensus model of the previous round. This helps control the direction of node updates, preventing drastic oscillations or divergences in the entire model training process due to abnormal or malicious updates from individual nodes, thereby enhancing the stability of federated training.

[0050] (S3-5) Based on the global mean prototype and the global variance prototype, the model update of the effective client set is subjected to fine-grained weighting and anomaly filtering through the hierarchical prototype alignment loss function to complete fault-tolerant aggregation and generate the current round global model.

[0051] This is the core step of the aggregation process. The server is designed and uses a hierarchical prototype alignment loss function to guide the aggregation process. This function first uses the calculated global mean prototype and global variance prototype as reference standards. The first layer of alignment is quality assessment: it calculates the similarity between the local feature prototype provided by each valid client and the global mean prototype. The higher the similarity, the more consistent the client's data is with the global consensus, and the higher the weight its model update receives during aggregation. The second layer of alignment is anomaly filtering: it checks the degree of matching between each client's local prototype and the global variance prototype. If a client's prototype distribution characteristics differ significantly from the global distribution characteristics (variance), exceeding a preset reasonable threshold, it will be judged as a potential anomaly (possibly due to extremely poor data quality or malicious attacks), and its update may be downweighted or even directly removed. Finally, based on the weights obtained from the first layer of alignment, the server performs a weighted average of the model updates of the clients that have passed the second layer of filtering, and incorporates the constraint of proximate terms to finally calculate the robust global model parameters after this round of updates.

[0052] Furthermore, in S3-5, the model update of the effective client set is subjected to fine-grained weighting and anomaly filtering through the hierarchical prototype alignment loss function, specifically including the following steps: (S3-5-1) First layer alignment: Calculate the cosine similarity between the local category feature prototype uploaded by each edge node and the global mean prototype to generate the first weight vector.

[0053] After obtaining the local feature prototypes and global mean prototypes from all clients, the server begins the first-level alignment. Specifically, for each client's uploaded local feature prototype for a specific fault category, the server calculates the cosine similarity between its vector and the corresponding global mean prototype vector. Cosine similarity is a metric that measures the similarity in direction between two vectors, with values ​​ranging from -1 to 1; a higher value indicates a more consistent direction. After calculation, each client receives a similarity score (or scores corresponding to multiple categories). Based on these scores, the server generates a first-level weight vector using a normalization function (such as the Softmax function). Each element in this vector corresponds to a client, and its value represents the client's initial relative importance in this round of updates; clients with higher scores will be assigned higher initial weights in subsequent aggregations. The weight vector for the first-level alignment... In the middle, node weight The calculation is as follows: .

[0054] in, Represents the cosine similarity function; This represents the normalization function, which converts similarity into a weighted distribution.

[0055] (S3-5-2) Second layer alignment: Calculate the distribution difference between each local category feature prototype and the global variance prototype, filter the prototypes based on the preset anomaly threshold, and remove abnormal local prototypes whose difference exceeds the threshold.

[0056] After obtaining the initial weights, the server performs a second, more rigorous alignment. The goal here is to identify potentially anomalous clients that are severely inconsistent with the overall data distribution characteristics. The server calculates the distribution dissimilarity between each local feature prototype and the global variance prototype. Distribution dissimilarity is a statistical distance measure used to quantify the degree of difference between the discrete patterns of two probability distributions or a set of data and their centroids. Specifically, this can be achieved by calculating the Mahalanobis distance of the prototype to the global distribution center, or by calculating the divergence between the local prototype set and the distribution described by the global variance (such as KL divergence). The system presets an anomaly threshold. If a client's distribution dissimilarity exceeds this threshold, the client's local prototype is determined to be severely inconsistent with the federated overall knowledge distribution, possibly due to extremely poor data quality, severe non-identical distribution, or malicious attack. The server marks the corresponding local prototype of that client as an anomaly and removes it from the valid candidates for this round of aggregation.

[0057] In the second layer of alignment, nodes Local prototype The following conditions must be met for the application to be considered valid: ,in, This is a distribution difference measure function. This is a preset abnormal threshold.

[0058] (S3-5-3) Aggregation execution: Based on the first weight vector, the model updates corresponding to the edge nodes after the second layer alignment filter are weighted and averaged to complete the aggregation generation of the global model.

[0059] After the initial two layers of alignment and weighting, the server obtains a filtered list of valid clients and their corresponding first weight vectors. In the final aggregation execution step, the server extracts the model updates (e.g., changes in model parameters or gradients) from all valid clients. Then, based on the weight proportions of these clients in the first weight vector, a weighted average is calculated for their model updates. Specifically, each client's update is first multiplied by its weight, then the weighted updates from all clients are summed together, and finally normalized. This weighted average result constitutes the final increment of the current round of federated learning iterations. The server applies this increment to the global model parameters from the previous round, thereby generating and outputting a new, fault-tolerant global model that integrates high-quality knowledge from multiple sources.

[0060] S4. The global model and global mean prototype are distributed to each edge node. Each node, based on its local multimodal inspection data and under the dual constraints of the global mean prototype and global variance prototype, performs personalized fine-tuning of the global model using a temporal attention mechanism to obtain a fault identification model adapted to the local environment and fault evolution trend. Personalized distribution and adaptive fine-tuning are key steps in achieving localized adaptation, completing global-local knowledge transfer and adaptation. After receiving the global model (consensus knowledge), nodes do not apply it directly. Instead, they use it as a starting point, guided by both the global mean prototype and the global variance prototype, and combined with their latest temporal data, to perform rapid fine-tuning. This ensures that the model maintains both the generalizability of the global consensus and the uniqueness of accurately adapting to the local environment. The dual prototype constraints prevent the fine-tuning process from deviating too far from the global consensus (aligning with the mean) while respecting reasonable local differences (matching variance). Simultaneously, the temporal attention mechanism enables the model to perceive the dynamic evolution of faults.

[0061] Specifically, in this embodiment, each node in S4, based on local multimodal inspection data, performs personalized fine-tuning of the global model under the dual constraints of the global mean prototype and the global variance prototype, combined with a temporal attention mechanism. This includes: (S4-1) First constraint stage: Using the global mean prototype, construct the prototype alignment loss to guide the features extracted by the node local model to move closer to the global feature center.

[0062] After receiving the global model and global mean prototype, edge nodes initiate personalized fine-tuning. In the first constraint phase, the nodes use the received global model as their initial state and train using local multimodal labeled data. To prevent the fine-tuned model from deviating excessively from the federated consensus, the nodes add a prototype alignment loss term to the training objective function. The core principle of this loss term is: for each local training sample, the node uses the current model to extract its features, and then calculates the difference (e.g., Euclidean distance) between this feature and the corresponding category's global mean prototype sent by the server. During training, by minimizing this difference through optimization algorithms (such as stochastic gradient descent), the local features extracted by the model can be gradually guided towards the global feature center (i.e., the global mean prototype) representing global consensus, thereby ensuring that the personalized model does not deviate from the correct path of global knowledge.

[0063] (S4-2) Second constraint stage: Using the global variance prototype, construct the distribution consistency loss to constrain the dispersion of the local feature distribution of the node to match the global distribution.

[0064] Simply conforming to the center is insufficient; the distribution of local data within a node may possess unique characteristics. In the second constraint phase, a distribution consistency loss is introduced as another constraint. This loss term aims to control the breadth or tightness of the local feature distribution, aligning it with the overall distribution pattern of the federation. Specifically, during training, the node calculates the statistical distribution characteristics of features extracted from a batch of local samples (e.g., calculating the variance of these features) and compares it to a global variance prototype issued by the server (representing the general dispersion of such features within the federation). The loss function measures the difference between these two. By minimizing this loss, nodes are encouraged to: appropriately broaden their distribution if local data features are concentrated (small variance) while the global distribution is broad (large variance); conversely, tighten their distribution to match the global distribution if local features are too dispersed. This ensures flexibility in knowledge transfer, allowing for reasonable local adaptation while avoiding severe distortion of the distribution pattern.

[0065] (S4-3) Temporal modeling stage: For the local multimodal inspection time series data of the node, use the temporal attention mechanism to extract the dependency relationship between features at different time steps and generate temporal enhanced features.

[0066] To capture dynamic changes in device status, nodes incorporate temporal information during fine-tuning. In the temporal modeling phase, nodes no longer process only snapshot data from a single inspection, but instead take as input a sequence of multimodal data acquired from multiple recent consecutive inspections (e.g., the past 5 inspections) of the same device. A model component called a temporal attention mechanism then comes into play. This mechanism can be understood as an intelligent information filter and integrator. It first analyzes the features extracted at each time step in the sequence, then automatically calculates and assigns different attention weights to different moments in the sequence. For example, the mechanism might assign higher weights to features from recent moments, or pay special attention to features from a historical moment that has undergone drastic changes. Finally, based on these weights, the mechanism selectively and weightedly fuses the features from different time steps to generate a temporally enhanced feature that implies time dependencies. This allows the model to perceive whether a fault is gradually forming or suddenly evolving.

[0067] (S4-4) Personalized fine-tuning stage: Based on the temporal enhancement features, under the joint constraints of the prototype alignment loss and the distribution consistency loss, the global model is trained locally for a limited number of rounds to obtain a personalized fault identification model adapted to the local environment and fault evolution trend.

[0068] Finally, the node integrates all components into a personalized fine-tuning phase. The training process begins with the received global model. For each local training batch, the node simultaneously performs three core operations: 1) Processing time-series data using the model, generating time-series augmented features, and performing classification or prediction, calculating the main task loss; 2) Calculating the difference between sample features and the global mean prototype, adding prototype alignment loss; 3) Calculating the difference between the batch feature distribution and the global variance prototype, adding distribution consistency loss. The node's total loss is the weighted sum of the above three losses. Through iterative training in a limited number of rounds (e.g., 10-20 rounds to avoid overfitting to small local data), the node optimizer continuously adjusts the model parameters. Ultimately, the node obtains a personalized fault identification model that absorbs global consensus knowledge, deeply integrates unique local time-series patterns and distribution characteristics, and is subject to dual constraints to prevent serious deviations. This model can better adapt to the specific environment of its region and make more accurate perceptions and judgments on the evolution trend of faults.

[0069] Among them, the total loss function in the personalized fine-tuning stage Defined as: , The definitions of each loss are as follows: , , , in, These are the weighting coefficients. For the current batch The feature set extracted from it. This is the variance calculation function. These are prototype alignment loss, distribution consistency loss, and task loss (classification / prediction), respectively. For the sample Fault category labels, Model on samples The predicted output, For the sample The true label.

[0070] Furthermore, the temporal attention mechanism used in the temporal modeling stage of S4-3 is a hybrid structure of multi-head self-attention mechanism and recurrent neural network, specifically including: (S4-3-1) Embedding layer, used to map multimodal inspection data at different time steps into a unified temporal feature vector: At the start of temporal modeling, the system first processes the raw multimodal data (such as a pair consisting of a visible light image and an infrared thermogram) collected at each time step. The features of each time step from the pre-trained cross-modal contrastive distillation network are input into an embedding layer. The embedding layer is essentially a learnable linear transformation or a small neural network, whose function is to map the high-level features extracted at each time step, which may vary in dimension or form, into a dense vector space of uniform dimension. After processing by this layer, the complex multimodal information of each time step is transformed into a fixed-length temporal feature vector, thus forming a regular temporal feature sequence that can be uniformly processed by subsequent modules.

[0071] (S4-3-2) Multi-head self-attention layer, used to capture long-term dependencies and key fault features across time steps: Next, the temporal feature sequence is fed into a multi-head self-attention layer. Multi-head self-attention is a powerful sequence modeling mechanism that allows the model to pay attention to information from all time steps in the sequence (including those long ago or later) in parallel and to varying degrees while processing information at the current time step. As multiple heads, it uses multiple sets of such attention computations in parallel, each focusing on different aspects (subspaces) of the input information. In this embodiment, this layer automatically analyzes the entire inspection sequence, identifying and focusing on features from historical or future moments that are crucial to the current fault diagnosis. For example, it might discover that a minor anomaly a week ago is a precursor to the current serious fault and assign higher weight to features from that moment. This mechanism enables it to effectively capture long-distance, non-adjacent long-term dependencies and highlight key fault features.

[0072] (S4-3-3) Gated recurrent unit layer, used for state modeling and information integration of attention-weighted temporal feature sequences.

[0073] After the multi-head self-attention layer performs global and flexible weighting and fusion of sequence information, the output sequence is input into the gated recurrent unit (ROU) layer. The Gated Recurrent Unit (GRU) is an efficient variant of the recurrent neural network, adept at sequential processing of sequence data and maintaining an internal state to remember historical information. This layer processes the attention-reweighted feature sequence step-by-step. Through internal gating mechanisms (update and reset gates), it determines how much past state information is retained in the current state and how much new input information is incorporated into the current state. This process models the temporal dynamics, smoothly integrates information, and encodes the evolution patterns of the sequence, making it particularly suitable for modeling gradual changes in device states.

[0074] (S4-3-4) Use the parameters of the global model as the initialization starting point.

[0075] Before personalized fine-tuning begins, all parameters (i.e., weights and biases) of the global model that have just been aggregated and updated, received by the edge nodes from the central server, are directly loaded into the local model as its initial parameters. This means that personalized fine-tuning does not start from scratch or random initialization, but rather starts from a robust baseline model that has already incorporated knowledge from numerous nodes in the federation, providing a high-quality starting point for fast and efficient fine-tuning.

[0076] (S4-3-5) In each local training iteration, the weighted sum of the prototype alignment loss, distribution consistency loss and temporal classification loss is calculated as the total loss.

[0077] In each round of local iterative training, nodes use a small batch of local time-series data. After the model forward propagates, the system simultaneously calculates three loss terms: prototype alignment loss (measures how close the features extracted by the current model are to the global mean prototype), distribution consistency loss (measures how well the distribution of the current batch of features matches the global variance prototype), and time-series classification loss (measures the accuracy of the model in classifying or predicting faults in time-series data, such as cross-entropy loss). These three values ​​are then weighted and summed according to a preset importance ratio to form a single, comprehensive total loss value, which comprehensively guides the direction of model updates.

[0078] (S4-3-6) The total loss is optimized using a meta-learning optimizer, enabling the model to quickly converge to the personalized optimal solution on a small amount of local data.

[0079] To enable the model to quickly adapt and find optimal personalized parameters with limited local data, the node uses a meta-learning optimizer to perform parameter updates. Unlike standard gradient descent optimizers, meta-learning optimizers are typically designed to learn more efficiently. For example, they can adaptively set differentiated learning rates for different parameters or memorize past update patterns through internal state, thus making smarter update decisions. In this embodiment, after receiving the calculated total loss, the optimizer uses the backpropagation algorithm to calculate the gradient and then applies its internal learning strategy to update the model parameters. This optimization method accelerates convergence, avoids overfitting or underfitting on limited data, and helps the model efficiently find a personalized optimal solution for the local data distribution and task objective within a limited number of local training epochs.

[0080] S5. Utilize the fine-tuned model to identify faults and assess trends in power equipment, and synchronize the identification results and fault evolution trajectories to the power equipment health management system. This diagnostic application and system integration represents the value output and closed-loop feedback step of the method. The personalized model obtained in step four is then put into practical application, enabling not only real-time fault identification but also trend prediction. More importantly, the results are structured and fed back to the power equipment health management system, updating the digital twin and generating maintenance recommendations, thereby embedding AI diagnostic capabilities into actual business management processes. This achieves end-to-end connectivity from model training to business decision-making, forming a business closed loop where diagnostic results drive management decisions, and management data (in the future) is then fed back to model training.

[0081] Specifically, in this embodiment, step S5 utilizes a fine-tuned model to identify faults and assess trends in power equipment, and synchronizes the identification results and fault evolution trajectories to the power equipment health management system. This includes: (S5-1) Fault Identification Stage: Input the real-time multimodal inspection data of the power equipment to be diagnosed into the personalized fault identification model, and output the fault type, location, and confidence level: When the inspection drone flies over the power equipment to be diagnosed (such as insulator strings on transmission lines), the system controls its onboard sensors to simultaneously acquire the latest visible light images and infrared thermal images. This real-time data is immediately input into a personalized fault identification model deployed on the drone or associated edge nodes. The model quickly analyzes the input data and outputs diagnostic results. The results are presented in a structured form, clearly providing three pieces of information: fault type (e.g., severe contamination on the surface of the insulator string), fault location (coordinate region in the image or 3D model of the equipment), and a confidence score (a value between 0 and 1, indicating the model's confidence in its judgment, e.g., 0.92).

[0082] (S5-2) Trend Assessment Phase: Based on the historical identification result sequence of the same equipment, a fault feature evolution curve is constructed, and a time series prediction model is used to predict the changes in fault state within a specific future time window: For the same device, the system retrieves all historical inspection and diagnostic records from a past period (e.g., the past 30 days), forming a time-ordered sequence of historical identification results. Based on this sequence, the system extracts key indicators (such as specific fault characteristic values ​​or confidence changes for each diagnosis) over time, forming a fault characteristic evolution curve. Subsequently, a specially trained time-series prediction model (such as a prediction network based on LSTM or Transformer) is loaded. This model uses this evolution curve as input and, by analyzing historical patterns, infers the changes in the device's fault state within a specific future time window (e.g., the next 7 days), predicting its possible development trend (e.g., the contamination level will continue to rise, expected to reach a high-risk threshold in 3 days).

[0083] (S5-3) System Synchronization Phase: The fault type, location, confidence level, and predicted fault state change trajectory are encapsulated into a standardized health status message and synchronized to the power equipment health management system through a secure interface. After diagnosis and prediction are completed, the system enters the integration phase. Information such as fault type, location coordinates, confidence score, and predicted future fault state change trajectory (i.e., trend assessment results) are packaged into a standardized health status message according to a predefined, machine-readable format. Then, this message is sent to the central server of the power equipment health management system via an encrypted, authenticated, and secure interface (such as an HTTPS-based API) to complete data synchronization.

[0084] (S5-4) The health management system updates the digital twin status of the corresponding device based on the received information and generates preventative maintenance suggestions or early warning signals: Upon receiving a standardized message, the health management system first parses the data. Based on the device identifier in the message, it locates the virtual image—the digital twin—corresponding to the physical device within the system. Subsequently, the system uses the latest diagnostic and predictive information from the message to update the status attributes of the digital twin (e.g., updating the pollution level of the insulator string from moderate to severe, along with a predicted trend). Based on the updated overall status, the system's decision engine automatically analyzes and may perform two operations: generating preventative maintenance recommendations (e.g., recommending cleaning the insulator strings on tower XX of line XX within 2 days), and / or generating early warning signals based on preset risk rules (e.g., issuing a high-risk warning and pushing it to the mobile terminals of relevant maintenance personnel or the dispatch dashboard), thereby driving precise and proactive maintenance actions.

[0085] Example 2

[0086] like Figure 2 As shown, this application provides a system architecture diagram for collaborative training and secure update of a distributed fault identification model for distribution network drones based on federated learning. It is applied to the collaborative training and secure update system for a distributed fault identification model for distribution network drones based on federated learning as described in Embodiment 1, including: multiple edge node devices 210, a central server device 220, a secure communication module 230, a personalized fine-tuning module 240, and a health management system interface module 250.

[0087] Multiple edge node devices 210 are deployed in various inspection drones or local stations to collect multimodal inspection data of power equipment and perform local model training, feature prototype calculation, privacy protection processing, and data compression and uploading.

[0088] The central server device 220 is communicatively connected to the multiple edge node devices and is used to decrypt and correct noise in the received encrypted information, perform global prototype calculation, dynamic client selection, model aggregation, and generate and distribute global models.

[0089] The secure communication module 230 is integrated into the edge node device and the central server device to realize layered secure communication based on homomorphic encryption and differential privacy, as well as adaptive data compression transmission based on communication bandwidth.

[0090] The personalized fine-tuning module 240 is deployed in each of the edge node devices and is used to perform personalized fine-tuning of the model after receiving the global model, combining local time series data and global prototype constraints.

[0091] The health management system interface module 250 is deployed in the edge node device or the central server device and is used to encapsulate and synchronize the fault identification and trend evaluation results of the fine-tuned model to the power equipment health management system.

[0092] During the implementation of this system, the components work together according to the following process: Data acquisition and local training phase: Edge node device 210 acquires power equipment inspection data through multimodal sensors mounted on the UAV, and performs cross-modal comparative distillation training and feature prototype calculation locally.

[0093] Secure upload phase: The secure communication module 230 performs homomorphic encryption, differential privacy noise addition and adaptive compression on the model parameters and feature prototypes, and uploads them to the central server device 220 via wireless network.

[0094] Global aggregation stage: After decrypting and noise correction of the received data, the central server device 220 calculates the global mean and variance prototypes, and achieves fault-tolerant aggregation by dynamically selecting alignment with the hierarchical prototypes through the client, thereby generating a global model.

[0095] Personalized fine-tuning stage: The central server distributes the global model to each edge node, and the personalized fine-tuning module 240 performs fine-tuning locally on each node by combining the temporal attention mechanism and dual prototype constraints to obtain a personalized fault recognition model.

[0096] Diagnosis and system integration phase: The health management system interface module 250 encapsulates the diagnostic results into standard messages and synchronizes them to the power equipment health management system to drive preventive maintenance decisions.

[0097] Through the above process, the system achieves a closed-loop system covering the entire chain from data collection, collaborative training, secure communication to intelligent diagnosis and operation and maintenance decision-making.

[0098] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0099] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0100] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0101] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0102] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0103] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0104] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0105] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0106] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for collaborative training and secure update of a distributed fault identification model for distribution network drones based on federated learning, characterized in that, Includes the following steps: Multiple edge nodes utilize locally collected multimodal inspection data of power equipment to train local fault identification models through a cross-modal comparative distillation network based on positive and negative samples, and calculate local category feature prototypes based on the distilled feature space. The edge nodes will upload the trained model parameters, model update information and local category feature prototypes to the central server after using homomorphic encryption and differential privacy noise injection for layered processing, and adaptive compression combined with communication bandwidth. After performing homomorphic decryption and differential privacy noise correction on the encrypted information, the central server calculates the global mean prototype and global variance prototype based on the local category feature prototype. Combined with dynamic client selection and near-term constraints, fine-grained fault-tolerant aggregation is achieved through hierarchical prototype alignment to generate a global model. The global model and global mean prototype are distributed to each edge node. Each node, based on local multimodal inspection data, performs personalized fine-tuning of the global model under the dual constraints of the global mean prototype and the global variance prototype, combined with the temporal attention mechanism, to obtain a fault identification model that adapts to the local environment and fault evolution trend. The fine-tuned model is used to identify faults and assess trends in power equipment, and the identification results and fault evolution trajectories are synchronized to the power equipment health management system.

2. The method according to claim 1, characterized in that, The training of the local fault identification model through a cross-modal contrastive distillation network based on positive and negative samples specifically includes: Construct positive sample pairs from inspection data of at least two different modes collected from the same power equipment at the same time. Multimodal inspection data of different power equipment or different fault categories are constructed into negative sample pairs; By using the positive and negative sample pairs, the cross-modal contrastive distillation network is constrained by the contrastive learning loss function, so that the features of different modes with the same fault state are close to each other in the feature space, while the features of different fault states are far apart, thereby optimizing the cross-modal feature fusion effect.

3. The method according to claim 2, characterized in that, The calculation of the local category feature prototype based on the distilled feature space specifically includes: Using the feature extractor of the trained cross-modal contrastive distillation network, forward propagation is performed on all local labeled samples to extract their output feature vectors at the fusion feature layer. Based on the fault category label of the sample, cluster analysis is performed on the feature vectors of all samples belonging to the same category; Calculate the feature vector of the cluster center for each category, and use the cluster center vector as the local feature prototype for that category.

4. The method according to claim 1, characterized in that, The process involves layered processing using homomorphic encryption and differential privacy noise injection, combined with adaptive compression based on communication bandwidth, before uploading to the central server. Specifically, this includes: First processing layer: The model parameters, model update information, and local category feature prototypes are encrypted using a homomorphic encryption algorithm; Second processing layer: Inject random noise that meets differential privacy requirements into the encrypted data; The third processing layer dynamically selects either gradient sparsification or quantization compression strategies based on the real-time communication bandwidth of the current edge node, and performs lossy compression on the data after injecting noise. The data, after being processed and compressed, is uploaded to the central server.

5. The method according to claim 1, characterized in that, After homomorphically decrypting and differentially private noise correcting the encrypted information, the central server calculates the global mean prototype and global variance prototype based on the local category feature prototype. Combining dynamic client selection and near-term constraints, it achieves fine-grained fault-tolerant aggregation through hierarchical prototype alignment to generate a global model, specifically including: The received encrypted information is homomorphically decrypted, and the noise introduced by differential privacy in the decrypted data is statistically corrected. Based on the local category feature prototypes uploaded by each edge node, the prototype mean of all nodes is calculated by category to obtain the global mean prototype; at the same time, the variance of the prototype distribution of each category is calculated to obtain the global variance prototype. Based on the computing resources, network status, and historical contribution of each edge node, the set of effective clients participating in this round of aggregation is dynamically selected. Introduce a proximal term into the objective function of model aggregation to constrain the distance between the local model update of each node and the previous round of global model; Based on the global mean prototype and global variance prototype, the model update of the effective client set is subjected to fine-grained weighting and anomaly filtering through the hierarchical prototype alignment loss function, thereby completing fault-tolerant aggregation and generating the current round global model.

6. The method according to claim 5, characterized in that, The step of applying fine-grained weighting and anomaly filtering to the model updates of the effective client set using the hierarchical prototype alignment loss function specifically includes the following steps: First layer alignment: Calculate the cosine similarity between the local category feature prototype uploaded by each edge node and the global mean prototype to generate the first weight vector; Second-level alignment: Calculate the distribution difference between each local category feature prototype and the global variance prototype, filter the prototypes based on a preset anomaly threshold, and remove abnormal local prototypes whose difference exceeds the threshold. Aggregation execution: Based on the first weight vector, the model updates corresponding to the edge nodes after the second-layer alignment filtering are weighted and averaged to complete the aggregation generation of the global model.

7. The method according to claim 1, characterized in that, Each node, based on local multimodal inspection data, performs personalized fine-tuning of the global model under the dual constraints of the global mean prototype and the global variance prototype, combined with a temporal attention mechanism. Specifically, this includes: First constraint stage: Using the global mean prototype, construct prototype alignment loss to guide the features extracted by the node local model to move closer to the global feature center; Second constraint stage: Using the global variance prototype, construct the distribution consistency loss to constrain the dispersion of the local feature distribution of the node to match the global distribution; Temporal modeling stage: For the multimodal inspection time-series data of the node, the temporal attention mechanism is used to extract the dependencies between features at different time steps and generate temporal enhanced features; Personalized fine-tuning stage: Based on the temporal enhancement features, under the joint constraints of the prototype alignment loss and the distribution consistency loss, the global model is trained locally for a limited number of rounds to obtain a personalized fault identification model adapted to the local environment and fault evolution trend.

8. The method according to claim 7, characterized in that, The temporal attention mechanism used in the temporal modeling stage is a hybrid structure of multi-head self-attention mechanism and recurrent neural network, specifically including: The embedding layer is used to map multimodal inspection data at different time steps into a unified temporal feature vector; A multi-head self-attention layer is used to capture long-term dependencies and key fault features across time steps; The gated recurrent unit layer is used for state modeling and information integration of attention-weighted temporal feature sequences; The limited rounds of local training in the personalized fine-tuning phase employ a meta-learning initialization strategy, specifically: The parameters of the global model are used as the initialization starting point; In each local training iteration, the weighted sum of the prototype alignment loss, distribution consistency loss, and temporal classification loss is calculated as the total loss; The total loss is optimized using a meta-learning optimizer, enabling the model to quickly converge to a personalized optimal solution on a small amount of local data.

9. The method according to claim 1, characterized in that, The process of using a fine-tuned model to identify faults and assess trends in power equipment, and synchronizing the identification results and fault evolution trajectories to the power equipment health management system, specifically includes: Fault identification phase: Input the real-time multimodal inspection data of the power equipment to be diagnosed into the personalized fault identification model, and output the fault type, location and confidence level; Trend assessment phase: Based on the historical identification result sequence of the same device, a fault feature evolution curve is constructed, and a time series prediction model is used to predict the changes in fault state within a specific future time window; System synchronization phase: The fault type, location, confidence level, and predicted fault state change trajectory are encapsulated into a standardized health status message and synchronized to the power equipment health management system through a secure interface; the health management system updates the digital twin status of the corresponding equipment based on the received information and generates preventive maintenance suggestions or early warning signals.

10. A collaborative training and security update system for a distributed fault identification model for distribution network drones based on federated learning, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: Multiple edge node devices are deployed in various inspection drones or local stations to collect multimodal inspection data of power equipment and perform local model training, feature prototype calculation, privacy protection processing, and data compression and uploading. The central server device is communicatively connected to the multiple edge node devices and is used to decrypt and correct noise for received encrypted information, perform global prototype calculation, dynamic client selection, model aggregation, and generate and distribute global models. A secure communication module is integrated into the edge node device and the central server device to realize layered secure communication based on homomorphic encryption and differential privacy, as well as adaptive data compression transmission based on communication bandwidth. The personalized fine-tuning module is deployed in each of the aforementioned edge node devices and is used to perform personalized fine-tuning of the model after receiving the global model, combining local time series data and global prototype constraints. The health management system interface module, deployed in the edge node device or central server device, is used to encapsulate and synchronize the fault identification and trend assessment results of the fine-tuned model to the power equipment health management system.

Images