A Distributed Cooperative Scheduling Method and System for Multi-Robots in Hydropower Stations Based on Edge Federated Learning

Abstract

This invention relates to the field of intelligent collaborative scheduling technology for hydropower station robot clusters, and discloses a distributed collaborative scheduling method and system for multiple robots in hydropower stations based on edge federated learning. The method includes: obtaining the working mode definitions of robots within each hydropower station participating in cross-hydropower station federated learning; generating a multi-granularity model library including high-precision complete models, medium-precision pruned models, and low-precision simplified models, maintaining a unique parameter structure at the high-level decision-making layer; dividing the working modes into a high-real-time mode group and a high-security mode group according to the relationship between real-time requirement weights and security requirement weights; each hydropower station combining globally shared layer parameters with its local unique layer parameters, and the robot loading the corresponding granularity model from its local model library according to the current working mode. This invention achieves the technical effect of differentiated optimization of real-time and security requirements while ensuring parameter privacy and security.

Description

Technical Field

[0001] This invention relates to the field of intelligent collaborative scheduling technology for hydropower station robot clusters, and more specifically, to a distributed collaborative scheduling method for multiple robots in hydropower stations based on edge federated learning. Background Technology

[0002] In hydropower station operation and maintenance scenarios, multiple hydropower stations need to collaboratively train a unified fault diagnosis model to improve overall operation and maintenance efficiency. Robots at each hydropower station switch between different modes based on actual work needs, including routine inspection mode, emergency response mode, and energy-saving maintenance mode, using neural network models of different granularities in different modes. Since different hydropower stations are managed by different operating entities, and their respective operational data and model parameters are considered trade secrets, centralized aggregation servers are not trusted. Therefore, decentralized, secure multi-party computation technology is required to achieve cross-hydropower station federated learning.

[0003] Existing technologies employ uniform encryption strength and aggregation strategies to process model parameters across all working modes. They typically use fully homomorphic encryption algorithms to protect parameter privacy and require consistent model parameter dimensions to perform secure multi-party computation aggregation.

[0004] However, existing technologies have the following technical problems: First, uniform strong encryption leads to excessively high aggregation latency in emergency mode. Robots in emergency response mode need to quickly aggregate and update high-precision complete models to cope with sudden failures, but the computational overhead of fully homomorphic encryption is huge, resulting in excessively high aggregation latency and affecting the speed of failure response. Second, inconsistent parameter dimensions lead to loss of aggregation efficiency. Robots in regular inspection mode use trimmed models, whose parameter dimensions are different from the complete model. Safety aggregation across hydropower stations requires filling and alignment, resulting in a large amount of redundant computation and communication overhead. Third, encryption strength is not optimized according to different requirements. The same encryption strength is used for model parameters of all modes and granularities. It fails to optimize according to the differences in security and real-time requirements of different working modes, resulting in excessive encryption in regular mode consuming resources and encryption latency in emergency mode affecting timeliness. Summary of the Invention

[0005] This application provides a distributed collaborative scheduling method and system for multiple robots in hydropower stations based on edge federated learning, which solves the technical problems in the background art where uniform encryption strength leads to excessively high aggregation delay in emergency mode and inconsistent parameter dimensions lead to loss of aggregation efficiency.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to a first aspect of the embodiments of this application, a distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning is provided, including: Obtain the working mode definitions of robots in each hydropower station participating in cross-hydropower station federated learning. The working mode definitions include routine inspection mode, emergency response mode and energy-saving maintenance mode, and determine the real-time requirement weight and safety requirement weight of each working mode. Based on the performance requirements and energy consumption constraints of each model, a multi-granularity model library is generated, including high-precision complete models, medium-precision pruned models, and low-precision simplified models. Each granularity model shares the parameter structure of the underlying feature extraction layer and maintains a unique parameter structure in the high-level decision layer. The parameter index set of the shared layer and the parameter index set of the unique layer of each granularity model are determined. Within each hydropower station, local federated aggregation is performed on robots operating in the same mode. A weighted average is calculated based on the number of training samples for each robot to generate hydropower station-level aggregation parameters. Based on the relationship between the weights of real-time requirements and security requirements, the working modes are divided into a high real-time mode group and a high security mode group. For the hydropower station-level aggregate parameters of the high real-time mode group, the shared layer parameters are extracted, a polynomial is constructed on a finite domain and the evaluation point value is calculated, and the parameters are divided into secret fragments. For the hydropower station-level aggregate parameters of the high security mode group, the shared layer parameters are extracted and the ciphertext parameters are generated using a fully homomorphic encryption algorithm. The system performs encrypted aggregation across hydropower stations. After the secret fragment exchange in the high real-time mode, the original parameters of each hydropower station are restored using the Lagrange interpolation formula, and a weighted average is calculated based on the number of training samples to generate global shared layer parameters. For the ciphertext parameters in the high security mode, the system performs homomorphic addition in the ciphertext space to calculate a weighted sum, and then decrypts the data using a distributed decryption protocol to generate global shared layer parameters. Each hydropower station combines the globally shared layer parameters with its local unique layer parameters to generate a multi-granularity global model library, which is then distributed to the robot. The robot loads the model of the corresponding granularity from the local model library according to the current working mode.

[0008] According to one embodiment of this application, determining the real-time requirement weight and security requirement weight for each working mode specifically includes: Obtain the maximum tolerable aggregate latency and the baseline aggregate latency for the working mode; Calculate the ratio of the maximum tolerable aggregation delay to the sum of the maximum tolerable aggregation delay and the baseline aggregation delay; Subtracting the ratio from 1 yields the real-time demand weight; The security requirement weight is equal to 1 minus the real-time requirement weight; Among them, the maximum tolerable aggregation latency is the maximum allowed time from initiating aggregation to completing model update for each working mode, which is pre-configured according to the urgency of the working mode; the baseline aggregation latency is the theoretical aggregation latency when using the lowest encryption strength, which is obtained through pre-testing as a normalized baseline value.

[0009] According to one embodiment of this application, determining the parameter index set of the shared layer and the parameter index set of each granularity model's unique layer specifically includes: Analyze the network architecture of each granular model layer by layer, starting from the input layer, and check whether the layer type, convolution kernel size, number of channels and activation function of each layer are completely consistent; The network layers that meet the consistency conditions are marked as shared layers, and a set of shared layer indexes is generated. High-level decision-making layers that do not meet the consistency condition are marked as unique layers, and a unique layer index set is generated for each granularity model. Among them, the shared low-level feature extraction layer of each granularity model consists of multiple convolutional layers and pooling layers, which are used to extract general time-frequency domain features of device operating status; each granularity model maintains a unique layer parameter structure in the high-level decision layer, which consists of fully connected layers, and contains different numbers of fully connected layers according to the model granularity.

[0010] According to one embodiment of this application, the step of extracting shared layer parameters from the hydropower station-level aggregated parameters of the high real-time mode group, constructing a polynomial over a finite domain and calculating the evaluation point values, and dividing it into secret fragments specifically includes: Each hydropower station generates a local public-private key pair and exchanges public keys to build an encrypted communication channel; For each parameter in the shared layer parameter vector, a polynomial is constructed over a finite field. The constant term of the polynomial is the parameter value, and the coefficients of the remaining terms are uniformly and randomly selected in the finite field. The degree of the polynomial is equal to the safety threshold minus 1. Calculate the values ​​of the polynomial at different evaluation points for the number of participating hydropower stations, and generate the corresponding number of secret fragments; Each secret segment is sent to its corresponding participating hydropower station; Among these requirements, the prime numbers in the finite field must be greater than the maximum absolute value of the parameter to ensure the uniqueness of the mapping of the parameter value within the finite field; the random coefficients are generated by a cryptographically secure random number generator to ensure the unpredictability of secret sharding; and the security threshold ranges from at least 2 to no more than the number of hydropower stations involved.

[0011] According to one embodiment of this application, the execution of cross-hydropower station encrypted aggregation, after secret fragment exchange in the high real-time mode, uses the Lagrange interpolation formula to recover the original parameters of each hydropower station and calculates a weighted average based on the number of training samples to generate global shared layer parameters, specifically includes: Each hydropower station receives secret fragments from other hydropower stations and selects at least a security threshold number of secret fragments. The original parameter values ​​are restored using the Lagrange interpolation formula. Specifically, the polynomial values ​​corresponding to each secret partition are multiplied by the corresponding Lagrange interpolation coefficients and then summed. The Lagrange interpolation coefficients are the product of the evaluation points of other partitions except the current partition divided by the product of the differences between the evaluation points. After restoring the original shared layer parameters of each hydropower station, the total number of training samples of each hydropower station is used as the weight to calculate the weighted average of the original parameters of all hydropower stations, thus obtaining the global shared layer parameters.

[0012] According to one embodiment of this application, the ciphertext parameters of the high-security mode are subjected to homomorphic addition to calculate a weighted sum in the ciphertext space, and then decrypted using a distributed decryption protocol to generate globally shared layer parameters, specifically including: Calculate the aggregate weight of each hydropower station. The aggregate weight is equal to the number of training samples for that hydropower station divided by the sum of the number of training samples for all participating hydropower stations. The aggregate weight is determined based on the proportion of the number of training samples for each hydropower station to the total number of samples. In the ciphertext space, the products of the ciphertext parameters of each hydropower station and their corresponding aggregation weights are summed, and the global ciphertext parameters are calculated using the additive homomorphic property of fully homomorphic encryption. Each hydropower station uses its own private key to participate in the distributed decryption protocol, jointly decrypting the global ciphertext parameters to obtain the global shared layer parameters in plaintext form.

[0013] According to one embodiment of this application, each hydropower station combines globally shared layer parameters with locally unique layer parameters to generate a multi-granularity global model library, specifically including: The unique layer parameters of each granularity model are aggregated independently within each hydropower station and do not participate in the transmission between hydropower stations. Each hydropower station assigns the global shared layer parameters to the shared bottom feature extraction layer of the model and assigns the unique layer parameters to the high-level decision layer of the model according to the model granularity used in its local mode, thus completing the parameter loading of the global model. Generate a multi-granularity global model library that includes high-precision complete models, medium-precision cropped models, and low-precision minimalist models.

[0014] According to one embodiment of this application, a consistency verification step is further included after performing cross-hydropower station cryptographic aggregation: Each hydropower station calculates the hash value of the reconstructed or decrypted global shared layer parameters, exchanges the hash values ​​with each other, and compares them. When an inconsistency is detected in the secret fragment sent by a hydropower station or the encryption parameter verification fails, the distributed verification protocol is triggered, requiring each hydropower station to resubmit its intermediate calculation results. Each normal hydropower station independently verifies the intermediate calculation results of the suspicious node. If the verification passes, it votes in favor; if the verification fails, it votes against. When the number of opposing votes exceeds half of the total number of participants, the suspicious node is determined to be an abnormal node. Isolate the abnormal node from the participant list, re-execute the aggregation process, and eliminate the influence of malicious participants.

[0015] According to one embodiment of this application, a dynamic optimization step is further included after the multi-granularity model library is distributed: Each hydropower station records the communication overhead, computation time, and model performance indicators for cross-hydropower station aggregation. The communication overhead is the total number of bytes transmitted by each hydropower station in the current round, the computation time is the total time taken from the start of encryption to the completion of aggregation in the current round, and the model performance indicator is the weighted average of the fault diagnosis accuracy of the global model on the local validation set of each hydropower station. The model performance index, computation time, and communication overhead are respectively subjected to mean normalization based on the range. The comprehensive evaluation index is calculated as the weighted difference between the normalized model performance index and the computation time and communication overhead. The sum of the weight coefficients is 1 and all are greater than 0. When the comprehensive evaluation index is lower than the preset threshold, the values ​​of the real-time requirement weight and the security requirement weight are adjusted, and the high real-time mode group and the high security mode group are reclassified. Based on the adjusted pattern grouping update pattern-encryption type mapping strategy, the new encryption method selection rules are applied in the next round of federated learning.

[0016] According to a first aspect of the embodiments of this application, a distributed collaborative scheduling system for multi-robots in a hydropower station based on edge federated learning is provided, for performing the method described in the first aspect, the system comprising: The mode requirement weight calculation module is used to calculate the real-time requirement weight and the security requirement weight based on the latency parameter of the working mode. The multi-granularity model library generation module is used to generate multi-granularity models with shared underlying parameter structures and unique layer parameter structures; The local federated aggregation module is used to perform local federated aggregation within the hydropower station to generate hydropower station-level aggregation parameters; The block encryption module is used to group the working modes according to the weight of real-time requirements and the weight of security requirements, and select the corresponding encryption algorithm to encrypt the shared layer parameters. The cross-hydropower station encryption aggregation module is used to perform interpolation recovery on secret fragments and / or perform homomorphic aggregation on ciphertext parameters to generate global shared layer parameters; The model combination and distribution module is used to combine globally shared layer parameters and locally unique layer parameters to generate a multi-granularity global model library and distribute it to the robot.

[0017] Compared with existing technologies, the beneficial effects of adopting the above technical solution are as follows: This invention, through the construction of a multi-granularity model library and the design of a shared parameter structure, identifies the shared underlying feature extraction layer and the unique layer parameter structure of each granularity model. It performs cross-hydropower station encrypted aggregation only on the shared layer parameters, while keeping the unique layer parameters updated locally. This solves the problem of filling and alignment efficiency loss caused by the inconsistency of parameter dimensions of different granularity models, and achieves the technical effect of avoiding filling and alignment operations of unique layer parameters and reducing redundant calculations and communication overhead.

[0018] This invention solves the technical problem of the triple contradiction between security, efficiency and flexibility in cross-institutional federated learning under multi-working mode and multi-granularity environment by adopting a pattern-aware selective aggregation strategy, dynamically balancing encryption strength according to the weights of real-time requirements and security requirements, and ensuring the security of decentralized aggregation through the application of secret sharing and homomorphic encryption in grouping. The distributed verification protocol detects malicious node tampering. It achieves the technical effect of differentiating the optimization of real-time and security requirements while ensuring parameter privacy and security. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0020] Figure 1 This is a flowchart of the multi-robot distributed collaborative scheduling method for hydropower stations based on edge federated learning, as described in this invention.

[0021] Figure 2 This is a mixed bar chart and line chart comparing the performance of the multi-granularity model of the present invention.

[0022] Figure 3 This is a bar chart comparing the aggregation performance of different encryption methods of the present invention.

[0023] Figure 4 This is a line graph showing the convergence curve of the accuracy of the federated learning rounds in this invention.

[0024] Figure 5 This is a bar chart showing the distribution of training samples and operating modes of various hydropower stations according to the present invention.

[0025] Figure 6 This is a scatter plot showing the relationship between the model parameters and encryption calculation time of this invention.

[0026] Figure 7 This is a line graph showing the dynamic optimization trend of the comprehensive evaluation index of this invention.

[0027] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0028] The embodiments of this application are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar modules or modules having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. Rather, the embodiments of this application include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0029] At least one embodiment of the present invention discloses a distributed cooperative scheduling method for multiple robots in hydropower stations based on edge federated learning, such as... Figure 1 As shown, it includes the following steps: Step 100: Obtain the work mode definition and determine the requirement weight of each work mode.

[0030] Obtain a list of edge servers from each hydropower station participating in the cross-hydropower station federated learning and establish a point-to-point communication topology. Obtain the working mode definitions of robots within each hydropower station, including routine inspection mode, emergency response mode, and energy-saving maintenance mode. Extract real-time requirement weights for each working mode. And security requirement weight ,in, .

[0031] In this embodiment, the weight of real-time requirements And security requirement weight The values ​​of both are in the range of [0,1], and the sum of the two is equal to 1, which ensures the normalization constraint of the demand weight.

[0032] Furthermore, the weight of real-time requirements. The response time requirement is determined based on the working mode; specifically, it is determined by obtaining the maximum tolerable aggregate latency for the working mode. Aggregate delay with benchmark ,calculate The stricter the response time requirement, the better. The higher the value, the greater the weight of security requirements. .

[0033] It should be noted that the maximum tolerable aggregation delay... This is the maximum allowed time, in seconds, from initiating aggregation to completing model updates for each work mode. It is pre-configured based on the urgency of the work mode; emergency response modes are typically set to 5-10 seconds, and regular inspection modes to 30-60 seconds. Baseline aggregation latency. The theoretical aggregate latency with the lowest encryption strength is obtained through pre-testing and serves as a normalization baseline.

[0034] In this embodiment, the Three Gorges Hydropower Station, Gezhouba Hydropower Station, and Xiangjiaba Hydropower Station are taken as examples to illustrate participation in federated learning. After each hydropower station establishes a point-to-point communication topology, the working mode definition data is obtained as shown in Table 1.

[0035] Table 1 Working Mode Definition and Delay Parameters

[0036] According to the formula The demand weights for each mode were calculated, and the results are shown in Table 2.

[0037] Table 2 Calculation results of demand weights for each work mode

[0038] Step 200: Generate a multi-granularity model library and determine the shared parameter structure.

[0039] Based on the performance requirements and energy consumption constraints of each working mode, a neural network architecture search algorithm is used to generate multiple candidate models of varying complexity, including a high-precision complete model. Medium-precision trimming model and low-precision minimalist model .

[0040] Specifically, the performance requirements parameters and energy consumption constraint thresholds of each working mode are used as inputs to the neural network architecture search algorithm, and the output is a multi-granular neural network architecture configuration that satisfies the constraints. Different architecture combinations are explored in the search space of network depth and width.

[0041] Furthermore, when executing the neural network architecture search algorithm, the structure of the shared bottom feature extraction layer is first fixed. Then, for each granularity model, different combinations of the number of fully connected layers and the number of neurons are searched at the high-level decision layer. The fault diagnosis accuracy and inference energy consumption of each candidate architecture are evaluated on the validation set. The architecture that meets the performance requirements of each mode and whose energy consumption is lower than the constraint threshold is retained as the corresponding granularity model.

[0042] In this embodiment, each granularity model adopts a convolutional neural network architecture. The input layer receives multi-channel time-series monitoring data from the hydropower station equipment. Before inputting the data into the neural network, Z-score standardization is performed on the data from each sensor channel to eliminate the dimensional differences between different physical quantities. The output layer outputs the probability distribution of fault categories. Each granularity model shares the parameter structure of the underlying feature extraction layer. The shared low-level feature extraction layer consists of multiple convolutional and pooling layers, used to extract general time-frequency domain features of the device's operating status. Each granularity model maintains its unique layer parameter structure at the higher-level decision layers. ,in This indicates the model granularity type. The high-level decision layer consists of fully connected layers, and the number of fully connected layers varies depending on the model granularity.

[0043] Specifically, a high-precision complete model The high-level decision-making layer contains multiple fully connected layers, and the medium-precision pruning model... The high-level decision-making layer contains fewer fully connected layers, resulting in a low-precision, minimalist model. The high-level decision-making layer contains only a single fully connected layer. Each granularity model is trained using supervised learning, employing the cross-entropy loss function. The difference between the predicted fault category probability and the actual fault label is measured, where For fault category index, This represents the total number of fault categories. One-hot encoding of the real label. The first output of the model Fault probability. The Adam optimization algorithm is used to update the model parameters.

[0044] After determining the structure of each granularity model, it is also necessary to determine the parameter index set for the shared layer and the parameter index set for the unique layers of each granularity model. This includes the following steps: Step 201: Analyze the network architecture of each granularity model and identify the preceding layers of each model, starting from the input layer. Layered network layer.

[0045] Furthermore, parameters This is determined by checking the consistency of the network layer structure of each granularity model layer by layer, starting from the input layer. When inconsistencies in layer type, kernel size, or number of channels are detected in a certain layer, the number of all consistent layers preceding the inconsistent layer is taken as the threshold. The value of .

[0046] Step 202: For each model's pre-model... Each layer is checked to ensure that the kernel size, number of channels, and activation function are completely consistent.

[0047] Step 203: Mark the network layers that meet the consistency condition as shared layers and generate a shared layer index set. ,in 、 、 They are the 1st, 2nd, and 3rd respectively. Indexes of a shared layer.

[0048] Step 204: For high-level decision-making layers that do not meet the consistency condition, mark them as unique layers and generate a unique layer index set for each granularity model. .

[0049] For the task of diagnosing equipment faults in hydropower stations, the neural network architecture search algorithm generates model architectures of three granularities, as shown in Table 3.

[0050] Table 3 Multi-granularity model architecture configuration

[0051] Each granularity model shares a common underlying feature extraction layer, consisting of a first convolutional layer (Conv1), a first pooling layer, a second convolutional layer (Conv2), and a second pooling layer. The shared layer parameter index set is as follows: The total number of shared layer parameters is 51,972. The number of fully connected layers in the high-level decision-making layer varies for each granularity model, and the unique layer parameter index sets are as follows: 、 and .

[0052] Figure 2 The performance of three different granularity models (full model, pruned model, and minimal model) is compared in terms of parameter count, inference energy consumption, and validation accuracy.

[0053] Step 300: Execute local federated aggregation to generate hydropower station-level aggregation parameters.

[0054] Obtain the current working mode tags of robots in each hydropower station. and the corresponding model granularity used Within each hydropower station, local federated aggregation is performed on robots operating in the same mode, generating mode-specific hydropower station-level aggregation parameters. ,in Indicates the hydroelectric power station number. Indicates the working mode.

[0055] Specifically, when performing local federated aggregation, the local model parameters of each robot and the corresponding number of training samples are used as inputs, and the output is the hydropower station-level aggregation parameters. The parameters are fused by calculating a weighted average.

[0056] For example, hydroelectric power stations Inner mode Model parameters of each robot and number of training samples Calculate the aggregation parameters at the hydropower station level ,in Indexing for robots, For hydroelectric power station Inner mode A collection of robots.

[0057] The working modes of robots in the three hydroelectric power stations are distributed as shown in Table 4 during a certain round of federated learning.

[0058] Table 4. Working modes and training sample size of robots at each hydropower station

[0059] Each hydropower station performs local federated aggregation on robots operating in the same mode. For example, the Three Gorges Hydropower Station calculates hydropower station-level aggregation parameters for robots R1-1 and R1-2 in emergency response mode. The aggregation weights are 0.533 and 0.467, respectively. Through local federated aggregation, hydropower station-level aggregation parameters for each mode of hydropower station are generated: Three Gorges Hydropower Station generation... and Gezhouba Hydropower Station , and Xiangjiaba Hydropower Station generated and .

[0060] Figure 5 The distribution of training sample size is shown for three hydropower stations (Three Gorges, Gezhouba, and Xiangjiaba) under different operating modes (emergency response, routine inspection, and energy-saving maintenance).

[0061] Step 400: Select the encryption method according to the working mode group and generate encryption parameters.

[0062] Based on the relative weights of real-time performance requirements and security requirements, the operating modes are divided into a high-real-time mode group and a high-security mode group. When the working mode is in the high real-time mode group, it will be classified as such; when At that time, the working mode will be classified into the high security mode group.

[0063] For high real-time mode groups, the hydropower station-level aggregated parameters Extract the shared layer parameters. The Shamir secret sharing algorithm is used to segment the parameters as follows: A secret fragment ,in , , They are the 1st, 2nd, and 3rd respectively. A secret fragment, To determine the number of participating hydropower stations, any Each fragment can restore the original parameters. The threshold value is the security threshold. The Shamir secret sharing algorithm takes the shared layer parameter vector and the security threshold as inputs and outputs multiple secret slices, with parameter partitioning achieved by constructing a polynomial over a finite field.

[0064] In this embodiment, the safety threshold The range of values ​​is The range of values ​​for the security threshold ensures that at least two shards are required to recover parameters to provide basic security, and that the number of shards required does not exceed the total number of participants.

[0065] For hydropower station-level aggregate parameters of the high-safety mode group Extract the shared layer parameters. The parameters are encrypted using a fully homomorphic encryption algorithm to generate ciphertext parameters. ,in The public key is used as the input. The fully homomorphic encryption algorithm takes the shared layer parameter vector and the public key as inputs and outputs ciphertext parameters, supporting addition operations in the ciphertext space.

[0066] Furthermore, this embodiment also presents the segmentation process of the Shamir secret sharing algorithm, as follows: Step 401: Each hydropower station generates a local public-private key pair. They exchange public keys to establish an encrypted communication channel.

[0067] Step 402: For the shared layer parameter vector Each parameter In a finite field upper structure polynomial of degree ,in For parameter index, These are random coefficients.

[0068] In this embodiment, the finite field In For a large prime number, it must satisfy the following condition: To ensure the uniqueness of the mapping of parameter values ​​within a finite domain. Random coefficients. In a finite field The values ​​are selected uniformly and randomly, and the range is [missing information]. It is generated using a cryptographically secure random number generator to ensure the unpredictability of secret fragments.

[0069] Step 403: Calculate the polynomial in The values ​​of each different point ,in 、 、 They are polynomials At point 1, point 2, point The value at that location is used to generate A secret fragment.

[0070] Step 404: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require A secret fragment Send to the One participating hydropower station, among which This is a sharded index.

[0071] Taking the demand weight calculation results from Table 2 as an example, the working modes are grouped. Emergency response mode... It is classified into the high real-time mode group; the regular inspection mode is... It is classified into the high-security mode group; the energy-saving maintenance mode... They were categorized into the high-security mode group. For the emergency response mode, the shared layer parameters were extracted from the hydropower station-level aggregate parameters of the three hydropower stations.

[0072] Since the total number of shared layer parameters is 51,972, the Three Gorges Hydropower Station extracted... 51,972 parameter values. Set the security threshold. Number of hydropower stations In a finite field Construct a first-order polynomial for each parameter. Taking the Three Gorges Hydropower Station as an example, the parameters of the first shared layer... Random selection coefficient Construct polynomial Calculate the values ​​of the three evaluation points. 、 、 Generate three secret fragments , and The data is sent to three hydroelectric power stations respectively. Each of the three hydroelectric power stations generates a secret fragment set for the emergency response mode. , and ,in 、 、 These are secret fragments generated by hydropower station 1 for the emergency response mode and sent to hydropower stations 1, 2, and 3, respectively. For the regular inspection mode, the three hydropower stations use a fully homomorphic encryption algorithm to generate ciphertext parameters respectively. 、 and For the energy-saving maintenance mode, the Gezhouba Hydropower Station generates encrypted parameters. .

[0073] Figure 3 The performance differences between Shamir secret sharing and fully homomorphic encryption in three dimensions—aggregate computation time, communication overhead, and security strength—were compared.

[0074] Figure 6 The computational time relationship of encryption methods (Shamir secret sharing and fully homomorphic encryption) is shown for models with different parameter scales.

[0075] Step 500: Perform cross-hydropower station encrypted aggregation to generate global shared layer parameters.

[0076] For secret sharding in high real-time mode, each hydropower station sends its shards to other participating hydropower stations. Upon receiving shards from other hydropower stations, it performs homomorphic addition. Specifically, for each received shard... Each segment ,in 、 、 They are from the 1st, 2nd, and 3rd respectively. The secret segmentation of a hydroelectric power station, using the Lagrange interpolation formula to restore the original parameters:

[0077] in For hydroelectric power station index, For auxiliary index variables in interpolation calculations, The evaluation points corresponding to the partitions, For the first Each hydropower station has parameters The contribution value is calculated. The aforementioned Lagrange interpolation formula takes as input at least the number of secret slices and corresponding evaluation points equal to the safety threshold, and outputs the restored original parameter values. A weighted average of the contributions from all hydropower stations is calculated to generate global shared layer parameters. .

[0078] Furthermore, the original parameters of each hydropower station were restored using Lagrange interpolation. Then, the total number of training samples from each hydropower station was used. As weights, calculate the global shared layer parameters. ,in For hydroelectric power station index, The number of participating hydropower stations.

[0079] For the encryption parameters in the high-security mode, each hydropower station performs homomorphic addition aggregation in the ciphertext space to calculate the weighted sum of the encryption parameters:

[0080] in For hydroelectric power station index, For the first The aggregate weight of each hydropower station To meet the requirements for participating in the number of hydropower stations .

[0081] In this embodiment, aggregate weights According to the Number of training samples for each hydropower station Calculation, specifically ,in Index the hydropower stations to ensure that hydropower stations with large training data volumes contribute higher weights in the global model.

[0082] In this embodiment, the input to homomorphic addition aggregation is the ciphertext parameters and aggregation weights of each hydropower station, and the output is the global ciphertext parameters. The weighted sum is directly calculated in the ciphertext space using the homomorphic property of addition in fully homomorphic encryption. Each hydropower station uses its own private key. Participate in the distributed decryption protocol to jointly decrypt. Obtain global shared layer parameters The distributed decryption protocol takes global ciphertext parameters and the private keys of each hydropower station as inputs and outputs global shared layer parameters in plaintext form, avoiding the risk of single-point leakage through multi-party collaborative computation.

[0083] In emergency response mode, after the hydropower stations exchange secret fragments, the Three Gorges Hydropower Station receives three fragments from itself, the Gezhouba Hydropower Station, and the Xiangjiaba Hydropower Station. Taking the first shared layer parameter as an example, the fragment received by the Three Gorges Hydropower Station is... 、 、 When using the Lagrange interpolation formula to restore the original parameters, select any two slices (that meet the safety threshold). For example, select and Calculate the recovery value The original parameter values ​​are then restored. Each hydropower station calculates its aggregate weight based on the total number of training samples in Table 4. The aggregate weight for the emergency response mode is... 、 、 Calculate the global shared layer parameters for emergency response mode. .

[0084] For the routine inspection model, the aggregate weights of the three participating hydropower stations (Three Gorges, Gezhouba, and Xiangjiaba) are respectively... 、 、 Each hydropower station calculates a weighted sum in the encrypted space. Then, the global shared layer parameters are obtained through a distributed decryption protocol. .

[0085] Step 600: Combine the global shared layer parameters and the unique layer parameters to generate a multi-granularity global model library.

[0086] For the unique layer parameters of each model granularity Each hydropower station operates independently and does not participate in inter-hydropower station data transmission. Each hydropower station uses a model with specific granularity tailored to its local conditions. , global shared layer parameters With unique layer parameters Combine to generate a multi-granularity global model library .

[0087] Furthermore, the parameter combination operation specifically includes: for each granularity model , global shared layer parameters Assigning values ​​to the shared underlying feature extraction layer of the model, and transferring the unique layer parameters The values ​​are assigned to the higher-level decision-making layer of the model to complete the loading of global model parameters.

[0088] Figure 4 The study demonstrates the convergence trends of local and global model accuracy for three hydropower stations (Three Gorges, Gezhouba, and Xiangjiaba) during 20 rounds of federated learning.

[0089] Step 700: Deploy the multi-granularity global model library to the robot and perform mode perception model switching.

[0090] Each hydropower station distributes a multi-granularity global model library to its local robot, which then stores all models locally. The robot then uses tags based on its current working mode. The system loads models of the corresponding granularity from the local model library into memory. When the robot switches working modes, it unloads the current model and loads the model corresponding to the new mode, achieving a fast switching process.

[0091] Furthermore, the mapping relationship between working modes and model granularity is determined by a preset mapping table. The emergency response mode is mapped to a high-precision complete model to ensure diagnostic accuracy, the routine inspection mode is mapped to a medium-precision trimmed model to balance accuracy and computational efficiency, and the energy-saving maintenance mode is mapped to a low-precision minimalist model to minimize energy consumption.

[0092] When the robot uses the loaded model to diagnose faults based on real-time collected equipment monitoring data, the model outputs the probability distribution of each fault category. The category with the highest probability is selected as the predicted fault category, and a state is considered normal when the maximum probability is below the confidence threshold. Specifically, the confidence threshold ranges from (0,1) and is set according to the false alarm tolerance of each operating mode. Emergency response mode typically uses a lower confidence threshold to improve the fault detection rate, while regular inspection mode uses a higher confidence threshold to reduce the false alarm rate.

[0093] Each hydropower station will distribute the updated multi-granularity global model library to its local robots. Taking the Three Gorges Hydropower Station as an example, robot R1-1 is currently in emergency response mode and is loading the complete model from its local model library. Loading into memory. After robot R1-1 completes the emergency troubleshooting task, the working mode switches from emergency response mode to regular inspection mode. The robot unloads the complete model and loads the trimmed model from the local model library. The model switching process took 0.8 seconds to load into memory. Robot R1-1 used a pruned model to diagnose faults based on real-time collected equipment monitoring data. Input data included multi-channel time-series data such as bearing vibration signals, turbine pressure fluctuations, and generator temperature. The model output a probability distribution for four fault categories: 'Normal 0.12, Bearing Fault 0.73, Abnormal Turbine Vibration 0.08, Generator Overheating 0.07'. The category with the highest probability, 'Bearing Fault', was selected as the prediction result. Since the maximum probability of 0.73 was higher than the confidence threshold of 0.65 set for the conventional inspection mode, it was determined to be a bearing fault, and an alarm message was generated.

[0094] In this embodiment of the application, in order to improve the safety of cross-hydropower station aggregation, the following steps are included after step 500: Step 510: Each hydropower station performs consistency verification on the reconstructed or decrypted global shared layer parameters. Specifically, each hydropower station calculates the hash value of the global parameters. They exchange hash values ​​and compare them. The aforementioned hash function... Input a global shared layer parameter vector and output a fixed-length digital digest value. A one-way mapping ensures that small differences in parameters result in significantly different digest values.

[0095] Step 520: When an inconsistency in the secret fragment sent by a hydropower station or a failure to verify the encryption parameters is detected, the distributed verification protocol is triggered. The distributed verification protocol requires each hydropower station to resubmit its intermediate calculation results, and the abnormal node is confirmed by a majority vote of the hydropower stations.

[0096] Furthermore, in the voting confirmation mechanism of most hydropower stations, each normal hydropower station independently verifies the intermediate calculation results of suspicious nodes. If the verification is successful, a vote of approval is cast; if the verification fails, a vote of disapproval is cast. When the number of disapproval votes exceeds half of the total number of participants, the suspicious node is determined to be an abnormal node.

[0097] Step 530: Isolate the abnormal node from the participant list and re-execute the aggregation process of step 500 to eliminate the influence of malicious participants.

[0098] In this embodiment of the application, in order to dynamically optimize the mode-encryption mapping strategy, the following steps are included after step 700: Step 710: Each hydropower station records the communication overhead of cross-hydropower station aggregation. Calculation time and model performance metrics under each mode .

[0099] Specifically, communication overhead Calculation time and model performance metrics These are all statistical values ​​for the current federated learning round, including communication overhead. The calculation time is the total number of bytes transmitted by each hydropower station in the current round. This represents the total time taken from the start of encryption to the completion of aggregation in the current round, in seconds.

[0100] Furthermore, model performance metrics This is the weighted average of the fault diagnosis accuracy of the global model after the current round of updates on the local validation sets of each hydropower station. The calculation formula is as follows: ,in For hydroelectric power station index, To determine the number of participating hydropower stations, For the first Number of validation set samples for each hydropower station For the first Accuracy on the validation set of a hydropower station.

[0101] Step 720: Analyze model performance metrics Calculation time and communication overhead Perform mean normalization based on the range separately to eliminate dimensional differences between different physical quantities. Calculate the comprehensive evaluation index. ,in These are the weighting coefficients.

[0102] Specifically, weighting coefficients satisfy and Each coefficient is configured according to the importance preference of the hydropower station for model performance, computational efficiency, and communication costs.

[0103] Step 730: When the comprehensive evaluation indicators When the threshold is lower than the preset threshold, adjust the weight of real-time requirements. And security requirement weight The values ​​were used to reclassify the high real-time mode group and the high security mode group.

[0104] Figure 7 The dynamic trends of model performance metrics, computation time, communication overhead, and comprehensive evaluation metric Q were shown during 30 rounds of federated learning.

[0105] In this embodiment, the preset threshold is based on the historical federated learning rounds. The statistical distribution of the values ​​is determined and is usually set to historical. The mean value minus one standard deviation, when the current round A strategy adjustment is triggered when the value is significantly lower than the historical level.

[0106] This application also provides a distributed collaborative scheduling system for multiple robots in hydropower stations based on edge federated learning. The system includes: a mode requirement weight calculation module, used to calculate real-time requirement weights and security requirement weights based on the latency parameters of the working mode; a multi-granularity model library generation module, used to generate multi-granularity models with shared underlying parameter structures and unique layer parameter structures; a local federated aggregation module, used to perform local federated aggregation within the hydropower station to generate hydropower station-level aggregated parameters; a group encryption module, used to group modes according to real-time requirement weights and security requirement weights and select corresponding encryption algorithms to encrypt shared layer parameters; a cross-hydropower station encryption aggregation module, used to perform interpolation recovery on secret fragments and / or perform homomorphic aggregation on ciphertext parameters to generate global shared layer parameters; and a model combination and distribution module, used to combine global shared layer parameters and local unique layer parameters to generate a multi-granularity global model library and distribute it to the robots.

[0107] Based on the same technical concept, embodiments of this application also provide an electronic device that can implement the distributed collaborative scheduling method for multi-robots in hydropower stations based on edge federated learning provided in the above embodiments of the present invention. In one embodiment, the electronic device can be a server, a terminal device, or other electronic devices. Figure 8 As shown, the electronic device may include: At least one processor and a memory connected to the at least one processor. In this embodiment of the invention, the specific connection medium between the processor and the memory is not limited. Figure 8 The example used is the connection between the processor and memory via a bus. The bus... Figure 8The connections between other components are indicated by thick lines and are for illustrative purposes only, not as limiting information. Buses can be divided into address buses, data buses, control buses, etc., but for ease of representation, [the specific bus type is not shown here]. Figure 8 The processor is represented by a single thick line, but this does not imply that there is only one bus or one type of bus. Alternatively, a processor can also be called a controller; there are no restrictions on the name.

[0108] In this embodiment of the invention, the memory stores instructions executable by at least one processor. By executing the instructions stored in the memory, the at least one processor can execute the previously discussed distributed collaborative scheduling method for multi-robots in hydropower stations based on edge federated learning. The processor can implement... Figure 8 The functions of each module in the device shown.

[0109] The processor is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory and calling data stored in memory, it can monitor the device's various functions and process data, thereby enabling overall monitoring of the device.

[0110] In an alternative design, the processor may include one or more processing units. The processor may integrate an application processor and a modem processor, wherein the application processor primarily handles the operating system, user interface, and applications, while the modem processor primarily handles wireless communication. It is understood that the modem processor may also not be integrated into the processor. In some embodiments, the processor and memory may be implemented on the same chip; in some embodiments, they may also be implemented separately on separate chips.

[0111] The processor can be a general-purpose processor, such as a CPU, digital signal processor, application-specific integrated circuit, field-programmable array, or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning disclosed in the embodiments of this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0112] Memory, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory can include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic memory, magnetic disk, optical disk, etc. Memory is any other medium capable of carrying or storing desired program code in the form of instructions or data structures, and accessible by a computer, but is not limited thereto. In embodiments of the present invention, memory can also be a circuit or any other device capable of implementing storage functions, used to store program instructions and / or data.

[0113] By designing and programming the processor, the code corresponding to the distributed collaborative scheduling method for multi-robot hydropower stations based on edge federated learning described in the foregoing embodiments can be embedded into the chip, enabling the chip to execute the steps of the method described in the foregoing embodiments during runtime. How to design and program the processor is a technique well-known to those skilled in the art and will not be elaborated upon here.

[0114] Based on the same inventive concept, embodiments of the present invention also provide a storage medium storing computer instructions, which, when executed on a computer, cause the computer to execute the previously described distributed collaborative scheduling method for multi-robots in hydropower stations based on edge federated learning.

[0115] In some alternative embodiments, the present invention also provides a method for distributed collaborative scheduling of multiple robots in a hydropower station based on edge federated learning, which can also be implemented as a program product including program code. When the program product is run on a device, the program code is used to cause the control device to perform the steps in the method for distributed collaborative scheduling of multiple robots in a hydropower station based on edge federated learning according to various exemplary embodiments of the present invention as described in this specification.

[0116] It should be noted that although several units or sub-units of the apparatus have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the invention, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units. Furthermore, although the operation of the method of the invention is described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a server, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] Program code for performing the operations of this invention can be written using any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0120] In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A distributed cooperative scheduling method for multiple robots in a hydropower station based on edge federated learning, characterized in that, include: Obtain the working mode definitions of robots in each hydropower station participating in cross-hydropower station federated learning. The working mode definitions include routine inspection mode, emergency response mode and energy-saving maintenance mode, and determine the real-time requirement weight and safety requirement weight of each working mode. Based on the performance requirements and energy consumption constraints of each model, a multi-granularity model library is generated, including high-precision complete models, medium-precision pruned models, and low-precision simplified models. Each granularity model shares the parameter structure of the underlying feature extraction layer and maintains a unique parameter structure in the high-level decision layer. The parameter index set of the shared layer and the parameter index set of the unique layer of each granularity model are determined. Within each hydropower station, local federated aggregation is performed on robots operating in the same mode. A weighted average is calculated based on the number of training samples for each robot to generate hydropower station-level aggregation parameters. Based on the relationship between the weights of real-time requirements and security requirements, the working modes are divided into a high real-time mode group and a high security mode group. For the hydropower station-level aggregate parameters of the high real-time mode group, the shared layer parameters are extracted, a polynomial is constructed on a finite domain and the evaluation point value is calculated, and the parameters are divided into secret fragments. For the hydropower station-level aggregate parameters of the high security mode group, the shared layer parameters are extracted and the ciphertext parameters are generated using a fully homomorphic encryption algorithm. Perform encrypted aggregation across hydropower stations. After secret fragment exchange of high real-time mode, use Lagrange interpolation formula to restore the original parameters of each hydropower station and calculate weighted average based on the number of training samples to generate global shared layer parameters. The ciphertext parameters in the high-security mode are calculated by homomorphic addition in the ciphertext space to generate a weighted sum, and then decrypted using a distributed decryption protocol to generate globally shared layer parameters. Each hydropower station combines the globally shared layer parameters with its local unique layer parameters to generate a multi-granularity global model library, which is then distributed to the robot. The robot loads the model of the corresponding granularity from the local model library according to the current working mode.

2. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The determination of the real-time requirement weight and security requirement weight for each working mode specifically includes: Obtain the maximum tolerable aggregate latency and the baseline aggregate latency for the working mode; Calculate the ratio of the maximum tolerable aggregation delay to the sum of the maximum tolerable aggregation delay and the baseline aggregation delay; Subtracting the ratio from 1 yields the real-time demand weight; The security requirement weight is equal to 1 minus the real-time requirement weight; Among them, the maximum tolerable aggregation latency is the maximum allowed time from initiating aggregation to completing model update for each working mode, which is pre-configured according to the urgency of the working mode; the baseline aggregation latency is the theoretical aggregation latency when using the lowest encryption strength, which is obtained through pre-testing as a normalized baseline value.

3. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The determination of the parameter index set for the shared layer and the parameter index set for each granularity model's unique layer specifically includes: Analyze the network architecture of each granular model layer by layer, starting from the input layer, and check whether the layer type, convolution kernel size, number of channels and activation function of each layer are completely consistent; The network layers that meet the consistency conditions are marked as shared layers, and a set of shared layer indexes is generated. High-level decision-making layers that do not meet the consistency condition are marked as unique layers, and a unique layer index set is generated for each granularity model. Among them, the shared low-level feature extraction layer of each granularity model consists of multiple convolutional layers and pooling layers, which are used to extract general time-frequency domain features of device operating status; each granularity model maintains a unique layer parameter structure in the high-level decision layer, which consists of fully connected layers, and contains different numbers of fully connected layers according to the model granularity.

4. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The process of extracting shared layer parameters from the hydropower station-level aggregated parameters of the high real-time mode group, constructing a polynomial over a finite domain and calculating the evaluation point values, and dividing the data into secret slices specifically includes: Each hydropower station generates a local public-private key pair and exchanges public keys to build an encrypted communication channel; For each parameter in the shared layer parameter vector, a polynomial is constructed over a finite field. The constant term of the polynomial is the parameter value, and the coefficients of the remaining terms are uniformly and randomly selected in the finite field. The degree of the polynomial is equal to the safety threshold minus 1. Calculate the values ​​of the polynomial at different evaluation points for the number of participating hydropower stations, and generate the corresponding number of secret fragments; Each secret segment is sent to its corresponding participating hydropower station; Among these requirements, the prime numbers in the finite field must be greater than the maximum absolute value of the parameter to ensure the uniqueness of the mapping of the parameter value within the finite field; the random coefficients are generated by a cryptographically secure random number generator to ensure the unpredictability of secret sharding; and the security threshold ranges from at least 2 to no more than the number of hydropower stations involved.

5. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The process of performing encrypted aggregation across hydropower stations involves exchanging secret fragments in a high-real-time mode, recovering the original parameters of each hydropower station using the Lagrange interpolation formula, and calculating a weighted average based on the number of training samples to generate global shared layer parameters. Specifically, this includes: Each hydropower station receives secret fragments from other hydropower stations and selects at least a security threshold number of secret fragments. The original parameter values ​​are restored using the Lagrange interpolation formula. Specifically, the polynomial values ​​corresponding to each secret partition are multiplied by the corresponding Lagrange interpolation coefficients and then summed. The Lagrange interpolation coefficients are the product of the evaluation points of other partitions except the current partition divided by the product of the differences between the evaluation points. After restoring the original shared layer parameters of each hydropower station, the total number of training samples of each hydropower station is used as the weight to calculate the weighted average of the original parameters of all hydropower stations, thus obtaining the global shared layer parameters.

6. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The ciphertext parameters of the high-security mode are calculated by homomorphic addition in the ciphertext space to generate a weighted sum, and then decrypted using a distributed decryption protocol to generate globally shared layer parameters, specifically including: Calculate the aggregate weight of each hydropower station. The aggregate weight is equal to the number of training samples for that hydropower station divided by the sum of the number of training samples for all participating hydropower stations. The aggregate weight is determined based on the proportion of the number of training samples for each hydropower station to the total number of samples. In the ciphertext space, the products of the ciphertext parameters of each hydropower station and their corresponding aggregation weights are summed, and the global ciphertext parameters are calculated using the additive homomorphic property of fully homomorphic encryption. Each hydropower station uses its own private key to participate in the distributed decryption protocol, jointly decrypting the global ciphertext parameters to obtain the global shared layer parameters in plaintext form.

7. The distributed cooperative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, Each hydropower station combines globally shared layer parameters with locally unique layer parameters to generate a multi-granularity global model library, specifically including: The unique layer parameters of each granularity model are aggregated independently within each hydropower station and do not participate in the transmission between hydropower stations. Each hydropower station assigns the global shared layer parameters to the shared bottom feature extraction layer of the model and assigns the unique layer parameters to the high-level decision layer of the model according to the model granularity used in its local mode, thus completing the parameter loading of the global model. Generate a multi-granularity global model library that includes high-precision complete models, medium-precision cropped models, and low-precision minimalist models.

8. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, Following the execution of cross-hydropower station cryptographic aggregation is a consistency verification step: Each hydropower station calculates the hash value of the reconstructed or decrypted global shared layer parameters, exchanges the hash values ​​with each other, and compares them. When an inconsistency is detected in the secret fragment sent by a hydropower station or the encryption parameter verification fails, the distributed verification protocol is triggered, requiring each hydropower station to resubmit its intermediate calculation results. Each normal hydropower station independently verifies the intermediate calculation results of the suspicious node. If the verification passes, it votes in favor; if the verification fails, it votes against. When the number of opposing votes exceeds half of the total number of participants, the suspicious node is determined to be an abnormal node. Isolate the abnormal node from the participant list, re-execute the aggregation process, and eliminate the influence of malicious participants.

9. The distributed collaborative scheduling method for multiple robots in a hydropower station based on edge federated learning according to claim 1, characterized in that, The process of distributing the multi-granularity model library also includes a dynamic optimization step: Each hydropower station records the communication overhead, computation time, and model performance indicators for cross-hydropower station aggregation. The communication overhead is the total number of bytes transmitted by each hydropower station in the current round, the computation time is the total time taken from the start of encryption to the completion of aggregation in the current round, and the model performance indicator is the weighted average of the fault diagnosis accuracy of the global model on the local validation set of each hydropower station. The model performance index, computation time, and communication overhead are respectively subjected to mean normalization based on the range. The comprehensive evaluation index is calculated as the weighted difference between the normalized model performance index and the computation time and communication overhead. The sum of the weight coefficients is 1 and all are greater than 0. When the comprehensive evaluation index is lower than the preset threshold, the values ​​of the real-time requirement weight and the security requirement weight are adjusted, and the high real-time mode group and the high security mode group are reclassified. Based on the adjusted pattern grouping update pattern-encryption type mapping strategy, the new encryption method selection rules are applied in the next round of federated learning.

10. A distributed collaborative scheduling system for multi-robots in hydropower stations based on edge federated learning, used to execute the method as described in any one of claims 1 to 9, characterized in that, The system includes: The mode requirement weight calculation module is used to calculate the real-time requirement weight and the security requirement weight based on the latency parameter of the working mode. The multi-granularity model library generation module is used to generate multi-granularity models with shared underlying parameter structures and unique layer parameter structures; The local federated aggregation module is used to perform local federated aggregation within the hydropower station to generate hydropower station-level aggregation parameters; The block encryption module is used to group patterns according to real-time requirement weights and security requirement weights, and select the corresponding encryption algorithm to encrypt the shared layer parameters. The cross-hydropower station encryption aggregation module is used to perform interpolation recovery on secret fragments and / or perform homomorphic aggregation on ciphertext parameters to generate global shared layer parameters; The model combination and distribution module is used to combine globally shared layer parameters and locally unique layer parameters to generate a multi-granularity global model library and distribute it to the robot.

Images