A power equipment operation state diagnosis system and method based on distributed learning

Abstract

The application discloses a power equipment operation state diagnosis system and method based on distributed learning, relates to the technical field of power operation and maintenance, and comprises a power operation and maintenance platform, a cloud server and power equipment. The power equipment collects real-time operation data of itself, predicts the operation state of the power equipment, generates operation and maintenance state information, trains an operation state diagnosis model, sequentially performs encryption and DEK+KEK encryption encapsulation processing on model parameters of the local model, and generates a local model digest. The cloud server performs encryption parameter security aggregation and integrity verification, generates aggregated global model parameters. The power equipment further adjusts the aggregated global model parameters to determine the operation state diagnosis model. The power operation and maintenance platform generates operation and maintenance parameters of each power equipment and delivers the operation and maintenance parameters to corresponding power equipment. The application can form a collaborative processing process of "data processing-local training-security aggregation-global model adaptation-dynamic operation and maintenance adjustment".

Description

Technical Field

[0001] This application relates to the field of power operation and maintenance technology, and in particular to a power equipment operation status diagnosis system and method based on distributed learning. Background Technology

[0002] In related technologies, power equipment condition diagnosis and data transmission solutions mainly rely on centralized model training, traditional static encryption, and fixed operation and maintenance strategies. Although these can meet basic needs in simple power grid environments, with the surge in the number of terminal devices and the increase in operating complexity, the technical system has gradually exposed several prominent problems: The diagnostic models suffer from weak generalization ability and insufficient real-time performance: Most diagnostic solutions adopt a "terminal data collection - cloud-based centralized training" architecture. On the one hand, power terminals are widely distributed (such as remote transformer substations and distributed photovoltaic sites), and the uploading of a large amount of raw multi-parameter data requires high bandwidth, which can easily cause transmission congestion during peak electricity consumption periods, resulting in prolonged model training latency (a single training session often takes more than 30 minutes), which cannot meet the real-time requirements of fault diagnosis. On the other hand, the load characteristics of terminals vary significantly under different scenarios (such as the load fluctuations in residential transformer substations being gradual, while the load impacts in industrial transformer substations are large), making it difficult for centralized models to adapt to local conditions. The diagnostic accuracy for scenario-specific faults such as 485 loose connections and circuit interference is generally less than 70%, which can easily lead to the problem of "different effects from the same model," failing to meet the needs of differentiated operation and maintenance.

[0003] The data security and privacy protection system is inadequate: Power multi-parameter data contains sensitive information (such as user electricity load curves and power grid topology correlation data), and existing security protection solutions have significant shortcomings: During the training phase, directly uploading raw data to the cloud can easily lead to the leakage of user privacy, and there is a lack of effective data anonymization mechanisms; During the parameter transmission phase, most rely on basic TLS encryption and do not design a dedicated verification mechanism for model parameters, which poses a risk of malicious nodes forging parameters and cloud tampering with aggregation results; During the terminal access phase, static password authentication is mostly used, which is easily cracked and counterfeited. Counterfeit terminals may upload false data after access, interfering with the global diagnostic results and seriously threatening the integrity and transmission security of power grid data.

[0004] There are still obvious shortcomings in the research on basic homomorphic encryption technology protection models: there is a lack of dynamic trust management among distributed learning nodes, the homomorphic encryption operation has a large overhead (the terminal takes more than 5 seconds to encrypt once), and it cannot adapt to the computing power requirements of lightweight terminals (such as old smart meters); moreover, most studies only focus on a single link of "diagnosis" or "transmission", and cannot fundamentally solve the problem of coordinating efficiency, security and collaboration. Summary of the Invention

[0005] The purpose of this application is to provide a power equipment operation status diagnosis system and method based on distributed learning, which can ensure the safety of the system and form a collaborative processing flow of "data processing - local training - security aggregation - global model adaptation - dynamic operation and maintenance adjustment".

[0006] To achieve the above objectives, this application provides the following solution: In the first aspect, this application provides a power equipment operation status diagnosis system based on distributed learning, including a power operation and maintenance platform, a cloud server, and various power equipment; the cloud server and each power equipment communicate using a two-factor authentication and key negotiation mechanism.

[0007] The power equipment is used to collect its own operating data in real time, and based on its own operating data and the current operating status diagnostic model, predict the operating status of the power equipment, generate operation and maintenance status information, train the operating status diagnostic model to generate a local model, and perform encryption and DEK+KEK encryption encapsulation processing on the model parameters of the local model in sequence, and generate a local model summary.

[0008] The cloud server is used to receive encapsulation information and local model digests uploaded by various power devices through the TLS channel, and to perform secure aggregation and integrity verification of encrypted parameters to generate aggregated global model parameters.

[0009] The power equipment is also used to adjust the aggregated global model parameters to update the operational status diagnostic model.

[0010] The power operation and maintenance platform is used to generate operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distribute them to the corresponding power devices.

[0011] Secondly, this application provides a method for diagnosing the operating status of power equipment based on distributed learning, including: Power equipment collects its own operating data in real time, and predicts the operating status of the power equipment and generates operation and maintenance status information based on its own operating data and the current operating status diagnostic model. In addition, it trains the operating status diagnostic model to generate a local model, performs encryption and DEK+KEK encryption encapsulation processing on the model parameters of the local model in sequence, and generates a local model summary.

[0012] The cloud server receives the encapsulation information and local model digest uploaded by each power device through the TLS channel, performs secure aggregation and integrity verification of encrypted parameters, and generates aggregated global model parameters.

[0013] The power equipment adjusts the aggregated global model parameters to update the operational status diagnostic model.

[0014] The power operation and maintenance platform generates operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distributes these parameters to the corresponding power devices.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a power equipment operation status diagnosis system and method based on distributed learning. The power equipment operation status diagnosis system includes a power operation and maintenance platform, a cloud server, and various power equipment. It collects real-time operation data from the power equipment, predicts the operating status of the power equipment, trains an operation status diagnosis model, encrypts the model parameters of the local model, and generates a local model summary. The cloud server performs secure aggregation and integrity verification of the encrypted parameters. The power equipment also adjusts the aggregated global model parameters (global model adaptation), and the power operation and maintenance platform generates operation and maintenance parameters for the power equipment, forming a collaborative processing flow of "data processing—local training—secure aggregation—global model adaptation—dynamic operation and maintenance adjustment." The cloud server and each power equipment communicate using a two-factor authentication and key negotiation mechanism, effectively avoiding the risks of malicious nodes forging nodes, uploading false data, interfering with global diagnostic results, and tampering with aggregation results. The encrypted model parameters are double-encrypted using DEK+KEK technology and transmitted through a TLS channel to ensure data security. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the structure of a power equipment operation status diagnosis system based on distributed learning in one embodiment of this application; Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] Explanation of technical terms: KARN, Knowledge Augmented Reasoning Network, is a type of knowledge-augmented recursive reasoning network.

[0019] MDP, Markov Decision Process.

[0020] TLS, Transport Layer Security.

[0021] AES, Advanced Encryption Standard.

[0022] SM2, SM2 Elliptic Curve Cryptography, is a national standard elliptic curve public-key cryptographic algorithm.

[0023] STFT, Short-Time Fourier Transform, is used to extract frequency domain features.

[0024] SGD, or Stochastic Gradient Descent.

[0025] TA, Trust Authority, is responsible for key generation, identity authentication, and security management.

[0026] DEK, Data Encryption Key, is used to directly encrypt model parameters and data.

[0027] KEK, Key Encryption Key, is a key used to encrypt data and implement layered encryption.

[0028] CRC, Cyclic Redundancy Check, is used to verify the integrity of data transmission.

[0029] Softmax is the normalized exponential function.

[0030] Sigmoid, a sigmoid activation function.

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] In one exemplary embodiment, such as Figure 1As shown, a power equipment operation status diagnosis system based on distributed learning is provided, including a power operation and maintenance platform, a cloud server, and various power equipment; the cloud server and each power equipment communicate using a two-factor authentication and key negotiation mechanism.

[0034] The power equipment is used to collect its own operating data in real time, and based on its own operating data and the current operating status diagnostic model, predict the operating status of the power equipment, generate operation and maintenance status information, train the operating status diagnostic model to generate a local model, and perform encryption and DEK+KEK encryption encapsulation processing on the model parameters of the local model in sequence, and generate a local model summary.

[0035] The cloud server is used to receive encapsulation information and local model digests uploaded by various power devices through the TLS channel, and to perform secure aggregation and integrity verification of encrypted parameters to generate aggregated global model parameters.

[0036] The power equipment is also used to adjust the aggregated global model parameters to update the operational status diagnostic model.

[0037] The power operation and maintenance platform is used to generate operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distribute them to the corresponding power devices.

[0038] The aforementioned power equipment operation status diagnosis system based on distributed learning can form a collaborative processing flow of "data processing - local training - secure aggregation - global model adaptation - dynamic operation and maintenance adjustment"; it can effectively avoid the risks of malicious nodes forging nodes, uploading false data, interfering with global diagnostic results and tampering with aggregation results, and ensure the security of transmitted data.

[0039] As an example, both the power operation and maintenance platform and the cloud server are bidirectionally connected to each power device.

[0040] The power equipment operation status diagnosis system based on distributed learning adopts a four-layer distributed collaborative architecture, consisting of a perception layer, a distributed layer, an aggregation layer, and an application layer from bottom to top. Power equipment simultaneously serves as both the perception layer (collecting its own operation data in real time) and the distributed layer (training local models on local servers to predict operation status). The cloud server acts as the aggregation layer, aggregating parameters from the local models. The power operation and maintenance platform serves as the application layer, generating operation and maintenance parameters for each power device. With the cloud server at its core and multiple power devices as nodes, the system forms an integrated system of "local terminal training, secure cloud aggregation, global collaborative diagnosis, and closed-loop intelligent operation and maintenance."

[0041] The system is constructed using a four-layer architecture: "Perception Layer - Distributed Layer - Aggregation Layer - Application Layer". The Perception Layer includes... The system collects electrical, communication, and environmental parameters from individual power devices (smart meters, dedicated transformer terminals, etc.). The distributed layer, acting as the terminal client (local server), performs data preprocessing (RC + 5th order Butterworth filtering), local improved KARN training, and parameter encryption. The aggregation layer, including a private cloud server, is responsible for the global model, double-mask secure aggregation (tolerating offline operation of some power devices), and verifiable generation. The application layer interfaces with the operation and maintenance platform, forming a closed loop. System interaction follows a "distributed-first" approach, with terminals transmitting encrypted parameters, and the cloud (cloud server) only aggregating and distributing them.

[0042] The perception layer consists of numerous power terminal nodes (power equipment), including smart meters, data acquisition terminals, and distributed energy controllers. It is responsible for real-time acquisition and preliminary preprocessing of multiple parameters such as voltage, current, bit error rate, communication latency, and electromagnetic interference. The distributed layer is deployed within each power terminal node, carrying a lightweight distributed learning client. It performs multi-parameter feature extraction, local training of the improved KARN model, data privacy protection, and model parameter encryption. It does not upload raw data, only encrypted model update parameters. The aggregation layer consists of a private cloud server and a Trust Center (TA). It handles global model initialization, secure aggregation of encrypted parameters, integrity verification, zero-knowledge proof generation, and dynamic node management, ensuring the security and reliability of the distributed learning process. The application layer interfaces with the power operation and maintenance platform. Based on the global model, it completes multi-parameter fault collaborative diagnosis, fault location, operation and maintenance strategy generation, and dynamic adjustment of sampling and transmission parameters, forming a closed-loop system of "acquisition—training—aggregation—diagnosis—optimization."

[0043] In another exemplary embodiment, regarding the real-time acquisition of its own operational data and the prediction of the operating status of the power equipment based on its own operational data and the current operational status diagnostic model, the power equipment is used for: It collects its own operational data in real time to form an operational data stream; the operational data includes various types of operational data.

[0044] The running data stream is sliced ​​using a sliding time window, and the validity of the data within each sliding time window is determined by multi-parameter cross-validation, thus extracting the valid data.

[0045] Features are extracted sequentially from the valid data within each sliding time window to generate corresponding feature datasets. The feature datasets include time-domain features, frequency-domain features, and correlation features corresponding to each type of running data. The correlation features include Pearson correlation coefficient and anomaly index.

[0046] Each feature dataset is sequentially input into the operational status diagnostic model to obtain the operational status of the power equipment.

[0047] The expression for determining the validity of multiple parameters is as follows: ; The expression for the Pearson correlation coefficient is: ; The expression for the anomaly index is: ; In the formula, Indicates the first n Power equipment exist A valid indicator of the runtime data at any given time. A value of 1 indicates electrical equipment. exist The runtime data at any given time is valid data and the terminal is trainable; Indicates the number of data types being processed. Indicates in Time of the first n The first power equipment k In the Pearson correlation coefficient formula, for this type of operational data, and These represent the first and second elements within the preset sliding window. The first power equipment Type of running data sequence and the first This type of operation runs a sequence of data, rather than data values ​​at a single moment. Indicates the first k The valid range of this type of runtime data For the first k The minimum valid value of this type of runtime data. For the first k The maximum valid value of this type of runtime data. Indicates the first n Training status indicators for each power device; This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. In the multi-parameter validity determination formula, the corresponding indicator function takes the value 1 when the running data is within the valid range. In the anomaly index formula, the corresponding indicator function takes the value 1 when the running data exceeds the valid range. Indicates the first The first power equipment Type of running data and the first The Pearson correlation coefficient between the two types of operational data, where , and usually , used to calculate the first n The first power equipment Type of runtime data and the first Linear correlation of various types of operational data (t) represents the th n Each electrical device at time The first collection k Types of runtime data, Indicates the first n The first power equipment m For this type of operational data, Cov(·) represents the covariance function. Both represent the variance function and are calculated based on the sequence of running data within a preset sliding window. This represents an anomaly index, which is a weighted quantification of anomalies across all operational data. n Real-time anomaly level of individual power equipment The larger the value, the more significant the number. Each electrical device The higher the degree of operational anomaly at any given moment, Indicates the first k The weights of the types of runtime data satisfy .

[0048] Short-term time-series statistics for each type of operational data are extracted to reflect the stability of operational data fluctuations. Core indicators include the moving average and peak deviation, using a window length. The sliding window calculates the first... The first power equipment Mean of the data for each type of operation: ; In the formula, For power equipment At any moment The first collection The mean of the types of running data, For power equipment At any moment The first collection Real-time values ​​of class runtime data. Corresponding voltage Corresponding current , Corresponding bit error rate , Corresponding communication delay , Corresponding electromagnetic interference ; The sampling frequency for running data; This is the index of the sampling points within the window.

[0049] Peak deviation: The maximum deviation between the real-time value and the moving average of the calculated operating data, reflecting sudden anomalies. ; For power equipment At any moment The first collection Peak deviation of various types of running data.

[0050] Short-time Fourier transform is used to extract the frequency domain characteristics of the operating data, capturing periodic electromagnetic interference and load fluctuations. The core indicators include the main frequency and the main frequency amplitude. clock speed It is in the window Within a given range, the frequency corresponding to the maximum power spectral density is given by the formula: ; In the formula For running data At any moment ,frequency The power spectral density at that point was obtained through STFT calculation, with the frequency range limited to [frequency range missing]. (Nyquist frequency), which conforms to the frequency distribution of electromagnetic interference (1kHz~100MHz) and load fluctuation (0.1Hz~50Hz) in power systems.

[0051] main frequency amplitude Corresponding clock frequency The peak value of the power spectral density reflects the intensity of that frequency component.

[0052] The extracted feature dataset is denoted as .

[0053] As an example, the number of data types K is 5, corresponding to voltage. Current Bit error rate Communication delay Electromagnetic interference The valid range of operating data, for example, voltage. Bit error rate , Training status flags This indicates that CPU utilization is ≤70% and memory remaining is ≥20%.

[0054] like If so, the corresponding runtime data will be stored in a local encrypted database (SQLCipher encrypted); if If the data collection is triggered again (retry ≤ 3 times, 100ms interval), and the data collection is failed repeatedly, an alarm will be sent to the cloud.

[0055] In another exemplary embodiment, the method for determining the validity of multiple parameters and generating the feature dataset can ensure the quality of distributed training data. It determines the validity of collected parameters, retains only valid data, and updates the local power equipment historical fault dataset for local training (training the operational status diagnostic model to generate the local model). For operational data that fails the validity determination, the power equipment does not write it into the local power equipment historical fault dataset, nor does it participate in operational status prediction or local model training. The system prioritizes re-collection; if continuous re-collection still does not meet the validity range, the parameter is marked as an abnormal input and used only for fault alarms and log recording, not as a basis for model parameter updates, to avoid abnormal sampled values ​​polluting the local training process.

[0056] In another exemplary embodiment, the power equipment also includes a local server.

[0057] The local server is used to update the local power equipment historical fault dataset based on the generated feature dataset and the actual operating status of the corresponding power equipment.

[0058] When the model update conditions are met, the operation status diagnosis model is trained using the local power equipment historical fault dataset (including labeled samples such as 485 disconnection, loose connection, circuit interference, etc., with a sample size of ≥1000 groups), generating a local model. The model parameters of the local model are then encrypted and encapsulated using DEK+KEK encryption. Finally, a local model summary is generated based on the model parameters of the local model.

[0059] In the training process of the operational status diagnostic model, differential privacy technology is used to add noise to the parameter gradients of each network layer during backpropagation.

[0060] The aggregated global model parameters are adjusted, and the runtime status diagnostic model is determined based on the adjusted global model parameters.

[0061] As an example, the local server uses the symmetric encryption algorithm AES-256 to encrypt the model parameters of the local model; it encapsulates the data encryption key DEK used to encrypt the local model parameters using KEK encryption to obtain encapsulation information; and transmits the encapsulation information, local model digest, predicted running status, signature, and timestamp to the cloud server through a TLS channel; the TLS channel is established between the cloud server and each of the power devices based on a shared session key.

[0062] As an example, the model update condition is reaching the model update cycle, or the predicted running state is an abnormal fault state.

[0063] As an example, during the training of the runtime diagnostic model, the following methods are employed: - Differential privacy techniques add Laplacian noise to the model gradient, as shown in the formula: ; In the formula, Indicates the first The gradient of the loss function after adding differential privacy noise to a power device. Indicates the first The gradient of the original loss function is calculated for each power device based on local training samples. The loss function uses mean squared error. Indicates the first The first power equipment The actual fault label corresponding to each training sample This represents the corresponding model prediction label; The noise intensity coefficient, Balancing privacy and model accuracy; This indicates a mean of 0 and a scale parameter of . The Laplacian noise, and the noise having the same gradient dimension as the original loss function, represents gradient sensitivity. Indicates privacy budget, The smaller, The larger the value, the greater the amplitude of the added noise, and the stronger the privacy protection.

[0064] Stochastic gradient descent (SGD) is employed, and the first step is optimized based on the noisy gradient. Local model parameters of each power device The iterative formula is: ; In the formula, Indicates the first Local model parameters at the next iteration Indicates the first The local model parameters after the next iteration. This represents the noisy gradient used for parameter updates. The learning rate is set to 50 iterations (to ensure model convergence, with a single training iteration time ≤ 10s, meeting the real-time requirements of power equipment). The convergence criterion is the difference in the loss function between two adjacent iterations. .

[0065] After the power equipment completes local training, the model parameters of the local model are... Encryption is performed, and a verifiable digest (local model digest) is generated for integrity verification during cloud server aggregation. The symmetric encryption algorithm AES-256 is used. Encryption, using the negotiated session key. Derivative key The encrypted parameters are denoted as Referring to the hash mechanism of verifiable distributed learning, the formula for generating the summary used for aggregation verification is improved as follows: ; In the formula, Indicates the first The local model parameters obtained from training each power device. Indicates the first The two-element hash digest corresponding to the local model parameters of each power device is output as a tuple; and For a large prime number Multiplication cyclic group, , These are the generators of the corresponding cyclic group, and the exponentiation operation modulo... Under the meaning of; An improved one-way hash function (input is model parameters) With power equipment (To prevent parameter tampering and identity fraud). This is a string concatenation operator; the digest is used to support subsequent aggregate digest consistency verification.

[0066] Power equipment uses DEK+KEK encryption technology to encrypt parameters. Perform encryption and encapsulation, generate encapsulation information, and combine the encapsulation information with the verification digest. The information is packaged and uploaded to the cloud server. The upload process is carried out through a TLS channel to ensure that the parameters are not stolen or tampered with during transmission.

[0067] In another exemplary embodiment, the cloud server includes a cloud processing unit; the cloud processing unit is configured to: Verify the signatures and local model summaries of each power device and determine the set of online power devices for the current round.

[0068] The encapsulation information of each online power device is processed using KEK decapsulation technology and DEK decryption technology to obtain the corresponding encrypted model parameters.

[0069] A double-mask protocol is used to decrypt each encrypted model parameter to determine the corresponding local model parameters.

[0070] The model parameters of each local model are securely aggregated using a double mask to determine the global model parameters, and zero-knowledge proofs and timestamps corresponding to the global model parameters are generated; the zero-knowledge proofs include aggregate digest consistency proofs and noise mask validity proofs.

[0071] As an example, the cloud server dynamically manages the access / exit of power equipment. Online power equipment refers to power equipment that has completed device authentication within the current aggregation round, uploaded the model parameter package within the preset upload time limit, and whose signature verification and parameter package integrity verification have both passed; the definition of the first Wheel online power equipment collection ,in, This indicates a collection of registered power equipment. The power equipment in the data refers to the power equipment that has completed model parameter upload and passed signature verification, model digest verification, and parameter package integrity verification within the current round; the access time, offline time, and parameter upload status of each power equipment are recorded; only when... Aggregation should be initiated immediately; otherwise, the waiting time should be extended, but not exceeding 5 minutes. The connection time of each power device should be recorded. Offline time With parameter upload status; only when the current number of parameters is uploaded. Wheel online power equipment set meets Aggregation is initiated at a specific time; among which, This represents the minimum threshold number of online power devices required to initiate the current round of secure aggregation. If If so, the waiting time will be extended, but the waiting time will not exceed 5 minutes. A unique allocation will be made for newly added power equipment. ("District Number + Device Type + Serial Number" 16-byte encoding) to bind hardware identifiers to prevent counterfeiting.

[0072] As an example, the cloud server employs a double-masking protocol for secure aggregation, ensuring that the global model can still be correctly recovered even if some nodes are offline or subjected to malicious attacks. The server generates zero-knowledge proofs corresponding to the parameters of the global model, allowing each terminal to verify the correctness and integrity of the aggregation process, preventing server tampering or deception. The specific process includes: S31, the cloud server calls the pre-stored public key of the power equipment. Verify the signature of power equipment Validity ( The upload timestamp is a parameter (valid duration ≤ 5 minutes). The formula is: ; .

[0073] Only when the verification result is If the power equipment is deemed legitimate, the parameter packet is discarded and an illegal access log is recorded.

[0074] S32, the CRC32 checksum of the encrypted parameter packet calculated by the cloud server, compared with the checksum uploaded by the power equipment. The comparison formula is as follows: .

[0075] If they are not equal, it is determined that data transmission has been tampered with or lost, and a "parameter retransmission" command is sent to the power equipment. In the formula, This represents the digital signature generated by the nth power device; This indicates the use of the terminal private key of the nth power device. Perform an SM2 digital signature on the content within the parentheses; This represents the terminal private key corresponding to the nth power device. This represents the terminal public key corresponding to the nth power device; This represents the encrypted local model parameters uploaded by the nth power device; This represents the model summary corresponding to the local model parameters of the nth power device; This represents the timestamp of the parameter packet uploaded by the nth power device; This represents the string concatenation operator; This represents the signature verification function. Outputting True indicates that the signature verification passed, and outputting False indicates that the signature verification failed. Represents the cyclic redundancy check function; This represents the CRC32 checksum calculated by the nth power device for the uploaded parameter packet and uploaded along with the parameter packet; This indicates that the CRC32 checksum recalculated by the cloud server is consistent with... Perform a consistency comparison.

[0076] Next, the cloud server aggregates the power equipment nodes with legitimate identities and complete data to form the first... The set of online nodes in rounds of iteration ,like ( For distributed thresholds, the power equipment in the distribution area Distribution network power equipment If the power supply fails to upload, the waiting time will be extended by ≤5 minutes and the equipment that failed to upload power will be notified; if the power supply fails to upload power after the waiting period... If the current aggregation round is terminated, an "insufficient nodes" alert will be recorded; if Then proceed to the subsequent aggregation process.

[0077] S33, Encryption parameters uploaded by power equipment The actual result is an encrypted combination of "original parameters + double mask". After decryption (for cloud servers) Derivative Key Decryption )get: ; In the formula, For power equipment Local model parameters, The decrypted parameters with a mask. For the first Paired random masks for each power device For the first Noise masking for electrical equipment.

[0078] S34, Cloud Servers Introduce Accuracy Weighting for Power Equipment Fault Diagnosis , , For the first The fault diagnosis accuracy (i.e., the accuracy of predicted operating status results) of the local model for each power device is statistically derived from the training results. The aggregation formula for improving the parameter influence of high-confidence power devices is as follows: ; In the formula, Indicates the first A collection of online power devices participating in the aggregation. Represents a set any power device index in, Indicates the first Accuracy of local model fault diagnosis for individual power equipment Indicates the first Accuracy of local model fault diagnosis for individual power equipment Indicates the first The accuracy weight of each power device in the current round of aggregation For the first Round global model parameters; (because The shared random numbers between nodes cancel each other out symmetrically, and the weighted average is still 0). The noise mask is weighted and summed, and then the cloud server recovers the power equipment via the previously deployed Shamir secret sharing protocol. Already Split into (A secret share; the cloud server can reconstruct it by collecting ≥t shares).

[0079] S35, the final simplified aggregation formula is: ; In the formula, This represents the weighted sum of noise masks reconstructed by the cloud server based on the Shamir secret sharing protocol; due to Paired random masks cancel each other out in the weighted aggregation, and the cloud server only needs to subtract the weighted sum of the reconstructed noise masks. You can get the first Round global model parameters .

[0080] S36, Zero-Knowledge Proof Construction: In the process of constructing a zero-knowledge proof, weights... When used for hash exponentiation, quantization encoding with a preset precision is performed before participating in the exponentiation calculation to ensure that the exponentiation operation is performed within the domain of the cyclic group. The Proof includes proof of aggregate digest consistency and proof of noise mask validity; the core verification relationship is based on the multiplicative homomorphism design of Homomorphic Hash. ; To simplify power equipment verification, the cloud server packages the Proof as follows: ; In the formula, Sign the cloud server with SM2 (to prevent Proof from being tampered with). Next, the cloud server will... The data is transmitted to each power device via a secure channel, while a lightweight verification interface is provided. The power device only needs to execute the verification; if the verification passes, the power device accepts the data. Used for subsequent diagnosis; if it fails, the power equipment sends an aggregated anomaly alarm to the cloud server and refuses to load the global model.

[0081] In another exemplary embodiment, the cloud processing unit is further configured to: The improved KARN model was trained using a cloud-based historical fault dataset of power equipment to obtain global model initialization parameters.

[0082] The improved KARN model includes a recursive network layer and a learnable activation function layer; the learnable activation function layer adopts an improved time basis function, which is used to adaptively fit the nonlinear fluctuation characteristics of power equipment; the recursive network layer adopts an improved memory mechanism, which is used to extract long-term time dependencies to filter out electromagnetic interference noise.

[0083] The expression for the improved time basis function is: ; ; The expression for the improved memory mechanism is: ; In the formula, Indicates the first Each electrical device The load-adaptive time basis function output at time t. express t Time of the first n Load normalization adjustment factor for each power device and They represent the first n Each electrical device t The load power and rated power at any given time; where, By the n The calculations for each power device are based on locally collected operating data such as voltage and current. The determination is based on the power equipment registration information or the system's initial configuration information. It should be noted that during the online distributed learning phase, the power equipment does not continuously upload real-time raw operating data to the cloud server. The cloud server primarily receives encrypted model parameters, local model summaries, predicted operating status, signatures, and timestamps. During the offline model initialization phase, the cloud processing unit retrieves historical training samples from the cloud-based historical fault database. These historical training samples include historical feature datasets, load power, rated load power, electromagnetic interference intensity, and measured fault labels, used for initializing and training the improved KARN model. This represents the Sigmoid linear unit activation function (boosting gradient flow). , Let represent the weight matrix and bias vector of the time basis function, respectively. express t Time of the first n A feature vector / feature dataset for a power device, consisting of time-domain, frequency-domain, and related features. and They represent t Time and t -1 moment n The hidden status of an electrical device To initialize as a zero vector, and These represent the hidden state weight matrix and the feature weight matrix, respectively. This indicates the bias term.

[0084] The improved time basis function dynamically adjusts the basis function's ability to capture temporal features. Used for high load ( The improved time basis function is made more sensitive to current and power characteristics. The improved memory mechanism integrates fault association features from the previous time step, optimizes hidden state updates, and ensures that the model remembers short-term fault trends.

[0085] As an example, during the model initialization phase, power equipment extracts and de-identifies features from data within a preset time window before and after a fault, based on locally stored historical fault operation data, forming historical feature samples and corresponding measured fault labels. The historical feature samples include time-domain features, frequency-domain features, correlation features, and load normalization adjustment factors. The measured fault labels are determined by equipment self-inspection results, on-site maintenance results, or historical operation and maintenance feedback results. The power operation and maintenance platform aggregates the historical feature samples and measured fault labels reported by each power equipment and stores them in a cloud-based historical fault database, forming a cloud-based historical fault dataset for power equipment. During the offline initialization phase, the cloud processing unit retrieves the cloud-based historical fault dataset for power equipment from the cloud-based historical fault database. The cloud-based historical fault dataset for power equipment includes historical feature samples extracted from and de-identified based on local historical operation data of the power equipment, and the measured fault labels corresponding to the historical feature samples.

[0086] As an example, the global model initialization is based on a cloud-based historical fault dataset of power equipment (including ≥100,000 samples of 485 disconnections, loose connections, circuit interference, etc.) to obtain the global model initialization parameters. A fault weight adaptation mechanism is introduced, with the following formula: ; In the formula, This indicates the total number of fault types, in this example. This includes 485 disconnection, loose connection, and circuit interference; Let be the fault weight coefficient, representing the th fault. The importance weights or prior weights of the fault classes satisfy the following conditions: ,and ; For the first The sample size of a fault class represents the number of historical faults of power equipment in the cloud-based dataset. Number of fault samples; For the first The initial parameters of the local model obtained from training or initializing fault-like samples, and Dimensions are consistent. The data is sent to each power device via a TLS 1.3 channel, with an attached SM2 digital signature. Each power device then verifies the data and loads it.

[0087] In another exemplary embodiment, the cloud server further includes a Trust Center (TA); the two-factor authentication and key negotiation mechanism includes: The power operation and maintenance platform is used to assign a unique ID to each of the power devices.

[0088] The Trust Center (TA) is used for: Determine the basic parameters; the basic parameters include large prime numbers ( (bit), multiplicative cyclic group and its corresponding generator.

[0089] A signature key pair is generated based on the aforementioned basic parameters and then sent to the cloud processing unit.

[0090] The system verifies the identity request submitted by the newly added power device, and generates a terminal key pair upon successful verification. The private key from the terminal key pair is then distributed to the corresponding power device (e.g., via USB-Key distribution). ).

[0091] Generate a shared session key and configure Shamir secret sharing parameters.

[0092] The cloud processing unit is also used for: The device authentication information sent by the newly added power equipment is verified based on the pre-stored power equipment registration information; the device authentication information includes the corresponding unique ID and the private key in the terminal key pair.

[0093] Once the device authentication information is verified, an encrypted random token is sent to the corresponding power device. The power device decrypts the encrypted random token based on the shared session key to obtain a random token, and generates a response verification value based on the random token and the terminal key. The response verification value is then returned to the cloud processing unit. The cloud processing unit performs a consistency check on the response verification value based on the random token it sent and the pre-stored power device registration information. If the check passes, the corresponding power device is allowed to access the network.

[0094] As an example, the device authentication information includes a corresponding unique ID, an authentication timestamp, and an identity signature generated using the terminal's private key on the unique ID and authentication timestamp.

[0095] Once the device authentication information is verified, the cloud processing unit generates a random token and a token validity period. Based on the shared session key negotiated with the corresponding power device, the random token is encrypted and sent to the corresponding power device. The power device decrypts the encrypted random token and generates a response verification value based on the decrypted random token, the corresponding unique ID, the shared session key, and the token validity period. The response verification value is then returned to the cloud processing unit.

[0096] The cloud processing unit regenerates the target response verification value based on the random token it sends, the pre-stored unique ID of the power device, the shared session key, and the token's validity period. It then compares the target response verification value with the response verification value returned by the power device. If they match and the token has not exceeded its validity period, the verification is considered successful, and the corresponding power device is allowed to access. If they do not match or the token has exceeded its validity period, the verification is considered unsuccessful. If the verification is successful, the corresponding power device is allowed to access. If the verification fails and the number of failures is less than the failure threshold, the corresponding device authentication information is re-verified. If the verification fails and the number of failures equals the failure threshold, the corresponding power device is locked.

[0097] As an example, the expressions for the signing key pair, terminal key pair, and shared session key are as follows: ; ; ; In the formula, represents a large prime number, Modulus Generators in a multiplicative group and These represent the private and public keys of the cloud server, respectively. and They represent the first The terminal private key and terminal public key of each power device. Indicates the area code, Indicates the negotiation timestamp (valid ≤ 30min). This indicates bit string concatenation. This represents the hash function SHA-256. This indicates a shared session key. Specifically, it will... and The shared session key derivation process is incorporated to implement anti-replay verification during the binding of power equipment to its respective transformer substation and the key negotiation process. It should be noted that the aforementioned public / private key generation, key negotiation, and hash derivation processes can be implemented using existing cryptographic algorithms. This embodiment uses a basic cryptographic algorithm for power equipment access authentication, transformer substation binding, timestamp-based anti-replay, and session key generation during subsequent encrypted parameter transmission.

[0098] During the key initialization phase, the Trust Center (TA) configures Shamir secret sharing parameters for each terminal, including the secret share threshold t, the number of participating nodes n, and the corresponding share distribution rules. After generating a noise mask, the power equipment splits it into n secret shares and distributes or submits them according to the threshold rules. This enables the cloud server to recover the noise mask when it collects no less than t valid shares, which is used to handle secure aggregation when some power equipment is offline.

[0099] When power equipment is first connected, it sequentially completes device authentication and dynamic token verification. During the device authentication phase, the power equipment sends authentication information to the cloud server: The cloud server uses pre-stored power equipment registration information to... The consistency of the tokens is verified. After successful verification, the dynamic token verification phase begins. The cloud server sends an encrypted random token to the power equipment. After the power equipment is decrypted, a response message is generated: cloud servers The system performs a verification process. If the verification passes, the power equipment is allowed to connect. If the number of authentication failures reaches 3, the power equipment is locked for at least 1 hour.

[0100] After the power equipment is certified, the power equipment and the cloud server complete communication encryption according to the sequence of "secure channel establishment - parameter encryption and encapsulation - parameter upload - cloud decryption and verification". First, the power equipment and the cloud server use a shared session key. A TLS 1.3 secure channel is established, with the transport layer supporting only strong encryption algorithms such as AES-256-GCM. Subsequently, before uploading the encrypted model parameters in each round, the power equipment randomly generates a data encryption key DEK and uses DEK to encrypt the local model parameters; simultaneously, the power equipment... The derived key is encrypted with the key KEK, and the DEK is then encrypted and encapsulated using the KEK (resulting in encapsulated information). The power equipment ultimately uploads the encapsulated information, its signature, predicted operating status, timestamp, and local model digest to the cloud server via a TLS 1.3 channel. Upon receiving the data, the cloud server first verifies the power equipment's signature and timestamp, then uses the KEK to decapsulate the DEK, and finally decrypts the model parameters to be aggregated (the encrypted model parameters) using the DEK, proceeding to the subsequent secure aggregation process.

[0101] In another exemplary embodiment, in adjusting the aggregated global model parameters and determining the runtime diagnostic model based on the adjusted global model parameters, the local server is used to: Based on the zero-knowledge proof, verify the aggregated global model parameters (i.e., verify whether the aggregated global model parameters are correctly generated by the local model parameters of the current round of online power equipment according to the double mask security aggregation rule).

[0102] If the aggregated global model parameters pass the verification, the global model parameters are adjusted in conjunction with the local adaptation factor.

[0103] Based on the adjusted global model parameters, the operational status diagnostic model is determined.

[0104] To avoid a decrease in diagnostic accuracy (accuracy in predicting operating status) of the global model due to differences in data distribution (e.g., area A is mainly residential power consumption, while area B is mainly industrial power consumption), the power equipment incorporates local historical fault datasets (including feature datasets). Using time-domain-frequency-relation features and a small number of labeled samples (≤50 groups), the global model is lightly fine-tuned, and a local adaptation factor is introduced. To balance global commonalities and local characteristics, the fine-tuning formula is as follows: ; Indicates the first n The adjusted global model parameters for each power device This represents the global model parameters after aggregation in the current round. Indicates the first n Local adaptation factors for each power device Ensure global model dominance (avoid overfitting to local noise). Indicates the first n Model parameters of the local model of a power device in the previous round. This represents the accuracy of the global model parameters after the previous round of aggregation (which can be calculated using 50 local samples). Indicates the number of the previous round n The accuracy of the operational status diagnostic model for individual power equipment. The accuracy weights represent the weights of the local model. The higher the accuracy of the local model, the greater the influence of its parameters on the adapted model (the model whose parameters are determined by the adjusted global model, i.e., the model applied to predict the running state).

[0105] As an example, power equipment utilizes the previous round of model adaptation to enter the online cyclic diagnostics (predicting the operating status of the power equipment) and model iterative update (determining the local model) stages. After completing the first... After a round of global model aggregation verification and local adaptation, the power equipment obtains the parameters of the previous round of adaptation model. And use this model as the first The currently available models for online diagnostics. (Continue to the next section) After each diagnostic cycle, the power equipment operates according to the current sampling frequency. Continuously collect real-time multi-run data, and use a sliding time window of 5 seconds as a diagnostic cycle. For the first... During each diagnostic cycle, the power equipment performs a feature extraction process on newly acquired operational data such as voltage, current, bit error rate, communication latency, and electromagnetic interference within the window, resulting in a feature dataset for online inference, denoted as: ; In the formula, For the first Within the first diagnostic cycle The first power equipment Real-time moving average of class-running data; This represents the real-time peak deviation. Clock frequency; Frequency amplitude value; For the first Class running data and the first Real-time Pearson correlation coefficients between class-based runtime data; This is a real-time multi-operational data anomaly index used to determine whether there are abnormal inputs to power equipment.

[0106] In the During the online cyclic diagnostic process, the power equipment first utilizes the adaptation model from the previous round. For real-time feature sets Perform online reasoning to complete the current round of anomaly detection, fault type classification, and fault root cause localization. Subsequently, the power equipment will... The real-time multi-run data that passes the validity check is written to the local incremental training cache and together with the local historical failure samples to form the local training dataset for this round. If the model update cycle is reached, or the current round of diagnostic results indicates an abnormal fault state, then the power equipment will be based on... Adaptation model for the previous round Perform local incremental training to obtain the first... Round local model parameters It performs differential privacy perturbation, parameter encryption encapsulation, and verifiable digest generation, and then... The encryption model parameters, verification digest, diagnostic results, timestamps, and power equipment signatures are uploaded to the cloud server.

[0107] The cloud server re-evaluates the set of online power devices for each round of aggregation. And only perform secure aggregation on the parameters of power equipment that are legally identified, have complete data, and meet the aggregation threshold, to obtain the first... Round global model parameters After receiving and verifying the global model parameters and Proof returned by the cloud server, the power equipment performs lightweight adaptation by combining local adaptation factors to obtain the [the desired result]. Wheel Adaptation Model and take it as the first The input model for online diagnosis is used in rounds. This forms a closed loop of "aggregation - model adjustment - data collection - diagnosis - training - uploading - re-aggregation".

[0108] In the During each diagnostic cycle, the power equipment uses the adaptation model from the previous cycle. For real-time feature sets Perform reasoning and output the probability that the current power equipment is in an abnormal state. This step is used to complete the normal / abnormal binary classification judgment. Mapped to anomaly probability via the Sigmoid function The formula is: In the formula, This represents the probability that the nth power device is in an abnormal state, and its value ranges from ([0,1]). Represents the Sigmoid function; This represents the feature representation of the nth power device output by the hidden layer during the real-time diagnostic phase. and These represent the weight parameters and bias parameters of the anomaly detection output layer, respectively. If If the condition is met, the nth power device is determined to be in an abnormal state and enters the fault type classification; if If the condition is met, the nth electrical device is determined to be in normal condition, and the diagnosis ends. To preset the anomaly detection threshold, in this example, we take... .

[0109] In another exemplary embodiment, the operating status includes anomaly probability and fault type probability; the power equipment is also used to determine the fault type and fault cause based on the fault type probability and Pearson correlation coefficient, and output corresponding operation and maintenance suggestions.

[0110] The operational status diagnostic model includes an output layer, which comprises an anomaly detection output layer and a fault type classification output layer. The anomaly detection output layer outputs the hidden layer data using a Sigmoid function. Mapped to anomaly probability When the probability of an anomaly is greater than or equal to the preset anomaly judgment threshold, the probability of each fault type is output through the Softmax function; the fault types include line disconnection, line loose connection and circuit interference.

[0111] The expression for the probability of the fault type is: ; Indicates the fault type index. These correspond to line disconnection, line loose connection, and circuit interference, respectively. The formula for determining the main cause of the fault is: ; In the formula, Indicates the first n The first electrical device corresponding to the first k Types of faults The probability, and Indicates the fault type Classification parameters, and Indicates the fault type Classification parameters, This represents the output information of the hidden layer. Representing the feature dataset and fault type The correlation, Represents the covariance function. Represents the variance function. Indicating the diagnostic stage n A feature dataset of power equipment Indicates a fault The tag. Take The category corresponding to the maximum value is used as the running status result, such as This indicates a 485 dummy connection.

[0112] As an example, power equipment outputs targeted operation and maintenance suggestions based on the fault type and its root cause, adapting to power system operation and maintenance scenarios: If the diagnosis is 485 disconnection ( The main reason is a line interruption. It is recommended to replace the wires with RVVSP 2×0.75mm wires conforming to DL / T 645-2007 standard. 2 For shielded cables, prioritize checking the joints between underground pipelines and concealed wiring in walls; If the diagnosis is a false connection (485) The main cause is abnormal contact resistance. It is recommended to "clean the oxide layer of the terminals with anhydrous ethanol, tighten the screws with a torque of 0.8-1.2 N·m, and apply anti-oxidation and anti-rust grease". If diagnosed as circuit interference ( The main cause is electromagnetic radiation. It is recommended to "install an LC low-pass filter (cutoff frequency 12MHz) next to the interference source (such as the frequency converter) and adjust the distance between the RS485 line and the power cable to ≥0.3m".

[0113] Diagnostic results and recommendations can be uploaded to the cloud server through a pre-established secure channel. After receiving the diagnostic results, the cloud server associates and stores them with the power equipment ID, timestamp, cause of failure, and maintenance recommendations, and synchronizes them to the maintenance platform as input for subsequent dynamic sampling, model update cycle adjustment, and maintenance strategy optimization.

[0114] Power equipment completed the first After each round of fault diagnosis, the diagnosis results, fault category, dominant fault operation data, operation and maintenance suggestions, real-time sampling frequency, communication status, and data validity flag are written to the local diagnostic cache. For real-time multi-operation data that passes the validity check, the power equipment uses it as the first... Incremental training samples are used in rounds; data that fails the validity check but triggers anomaly alarms are only used for fault alarms and log recording, and are not used as the basis for updating model parameters, so as to avoid abnormal sampling values ​​from polluting the local training process.

[0115] When the model update cycle is met, or when abnormal faults such as 485 disconnection, 485 loose connection, or circuit interference are detected, the power equipment is based on the first... The newly added samples and local historical fault samples are used to incrementally train the current adapted model, and the first batch of samples is generated. The local model parameters, parameter hash values, verification digests, and power equipment signatures are uploaded to the cloud server to participate in a new round of security aggregation. The cloud server completes the first... After the first round of aggregation and returning the global model, the power equipment verifies the aggregation results and performs local adaptation to generate the model for the second round. The diagnostic model is adapted for each round of diagnostics. Therefore, the diagnostic results are not only used for current fault handling, but also serve as feedback input for subsequent model updates, sampling frequency adjustments, model update cycle adjustments, and transmission strategy optimizations, ensuring a continuous closed-loop iterative process of "diagnosis—training—aggregation—model update" during the online diagnostic phase.

[0116] In another exemplary embodiment, the operation and maintenance parameters include the sampling frequency and model update cycle of each power device; in generating the operation and maintenance parameters of each power device based on the measured operating status and the operation and maintenance status information of each power device, the power operation and maintenance platform is used for: The measured operating status and maintenance status information sent by each power device are obtained for each round; the maintenance status information includes load power, electromagnetic interference intensity and predicted operating status.

[0117] As an example, the load power is calculated by each power device based on its own collected operating data. The measured operating status is determined by on-site maintenance results, equipment self-inspection results, or operation and maintenance feedback results, and can be uploaded to the power operation and maintenance platform through a human-machine interface.

[0118] Based on the load power, electromagnetic interference intensity, and predicted operating status of each power device in the current round, adjust the sampling frequency of each power device in the next round.

[0119] Based on the electromagnetic interference intensity, load power, predicted operating status and measured operating status of each power device in the current cycle, calculate the load fluctuation, electromagnetic interference fluctuation and status diagnosis accuracy of each power device in the current cycle, and adjust the model update cycle of each power device in the next cycle.

[0120] The formula for adjusting the sampling frequency is as follows: ; In the formula, Indicates the first The sampling frequency of each power device was adjusted. Indicates the reference sampling frequency; This represents the load power of the nth electrical device; This represents the rated load power of the nth electrical device; This represents the electromagnetic interference intensity of the nth power device; Indicates the maximum permissible electromagnetic interference value; Indicates the fault level weight. Indicates the first Each power device corresponds to a fault type Level coefficient, fault type By the The predicted operating status of each power device is determined. Electromagnetic interference weight.

[0121] The formula for adjusting the model update cycle is: ; ; In the formula, This represents the model update cycle after the nth power device is adjusted. Indicates the baseline model update cycle. This represents the environmental fluctuation coefficient of the nth power device. This represents the load fluctuation of the nth power device within a preset time period. This represents the electromagnetic interference fluctuation of the nth power device within a preset time period. This represents the accuracy of the status diagnosis of the nth power device in the current round; It is a constant used to compensate for the denominator when the accuracy is 0.

[0122] As an example, the default reference sampling frequency is 500Hz, based on the time-series data requirements of the KARN model. Highest priority 485 disconnected 485 virtual connection Circuit interference ,normal , , , .

[0123] In another exemplary embodiment, after obtaining the online diagnostic results, the system abstracts the current power operation state and communication state into a state space of a Markov Decision Process (MDP) for subsequent dynamic decisions on transmission power and data priority. ; In the formula, For load power ratings (low / medium / high, corresponding to 0.3 / 0.7 / 1.0). Electromagnetic interference level (low / medium / high, corresponding to 0.2 / 0.6 / 1.0). The fault level (as described above). The real-time bit error rate (normalized to [0,1]). The communication delay is normalized to [0,1] with a maximum value of 50ms.

[0124] Action space Actionable steps focusing on the power sector: ; In the formula, For transmission power adjustment ( , (Rated transmission power of electrical equipment, such as 1W). Assign priority to data (priority of critical alarm data = 1, priority of ordinary monitoring data = 0.5).

[0125] reward function Taking into account safety, efficiency, and energy consumption, a power scenario weight (safety first) is introduced, and the formula is as follows: ; In the formula, Security weight (the lower the error rate, the higher the reward); Efficiency weight (lower latency, higher reward); Energy consumption weight (the higher the power, the higher the penalty); Safety metrics (normalized, range [0,1]); Efficiency index (normalized, range [0,1]); Energy consumption index (normalized, range [0.5, 1.2]).

[0126] The optimal strategy is solved using a value iteration algorithm. The iterative formula is: ; In the formula, state The optimal value function; Discount factor, ( Execute action After state Transferred to The probability. The convergence condition for iteration is the difference between the value functions of two adjacent rounds. After convergence, the optimal strategy is output. For example, when the fault level is high, the power is increased to reduce the bit error rate; when the fault is normal, the power is reduced to reduce energy consumption.

[0127] The protection Markov decision optimization method includes a state space containing five types of parameters such as load level and interference level, and an action space covering transmission power (0.5-1.2 times rated) and data priority (alarm 1, monitoring 0.5). The reward function balances safety (weight 0.5), efficiency (0.3), and energy consumption (0.2). The optimal strategy is solved using a value iteration algorithm (convergence difference < 10). -4 It also protects the local adaptation of the global model by using an adaptation factor of 0.6-0.8, combined with a small number of local samples (≤50 groups) and accuracy weights to generate adaptation parameters, thus solving the problem of data distribution differences.

[0128] In another exemplary embodiment, the power operation and maintenance platform also records operation and maintenance data for each round (operation and maintenance feedback log): including power equipment ID, sampling frequency, model update cycle, transmission power, diagnostic results, and operation and maintenance effect; the index uses SHA-256. The operation and maintenance effect is analyzed monthly. When the fault repair rate is lower than 90% or the communication energy consumption exceeds the rated value by 110%, the power operation and maintenance platform adjusts the corresponding adjustment parameters of sampling frequency, model update cycle and transmission power to achieve continuous optimization of operation and maintenance strategy.

[0129] In another exemplary embodiment, the cloud server also records audit logs of the distributed learning process. The audit logs are maintained throughout the entire process, including model initialization, parameter uploading, security aggregation, proof verification, and model deployment, and are updated once after each round of security aggregation. For the... The distributed learning process is repeated several times, and the log content includes the terminal ID and the training round. Online terminal collection Model parameter hash, encryption parameters, etc., the encryption parameters include verification results, aggregate verification results, Proof verification results and timestamps.

[0130] Audit logs are primarily used to record key data during model initialization, parameter uploading, security aggregation, and verification processes, facilitating subsequent identification of abnormal parameter sources, tracing of the aggregation and verification process, and security auditing.

[0131] The audit logs are stored locally using RAID5 and preserved via a power consortium blockchain. The log indexing formula is as follows: ; In the formula, Use the SHA-256 hash function; For the first A unique identifier for each terminal; This is the current training round; For the first Wheel model parameter hashing; For aggregated verification results or Proof verification results, including "Pass" or "Fail"; This is the timestamp for the current round. This index is used for fault attribution, parameter tracking, and security auditing.

[0132] This application's embodiments integrate an improved KARN model, multi-feature extraction, a four-layer distributed architecture for collaborative interaction, dual-mask security aggregation, and Markov decision-making for dynamic operation and maintenance, forming a collaborative processing system for the entire power multi-parameter process. Simultaneously, through bidirectional linkage between local terminal training and cloud-based security aggregation, real-time status awareness, and adaptive policy adjustment, the system can optimize model training parameters, data transmission strategies, and security protection mechanisms in real time based on grid load fluctuations, electromagnetic interference intensity, and fault levels. Ultimately, this achieves precise power multi-parameter diagnosis, secure data transmission, and intelligent operation and maintenance decision-making, providing technical support for the stable operation of distribution network areas, distributed energy access points, and other scenarios.

[0133] In another exemplary embodiment of this application, a power distribution network operation and maintenance environment, including a 485 communication terminal, power data acquisition equipment, edge computing nodes, and a cloud-based model aggregation server, is selected as the test object in a verification scenario. A comparative test environment is constructed to compare the centralized diagnostic scheme with the embodiment of this application. The comparative scheme uses a centralized method of uploading multi-source raw operating data and uniformly training the diagnostic model in the cloud. This embodiment uses a distributed learning architecture, where the terminal or edge node trains the model based on local data, only uploading encrypted model parameters, and the cloud completes global model aggregation and distribution. During the test, three typical power fault scenarios—485 disconnection, 485 loose connection, and circuit interference—are set up. Multi-source operating data such as voltage, current, communication status, terminal response, and interference intensity are collected, and various fault samples are manually labeled as the basis for comparing the model diagnostic results.

[0134] The specific testing method is as follows: Training and testing sample sets are constructed for each type of fault. The fault categories output by the model are compared with the pre-labeled real fault categories, and the diagnostic accuracy is calculated based on the ratio of correctly identified samples to the total number of test samples. Simultaneously, the time required for the terminal to complete one round of model training, and the total time spent on cloud-based model parameter aggregation, updating, and distribution, are recorded and compared with the training time of a comparative model. Test results show that the collaborative design of a distributed learning architecture, multi-layered security protection, and dynamic operation and maintenance decision-making achieves a significant breakthrough in addressing the pain points of existing power multi-parameter (corresponding to various operational data) diagnosis and transmission solutions. In terms of diagnostic performance, the improved KARN model, combined with a local adaptation and fine-tuning mechanism, achieves a diagnostic accuracy of over 92% for three typical power faults: 485 disconnection, 485 loose connection, and circuit interference. This represents a significant improvement compared to the less than 70% accuracy of existing centralized models. Meanwhile, the time taken for a single round of local training on the terminal is controlled within 10 seconds, and the total latency for global model aggregation and distribution does not exceed 30 seconds. Compared to the comparative case where a single training session takes more than 30 minutes, the latency is reduced by more than 95%, fully meeting the actual needs for second-level response to power faults.

[0135] In terms of data security, a three-tier protection system of "differential privacy + double-layer encryption + HomomorphicHash verification" has been constructed. During the online distributed learning process, the power equipment does not continuously upload complete real-time raw operating data such as voltage, current, and user load curves to the cloud server. The power operation and maintenance platform only receives necessary operation and maintenance status information such as load power, electromagnetic interference intensity, predicted operating status, and measured operating status calculated locally by the power equipment, which is used for dynamic adjustment of sampling frequency and model update cycle. The cloud server primarily receives encrypted model parameters, local model summaries, signatures, and timestamps for secure aggregation and integrity verification, thereby reducing the risk of sensitive information leakage caused by directly uploading complete original operating data. Through a dual-mask protocol and the Shamir secret sharing protocol, even if 30% of the terminals are offline or under malicious attack, the global model can still recover normally, improving the ability to resist malicious nodes. Terminal access adopts a two-factor mechanism of "hardware-bound unique ID + device authentication + dynamic token verification," and is automatically locked after more than 3 authentication failures, reducing the risk of counterfeit terminal access and fake data. The entire process log uses RAID5 local storage and power consortium blockchain for evidence storage, meeting the data compliance audit requirements of the power industry and reducing the risk of sensitive information leakage caused by direct transmission of original operating data.

[0136] In the verification scenario, the comparative example uses direct uploading of raw data, a fixed sampling frequency, a fixed transmission power, and manual inspection and handling. This application's embodiment uses an encrypted model parameter uploading method, uploading approximately 10KB of model parameters per round. The sampling frequency is dynamically adjusted within the 100Hz to 1000Hz range based on load level, fault level, and electromagnetic interference status. Simultaneously, transmission power is optimized through a Markov decision process, and communication resources are allocated according to service priority. During testing, typical operation and maintenance scenarios such as loose wiring terminals, communication anomalies, device offline, operating parameter exceeding limits, and enhanced electromagnetic interference are set up. Indicators such as bandwidth usage, overall terminal energy consumption, key alarm data transmission latency, resource scheduling response capability, fault discovery to repair time, number of manual interventions, number of terminals managed by operation and maintenance personnel, and average device failure rate are statistically analyzed. Verification results show that, in terms of transmission efficiency, since power equipment mainly uploads lightweight information such as encrypted local model parameters, model summaries, predicted operating status, signatures, and timestamps, the amount of data uploaded per round is approximately 10KB, rather than uploading large amounts of raw operating data such as voltage, current, and load curves. Therefore, the amount of transmitted data and bandwidth usage can be reduced, effectively avoiding transmission congestion during peak electricity consumption periods. Instead of uploading large amounts of raw data, bandwidth usage is reduced by 60% compared to the comparison, effectively avoiding transmission congestion during peak electricity consumption periods. By dynamically adjusting the sampling frequency (100-1000Hz range) based on load, fault level, and electromagnetic interference, and optimizing transmission power through Markov Decision Process (MDP), the overall energy consumption of the terminal is reduced by 35%, extending the equipment's battery life. Dedicated communication resources are allocated to services of different priorities, and the latency of key alarm data transmission is controlled within 50 milliseconds, improving resource scheduling flexibility by 50%.

[0137] In terms of operation and maintenance costs, the diagnostic results directly include scenario-based operation and maintenance suggestions. For example, when there is a 485 faulty connection, the operation guidelines clearly state that "clean the terminals with anhydrous ethanol and tighten the screws with a torque of 0.8-1.2 N·m". The average fault repair time has been shortened from 2-3 hours in the existing solution to less than 45 minutes. The automated closed-loop operation and maintenance system reduces manual intervention. The number of terminals managed by the average distribution network maintenance personnel has increased from 50 to 80, reducing labor costs by 40%. The average annual equipment failure rate has also decreased from 8% to 3.5%, significantly reducing the operation and maintenance pressure and economic losses of the power system.

[0138] In one exemplary embodiment, a method for diagnosing the operating status of power equipment based on distributed learning is provided, comprising: Power equipment collects its own operating data in real time, and predicts the operating status of the power equipment and generates operation and maintenance status information based on its own operating data and the current operating status diagnostic model. In addition, it trains the operating status diagnostic model to generate a local model, performs encryption and DEK+KEK encryption encapsulation processing on the model parameters of the local model in sequence, and generates a local model summary.

[0139] The cloud server receives the encapsulation information and local model digest uploaded by each power device through the TLS channel, performs secure aggregation and integrity verification of encrypted parameters, and generates aggregated global model parameters.

[0140] The power equipment adjusts the aggregated global model parameters to update the operational status diagnostic model.

[0141] The power operation and maintenance platform generates operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distributes these parameters to the corresponding power devices.

[0142] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 2 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media.

[0143] This computer device serves as a power operation and maintenance platform, cloud server, or power equipment in the distributed learning-based power equipment operation status diagnosis system described in any of the above embodiments. The database of this computer device stores data from the corresponding power operation and maintenance platform, cloud server, or power equipment. The input / output interface of this computer device is used for information exchange between the processor and external devices. The communication interface of this computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it can implement the functions of the corresponding power operation and maintenance platform, cloud server, or power equipment.

[0144] Those skilled in the art will understand that Figure 2 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0145] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0146] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0147] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations and be authorized by the owner of the corresponding device.

[0148] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0149] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0150] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0151] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A power equipment operation status diagnosis system based on distributed learning, characterized in that, The system includes a power operation and maintenance platform, a cloud server, and various power equipment. The cloud server communicates with various power devices using a two-factor authentication and key negotiation mechanism. The power equipment is used to collect its own operating data in real time, and predict the operating status of the power equipment and generate operation and maintenance status information based on its own operating data and the current operating status diagnostic model. In addition, the operating status diagnostic model is trained to generate a local model, and the model parameters of the local model are encrypted and DEK+KEK encrypted and encapsulated in sequence, and a local model summary is generated. The cloud server is used to receive encapsulation information and local model digests uploaded by various power devices through the TLS channel, and to perform encrypted parameter security aggregation and integrity verification to generate aggregated global model parameters; The power equipment is also used to adjust the aggregated global model parameters to update the operational status diagnostic model; The power operation and maintenance platform is used to generate operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distribute them to the corresponding power devices.

2. The power equipment operation status diagnosis system based on distributed learning according to claim 1, characterized in that, In terms of real-time acquisition of its own operational data and prediction of the operating status of power equipment based on its own operational data and a current operational status diagnostic model, the power equipment is used for: It collects its own operational data in real time to form an operational data stream; the operational data includes various types of operational data. The running data stream is sliced ​​using a sliding time window, and the validity of the data in each sliding time window is determined by multi-parameter cross-validation to extract valid data. Features are extracted sequentially from the valid data within each sliding time window to generate a corresponding feature dataset. The feature dataset includes time-domain features, frequency-domain features, and correlation features corresponding to each type of running data. The correlation features include Pearson correlation coefficient and anomaly index. Each feature dataset is sequentially input into the operational status diagnostic model to obtain the operational status of the power equipment.

3. The power equipment operation status diagnosis system based on distributed learning according to claim 2, characterized in that, The power equipment also includes a local server; The local server is used to update the local power equipment historical fault dataset based on the generated feature dataset and the actual operating status of the corresponding power equipment. When the model update conditions are met, the operation status diagnosis model is trained using the local power equipment historical fault dataset, a local model is generated, and the model parameters of the local model are encrypted and encapsulated using DEK+KEK encryption. In addition, a local model summary is generated based on the model parameters of the local model. In the training process of the operational status diagnostic model, differential privacy technology is used to add noise to the parameter gradients of each network layer during backpropagation; The aggregated global model parameters are adjusted, and the runtime status diagnostic model is determined based on the adjusted global model parameters.

4. The power equipment operation status diagnosis system based on distributed learning according to claim 3, characterized in that, The cloud server includes a cloud processing unit; The cloud processing unit is used for: Verify the signatures and local model summaries of each power device and determine the set of online power devices for the current round; The encapsulation information of each online power device is processed using KEK decapsulation technology and DEK decryption technology to obtain the corresponding encrypted model parameters; A double-mask protocol is used to decrypt each encrypted model parameter to determine the corresponding local model parameters; The model parameters of each local model are securely aggregated using a double mask to determine the global model parameters, and zero-knowledge proofs and timestamps corresponding to the global model parameters are generated; the zero-knowledge proofs include aggregate digest consistency proofs and noise mask validity proofs.

5. The power equipment operation status diagnosis system based on distributed learning according to claim 4, characterized in that, The cloud processing unit is also used for: The improved KARN model was trained based on a cloud-based historical fault dataset of power equipment to obtain global model initialization parameters; The improved KARN model includes a recursive network layer and a learnable activation function layer; the learnable activation function layer adopts an improved time basis function, which is used to adaptively fit the nonlinear fluctuation characteristics of power equipment; The recursive network layer employs an improved memory mechanism to extract long-term dependencies in order to filter out electromagnetic interference noise.

6. The power equipment operation status diagnosis system based on distributed learning according to claim 4, characterized in that, The cloud server also includes a Trust Center (TA); The two-factor authentication and key negotiation mechanism includes: The power operation and maintenance platform is used to assign a unique ID to each of the power devices. The Trust Center (TA) is used for: Determine the basic parameters; A signature key pair is generated based on the aforementioned basic parameters and then sent to the cloud processing unit. The identity request submitted by the newly added power equipment is verified, and a terminal key pair is generated after the verification is successful. The private key in the terminal key pair is then sent to the corresponding power equipment. Generate a shared session key and configure Shamir secret sharing parameters; The cloud processing unit is also used for: Verify the device authentication information sent by the newly added power equipment based on the pre-stored power equipment registration information; Once the device authentication information is verified, an encrypted random token is sent to the corresponding power device, and the information returned by the power device is verified. If the verification is successful, the corresponding power device is allowed to access. If the verification fails and the number of failures is less than the failure threshold, the corresponding device authentication information is re-verified. If the verification fails and the number of failures equals the failure threshold, the corresponding power device is locked.

7. The power equipment operation status diagnosis system based on distributed learning according to claim 4, characterized in that, In adjusting the aggregated global model parameters and determining the runtime status diagnostic model based on the adjusted global model parameters, the local server is used for: Based on the zero-knowledge proof, the aggregated global model parameters are verified. If the aggregated global model parameters pass the verification, the global model parameters are adjusted in conjunction with the local adaptation factor; Based on the adjusted global model parameters, determine the operational status diagnostic model; The formula for adjusting the global model parameters is as follows: ； Indicates the first n The adjusted global model parameters for each power device This represents the global model parameters after aggregation in the current round. Indicates the first n Local adaptation factors for each power device Indicates the first n Model parameters of the local model of a power device in the previous round. This represents the accuracy of the global model parameters after the previous round of aggregation. Indicates the number of the previous round n The accuracy of the operational status diagnostic model for individual power equipment.

8. The power equipment operation status diagnosis system based on distributed learning according to claim 4, characterized in that, The operating status includes anomaly probability and fault type probability; the power equipment is also used to determine the fault type and fault cause based on the fault type probability and Pearson correlation coefficient, and output corresponding operation and maintenance suggestions; The operational status diagnostic model includes an output layer, which maps the data of the learnable activation function layer to anomaly probabilities using the Sigmoid function. When the anomaly probability is greater than a preset anomaly value, the output layer outputs the probabilities of each fault type using the Softmax function. The fault types include line disconnection, line loose connection, and circuit interference.

9. The power equipment operation status diagnosis system based on distributed learning according to claim 1, characterized in that, The operation and maintenance parameters include the sampling frequency and model update cycle of each power device; in generating the operation and maintenance parameters of each power device based on the measured operating status and the operation and maintenance status information of each power device, the power operation and maintenance platform is used for: The measured operating status and maintenance status information sent by each power device are obtained in each round; the maintenance status information includes load power, electromagnetic interference intensity and predicted operating status. Based on the load power, electromagnetic interference intensity, and predicted operating status of each power device in the current round, adjust the sampling frequency of each power device in the next round; Based on the electromagnetic interference intensity, load power, predicted operating status and measured operating status of each power device in the current cycle, calculate the load fluctuation, electromagnetic interference fluctuation and status diagnosis accuracy of each power device in the current cycle, and adjust the model update cycle of each power device in the next cycle.

10. A method for diagnosing the operating status of power equipment based on distributed learning, characterized in that, The method for diagnosing the operating status of power equipment includes: Power equipment collects its own operating data in real time, and predicts the operating status of the power equipment and generates operation and maintenance status information based on its own operating data and the current operating status diagnosis model. In addition, it trains the operating status diagnosis model to generate a local model, performs encryption and DEK+KEK encryption encapsulation processing on the model parameters of the local model in sequence, and generates a local model summary. The cloud server receives the encapsulation information and local model digest uploaded by each power device through the TLS channel, and performs encrypted parameter security aggregation and integrity verification to generate aggregated global model parameters; The power equipment adjusts the aggregated global model parameters to update the operational status diagnostic model; The power operation and maintenance platform generates operation and maintenance parameters for each power device based on the measured operating status and operation and maintenance status information of each power device, and then distributes these parameters to the corresponding power devices.

Images