Current sensor fault isolation and signal reconstruction method, device, equipment and medium

Abstract

The invention relates to the technical field of intelligent sensing and power equipment predictive maintenance, and discloses a current sensor fault isolation and signal reconstruction method, device and equipment and a medium. The method comprises the following steps: firstly, initializing a federated learning system, constructing a public calibration data set which does not contain private data of a client in a central server, and initializing a global diagnosis model; then, carrying out federal knowledge fusion training based on the public calibration data set and the global diagnosis model, and repeating to convergence through client local training, soft label uploading and server knowledge distillation updating of the model; then acquiring real-time measurement signals of multiple current sensors, inputting the real-time measurement signals into a convergence model to obtain a health state score, and identifying and isolating a fault sensor through a dynamic threshold strategy; and finally, calculating a consensus true value in combination with the health state score of the health sensor, and performing online correction and reconstruction on the output signal of the fault sensor. Data privacy protection and fault processing accuracy are considered, and system reliability is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent sensing and predictive maintenance technology for power equipment, and more specifically, to a method, apparatus, device, and medium for fault isolation and signal reconstruction of a current sensor. Background Technology

[0002] As a key sensing device in power systems, industrial IoT, and other fields, the reliability of current sensors in fault diagnosis and signal reconstruction directly determines the operational safety of the system. Existing technologies mainly rely on three approaches: methods based on hardware redundancy achieve fault identification through multi-sensor voting, but significantly increase costs and wiring complexity, and have poor applicability in space-constrained scenarios; methods based on analytical models rely on accurate mathematical models of the system, but modeling complex power electronic systems is difficult, and they are sensitive to interference, resulting in insufficient practical reliability; methods based on centralized data-driven approaches require the collection of raw data to train the model, which poses risks of privacy leaks and incurs high communication and storage costs due to massive data transmission.

[0003] While federated learning, which has emerged in recent years, can protect privacy, it is difficult to cope with the heterogeneity of data and fault modes at each node when applied directly, resulting in a decline in the generalization performance of the model. Furthermore, it lacks a design for the closed-loop task of "fault isolation-signal reconstruction", and cannot effectively integrate deep fault knowledge across nodes, making it difficult to achieve accurate fault perception and high-quality signal recovery under the premise of privacy protection. Summary of the Invention

[0004] The main purpose of this application is to provide a method for fault isolation and signal reconstruction of current sensors, which aims to solve the technical problems in current sensor fault diagnosis and signal reconstruction, such as high cost, poor reliability, and privacy risks in existing technologies, and insufficient generalization, lack of closed-loop design, and difficulty in balancing privacy and accurate fault handling in federated learning applications.

[0005] The first aspect of this application proposes a method for fault isolation and signal reconstruction of a current sensor, including:

[0006] The federated learning system is initialized by building a public calibration dataset on the central server that does not contain client-private data, and by initializing the global diagnostic model.

[0007] Federated knowledge fusion training is carried out based on public calibration dataset and global diagnostic model. The central server distributes global diagnostic model to client. Client uses local data to train ontology to update model and generates soft tags for uploading. The central server uses knowledge distillation to fuse soft tags to update global diagnostic model, and repeats until convergence.

[0008] The real-time measurement signals of multiple current sensors are acquired, and the real-time measurement signals are input into the converged global diagnostic model to obtain the health status score of each sensor. Based on the health status score, a dynamic threshold strategy is used to identify and isolate faulty sensors.

[0009] Based on the health sensors other than the isolated faulty sensors, a consensus true value is calculated by combining the health status scores of each health sensor. The consensus true value is then used to perform online correction and reconstruction of the output signal of the faulty sensor.

[0010] Furthermore, the steps of initializing the federated learning system, constructing a public calibration dataset on the central server that does not contain client-private data, and initializing the global diagnostic model include:

[0011] Collect standard current signals and historical operating data segments of sensors under healthy conditions, perform standardization processing on the standard current signals and data segments, and generate a common calibration dataset covering typical operating conditions of the system;

[0012] Based on the feature distribution of the public calibration dataset, configure the network architecture parameters of the global diagnostic model, clarify the hierarchical structure of the feature extraction layer and the fault classification layer, and divide the globally shared parameters and personalized parameters.

[0013] The parameters of the global diagnostic model are assigned values ​​using a preset initialization strategy to obtain an initial global diagnostic model.

[0014] Furthermore, the steps of conducting federated knowledge fusion training based on the public calibration dataset and the global diagnostic model, distributing the global diagnostic model to the client, the client training the ontology using local data to update the model and generating soft tags for uploading, and the central server updating the global diagnostic model by fusing the soft tags with knowledge distillation, and repeating until convergence, include:

[0015] The central server distributes the current global diagnostic model to each client. The client builds a hybrid model architecture containing globally shared parameters and personalized parameters, trains it using local private data and multi-objective optimization functions, and obtains a locally updated model.

[0016] The client uses the local update model to infer the public calibration dataset, generates soft labels containing fault knowledge, and uploads them to the central server;

[0017] The central server uses knowledge distillation technology to integrate all soft tags and calculates global model update parameters through weighted aggregation based on the aggregation weight of each client.

[0018] The central server adjusts the global diagnostic model based on the updated parameters, repeating the above steps until the performance of the global diagnostic model converges.

[0019] Furthermore, the objective optimization function introduces a feature decoupling loss, decomposing model features into globally shared features and client-specific features. These client-specific features are used only in the local refactorer and do not participate in federated aggregation.

[0020] The feature decoupling loss function is:

[0021]

[0022] in, Task loss weighting coefficient; To compare the learning loss weight coefficients; These are the orthogonal constraint loss weight coefficients; The feature decoupling loss function; The task loss function; To learn the loss function by comparison; The orthogonal constraint loss function;

[0023] The task loss function is:

[0024]

[0025] in, The cross-entropy loss function; Let i be the amount of local data on client i;

[0026] The contrastive learning loss function is:

[0027]

[0028] in, The set of positive sample pairs indexes; For the set of indexes of all samples; The feature representation of sample i; This is a temperature parameter used to control the smoothness of the distribution;

[0029] The orthogonal constraint loss function is:

[0030]

[0031] in, This is the global feature matrix; It is a local feature matrix; It is the Frobenius norm.

[0032] Further, the steps of acquiring real-time measurement signals from multiple current sensors, inputting these real-time measurement signals into a converged global diagnostic model to obtain a health status score for each sensor, and identifying and isolating faulty sensors based on the health status scores using a dynamic threshold strategy, include:

[0033] The real-time measurement signals of multiple current sensors are acquired, and the real-time measurement signals are input into the converged global diagnostic model. The health status score of each sensor is generated through feature extraction and a health scoring layer.

[0034] A dynamic isolation threshold is calculated based on the health status score, and an anomaly score is generated by checking the consistency of features between sensors.

[0035] Based on the health status score, dynamic isolation threshold, and anomaly score, faulty sensors are comprehensively identified and isolated.

[0036] Furthermore, the step of calculating a consensus true value based on health sensors other than the isolated faulty sensor, combined with the health status scores of each health sensor, and using the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor, includes:

[0037] Collect measurement signals from unisolated health sensors, combine the health status scores of each health sensor with the feature similarity between sensors, and assign reliability weights and consistency weights.

[0038] Based on the aforementioned reliability weight and consistency weight, a fusion weight is constructed, and the measurement signals from the health sensor are weighted and fused to generate a consensus true value.

[0039] An adaptive correction model is introduced, which combines the depth features of the health sensor with the health status score to calculate a personalized correction term. The output signal of the faulty sensor is then reconstructed after optimizing the consensus true value.

[0040] Furthermore, the reliability weight and consistency weight are obtained by the following formula:

[0041]

[0042] in, Reliability weights based on health scores; A collection of health sensors; Assess the health of sensor i;

[0043]

[0044] in, Consistency weight; Let be the similarity between sensor i and the average features of most sensors.

[0045] A second aspect of this application also proposes a current sensor fault isolation and signal reconstruction device, comprising:

[0046] The initialization module is used to initialize the federated learning system, build a public calibration dataset on the central server that does not contain client-private data, and initialize the global diagnostic model.

[0047] The federated training module is used to conduct federated knowledge fusion training based on a public calibration dataset and a global diagnostic model. The central server distributes the global diagnostic model to the client, the client trains the ontology using local data to update the model, and generates soft labels for uploading. The central server uses knowledge distillation to fuse the soft labels to update the global diagnostic model, and repeats the process until convergence.

[0048] The fault isolation module is used to acquire real-time measurement signals from multiple current sensors, input the real-time measurement signals into a converged global diagnostic model to obtain a health status score for each sensor, and identify and isolate faulty sensors based on the health status score using a dynamic threshold strategy.

[0049] The signal reconstruction module is used to calculate a consensus true value based on the health sensors other than the isolated faulty sensor and the health status scores of each health sensor, and to use the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

[0050] A third aspect of this application provides a current sensor fault isolation and signal reconstruction device, comprising:

[0051] processor;

[0052] Memory, used to store computer programs;

[0053] When the processor executes the computer program, it implements the current sensor fault isolation and signal reconstruction method described in any of the above embodiments.

[0054] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the current sensor fault isolation and signal reconstruction method described in any of the above embodiments.

[0055] The first aspect of this plan brings the following benefits:

[0056] This application lays the foundation by initially constructing a public dataset and global model without private data; then, through federated training, client-side local training, soft label uploading, and server distillation updates, it balances privacy protection and model generalization; subsequently, it uses a converged model to generate a health score and dynamically isolates faults using thresholds; finally, it combines the health score to calculate a consensus value to reconstruct the signal. This forms a closed loop of "training-diagnosis-reconstruction," solving the problems of high cost, privacy risks, and insufficient generalization of existing technologies, while balancing privacy and accurate fault handling. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a current sensor fault isolation and signal reconstruction method according to an embodiment of this application;

[0058] Figure 2 This is a schematic diagram of the structure of a current sensor fault isolation and signal reconstruction device according to an embodiment of this application;

[0059] Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application;

[0060] Figure 4 This is a system architecture diagram of a current sensor fault isolation and signal reconstruction according to an embodiment of this application;

[0061] Figure 5 This is a schematic diagram illustrating the workflow of a current sensor fault isolation and signal reconstruction method according to an embodiment of this application.

[0062] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0065] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0066] Reference Figure 1 , Figure 4 , Figure 5 This application provides a method for fault isolation and signal reconstruction of a current sensor, including:

[0067] S1: Initialize the federated learning system, build a public calibration dataset on the central server that does not contain client private data, and initialize the global diagnostic model;

[0068] S2: Federated knowledge fusion training is carried out based on the public calibration dataset and the global diagnostic model. The central server distributes the global diagnostic model to the client. The client uses local data to train the ontology to update the model and generates soft labels for uploading. The central server uses knowledge distillation to fuse the soft labels to update the global diagnostic model and repeats until convergence.

[0069] S3: Acquire the real-time measurement signals of multiple current sensors, input the real-time measurement signals into the converged global diagnostic model to obtain the health status score of each sensor, and identify and isolate faulty sensors based on the health status score using a dynamic threshold strategy.

[0070] S4: Based on the health sensors other than the isolated faulty sensors, calculate the consensus true value by combining the health status scores of each health sensor, and use the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

[0071] like Figure 4 This paper showcases a multi-current sensor system architecture based on federated knowledge fusion, comprising a cloud-based federated fusion center and multiple distributed client sensor nodes. The core cloud component is a globally shared encoder, responsible for fusing the shared knowledge uploaded by each node and distributing the updated global model. Each client node contains a local private encoder and a personalized decoder, respectively handling sensor-specific features and completing localized diagnostic tasks. Solid arrows in the diagram represent the global model distribution process, while dashed arrows show the upload paths of shared knowledge from each node. The overall architecture achieves a balance between knowledge fusion and privacy protection by separating shared and private components.

[0072] like Figure 5The system consists of a cloud server and multiple client sensor nodes (Client 1 to Client N). The overall process covers two parts: federated training and online diagnosis and calibration. In the federated training phase, the cloud server first performs global model initialization and distributes the initial model to each client. Each client trains its local model based on its private data and generates soft labels based on a public calibration database, then uploads the soft labels to the cloud. The cloud uses knowledge distillation and fusion methods to integrate the knowledge from each client, updating the global model and forming a continuously optimized federated learning loop. In the online diagnosis and calibration phase, the trained global model is deployed to each sensor node, enabling parallel global model diagnosis of simultaneously measured signals from multiple sensors, thereby completing faulty sensor isolation and signal correction and reconstruction.

[0073] In step S1, the federated learning system is initialized. The core of this step is to build a public calibration dataset on the central server, eliminating client-side private data, and to initialize the global diagnostic model, providing a unified benchmark and initial model framework for subsequent federated knowledge fusion training. Taking the current sensor monitoring scenario of a new energy power plant cluster as an example, the public calibration dataset is first constructed by collecting standard current signals during normal power plant operation (such as a 50Hz pure sine wave or a composite waveform containing 3rd / 5th harmonics), as well as historical operating data fragments from 10 healthy current sensors over the past 6 months (only common features such as current amplitude and frequency are retained, while private information such as power plant production data and sensor numbers are removed). These data are then standardized (e.g., subtracting the mean current of 200A and dividing by the standard deviation of 50A), ultimately generating a public calibration dataset with 3000 samples (used to unify the soft label generation benchmark), covering typical operating conditions such as full load, half load, and start-up / shutdown of the power plant. Next, the global diagnostic model architecture was configured: the global shared encoder feature dimension was set to 128, and the personalized feature dimension to 64. The hierarchical structure of the feature extraction layer (using a 3-layer CNN) and the fault classification layer (2 fully connected layers) was defined, and global shared parameters (CNN layer parameters) and personalized parameters (local parameters of the fully connected layers) were divided. Finally, zero-mean Gaussian distribution (initialized standard deviation 0.01) was used to assign values ​​to the global diagnostic model parameters, and the basic training parameters were set: a total of 150 communication rounds, 40% of clients participating in each round, 4 local training cycles on the client side, and an initial learning rate of 0.001 adjusted using a cosine annealing strategy, thus completing the initialization. This step avoids data privacy leaks by constructing a public calibration dataset without private data and ensures a unified training and evaluation benchmark for all clients; it also reasonably configures the model architecture and parameters, balancing global knowledge sharing and client-specific needs, providing a high-quality initial model for subsequent federated training, and ensuring the accuracy and efficiency of subsequent fault diagnosis and signal reconstruction.

[0074] In step S2, which is federated knowledge fusion training, the core step of cross-node knowledge collaboration under privacy protection, it needs to be executed iteratively in combination with the public calibration dataset initialized in S1 and the global diagnostic model. Continuing with the new energy power plant scenario, the central server first distributes the global diagnostic model initialized in S1 (including a 128-dimensional shared encoder and 64-dimensional personalized parameters) to 20 sensor clients. Each client constructs a hybrid model of "global shared parameters + local personalized parameters", using local private data (such as the current measurement data and historical fault records of the corresponding sensor for the past 3 months) and trains it through a multi-objective optimization function: calculating the task loss with cross-entropy (weight coefficient 0.6), contrastive learning loss (weight coefficient 0.2, temperature parameter 0.1), and orthogonal constraint loss (weight coefficient 0.2) to construct the feature decoupling loss. After training for 4 cycles, a local updated model is obtained. Subsequently, each client uses the local updated model to infer the 3000-sample public calibration dataset constructed in S1, sets the distillation temperature T=5 to generate soft labels (including the probability distribution of sensor health / fault status) and uploads them to the server. The server calculates the KL divergence (knowledge distillation loss) between 20 client-side soft labels and the global model's predictions. Aggregation weights are assigned based on the local training loss of each client (lower loss means higher weight, adjustment coefficient α=0.8), and the global model parameters are updated through weighted aggregation. The "model deployment - local training - soft label upload - model update" process is repeated until, after 120 rounds, the global model's fault identification accuracy on the public dataset stabilizes above 98.5%, at which point the model is considered converged. This step avoids raw data leakage through soft label upload, achieves efficient cross-node knowledge fusion by combining knowledge distillation and feature decoupling, and dynamically allocates aggregation weights to prioritize the absorption of high-quality client knowledge, giving the global model both privacy and high generalization ability, providing reliable model support for subsequent fault isolation.

[0075] In step S3, which involves online fault isolation, a health assessment and fault determination are performed on the real-time signals of multiple current sensors based on the global diagnostic model converged in S2 (fault identification accuracy of 98.5%). Continuing with the new energy power plant scenario, the system simultaneously acquires real-time measurement signals from 20 current sensors (sampling frequency 50Hz, including features such as current amplitude and waveform distortion rate). These signals are input into the converged global model: first, deep features are extracted using a 128-dimensional global shared encoder; then, a health scoring layer (Sigmoid activation) generates health status scores for each sensor. Eighteen sensors have scores between 0.85 and 0.98, while two sensors (numbered 3 and 15) have scores of 0.32 and 0.28, respectively. Next, the dynamic isolation threshold is calculated: based on the mean of the 20 sensor health scores (0.82, standard deviation 0.15), and assuming a sensitivity coefficient k=2, the threshold is calculated to be 0.52 according to the dynamic isolation threshold formula. Simultaneously, a consistency check was performed: the cosine similarity of each sensor's characteristics with the other 19 sensors was calculated. The average similarity of sensors numbered 3 and 15 with most sensors was 0.35 and 0.29, respectively, and the anomaly scores (1 - average similarity) reached 0.65 and 0.71 (far exceeding the normal threshold of 0.3). Based on the comprehensive judgment: sensors numbered 3 and 15 had health scores below the dynamic threshold and high anomalies, and were marked as faulty sensors and immediately isolated, their signals no longer participating in subsequent data processing. This step quantifies the sensor status through health scores, adapts to changes in operating conditions by combining dynamic thresholds, and reduces the risk of misjudgment with consistency checks, achieving accurate identification and rapid isolation of faulty sensors, preventing faulty signals from interfering with the system, and clearing obstacles for subsequent signal reconstruction.

[0076] In step S4, which is an adaptive signal reconstruction, a replacement signal is generated using data from the remaining 18 healthy sensors based on the faulty sensors (numbered 3 and 15) isolated in S3. Continuing with the new energy power plant scenario, two types of weights for the healthy sensors are first calculated: the reliability weight is allocated based on the health score (0.85-0.98) in S3, with sensors having the highest weight (6.2%) for a score of 0.98; the consistency weight is calculated based on the feature similarity (0.82-0.95) with most sensors, with sensors having a similarity of 0.95 accounting for 5.8% of the weight. A fusion coefficient δ=0.6 is set, and the two types of weights are weighted to obtain a comprehensive fusion weight. Then, the consensus true value is calculated: the real-time current readings (range 180-220A) of the 18 healthy sensors are weighted and summed according to the fusion weights to obtain a preliminary estimate of 202.3A. Finally, personalized calibration is performed: the calibration strength coefficient β=0.7, a dedicated calibration network is invoked, and personalized calibration terms (-1.2A to +0.8A) are calculated by combining the depth features and health scores of the health sensors. The consensus estimate is optimized, and the reconstructed signals of faulty sensors 3 and 15 are 201.5A and 202.1A, respectively. The errors between these signals and the actual calibration value (201.8A) are both less than 0.5%, meeting the power plant's requirements for current measurement accuracy (error allowable range ±1%). This step balances sensor health and data consistency through dual weight allocation, and improves signal accuracy through two-stage reconstruction optimization, achieving high-quality recovery of faulty sensor outputs and ensuring stable system operation even when sensors fail.

[0077] In one embodiment, the steps of initializing the federated learning system, constructing a public calibration dataset on a central server that does not contain client-private data, and initializing a global diagnostic model include:

[0078] S10: Collect standard current signals and historical operating data segments of sensors under healthy conditions, standardize the standard current signals and data segments, and generate a common calibration dataset covering typical operating conditions of the system.

[0079] S11: Based on the feature distribution of the public calibration dataset, configure the network architecture parameters of the global diagnostic model, clarify the hierarchical structure of the feature extraction layer and the fault classification layer, and divide the globally shared parameters and personalized parameters.

[0080] S12: Assign values ​​to the parameters of the global diagnostic model using a preset initialization strategy to obtain an initial global diagnostic model.

[0081] In this embodiment, the focus of the current sensor monitoring system in the new energy power plant cluster is on the construction of a public calibration dataset, which is the foundation for ensuring the fairness and privacy of subsequent federated learning. Specifically, two types of core data are collected first: one type is standard current signals generated in the laboratory, including a pure sine wave of 50Hz (amplitude range 100A-300A, sampled at 20A intervals), a composite waveform containing the 3rd harmonic (amplitude accounting for 20% of the fundamental wave), and a composite waveform containing the 5th harmonic (amplitude accounting for 10% of the fundamental wave), generating a total of 1500 standard signal samples; the other type is the operating data segments of 10 fault-free current sensors selected from the power plant's historical database. Continuous 10-minute data points from different time periods each day over the past 6 months (morning peak 8:00 AM, noon 12:00 PM, evening peak 6:00 PM, and 2:00 AM) are selected, and 5-second segments without abnormal fluctuations are extracted, totaling 1500 samples. Private data such as sensor numbers, installation locations, and corresponding unit information are strictly removed. All samples were standardized: the reference mean of the standard current signal was calculated to be 200A and the reference standard deviation to be 50A. The samples were then converted using the formula "(sample value - 200) / 50" to generate a public calibration dataset X containing 3000 samples. pub ={x j |j=1,2,...,3000}, fully covering typical operating conditions of power plants such as full load (250A-300A), half load (150A-200A), and start-up phase (100A-150A fluctuation), ensuring that the data distribution is highly matched with the actual operating scenario.

[0082] The model architecture parameters are configured as follows: the feature dimension of the globally shared encoder is set to 128, and the dimension of the personalized features is set to 64. A public calibration dataset containing 1000 to 5000 samples is constructed, with each training batch containing 32 to 64 samples. Public calibration dataset construction:

[0083]

[0084] in, This is a public calibration dataset and does not contain any private data. The j-th common sample is derived from a standard current signal or historical data from a health sensor. Size of the public dataset; To cover typical operating conditions, the global model is initialized:

[0085]

[0086] in, Global model parameters for initialization; It follows a zero-mean Gaussian distribution. To initialize the standard deviation.

[0087] The network architecture parameters of the global diagnostic model are configured to achieve a balance between "general knowledge sharing" and "node-specific preservation". Based on the feature distribution of the public dataset (key features such as harmonic components and amplitude change rate of the current waveform), the model adopts a two-level architecture of "feature extraction layer + fault classification layer": The feature extraction layer consists of three layers of convolutional neural network (CNN). The first layer has a kernel size of 3×3 and 64 output channels, the second layer has a kernel size of 3×3 and 128 output channels, and the third layer outputs a 128-dimensional feature vector after global average pooling. These parameters are set as globally shared parameters to learn common fault features across sensors (such as zero-point drift, sensitivity reduction, etc.). The fault classification layer contains two fully connected network layers. The first layer concatenates the 128-dimensional shared features with 64-dimensional personalized features (generated by training on local data from the client). The second layer outputs a binary classification result of health / fault. The parameters corresponding to the 64-dimensional personalized features are client-private parameters used to adapt to individual differences of different sensors (such as factory calibration deviation). At the same time, the training batch size is set to 64 samples, which ensures the representativeness of features in a single iteration and avoids overloading the client device's memory.

[0088] The model parameters were initialized and the training strategy was set to provide a stable starting point for federated training. Model parameter initialization ensured that all weight parameters followed a Gaussian distribution with a mean of 0 and a standard deviation of 0.01, avoiding gradient explosion caused by excessively large initial values. Regarding the training strategy, the total number of communication rounds was set to 150 (based on previous testing, this number of rounds allowed the model's accuracy to converge stably on the public dataset). In each round, 40% (8 clients) of the 20 clients were randomly selected to participate in training, ensuring data diversity while reducing communication pressure. The client-side local training cycle was 4 times (each cycle containing 100 iterations), balancing local fitting performance with training time. The initial learning rate was set to 0.001, and the preset initialization strategy was cosine annealing, with the learning rate dynamically adjusted using cosine annealing. The cosine annealing strategy is as follows:

[0089]

[0090] in, Let be the learning rate at the t-th iteration; This represents the minimum learning rate. This represents the initial maximum value of the learning rate; This represents the current iteration number; This represents the total number of iterations in one training cycle. It is a cosine function. =0.0001、 =0.001、 =100 (total iterations in one cycle). Through the periodic decay of the learning rate, rapid convergence is achieved in the early stages of training, followed by fine-tuning in the later stages, reducing the risk of overfitting. This initialization step ensures a unified federated training benchmark by constructing a standardized dataset without privacy data; it rationally divides shared and personalized parameters, balancing model generalization and node adaptability; and it scientifically sets training strategies to ensure efficient model convergence, laying a high-quality foundation for subsequent federated knowledge fusion while mitigating the risk of privacy leakage from the outset.

[0091] In one embodiment, the steps of federated knowledge fusion training based on a public calibration dataset and a global diagnostic model, whereby a central server distributes the global diagnostic model to the client, the client trains an ontology using local data to update the model, generates soft tags and uploads them, and the central server updates the global diagnostic model by fusing the soft tags with knowledge distillation, and repeats this process until convergence, include:

[0092] S20: The central server distributes the current global diagnostic model to each client. The client builds a hybrid model architecture containing globally shared parameters and personalized parameters, trains it using local private data and multi-objective optimization functions, and obtains a locally updated model.

[0093] S21: The client infers on the public calibration dataset through the local update model, generates soft labels containing fault knowledge, and uploads them to the central server;

[0094] S22: The central server uses knowledge distillation technology to integrate all soft tags, and calculates the global model update parameters through weighted aggregation based on the aggregation weight of each client;

[0095] S23: The central server adjusts the global diagnostic model according to the updated parameters and repeats the above steps until the performance of the global diagnostic model converges.

[0096] In this embodiment, the central server distributes the global diagnostic model (containing 128-dimensional globally shared parameters and 64-dimensional personalized parameter interfaces) initialized in stage S1 to 20 sensor clients. Each client constructs a hybrid model architecture, where the globally shared part learns common features across sensors (such as measurement deviation patterns caused by temperature changes), and the personalized part captures individual differences between sensors (such as the inherent 1.2A offset of sensor 3 due to its installation location).

[0097] Among them, building a personalized model architecture:

[0098]

[0099] in, The model prediction output for client i; These are globally shared parameters (participating in federated aggregation); Personalized parameters for client i (not included in aggregation); This is a parameter concatenation operation to achieve feature decoupling;

[0100] Taking client 3 as an example, its local private data contains 2000 records from the past 3 months (1800 normal data, with current stable at 200±3A; 200 fault data, including a 5-8A deviation caused by zero-point drift). Multi-objective optimization was used during training: task loss (weight 0.6) ensured basic diagnostic accuracy; contrastive learning loss (weight 0.2) reduced the feature distance of normal samples to within 0.1 and expanded the distance of fault samples to above 0.8; and orthogonal constraint loss (weight 0.2) reduced the correlation between shared and personalized features to below 0.1. After 4 training cycles, the local model achieved an accuracy of 97.3% on the private data.

[0101] Furthermore, the objective optimization function introduces a feature decoupling loss, which decomposes the model features into globally shared features and client-specific features. These client-specific features are used only in the local reconstructor and do not participate in federated aggregation.

[0102] The feature decoupling loss function is:

[0103]

[0104] in, Task loss weighting coefficient; To compare the learning loss weight coefficients; These are the orthogonal constraint loss weight coefficients; The feature decoupling loss function; The task loss function; To learn the loss function by comparison; The orthogonal constraint loss function;

[0105] The task loss function is:

[0106]

[0107] in, The cross-entropy loss function; Let i be the amount of local data on client i;

[0108] The contrastive learning loss function is:

[0109]

[0110] in, The set of positive sample pairs indexes; For the set of indexes of all samples; The feature representation of sample i; This is a temperature parameter used to control the smoothness of the distribution;

[0111] The orthogonal constraint loss function is:

[0112]

[0113] in, This is the global feature matrix; It is a local feature matrix; It is the Frobenius norm.

[0114] Each client uses a locally updated model to infer the 3000-sample public calibration dataset constructed by S1, and sets the distillation temperature to 5 to generate soft labels (with a smoother probability distribution).

[0115] The formula for generating soft tags is as follows:

[0116]

[0117] in, Predict the probability of the soft label of class k for client i; For distillation temperature, T>1 produces a smoother probability distribution; Number of fault categories;

[0118] For example, for the sample of "200A standard signal containing the 3rd harmonic", client 3 outputs a health probability of 92% and a fault probability of 8%, client 15 outputs a health probability of 91% and a fault probability of 9%, and client 7 outputs a health probability of 93% and a fault probability of 7%. These soft tags are uploaded after encryption, and the original data is always stored locally to avoid leaking the privacy of power plant operation.

[0119] After the server collects soft tags from 8 participating clients, it calculates the knowledge distillation loss (a measure of the difference between the soft tags and the global model prediction). The difference between the soft tags of client 3 and the global model prediction is 0.04, and that of client 15 is 0.05.

[0120] in,

[0121] in, The knowledge distillation loss function; The number of clients participating in the training; The Kullback-Leibler divergence;

[0122]

[0123] in, Let i be the aggregate weight for client i; This is the weighting adjustment coefficient;

[0124] Global model update is achieved through weighted aggregation:

[0125]

[0126] in, These are the global parameters updated in the (t+1)th round; Let i be the global parameters of client i after the t-th round of training;

[0127] Weights are assigned based on loss (lower loss means higher weight). Client 3 has a weight of 0.13, client 15 has a weight of 0.12, and the remaining clients have weights between 0.11 and 0.14. After updating the global model parameters through weighted aggregation, the new model is distributed to the clients. This process is repeated 120 times. The accuracy of the global model on the public dataset increases from the initial 65% to 98.5% and remains stable for 3 rounds, indicating convergence. This step preserves sensor individuality through a hybrid model, protects data privacy through a soft labeling mechanism, and prioritizes the absorption of high-quality knowledge through dynamic weight aggregation. The resulting global model balances diagnostic accuracy (98.5%) and generalization ability, providing reliable support for subsequent fault handling.

[0128] In one embodiment, the steps of acquiring real-time measurement signals from multiple current sensors, inputting the real-time measurement signals into a converged global diagnostic model to obtain a health status score for each sensor, and identifying and isolating faulty sensors based on the health status scores using a dynamic threshold strategy include:

[0129] S30: Acquire real-time measurement signals from multiple current sensors, input the real-time measurement signals into the converged global diagnostic model, and generate health status scores for each sensor through feature extraction and a health scoring layer.

[0130] S31: Calculate the dynamic isolation threshold based on the health status score, and generate an anomaly score by checking the consistency of features between sensors.

[0131] S32: Based on the health status score, dynamic isolation threshold and anomaly score, comprehensively determine and isolate the faulty sensor.

[0132] In this embodiment, in the current monitoring system of the new energy power plant cluster, the system collects measurement signals from 20 current sensors in real time (sampling frequency 50Hz, each frame contains 1024 sampling points, covering features such as current amplitude and waveform distortion rate). These signals are input into a global diagnostic model that has been trained and converged through the S2 stage (achieving a fault identification accuracy of 98.5% on a public dataset). The model first processes the signals through a 128-dimensional global feature extractor to extract deep features such as harmonic content, noise level, and response delay. Then, through a health scoring layer (equipped with a Sigmoid activation function), these high-dimensional features are mapped to the [0,1] interval to generate a health status score—the closer the score is to 1, the healthier the sensor is.

[0133] The health status score is obtained by the following formula:

[0134]

[0135] in, Assess the health of sensor i; Use the Sigmoid activation function; Weights for the health score layer; Bias for the health scoring layer; This is a global feature extractor; the scoring function projects the high-dimensional features of the sensor signal onto the [0,1] interval, where a higher score indicates a better sensor condition.

[0136] In actual operation, the scores of the 18 sensors were stable: sensor 1 scored 0.92 under full load, sensor 5 scored 0.88 under half load, and sensor 8 scored 0.95 during the startup phase; while the scores of sensor 3 and sensor 15 were consistently low, at 0.32 and 0.28 respectively, which were significantly deviated from the normal range.

[0137] Based on the health scores of 20 sensors, statistical parameters were first calculated: the mean health score was 0.82, and the standard deviation was 0.15. To adapt to power plant load fluctuations (such as large current surges during the morning peak hours), a dynamic threshold strategy was adopted, using a sensitivity coefficient of 2 (verified through multiple tests to balance fault detection rate and false alarm rate). The threshold was calculated using the formula "Dynamic isolation threshold = mean - sensitivity coefficient × standard deviation", resulting in a threshold of 0.82 - 2 × 0.15 = 0.52. Simultaneously, feature consistency testing was used for verification: the cosine similarity (quantifying the degree of matching of signal features) of each sensor with the other 19 sensors was calculated. The average similarity of sensor 3 with most sensors was 0.35, and that of sensor 15 was 0.29, while the average similarity of normal sensors was above 0.7. Further calculation of the anomaly score (1 minus the average similarity) yielded a score of 0.65 for sensor 3 and 0.71 for sensor 15, significantly higher than the anomaly level of below 0.3 for normal sensors, indicating that the measurement data of these two sensors deviated significantly from the mainstream trend of the system.

[0138] To adapt to changes in system operating status, a dynamic threshold setting method is adopted:

[0139]

[0140] in, The dynamic isolation threshold; The mean of the health score; This is the sensitivity coefficient; The standard deviation of the health score; the final judgment criteria are as follows:

[0141]

[0142] in, This is a fault indicator; 1 indicates a fault, and 0 indicates normal operation.

[0143] By calculating the cosine similarity of the feature representations of each sensor:

[0144]

[0145] in, Let cosine similarity be the feature similarity between sensors i and j; further calculate the anomaly score for each sensor by quantifying the consistency between sensor readings:

[0146]

[0147] in, An anomaly score is assigned to sensor i, with a higher value indicating greater anomaly. This score reflects the degree of deviation of a specific sensor from the majority of sensors in the system, providing important auxiliary verification for health assessment results and effectively reducing the risk of false alarms.

[0148] According to the fault determination criteria: when a sensor's health score is below the dynamic threshold and its anomaly score is above 0.3 (an empirically verified reliable threshold), it is marked as faulty (fault identifier = 1). The health scores of sensors 3 and 15 (0.32, 0.28) are both below 0.52, and their anomaly scores (0.65, 0.71) are both above 0.3, fully meeting the fault conditions. The system immediately triggers the isolation mechanism, removing the signals from these two sensors from the data bus and simultaneously sending a fault alarm (including sensor number, abnormal period, characteristic deviation value, etc.) to the maintenance terminal. This ensures that the main control system only receives valid data from 18 healthy sensors, preventing fault signals from interfering with the power plant's power regulation and safety decisions. This step, through health score quantification, dynamic threshold adaptation to operating conditions, and consistency verification, achieves accurate identification (98.2% accuracy) and rapid isolation of faulty sensors (response time <100ms), effectively reducing the false alarm rate (<1%) and ensuring system data reliability.

[0149] In one embodiment, the step of calculating a consensus true value based on health sensors other than the isolated faulty sensor, combined with the health status scores of each health sensor, and using the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor includes:

[0150] S40: Collect measurement signals from unisolated health sensors, combine the health status scores of each health sensor with the feature similarity between sensors, and assign reliability weights and consistency weights.

[0151] S41: Construct a fusion weight based on the reliability weight and consistency weight, and perform weighted fusion on the measurement signals of the health sensor to generate a consensus true value;

[0152] S42: An adaptive correction model is introduced, which combines the depth features of the health sensor with the health status score to calculate a personalized correction term. The output signal of the faulty sensor is then reconstructed after optimizing the consensus true value.

[0153] In this embodiment, in the current monitoring system of the new energy power plant cluster, faulty sensors 3 and 15 have been isolated in the previous S3 phase. The system prioritizes collecting real-time measurement signals from the remaining 18 healthy sensors (the sampling time is synchronized with the faulty sensor's failure time to ensure consistent data timestamps; the measured value range is between 198A and 205A, covering the typical current range during half-load operation of the power plant). Weight allocation is based on two dimensions: "reliability" and "consistency."

[0154] The reliability weight and consistency weight are obtained by the following formula:

[0155]

[0156] in, Reliability weights based on health scores; A collection of health sensors; Assess the health of sensor i;

[0157]

[0158] in, Consistency weight; Let be the similarity between sensor i and the average features of most sensors.

[0159] The reliability weight is directly related to the health status score generated in stage S3. A higher score indicates more stable sensor performance, and the higher the weight percentage. For example, sensor 8 has a health score of 0.95 (the highest among the 18 sensors), and after normalization, its weight percentage is 6.2%; sensor 5 has a health score of 0.88, with a weight percentage of 5.7%; and sensor 12 has a health score of 0.85, with a weight percentage of 5.5%. The total reliability weight of the 18 sensors is strictly 1, ensuring that the allocation logic is fair and consistent with the actual health level of the sensors. The consistency weight is calculated based on the feature similarity between sensors. By comparing the cosine similarity of the 128-dimensional deep features of each sensor with the other 17 healthy sensors, the matching degree between its data and the group trend is quantified. Sensor 1 has an average similarity of 0.92 with most sensors (the highest), with a consistency weight percentage of 5.9%; sensor 7 has an average similarity of 0.89, with a consistency weight percentage of 5.5%; while sensor 9, with a measured value of 198A (slightly deviating from the group average of 202A), has an average similarity of 0.81, with a consistency weight percentage of only 5.1%, thus weakening the impact of deviation data on subsequent fusion.

[0160] Based on power plant operation experience (data stability should be prioritized under half-load conditions), the fusion coefficient is set to 0.6. The comprehensive weight is calculated according to the logic of "fusion weight = 0.6 × reliability weight + 0.4 × consistency weight": the fusion weight of sensor 8 is 0.6 × 6.2% + 0.4 × 5.8% = 6.1%; the fusion weight of sensor 1 is 0.6 × 5.6% + 0.4 × 5.9% = 5.8%; and the fusion weight of sensor 9 is 0.6 × 5.3% + 0.4 × 5.1% = 5.2%.

[0161] The fusion weights are obtained by the following formula:

[0162]

[0163] in, For an adjustable fusion coefficient, a dynamic balance is achieved between reliability weights based on individual health status and synergy weights based on group consistency. Based on this, a set of health sensors is used... The readings from each sensor are weighted and fused to generate a consensus estimate of the true current value.

[0164]

[0165] Subsequently, the real-time measurement signals from 18 health sensors were weighted and summed according to their fusion weights. Specifically, the measurement value of each sensor was first multiplied by its fusion weight, and then all results were summed—for example, sensor 8 (204A × 6.1%), sensor 1 (202A × 5.8%), sensor 9 (198A × 5.2%), etc.—ultimately generating a consensus true value of 202.3A. This value incorporates effective measurement information from the health sensors while suppressing interference from slightly off-target data from sensors like sensor 9 through low weighting, ensuring that the preliminary estimate closely approximates the actual current level.

[0166] To balance the stability of consensus estimation with the need for sensor individual bias compensation, a correction strength coefficient of 0.7 is set (so that the consensus true value dominates the final result). A dedicated correction network is invoked, and the 128-dimensional deep features of the health sensors (such as waveform harmonic distribution and noise fluctuation characteristics) and real-time health scores are input to calculate personalized correction terms: Sensor 8 has a slight positive bias due to long-term operation, and the correction term is +0.2A; Sensor 1 has a slight negative bias, and the correction term is -0.1A; Combining the correction terms of all health sensors, a global basic correction term of -0.3A is generated.

[0167] To further improve the accuracy and adaptability of the reconstructed signal, the system introduces an adaptive correction model to optimize the initial estimate:

[0168]

[0169] in, To correct the intensity coefficient, the weight of the consensus estimate and the sensor's raw reading in the final output is flexibly adjusted; These are personalized correction items, which are provided by a dedicated correction network. Sensor-based depth feature representation and their health status The calculation shows that:

[0170]

[0171] in, For output items; The depth feature vector of sensor i extracted from the global model; Provide a real-time health score for sensor i; To correct the set of trainable parameters for the network; this design enables the system to learn and compensate for the inherent bias characteristics of specific sensors, thereby achieving more accurate signal reconstruction.

[0172] Simultaneously, based on historical data from the faulty sensors (sensor 3 previously showed no significant deviation, while sensor 15 exhibited an inherent deviation of -0.2A), the consensus true values ​​were optimized: the reconstructed value for faulty sensor 3 was 202.3A (consensus true value) × 0.7 + (202.3A - 0.3A) × 0.3 = 202.1A; the reconstructed value for faulty sensor 15 was 202.3A × 0.7 + (202.3A - 0.3A - 0.2A) × 0.3 = 201.8A. Verification by the power plant's on-site calibration equipment confirmed that the actual current value at that moment was 202.0A. The errors between the reconstructed and actual values ​​were 0.05% and 0.1%, respectively, far below the power plant's ±1% accuracy requirement for current measurement, fully meeting the data requirements for subsequent load dispatching and safety monitoring. This step balances individual sensor reliability and group consistency through dual weight allocation. After adaptive correction to compensate for deviations, the reconstructed signal error is controlled within 0.1%, ensuring data continuity during faults while meeting high precision requirements, thus providing reliable data support for the stable operation of the power plant.

[0173] refer to Figure 2 A current sensor fault isolation and signal reconstruction device, comprising:

[0174] Initialization module 100 is used to initialize the federated learning system, build a public calibration dataset on the central server that does not contain client private data, and initialize the global diagnostic model;

[0175] The federated training module 200 is used to carry out federated knowledge fusion training based on the public calibration dataset and the global diagnostic model. The central server distributes the global diagnostic model to the client. The client uses local data to train the ontology to update the model and generates soft labels for uploading. The central server uses knowledge distillation to fuse the soft labels to update the global diagnostic model and repeats the process until convergence.

[0176] The fault isolation module 300 is used to acquire real-time measurement signals from multiple current sensors, input the real-time measurement signals into a converged global diagnostic model to obtain a health status score for each sensor, and identify and isolate faulty sensors based on the health status score using a dynamic threshold strategy.

[0177] The signal reconstruction module 400 is used to calculate a consensus true value based on the health sensors other than the isolated faulty sensor and the health status scores of each health sensor, and to use the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

[0178] Furthermore, the aforementioned initialization module 100 includes:

[0179] The public calibration dataset construction unit is used to collect standard current signals and historical operating data segments of sensors under healthy conditions, and to standardize the standard current signals and data segments to generate a public calibration dataset covering typical operating conditions of the system.

[0180] The model architecture parameter configuration unit is used to configure the network architecture parameters of the global diagnostic model based on the feature distribution of the public calibration dataset, clarify the hierarchical structure of the feature extraction layer and the fault classification layer, and divide the globally shared parameters and personalized parameters.

[0181] The global model parameter initialization unit is used to assign values ​​to the parameters of the global diagnostic model using a preset initialization strategy to obtain an initial global diagnostic model.

[0182] Furthermore, the aforementioned federal training module 200 includes:

[0183] The client-side local model training unit is used by the central server to distribute the current global diagnostic model to each client. The client builds a hybrid model architecture containing globally shared parameters and personalized parameters, and trains it using local private data and multi-objective optimization functions to obtain a locally updated model.

[0184] The soft tag generation and uploading unit is used by the client to infer the public calibration dataset through the local update model, generate soft tags containing fault knowledge, and upload them to the central server.

[0185] The global model parameter update unit is used by the central server to integrate all soft tags using knowledge distillation technology, and calculate the global model update parameters through weighted aggregation based on the aggregation weight of each client.

[0186] The model convergence determination and iteration unit is used by the central server to adjust the global diagnostic model according to the updated parameters, and repeat the above steps until the performance of the global diagnostic model converges.

[0187] Furthermore, the objective optimization function described above introduces a feature decoupling loss, which decomposes the model features into globally shared features and client-specific features. The client-specific features are used only in the local reconstructor and do not participate in federated aggregation.

[0188] The feature decoupling loss function is:

[0189]

[0190] in, Task loss weighting coefficient; To compare the learning loss weight coefficients; These are the orthogonal constraint loss weight coefficients; The feature decoupling loss function; The task loss function; To learn the loss function by comparison; The orthogonal constraint loss function;

[0191] The task loss function is:

[0192]

[0193] in, The cross-entropy loss function; Let i be the amount of local data on client i;

[0194] The contrastive learning loss function is:

[0195]

[0196] in, The set of positive sample pairs indexes; For the set of indexes of all samples; The feature representation of sample i; This is a temperature parameter used to control the smoothness of the distribution;

[0197] The orthogonal constraint loss function is:

[0198]

[0199] in, This is the global feature matrix; It is a local feature matrix; It is the Frobenius norm.

[0200] Furthermore, the aforementioned fault isolation module 300 includes:

[0201] The sensor health status score generation unit is used to acquire real-time measurement signals from multiple current sensors, input the real-time measurement signals into the converged global diagnostic model, and generate health status scores for each sensor through feature extraction and a health score layer.

[0202] The dynamic threshold and anomaly score calculation unit is used to calculate the dynamic isolation threshold based on the health status score, and at the same time generate an anomaly score through the consistency test of features between sensors.

[0203] The fault sensor comprehensive judgment and isolation unit is used to comprehensively judge and isolate faulty sensors based on the health status score, dynamic isolation threshold and anomaly score.

[0204] Furthermore, the aforementioned signal reconstruction module 400 includes:

[0205] The health sensor weight allocation unit is used to collect measurement signals from unisolated health sensors, and combine the health status scores of each health sensor with the feature similarity between sensors to allocate reliability weights and consistency weights.

[0206] The consensus true value generation unit is used to construct a fusion weight based on the reliability weight and consistency weight, and to perform weighted fusion on the measurement signals of the health sensor to generate a consensus true value.

[0207] The fault signal adaptive reconstruction unit is used to introduce an adaptive correction model, combine the depth features of the health sensor with the health status score to calculate personalized correction terms, and reconstruct the output signal of the fault sensor after optimizing the consensus true value.

[0208] Furthermore, the reliability weight and consistency weight mentioned above are obtained by the following formula:

[0209]

[0210] in, Reliability weights based on health scores; A collection of health sensors; Assess the health of sensor i;

[0211]

[0212] in, Consistency weight; Let be the similarity between sensor i and the average features of most sensors.

[0213] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3As shown, this computer device includes a processor, memory, network interface, and database connected via a bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operations, computer programs, and the database. The internal memory provides an environment for the operation of the operations and computer programs stored in the non-volatile storage media. The database stores data such as current sensor fault isolation and signal reconstruction methods. The network interface is used for communication with external terminals via a network connection. When executed by a processor, this computer program implements a method for fault isolation and signal reconstruction of current sensors, including the following steps: initializing a federated learning system, constructing a public calibration dataset without client-private data on a central server, and initializing a global diagnostic model; conducting federated knowledge fusion training based on the public calibration dataset and the global diagnostic model, the central server distributing the global diagnostic model to the client, the client using local data to train an ontology to update the model, generating soft tags and uploading them, the central server updating the global diagnostic model by fusing the soft tags with knowledge distillation, and repeating this process until convergence; acquiring real-time measurement signals from multiple current sensors, inputting the real-time measurement signals into the converged global diagnostic model to obtain a health status score for each sensor, identifying and isolating faulty sensors based on the health status scores using a dynamic threshold strategy; calculating a consensus true value based on the health status scores of the healthy sensors other than the isolated faulty sensors, and using the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

[0214] One embodiment of this application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements a method for fault isolation and signal reconstruction of a current sensor, including the following steps: initializing a federated learning system, constructing a public calibration dataset without client-private data on a central server, and initializing a global diagnostic model; conducting federated knowledge fusion training based on the public calibration dataset and the global diagnostic model, the central server distributing the global diagnostic model to the client, the client training an ontology update model using local data, generating and uploading soft tags, the central server updating the global diagnostic model with knowledge distillation fusion of soft tags, and repeating until convergence; acquiring real-time measurement signals from multiple current sensors, inputting the real-time measurement signals into the converged global diagnostic model to obtain health status scores for each sensor, identifying and isolating faulty sensors based on the health status scores using a dynamic threshold strategy; calculating a consensus true value based on healthy sensors other than the isolated faulty sensors, combined with the health status scores of each healthy sensor, and using the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

[0215] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0216] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for fault isolation and signal reconstruction of a current sensor, characterized in that, include: The federated learning system is initialized by building a public calibration dataset on the central server that does not contain client-private data, and by initializing the global diagnostic model. Federated knowledge fusion training is carried out based on public calibration dataset and global diagnostic model. The central server distributes global diagnostic model to client. Client uses local data to train ontology to update model and generates soft tags for uploading. The central server uses knowledge distillation to fuse soft tags to update global diagnostic model, and repeats until convergence. The real-time measurement signals of multiple current sensors are acquired, and the real-time measurement signals are input into the converged global diagnostic model to obtain the health status score of each sensor. Based on the health status score, a dynamic threshold strategy is used to identify and isolate faulty sensors. Based on the health sensors other than the isolated faulty sensors, a consensus true value is calculated by combining the health status scores of each health sensor. The consensus true value is then used to perform online correction and reconstruction of the output signal of the faulty sensor.

2. The current sensor fault isolation and signal reconstruction method according to claim 1, characterized in that, The steps of initializing the federated learning system, constructing a public calibration dataset on the central server that does not contain client-private data, and initializing the global diagnostic model include: Collect standard current signals and historical operating data segments of sensors under healthy conditions, perform standardization processing on the standard current signals and data segments, and generate a common calibration dataset covering typical operating conditions of the system; Based on the feature distribution of the public calibration dataset, configure the network architecture parameters of the global diagnostic model, clarify the hierarchical structure of the feature extraction layer and the fault classification layer, and divide the globally shared parameters and personalized parameters. The parameters of the global diagnostic model are assigned values ​​using a preset initialization strategy to obtain an initial global diagnostic model.

3. The current sensor fault isolation and signal reconstruction method according to claim 1, characterized in that, The steps of federated knowledge fusion training based on a public calibration dataset and a global diagnostic model, whereby the central server distributes the global diagnostic model to the client, the client trains an ontology using local data to update the model, generates soft tags and uploads them, and the central server updates the global diagnostic model by fusing the soft tags with knowledge distillation, and repeats until convergence, include: The central server distributes the current global diagnostic model to each client. The client builds a hybrid model architecture containing globally shared parameters and personalized parameters, trains it using local private data and multi-objective optimization functions, and obtains a locally updated model. The client uses the local update model to infer the public calibration dataset, generates soft labels containing fault knowledge, and uploads them to the central server; The central server uses knowledge distillation technology to integrate all soft tags and calculates global model update parameters through weighted aggregation based on the aggregation weight of each client. The central server adjusts the global diagnostic model based on the updated parameters, repeating the above steps until the performance of the global diagnostic model converges.

4. The current sensor fault isolation and signal reconstruction method according to claim 3, characterized in that, The objective optimization function introduces a feature decoupling loss, which decomposes model features into globally shared features and client-specific features. These client-specific features are used only in the local reconstructor and do not participate in federated aggregation. The feature decoupling loss function is: in, Task loss weighting coefficient; To compare the learning loss weight coefficients; These are the orthogonal constraint loss weight coefficients; The feature decoupling loss function; The task loss function; To learn the loss function by comparison; The orthogonal constraint loss function; The task loss function is: in, The cross-entropy loss function; Let i be the amount of local data on client i; The contrastive learning loss function is: in, The set of positive sample pairs indexes; For the set of indexes of all samples; The feature representation of sample i; This is a temperature parameter used to control the smoothness of the distribution; The orthogonal constraint loss function is: in, This is the global feature matrix; It is a local feature matrix; It is the Frobenius norm.

5. The current sensor fault isolation and signal reconstruction method according to claim 1, characterized in that, The steps of acquiring real-time measurement signals from multiple current sensors, inputting these signals into a converged global diagnostic model to obtain a health status score for each sensor, and identifying and isolating faulty sensors based on these health status scores using a dynamic threshold strategy include: The real-time measurement signals of multiple current sensors are acquired, and the real-time measurement signals are input into the converged global diagnostic model. The health status score of each sensor is generated through feature extraction and health scoring layer. A dynamic isolation threshold is calculated based on the health status score, and an anomaly score is generated by checking the consistency of features between sensors. Based on the health status score, dynamic isolation threshold, and anomaly score, faulty sensors are comprehensively identified and isolated.

6. The current sensor fault isolation and signal reconstruction method according to claim 1, characterized in that, The step of calculating a consensus true value based on health sensors other than the isolated faulty sensor, and combining the health status scores of each health sensor, and then using the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor includes: Collect measurement signals from unisolated health sensors, combine the health status scores of each health sensor with the feature similarity between sensors, and assign reliability weights and consistency weights. Based on the aforementioned reliability weight and consistency weight, a fusion weight is constructed, and the measurement signals from the health sensor are weighted and fused to generate a consensus true value. An adaptive correction model is introduced, which combines the depth features of the health sensor with the health status score to calculate a personalized correction term. The output signal of the faulty sensor is then reconstructed after optimizing the consensus true value.

7. The current sensor fault isolation and signal reconstruction method according to claim 6, characterized in that, The reliability weight and consistency weight are obtained by the following formula: in, Reliability weights based on health scores; A collection of health sensors; Assess the health of sensor i; in, Consistency weight; Let be the similarity between sensor i and the average features of most sensors.

8. A current sensor fault isolation and signal reconstruction device, characterized in that, include: The initialization module is used to initialize the federated learning system, build a public calibration dataset on the central server that does not contain client-private data, and initialize the global diagnostic model. The federated training module is used to conduct federated knowledge fusion training based on a public calibration dataset and a global diagnostic model. The central server distributes the global diagnostic model to the client, the client trains the ontology using local data to update the model, and generates soft labels for uploading. The central server uses knowledge distillation to fuse the soft labels to update the global diagnostic model, and repeats the process until convergence. The fault isolation module is used to acquire real-time measurement signals from multiple current sensors, input the real-time measurement signals into a converged global diagnostic model to obtain a health status score for each sensor, and identify and isolate faulty sensors based on the health status score using a dynamic threshold strategy. The signal reconstruction module is used to calculate a consensus true value based on the health sensors other than the isolated faulty sensor and the health status scores of each health sensor, and to use the consensus true value to perform online correction and reconstruction of the output signal of the faulty sensor.

9. A current sensor fault isolation and signal reconstruction device, characterized in that, include: processor; Memory, used to store computer programs; Wherein, when the processor executes the computer program, it implements the current sensor fault isolation and signal reconstruction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the current sensor fault isolation and signal reconstruction method according to any one of claims 1 to 7.

Images