Current transformer multi-physical quantity signal decoupling method and system based on optimized CNN

Abstract

This invention relates to the field of power equipment condition monitoring and signal processing technology, specifically disclosing a method for decoupling multiple physical quantity signals in a current transformer based on an optimized CNN. The method includes the following steps: acquiring a composite sensing signal of temperature, pressure, and hydrogen concentration collected by a fiber optic grating sensor within the current transformer; constructing a convolutional neural network model for signal decoupling; optimizing the hyperparameters of the convolutional neural network model using the Osprey optimization algorithm to obtain an optimal hyperparameter combination; configuring and training the convolutional neural network model using the optimal hyperparameter combination to obtain a trained decoupling model; inputting the composite sensing signal to be decoupled into the trained decoupling model, and outputting decoupled independent temperature, pressure, and hydrogen concentration signals. This invention is beneficial for improving the decoupling accuracy of temperature, pressure, and hydrogen concentration in the FBG composite sensing signal and effectively overcoming the cross-sensitivity problem.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and signal processing technology, and in particular to a method and system for decoupling multiple physical quantity signals of current transformers based on optimized CNN. Background Technology

[0002] Fiber Bragg grating (FBG) sensors are widely used for online monitoring of multiple parameters of the internal state of high-voltage power equipment such as current transformers due to their advantages such as resistance to electromagnetic interference, corrosion resistance, and ease of networking. They can simultaneously monitor key parameters such as temperature, pressure, and hydrogen concentration. However, the inherent cross-sensitivity of FBG sensors is the main technical bottleneck in their practical applications: the wavelength drift of an FBG is simultaneously affected by changes in multiple physical quantities such as temperature, strain (pressure), and hydrogen concentration, making it difficult to accurately separate the composite signals from a single sensor, which seriously affects the accuracy and reliability of monitoring individual parameters.

[0003] To demodulate multi-parameter information in FBG sensors, existing technologies mainly rely on designing sensors with special structures (such as dual gratings or special optical fibers) or establishing complex joint demodulation mathematical models. These methods are often costly, complex, and their decoupling accuracy is easily affected by model errors and the environment. In recent years, deep learning-based signal processing methods have provided new ideas for solving cross-sensitivity problems. For example, Convolutional Neural Networks (CNNs) have been attempted for regression and decoupling of FBG signals due to their powerful feature extraction capabilities. However, the performance of CNN models is highly dependent on the setting of their hyperparameters (such as learning rate and kernel size). Inappropriate hyperparameters can lead to non-convergence, slow convergence, or getting stuck in local optima, which seriously affects the final decoupling accuracy. Currently, CNN hyperparameters are mostly set manually based on experience or by grid search, which suffers from blind parameter tuning, large computational load, and difficulty in obtaining the global optimum, failing to meet the requirements of high-precision and high-reliability status monitoring of current transformers.

[0004] Therefore, how to intelligently and automatically optimize the CNN model to efficiently and accurately decouple the temperature-pressure-hydrogen concentration composite signal from the current transformer FBG sensor has become a pressing technical problem in this field. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method and system for decoupling multiple physical quantity signals of current transformers based on optimized CNN. By using an intelligent optimization algorithm to automatically find the optimal hyperparameters of the CNN model, the decoupling accuracy of temperature, pressure and hydrogen concentration in FBG composite sensing signals is significantly improved, effectively overcoming the cross-sensitivity problem, thereby providing a reliable data foundation for accurate condition assessment and fault early warning of current transformers.

[0006] To achieve the above objectives, on the one hand, the present invention provides a method for decoupling multiple physical quantity signals of a current transformer based on optimized CNN, comprising the following steps: Acquire the temperature-pressure-hydrogen concentration composite sensing signal collected by the fiber optic grating sensor inside the current transformer; Construct a convolutional neural network model for signal decoupling; The hyperparameters of the convolutional neural network model are optimized using the Osprey optimization algorithm to obtain the optimal combination of hyperparameters; The convolutional neural network model is configured and trained using the optimal hyperparameter combination to obtain a trained decoupling model. The composite sensing signal to be decoupled is input into the trained decoupling model, which outputs the decoupled independent temperature signal, independent pressure signal, and independent hydrogen concentration signal.

[0007] Furthermore, the optimization of the hyperparameters of the convolutional neural network model using the Osprey optimization algorithm specifically involves: The learning rate and kernel size of the convolutional neural network model are used as optimization variables. The Osprey optimization algorithm is used to iteratively search within the set search space to minimize the model prediction error and determine the optimal learning rate and kernel size.

[0008] Furthermore, the Osprey optimization algorithm is used to perform an iterative search within the defined search space, specifically including the following steps: S31. Initialize the osprey population and use the mean squared error of the convolutional neural network model on the validation set as the fitness function. S32. For the i-th osprey, determine the set of fish locations it can choose based on its fitness value. The set includes the positions of individuals in the population with better fitness than the current osprey and the globally optimal position. ; S33. Perform a global exploration phase location update: based on the formula Calculate candidate positions And compare it with the current position, updating the position according to the better one; where x i，j Let r be the j-th dimension component at the current position. i，j A random number in the range (0, 1). From The j-th dimension component of the randomly selected fish position, I i，j It is a random integer in {1, 2}; S34. Perform location update for the partial mining phase: according to the formula Calculate candidate positions It compares the current position with the current position and updates the position of the better one; where t is the current iteration number, and lb is the position of the current iteration number. j and ub j Let be the lower and upper bounds for the search of the j-th dimension parameter; S35. Repeat steps S32 to S34 until the preset number of iterations is met, and output the hyperparameter combination corresponding to the global optimal position as the optimal hyperparameter combination.

[0009] Furthermore, in step S31, the osprey position vector is a hybrid encoding vector, comprising a continuous parameter part and a discrete parameter part; wherein, the continuous parameter part includes at least one of learning rate, regularization coefficient, and dropout rate; and the discrete parameter part includes at least one of number of convolutional layers, number of convolutional kernels, and convolutional kernel size. In steps S33 and S34, the dimension values ​​corresponding to the discrete parameter part are rounded and boundary constraints are applied.

[0010] Furthermore, the convolutional neural network model is a multi-branch attention network structure, including: A shared feature extraction layer is used to perform preliminary feature extraction on the input composite sensing signal; Multiple feature decoupling branches are set up in parallel to extract features that are sensitive to different physical quantities, including temperature, pressure and hydrogen concentration. An attention fusion layer is used to perform weighted fusion of the output features of the multiple feature decoupled branches, wherein the fusion weights are generated through a learnable attention mechanism.

[0011] Furthermore, the discrete parameter section also includes weight bias parameters for initializing the attention fusion layer; the fusion weight 'a' of the attention fusion layer is determined by the formula... Calculate, where, It is a learnable weight matrix, and h is the concatenated vector of output features from each branch. The initial value is set by the weight bias parameter obtained through optimization.

[0012] Furthermore, the output of the convolutional neural network model also includes prediction uncertainty estimates corresponding to the decoupling values ​​of each physical quantity; The loss function L used to train the convolutional neural network model is: in, , Let N be the estimated uncertainty of the prediction for the m-th physical quantity, and N be the sample size. and These are the actual value and the predicted value of the m-th physical quantity in the k-th sample, respectively.

[0013] Furthermore, in step S31, the fitness function is the weighted mean square error of the convolutional neural network model's predictions of temperature, pressure, and hydrogen concentration on the validation set.

[0014] On the other hand, the present invention also provides a multi-physical quantity signal decoupling system for current transformers based on optimized CNN, comprising: The signal acquisition module is used to acquire the temperature-pressure-hydrogen concentration composite sensing signal collected by the fiber optic grating sensor inside the current transformer. The model building and optimization module is used to execute the aforementioned methods to build and train the optimal decoupled model; The signal decoupling module is used to decouple the input composite sensing signal using the decoupling model and output independent temperature, pressure and hydrogen concentration signals.

[0015] Furthermore, the system also includes: The human-machine interface provides a graphical interface for configuring system parameters and visually displaying real-time curves and historical trends of decoupled temperature, pressure, and hydrogen concentration; as well as... External communication unit, used for external communication.

[0016] The present invention provides a method and system for decoupling multiple physical quantity signals of a current transformer based on optimized CNN, which has at least the following beneficial technical effects compared with the prior art: First, by automatically and globally optimizing the key hyperparameters of the CNN model using the Osprey optimization algorithm, the blindness of manual parameter tuning is overcome, enabling the model to achieve optimal performance. Experiments show that, compared with the traditional CNN model, the method of this invention (OOA-CNN) can reduce the maximum absolute error of hydrogen concentration decoupling by 94.8%, temperature decoupling error by 48.1%, and pressure decoupling error by 11.5%, and all error indicators (MAE, MSE, RMSE, MAPE) are significantly reduced. Second, by deeply integrating intelligent optimization algorithms with deep learning models, the entire process from signal input to parameter optimization and high-precision decoupling is automated, reducing reliance on professional experience and enhancing the adaptability and robustness of the method under different working conditions. Third, this invention directly addresses specific engineering problems in FBG monitoring of current transformers, providing decoupling signals with high accuracy and reliability, which can be directly used for equipment condition assessment and fault early warning, and has significant engineering application value. At the same time, the proposed optimization schemes such as hybrid coding optimization, multi-branch attention structure and uncertainty estimation further improve the performance of the method and the richness of output information. Attached Figure Description

[0017] Figure 1This is a flowchart of the signal decoupling method provided by the present invention; Figure 2 Flowchart of algorithm implementation for OOA-CNN model optimization and decoupling; Figure 3 The mean square error convergence curves of the Osprey optimization algorithm at different iteration numbers are shown. Figure 4 A comparison of the mean square error under different osprey population sizes; Figure 5 A diagram showing the comparison of temperature decoupling error of the CNN model before and after OOA optimization; Figure 6 A comparative diagram showing the stress decoupling error of the CNN model before and after OOA optimization; Figure 7 A diagram showing the comparison of the decoupling error of the CNN model for hydrogen concentration before and after OOA optimization. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] First, it should be noted that the Osprey Optimization Algorithm (OOA) is a novel metaheuristic optimization algorithm inspired by the hunting behavior of ospreys in nature. The core idea of ​​this algorithm is to solve complex optimization problems by simulating the efficient hunting strategies of ospreys using mathematical models. Its search process is mainly divided into two phases to balance the exploration (finding new potential regions) and development (refined searching within known high-quality regions) of the solution space. The exploration phase simulates the process of an osprey flying through the air, discovering and swooping down to hunt. In the algorithm, each osprey represents a potential solution, and they update their positions according to certain strategies (such as random selection or moving towards better solutions) to explore better regions in the solution space. The development phase simulates the process of an osprey bringing its prey to a safe and suitable location to enjoy it after capture. In the algorithm, this corresponds to a more refined local search near the currently found better solutions in order to find the optimal solution within that region.

[0020] Example 1 This embodiment details a specific method for optimizing the hyperparameters of a convolutional neural network based on the Osprey optimization algorithm to achieve decoupling of the temperature-pressure-hydrogen concentration signals of a current transformer.

[0021] Step S101: Data acquisition and preprocessing.

[0022] Acquire composite sensing signals (time-series data) from the fiber optic grating (FBG) sensor inside the current transformer, which contain cross-sensitive information on temperature, pressure, and hydrogen concentration. Divide the obtained experimental data into training and test sets in a 4:1 ratio. To eliminate dimensional differences and accelerate model convergence, normalize the data in the training and test sets respectively, scaling the values ​​of each feature dimension (i.e., composite signal sequence) to the [0, 1] interval.

[0023] Step S102: Construct the CNN model and define the optimization objective.

[0024] Construct a convolutional neural network model for multi-output regression. The construction of a convolutional neural network model usually includes steps such as defining inputs and outputs, designing the network structure, and automatic optimization and training, which are well known to those skilled in the art and will not be elaborated here.

[0025] The convolutional neural network model structure in this embodiment includes an input layer, two convolutional layers (each convolutional layer is followed by a ReLU activation function and a max pooling layer), a flattening layer, and a fully connected output layer; the model output consists of three neurons, which correspond to the decoupled estimates of temperature (T), pressure (P), and hydrogen concentration (H2), respectively.

[0026] In this embodiment, the learning rate (lr) and kernel size (Ksize) of the CNN model are defined as hyperparameters to be optimized; the search space for the learning rate is set as [ , The search space for the convolution kernel size is a discrete value {3, 5, 7}.

[0027] Step S103: Use the Osprey optimization algorithm to optimize hyperparameters.

[0028] This step is crucial for improving the decoupling accuracy of the method shown in this embodiment. It automatically finds the optimal hyperparameter combination through the OOA algorithm. The specific optimization process is as follows: 1. Initialize OOA algorithm parameters: Determine the algorithm parameters through preliminary experiments; such as... Figure 3 As shown in the (error convergence curve), after testing, the fitness function (MSE) has fully converged when the number of iterations reaches 40; therefore, the maximum number of iterations T is set to 40. Figure 4 (Population size optimization diagram) As shown, the average MSE under different population sizes was tested, and it was determined that the best effect was achieved when the population size N=43. Therefore, the population size N=43 was set. The population was randomly initialized, and the position vector of each osprey was defined as Xi=[lri, Ksizei].

[0029] 2. Define the fitness function: The fitness function F(Xi) is used to evaluate the performance of each Osprey position (i.e., a set of hyperparameters). In this embodiment, F(Xi) is defined as: the sum of the mean squared errors of the model's predictions of temperature, pressure, and hydrogen concentration after training the CNN model on a separately reserved validation set (a small number of rounds) using the hyperparameters corresponding to position Xi, i.e., minimizing F(Xi).

[0030] 3. Iterative optimization: For each individual i (i=1, 2, ..., 43) in each generation t (t=1, 2, ..., 40), perform the following operation: a) Determine the fish set: For the i-th osprey, the set of fish positions FP that it can choose. i Defined by the following formula: Among them, X best For the historical global optimal position, F k and F i These are the fitness values ​​for the corresponding positions.

[0031] b) Global Exploration Phase: From FP i Randomly select a location as the target fish SF i ; Calculate the j-th dimension component of the new candidate position according to the following formula. : Where j = 1, 2; r i，j ~U(0,1); I i，j Let be a random integer in {1, 2}; calculate the fitness F(XiP1) of the candidate position. If F( ) <F(X i If X is updated, then X is updated. i = .

[0032] c) Local mining stage: Calculate new candidate locations according to the following formula. The j-th dimension component : Among them, lb j and ub j These are the lower and upper bounds of the search in the j-th dimension; calculate F( If F( ) <F(X i If X is updated, then X is updated. i = .

[0033] 4. Output the optimal solution: After the iteration is complete, output the global optimal position X. best Its corresponding [lropt Ksize opt This is the optimal combination of hyperparameters.

[0034] Step S104: Model training and decoupling.

[0035] By using the optimal hyperparameter configuration and training the CNN model, the final OOA-CNN decoupling model is obtained for signal decoupling.

[0036] Experimental results verification: The OOA-CNN model in this embodiment is compared with an unoptimized CNN model.

[0037] like Figure 5-7 As shown, the OOA-CNN model exhibits superior decoupling accuracy in temperature and hydrogen decoupling scenarios, with a maximum absolute error of 1.94℃, a 48.1% reduction compared to the traditional CNN model's 3.74℃. In pressure decoupling, the maximum absolute error of OOA-CNN is further reduced to 1.38kPa, an 11.5% reduction compared to the CNN model (1.55kPa). In hydrogen concentration decoupling, the maximum absolute error is 0.09uL / L, a 94.8% reduction compared to the traditional CNN model's 1.72uL / L.

[0038] The decoupling effects of different models were further compared using four metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), as shown in Tables 1, 2, and 3.

[0039] Table 1 Comparison of temperature decoupling effects of different models Table 2 Comparison of pressure decoupling effects of different models Table 3 Comparison of decoupling effects of different models with hydrogen concentration Experimental data fully demonstrate that the OOA-CNN method provided in this embodiment outperforms the undecoupled state and traditional CNN models in all four key error indicators (MAE, MSE, RMSE, and MAPE) for temperature, pressure, and hydrogen concentration, effectively solving the cross-sensitivity problem of FBG sensors.

[0040] Example 2 This embodiment enhances the search space and network model structure of the optimization algorithm based on Embodiment 1 to achieve better performance.

[0041] Building upon Example 1, the search space of the Osprey optimization algorithm is expanded from two hyperparameters to a hybrid encoded vector containing both continuous and discrete parameters. Specifically, the position vector of each Osprey is defined as Xi = [Ci, Di]. Here, Ci represents the continuous parameter portion, such as the learning rate lr, L2 regularization coefficient, and dropout rate. rate Di represents the discrete integer parameter part, such as the number of convolutional layers L, the number of convolutional kernels nfilters, and the convolutional kernel size Ksize.

[0042] During the OOA iteration process, after the position update of the aforementioned formula is executed, the floor operation is performed on each discrete dimension value in the vector Di, and it is checked whether it is within the preset set of discrete values. If it is out of bounds, it is constrained to the nearest boundary value, so that the algorithm can automatically and collaboratively optimize hyperparameters and network architecture.

[0043] In addition, a dedicated multi-branch attention convolutional neural network (AM-CNN) can be built to replace the base CNN. This network includes: Shared Feature Extraction Layer: A basic convolutional layer used to extract shared low-level features F from the input composite signal. shared .

[0044] Parallel decoupled branches: three structurally identical but parameter-independent convolutional sub-network branches (Branch) T Branch P Branch H Each branch starts with F. shared As input, it learns deep features related to temperature, pressure or hydrogen concentration through several layers of convolution, and outputs feature vectors respectively.

[0045] Learnable attention fusion layer: concatenates the feature vectors of the three branches into h=[f t f p f h The fusion weights for each branch feature are automatically calculated using a learnable attention module. Where h is the concatenation of the feature vectors of each branch. It is a trainable weight matrix, b a It is a trainable bias vector. a=[a t a p a h This refers to the normalized attention weights. Ultimately, the weighted fusion feature z=a t *ft +a p *f p +a h *f h It is fed into the fully connected layer for final prediction; bias vector b a The initial values ​​can be initialized using a set of parameters obtained from OOA optimization, thereby introducing domain prior knowledge into network training.

[0046] Example 3 This embodiment enhances the output information of the model based on embodiment 1 or 2, enabling it to assess the credibility of the decoupling results.

[0047] Modify the output layer of a CNN or AM-CNN model to output the predicted value of each physical quantity simultaneously. ) and its prediction uncertainty log variance (s=log()); that is, the final output is: [ , , , , , ], where is the standard deviation, representing uncertainty.

[0048] To train this model, a loss function based on maximum likelihood estimation is used, which can simultaneously optimize prediction accuracy and uncertainty calibration: in, , Let N be the estimated uncertainty of the prediction for the m-th physical quantity, and N be the sample size. and These are the actual value and the predicted value of the m-th physical quantity in the k-th sample, respectively.

[0049] After training, the model provides decoupling values ​​for each input. It can also provide an estimate of its uncertainty, providing an important reliability reference for subsequent equipment condition assessment decisions.

[0050] Example 4 This embodiment provides a current transformer multi-physical quantity signal decoupling system that implements the above method. The system can be deployed in the local monitoring unit of a substation or the server of a remote control center. It can receive raw sensing signals, automatically run intelligent decoupling algorithms, and output clean physical quantity data that can be used for condition assessment.

[0051] The decoupling system includes a signal acquisition module, a model building and optimization module, and a signal decoupling module.

[0052] The signal acquisition module serves as the interface between the system and the physical world, responsible for collecting and receiving raw data. This module can be directly connected to the fiber Bragg grating demodulator via an industrial communication interface (such as Ethernet, RS-485, or fiber optic interface). The demodulator monitors the reflection spectrum of the FBG sensor embedded inside the current transformer in real time and converts the spectral shift into a preliminary, undecoupled composite sensing digital signal sequence. This sequence simultaneously contains mixed information caused by changes in temperature, pressure, and hydrogen concentration. The core function of the signal acquisition module is data access and buffering; it continuously monitors the data stream from the demodulator, receiving and buffering the data at a preset sampling period in a signal S. raw (t), where t represents the time series, and its output is a regular, timestamped raw composite signal data packet, ready to be sent to the next module for processing.

[0053] The model building and optimization module integrates the core algorithms of this system and is responsible for training, optimizing, and preparing the decoupled model. This module relies on high-performance processors (such as CPUs or GPU-accelerated computing cards) and memory (RAM, solid-state drives). The memory stores the system software and algorithm programs, including the basic architecture code of the CNN model, the implementation code of the OOA optimization algorithm, and the historical training dataset. Its final output is an executable, optimized OOA-CNN decoupled model instance, which resides in memory and is in a state of waiting for inference.

[0054] The signal decoupling module uses an optimized model to perform online decoupling calculations on real-time or historical data; this module receives the real-time signal S from the preprocessing submodule. raw (t), or input the stored historical data batches into the loaded OOA-CNN model instance; the model processes the input signal in a forward propagation manner, and finally generates three independent regression values ​​in the output layer. The module outputs the final decoupling result vector Y(t)=[T(t), P(t), H2(t)], where: T(t) is the decoupled temperature signal (unit: ℃), P(t) is the decoupled pressure signal (unit: kPa), and H2(t) is the decoupled hydrogen concentration signal (unit: L / L).

[0055] The three core modules mentioned above work collaboratively within the system. In addition, the system may include a necessary human-machine interface and external communication unit. The human-machine interface provides a graphical interface for configuring system parameters (such as sampling rate and alarm thresholds), visually displaying real-time curves and historical trends of decoupled temperature, pressure, and hydrogen concentration, and can overlay uncertainty intervals. The external communication unit serves as a bridge between the system and the higher-level management system. Following standard power system communication protocols (such as IEC 61850), it encapsulates the decoupled standard structured data into GOOSE or SV messages, or uploads them to the substation monitoring system or equipment condition assessment master station via the IEC 104 protocol, enabling higher-level fault early warning, life assessment, and operation and maintenance decisions.

[0056] Finally, it should be noted that this article uses specific examples to illustrate the principles and implementation methods of the present invention. The above description of the embodiments is only for the purpose of helping to understand the core ideas of the present invention. Without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A method for decoupling multiple physical quantity signals in a current transformer based on optimized CNN, characterized in that, Includes the following steps: Acquire the temperature-pressure-hydrogen concentration composite sensing signal collected by the fiber optic grating sensor inside the current transformer; Construct a convolutional neural network model for signal decoupling; The hyperparameters of the convolutional neural network model are optimized using the Osprey optimization algorithm to obtain the optimal combination of hyperparameters; The convolutional neural network model is configured and trained using the optimal hyperparameter combination to obtain a trained decoupling model. The composite sensing signal to be decoupled is input into the trained decoupling model, which outputs the decoupled independent temperature signal, independent pressure signal, and independent hydrogen concentration signal.

2. The method according to claim 1, characterized in that, The optimization of the hyperparameters of the convolutional neural network model using the Osprey optimization algorithm specifically involves: The learning rate and kernel size of the convolutional neural network model are used as optimization variables. The Osprey optimization algorithm is used to iteratively search within the set search space to minimize the model prediction error and determine the optimal learning rate and kernel size.

3. The method according to claim 2, characterized in that, The Osprey optimization algorithm iteratively searches within a defined search space, specifically including the following steps: S31. Initialize the osprey population and use the mean squared error of the convolutional neural network model on the validation set as the fitness function. S32. For the i-th osprey, determine the set of fish locations it can choose based on its fitness value. The set includes the positions of individuals in the population with better fitness than the current osprey and the globally optimal position. ; S33. Perform a global exploration phase location update: based on the formula Calculate candidate positions And compare it with the current position, updating the position according to the better one; where x i，j Let r be the j-th dimension component at the current position. i，j A random number in the range (0, 1). From The j-th dimension component of the randomly selected fish position, I i，j It is a random integer in {1, 2}; S34. Perform location update for the partial mining phase: according to the formula Calculate candidate positions It compares the current position with the current position and updates the position of the better one; where t is the current iteration number, and lb is the position of the current iteration number. j and ub j Let be the lower and upper bounds for the search of the j-th dimension parameter; S35. Repeat steps S32 to S34 until the preset number of iterations is met, and output the hyperparameter combination corresponding to the global optimal position as the optimal hyperparameter combination.

4. The method according to claim 3, characterized in that, In step S31, the osprey position vector is a hybrid encoding vector, including a continuous parameter part and a discrete parameter part; wherein, the continuous parameter part includes at least one of learning rate, regularization coefficient, and dropout rate; and the discrete parameter part includes at least one of number of convolutional layers, number of convolutional kernels, and convolutional kernel size. In steps S33 and S34, the dimension values ​​corresponding to the discrete parameter part are rounded and boundary constraints are applied.

5. The method according to claim 4, characterized in that, The convolutional neural network model is a multi-branch attention network structure, including: A shared feature extraction layer is used to perform preliminary feature extraction on the input composite sensing signal; Multiple feature decoupling branches are set up in parallel to extract features that are sensitive to different physical quantities, including temperature, pressure and hydrogen concentration. An attention fusion layer is used to perform weighted fusion of the output features of the multiple feature decoupled branches, wherein the fusion weights are generated through a learnable attention mechanism.

6. The method according to claim 5, characterized in that, The discrete parameter section also includes weight bias parameters for initializing the attention fusion layer; the fusion weight 'a' of the attention fusion layer is determined by the formula... Calculate, where, It is a learnable weight matrix, and h is the concatenated vector of output features from each branch. The initial value is set by the weight bias parameter obtained through optimization.

7. The method according to claim 1, characterized in that, The output of the convolutional neural network model also includes prediction uncertainty estimates corresponding to the decoupling values ​​of each physical quantity. The loss function L used to train the convolutional neural network model is: in, , Let N be the estimated uncertainty of the prediction for the m-th physical quantity, and N be the sample size. and These are the actual value and the predicted value of the m-th physical quantity in the k-th sample, respectively.

8. The method according to claim 3, characterized in that, In step S31, the fitness function is the weighted mean square error of the convolutional neural network model's predictions of temperature, pressure, and hydrogen concentration on the validation set.

9. A multi-physical quantity signal decoupling system for a current transformer based on optimized CNN, characterized in that, include: The signal acquisition module is used to acquire the temperature-pressure-hydrogen concentration composite sensing signal collected by the fiber optic grating sensor inside the current transformer. The model building and optimization module is used to execute the method of any one of claims 1 to 8 to build and train the optimal decoupled model; The signal decoupling module is used to decouple the input composite sensing signal using the decoupling model and output independent temperature, pressure and hydrogen concentration signals.

10. The system according to claim 9, characterized in that, The system also includes: The human-machine interface provides a graphical interface for configuring system parameters and visually displaying real-time curves and historical trends of decoupled temperature, pressure, and hydrogen concentration; as well as... External communication unit, used for external communication.

Images