Method and system for monitoring safety state of lead-acid battery of backup power supply of transformer substation

Abstract

The invention discloses a transformer substation backup power supply lead-acid battery safety state monitoring method and system, belongs to the technical field of transformer substation operation and maintenance, and solves the problems of insufficient sample size, inaccurate feature extraction and low processing efficiency in battery safety state monitoring in the prior art. According to the method, the generative adversarial network algorithm based on robust estimation is adopted to perform data expansion, the neural network is updated by adopting the adaptive fluctuation optimization algorithm to perform feature extraction, and the extracted features are subjected to dimensionality reduction by utilizing the self-encoding neural network based on manifold mapping learning; classifying the dimensionality-reduced data through a support vector machine classification algorithm based on Landau quantization, and using a classification result for monitoring the safety state of the battery; the problems that traditional data acquisition is difficult and samples are insufficient are solved, training errors and model volatility are reduced, data dimensions are reduced, processing efficiency and operation loads are optimized, and classification precision and stability are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of substation operation and maintenance, and relates to a method and system for monitoring the safety status of a lead-acid battery of a backup power supply of a substation. Background Art

[0002] As an important component of the backup power supply of substations, the long-term safety of lead-acid batteries has become an important issue. Lead-acid batteries may have various problems such as rapid aging during use. If these problems are not discovered and handled in time, they may lead to battery performance degradation and even safety accidents. Therefore, it is crucial to ensure the safe operation of lead-acid batteries for substation backup power under various working conditions.

[0003] Traditional battery monitoring systems rely on simple sensor data collection and threshold judgment, which is often inadequate when dealing with complex or hidden battery failures. In addition, existing technologies have problems in data collection and processing, such as insufficient sample size, inaccurate feature extraction, and low processing efficiency, which limit the effectiveness and reliability of the system in practical applications. Summary of the invention

[0004] The technical solution of the present invention is used to solve the problems of insufficient sample size, inaccurate feature extraction and low processing efficiency in the battery safety status monitoring technology in the prior art.

[0005] The present invention solves the above technical problems through the following technical solutions:

[0006] A method for monitoring the safety status of a lead-acid battery for a backup power supply of a transformer substation comprises the following steps:

[0007] Step 1, collecting battery electrochemical impedance spectroscopy data and storing the collected data;

[0008] Step 2: Label the collected data. The labeling categories include capacity and battery status. The battery status includes normal operation, battery aging, and prohibited use.

[0009] Step 3: Use a generative adversarial network algorithm based on robust estimation to perform data expansion to generate new training samples;

[0010] Step 4: Adopting the adaptive wave optimization algorithm to update the weights and biases in the neural network, simulating the propagation and energy distribution characteristics of the wave, and extracting features from the battery electrochemical impedance spectrum data;

[0011] Step 5: Use the autoencoder neural network based on manifold mapping learning to reduce the dimension of the extracted features;

[0012] Step 6: Classify the reduced-dimensional data using a support vector machine classification algorithm based on Landau quantization to determine the safety status of the battery, and use the classification results to monitor the safety status of the battery.

[0013] Furthermore, the method for collecting the battery electrochemical impedance spectroscopy data described in step 1 and storing the collected data is as follows: using an electrochemical impedance spectroscopy acquisition chip to collect solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W from the battery system, the collected data is stored in JSON format.

[0014] Furthermore, the method for annotating the collected data described in step 2 is as follows: the battery capacity is determined by constant current charging and discharging the battery, and different capacity stages are annotated from 100% to 70% capacity, and a set of capacity and solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W data sets are obtained every 1% to 2% for manual annotation.

[0015] Furthermore, the method for data expansion using the robust estimation-based generative adversarial network algorithm described in step 3 is as follows:

[0016] (1) Initialize the network parameters of the generator and discriminator. The initialization is randomly assigned according to the Gaussian distribution. For the generator initialization, it is expressed as:

[0017]

[0018] And, for discriminator initialization, it is expressed as:

[0019]

[0020] In the formula, θ g is the parameter of the generator, σ g is the standard deviation of the generator parameters, θ d is the parameter of the discriminator, σ d is the standard deviation of the discriminator parameters;

[0021] (2) Perform iterative training. The generator G receives random noise z from the latent space. The random noise z follows a standard normal distribution. The generator maps the noise to the data space through learning, imitating the distribution of battery monitoring data, which can be expressed as:

[0022] x gen =G(z;θ g )

[0023]

[0024] In the formula, x genis the generated data, z is the input random noise vector, and I is the identity matrix;

[0025] (3) The generated data and the actual collected battery monitoring data are input into the discriminator. The discriminator evaluates the authenticity of these data and updates its parameters through backpropagation. The training goal of the discriminator D is to maximize the distinction between the real data x and the generated data x gen The ability is achieved by minimizing the following loss function:

[0026]

[0027] The expected calculation formula for real data is:

[0028]

[0029] And, the expected calculation formula for generating data is:

[0030]

[0031] In the formula, p data is the distribution of the real data, p z is the distribution of noise;

[0032] (4) Based on the feedback from the discriminator, the generator updates its parameters to generate data that is more difficult for the discriminator to recognize. The goal of the generator G is to make the discriminator D unable to distinguish between real data and generated data, which is achieved by minimizing the following loss function:

[0033]

[0034] Where, L QGAN is the uncertainty measurement loss term;

[0035] For the uncertainty evaluation of the data G(z) generated by the generator, the calculation formula is:

[0036] Sr(G(z))=-Tr(G(z)×logG(z))

[0037] Where Sr() is the uncertainty assessment function, Tr() represents the trace operation of the matrix;

[0038] (5) Uncertainty measurement loss term L QGAN The calculation formula is:

[0039]

[0040] In the formula, γ ed is the uncertainty measure loss coefficient;

[0041] (6) In each iteration, the quality of the generated data in the current training batch is evaluated, and the learning rates of the generator and discriminator are dynamically adjusted. The adjustment method is expressed as:

[0042]

[0043] In the formula, and are the new and old learning rates respectively; γ vh is the learning rate adjustment factor; L previous and L current are the previous and current loss function values ​​respectively; ΔL is the estimated amount of loss function change;

[0044] (7) Repeat the above steps until a preset stop iteration condition is met, where the preset stop iteration condition is reaching a preset maximum number of iterations;

[0045] (8) Use similarity metrics to evaluate the quality of generated data and adjust the training strategy accordingly, expressed as:

[0046] Q(x f )→update(Q,θ g ′,θ d ′)

[0047] Where Q() is the quality assessment function; x f is to generate data; update() is a function to adjust the training strategy based on quality evaluation;

[0048] When the evaluation value of the quality evaluation function is greater than the preset threshold, the data is retained; otherwise, the generated data is discarded.

[0049] Furthermore, the quality assessment function Q() is implemented by calculating the similarity between the generated data and the data in the real data set, which is expressed as:

[0050]

[0051] Among them, sim() represents the similarity measurement function, and the measurement method is implemented based on the Euclidean distance calculation method; represents the i-th sample in the real data set, and N is the total number of samples.

[0052] Furthermore, the method described in step 4 of using the adaptive fluctuation optimization algorithm to update the weights and biases in the neural network, simulating the propagation and energy distribution characteristics of fluctuations, and extracting features from the battery electrochemical impedance spectrum data is as follows:

[0053] (1) Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is the ReLU activation function.

[0054] (2) Set the initial parameters of the neural network, including the weight w and bias b of the neural network, and initialize the wave source S i The position x i , intensity a i and the initial phase φ i , assuming the total number of wave sources is n, the initialization is expressed as:

[0055] θ0={w0,b0}

[0056] S i (x i ,a i ,φ i ),i=1,2,…,n

[0057] In the formula, x i Usually chosen as the initial position of the parameter, a i and φ i Randomly set its initial value; S i is the i-th wave source, with position x i , intensity a i and the initial phase φ i composition;

[0058] (3) Calculate the loss L through forward propagation. Perform forward propagation on the input X using the parameter θ, calculate the output Y and compare it with the target value T. The calculation formula is:

[0059] Y = frs(X;θ)

[0060] L=Loss(Y,T)

[0061] In the formula, frs() represents the network structure, and the output layer of the network structure uses the preset Softmax classifier to obtain the predicted label; Loss(Y,T) is the cross entropy loss;

[0062] (4) Calculate the gradient of each parameter based on the loss function The calculation formula is:

[0063]

[0064] (5) The chain method is used to calculate the gradient. Assuming that the network output Y passes through the Softmax function, the gradient calculation formula for the weight w is:

[0065]

[0066] In the formula, yk is the output of Softmax, z k It is the linear output before the Softmax layer;

[0067] (6) The calculation formula is:

[0068]

[0069] and, The calculation formula is:

[0070]

[0071] and, The calculation formula is:

[0072]

[0073] (7) According to the direction and magnitude of the gradient, a wave source is generated at the corresponding parameter position. Each wave source S i The strength is updated to The calculation formula is:

[0074]

[0075] In the formula, α ck is the learning rate, β bp is the attenuation coefficient, γ bp is the amplification factor;

[0076] (8) The fluctuations of all wave sources are superimposed in the parameter space. The update amount of each parameter is determined by the superposition result of the fluctuations of all wave sources at that point. The parameter update method is expressed as:

[0077]

[0078] Where k is the wave number, ω is the angular frequency, which indicates the propagation speed and direction of the wave; Re() is the ReLU activation function, τ i,l is the fluctuation threshold;

[0079] (9) Fluctuation threshold τ of each wave source i,l The layer sensitivity index is adjusted dynamically, and the adjustment method is expressed as:

[0080]

[0081] Where τ0 is the initial fluctuation threshold, S l is the layer sensitivity index, κ pq is the adjustment factor;

[0082] (10) Layer sensitivity index S of the lth layer lThe norm of the gradient output by this layer is calculated to measure the contribution of this layer to the overall network error. The calculation formula is:

[0083]

[0084] In the formula, is the norm of the parameter gradient of the lth layer, Lc is the number of network layers;

[0085] (11) According to the loss reduction of this iteration, the attenuation coefficient β and amplification coefficient γ of each wave source are dynamically adjusted. The calculation formula is:

[0086]

[0087] In the formula, are the attenuation coefficient and amplification coefficient of the previous iteration, λ bp ,μ bp are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, respectively, and ΔL is the change in loss;

[0088] (12) Repeat the above steps until the preset stop iteration condition is met.

[0089] Furthermore, the method described in step 5 for reducing the dimension of the extracted features using the autoencoder neural network based on manifold mapping learning is as follows:

[0090] (1) Initialize the network parameters. Use Gaussian distribution to initialize the weight w and bias b parameters of the autoencoder, expressed as:

[0091]

[0092] b i =0

[0093] In the formula, w i represents the weight of the i-th layer, σ ch represents standard deviation;

[0094] (2) Through the autoencoder, the data x is mapped to the low-dimensional feature space z as follows:

[0095] z=Sig(W e x+b e )

[0096] And, the decoder attempts to reconstruct the original input The way is expressed as:

[0097]

[0098] Where Sig() is the activation function; W e and b eare the weight and bias of the encoder respectively; W d and b d are the weights and biases of the decoder;

[0099] (3) In the encoded low-dimensional space, the Riemann factor R is used to implement manifold learning. The calculation formula of the feature z of the Riemann factor constraint encoding is:

[0100]

[0101] In the formula, μ cy represents the preset manifold center, σ R is a parameter that controls the compactness of the manifold;

[0102] (4) The loss function not only considers the reconstruction error, but also the Riemann manifold constraint. The reconstruction error ensures that the data can be effectively restored. The manifold constraint strengthens the geometric structure of the feature space. The calculation formula of the loss function is:

[0103]

[0104] In the formula, λ cm is the regularization coefficient;

[0105] (5) Reconstruction error The calculation formula is:

[0106]

[0107] In the formula, z i is the value of z in the i-th dimension, μ i is the value of the preset manifold center in the i-th dimension;

[0108] (6) Based on the loss function, adjust the network parameters through the back propagation algorithm;

[0109] (7) Gradient It is calculated by the error back propagation method, and the calculation formula is:

[0110]

[0111] and, Through the decoder weight W d The transposed representation of is calculated as follows:

[0112]

[0113] (8) Since z is W e The result obtained by the activation function after the linear combination of and x is The calculation formula is:

[0114]

[0115] In the formula, Sig ′ () is the derivative of the Sigmoid activation function, diag(v) represents a diagonal matrix with vector v as the diagonal element, x T is the transpose of x;

[0116] (9) Use the gradient descent method to update the parameters. The update method is expressed as:

[0117]

[0118] In the formula, α rv is the learning rate, t represents the number of iterations, is the updated weight parameter, is the weight parameter before updating;

[0119] (10) Repeat the above steps until the preset stop iteration condition is met.

[0120] Furthermore, the weight W e The calculation formula is pre-set by principal component analysis:

[0121]

[0122] Where V k is the covariance matrix X T The matrix consisting of the first k eigenvectors obtained from X, Σ k is the corresponding diagonal matrix of eigenvalues.

[0123] Furthermore, the method for classifying the reduced-dimensional data by the support vector machine classification algorithm based on Landau quantization described in step 6 is specifically as follows:

[0124] (1) Initialize the quantum state of the training data. Suppose the initial i-th sample data vector is And q ij Represents the jth feature of the i-th sample, converted into the quantum state |ψ i The calculation formula of > is:

[0125]

[0126] In the formula, α ij is the complex amplitude, satisfying |q ij > represents the characteristic q ij The corresponding quantum ground state;

[0127] (2) The initial quantized hyperplane is constructed by the Landau quantization principle. The probability amplitude of the quantum state is used to evaluate its relative position with the hyperplane to determine the initial support vector. The quantum hyperplane |w> is defined as:

[0128]

[0129] In the formula, γ k is the weight of the hyperplane vector, |v k > is the hyperplane vector;

[0130] (3) The calculation formula for the decision boundary between the quantum state and the hyperplane is:

[0131]

[0132] In the formula, <ψ i |w> is the inner product of the quantum state and the hyperplane, f() is the decision boundary function, ij |v k > is the inner product of the feature and the hyperplane;

[0133] (4) Weight γ of the quantum hyperplane k The calculation formula is determined by principal component analysis:

[0134]

[0135] In the formula, e k is the kth eigenvalue of the data covariance matrix, e l is the lth eigenvalue of the data covariance matrix;

[0136] (5) In each iteration, the distance and angle between each support vector and other data points are checked, and fission operations are performed according to the clustering density and edge distance to refine the decision boundary; for each support vector, the fission process is achieved by adjusting its quantum state, which can be expressed as:

[0137] |ψ i ′ >=|ψ i >+β bp ∑ m≠i δ im |ψ m >

[0138] In the formula, β bp is the fission factor, δ im is the inverse coefficient of the distance between i and m;

[0139] (6) Distance inverse ratio coefficient δ im Dynamically adjust according to the distance between the ith data point and the mth data point. The calculation formula is: ​

[0140]

[0141] in, Represents data points and data points The square of the Euclidean distance between ∈ is a constant;

[0142] (7) For data groups that exhibit similar characteristics, a fusion operation is performed to merge the representations of these data points to enhance the model's response to these characteristics. The quantum state merging operation of similar data points is expressed as:

[0143]

[0144] Where M i is a set of quantum states similar to , σ m is the weight assigned according to similarity;

[0145] (8) Weight σ m It is calculated based on the similarity and the local density of the data points and is expressed as:

[0146]

[0147] Where, d mM It is the data point m to the set M i The average distance to other points in

[0148] (9) Adjust the state of the qubit through quantum gate operations to optimize the accuracy of the classification results, expressed as:

[0149] U(θ)|ψ i ″>=cos(θ)|ψ i ″>+sin(θ)∑ m≠i |ψ m >

[0150] Where θ is the rotation angle of the quantum gate;

[0151] (10) Repeat the above steps until the preset stop iteration condition is met.

[0152] Furthermore, the quantum gate operation U(θ) is used to adjust the state of the quantum bit, and the calculation method is:

[0153] U(θ)=e -iθH

[0154] Where H is a specific Hamiltonian.

[0155] The advantages of the present invention are:

[0156] (1) The insufficient training data is expanded by using a generative adversarial network based on robust estimation, which can effectively deal with outliers and noise in monitoring data. The data expanded by the generative adversarial network approximates the real data distribution, enhances the diversity of the data set and the effectiveness of model training, and solves the problems of difficulty in traditional data collection and insufficient samples.

[0157] (2) Adopting adaptive fluctuation optimization algorithm to update weights and biases in neural network, simulating the propagation and energy distribution characteristics of fluctuations, effectively extracting the complex fluctuation and nonlinear characteristics of battery electrochemical impedance spectroscopy monitoring data, and accurately extracting key features in complex data background, thus reducing training errors and model fluctuations;

[0158] (3) Utilizing an autoencoder neural network based on manifold mapping learning, the representation of the feature space is optimized through Riemann factors, thereby improving the intrinsic structure learning ability and accurate reconstruction ability of the data; the autoencoder structure reduces the dimension of the data while maintaining key information during feature dimensionality reduction, thereby optimizing processing efficiency and computational load.

[0159] (3) The Landau quantization-based support vector machine algorithm is adopted to adjust the support vector and classification hyperplane through quantum state transformation and fission and fusion concepts, refine the decision boundary, and enhance the model's generalization ability for new data; the Landau quantization support vector machine improves the classification accuracy and stability through precise quantum state adjustment and data feature merging. BRIEF DESCRIPTION OF THE DRAWINGS

[0160] Figure 1 It is a flow chart of a method for monitoring the safety status of a lead-acid battery of a backup power supply of a substation according to a first embodiment of the present invention;

[0161] Figure 2 This is a feature extraction training flow chart of a method for monitoring the safety status of a lead-acid battery for backup power supply in a substation according to the first embodiment of the present invention;

[0162] Figure 3 It is a feature dimension reduction training flow chart of the safety status monitoring method of the backup power supply lead-acid battery of the substation according to the first embodiment of the present invention;

[0163] Figure 4 The data diagram of the battery electrochemical impedance spectrum under different battery capacities of the safety status monitoring method of the lead-acid battery of the backup power supply of the substation according to the first embodiment of the present invention;

[0164] Figure 5 This is a structural block diagram of a safety status monitoring system for a lead-acid battery of a backup power supply for a substation according to a second embodiment of the present invention. DETAILED DESCRIPTION

[0165] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0166] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0167] Embodiment 1

[0168] like Figure 1 As shown, a method for monitoring the safety status of a lead-acid battery of a backup power supply for a substation according to an embodiment of the present invention comprises the following steps:

[0169] 1. Data Collection

[0170] An electrochemical impedance spectroscopy acquisition chip is used to collect data from the battery system. The electrochemical impedance spectroscopy acquisition chip collects the real impedance and imaginary impedance of the lead-acid battery of the substation backup power supply to obtain but not limited to the solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W. The collected data is stored in JSON format.

[0171] In this embodiment, an example of storing a piece of data in JSON format is as follows:

[0172] {"atime":"2024-06-21T12:00:00Z","aRs":0.035,"aRct":0.045,"aCdl":0.42,"aW":0.17}; The fields and attribute values ​​of this data are: aRs: 0.035mΩ (solution impedance); aRct: 0.045mΩ (charge transfer impedance); aCdl: 0.42mH (double-layer capacitance); aW: 0.17mΩ (Warburg impedance).

[0173] It should be noted that this embodiment only describes one data format and type. In actual applications, the attributes of data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0174] 2. Data annotation

[0175] The battery capacity is determined by charging and discharging the battery at a constant current, and different capacity stages are marked from 100% to 70%. A set of data sets of capacity and solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W are obtained every 1% for manual annotation. The annotation categories include: capacity (in percentage) and battery status. The battery status includes normal operation, battery aging and prohibited use, a total of 3 categories.

[0176] 3. Data expansion

[0177] The collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect the accuracy of the model. This embodiment uses a generative adversarial network algorithm based on robust estimation to perform data expansion to generate new training samples; the generative adversarial network algorithm based on robust estimation includes two parts: a generator (G) and a discriminator (D). The task of the generator is to generate new data that is as close to the real data as possible, and the task of the discriminator is to distinguish the difference between the generated data and the real data. In order to improve the quality of the generated data, this embodiment adopts a strategy based on robust estimation in the loss function of the generator and the discriminator, which helps the algorithm to better handle outliers and noise data in the battery monitoring data.

[0178] Specifically, the training process of the generative adversarial network algorithm based on robust estimation is as follows:

[0179] (1) Initialize the network parameters of the generator and discriminator. The initialization is randomly assigned according to the Gaussian distribution. For the generator initialization, it is expressed as:

[0180]

[0181] And, for discriminator initialization, it is expressed as:

[0182]

[0183] In the formula, θ g is the parameter of the generator, σ g is the standard deviation of the generator parameters, θ d is the parameter of the discriminator, σ d is the standard deviation of the discriminator parameters.

[0184] (2) Perform iterative training. The generator G receives random noise z from the latent space. The random noise z follows a standard normal distribution. The generator maps the noise to the data space through learning, imitating the distribution of battery monitoring data, which can be expressed as:

[0185] x gen =G(z;θ g )

[0186]

[0187] In the formula, x gen is the generated data, z is the input random noise vector, and I is the identity matrix;

[0188] (3) The generated data and the actual collected battery monitoring data are input into the discriminator. The discriminator evaluates the authenticity of these data and updates its parameters through backpropagation. The training goal of the discriminator D is to maximize the distinction between the real data x and the generated data x gen The ability is achieved by minimizing the following loss function:

[0189]

[0190] Among them, the expected calculation method of real data is expressed as:

[0191]

[0192] And, the expected calculation method for generating data is expressed as:

[0193]

[0194] In the formula, p data is the distribution of the real data, p z is the distribution of noise.

[0195] (4) Based on the feedback from the discriminator, the generator updates its parameters to generate data that is more difficult for the discriminator to recognize. The goal of the generator G is to make the discriminator D unable to distinguish between real data and generated data, which is achieved by minimizing the following loss function:

[0196]

[0197] Where, L QGAN is the uncertainty measure loss term.

[0198] For the uncertainty evaluation of the data G(z) generated by the generator, the calculation method is expressed as:

[0199] Sr(G(z))=-Tr(G(z)×logG(z))

[0200] Where Sr() is the uncertainty assessment function, and Tr() represents the trace operation of the matrix.

[0201] (5) Uncertainty measurement loss term L QGAN The calculation method is expressed as:

[0202]

[0203] In the formula, γed is the uncertainty measure loss coefficient. Preferably, γ ed Set to 3.

[0204] (6) In each iteration, the quality of the generated data in the current training batch is evaluated, and the learning rates of the generator and discriminator are dynamically adjusted. The adjustment method is expressed as:

[0205]

[0206] In the formula, and are the new and old learning rates respectively; γ vh is the learning rate adjustment factor; L previous and L current are the previous and current loss function values ​​respectively; ΔL is the estimated loss function change. Preferably, γ vh Set to 0.3.

[0207] (7) Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0208] (8) Use similarity metrics to evaluate the quality of generated data and adjust the training strategy accordingly, which can be expressed as:

[0209] Q(x f )→update(Q,θ g ′,θ d ′)

[0210] Where Q() is the quality assessment function; x f is used to generate data; update() is a function that adjusts the training strategy based on quality evaluation.

[0211] The quality assessment function Q() is realized by calculating the similarity between the generated data and the data in the real data set, which is expressed as:

[0212]

[0213] Among them, sim() represents the similarity measurement function, and the measurement method is implemented based on the Euclidean distance calculation method; represents the i-th sample in the real data set, and N is the total number of samples.

[0214] When the evaluation value of the quality evaluation function is greater than a preset threshold, the data is retained; otherwise, the generated data is discarded.

[0215] After the data expansion model training is completed, the trained data expansion model is used to increase the number of samples. In one embodiment, assuming that the original collected samples are 800, and the data expansion model expands and generates 200 samples, the expanded data set contains 1000 samples.

[0216] 4. Feature extraction

[0217] The expanded data is input into the feature extraction model for training. This embodiment adopts a neural network feature extraction algorithm based on adaptive fluctuation optimization. Inspired by the fluctuation phenomenon, the fluctuation presents the characteristics of periodic change and energy transfer, and can transmit and transform energy forms in different media. The adaptive fluctuation optimization algorithm iteratively updates the weights and biases in the network by simulating the propagation and energy distribution characteristics of the fluctuation. Each parameter update is not only affected by the current gradient, but also by the cumulative influence of the fluctuation of the previous parameter update, which is similar to the fluctuation influence caused by the superposition of wave sources, making the parameter update process smoother and more adaptive, effectively avoiding the oscillation and local minimum problems in the traditional gradient descent method.

[0218] Specifically, Figure 2 As shown in the figure, the training process of the neural network feature extraction algorithm based on adaptive fluctuation optimization is as follows:

[0219] (1) Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is the ReLU activation function.

[0220] (2) Let θ be the parameters of the neural network, including weight w and bias b. Set the initial parameters θ0 of the neural network, including the weight w and bias b of the neural network, and initialize the wave source S i The position x i , intensity a i and the initial phase φ i , assuming the total number of wave sources is n, the initialization is expressed as:

[0221] θ0={w0,b0}

[0222] S i (x i ,a i ,φ i ),i=1,2,…,n

[0223] In the formula, x i Usually chosen as the initial position of the parameter, a i and φ i Randomly set its initial value; S i is the i-th wave source, with position x i , intensity ai and the initial phase φ i composition.

[0224] (3) Calculate the loss L through forward propagation. Perform forward propagation on the input X using the parameter θ, calculate the output Y and compare it with the target value T. The calculation formula is:

[0225] Y = frs(X;θ)

[0226] L=Loss(Y,T)

[0227] In the formula, frs() represents the network structure, and the output layer of the network structure uses the preset Softmax classifier to obtain the predicted label; Loss(Y,T) is the cross entropy loss.

[0228] (4) Calculate the gradient of each parameter based on the loss function The calculation formula is:

[0229]

[0230] (5) The chain method is used to calculate the gradient. Assuming that the network output Y passes through the Softmax function, the gradient calculation formula for the weight w is:

[0231]

[0232] In the formula, y k is the output of Softmax, z k It is the linear output before the Softmax layer.

[0233] (6) The calculation formula is:

[0234]

[0235] and, The calculation formula is:

[0236]

[0237] and, The calculation formula is:

[0238]

[0239] (7) According to the direction and magnitude of the gradient, a wave source is generated at the corresponding parameter position. Each wave source S i The strength is updated to The calculation formula is:

[0240]

[0241] In the formula, αck is the learning rate, β bp is the attenuation coefficient, γ bp is the amplification factor. Preferably, α ck Set to 0.01.

[0242] (8) The fluctuations of all wave sources are superimposed in the parameter space. The update amount of each parameter is determined by the superposition result of the fluctuations of all wave sources at that point. The parameter update method is expressed as:

[0243]

[0244] Where k is the wave number, ω is the angular frequency, which indicates the propagation speed and direction of the wave; Re() is the ReLU activation function, τ i,l is the fluctuation threshold.

[0245] (9) Fluctuation threshold τ of each wave source i,l The layer sensitivity index is adjusted dynamically, and the adjustment method is expressed as:

[0246]

[0247] Where τ0 is the initial fluctuation threshold, S l is the layer sensitivity index, κ pq is the adjustment coefficient. Preferably, τ0 is set to 0.1, κ pq Set to 3.

[0248] (10) Layer sensitivity index S of the lth layer l The norm of the gradient output by this layer is calculated to measure the contribution of this layer to the overall network error. The calculation formula is:

[0249]

[0250] In the formula, is the norm of the parameter gradient of the lth layer, and Lc is the number of network layers.

[0251] (11) According to the loss reduction of this iteration, the attenuation coefficient β and amplification coefficient γ of each wave source are dynamically adjusted. The calculation formula is:

[0252]

[0253] In the formula, are the attenuation coefficient and amplification coefficient of the previous iteration, λ bp ,μ bp are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, respectively, ΔL is the change in loss; preferably, the parameter λ bp and μ bp Set to 0.1 and 0.05 respectively.

[0254] (12) Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0255] 5. Feature Dimensionality Reduction

[0256] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. This embodiment uses an autoencoder neural network based on manifold mapping learning to perform feature dimensionality reduction. The autoencoder neural network not only learns the intrinsic structure of the data in the encoding stage, but also attempts to reconstruct the input through the decoding process to ensure that the features after dimensionality reduction can still retain key information. The autoencoder neural network model structure consists of an encoder and a decoder. The encoder is responsible for mapping the high-dimensional input data to a low-dimensional feature space, and the decoder reconstructs the original input data from this low-dimensional space. The present invention uses Riemann factors based on manifold learning to optimize the representation of the feature space. By utilizing the geometric properties on the Riemann manifold, the local structure of the data can be better understood and more accurate data reconstruction can be achieved.

[0257] Specifically, Figure 3 As shown in the figure, the training process of the autoencoder neural network algorithm based on manifold mapping learning is as follows:

[0258] (1) Initialize the network parameters. Use Gaussian distribution to initialize the weight w and bias b parameters of the autoencoder, expressed as:

[0259]

[0260] b i =0

[0261] In the formula, w i represents the weight of the i-th layer, σ ch represents the standard deviation. Preferably, n in is the number of nodes in the input layer.

[0262] (2) During the encoding process, the input data is mapped to a low-dimensional feature space through the encoder; during the decoding process, the decoder attempts to reconstruct the original data. Specifically, through the autoencoder, the data x is mapped to the low-dimensional feature space z as follows:

[0263] z=Sig(W e x+b e )

[0264] And, the decoder attempts to reconstruct the original input can be expressed as:

[0265]

[0266] Where Sig() is the activation function; W e and b e are the weight and bias of the encoder respectively; W d and b d are the decoder weights and biases.

[0267] Weight W e The calculation formula is pre-set by principal component analysis:

[0268]

[0269] Where V k is the covariance matrix X T The matrix consisting of the first k eigenvectors obtained from X, Σ k is the corresponding diagonal matrix of eigenvalues.

[0270] (3) In the encoded low-dimensional space, the Riemann factor R is used to implement manifold learning. The calculation formula of the feature z of the Riemann factor constraint encoding is:

[0271]

[0272] In the formula, μ cy represents the preset manifold center, σ R is a parameter that controls the compactness of the manifold. Preferably, μ cy Set to the mean vector of the low-dimensional space, σ R Set to 0.1.

[0273] (4) The loss function not only considers the reconstruction error, but also the Riemann manifold constraint. The reconstruction error ensures that the data can be effectively restored. The manifold constraint strengthens the geometric structure of the feature space. The calculation formula of the loss function is:

[0274]

[0275] In the formula, λ cm is the regularization coefficient. Preferably, λ cm Set to 0.3.

[0276] (5) Reconstruction error The calculation formula is:

[0277]

[0278] In the formula, z i is the value of z in the i-th dimension, μ i is the value of the preset manifold center in the i-th dimension.

[0279] (6) Based on the loss function, the network parameters are adjusted through the back propagation algorithm to optimize the performance of the encoder and decoder. e For example, the calculation formula of its gradient is:

[0280]

[0281] (7) Gradient It is calculated by the error back propagation method, and the calculation formula is:

[0282]

[0283] and, Through the decoder weight W d The transposed representation of is calculated as follows:

[0284]

[0285] (8) Since z is W e The result obtained by the activation function after the linear combination of and x is The calculation formula is:

[0286]

[0287] In the formula, Sig ′ () is the derivative of the Sigmoid activation function, diag(v) represents a diagonal matrix with vector v as the diagonal element, x T is the transpose of x.

[0288] (9) Use the gradient descent method to update the parameters. The update method is expressed as:

[0289]

[0290] In the formula, α rv is the learning rate, t represents the number of iterations, is the updated weight parameter, is the weight parameter before updating. Preferably, α rv Set to 0.05.

[0291] (10) Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0292] 6. Data classification and battery safety status monitoring

[0293] The reduced data is input into the classifier to train the classifier model. This embodiment adopts a support vector machine classification algorithm based on Landau quantization. The traditional support vector machine maximizes the margin between different categories of data by finding a hyperplane. The present invention optimizes the selection of support vectors and the definition of classification hyperplanes through a quantum state conversion strategy, and further optimizes the classification process by adopting the concepts of fission and fusion. During the training process, the model dynamically adjusts the support vectors through a fission mechanism, that is, splits in the data space to form a more detailed decision boundary. At the same time, the fusion mechanism allows the model to merge similar features in multiple data set partitions to enhance the model's generalization ability for new data. This two-way dynamic adjustment mechanism provides the model with higher flexibility and accuracy.

[0294] Specifically, the training process of the support vector machine classification algorithm based on Landau quantization is as follows:

[0295] (1) Initialize the quantum state of the training data. Each sample is represented by a quantum bit, which has multiple possible states corresponding to different classification labels. Specifically, let the initial i-th sample data vector be And q ij Represents the jth feature of the i-th sample, converted into the quantum state |ψ i The calculation formula of > is:

[0296]

[0297] In the formula, α ij is the complex amplitude, satisfying |q ij > represents the characteristic q ij The corresponding quantum ground state.

[0298] (2) The initial quantized hyperplane is constructed by the Landau quantization principle, and the probability amplitude of the quantum state is used to evaluate its relative position with the hyperplane to determine the initial support vector. Specifically, the quantum hyperplane |w> is defined as:

[0299]

[0300] In the formula, γ k is the weight of the hyperplane vector, |v k > is the hyperplane vector.

[0301] (3) The calculation formula for the decision boundary between the quantum state and the hyperplane is:

[0302]

[0303] In the formula, <ψ i |w> is the inner product of the quantum state and the hyperplane, f() is the decision boundary function, <qij |v k > is the inner product of the feature and the hyperplane.

[0304] (4) Weight γ of the quantum hyperplane k The calculation formula is determined by principal component analysis:

[0305]

[0306] In the formula, e k is the kth eigenvalue of the data covariance matrix, e l is the lth eigenvalue of the data covariance matrix.

[0307] (5) In each iteration, the distance and angle between each support vector and other data points are checked, and fission operations are performed according to the cluster density and edge distance to refine the decision boundary. Specifically, for each support vector, the fission process is achieved by adjusting its quantum state, which can be expressed as:

[0308] |ψ i ′ >=|ψ i >+β bp ∑ m≠i δ im |ψ m >

[0309] In the formula, β bp is the fission factor, δ im is the inverse coefficient of the distance between i and m. Preferably, β bp Set to 3.

[0310] (6) Distance inverse ratio coefficient δ im Dynamically adjust according to the distance between the ith data point and the mth data point. The calculation formula is:

[0311]

[0312] here, Represents data points and data points The square of the Euclidean distance between ∈ is a small constant. Preferably, ∈ is set to 0.0001.

[0313] (7) For data groups that exhibit similar characteristics, a fusion operation is performed to merge the representations of these data points to enhance the model's response to these characteristics. The quantum state merging operation of similar data points is expressed as:

[0314]

[0315] Where M i is a set of quantum states similar to , σm is the weight assigned according to the similarity.

[0316] (8) Weight σ m It is calculated based on the similarity and the local density of the data points and is expressed as:

[0317]

[0318] Where, d mM It is the data point m to the set M i The average distance to other points in .

[0319] (9) Adjust the state of the qubit through quantum gate operations to optimize the accuracy of the classification results, expressed as:

[0320] U(θ)|ψ i ″)=cos(θ)|ψ i ″>+sin(θ)∑ m≠i |ψ m >

[0321] Where θ is the rotation angle of the quantum gate.

[0322] The quantum gate operation U(θ) is used to adjust the state of the qubit and is calculated as:

[0323] U(θ)=e -iθH

[0324] Wherein, H is a specific Hamiltonian. Preferably, a Pauli-Z gate is used to realize phase rotation in quantum computing, and θ controls the angle of rotation.

[0325] (10) Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0326] The trained model is used to process new samples to monitor the battery safety status. In one embodiment, the input data of the new sample is first fed into the trained feature extraction model to obtain a first feature vector; further, the first feature vector is fed into the trained feature dimension reduction model to obtain a second feature vector; further, the second feature vector is fed into the trained classifier model to obtain classification categories, namely, battery capacity and battery safety status. In this embodiment, the capacity (in percentage) and battery status include normal operation, battery aging, and prohibited use, a total of 3 categories.

[0327] Test verification

[0328] like Figure 4As shown, the battery capacity is determined by constant current charge and discharge. According to the aging process of the battery, the capacity detection and electrochemical impedance spectrum detection are performed every 1% to 2% from 100% until the battery capacity reaches 70% of the rated capacity, and the battery electrochemical impedance spectrum data under different battery capacities are obtained. The impedance spectra of different capacities are analyzed to obtain the solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W data of batteries with different capacities. Through data annotation, cleaning, expansion and feature extraction, the correlation between the battery aging state and the characteristic value of the impedance spectrum is clarified, the electrochemical impedance spectrum of the battery can be input, and the aging state of the battery can be obtained quickly and accurately, which improves the accuracy of the state monitoring of the lead-acid battery on site and further improves the operation safety level of the power grid.

[0329] Embodiment 2

[0330] like Figure 5 As shown, a substation backup power supply lead-acid battery safety status monitoring system according to an embodiment of the present invention includes: a data test analysis module, a data annotation expansion module, a feature extraction calculation module, and a result output module;

[0331] The data test analysis module is used to analyze the impedance spectrum to obtain the lead-acid battery solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W;

[0332] The data annotation expansion module is used to perform data expansion based on a generative adversarial network algorithm based on robust estimation to generate new training samples;

[0333] The feature extraction calculation module is used to perform dimensionality reduction calculation on the extracted features based on the autoencoder neural network learned by manifold mapping;

[0334] The result output module is used to output the classification result and to monitor the battery safety status.

[0335] Embodiment 3

[0336] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the method for monitoring the safety status of a lead-acid battery of a backup power supply of a substation in Example 1, and the processor is configured to execute the program stored in the memory.

[0337] Embodiment 4

[0338] A storage medium stores a computer program, which, when executed by a processor, executes the steps of the method for monitoring the safety status of a lead-acid battery of a backup power supply for a substation in Embodiment 1.

[0339] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation, characterized in that: The following steps are involved: Step 1, collecting battery electrochemical impedance spectroscopy data and storing the collected data; Step 2: Label the collected data. The labeling categories include capacity and battery status. The battery status includes normal operation, battery aging, and prohibited use. Step 3: Use a generative adversarial network algorithm based on robust estimation to perform data expansion to generate new training samples; Step 4: Adopting the adaptive wave optimization algorithm to update the weights and biases in the neural network, simulating the propagation and energy distribution characteristics of the wave, and extracting features from the battery electrochemical impedance spectrum data; Step 5: Use the autoencoder neural network based on manifold mapping learning to reduce the dimension of the extracted features; Step 6: Classify the reduced-dimensional data using a support vector machine classification algorithm based on Landau quantization to determine the safety status of the battery, and use the classification results to monitor the safety status of the battery.

2. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method for collecting the battery electrochemical impedance spectroscopy data described in step 1 and storing the collected data is as follows: using an electrochemical impedance spectroscopy acquisition chip to collect solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W from the battery system, the collected data is stored in JSON format.

3. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method for annotating the collected data described in step 2 is as follows: the battery capacity is determined by constant current charging and discharging the battery, and different capacity stages are annotated from 100% to 70%. A set of capacity and solution impedance Rs, charge transfer impedance Rct, double layer capacitance Cdl and Warburg impedance W data sets are obtained every 1% for manual annotation.

4. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method for data expansion using the robust estimation-based generative adversarial network algorithm described in step 3 is as follows: (1) Initialize the network parameters of the generator and discriminator. The initialization is randomly assigned according to the Gaussian distribution. For the generator initialization, it is expressed as: And, for discriminator initialization, it is expressed as: In the formula, θ g is the parameter of the generator, σ g is the standard deviation of the generator parameters, θ d is the parameter of the discriminator, σ d is the standard deviation of the discriminator parameters; (2) Perform iterative training. The generator G receives random noise z from the latent space. The random noise z follows a standard normal distribution. The generator maps the noise to the data space through learning, imitating the distribution of battery monitoring data, which can be expressed as: x gen =G(z;θ g ) In the formula, x gen is the generated data, z is the input random noise vector, and I is the identity matrix; (3) The generated data and the actual collected battery monitoring data are input into the discriminator. The discriminator evaluates the authenticity of these data and updates its parameters through backpropagation. The training goal of the discriminator D is to maximize the distinction between the real data x and the generated data x gen The ability is achieved by minimizing the following loss function: The expected calculation formula for real data is: And, the expected calculation formula for generating data is: In the formula, p data is the distribution of the real data, p z is the distribution of noise; (4) Based on the feedback from the discriminator, the generator updates its parameters to generate data that is more difficult for the discriminator to recognize. The goal of the generator G is to make the discriminator D unable to distinguish between real data and generated data, which is achieved by minimizing the following loss function: Where, L QGAN is the uncertainty measurement loss term; For the uncertainty evaluation of the data G(z) generated by the generator, the calculation formula is: Sr(G(z))=-Tr(G(z)×logG(z)) Where Sr() is the uncertainty assessment function, Tr() represents the trace operation of the matrix; (5) Uncertainty measurement loss term L QGAN The calculation formula is: In the formula, γ ed is the uncertainty measure loss coefficient; (6) In each iteration, the quality of the generated data in the current training batch is evaluated, and the learning rates of the generator and discriminator are dynamically adjusted. The adjustment method is expressed as: In the formula, and are the new and old learning rates respectively; γ vh is the learning rate adjustment factor; L previous and L current are the previous and current loss function values ​​respectively; ΔL is the estimated amount of loss function change; (7) Repeat the above steps until a preset stop iteration condition is met, where the preset stop iteration condition is reaching a preset maximum number of iterations; (8) Use similarity metrics to evaluate the quality of generated data and adjust the training strategy accordingly, expressed as: Q(x f )→update(Q,θ g ′,θ d ′) Where Q() is the quality assessment function; x f is to generate data; update() is a function to adjust the training strategy based on quality evaluation; When the evaluation value of the quality evaluation function is greater than the preset threshold, the data is retained; otherwise, the generated data is discarded.

5. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 4, characterized in that: The quality assessment function Q() is implemented by calculating the similarity between the generated data and the data in the real data set, which is expressed as: Among them, sim() represents the similarity measurement function, and the measurement method is implemented based on the Euclidean distance calculation method; represents the i-th sample in the real data set, and N is the total number of samples.

6. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method described in step 4 of using the adaptive fluctuation optimization algorithm to update the weights and biases in the neural network, simulating the propagation and energy distribution characteristics of fluctuations, and extracting features from the battery electrochemical impedance spectrum data is as follows: (1) Define the neural network structure. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is the ReLU activation function. (2) Set the initial parameters of the neural network, including the weight w and bias b of the neural network, and initialize the wave source S i The position x i , intensity a i and the initial phase φ i , assuming the total number of wave sources is n, the initialization is expressed as: S i (x i ,a i ,f i ),i=1,2,…,n In the formula, x i Usually chosen as the initial position of the parameter, a i and φ i Randomly set its initial value; S i is the i-th wave source, with position x i , intensity a i and the initial phase φ i composition; (3) Calculate the loss L through forward propagation. Perform forward propagation on the input X using the parameter θ, calculate the output Y and compare it with the target value T. The calculation formula is: Y = frs(X;θ) L=Loss(Y,T) In the formula, frs() represents the network structure, and the output layer of the network structure uses the preset Softmax classifier to obtain the predicted label; Loss(Y,T) is the cross entropy loss; (4) Calculate the gradient of each parameter based on the loss function The calculation formula is: (5) The chain method is used to calculate the gradient. Assuming that the network output Y passes through the Softmax function, the gradient calculation formula for the weight w is: In the formula, y k is the output of Softmax, z k It is the linear output before the Softmax layer; (6) The calculation formula is: and, The calculation formula is: and, The calculation formula is: (7) According to the direction and magnitude of the gradient, a wave source is generated at the corresponding parameter position. Each wave source S i The strength is updated to The calculation formula is: In the formula, α ck is the learning rate, β bp is the attenuation coefficient, γ bp is the amplification factor; (8) The fluctuations of all wave sources are superimposed in the parameter space. The update amount of each parameter is determined by the superposition result of the fluctuations of all wave sources at that point. The parameter update method is expressed as: Where k is the wave number, ω is the angular frequency, which indicates the propagation speed and direction of the wave; Re() is the ReLU activation function, τ i,l is the fluctuation threshold; (9) Fluctuation threshold τ of each wave source i,l The layer sensitivity index is adjusted dynamically, and the adjustment method is expressed as: Where τ0 is the initial fluctuation threshold, S l is the layer sensitivity index, κ pq is the adjustment factor; (10) Layer sensitivity index S of the lth layer l The norm of the gradient output by this layer is calculated to measure the contribution of this layer to the overall network error. The calculation formula is: In the formula, is the norm of the parameter gradient of the lth layer, Lc is the number of network layers; (11) According to the loss reduction of this iteration, the attenuation coefficient β and amplification coefficient γ of each wave source are dynamically adjusted. The calculation formula is: In the formula, β bp0 ,γ bp0 are the attenuation coefficient and amplification coefficient of the previous iteration, λ bp ,μ bp are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, respectively, and ΔL is the change in loss; (12) Repeat the above steps until the preset stop iteration condition is met.

7. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method described in step 5 for reducing the dimension of the extracted features using the autoencoder neural network based on manifold mapping learning is as follows: (1) Initialize the network parameters. Use Gaussian distribution to initialize the weight w and bias b parameters of the autoencoder, expressed as: b i ＝0 In the formula, w i represents the weight of the i-th layer, σ ch represents standard deviation; (2) Through the autoencoder, the data x is mapped to the low-dimensional feature space z as follows: z=Sig(W e x+b e ) And, the decoder attempts to reconstruct the original input The way is expressed as: Where Sig() is the activation function; W e and b e are the weight and bias of the encoder respectively; W d and b d are the weights and biases of the decoder; (3) In the encoded low-dimensional space, the Riemann factor R is used to implement manifold learning. The calculation formula of the feature z of the Riemann factor constraint encoding is: In the formula, μ cy represents the preset manifold center, σ R is a parameter that controls the compactness of the manifold; (4) The loss function not only considers the reconstruction error, but also the Riemann manifold constraint. The reconstruction error ensures that the data can be effectively restored. The manifold constraint strengthens the geometric structure of the feature space. The calculation formula of the loss function is: In the formula, λ cm is the regularization coefficient; (5) Reconstruction error The calculation formula is: In the formula, z i is the value of z in the i-th dimension, μ i is the value of the preset manifold center in the i-th dimension; (6) Based on the loss function, adjust the network parameters through the back propagation algorithm; (7) Gradient It is calculated by the error back propagation method, and the calculation formula is: and, Through the decoder weight W d The transposed representation of is calculated as follows: (8) Since z is W e The result obtained by the activation function after the linear combination of and x is The calculation formula is: In the formula, Sig ′ () is the derivative of the Sigmoid activation function, diag(v) represents a diagonal matrix with vector v as the diagonal element, x T is the transpose of x; (9) Use the gradient descent method to update the parameters. The update method is expressed as: In the formula, α rv is the learning rate, t represents the number of iterations, is the updated weight parameter, is the weight parameter before updating; (10) Repeat the above steps until the preset stop iteration condition is met.

8. The method for monitoring the safety status of a lead-acid battery for backup power supply of a substation according to claim 7, characterized in that: The weight W e The calculation formula is pre-set by principal component analysis: Where V k is the covariance matrix X T The matrix consisting of the first k eigenvectors obtained from X, Σ k is the corresponding diagonal matrix of eigenvalues.

9. The method for monitoring the safety status of a lead-acid battery for a backup power supply of a substation according to claim 1, characterized in that: The method described in step 6 for classifying the reduced-dimensional data by the support vector machine classification algorithm based on Landau quantization is as follows: (1) Initialize the quantum state of the training data. Suppose the initial i-th sample data vector is And q ii Represents the jth feature of the i-th sample, converted into the quantum state |ψ i The calculation formula of > is: In the formula, α ij is the complex amplitude, satisfying Represents feature q ij The corresponding quantum ground state; (2) The initial quantized hyperplane is constructed by the Landau quantization principle. The probability amplitude of the quantum state is used to evaluate its relative position with the hyperplane to determine the initial support vector. The quantum hyperplane |w> is defined as: In the formula, γ k is the weight of the hyperplane vector, |v k > is the hyperplane vector; (3) The calculation formula for the decision boundary between the quantum state and the hyperplane is: In the formula, <ψ i |w> is the inner product of the quantum state and the hyperplane, f() is the decision boundary function, ij |v k > is the inner product of the feature and the hyperplane;​ (4) Weight γ of the quantum hyperplane k The calculation formula is determined by principal component analysis: In the formula, e k is the kth eigenvalue of the data covariance matrix, e l is the lth eigenvalue of the data covariance matrix; (5) In each iteration, the distance and angle between each support vector and other data points are checked, and fission operations are performed according to the clustering density and edge distance to refine the decision boundary; for each support vector, the fission process is achieved by adjusting its quantum state, which can be expressed as: In the formula, β bp is the fission factor, δ im is the inverse coefficient of the distance between i and m; (6) Distance inverse ratio coefficient δ im Dynamically adjust according to the distance between the ith data point and the mth data point. The calculation formula is: in, Represents data points and data points The square of the Euclidean distance between ∈ is a constant; (7) For data groups that exhibit similar characteristics, a fusion operation is performed to merge the representations of these data points to enhance the model's response to these characteristics. The quantum state merging operation of similar data points is expressed as: Where M i is a set of quantum states similar to , σ m is the weight assigned according to similarity; (8) Weight σ m It is calculated based on the similarity and the local density of the data points and is expressed as: Where, d mM It is the data point m to the set M i The average distance to other points in (9) Adjust the state of the qubit through quantum gate operations to optimize the accuracy of the classification results, expressed as: U(θ)|ψ i ″>=cos(θ)|ψ i ″>+sin(θ)∑ m≠i |ψ m > Where θ is the rotation angle of the quantum gate; (10) Repeat the above steps until the preset stop iteration condition is met.

10. The method for monitoring the safety status of a lead-acid battery for backup power supply of a substation according to claim 9, characterized in that: The quantum gate operation U(θ) is used to adjust the state of the qubit and is calculated as: U(θ)=e -iθH Where H is a specific Hamiltonian.

Images