A power distribution facility anomaly detection method based on deep learning

Abstract

The application relates to a power distribution facility anomaly detection method based on deep learning, and relates to the technical field of power distribution facility anomaly detection. The application collects electric characteristic signals of a power distribution facility at a set sampling rate, selects electric characteristic signals of a non-anomaly time period located on a slow spectrum submanifold as training data, minimizes a loss function, uses a full connection network to learn a mapping matrix from a nonlinear feature vector of the electric characteristic signals to a latent variable on a low-dimensional slow spectrum submanifold of the electric characteristic signals and a linear dynamics matrix in a latent space, in application, collects electric characteristic signals, uses the last corresponding latent variable of the electric characteristic signals and a predicted future latent variable obtained by iteration using the linear dynamics matrix, reconstructs a predicted non-anomaly electric characteristic signal through inverse transformation between the latent variable and the electric characteristic signal, and performs anomaly analysis by using the difference between the predicted non-anomaly electric characteristic signal and an actual electric characteristic signal.

Description

Technical Field

[0001] This invention relates to the field of power distribution facility anomaly detection technology, and in particular to a deep learning-based power distribution facility anomaly detection method. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Arcing faults, transformer winding deformation, and inter-turn short circuits in power systems are all typical abnormal events in nonlinear systems. If these abnormal events are not detected and handled in a timely manner in their early stages, they can often lead to equipment damage, production interruptions, property losses, and even personal injury or death.

[0004] However, early anomaly detection in nonlinear power distribution systems faces the following common challenges: First, early anomaly signals are weak. In the initial stage of anomaly development, observed system signals (such as current, voltage, vibration, etc.) are extremely similar to normal operating conditions, with fluctuations typically smaller than the signal noise level. Traditional amplitude threshold-based methods cannot effectively identify these anomalies. Second, nonlinear dynamics are complex. Almost all real-world engineering systems are nonlinear, exhibiting complex behaviors such as saturation, hysteresis, and chaos. Linear models cannot accurately describe their dynamic characteristics, while nonlinear models suffer from difficulties in parameter identification and excessive computational complexity. Third, anomaly samples are scarce. In practical engineering, real anomaly data is difficult to obtain (limited by safety conditions, equipment costs, or the rarity of anomaly events), making training methods relying on a large number of labeled anomaly samples difficult to implement. Fourth, high real-time requirements. Many application scenarios (such as power system protection and aero-engine monitoring) require responses within milliseconds or even microseconds, which traditional complex algorithms struggle to meet.

[0005] Existing statistical process control methods for anomaly analysis, such as Shewhart control charts, cumulative sum control charts, and exponentially weighted moving average control charts, determine anomalies by monitoring whether observed values ​​exceed statistical control limits. These methods assume the system is linear or that the observed values ​​are independent, making them difficult to handle nonlinear systems with strong temporal correlations. Furthermore, statistical process control methods have limited sensitivity in detecting early, weak shifts, typically requiring anomalies to be 2–3 times the standard deviation for reliable detection.

[0006] Existing signal processing methods for anomaly analysis, such as Fourier transform, wavelet transform, and Hilbert-Huang transform, identify anomalies by analyzing the signal's spectrum and time-frequency energy distribution. These methods have limitations: they require prior knowledge of the frequency band characteristics corresponding to the anomaly, and they are poorly adaptable to unknown anomaly types; early anomalies are typically very weak in the spectrum, resulting in low signal-to-noise ratios; and for non-stationary signals, time-frequency analysis has high computational complexity.

[0007] In recent years, Support Vector Machines (SVM), Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), and Long Short-Term Memory Networks (LSTM) have been widely used in anomaly detection. The main limitations of these methods include: the need for a large number of labeled normal and anomalous samples for training, while anomalous samples are extremely scarce in real-world engineering scenarios; and the model's generalization ability is greatly affected by the distribution of the training data, with performance dropping sharply for unseen working conditions or anomaly types.

[0008] Therefore, a deep learning-based method for detecting anomalies in power distribution facilities is needed to solve the above problems. Summary of the Invention

[0009] To solve the above-mentioned technical problems, or at least partially solve them, the present invention provides a deep learning-based method for detecting anomalies in power distribution facilities.

[0010] In a first aspect, the present invention provides a deep learning-based method for detecting anomalies in power distribution facilities, comprising: At a set sampling rate, electrical characteristic signals of power distribution facilities are collected, and electrical characteristic signals during periods without anomalies are selected as training data. The training data are located on a slow spectral manifold. With the goal of minimizing the loss function, a fully connected network is used to learn the mapping matrix from the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamic matrix in the latent space. The obtained linear dynamic matrix is ​​constrained to make the dynamic system of the electrical feature converge. In application, electrical characteristic signals are collected, and the latent variables corresponding to the electrical characteristic signals and the future latent variables are predicted by iterative linear dynamic matrix calculation. Predicting anomaly-free electrical characteristic signals by reconstructing them through inverse transformation between latent variables and electrical characteristic signals; Anomalies in power distribution facilities can be detected by using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals.

[0011] Furthermore, the process of constructing the nonlinear eigenvector of the electrical characteristic signal includes: According to Takens' embedding theorem, the training data is time-delayed embedded to reconstruct a high-dimensional state vector. The dimension of the high-dimensional state vector is no less than twice the dimension of the slow spectral manifold M, and it is optimized as follows: For an n-dimensional high-dimensional state vector, the distance to its nearest neighbor is calculated. When the dimension of the high-dimensional state vector increases by 1, calculate the new distance to its nearest neighbor. , among which, if If the percentage exceeds the set threshold, it is determined to be a false nearest neighbor, and the value of dimension n is set to reduce the proportion of false nearest neighbors to below the set threshold. For each high-dimensional state vector, perform p-order polynomial feature expansion and then combine the polynomial features to obtain a nonlinear feature vector.

[0012] Furthermore, the dimension of the slow spectral manifold M is obtained through the singular value decomposition of the nonlinear eigenvectors, including: centering the nonlinear eigenvectors; performing singular value decomposition on the centered nonlinear eigenvectors; taking the largest singular value decomposition eigenvalues ​​in the sorted order to calculate the cumulative energy percentage; and taking the smallest number of eigenvalues ​​that satisfy the cumulative energy percentage threshold as the dimension of the slow spectral manifold M.

[0013] Furthermore, by minimizing the loss function, deep learning is used to map the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamical matrix in the latent space, including: Suppose the latent variables on the characteristic slow spectral manifold M of the electrical signal With nonlinear eigenvectors The mapping relationship between them is as follows: ,in, For mapping matrix , where is the matrix to be determined; Latent variables on slow spectral manifold M Following linear dynamics, the linear dynamic relationship of continuous electrical signal characteristics is linearized and discretized to obtain: ; To describe latent variables The linear dynamic matrix of the dynamic relationship is the matrix to be determined; Fully connected networks output a prediction mapping matrix based on nonlinear eigenvectors. and predicting linear dynamics matrix ; During training, the parameters of the fully connected network are adjusted to enable it to predict the mapping matrix. and predicting linear dynamics matrix To minimize the loss function.

[0014] Furthermore, the loss function is: ; in Here, is the regularization coefficient, and N is the number of training samples. It is the Frobenius norm.

[0015] Furthermore, the nonlinear eigenvectors of electrical characteristic signals and latent variables Mapping relationship between and nonlinear eigenvectors Given a linear term containing an electrical characteristic signal, there exists an inverse transform between the latent variables and the electrical characteristic signal. Let the linear combination from the latent variables to recover the electrical characteristic signal be defined as... Where W and b are the weight matrix and bias of the inverse transform, respectively, and W and b are obtained by minimizing the reconstruction error of the point feature signals on the training set: ; in, Let N be the mapping matrix, and N be the number of training data samples. This represents the true value of the corresponding electrical characteristic signal.

[0016] Furthermore, restricting the linear dynamic matrix to make the dynamic system of electric characteristics converge includes: Perform eigenvalue decomposition on the learned linear dynamics matrix; The eigenvalues ​​with a modulus greater than 1 in the diagonal matrix of eigenvalues ​​obtained by eigenvalue decomposition are projected onto the unit circle, their arguments are preserved, and the magnitude growth is eliminated to obtain the target eigenvalues. The target diagonal matrix is ​​obtained by reconstructing the diagonal matrix using the target eigenvalues. The target's linear dynamics matrix is ​​reconstructed using the target's diagonal matrix and the eigenvector matrix of the linear dynamics matrix.

[0017] Furthermore, detecting anomalies in power distribution facilities by utilizing the difference between predicted anomaly-free electrical characteristic signals and actual electrical characteristic signals includes: Calculate the absolute value sequence of prediction errors Calculate the mean of the absolute values ​​of the prediction errors. and standard deviation ; Using the mean of the absolute values ​​of prediction errors and standard deviation Standardize the error: ; Statistical positive offset and negative offset : , , in, Allowable drift parameters; initial positive and negative offsets are set to zero; An abnormal alarm is triggered when the positive or negative offset exceeds the control limit.

[0018] Furthermore, after determining the standardized error, the outlier probability corresponding to the current error is determined: ; in, The cumulative distribution function of the standard normal distribution. It is the error function; The confidence interval of the predicted electrical characteristic signal is determined based on the variance of the predicted electrical characteristic signal, including: The variance of the predicted electrical characteristic signal is: ,in, To predict the variance of electrical characteristic signals, To predict the variance of the error in the electrical characteristic signal, W is the weight matrix of the inverse transform. For latent variable covariance; The standard deviation is determined based on the variance of the predicted electrical characteristic signal, and the confidence interval is determined based on the standard deviation and the predicted electrical characteristic signal. The anomaly probability and confidence interval are output.

[0019] Secondly, the present invention provides a deep learning-based power distribution facility anomaly detection device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit; the storage unit storing a computer program, which, when executed by the processing unit, implements the deep learning-based power distribution facility anomaly detection method.

[0020] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art: This invention minimizes a loss function and utilizes a fully connected network to learn the mapping matrix from the nonlinear eigenvectors of an electrical characteristic signal to the latent variables on a low-dimensional slow spectral manifold of the electrical characteristic signal, as well as the linear dynamic matrix in the latent space. In application, the electrical characteristic signal is acquired, and the latent variables corresponding to the last electrical characteristic signal are used, along with the predicted future latent variables obtained through iterative calculation using the linear dynamic matrix. An anomaly-free electrical characteristic signal is reconstructed through the inverse transformation between the latent variables and the electrical characteristic signal. Anomaly analysis is performed using the difference between the predicted anomaly-free electrical characteristic signal and the actual electrical characteristic signal, enabling the detection and early warning of anomalies using only anomaly-free data. This invention amplifies weak perturbations through polynomial boosting, enabling early warning before visible changes appear in the observed anomaly signal. This invention requires no anomaly samples, reducing data acquisition costs and making it suitable for various power distribution facilities where anomaly samples are scarce. The original observed signal can be accurately reconstructed from the latent space, and the prediction error has a clear physical meaning, facilitating understanding and verification by engineers. The online application only involves low-dimensional matrix multiplication and polynomial evaluation, making it suitable for embedded real-time implementation. This invention has significant advantages in terms of reversibility, sensitivity, false alarm rate, and computational complexity, and has industrial practical value and prospects for promotion. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart of a deep learning-based anomaly detection method for power distribution facilities is provided in an embodiment of the present invention; Figure 2 A flowchart illustrating how a fully connected network learns a mapping matrix from nonlinear feature vectors to low-dimensional latent variables and a linear dynamic matrix in the latent space, as provided in an embodiment of the present invention. Figure 3 A flowchart illustrating the convergence of the dynamic system of the electric characteristic obtained by limiting the linear dynamic matrix provided in the embodiments of the present invention; Figure 4 A flowchart for detecting power distribution facility anomalies using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals, provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a deep learning-based power distribution facility anomaly detection device provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0026] Example 1 like Figure 1 As shown, the present invention provides a deep learning-based method for detecting anomalies in power distribution facilities, comprising: S100 collects electrical characteristic signals of power distribution facilities at a set sampling rate, selects electrical characteristic signals from time periods without abnormalities as training data, and the training data is located on a slow spectral manifold.

[0027] In practical implementation, let's take current as an example, using it as an electrical characteristic signal to indicate potential anomalies. The dynamic behavior of current is modeled as the following nonlinear system: ; in, Let be the current at time t. Let be the derivative of the current at time t. For the dynamic relationship to be determined, let . As the equilibrium point, Smooth, its Jacobian matrix J eigenvalues They can be divided into two groups: slow eigenvalues fast eigenvalues And spectral gaps exist: .

[0028] In the absence of anomalies, at the equilibrium point of a nonlinear system, there exists an m-dimensional slow spectral manifold M, where m is smaller than the dimension of the nonlinear system. The slow spectral manifold M is tangent to the slow characteristic space, i.e., it is spanned by the slow characteristics of the dominant nonlinear system and is locally invariant. On the slow spectral manifold M, a j-dimensional nonlinear system can be simplified to m-dimensional dynamics, and a transformation relationship between latent variables and currents on the slow spectral manifold M can be established, such that the dynamics of the nonlinear system are approximately linear under the latent variables.

[0029] The dimension of the slow spectral manifold M is obtained through the singular value decomposition of the nonlinear eigenvectors, including: centering the nonlinear eigenvectors; performing singular value decomposition on the centered nonlinear eigenvectors; taking the largest singular value decomposition eigenvalues ​​in the sorted order to calculate the cumulative energy percentage; and taking the smallest number of eigenvalues ​​that satisfy the cumulative energy percentage threshold as the dimension of the slow spectral manifold M.

[0030] For the power grid, the dominant slow characteristic is the power grid fundamental frequency and its second harmonic.

[0031] S200, according to Takens' embedding theorem, performs time-delay embedding on the training data and reconstructs it into a high-dimensional state vector.

[0032] Taking current signals as an example, the discrete quantity of the sampled current signal is: Its reconstructed high-dimensional state vector is: ;in, For the k-th sampled current of the training data, For the time-delayed embedding The high-dimensional state vector is defined by n, where n is the dimension of the high-dimensional state vector. The dimension of the high-dimensional state vector is no less than twice the dimension of the slow spectral manifold M. In specific implementation, for an n-dimensional high-dimensional state vector, its distance to its nearest neighbor is calculated. When the dimension of the high-dimensional state vector increases by 1, calculate the new distance to its nearest neighbor. , among which, if If the percentage exceeds the set threshold, it is determined to be a false nearest neighbor, and the dimension n is adjusted to reduce the proportion of false nearest neighbors to below the set threshold.

[0033] According to Takens' embedding theorem, for a given dynamical system, even if only one sequence of the system (current sequence) can be observed, a phase space topologically equivalent to the original system can be reconstructed by time delaying that sequence. This reconstructs a one-dimensional electrical characteristic signal sequence into a high-dimensional state vector. The reconstructed high-dimensional state vector retains the topological structure of the original attractor, providing sufficient state information for subsequent nonlinear transformations.

[0034] S300, perform a p-order polynomial feature expansion on each high-dimensional state vector to obtain a nonlinear feature vector. .

[0035] In specific implementation, nonlinear eigenvectors It contains first-order terms, second-order terms, ..., p-th-order terms of a high-dimensional state vector. The first-order term of the high-dimensional state vector is... The quadratic terms of a high-dimensional state vector contain squared terms and quadratic terms, such as... The cubic terms of a high-dimensional state vector include cubic terms, quadratic terms, cubic cross terms, and so on.

[0036] The observed evolution of any nonlinear dynamical system can be described by an infinite-dimensional linear operator. A polynomial basis is a set of bases for an infinite-dimensional space, which, when truncated to a finite order, can be approximated by the Koopman operator. The Weierstrass approximation theorem guarantees that continuous functions on compact sets can be uniformly approximated by polynomials. Therefore, for dynamics on attractors, there exists a polynomial lift that makes the linear approximation error arbitrarily small.

[0037] Processes S200 and S300 construct the nonlinear eigenvectors of the electrical characteristic signal. The nonlinear eigenvectors in this application... A nonlinear extension to the Takens reconstruction. The reconstructed phase space of the system typically shows a linear or approximately linear dynamic system. However, the real dynamic evolution equations are often nonlinear. This involves constructing nonlinear eigenvectors. In essence, this involves a high-dimensional mapping. In a higher-dimensional feature space, the complex nonlinear relationships that were originally complex in the low-dimensional space become linearly separable or can be fitted using linear regression. For current signals, the higher-order terms in the nonlinear eigenvector obtained by polynomial expansion can amplify minute current changes and are extremely sensitive to weak precursors of current anomalies. The cross terms in the nonlinear eigenvector obtained by polynomial expansion capture the nonlinear interactions between variables; the computation is simple and easy to implement in real time.

[0038] S400 aims to minimize the loss function by using a fully connected network to learn the mapping matrix from nonlinear eigenvectors to low-dimensional latent variables and the linear dynamic matrix in the latent space. The obtained linear dynamic matrix is ​​constrained to ensure that the dynamic system of the electric feature converges in the long term.

[0039] like Figure 2 As shown, in the specific implementation process, let the latent variables on the slow spectral manifold M of the electrical signal characteristics be... With nonlinear eigenvectors The mapping relationship between them is as follows: ,in, For mapping matrix , where is the matrix to be determined. These represent the slow spectral manifold M and the nonlinear eigenvectors, respectively. Dimensions.

[0040] Latent variables on slow spectral manifold M Following linear dynamics, the linearized and discretized continuous linear dynamic relationship yields: ; To describe latent variables The linear dynamic matrix of the dynamic relationship is the matrix to be determined.

[0041] Construct a fully connected network, which is based on nonlinear feature vectors. To output the mapping matrix and linear dynamics matrix .

[0042] Based on the prediction mapping matrix output by the fully connected network and predicting linear dynamics matrix Based on the above relationship, the nonlinear feature vectors of the training data are used. Determine the latent variables for prediction And predict the results of the next time step for the latent variables. , specifically , .

[0043] During training, the parameters of the fully connected network are adjusted to enable it to predict the mapping matrix. and predicting linear dynamics matrix To minimize the loss function: ; in is the regularization coefficient, and N is the number of training samples.

[0044] By using only training data without anomalies to train a fully connected network to learn the dynamics of electrical characteristic signals under anomaly-free conditions, the problem of scarce abnormal data in practical engineering is solved.

[0045] The learned linear dynamics matrix The eigenvalues ​​may have a modulus greater than 1, leading to long-term prediction divergence. This application restricts the obtained linear dynamic matrix to ensure convergence of the dynamic system of the electrical characteristic, such as... Figure 3 As shown, the process includes: The learned linear dynamics matrix Perform eigenvalue decomposition: ,in, Eigenvalues The resulting diagonal matrix , This is the eigenvector matrix.

[0046] Projecting the eigenvalues ​​with a modulus greater than 1 in the diagonal matrix of eigenvalues ​​obtained from eigenvalue decomposition onto the unit circle, preserving their arguments, and eliminating amplitude growth yields the target eigenvalues: ; The target eigenvalue projects eigenvalues ​​with a magnitude greater than 1 onto the unit circle, preserving their argument and eliminating magnitude growth.

[0047] Reconstruct the diagonal matrix using the target eigenvalues ​​to obtain the target diagonal matrix: ; Using the target diagonal matrix and the linear dynamic matrix The eigenvector matrix reconstructs the linear dynamics matrix of the target: .

[0048] Ideally, linearization of an anomaly-free system should result in a spectrum that lies on or inside the unit circle. Eigenvalues ​​with a modulus greater than 1 typically originate from noise or truncation errors. Forcing eigenvalues ​​with moduli greater than 1 to have a modulus of 1 preserves the system's oscillatory characteristics while preventing divergence. Any deviation from the stable trajectory can be attributed to external disturbances, thus enabling reliable early warning.

[0049] When the S500 is applied, it acquires electrical characteristic signals and uses the latent variables corresponding to these signals and iteratively predicts future latent variables using a linear dynamic matrix. .

[0050] S600 reconstructs and predicts anomaly-free electrical characteristic signals through an inverse transformation between latent variables and electrical characteristic signals. In specific implementation, to reconstruct the electrical characteristic signals from the latent variables, the nonlinear eigenvectors of the electrical characteristic signals are... and latent variables There is a mapping relationship and nonlinear eigenvectors Given a linear term containing an electrical characteristic signal, there exists a linear combination from which the electrical characteristic signal can be recovered, i.e., an inverse transform between the latent variables and the electrical characteristic signal. Let the linear combination from which the electrical characteristic signal can be recovered be... Where W and b are the weight matrix and bias, respectively, and W and b are obtained by minimizing the reconstruction error of the point feature signals on the training set: .

[0051] The S700 detects anomalies in power distribution facilities by utilizing the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals. Figure 4 As shown, the process includes: Calculate the absolute value sequence of prediction errors Calculate the mean of the absolute values ​​of the prediction errors. and standard deviation .

[0052] Using the mean of the absolute values ​​of prediction errors and standard deviation Standardize the error: ; Statistical positive offset using standardized error and negative offset : , , in, The allowable drift parameter is set to 0.5~1.0; the initial positive and negative offsets are set to zero.

[0053] An abnormal alarm is triggered when the positive or negative offset exceeds the control limit.

[0054] Permissible drift parameters and control limits can be set through Monte Carlo simulation or empirical methods.

[0055] To prevent frequent alarms, a blanking time is set, meaning that after an alarm is triggered, it will not re-alarm within this time period. The above anomaly identification method is highly sensitive to persistent, small error shifts and is robust to transient noise, significantly outperforming the fixed threshold method.

[0056] In the specific implementation process, the mean of the absolute value of the prediction error is also used. and standard deviation After calculating the standardized error, determine the anomaly probability corresponding to the current error: ; in, The cumulative distribution function of the standard normal distribution. This is the error function.

[0057] The confidence interval of the predicted electrical characteristic signal is determined based on the variance of the predicted values. The predicted current electrical characteristic signal is affected by two sources of uncertainty: the randomness of the prediction error itself, represented by the variance; and the estimation uncertainty of the latent variables, represented by the covariance of the latent variables. To express.

[0058] The variance of the predicted electrical characteristic signal is: ; The standard deviation is determined based on the variance of the predicted electrical characteristic signal. The confidence interval is then determined based on the standard deviation and the predicted electrical characteristic signal. For example, the 95% confidence interval corresponds to the range of the predicted electrical characteristic signal minus 1.96 times the standard deviation and the predicted electrical characteristic signal plus 1.96 times the standard deviation.

[0059] The anomaly probability and confidence interval are output.

[0060] Example 2 See Figure 4 As shown, this embodiment of the invention provides a deep learning-based power distribution facility anomaly detection device, comprising: at least one processing unit, the processing unit being connected to a storage unit via a bus unit; the storage unit, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the software program, computer-executable program, and module corresponding to the deep learning-based power distribution facility anomaly detection method in this embodiment of the invention. The processing unit implements the aforementioned deep learning-based power distribution facility anomaly detection method by running the software program, computer-executable program, and module stored in the storage unit, including: At a set sampling rate, electrical characteristic signals of power distribution facilities are collected, and electrical characteristic signals during periods without anomalies are selected as training data. The training data are located on a slow spectral manifold. With the goal of minimizing the loss function, a fully connected network is used to learn the mapping matrix from the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamic matrix in the latent space. The obtained linear dynamic matrix is ​​constrained to ensure that the dynamic system of the electrical feature converges in the long term. In application, electrical characteristic signals are collected, and the latent variables corresponding to the electrical characteristic signals and the future latent variables are predicted by iterative linear dynamic matrix calculation. Predicting anomaly-free electrical characteristic signals by reconstructing them through inverse transformation between latent variables and electrical characteristic signals; Anomalies in power distribution facilities can be detected by using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals.

[0061] Of course, the computer program stored in the storage unit of the deep learning-based power distribution facility anomaly detection device provided in the embodiments of the present invention is not limited to the method operation described above, but can also execute related operations in the deep learning-based power distribution facility anomaly detection method provided in any embodiment of the present invention.

[0062] Example 3 This invention provides a computer-readable storage medium storing a computer program. When executed, the computer program implements the deep learning-based power distribution facility anomaly detection method, comprising: At a set sampling rate, electrical characteristic signals of power distribution facilities are collected, and electrical characteristic signals during periods without anomalies are selected as training data. The training data are located on a slow spectral manifold. With the goal of minimizing the loss function, a fully connected network is used to learn the mapping matrix from the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamic matrix in the latent space. The obtained linear dynamic matrix is ​​constrained to ensure that the dynamic system of the electrical feature converges in the long term. In application, electrical characteristic signals are collected, and the latent variables corresponding to the electrical characteristic signals and the future latent variables are predicted by iterative linear dynamic matrix calculation. Predicting anomaly-free electrical characteristic signals by reconstructing them through inverse transformation between latent variables and electrical characteristic signals; Anomalies in power distribution facilities can be detected by using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals.

[0063] In the embodiments provided by this invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, structures, or units, and may be electrical, mechanical, or other forms.

[0064] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0065] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0066] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for detecting anomalies in power distribution facilities based on deep learning, characterized in that, include: At a set sampling rate, electrical characteristic signals of power distribution facilities are collected, and electrical characteristic signals during periods without anomalies are selected as training data. The training data are located on a slow spectral manifold. With the goal of minimizing the loss function, a fully connected network is used to learn the mapping matrix from the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamic matrix in the latent space. The obtained linear dynamic matrix is ​​constrained to ensure that the dynamic system of the electrical feature converges in the long term. In application, electrical characteristic signals are collected, and the latent variables corresponding to the electrical characteristic signals and the future latent variables are predicted by iterative linear dynamic matrix calculation. Predicting anomaly-free electrical characteristic signals by reconstructing them through inverse transformation between latent variables and electrical characteristic signals; Anomalies in power distribution facilities can be detected by using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals.

2. The deep learning-based anomaly detection method for power distribution facilities according to claim 1, characterized in that, The process of constructing the nonlinear feature vector of the electrical feature signal includes: According to Takens' embedding theorem, the training data is time-delayed embedded to reconstruct a high-dimensional state vector. The dimension of the high-dimensional state vector is no less than twice the dimension of the slow spectral manifold M, and it is optimized as follows: For an n-dimensional high-dimensional state vector, the distance to its nearest neighbor is calculated. When the dimension of the high-dimensional state vector increases by 1, calculate the new distance to its nearest neighbor. , among which, if If the percentage exceeds the set threshold, it is determined to be a false nearest neighbor, and the value of dimension n is set to reduce the proportion of false nearest neighbors to below the set threshold. For each high-dimensional state vector, perform p-order polynomial feature expansion and then combine the polynomial features to obtain a nonlinear feature vector.

3. The deep learning-based anomaly detection method for power distribution facilities according to claim 2, characterized in that, The dimension of the slow spectral manifold M is obtained through the singular value decomposition of the nonlinear eigenvectors, including: centering the nonlinear eigenvectors; performing singular value decomposition on the centered nonlinear eigenvectors; taking the largest singular value decomposition eigenvalues ​​in the sorted order to calculate the cumulative energy percentage; and taking the smallest number of eigenvalues ​​that satisfy the cumulative energy percentage threshold as the dimension of the slow spectral manifold M.

4. The deep learning-based anomaly detection method for power distribution facilities according to claim 1, characterized in that, By minimizing the loss function, deep learning is used to map the nonlinear eigenvectors of the electrical feature signal to the latent variables on the low-dimensional slow spectral manifold of the electrical feature signal, as well as the linear dynamical matrix in the latent space, including: Suppose the latent variables on the characteristic slow spectral manifold M of the electrical signal With nonlinear eigenvectors The mapping relationship between them is as follows: ,in, For mapping matrix , where is the matrix to be determined; Latent variables on slow spectral manifold M Following linear dynamics, the linear dynamic relationship of continuous electrical signal characteristics is linearized and discretized to obtain: ; To describe latent variables The linear dynamic matrix of the dynamic relationship is the matrix to be determined; Fully connected networks output a prediction mapping matrix based on nonlinear eigenvectors. and predicting linear dynamics matrix ; During training, the parameters of the fully connected network are adjusted to enable it to predict the mapping matrix. and predicting linear dynamics matrix To minimize the loss function.

5. The deep learning-based anomaly detection method for power distribution facilities according to claim 4, characterized in that, The loss function is: ； in Here, is the regularization coefficient, and N is the number of training samples. It is the Frobenius norm.

6. The deep learning-based anomaly detection method for power distribution facilities according to claim 1, characterized in that, Nonlinear eigenvectors of electrical characteristic signals and latent variables Mapping relationship between and nonlinear eigenvectors Given a linear term containing an electrical characteristic signal, there exists an inverse transform between the latent variables and the electrical characteristic signal. Let the linear combination from the latent variables to recover the electrical characteristic signal be defined as... Where W and b are the weight matrix and bias of the inverse transform, respectively, and W and b are obtained by minimizing the reconstruction error of the point feature signals on the training set: ； in, Let N be the mapping matrix, and N be the number of training data samples. This represents the true value of the corresponding electrical characteristic signal.

7. The deep learning-based anomaly detection method for power distribution facilities according to claim 1, characterized in that, Restricting the linear dynamic matrix to ensure convergence of a dynamic system with electric characteristics includes: Perform eigenvalue decomposition on the learned linear dynamics matrix; The eigenvalues ​​with a modulus greater than 1 in the diagonal matrix of eigenvalues ​​obtained by eigenvalue decomposition are projected onto the unit circle, their arguments are preserved, and the magnitude growth is eliminated to obtain the target eigenvalues. The target diagonal matrix is ​​obtained by reconstructing the diagonal matrix using the target eigenvalues. The target's linear dynamics matrix is ​​reconstructed using the target's diagonal matrix and the eigenvector matrix of the linear dynamics matrix.

8. The deep learning-based anomaly detection method for power distribution facilities according to claim 1, characterized in that, Detecting power distribution facility anomalies using the difference between predicted abnormal electrical characteristic signals and actual electrical characteristic signals includes: Calculate the absolute value sequence of prediction errors Calculate the mean of the absolute values ​​of the prediction errors. and standard deviation ; Using the mean of the absolute values ​​of prediction errors and standard deviation Standardize the error: ； Statistical positive offset and negative offset : ， ， in, Allowable drift parameters; initial positive and negative offsets are set to zero; An abnormal alarm is triggered when the positive or negative offset exceeds the control limit.

9. The deep learning-based anomaly detection method for power distribution facilities according to claim 8, characterized in that, After standardizing the error, determine the anomaly probability corresponding to the current error: ； in, The cumulative distribution function of the standard normal distribution. It is the error function; The confidence interval of the predicted electrical characteristic signal is determined based on the variance of the predicted electrical characteristic signal, including: The variance of the predicted electrical characteristic signal is: ,in, To predict the variance of electrical characteristic signals, To predict the variance of the error in the electrical characteristic signal, W is the weight matrix of the inverse transform. For latent variable covariance; The standard deviation is determined based on the variance of the predicted electrical characteristic signal, and the confidence interval is determined based on the standard deviation and the predicted electrical characteristic signal. The anomaly probability and confidence interval are output.

10. A deep learning-based anomaly detection device for power distribution facilities, characterized in that, include: At least one processing unit, wherein the processing unit is connected to the storage unit via a bus unit; The storage unit stores a computer program, which, when executed by the processing unit, implements the deep learning-based power distribution facility anomaly detection method as described in any one of claims 1-9.

Images