Grid-connected inverter fault diagnosis method based on dynamic response

Abstract

The invention discloses a grid-connected inverter fault diagnosis method based on dynamic response, and particularly relates to the technical field of grid-connected inverter fault diagnosis, and the method comprises the steps: decomposing a dynamic response signal of a grid-connected inverter through fractional order wavelet transform, dynamically adjusting a fractional order parameter based on weak power grid impedance estimation, screening a sub-band signal through Tsallis entropy, and carrying out the fault diagnosis of the grid-connected inverter. Generating a time-frequency characteristic tensor; processing the time-frequency feature tensor by adopting dynamic diffusion mapping, and outputting a spectrum embedding vector and an abnormal scale index; analyzing a spectrum embedding vector by using a causal variation auto-encoder, constructing a cross-scale causal graph based on a structural causal model, initializing prior distribution by using a time-frequency feature tensor, and outputting a causal edge weight and a potential feature subset; the potential feature subsets are fused through a physical constraint neural differential equation, a fault feature vector is generated through random perturbation regularization, the fault type and position are output, and the problem that the false alarm rate is high due to insufficient resolution and noise interference under a weak power grid is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of grid-connected inverter fault diagnosis, and more particularly to a grid-connected inverter fault diagnosis method based on dynamic response. Background Art

[0002] With the continuous expansion of renewable energy grid integration, grid-connected inverters are playing an increasingly important role in modern power systems. In weak grid environments, the stability and reliability of inverters are crucial. However, due to the dynamic characteristics of the grid and the complexity of inverter faults, traditional fault diagnosis methods face significant challenges. Existing fault diagnosis techniques mostly rely on frequency-domain or time-domain analysis, using methods such as Fourier transforms and wavelet transforms to extract signal features and identify fault type and location. However, these methods often suffer from insufficient diagnostic accuracy and poor robustness in complex grid environments, especially those with strong dynamic responses and nonlinear noise. These methods struggle to adapt to the rapid changes in the grid environment and the diverse fault modes.

[0003] Current fault diagnosis techniques often overlook changes in grid impedance and the multi-scale and non-Gaussian characteristics of signals at different time scales. Under weak grid conditions, the dynamic changes in grid impedance and inverter fault modes may exhibit different characteristics at multiple time scales, resulting in poor performance of traditional diagnostic methods when dealing with cross-scale signals. Existing diagnostic systems are often unable to effectively identify these cross-scale coupling characteristics, especially during transient anomalies and long-term evolution, resulting in reduced fault diagnosis accuracy. While traditional diagnostic methods can identify faults at a single time scale, they have significant deficiencies in capturing cross-scale characteristics and handling nonlinear issues. Therefore, how to comprehensively consider the multi-scale characteristics of signals, non-Gaussian noise, and grid impedance changes has become a key technical issue in improving the accuracy and robustness of grid-connected inverter fault diagnosis. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a grid-connected inverter fault diagnosis method based on dynamic response. By combining fractional-order wavelet transform, dynamic diffusion mapping and physical constraint neural differential equation technology, multi-scale fault feature extraction and high-precision diagnosis are realized to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for diagnosing grid-connected inverter faults based on dynamic response, comprising the following steps: The dynamic response signal of the grid-connected inverter is decomposed by fractional-order wavelet transform. The fractional-order parameters are dynamically adjusted based on the weak grid impedance estimation. The short-, medium-, and long-time scale sub-band signals are screened using Tsallis entropy to generate a time-frequency feature tensor. Dynamic diffusion mapping is used to process the time-frequency feature tensor, generating anomaly scores through diffusion distance on the Riemannian manifold, mining cross-scale anomaly patterns with an adaptive diffusion kernel, optimizing mapping accuracy through the width of the adaptive diffusion kernel, and outputting spectral embedding vectors and anomaly scale indices. A causal variational autoencoder is used to analyze spectral embedding vectors, construct a cross-scale causal graph based on a structural causal model, initialize the prior distribution with the time-frequency feature tensor, infer causal dependencies at short, medium, and long time scales, and output causal edge weights and potential feature subsets. The potential feature subsets are fused through physically constrained neural differential equations, the parameters are initialized with causal edge weights and inverter dynamic model, and random perturbation regularization is used to generate fault feature vectors, and the fault type and location are output.

[0006] Preferably, the fractional-order parameter is dynamically adjusted based on the weak grid impedance, and the value range of the fractional-order parameter is between 0 and 1. The acquisition process includes: Real-time estimation of weak grid impedance using recursive least squares method , the lower the weak grid impedance, the larger the fractional order parameter, which leads to the improvement of short-time sub-band resolution, the decrease of adaptive diffusion kernel width, and the focus on transient anomaly detection; the fractional order parameter is negatively correlated with the weak grid impedance.

[0007] Preferably, the cross-scale causal edge set is obtained from the abnormal scale index set , (i, j) indicates that the fault propagates from scale i to scale j, satisfying , 、 is the anomaly score corresponding to scale i, j, Represents the anomaly score threshold; the construction logic of the causal edge set E includes: Anomaly scale screening: Based on the anomaly score and anomaly score threshold, the anomaly scale index set is screened; Causal edge generation: Based on the inverter fault propagation law, cross-scale causal relationships are predefined; causal edges are only constructed for abnormal scales belonging to the abnormal scale index set.

[0008] Preferably, the anomaly score threshold There are two methods for determining t: static and dynamic. When the grid impedance fluctuation rate exceeds the threshold, the dynamic method is used; otherwise, the static method is used. The static method calculates the statistical distribution of the anomaly score based on historical normal data and obtains the sum of the mean and three standard deviations. The dynamic method calculates the quantile of the anomaly score in the past time period as the anomaly score threshold at the current time t.

[0009] Preferably, the rate of change of the feature over time is predicted by a physically constrained neuronal differential equation , to distinguish between transient interference and persistent faults, the dynamic equation of the physical constraint neuron differential equation is: in, The data-driven term for modeling a neural network takes as input the multi-scale latent feature vector x at time t and outputs the time-evolved gradient of the feature vector; is the causal edge weight, which indicates the strength of fault propagation from scale i to scale j; It represents the sensitivity of scale j features to scale i features and is calculated by back-propagation automatic differentiation technique; The inverter dynamic model equations include the phase-locked loop dynamic equations and the grid impedance equations; is the physical coupling coefficient, which is used to balance the contribution of data-driven terms and physical constraint terms.

[0010] Preferably, the physical coupling coefficient γ is determined by the initial coupling coefficient , input characteristic volatility and attenuation coefficient Jointly determined, the characteristic volatility includes the variance Var(x) of the multi-scale characteristic vector x, taking the attenuation coefficient The negative value of the product of the sum and variance Var(x) is an exponential function; The input variance Var(x) increases, the exponential term The absolute value of increases, resulting in a decrease in the exponential function value, a decrease in the γ value, a weakening of the physical constraint effect, and avoiding overfitting under noise interference; The input variance Var(x) decreases: the absolute value of the exponential term decreases, the exponential function value increases, the physical constraint effect is enhanced, the physical constraint neural differential equation more strictly follows the inverter dynamic equation, and the physical consistency of the diagnosis is improved; The way to obtain the physical coupling coefficient balances the weights of data-driven and physical constraints.

[0011] Preferably, the adjustment of the adaptive diffusion kernel width is achieved by combining the abnormal sensitivity coefficient and the grid impedance fluctuation rate; Calculate the time-frequency entropy change rate TER of the subband signal and obtain the time-frequency entropy change rate smoothing value TER by sliding average filtering MA ; Calculate the local variance ASV of the anomaly score by the formula Calculate the abnormal sensitivity coefficient ASC, where α is the adjustment factor; the width of the adaptive diffusion kernel is determined by the following formula Adjustment: in, represents the width of the adaptive diffusion kernel at time t, represents the initial kernel width; λ is the dynamic sensitivity coefficient; It represents the time domain derivative of the Tsallis entropy of the subband signal, reflecting the non-stationary rate of the subband signal; Represents a sliding average filter with a window length of Wt.

[0012] Preferably, when constructing a cross-scale causal graph based on a structural causal model, the loss function of the causal variational autoencoder includes a causal regularization term : in, The causal edge weight corresponding to the anomaly scale index (i, j) reflects the propagation intensity of the fault from scale i to j. It is predefined by the structural causal model and is a scalar weight. 、 is the output of the causal variational autoencoder 、 latent variables of multiple scales; Represents latent variables and Pearson correlation coefficient; Represents latent variables and The covariance of Represents the latent variables and The standard deviation of represents the square of the Frobenius norm, which is used to quantify the difference between weights and covariances; latent variables and It is the output of the causal variational autoencoder, an abstract representation of fault features at different time scales, and a low-dimensional potential feature obtained by nonlinearly mapping the spectral embedding vector.

[0013] Preferably, the latent variable is a low-dimensional feature representation of different time scales output by a causal variational autoencoder; the latent feature subset is composed of multiple latent variables, which is a feature set of abnormal scales screened out from the cross-scale causal graph; the latent variables are filtered by an abnormal score threshold to obtain the latent feature subset; a multi-scale latent feature vector x is obtained through fusion, and under physical constraints, a fault feature vector is obtained through evolution.

[0014] Preferably, after the physical constraint neural differential equation generates a comprehensive fault feature vector, it is input into the classifier to output the probability distribution of each fault type; the maximum value in the probability distribution is taken as the current diagnostic confidence p; a preset threshold p is set th , if p <p th , the diagnosis confidence is determined to be insufficient, triggering the backtracking mechanism, including: The adaptive diffusion kernel parameters are adjusted by the following formula: in, 、 Represent the width of the adaptive diffusion kernel before and after adjustment; Represents the backtracking gain coefficient.

[0015] Technical effects and advantages of the present invention: (1) The present invention uses fractional-order wavelet transform to decompose dynamic response signals, and dynamically adjusts fractional-order parameters in combination with real-time estimation of weak grid impedance to enhance the resolution of short-time transient features; uses Tsallis entropy to screen short / medium / long-time subband signals to generate time-frequency feature tensors; mines cross-scale abnormal patterns and outputs spectral embedding vectors, constructs cross-scale causal graphs through causal variational autoencoders, and analyzes fault propagation paths; uses physically constrained neural differential equations to generate fault feature vectors and classify and output them; and solves the problems of insufficient resolution of traditional methods in weak grids, high false alarm rate caused by noise interference, and difficulty in modeling cross-scale fault propagation.

[0016] (2) The adaptive diffusion kernel width adjustment steps of the present invention are as follows: the diffusion kernel width is dynamically optimized by the time-frequency entropy change rate and the local variance of the anomaly score to improve the signal-to-noise ratio of the weak fault feature; a causal regularization term is added to the causal variational autoencoder loss function to force the covariance of the latent variables to be consistent with the predefined causal edge weights, thereby solving the logical deviation problem caused by the causal graph relying on prior knowledge; the fault propagation path is aligned through the Pearson correlation coefficient constraint to reduce misjudgment; when the classification confidence is insufficient, the diffusion kernel width is triggered to backtrack, and the dynamic change rate of the feature and the consistency of the causal edge weight are integrated to generate a confidence enhancement coefficient to suppress false triggering caused by instantaneous interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the grid-connected inverter fault diagnosis method of the present invention.

[0018] Figure 2 This is a flow chart of the fault identification based on the cross-scale causal graph of the present invention. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0020] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0021] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0022] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0023] Example 1, see Figure 1 The embodiment of the present invention provides a grid-connected inverter fault diagnosis method based on dynamic response, which includes the following steps: Step 1: Decompose the dynamic response signal of the grid-connected inverter through fractional-order wavelet transform, dynamically adjust the fractional-order parameters based on the weak grid impedance estimation, and use Tsallis entropy to filter the short-time, medium-time, and long-time scale sub-band signals to generate the time-frequency feature tensor; The specific implementation process of step one is: in order to capture the cross-time scale characteristics of the dynamic response of the grid-connected inverter, the dynamic response signal (such as current, voltage, power) is decomposed into multiple scales using fractional-order wavelet transform, and the dynamic response signal is split into sub-band signals of short-time (transient current mutation), medium-time (oscillation attenuation) and long-time (aging trend) scales; the technical features include: a fractional-order adaptive strategy based on weak grid impedance fluctuation, dynamically adjusting the resolution of fractional-order wavelet transform decomposition through fractional-order parameters; using generalized information entropy (Tsallis entropy) to evaluate the effectiveness of sub-band signals and screen fault-related scales, including: dividing the sub-band signal into several amplitude intervals, counting the number of sub-band coefficients in each amplitude interval and dividing it by the total number of coefficients in the sub-band to construct a discrete probability distribution; after decomposition, each sub-band signal consists of a set of discrete values, which are called sub-band coefficients, depending on the decomposition level and type); each A subband coefficient corresponds to the amplitude information of the original signal at that scale (or frequency band) and at that moment. On this basis, a Tsallis entropy parameter q greater than zero and not equal to one is selected, and the probability values corresponding to each amplitude interval are processed according to the qth power, and summarized and calculated according to the Tsallis entropy definition formula to obtain the entropy value of the subband signal. The entropy values of the subband signals at each scale are compared, and the subband signals with significantly higher entropy values and concentrated information are retained (considered as fault-related scales), and the subband signals with lower entropy values and dominated by noise are eliminated. The fractional-order parameters are optimized in combination with real-time impedance estimation. The processing logic is as follows: in a weak power grid environment, the fractional-order parameters are dynamically adjusted through impedance estimation to adapt to the non-stationarity of the signal. Through Tsallis entropy sorting, the high-information quantum band signals are retained, and the noise-dominated scales are eliminated. The output is a multi-scale subband signal set (time-frequency feature tensor), which provides high-resolution input for subsequent anomaly mining. Explanation: A sliding window is used to intercept a time segment of the dynamic response signal, and the ratio of the energy of each sub-band signal to the total energy of the full frequency band is combined. In one possible embodiment, the window length of the sliding window is 2-3 times the fundamental wave period of the dynamic response signal, and the sliding step size is 1 / 5-1 / 3 of the window length. Step 2: Dynamic diffusion mapping is used to process the time-frequency feature tensor. Anomaly scores are generated through diffusion distance on the Riemannian manifold. Adaptive diffusion kernels are used to mine cross-scale anomaly patterns. The mapping accuracy is optimized by the width of the adaptive diffusion kernel. The spectral embedding vector and anomaly scale index are output. The specific implementation process of step 2 is as follows: Based on the multi-scale sub-band signal set in step 1, the abnormal features in each scale sub-band signal are detected through dynamic diffusion mapping to extract cross-scale fault patterns; dynamic diffusion mapping uses spectral embedding technology to map the sub-band signals of each scale into a low-dimensional space through the diffusion distance on the Riemann manifold, revealing abnormal patterns at short-term (such as current mutation caused by IGBT short circuit) and long-term (such as voltage trend caused by capacitor aging) scales, including: using diffusion distance to generate dynamic anomaly scores and quantify the degree of anomaly; the processing logic is as follows: taking the time-frequency feature tensor as input, constructing spectral embedding vectors of each scale through dynamic diffusion mapping; calculating anomaly scores through diffusion distance, and filtering out highly abnormal scales and feature vectors through anomaly scores; adjusting the diffusion kernel parameters in combination with power grid dynamics (such as load fluctuations) to enhance robustness; the output is a cross-scale anomaly pattern set (the cross-scale anomaly pattern set includes spectral embedding vectors and anomaly scale indexes), which provides low-dimensional features for subsequent causal modeling; The process of dynamic diffusion of spectral embedding technology is to construct a similarity matrix, a normalized transfer matrix, a diffusion operator and perform spectral decomposition based on the multi-scale sub-band signal set, and finally obtain the spectral embedding vector and anomaly score. Specifically, it includes: The specific implementation process is as follows: (1) Feature vector set: At the sth scale, the time-frequency characteristics of each sampling point at the scale of the dynamic response signal (such as current, voltage, power) of the grid-connected inverter are extracted through fractional-order wavelet transform; the time-frequency characteristics of the sth scale subband are defined as follows: , i=1…N; N represents the number of sampling points. In this way, each sampling point corresponds to a high-dimensional vector at the current scale, which is used to measure the similarity and abnormality between points in the future; (2) Similarity matrix: In order to measure the similarity between any two points (i.e., two eigenvectors) at the same scale, it is necessary to first construct an N×N similarity matrix; the elements in the i-th row and j-th column of the matrix are represents the i-th sample With the jth sample The similarity between them is defined as follows: , , Among them, σ(·) is the k-nearest neighbor density estimate, Adaptive adjustment based on the power grid dynamics is the scaling factor; in, It represents the Euclidean distance between the i-th eigenvector and the j-th eigenvector, that is, the straight-line distance between them when they are regarded as points in d-dimensional space. The smaller the distance, the more similar their time-frequency features are; the larger the distance, the more obvious the feature difference is. exp(·) is a Gaussian kernel function used to map the distance value to the similarity interval (0, 1). When the distance between two vectors is extremely small, Close to 1, it means high similarity; when the distance is large, Close to 0, indicating low similarity; (3) Transfer matrix and steady-state distribution Degree Matrix ; At the sth scale, for each sample point i, the sum of its similarities with all other samples is calculated first. This sum of similarities is called the degree of the point (denoted by Put the degree values of all points in a diagonal matrix to get the degree matrix In layman's terms, each row (or column) of the degree matrix corresponds to the "total similarity between this point and other points"; Transfer matrix: similarity matrix Each row is divided by the degree of the corresponding point , we get the transfer matrix ; Since we make the sum of all similarities in this row become 1, each row can be regarded as "the probability distribution of spreading from the current point to all other points"; in other words, The i-th row of the formula indicates that if a random diffusion is performed starting from the i-th point, the probability of diffusion to the j-th point is ”.

[0024] Steady-state distribution ; By dividing the degree of each time-frequency feature by the sum of all time-frequency feature degrees, we can obtain the probability that the random diffusion process will stay at each point in the long-term diffusion process; this probability reflects the importance or centrality of each long-term diffusion in the diffusion network; Diffusion operator and spectral embedding: A positive integer t is selected as the diffusion time. The diffusion time t means that when constructing the diffusion operator, the random diffusion process is iterated on the transfer matrix for t steps to capture the deeper structure of the data in time or space. Construct a t-step diffusion operator: select diffusion time t>0, Find the t-step diffusion operator by raising it to the power of t The t-step diffusion operator means: a row vector represents the probability distribution of reaching all other points after t-step diffusion from a starting point; as t increases, the diffusion operator will gradually integrate local neighborhood information, allowing the data to obtain a smoother and more coherent low-dimensional representation of the overall geometric structure; Spectral decomposition: Solving the eigenvalues and eigenvectors of the t-step diffusion operator, the spectral decomposition will obtain a set of eigenvalues and its corresponding normalized eigenvector ; ; Retain the first m eigenvalues and construct spectral embedding: In order to map the original sample points from a high-dimensional space to a lower-dimensional embedding space, it is necessary to retain only the most important features. Usually the first m eigenvalues closest to 1 are selected. , eigenvalue , ..., eigenvalue ; (5) Anomaly score and screening Calculate the diffusion distance, which is used to measure the sample point at the sth scale. and The difference between the distributions after t steps of diffusion; its meaning is: first calculate the difference between the distributions after t steps of diffusion; The probability distribution of diffusion to all other points after t steps, and After t steps of diffusion to all other points, the probability distribution is then calculated for the two sets of probability distributions in the weights Weighted Euclidean distance under (steady-state distribution); specific process: For each intermediate point k (k=1…N), take the i-th row element of the transfer matrix at the s-th scale after t-step power operation , and the j-th row element ; Calculate their difference , and then square the difference; Divide the square of the difference corresponding to all k by (steady-state probability of point k at the sth scale), and accumulate the sum; Finally, take the square root of the accumulated result and get ; This distance is smaller, indicating and The more similar they are on the diffusion network, the closer the probability distributions after diffusion are; a larger distance indicates that their corresponding diffusion distributions are quite different, and they may belong to different clusters or have anomalies; Anomaly score: for the i-th sample point at the s-th scale , its anomaly score Defined as: First, all other sample points at the sth scale (j=1…N) Calculate the diffusion distance ; Square each distance value and multiply it by the corresponding weight ; Finally, sum all items from j=1 to N and get ; Among them, the weight You can take the corresponding similarity in the similarity matrix , or we can directly take the steady-state distribution ; Intuitively, if If the diffusion distance is very different from most other points (the accumulated value of the squared distance after weighting is large), it means that it is more easily identified as an outlier at the sth scale. On the contrary, if the diffusion distance from the neighboring points is small, the anomaly score is low and it is more like normal data. Multi-scale screening: For each scale s=1, 2, ..., S, calculate the set of anomaly scores of all samples at that scale Define a statistical deviation metric function StatDist(·) to measure the difference or degree of deviation between the abnormal score distribution and the normal distribution at this scale. Commonly used metrics include: whether the mean and variance are significantly higher than other scales, whether the kurtosis or skewness of the distribution is significantly biased towards the tail, etc. Select the scale index s* that makes StatDist the largest among all scales: , this optimal scale s* is considered to be the scale that best reflects the fault or abnormal mode; Finally, the spectral embedding vector of the s*th scale is retained , as the low-dimensional feature output of cross-scale abnormal patterns.

[0025] Step 3: Use a causal variational autoencoder to analyze the spectral embedding vector, construct a cross-scale causal graph based on the structural causal model, initialize the prior distribution with the time-frequency feature tensor, infer causal dependencies at short, medium, and long time scales, and output causal edge weights and potential feature subsets; Explanation: Latent variables are low-dimensional feature representations of different time scales output by the causal variational autoencoder. The latent feature subset consists of multiple latent variables and is a collection of features at abnormal scales filtered out from the cross-scale causal graph. The latent variables are filtered by an anomaly score threshold to obtain the latent feature subset. The multi-scale latent feature vector x is obtained through fusion, and under physical constraints, it is evolved to obtain the fault feature vector. The logical description of the causal encoder is as follows: first, extract vectors of three scales, short-time, medium-time and long-time, from the spectral embedding vector obtained in step 2, and concatenate the three-scale vectors corresponding to the same moment as the input of the causal variational autoencoder; for each time scale, set a Gaussian prior distribution with zero mean and unit covariance as parameters to constrain the learning of latent variables; the encoding network receives the spectral embedding vectors of each scale, outputs the posterior distribution parameters (mean and variance) of the corresponding latent variables, and obtains the latent variables by sampling from the posterior distribution through reparameterization technology; at the same time, use a joint encoder to map the spliced global vector to a joint latent variable across scales. Assuming that there is a linear causal relationship between latent variables, they are represented in the form of a directed acyclic graph. By introducing structural constraints, the encoder, decoder and causal adjacency matrix are jointly optimized during the training process to ensure that each non-zero element in the adjacency matrix accurately represents the causal weight between the latent variables. According to the anomaly score calculated in step 2, it is mapped to the latent variable level, and the anomaly score is calculated for the latent dimension of each scale. The dimension with the highest score is selected according to a pre-set threshold to form a latent feature subset. The latent variables of each scale in the latent feature subset are combined according to a pre-defined fusion method to generate a multi-scale latent feature vector at the current moment. Under the physical model constraints of the grid-connected inverter, the fused latent feature vector is mapped to a fault feature vector through an evolution function for subsequent fault diagnosis and prediction.

[0026] The specific implementation process of step three is as follows: based on the cross-scale anomaly pattern set in step two, a causal variational autoencoder is constructed to infer the causal relationship between features at different time scales, and the cross-scale coupling effect of fault features is analyzed, including: taking the spectral embedding vector as input, the causal variational autoencoder constructs a potential causal graph through causal regularization (based on the structural causal model SCM) to characterize the causal dependencies at short, medium, and long time scales (such as transient faults triggering long-term instability); optimizing the latent space distribution through variational inference and quantifying the coupling strength; the processing logic is as follows: using the anomaly scale index as a constraint, the causal nodes of the causal variational autoencoder are screened to reduce the complexity of the structural causal model; the prior distribution of the causal variational autoencoder is initialized using the time-frequency feature tensor; through iterative updates of the time series, the dynamic interaction between the control loop (such as the phase-locked loop) and the fault features is captured, and the output is a cross-scale causal graph (causal edge weights and latent feature subsets), providing a structured representation for the final diagnosis; Step 4: Fusion of potential feature subsets through physically constrained neural differential equations, initialization of parameters with causal edge weights and inverter dynamic model, generation of fault feature vectors using random perturbation regularization, and output of fault type and location; See Figure 2The fault identification flow chart based on cross-scale causal graph, the specific implementation process of step four is: based on the cross-scale causal graph of step three, use the physical constraint neural differential equation to fuse cross-scale features and perform fault diagnosis, including: designing multi-scale physical constraint neural differential equations, fusing short-time, medium-time and long-time features through continuous-time dynamic modeling; initializing ODE parameters with the edge weights of the causal graph and the inverter dynamic model to enhance the physical consistency of the coupling features; improving the robustness of the model to weak grid noise through random perturbation regularization; the processing logic is: taking the causal graph and potential feature subset of step three as input, constructing a multi-scale feature tensor; the physical constraint neural differential equation fuses cross-scale features through adaptive time step and physical constraints to generate a comprehensive fault feature vector; using the classifier to output the fault type and location; if the diagnosis confidence is insufficient, backtracking the cross-scale abnormal pattern set of step two, re-triggering the causal modeling, and if the diagnosis confidence meets the requirements, outputting the fault diagnosis result (fault type, location, confidence).

[0027] It is necessary to further explain in the embodiment of the present invention that in step 1, the fractional-order parameter is dynamically adjusted based on the weak grid impedance. The value range of the fractional-order parameter is 0 to 1. The acquisition process includes: Real-time estimation of weak grid impedance using recursive least squares method , the lower the weak grid impedance, the larger the fractional-order parameter, resulting in improved short-time sub-band resolution, reduced width of the adaptive diffusion kernel, and focused transient anomaly detection; the fractional-order parameter is negatively correlated with the weak grid impedance; in a possible embodiment, the fractional-order parameter at time t is calculated by the following formula : in, It represents the maximum value of the weak grid impedance, and the hyperbolic tangent function constraint is used to ensure that the value range of the fractional-order parameter is between 0 and 1.

[0028] It is necessary to further explain in the embodiment of the present invention that in step 1, the subband signal is screened using Tsallis entropy to remove the noise-dominated scale, which specifically includes: use Represents subband signal Tsallis entropy, Represents the index of the subband signal, the noise dominant scale is the Tsallis entropy below the dynamic threshold The subband signal of the noise-dominated scale ; The screening conditions for the time-frequency feature tensor are , in, is the non-extensive entropy order, which is used to adjust the sensitivity to non-Gaussian features. When , Tsallis entropy has a selective suppression characteristic for non-Gaussian noise; in a possible embodiment, the value of q is ; Represents subband signal The normalized energy distribution probability at time k satisfies ; Represents subband signal The total number of time points, k is the time point index; dynamic threshold 、 Respectively represent the mean and standard deviation of Tsallis entropy in historical data; is the adjustment factor, and its value is .

[0029] In the embodiment of the present invention, it is necessary to further explain that the cross-scale causal edge set is obtained by the abnormal scale index set. , (i, j) indicates that the fault propagates from scale i to scale j, satisfying , 、 is the anomaly score corresponding to scale i, j, Represents the anomaly score threshold; the construction logic of the causal edge set E includes: Anomaly scale screening: Based on the anomaly score and anomaly score threshold, the anomaly scale index set is screened; Causal edge generation: Based on the inverter fault propagation law (e.g., IGBT short circuit → capacitor aging → voltage distortion), pre-define cross-scale causal relationships; only construct causal edges for abnormal scales belonging to the abnormal scale index set; Anomaly score threshold Methods for determining the anomaly score include static and dynamic methods. When the grid impedance fluctuation rate (the degree of fluctuation of the difference between the maximum grid impedance and the minimum grid impedance relative to the preset value within an integer multiple of the grid fundamental wave period) exceeds a threshold, such as 5%, the dynamic method is used; otherwise, the static method is used. The static method calculates the statistical distribution of the anomaly score based on historical normal data, obtaining the sum of the mean and three standard deviations. The dynamic method calculates the quantile of the anomaly score in the past time period as the anomaly score threshold at the current time t. For example, at each time point t, data within the past time length (such as 1 minute) is taken and the 0.99 quantile of these data is calculated as the dynamic anomaly score threshold at the current time t.

[0030] In the embodiment of the present invention, it is necessary to further explain that in step 4, the rate of change of the characteristic over time is predicted by the physical constraint neuron differential equation. , to distinguish between transient interference and persistent faults, the dynamic equation of the physical constraint neuron differential equation is: in, The data-driven term for modeling a neural network takes as input the multi-scale latent feature vector x at time t and outputs the time-evolved gradient of the feature vector; is the causal edge weight, which indicates the strength of fault propagation from scale i to scale j; It represents the sensitivity of scale j features to scale i features and is calculated by back-propagation automatic differentiation technique; The inverter dynamic model equations include the phase-locked loop dynamic equations and the grid impedance equations; is the physical coupling coefficient, which is , used to balance the contributions of data-driven terms and physical constraints; Explanation: By solving the physical constraint neuron differential equation and integrating the time evolution gradient, the multi-scale potential feature vector x is obtained as a result of its evolution over time, and the integrated fault feature vector is output; the integrated fault feature vector is input into the classifier to output the fault type and location; the dynamic evolution trajectory of the integrated feature vector is used to distinguish between transient interference and persistent faults; the feature vector of persistent faults will show a stable or cumulative abnormal pattern during the integration process, while the feature vector of transient interference will decay rapidly.

[0031] The explanation is that the phase-locked loop is the core control module of the grid-connected inverter. Its function is to track the phase and frequency of the grid voltage in real time to ensure that the inverter output current is synchronized with the grid voltage. The phase-locked loop dynamically adjusts the phase angle according to the grid voltage error: by measuring the instantaneous value of the grid voltage, the phase error between the actual voltage and the expected voltage is calculated. The error signal is converted into the adjustment rate of the phase angle through the proportional-integral controller, and finally the phase of the inverter output is strictly synchronized with the grid. The dynamic characteristics of the phase-locked loop directly affect the stability of the inverter. For example, when the grid voltage is distorted or the frequency suddenly changes, the phase-locked loop needs to converge quickly to avoid the grid current from losing step and triggering a protective shutdown. The grid impedance equation describes the interaction between the inverter output impedance and the grid impedance and is used to analyze system stability in weak grids (low short-circuit ratios). The ratio of the inverter output impedance to the grid impedance directly affects the grid-connected current characteristics. In weak grids (high grid impedance), the inverter must actively adjust control parameters (such as the current loop gain) to avoid resonance or harmonic amplification caused by impedance mismatch. By analyzing the impedance ratio (such as the Nyquist criterion), it is determined whether the system will oscillate at a specific frequency. The grid impedance equation is an important basis for fault diagnosis. For example, capacitor aging or inductor saturation can change the inverter output impedance, causing the impedance characteristics to deviate from the normal range, which in turn causes voltage / current harmonic anomalies.

[0032] In one possible embodiment, the fault diagnosis implementation process based on physical constraint neural differential equations includes: Step 401: Build a multi-scale feature fusion architecture: Integrate the potential feature subsets (including feature vectors at different scales, such as short-term, medium-term, and long-term) with the causal edge weights to construct a unified multi-scale feature tensor; align the resolutions of different time scales through interpolation or pooling techniques to ensure that the features are consistent along the time axis; and design the core architecture of the physically constrained neural differential equation. Step 402: Physical parameter initialization and dynamic adjustment: The causal edge weights are directly assigned as the initial parameters of the cross-scale propagation term. At the same time, the control coefficients of the phase-locked loop (such as proportional gain and integral coefficient) are initialized according to the inverter hardware parameter manual. The grid impedance value is updated in real time. To balance the weight of data-driven and physical constraints, a dynamic physical coupling coefficient is introduced: In strong power grid scenarios (impedance fluctuation <2%), the physical coupling coefficient is close to 1.0, and the model strictly follows the physical equations, making it suitable for high-precision diagnosis in stable environments; In weak grid scenarios (impedance fluctuation > 5%), the physical coupling coefficient is reduced to 0.5, enhancing the adaptability of data-driven terms and suppressing noise interference. The physical coupling coefficient is automatically adjusted through an exponential decay function to ensure a smooth transition. Step 403: Anti-interference training and dynamic solution: To improve the model's robustness to grid noise, controllable Gaussian noise is injected into the feature tensor. The noise intensity is positively correlated with the feature's inherent volatility. A fourth-order Runge-Kutta adaptive solver is used to calculate the feature evolution path, with the time step dynamically adjusted based on the feature change rate. For example, during rapid changes (such as when an IGBT shorts), the time step is shortened to capture transient details; during stable periods (such as when capacitors slowly age), the step is lengthened to improve computational efficiency. Finally, a comprehensive fault feature vector is output that integrates cross-scale information. Step 404: Fault decision making and closed-loop optimization: The comprehensive features are mapped to specific fault categories and locations through the Softmax classifier; the iterative termination conditions are set to avoid infinite loops; and through the synergy of multi-scale fusion and physical constraints, highly reliable fault diagnosis is achieved in complex noisy environments.

[0033] It should be further explained in the embodiment of the present invention that step 4 includes: fusing the potential feature subset through a physical constraint neural differential equation, using the causal edge weights and inverter dynamic model initialization parameters, dynamically adjusting the coupling strength between the data-driven term and the physical constraint term, generating a fault feature vector and outputting the fault type and location; wherein the physical constraint neural differential equation includes: Data-driven: Modeling the time evolution gradient of fault characteristics through neural networks; Causal propagation term: quantifies the propagation strength between features at different time scales based on the causal edge weights of the cross-scale causal graph; Physical constraints: The inverter's phase-locked loop control equation and grid impedance equation are embedded, and the evolution of constraint characteristics conforms to the laws of circuit physics. The data-driven term, the causal propagation term and the physical constraint term work together through a dynamic physical coupling coefficient, and the dynamic physical coupling coefficient is adaptively adjusted according to the grid impedance fluctuation rate.

[0034] In one possible embodiment, the physical coupling coefficient From the initial coupling coefficient , input feature volatility (variance of the multi-scale feature vector x) and decay coefficient (set to 0.1) jointly determine the initial coupling coefficient Set to 0.8, dynamically adjust the value of γ through the exponential decay function, and take the decay coefficient The negative value of the product of the sum and variance Var(x) is an exponential function, which is specifically expressed as follows: in, is the initial coupling coefficient, is the variance of the multi-scale feature vector, reflecting the volatility of the input features, is the attenuation coefficient; Formula interpretation: As the input variance Var(x) increases, the exponential term The absolute value of increases, resulting in a decrease in the exponential function value, a decrease in the γ value, a weakening of the physical constraint, and a greater reliance on data-driven terms in the physical constraint neural differential equation to avoid overfitting under noise interference. The input variance Var(x) decreases: the absolute value of the exponential term decreases, the exponential function value increases, and the γ value approaches the initial value of 0.8. The physical constraint effect is enhanced, and the physical constraint neural differential equation more strictly follows the inverter dynamic equation, improving the physical consistency of diagnosis. The way of obtaining the physical coupling coefficient balances the weights of data-driven and physical constraints, adapts to dynamic noise changes in weak power grid scenarios, and ensures the robustness of fault diagnosis.

[0035] Summary: This embodiment of the present invention achieves precise location of multi-scale faults in grid-connected inverters in weak power grid scenarios by constructing a basic diagnostic framework combining dynamic decomposition, anomaly detection, causal fusion, and physical constraint classification. First, a fractional-order wavelet transform is used to decompose the dynamic response signal. The fractional-order parameters are dynamically adjusted in combination with real-time estimation of the weak power grid impedance to enhance the resolution of short-term transient features. Tsallis entropy is used to filter short-, medium-, and long-duration subband signals, and a dynamic threshold is used to remove noise-dominated scales to generate a time-frequency feature tensor. Secondly, a dynamic diffusion map is used to calculate diffusion distances on a Riemannian manifold, mining cross-scale anomaly patterns and outputting spectral embedding vectors. Furthermore, a causal variational autoencoder is used to construct a cross-scale causal graph, analyze the fault propagation path, and extract potential feature subsets. Finally, a physically constrained neural differential equation is used to fuse the inverter dynamic model with data-driven features to generate and classify fault feature vectors. This embodiment addresses the problems of traditional methods in weak power grids, such as insufficient resolution, high false alarm rates due to noise interference, and difficulty in modeling cross-scale fault propagation. It improves the diagnostic accuracy of faults such as IGBT short circuits and capacitor aging.

[0036] The difference between Example 2 and Example 1 is that, in order to enhance the robustness of the solution, the following steps are further included: Adaptive diffusion kernel width adjustment steps: Dynamically optimize the diffusion kernel width by using the time-frequency entropy change rate and the local variance of the anomaly score to address the issue of insufficient sensitivity of the fixed kernel width to transient anomaly detection that may exist in Example 1. The kernel width adaptively shrinks / expands with grid impedance fluctuations (such as lightning strikes and load mutations), focusing on key frequency bands in strong noise environments and improving the signal-to-noise ratio of weak fault features. Dynamic causal edge weight constraint step: Add a causal regularization term to the causal variational autoencoder loss function to force the covariance of the latent variables to be consistent with the predefined causal edge weights. This resolves the logical deviation problem caused by the causal graph's reliance on prior knowledge in Example 1. Align the fault propagation path using the Pearson correlation coefficient constraint to reduce misjudgments (for example, misdiagnosing capacitor aging as voltage distortion). In the diagnostic backtracking and confidence enhancement step, when the classification confidence is insufficient, the diffusion kernel width backtracking is triggered, and the feature dynamic change rate (DCR) and the causal edge weight consistency (LCW) are integrated to generate the confidence enhancement coefficient (FCEC) to suppress false triggering caused by transient interference (such as lightning strikes); It should be further explained in the embodiment of the present invention that in the step 2, the adjustment of the adaptive diffusion kernel width is achieved by combining the abnormal sensitivity coefficient and the grid impedance fluctuation rate; Calculate the time-frequency entropy change rate TER of the subband signal and obtain the time-frequency entropy change rate smoothing value TER by sliding average filtering MA ; Calculate the local variance ASV of the anomaly score by the formula Calculate the abnormal sensitivity coefficient ASC, where α is the adjustment factor; the width of the adaptive diffusion kernel is determined by the following formula Adjustment: in, represents the width of the adaptive diffusion kernel at time t, represents the initial kernel width; λ is the dynamic sensitivity coefficient, which is 0.05≤λ≤0.2; It represents the time domain derivative of the Tsallis entropy of the subband signal, reflecting the non-stationary rate of the subband signal; Represents a sliding average filter with a window length of Wt.

[0037] It is necessary to further explain in the embodiment of the present invention that in step 3, when constructing a cross-scale causal graph based on the structural causal model, the loss function of the causal variational autoencoder includes a causal regularization term : in, The causal edge weight corresponding to the anomaly scale index (i, j) reflects the propagation intensity of the fault from scale i to j. It is predefined by the structural causal model and is a scalar weight. 、 is the output of the causal variational autoencoder 、 latent variables of multiple scales; Represents latent variables and Pearson correlation coefficient; Represents latent variables and The covariance of Represents the latent variables and The standard deviation of represents the square of the Frobenius norm, which is used to quantify the difference between weights and covariances; latent variables and It is the output of the causal variational autoencoder, an abstract representation of fault characteristics at different time scales, and a low-dimensional latent feature obtained by nonlinearly mapping the spectral embedding vector; The significance of the causal regularization term: The causal regularization term ensures that the structural causal model learns cross-scale causal relationships that conform to the laws of fault propagation by forcing the covariance between latent variables to be consistent with predefined causal weights.

[0038] In practical applications, the dynamic response signals of the grid-connected inverter, such as current, voltage, and power, are divided into M sampling segments according to a fixed-length window. The encoder outputs a latent variable vector for each window. Then, the value sequences of the i-th and j-th potential dimensions are extracted on all windows respectively. , calculate their sample covariance and normalize them with their respective sample standard deviations to obtain the empirical correlation coefficient The causal regularization term is to add the predefined causal weights to each pair of potential dimensions i, j. Correlation coefficient with experience The squared differences are accumulated, forcing the encoder to keep the co-variations between dimensions in the latent space consistent with the preset causal structure during training, thereby incorporating the true dynamic response signal and revealing how the encoder constrains the latent variables to learn the correct causal relationship.

[0039] In the embodiment of the present invention, it is necessary to further explain that after the physical constraint neural differential equation generates a comprehensive fault feature vector, it is input into a classifier (such as a Softmax layer) to output the probability distribution of each fault type; the maximum value in the probability distribution is taken as the current diagnostic confidence p; a preset threshold p is set. th (such as p th =0.9), if p <p th , then the diagnosis confidence is judged to be insufficient and the backtracking mechanism is triggered; In step 4, if the diagnosis confidence is insufficient to trigger backtracking, specifically if: When the fault type probability output by the classifier is the maximum When , adjust the adaptive diffusion kernel parameters of step 2, and the update method is: in, 、 Represent the width of the adaptive diffusion kernel before and after adjustment; Represents the backtracking gain coefficient, which is ; Indicates the diagnostic confidence threshold, which is .

[0040] It should be further explained in the embodiment of the present invention that in step 4, improving the diagnostic confidence by calculating the diagnostic confidence enhancement coefficient includes the following sub-steps: Calculate the dynamic change rate of the fault feature vector and quantify the time gradient norm of the feature evolution; express the dynamic change rate by the L2 norm of the calculation result of the physical constraint neural differential equation ; Calculate the local consistency of causal edge weights and use the standard deviation of causal edge weights to represent the local consistency parameter LCW of causal edge weights; By combining the dynamic changes of features and the stability of causal propagation, a comprehensive indicator for measuring the credibility of features is generated, which is proportional to the dynamic changes of features and inversely proportional to the volatility of causal propagation. The index is then mapped to a range of 0 to 1 to obtain the diagnostic confidence enhancement coefficient. Adjusting the classifier input weight according to the diagnosis confidence enhancement coefficient refers to: When the fault confidence enhancement coefficient FCEC is high, it means that the feature credibility is high. By increasing the weighting factor, the input feature vector is adjusted to enhance the confidence of the classifier output. When FCEC is low, it means that the feature credibility is low. By decreasing the weighting factor, the input feature vector is adjusted to reduce the confidence of the classifier output.

[0041] Summary: Based on Example 1, this embodiment of the present invention further improves the diagnostic robustness under complex working conditions by introducing a dynamic optimization mechanism and a confidence enhancement strategy. Specific improvements include: 1) Design an adaptive diffusion kernel width adjustment algorithm to dynamically optimize the kernel function by integrating the time-frequency entropy change rate and anomaly score variance to enhance the sensitivity of transient anomaly detection; 2) Adding a causal regularization term to the causal variational autoencoder forces the covariance of latent variables to be consistent with the predefined causal edge weights, solving the problem of fault propagation logic misalignment caused by model bias; 3) Construct a diagnostic backtracking mechanism. When the classification confidence falls below a threshold, it triggers dynamic backtracking of the diffusion kernel parameters and integrates the dynamic rate of change of features with causal consistency to generate a confidence enhancement coefficient. This embodiment addresses complex scenarios such as weak grid impedance fluctuations and confusion between transient interference and persistent faults. It solves the problems of traditional methods such as parameter rigidity, weakened causal logic, and high transient misjudgment rate, thereby improving fault location accuracy.

[0042] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fault diagnosis method for grid-connected inverter based on dynamic response, characterized in that: The following steps are involved: The dynamic response signal of the grid-connected inverter is decomposed by fractional-order wavelet transform. The fractional-order parameters are dynamically adjusted based on the weak grid impedance estimation. The short-, medium-, and long-time scale sub-band signals are screened using Tsallis entropy to generate a time-frequency feature tensor. Dynamic diffusion mapping is used to process the time-frequency feature tensor to generate anomaly scores, and an adaptive diffusion kernel is used to mine cross-scale anomaly patterns. The mapping accuracy is optimized by the width of the adaptive diffusion kernel, and the spectral embedding vector and anomaly scale index are output. A causal variational autoencoder is used to analyze spectral embedding vectors, construct a cross-scale causal graph based on a structural causal model, initialize the prior distribution with the time-frequency feature tensor, infer causal dependencies at short, medium, and long time scales, and output causal edge weights and potential feature subsets. The potential feature subsets are fused through physically constrained neural differential equations, and the parameters are initialized with causal edge weights and inverter dynamic model to generate fault feature vectors and output fault type and location.

2. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 1, characterized in that: The fractional-order parameters are dynamically adjusted based on the weak grid impedance. The value range of the fractional-order parameters is 0 to 1. The acquisition process includes: Real-time estimation of weak grid impedance using recursive least squares method , the lower the weak grid impedance, the larger the fractional order parameter, which leads to the improvement of short-time sub-band resolution, the decrease of adaptive diffusion kernel width, and the focus on transient anomaly detection; the fractional order parameter is negatively correlated with the weak grid impedance.

3. The method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 1, characterized in that: Obtain the cross-scale causal edge set from the abnormal scale index set , (i, j) indicates that the fault propagates from scale i to scale j, satisfying , 、 is the anomaly score corresponding to scale i, j, Represents the anomaly score threshold; the construction logic of the causal edge set E includes: Anomaly scale screening: Based on the anomaly score and anomaly score threshold, the anomaly scale index set is screened; Causal edge generation: Based on the inverter fault propagation law, cross-scale causal relationships are predefined; causal edges are only constructed for abnormal scales belonging to the abnormal scale index set.

4. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 3, characterized in that: Anomaly score threshold The determination methods include static method and dynamic method. When the grid impedance fluctuation rate exceeds the threshold, the dynamic method is used, otherwise the static method is used. The static method calculates the statistical distribution of anomaly scores based on historical normal data, obtaining the sum of the mean and three standard deviations. The dynamic method calculates the quantile of anomaly scores in the past time period and uses it as the anomaly score threshold at the current time t.

5. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 4, characterized in that: Predicting the rate of change of features over time using physically constrained neural differential equations , to distinguish between transient interference and persistent faults, the dynamic equation of the physical constraint neuron differential equation is: in, The data-driven term for modeling a neural network takes as input the multi-scale latent feature vector x at time t and outputs the time-evolved gradient of the feature vector; is the causal edge weight, which indicates the strength of fault propagation from scale i to scale j; It represents the sensitivity of scale j features to scale i features and is calculated by back-propagation automatic differentiation technique; The inverter dynamic model equations include the phase-locked loop dynamic equations and the grid impedance equations; is the physical coupling coefficient, which is used to balance the contribution of data-driven terms and physical constraint terms.

6. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 5, characterized in that: Physical coupling coefficient From the initial coupling coefficient , input characteristic volatility and attenuation coefficient Jointly determined, the characteristic volatility includes the variance Var(x) of the multi-scale characteristic vector x, taking the attenuation coefficient The negative value of the product of the sum and variance Var(x) is an exponential function; The input variance Var(x) increases, the exponential term The absolute value of increases, resulting in a decrease in the exponential function value, a decrease in the γ value, a weakening of the physical constraint effect, and avoiding overfitting under noise interference; The input variance Var(x) decreases: the absolute value of the exponential term decreases, the exponential function value increases, the physical constraint effect is enhanced, the physical constraint neural differential equation more strictly follows the inverter dynamic equation, and the physical consistency of the diagnosis is improved; The way to obtain the physical coupling coefficient balances the weights of data-driven and physical constraints.

7. The method for diagnosing faults of a grid-connected inverter based on dynamic response according to claim 1, characterized in that: The adaptive diffusion kernel width is adjusted by combining the abnormal sensitivity coefficient and the grid impedance fluctuation rate; Calculate the time-frequency entropy change rate TER of the subband signal and obtain the time-frequency entropy change rate smoothing value TER by sliding average filtering MA ; Calculate the local variance ASV of the anomaly score by the formula Calculate the abnormal sensitivity coefficient ASC, where α is the adjustment factor; the width of the adaptive diffusion kernel is determined by the following formula Adjustment: in, represents the width of the adaptive diffusion kernel at time t, represents the initial kernel width; λ is the dynamic sensitivity coefficient; It represents the time domain derivative of the Tsallis entropy of the subband signal, reflecting the non-stationary rate of the subband signal; Represents a sliding average filter with a window length of Wt.

8. The method for diagnosing faults of a grid-connected inverter based on dynamic response according to claim 1, characterized in that: When constructing a cross-scale causal graph based on a structural causal model, the loss function of the causal variational autoencoder includes a causal regularization term. : in, The causal edge weight corresponding to the anomaly scale index (i, j) reflects the propagation intensity of the fault from scale i to j. It is predefined by the structural causal model and is a scalar weight. 、 is the output of the causal variational autoencoder 、 latent variables of multiple scales; Represents latent variables and Pearson correlation coefficient; Represents latent variables and The covariance of Represents the latent variables and The standard deviation of represents the square of the Frobenius norm, which is used to quantify the difference between weights and covariances; latent variables and It is the output of the causal variational autoencoder, an abstract representation of fault features at different time scales, and a low-dimensional potential feature obtained by nonlinearly mapping the spectral embedding vector.

9. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 8, characterized in that: The latent variables are low-dimensional feature representations of different time scales output by the causal variational autoencoder; The latent feature subset consists of multiple latent variables and is a set of features of abnormal scales screened out in the cross-scale causal graph; The latent variables are filtered by anomaly score threshold to obtain the latent feature subset; The multi-scale potential feature vector x is obtained through fusion, and the fault feature vector is obtained through evolution under physical constraints.

10. A method for fault diagnosis of a grid-connected inverter based on dynamic response according to claim 8, characterized in that: After the physical constraint neural differential equation generates a comprehensive fault feature vector, it is input into the classifier to output the probability distribution of each fault type; the maximum value in the probability distribution is taken as the current diagnosis confidence p; Set the preset threshold p th , if p <p th , the diagnosis confidence is determined to be insufficient, triggering the backtracking mechanism, including: The adaptive diffusion kernel parameters are adjusted by the following formula: in, 、 Represent the width of the adaptive diffusion kernel before and after adjustment; Represents the backtracking gain coefficient.

Images