Intelligent diagnosis method and system for mutual inductor based on fusion of reinforcement learning and transfer learning

Abstract

Embodiments of the present disclosure disclose a method and system for intelligent diagnosis of mutual inductors based on reinforcement learning and transfer learning fusion. The specific implementation of the method includes: acquiring multi-channel time sequence monitoring data of the mutual inductor to assemble a dynamic space-time state tensor, inputting a domain adversarial feature extraction network to generate a low-dimensional domain alignment hidden feature vector; assembling the low-dimensional domain alignment hidden feature vector and the historical window difference mean vector to construct a state vector, inputting the target domain diagnosis strategy network to output a diagnosis instruction; combining the fault classifier soft output, the source domain class center normalized distance and the strategy entropy to calculate a composite reward signal, using a curriculum learning gate factor to scale the strategy gradient and updating the network parameters. The implementation eliminates the cross-domain distribution offset, reduces the interference of random noise on strategy update, and realizes the stable convergence of the diagnosis strategy under the unlabeled working condition.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of power equipment diagnostic technology, specifically to an intelligent diagnostic method and system for instrument transformers based on the fusion of reinforcement learning and transfer learning. Background Technology

[0002] High-voltage instrument transformers are crucial devices in power systems for power metering and operational status monitoring. Their operational status directly impacts the safe operation of the power grid. Existing technologies typically employ neural networks to process multi-channel monitoring signals from instrument transformers and use feature extraction networks to map simulation data and field-acquired data into a low-dimensional shared latent space. The feature distributions of the two types of data are then aligned to reduce the distribution differences between the source and target domains. Furthermore, reinforcement learning methods can be introduced to adaptively output diagnostic commands based on state characteristics, thereby achieving online intelligent diagnosis of the instrument transformers. However, in actual operating environments, the number of labeled fault samples available on-site is limited. Environmental noise, load fluctuations, and transient impacts from the opening and closing of high-voltage switches can cause irregular offsets in the mapping of monitoring signals to the low-dimensional latent space. Existing methods usually assume that the feature distribution in the latent space remains stable and that the variation patterns in each direction are basically consistent. Therefore, directly using the static latent features at the current moment as the state input of the Markov decision process can lead to errors when non-stationary noise causes the direction of latent feature change to deviate from the actual deterioration direction of the equipment. This can result in errors in the feature alignment reward calculated at a fixed distance, continuously interfering with the strategy gradient and causing significant oscillations or failure to converge in the parameter updates of the diagnostic strategy network.

[0003] To improve feature alignment accuracy, existing technologies have employed methods such as increasing the depth of the feature extraction network or extending the sampling window. However, simply improving the network's feature representation ability is insufficient to promptly identify abnormal shifts in latent features caused by transient shocks, and it is difficult to suppress policy parameter update biases caused by erroneous rewards. Therefore, these methods still struggle to address the problem of unstable convergence of diagnostic strategies when latent space features are unstable. For example, Chinese invention patent CN120822135B discloses an intelligent diagnostic method and system for instrument transformers that integrates deep learning and transfer learning. It achieves feature distribution alignment by constructing a spatial positioning compensation mapping table and calculating the maximum mean difference (MMD). However, this scheme relies on static feature space assumptions and deterministic coordinate mapping. When faced with transient electromagnetic shocks such as the opening and closing of high-voltage switches in substations, the low-dimensional latent space is prone to nonlinear manifold distortion. Furthermore, this method lacks real-time verification of the latent feature evolution direction and cannot adaptively adjust the fitting weights through reinforcement learning strategies in the unlabeled target domain. This can easily lead to model diagnostic shifts and training oscillations due to invalid feature alignment.

[0004] Therefore, how to reduce the impact of non-stationary latent feature shifts on reward calculation and enable the diagnostic policy network to maintain stable convergence in a target domain unlabeled and highly disturbed operating environment is a problem that existing intelligent diagnostic technology for transformers needs to solve. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose an intelligent diagnostic method and system for instrument transformers based on the fusion of reinforcement learning and transfer learning, in order to solve the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide an intelligent diagnostic method for current transformers based on the fusion of reinforcement learning and transfer learning. The method includes: step S1, acquiring multi-channel time-series monitoring data of the current transformer to be diagnosed, and assembling a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window; Step S2: Input the known label dataset of the source domain and the unlabeled dynamic spatiotemporal state tensor of the target domain into the adversarial feature extraction network, map to generate a low-dimensional domain-aligned latent feature vector, and extract the source domain class center of the known label dataset of the source domain in the latent space. Step S3: Concatenate the low-dimensional domain aligned latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. Step S4: Input the state vector into the target domain diagnostic strategy network and output a diagnostic action command containing the fault type and degradation degree evaluation coefficient. Step S5: Determine the diagnostic guidance component based on the soft output of the fault classifier; determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain-aligned latent feature vector to the source domain class center; combine the policy entropy of the target domain diagnostic policy network output distribution; calculate the weighted composite reward signal; and decrease the weighting coefficient of the domain fitting component in the same direction when the cosine of the projection angle is negative; update the target domain diagnostic policy network parameters by scaling the composite reward signal using the course learning gating factor that is monotonically increasing based on the training rounds; and output the instrument transformer diagnostic result.

[0008] Preferably, in step S1, acquiring multi-channel time-series monitoring data includes: collecting the current transformer secondary side ratio error sequence, the partial discharge ultra-high frequency pulse frequency sequence, and the dielectric loss factor change rate sequence; assembling the dynamic spatiotemporal state tensor includes: standardizing and normalizing the multi-channel time-series monitoring data according to a preset sampling frequency of 100Hz, extracting sampling points according to a set sliding time window to form a three-channel monitoring matrix, and splicing the three-channel monitoring matrix at the current sampling time with the time-domain differential matrix of multiple consecutive windows along the channel dimension to generate a dynamic spatiotemporal state tensor.

[0009] Preferably, in step S2, the domain adversarial feature extraction network includes a feature extractor, a fault classifier, and a domain discriminator. The process of mapping and generating low-dimensional domain-aligned latent feature vectors and extracting source domain class centers includes: inputting known-label samples from the source domain and unlabeled dynamic spatiotemporal state tensors from the target domain into the feature extractor in parallel to obtain latent feature vectors from the source domain and the target domain; calculating the mean of the latent feature vectors from the source domain according to the label category to generate source domain class centers; inputting the latent feature vectors from the source domain into the fault classifier to calculate the classification cross-entropy loss, and simultaneously inputting the latent feature vectors from the source domain and the target domain into the domain discriminator to calculate the domain discriminant binary classification loss; using a gradient reversal layer to apply a negative adversarial coefficient to the gradient of the domain discriminator during backpropagation, alternately updating the parameters of the feature extractor and the domain discriminator, and outputting low-dimensional domain-aligned latent feature vectors upon convergence.

[0010] Preferably, the parameter update process of the feature extractor further includes: introducing a maximum mean difference loss term based on a kernel function on the basis of classification cross-entropy loss and domain discriminant binary classification loss with negative adversarial coefficient; constructing a composite transfer loss function by weighting the maximum mean difference loss term with a preset penalty weight coefficient; reducing the distribution distance of latent features of the source domain and target domain in the regenerating kernel Hilbert space by minimizing the composite transfer loss function, so that the generated low-dimensional domain-aligned latent feature vector has domain independence while preserving the separability between fault classes.

[0011] Preferably, in step S3, the process of constructing the state vector and calculating the cosine value of the projection angle includes: step S31, obtaining the low-dimensional domain aligned latent feature vector at the current sampling time, extracting the latent feature vectors under multiple consecutive previous historical sampling windows, calculating the difference between the latent feature vectors of adjacent historical windows, and taking the arithmetic mean of the obtained difference vectors to obtain the first-order difference mean vector; step S32, performing vector concatenation between the low-dimensional domain aligned latent feature vector at the current sampling time and the first-order difference mean vector to generate the state vector; step S33, taking the sliding weighted mean of the latent feature difference vectors of multiple consecutive previous historical sampling windows according to the time decay weight, generating the historical tangent direction vector, calculating the dot product of the first-order difference mean vector and the historical tangent direction vector, dividing by the product of their magnitudes to obtain the cosine value of the projection angle.

[0012] Preferably, in step S5, the process of weighted calculation of the composite reward signal includes: under the unlabeled sampling condition in the target domain, inputting the state vector into the fault classifier to obtain the classification probability distribution of each type of fault, and multiplying the logarithm of the output class with the highest probability by the severity evaluation coefficient in the diagnostic action instruction as the diagnostic guidance component; calculating the reciprocal of the Euclidean distance between the state vector at the current sampling time and the class center of the source domain and performing normalization processing as the domain fitting component; calculating the policy entropy of the output action distribution of the target domain diagnostic policy network as the entropy regularization term; when the cosine value of the projection angle is greater than or equal to 0, combining the diagnostic guidance component, the domain fitting component, and the policy entropy according to the preset weights; when the cosine value of the projection angle is negative, adding the cosine value of the projection angle to 1 as the attenuation ratio, and summing the preset weights of the domain fitting component after proportional reduction in the same direction to generate the composite reward signal.

[0013] Preferably, in step S4, the diagnostic action command is composed of a combination of discrete fault types and continuous severity evaluation coefficients; wherein, the discrete fault types include the normal state of the transformer, the magnetic saturation state of the core, the partial discharge state of the insulation, and the ratio error drift state, and the continuous severity evaluation coefficients are continuous real numbers between 0.0 and 1.0; the target domain diagnostic strategy network adopts an Actor-Critic architecture, which includes a strategy network for outputting action distribution and a value network for estimating state value.

[0014] Preferably, the parameter update process of the target domain diagnostic policy network adopts the near-end policy optimization algorithm; by calculating the ratio of the new policy probability to the old policy probability, the magnitude is constrained by the preset upper and lower limit comparison values ​​of pruning to prevent the single-step policy update magnitude from being too large; the product of the pruned ratio and the generalized advantage estimate is used as the update target, and the gradient value after scaling by the course learning gating factor is combined with the adaptive moment estimation optimizer to iteratively update the weights of the policy network and the value network.

[0015] Preferably, in step S5, the process of outputting the instrument transformer diagnostic results includes: when the course learning gating factor reaches a preset truncation threshold and the smooth mean fluctuation range of the composite reward signal is less than a preset convergence threshold in multiple consecutive training rounds, stopping the network parameter iteration and locking the model parameters of the target domain diagnostic strategy network; converting the instrument transformer monitoring data collected in real time at the substation into a state vector input to the locked target domain diagnostic strategy network, and outputting the diagnostic results containing the determined fault category and quantitative degradation coefficient to the edge computing node in real time.

[0016] A current transformer intelligent diagnostic system based on the fusion of reinforcement learning and transfer learning, which is used to implement a current transformer intelligent diagnostic method based on the fusion of reinforcement learning and transfer learning, includes: The data acquisition and tensor assembly module is used to acquire multi-channel time-series monitoring data of the current transformer to be diagnosed and assemble a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window. The feature extraction and class center construction module is used to map the source domain known label dataset and the target domain unlabeled dynamic spatiotemporal state tensor input domain adversarial feature extraction network to generate low-dimensional domain aligned latent feature vectors, and extract the source domain class centers of the source domain known label dataset in the latent space. The state vector construction and tangent cosine calculation module is used to align the low-dimensional domain of the latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. The diagnostic strategy output module is used to input the state vector into the target domain diagnostic strategy network and output diagnostic action instructions containing fault type and degradation degree evaluation coefficients. The composite reward calculation and policy update module is used to determine the diagnostic guidance component based on the soft output of the fault classifier, determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain aligned latent feature vector to the source domain class center, combine the policy entropy of the target domain diagnostic policy network output distribution, calculate the composite reward signal in a weighted manner, and attenuate the weighting coefficients of the domain fitting component in the same direction when the cosine of the projection angle is negative, use the policy gradient derived from the composite reward signal by scaling the course learning gating factor based on the monotonically increasing training rounds, update the parameters of the target domain diagnostic policy network, and output the instrument transformer diagnostic results.

[0017] The above-described embodiments of this disclosure have the following beneficial effects: 1. In the intelligent diagnosis of instrument transformers using transfer learning fusion, by constructing a dynamic spatiotemporal state tensor and using a domain adversarial feature extraction network to align the features of labeled samples in the source domain and unlabeled data in the target domain, the feature distribution shift caused by environmental noise, load fluctuations, and differences in operating conditions can be reduced. On this basis, the low-dimensional domain-aligned latent feature vector at the current moment and the first-order difference mean vector of the historical window are used together as the state vector, so that the diagnostic state not only includes the current operating characteristics of the instrument transformer, but also reflects its degradation trend over time, providing a more stable, complete, and easily distinguishable state input for the diagnostic strategy network.

[0018] 2. By calculating the cosine of the projection angle between the first-order difference mean vector and the historical tangent direction vector, it is possible to determine whether the current change direction of the latent feature is consistent with the previous change trend. When transient impacts such as the opening and closing of high-voltage switches cause abnormal deviations in the latent feature, such deviations can be identified in time, avoiding the misinterpretation of random interference as a continuous deterioration of the transformer state, thereby improving the ability to distinguish between real fault changes and occasional noise under strong interference environment.

[0019] 3. Linking the cosine of the projection angle with the weighting coefficient of the domain fitting component reduces the influence of the domain fitting component on the composite reward signal when the direction of latent feature change is abnormal. This can reduce the interference of erroneous distance information on the policy gradient, thereby avoiding deviation of the parameter update direction caused by transient noise in the diagnostic policy network, reducing the risk of oscillation and divergence during training, and making the policy update in the unlabeled target domain more stable. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0021] Figure 1 This is a flowchart of the current transformer fault diagnosis method of the present invention; Figure 2 This is a diagram illustrating the composite reward calculation and strategy network update of the present invention. Figure 3 This is a graph showing the change of the learning gating factor of this invention with the number of training rounds. Detailed Implementation

[0022] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0023] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0024] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0025] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0026] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0027] Before performing any of the operations involving the collection, storage, processing, or use of the target images disclosed herein, the relevant organizations or individuals shall fulfill their obligations, including conducting information security impact assessments, informing the information subjects, and obtaining prior authorization and consent from the information subjects.

[0028] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] refer to Figure 1 The flowchart illustrates some embodiments of an intelligent diagnostic method for instrument transformers based on the fusion of reinforcement learning and transfer learning according to this disclosure. This intelligent diagnostic method for instrument transformers based on the fusion of reinforcement learning and transfer learning includes the following steps: Step S1: Obtain multi-channel time-series monitoring data of the current transformer to be diagnosed, and assemble a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window; Step S2: Input the known label dataset of the source domain and the unlabeled dynamic spatiotemporal state tensor of the target domain into the adversarial feature extraction network, map to generate a low-dimensional domain-aligned latent feature vector, and extract the source domain class center of the known label dataset of the source domain in the latent space. Step S3: Concatenate the low-dimensional domain aligned latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. Step S4: Input the state vector into the target domain diagnostic strategy network and output a diagnostic action command containing the fault type and degradation degree evaluation coefficient. Step S5: Determine the diagnostic guidance component based on the soft output of the fault classifier; determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain-aligned latent feature vector to the source domain class center; combine the policy entropy of the target domain diagnostic policy network output distribution; calculate the weighted composite reward signal; and decrease the weighting coefficient of the domain fitting component in the same direction when the cosine of the projection angle is negative; update the target domain diagnostic policy network parameters by scaling the composite reward signal using the course learning gating factor that is monotonically increasing based on the training rounds; and output the instrument transformer diagnostic result.

[0030] Preferably, in step S1, acquiring multi-channel time-series monitoring data includes: collecting the current transformer secondary side ratio error sequence, the partial discharge ultra-high frequency pulse frequency sequence, and the dielectric loss factor change rate sequence; assembling the dynamic spatiotemporal state tensor includes: standardizing and normalizing the multi-channel time-series monitoring data according to a preset sampling frequency of 100Hz, extracting sampling points according to a set sliding time window to form a three-channel monitoring matrix, and splicing the three-channel monitoring matrix at the current sampling time with the time-domain differential matrix of multiple consecutive windows along the channel dimension to generate a dynamic spatiotemporal state tensor.

[0031] Preferably, in step S2, the domain adversarial feature extraction network includes a feature extractor, a fault classifier, and a domain discriminator. The process of mapping and generating low-dimensional domain-aligned latent feature vectors and extracting source domain class centers includes: inputting known-label samples from the source domain and unlabeled dynamic spatiotemporal state tensors from the target domain into the feature extractor in parallel to obtain latent feature vectors from the source domain and the target domain; calculating the mean of the latent feature vectors from the source domain according to the label category to generate source domain class centers; inputting the latent feature vectors from the source domain into the fault classifier to calculate the classification cross-entropy loss, and simultaneously inputting the latent feature vectors from the source domain and the target domain into the domain discriminator to calculate the domain discriminant binary classification loss; using a gradient reversal layer to apply a negative adversarial coefficient to the gradient of the domain discriminator during backpropagation, alternately updating the parameters of the feature extractor and the domain discriminator, and outputting low-dimensional domain-aligned latent feature vectors upon convergence.

[0032] Preferably, the parameter update process of the feature extractor further includes: introducing a maximum mean difference loss term based on a kernel function on the basis of classification cross-entropy loss and domain discriminant binary classification loss with negative adversarial coefficient; constructing a composite transfer loss function by weighting the maximum mean difference loss term with a preset penalty weight coefficient; reducing the distribution distance of latent features of the source domain and target domain in the regenerating kernel Hilbert space by minimizing the composite transfer loss function, so that the generated low-dimensional domain-aligned latent feature vector has domain independence while preserving the separability between fault classes.

[0033] Preferably, in step S3, the process of constructing the state vector and calculating the cosine value of the projection angle includes: step S31, obtaining the low-dimensional domain aligned latent feature vector at the current sampling time, extracting the latent feature vectors under multiple consecutive previous historical sampling windows, calculating the difference between the latent feature vectors of adjacent historical windows, and taking the arithmetic mean of the obtained difference vectors to obtain the first-order difference mean vector; step S32, performing vector concatenation between the low-dimensional domain aligned latent feature vector at the current sampling time and the first-order difference mean vector to generate the state vector; step S33, taking the sliding weighted mean of the latent feature difference vectors of multiple consecutive previous historical sampling windows according to the time decay weight, generating the historical tangent direction vector, calculating the dot product of the first-order difference mean vector and the historical tangent direction vector, dividing by the product of their magnitudes to obtain the cosine value of the projection angle.

[0034] Preferably, in step S5, the process of weighted calculation of the composite reward signal includes: under the unlabeled sampling condition in the target domain, inputting the state vector into the fault classifier to obtain the classification probability distribution of each type of fault, and multiplying the logarithm of the output class with the highest probability by the severity evaluation coefficient in the diagnostic action instruction as the diagnostic guidance component; calculating the reciprocal of the Euclidean distance between the state vector at the current sampling time and the class center of the source domain and performing normalization processing as the domain fitting component; calculating the policy entropy of the output action distribution of the target domain diagnostic policy network as the entropy regularization term; when the cosine value of the projection angle is greater than or equal to 0, combining the diagnostic guidance component, the domain fitting component, and the policy entropy according to the preset weights; when the cosine value of the projection angle is negative, adding the cosine value of the projection angle to 1 as the attenuation ratio, and summing the preset weights of the domain fitting component after proportional reduction in the same direction to generate the composite reward signal.

[0035] Preferably, in step S4, the diagnostic action command is composed of a combination of discrete fault types and continuous severity evaluation coefficients; wherein, the discrete fault types include the normal state of the transformer, the magnetic saturation state of the core, the partial discharge state of the insulation, and the ratio error drift state, and the continuous severity evaluation coefficients are continuous real numbers between 0.0 and 1.0; the target domain diagnostic strategy network adopts an Actor-Critic architecture, which includes a strategy network for outputting action distribution and a value network for estimating state value.

[0036] Preferably, the parameter update process of the target domain diagnostic policy network adopts the near-end policy optimization algorithm; by calculating the ratio of the new policy probability to the old policy probability, the magnitude is constrained by the preset upper and lower limit comparison values ​​of pruning to prevent the single-step policy update magnitude from being too large; the product of the pruned ratio and the generalized advantage estimate is used as the update target, and the gradient value after scaling by the course learning gating factor is combined with the adaptive moment estimation optimizer to iteratively update the weights of the policy network and the value network.

[0037] Preferably, in step S5, the process of outputting the instrument transformer diagnostic results includes: when the course learning gating factor reaches a preset truncation threshold and the smooth mean fluctuation range of the composite reward signal is less than a preset convergence threshold in multiple consecutive training rounds, stopping the network parameter iteration and locking the model parameters of the target domain diagnostic strategy network; converting the instrument transformer monitoring data collected in real time at the substation into a state vector input to the locked target domain diagnostic strategy network, and outputting the diagnostic results containing the determined fault category and quantitative degradation coefficient to the edge computing node in real time.

[0038] A current transformer intelligent diagnostic system based on the fusion of reinforcement learning and transfer learning, which is used to implement a current transformer intelligent diagnostic method based on the fusion of reinforcement learning and transfer learning, includes: The data acquisition and tensor assembly module is used to acquire multi-channel time-series monitoring data of the current transformer to be diagnosed and assemble a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window. The feature extraction and class center construction module is used to map the source domain known label dataset and the target domain unlabeled dynamic spatiotemporal state tensor input domain adversarial feature extraction network to generate low-dimensional domain aligned latent feature vectors, and extract the source domain class centers of the source domain known label dataset in the latent space. The state vector construction and tangent cosine calculation module is used to align the low-dimensional domain of the latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. The diagnostic strategy output module is used to input the state vector into the target domain diagnostic strategy network and output diagnostic action instructions containing fault type and degradation degree evaluation coefficients. The composite reward calculation and policy update module is used to determine the diagnostic guidance component based on the soft output of the fault classifier, determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain aligned latent feature vector to the source domain class center, combine the policy entropy of the target domain diagnostic policy network output distribution, calculate the composite reward signal in a weighted manner, and attenuate the weighting coefficients of the domain fitting component in the same direction when the cosine of the projection angle is negative, use the policy gradient derived from the composite reward signal by scaling the course learning gating factor based on the monotonically increasing training rounds, update the parameters of the target domain diagnostic policy network, and output the instrument transformer diagnostic results.

[0039] Example 1: When the high-voltage current transformer in the intelligent distribution system of a 220kV substation is running online, the high-voltage switch opening and closing will generate high-frequency random pulses, and the load will fluctuate over time. The secondary side ratio error sequence of the current transformer, the frequency sequence of partial discharge ultra-high-frequency pulses, and the dielectric loss factor change rate sequence are collected to form multi-channel time-series monitoring data. Each sequence is standardized and normalized according to a 100Hz sampling frequency. 100 sampling points are extracted using a 1s sliding time window to form a three-channel monitoring matrix. For the three-channel monitoring matrix of the previous four consecutive historical sliding windows, the difference values ​​between corresponding elements in adjacent windows are calculated sequentially to generate four... The first-order temporal differential matrix has a dimension of 3 rows and 100 columns. Then, the three-channel monitoring matrix with a dimension of 3 rows and 100 columns at the current sampling time is stacked and cascaded with the four first-order temporal differential matrices with a dimension of 3 rows and 100 columns, while keeping the temporal topology of the 100 discrete sampling points unchanged. This constructs a two-dimensional data tensor with a dimension of 15 channels multiplied by 100 sampling points, that is, a data matrix with a feature dimension of 1500. This accurately describes the deterministic topological rules of the dynamic spatiotemporal state tensor in the discrete storage structure and data channel expansion, and assembles it to obtain a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window.

[0040] Using a laboratory-simulated accelerated aging dataset as the source domain's known-label dataset and substation-collected data as the target domain's unlabeled data, the two types of data are input in parallel into a domain adversarial feature extraction network. This network includes a one-dimensional convolutional neural network feature extractor based on residual structures, a fault classifier, and a domain discriminator. The dynamic spatiotemporal state tensors of the source domain's known-label dataset and the target domain's unlabeled dataset are mapped to source domain latent feature vectors and target domain latent feature vectors, respectively, by the feature extractor. The mean of the source domain latent feature vectors is calculated according to the label category to obtain 64-dimensional source domain class centers. During training, the classification cross-entropy loss, domain discriminant binary classification loss, and maximum mean difference loss are combined into a composite transfer loss function. A negative adversarial coefficient of 0.15 is applied to the domain discriminator through a gradient inversion layer, and the maximum mean difference loss term is weighted with a penalty weight of 0.05. The parameters of the feature extractor and the domain discriminator are updated alternately to obtain a 64-dimensional low-dimensional domain-aligned latent feature vector with a L2 value of 1.414.

[0041] Low-dimensional aligned latent feature vectors are extracted from the five consecutive historical sampling windows preceding the current sampling time. The differences between the latent feature vectors of adjacent historical windows are calculated sequentially, and the arithmetic mean of the resulting difference vectors is taken to obtain a 64-dimensional first-order difference mean vector with a second norm of 0.500. The 64-dimensional low-dimensional aligned latent feature vector at the current sampling time is concatenated with this first-order difference mean vector to construct the state vector of a 128-dimensional Markov decision process. For the 64-dimensional latent feature difference vectors of the four adjacent windows extracted from the previous five consecutive historical sampling windows, deterministic discrete weight coefficients are sequentially assigned, monotonically decreasing from near to far in terms of time distance. Specifically, the maximum weight coefficient is set to... The weights of the most recent difference vectors are 0.40, the second most recent is 0.30, the third is 0.20, and the furthest is 0.10, with the sum of the four discrete weight coefficients always equal to 1.00. Each latent feature difference vector is multiplied by its corresponding discrete weight coefficient, and then summed to obtain a 64-dimensional discrete moving weighted mean vector. This vector serves as a deterministic entity reflecting the discrete evolution trend of low-dimensional vectors in the latent feature space, generating a historical tangent direction vector with a second norm of 0.400. The dot product of the first-order difference mean vector and the historical tangent direction vector is 0.160. Dividing this dot product by the product of the magnitudes of the two vectors yields the cosine of the projection angle. .

[0042] exist At that time, the direction of change of the current latent feature is consistent with the trend of the previous change, and the dynamic domain fitting weights are... Maintaining a value of 0.200, when the circuit breaker closes and generates a sudden high-frequency random pulse, the dot product of the first-order difference mean vector at another sampling time and the historical tangent direction vector is -0.080, corresponding to... ;Will The sum of 0.600 obtained by adding 1 is used as the attenuation ratio, and the dynamic domain fitting weights are used for this purpose. Reducing the value from 0.200 to 0.120 decreases the weighted contribution of the domain fitting component in the composite reward signal from -0.0462 to -0.0277.

[0043] The 128-dimensional state vector is input into the target domain diagnostic strategy network using the Actor-Critic architecture. This network employs a near-end policy optimization algorithm, with both the internal policy network and value network having two fully connected hidden layers and using the LeakyReLU activation function. The policy network outputs diagnostic action commands composed of discrete fault types and continuous severity evaluation coefficients. The discrete fault types include, in order, the transformer normal state, the core magnetic saturation state, the insulation partial discharge state, and the ratio error drift state. The continuous severity evaluation coefficients are continuous real numbers between 0.0 and 1.0.

[0044] In the fault classifier of the domain adversarial feature extraction network, an input mapping layer corresponding to the input dimension is set before the shared classification layer. During source domain training, the 64-dimensional source domain latent feature vector enters the shared classification layer through the corresponding input mapping layer to calculate the classification cross-entropy loss. During target domain diagnosis, the 128-dimensional state vector is mapped to 64-dimensional classification features through another input mapping layer and then enters the same shared classification layer to output the classification probability distribution of each fault type, so that the current features and historical changes in the state vector participate in fault classification.

[0045] Under the unlabeled sampling condition in the target domain, the state vector is input into the fault classifier to obtain the classification probability distribution of the four fault types. After extracting the probability value of the class with the highest output probability, its range is truncated to between 0.01 and 1.00. Then, the natural logarithm of this truncated probability value is calculated to deterministically limit its value to the interval between -4.605 and 0.000. Next, the natural logarithm value is added to the value of 4.605 and a baseline shift is performed to convert it to a value between 0.000 and 4.605. The nonnegative bounded scalar is then normalized and scaled by dividing the translated nonnegative scalar by 4.605, mapping it to the dimensionless interval of 0.000 to 1.000. Finally, the normalized nonnegative guidance scalar is multiplied by the continuous severity assessment coefficient between 0.000 and 1.000 in the diagnostic action command to obtain the diagnostic guidance component. When calculating the domain fitting component, the first 64-dimensional sub-vectors are precisely decoupled from the 128-dimensional state vector, i.e., the 64-dimensional low-dimensional domain-aligned latent feature vector corresponding to the current sampling time, and discarded. The 64-dimensional first-order difference mean sub-vector is then used. The extracted 64-dimensional low-dimensional domain aligned latent feature vectors are then compared with the 64-dimensional source domain class center vectors constructed under the same 64-dimensional latent space topology. The Euclidean distance in the same dimension is calculated for each vector. The smallest distance value is found, and its reciprocal is calculated. This reciprocal of the nearest source domain class center distance is then normalized. The policy entropy is calculated based on the action distribution output by the target domain diagnostic policy network. The policy entropy value is then divided by the maximum achievable entropy value of 1.386 for the four types of faults. The normalized values ​​are scaled to the range of 0.000 to 1.000. The normalized non-negative diagnostic guidance component is assigned a preset weighting coefficient of 0.50, the normalized domain fitting component is assigned a preset weighting coefficient of 0.30, and the normalized policy entropy is assigned a preset weighting coefficient of 0.20. The dimensionless components multiplied by their respective weighting coefficients are then summed using a weighted method to construct a minimal single-step scalar whose numerical boundary is precisely locked within the dimensionless range of 0.000 to 1.000, thus obtaining the single-step composite reward signal. .

[0046] The course learning gating factor is calculated using a defined logistic normalization function with each training round. The specific calculation process is as follows: First, divide the current training round number by the maximum training round number (1000) to obtain a normalized round progress between 0.00 and 1.00. Next, subtract the midpoint offset of 0.50 from the normalized round progress and multiply by the growth rate constant 5.00 to obtain the exponential variable. Then, calculate the negative exponential power of the natural constant e, add 1.00 to this value as the denominator, and perform a division operation with 1.00 as the numerator to obtain a dimensionless gate between 0.000 and 1.000. The scalar control is set to a maximum of 1000 training epochs with a growth rate parameter of 5.0. A low gating factor is maintained for the first 200 epochs to limit the policy update magnitude in the target domain. As training progresses, the gating factor is gradually increased. After calculating the policy network loss based on the composite reward signal and the generalized advantage estimate, backpropagation is used to obtain the original target gradient of the policy network weight parameters. Before feeding this gradient into the adaptive moment estimator optimizer, the course learning gating factor corresponding to the current training epoch is multiplied element-wise with each component of the original target gradient to update the policy network and value network parameters with the scaled gradient values. The learning rate is set to 0.0003.

[0047] When the course learning gating factor reaches the preset truncation threshold, and the smoothed mean fluctuation range of the composite reward signal is less than the preset convergence threshold in multiple consecutive training rounds, parameter iteration stops, the model parameters of the target domain diagnostic strategy network are locked, and it is deployed on the substation edge computing node. For the standardized feature matrix of the target domain at the time of diagnosis, the four discrete fault probabilities output by the strategy network are 0.02, 0.03, 0.91, and 0.04, respectively, and the continuous severity evaluation coefficient is 0.85. Based on this, the agent generates diagnostic action instructions, determines the fault category as insulation partial discharge state, the quantitative degradation coefficient is 0.85, the corresponding diagnostic guidance component is -0.080, the Euclidean distance from the current low-dimensional domain aligned latent feature vector to the nearest source domain class center is 0.231, the policy entropy is 0.385, and finally the composite reward signal is obtained. The edge computing nodes output the fault category and quantitative degradation coefficient in real time.

[0048] Example 2: The current intelligent power distribution system test platform consists of a high-voltage current transformer physical test bench and a power system digital simulation node. It is used to test the fault diagnosis results and strategy convergence process when the target domain is unlabeled and there is strong noise interference. The test data includes a source domain known label dataset formed by laboratory simulated accelerated aging fault library, and a target domain unlabeled dataset formed by substation field collection. The multi-channel time-series monitoring data includes the current transformer secondary side ratio error sequence, partial discharge ultra-high frequency pulse frequency sequence, and dielectric loss factor change rate sequence. Among them, the measurement range of ratio error is -1.0% to +1.0%, the measurement range of partial discharge ultra-high frequency pulse frequency is 0 to 1000 times per second, and the measurement range of dielectric loss factor change rate is -0.05% to +0.05% per hour.

[0049] The partial discharge ultra-high frequency pulse frequency sequence is not directly acquired from the raw ultra-high frequency electromagnetic wave waveform at the hundreds of megahertz level. The front-end hardware demodulation unit performs peak detection and pulse counting on the raw signal output by the ultra-high frequency sensor, converting the nanosecond-level high-frequency pulses into discrete frequency values ​​per unit time. The secondary side ratio error sequence and the dielectric loss factor change rate sequence are processed by envelope detection and low-pass filtering, and their highest effective change frequency is limited to within 50Hz. The sampling frequency is set to 100Hz, and 100 sampling points are intercepted in the sliding time window, corresponding to a window duration of 1s. The single processing delay of the edge computing node is controlled within 50ms. Gaussian white noise with a signal-to-noise ratio of 20dB and 50Hz power frequency harmonic disturbance are superimposed on the target domain signal, and a high-frequency random pulse with a peak value of 500V generated by the closing of the high-voltage circuit breaker is injected.

[0050] During the training of the domain adversarial feature extraction network, the adversarial gradient inversion coefficient is set to 0.15. When the coefficient is below 0.05, there is still a significant distribution shift between the latent features of the source domain and the latent features of the target domain. When the coefficient is above 0.30, the negative gradient generated by backpropagation will weaken the feature extractor's ability to distinguish fault features. Therefore, 0.15 is used as the adversarial gradient inversion coefficient in this embodiment. The maximum mean difference penalty weight is set to 0.05, which is determined by measuring the distribution distance of 500 different samples in the regenerating kernel Hilbert space. When the penalty weight is below 0.01, the constraint on the higher-order moment distribution is insufficient. When the penalty weight is above 0.10, the latent feature distribution becomes overly smooth. Under the condition of 20dB signal-to-noise ratio, taking 0.05 can balance the separability between fault classes and domain independence.

[0051] The maximum number of training epochs for the target domain diagnostic policy network is set to 1000 epochs. The growth rate parameter of the course learning gating factor is set to 5.0. The course learning gating factor increases monotonically with the number of training epochs according to an S-shaped curve. It remains below 0.15 for the first 200 epochs to limit the magnitude of the policy gradient in the early stage of training. When training reaches 1000 epochs, the gating factor increases to above 0.98, so that the target domain diagnostic policy network gradually transitions to target domain adaptive updates.

[0052] After standardization and normalization, the target domain three-channel monitoring matrix is ​​concatenated with the temporal differential matrices of multiple consecutive windows along the channel dimension to form a dynamic spatiotemporal state tensor. This tensor is processed by a one-dimensional convolutional neural network feature extractor to obtain a 64-dimensional low-dimensional domain-aligned latent feature vector with a second norm of 1.414. The low-dimensional domain-aligned latent feature vectors of the previous five consecutive historical sampling windows are extracted. The difference between the latent feature vectors of adjacent historical windows is calculated and the arithmetic mean is obtained to obtain a 64-dimensional first-order difference mean vector with a second norm of 0.500. The low-dimensional domain-aligned latent feature vector at the current sampling time is concatenated with the first-order difference mean vector to construct the state vector of a 128-dimensional Markov decision process.

[0053] The sliding weighted mean of the latent feature difference vectors of continuous historical windows is calculated using time decay weights to obtain the historical tangent direction vector, whose second norm is 0.400. Under stationary conditions, the dot product of the first-order difference mean vector and the historical tangent direction vector is 0.160. The cosine of the projection angle is then calculated. Dynamic domain fitting weights Maintaining a value of 0.200, after the circuit breaker closes and a high-frequency random pulse is introduced, the dot product of the two vectors becomes -0.080, corresponding to... ;by The sum of 1 and 1 is used as the attenuation ratio, and the dynamic domain fitting weights are used as the weights. The weighted contribution of the domain fitting component decreased from -0.0462 to -0.0277 as it was reduced from 0.200 to 0.120.

[0054] A 128-dimensional state vector is input into the target domain diagnostic strategy network, which outputs diagnostic action instructions composed of discrete fault types and continuous severity evaluation coefficients. When acquiring the classification probability distribution of each fault type, the state vector undergoes input mapping processing to match the input dimension of the fault classifier, forming a 64-dimensional classification feature. This feature is then fed into the fault classifier, and the product of the logarithm of the highest output probability class and the continuous severity evaluation coefficient is obtained to yield the diagnostic guidance component. When calculating the domain fitting component, the 64-dimensional low-dimensional domain-aligned latent feature vector at the current sampling time is read from the state vector. The reciprocal of the Euclidean distance from this vector to the corresponding 64-dimensional source domain class center is calculated and normalized. The current diagnostic guidance component is -0.080, the Euclidean distance from the low-dimensional domain-aligned latent feature vector to the source domain class center is 0.231, and the strategy entropy of the target domain diagnostic strategy network's output action distribution is 0.385. After combining the diagnostic guidance component, domain fitting component, and strategy entropy according to preset weights, a single-step composite reward signal is obtained. .

[0055] In sample group 1 of this invention, the signal-to-noise ratio of the target domain signal is 20dB, and the cosine value of the projection angle is... The dynamic domain fitting weight is 0.800. The value is 0.200, representing a single-step composite reward signal. With a value of -0.122, the target domain diagnostic strategy network converged after 412 iterations, achieving a fault diagnosis accuracy of 96.5% and a mean square error of 0.038 for the continuous severity assessment coefficient.

[0056] In sample group 2 of this invention, the signal-to-noise ratio of the target domain signal is maintained at 20dB, and the high-frequency random pulse causes the cosine value of the projection angle to be... It dropped to -0.400, according to The sum of 1 and 1 reduces the fitting weights in the preset domain in the same direction; the fitting weights in the dynamic domain... The value was adjusted from 0.200 to 0.120, and the single-step composite reward signal was changed. With a mean square error of -0.104, the target domain diagnostic strategy network converged after 438 iterations, achieving a fault diagnosis accuracy of 95.8% and a mean square error of 0.042 for the continuous severity assessment coefficient.

[0057] In sample group 3 of this invention, the signal-to-noise ratio of the target domain signal drops to 10dB, and the cosine value of the projection angle... The value is -0.650. Following the same weight decay relationship, the dynamic domain fit weights are... Adjusted to 0.070, the resulting single-step composite reward signal With a mean square error of -0.091, the target domain diagnostic strategy network converged after 485 iterations, achieving a fault diagnosis accuracy of 94.2% and a mean square error of 0.051 for the continuous severity assessment coefficient.

[0058] In control group one, the linkage between the cosine value of the projection angle and the weighting coefficient of the domain fitting component was eliminated, and the signal-to-noise ratio was 20dB. At that time, dynamic domain fitting weights The value remains at 0.200, indicating a single-step compound reward signal. With a fixed weight of -0.126, the distance deviation generated by the high-frequency random pulse continues to enter the policy gradient. The target domain diagnostic policy network converges after 825 iterations, with a fault diagnosis accuracy of 87.3% and a mean square error of 0.115 for the continuous severity assessment coefficient.

[0059] Control group 2, without domain adversarial feature alignment processing, at a signal-to-noise ratio of 20dB, the cosine value of the projection angle. The dynamic domain fitting weight is 0.780. The value is 0.200, representing a single-step composite reward signal. The feature distribution offset between the source and target domains is -0.345, which is preserved in the low-dimensional latent space. The policy training converges after 980 iterations. The fault diagnosis accuracy is 72.1%, and the mean square error of the continuous severity assessment coefficient is 0.210.

[0060] In control group three, the scaling of the policy gradient by the course learning gating factor is removed. At a signal-to-noise ratio of 20 dB, the cosine value of the projection angle is... The dynamic domain fitting weight is 0.800. The value is 0.200, representing a single-step composite reward signal. The value is -0.210. The policy gradient in the early stage of training is directly used to update the network parameters. The target domain diagnostic policy network converges after 750 iterations. The fault diagnosis accuracy is 83.6%, and the mean square error of the continuous severity evaluation coefficient is 0.142.

[0061] In the out-of-range control group 1, the domain adversarial feature extraction network uses adversarial gradient reversal coefficients exceeding the effective value range and maximum mean difference penalty weights. The signal-to-noise ratio in this group is 20 dB, and the cosine of the projection angle is... The dynamic domain fitting weight is 0.810. The value is 0.200, representing a single-step composite reward signal. The value is -0.185, indicating that the classification boundary of the latent features is subject to strong domain alignment constraints. The target domain diagnostic strategy network converges after 680 iterations, with a fault diagnosis accuracy of 85.1% and a mean square error of 0.128 for the continuous severity assessment coefficient.

[0062] In the out-of-range control group 2, the signal-to-noise ratio was 20 dB, and the cosine value of the projection angle was... The dynamic domain fitting weight is 0.790. The value is 0.200, representing a single-step composite reward signal. The target domain diagnostic strategy network converged after 910 iterations with a value of -0.290, achieving a fault diagnosis accuracy of 76.4% and a mean square error of 0.188 for the continuous severity assessment coefficient.

[0063] At the same 20dB signal-to-noise ratio and Under the conditions, sample group 2 of the present invention will fit the dynamic domain weights. The value was reduced to 0.120, while the control group 1 maintained 0.200. Correspondingly, the number of strategy convergence rounds increased from 438 to 825, and the fault diagnosis accuracy decreased from 95.8% to 87.3%. When the present invention's sample group 1 was used as a reference for stable operating conditions, its strategy convergence rounds were 412, while the control group 1 with fixed domain fitting weights was 825. After canceling the domain adversarial feature alignment processing, the fault diagnosis accuracy of the control group 2 decreased to 72.1%. After canceling the course learning gating factor, the strategy convergence rounds of the control group 3 were extended to 750. The adversarial gradient reversal coefficient and the maximum mean difference penalty weight in the out-of-range control group 1 exceeded the effective value range, and the fault diagnosis accuracy was 85.1%. When the signal-to-noise ratio dropped to 10dB, the present invention's sample group 3 changed the dynamic domain fitting weights. The value was adjusted to 0.070, the fault diagnosis accuracy remained at 94.2%, and the mean square error of the continuous severity assessment coefficient was 0.051.

[0064] Example 3: In the edge computing node of the intelligent distribution system of a 220kV substation, the analog-to-digital conversion acquisition card collects the secondary side ratio error sequence, the partial discharge ultra-high frequency pulse frequency sequence, and the dielectric loss factor change rate sequence of the high-voltage current transformer at a sampling frequency of 100Hz. The sampled data is written to the ring storage buffer with a base address of 0x8000 to 0x812B through the direct memory access channel. After each 1s sliding time window is filled with 100 three-channel sampling points, the microprocessor concatenates the three-channel monitoring matrix at the current sampling time with the time domain differential matrix of the previous consecutive windows along the channel dimension to assemble a dynamic spatiotemporal state tensor. At the same time, a hardware interrupt service routine with a period of 50ms is triggered to load the tensor into the input register of the one-dimensional convolutional neural network feature extractor.

[0065] The one-dimensional convolutional neural network feature extractor maps the dynamic spatiotemporal state tensor into a 64-dimensional low-dimensional domain-aligned latent feature vector. The microprocessor reads the low-dimensional domain-aligned latent feature vectors of the previous five consecutive historical sampling windows, calculates the difference between the latent feature vectors of adjacent historical windows, and takes the arithmetic mean of the resulting difference vectors to obtain a 64-dimensional first-order difference mean vector. The low-dimensional domain-aligned latent feature vector at the current sampling time is concatenated with the first-order difference mean vector to construct the state vector of a 128-dimensional Markov decision process.

[0066] When calculating the historical tangent direction vector, each of the five historical sampling windows corresponds to a time decay weight. The sum of each weight is 1, and the weights decrease monotonically as the time distance between the corresponding window and the current sampling time increases. The microprocessor calculates the sliding weighted average of the latent feature difference vectors of the five historical sampling windows based on this set of weights to obtain the historical tangent direction vector. Then, the arithmetic logic unit calculates the cosine of the angle between the projection of the first-order difference mean vector and the historical tangent direction vector.

[0067] The source domain class center is maintained as a 64-dimensional vector. For a known label dataset of the source domain, the feature extractor first calculates the 64-dimensional latent feature vector of each source domain sample, and then calculates the arithmetic mean according to the label category to obtain the source domain class center corresponding to the normal state of the transformer, the magnetic saturation state of the iron core, the partial discharge state of the insulation, and the drift state of the transformer ratio error. When calculating the domain fitting component, the microprocessor reads the 64-dimensional low-dimensional domain-aligned latent feature vector of the current sampling time from the 128-dimensional state vector, calculates the Euclidean distance of the vector to each 64-dimensional source domain class center, takes the reciprocal of the distance to the nearest source domain class center and performs normalization. This calculation process does not expand the dimension of the source domain class center. The first-order difference mean vector is only used to construct the state vector and reflect the time change trend of the latent features.

[0068] The target domain diagnostic strategy network receives a 128-dimensional state vector and outputs a diagnostic action instruction composed of a combination of discrete fault types and continuous severity assessment coefficients. The microprocessor writes the discrete fault type into the high 2 bits of the control register 0x9000 and converts the continuous severity assessment coefficient into a 14-bit fixed-point number and writes it into the low 14 bits. When the discrete fault type corresponds to the insulation partial discharge state and the continuous severity assessment coefficient is 0.85, the microprocessor triggers the partial discharge alarm bus signal and sends the quantitative degradation coefficient of 0.85 to the substation monitoring terminal, completing sensor signal acquisition, buffer writing, dynamic spatiotemporal state tensor assembly, feature extraction, state vector construction, diagnostic action generation, and control instruction output.

[0069] Example 4: This example combines Figures 1 to 3 This paper describes an intelligent diagnostic method and system for mutual inductors based on the fusion of reinforcement learning and transfer learning, such as... Figure 1As shown, step S1 acquires multi-channel time-series monitoring data of the transformer to be diagnosed and assembles a dynamic spatiotemporal state tensor containing the current sampling window and historical evolution windows; step S2 inputs the known label dataset of the source domain and the unlabeled dynamic spatiotemporal state tensor of the target domain into the domain adversarial feature extraction network, maps to generate a low-dimensional domain-aligned latent feature vector, and extracts the source domain class center of the known label dataset in the latent space; step S3 concatenates the low-dimensional domain-aligned latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculates the projection of the first-order difference mean vector and the historical tangent direction vector. Step S4: Input the state vector into the target domain diagnostic strategy network and output diagnostic action instructions containing fault type and degradation degree evaluation coefficients. Step S5: The soft output of the fault classifier determines the diagnostic guidance component, the inverse of the normalized distance from the low-dimensional domain-aligned latent feature vector to the source domain class center determines the domain fitting component, and the strategy entropy of the target domain diagnostic strategy network output distribution is combined with the weighted calculation of the composite reward signal. When the cosine of the projection angle is negative, the weighting coefficient of the domain fitting component is reduced in the same direction. The strategy gradient derived from the composite reward signal is scaled using the course learning gating factor based on the monotonically increasing training rounds, and the parameters of the target domain diagnostic strategy network are updated. The instrument transformer diagnostic results are output.

[0070] like Figure 2 As shown, firstly, the state vector of the Markov decision process is constructed, and the cosine of the projection angle between the first-order difference mean vector and the historical tangent direction vector is calculated. When the cosine of the projection angle is greater than or equal to 0, the diagnostic guidance component, the domain fitting component, and the policy entropy are combined according to preset weights. When the cosine of the projection angle is negative, the cosine of the projection angle is added to 1 as the attenuation ratio, and the preset weights of the domain fitting component are proportionally reduced in the same direction and then summed. After the above decision branch processing, a composite reward signal is generated. When the course learning gating factor reaches the preset truncation threshold and the smooth mean fluctuation range of the composite reward signal in multiple consecutive training rounds is less than the preset convergence threshold, the network parameter iteration is stopped and the model parameters of the target domain diagnostic policy network are locked, and the mutual inductor diagnosis result is output. When the condition is not met, the policy gradient derived from the composite reward signal is scaled by the monotonically increasing course learning gating factor in the training rounds, the target domain diagnostic policy network parameters are updated, and the feedback is cyclically fed back to the state vector of the Markov decision process.

[0071] like Figure 3As shown, the evolution curve of the gating factor illustrates the corresponding evolution relationship between the course learning gating factor and the number of training rounds. The horizontal axis represents the number of training rounds in units of rounds, with sampling and labeling points of 0, 200, 400, 600, 800, and 1000 respectively. The vertical axis represents the dimensionless course learning gating factor, with values ​​and labeling points of 0, 0.2, 0.4, 0.6, 0.8, and 1 respectively. The course learning gating factor shows a monotonically increasing trend with the increase of training rounds. When the number of training rounds is in the range of 0 to 200, the course learning gating factor remains at a low level close to 0. When the number of training rounds increases from 200 to 800, the value of the course learning gating factor rises smoothly and approaches the maximum value of 1 when the number of training rounds reaches 1000.

[0072] Example 5: Before deploying the online diagnostic system for high-voltage current transformers at a 220kV substation, the analog-to-digital conversion acquisition card is connected to the signal channels for secondary side ratio error, partial discharge ultra-high frequency pulse, and dielectric loss factor of the current transformer. Baseline data is continuously collected for 24 hours under fault-free rated load. The mean value of the secondary side ratio error sequence is calculated and written as the zero-point offset compensation amount into the bias register with a base address of 0x8000.

[0073] After the system is powered on, the analog-to-digital converter acquisition card sequentially writes the three channel sampling points into the circular storage buffer at a sampling frequency of 100Hz. After the buffer is filled with five consecutive sliding time windows, i.e. 500 sampling points, the microprocessor extracts the low-dimensional domain-aligned latent feature vectors of the previous five historical sampling windows, calculates the initial first-order difference mean vector and the historical tangent direction vector, and concatenates the low-dimensional domain-aligned latent feature vector of the current sampling moment with the first-order difference mean vector to complete the state vector construction of the Markov decision process.

[0074] When the target domain diagnostic strategy network starts training, the training round parameters are reset to zero, the course learning gating factor is set to the initial value, the initial direction of the historical tangent direction vector is set to a unit vector in the same direction as the first-order difference mean vector, the cosine of the initial projection angle is greater than or equal to 0, the weighting coefficients of the domain fitting components are kept at the preset value, the microprocessor calculates the diagnostic guidance component based on the soft output of the fault classifier, calculates the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain aligned latent feature vector to the source domain class center at the current sampling time, calculates the strategy entropy based on the action distribution output by the target domain diagnostic strategy network, and then generates a composite reward signal according to the preset weights.

[0075] Under the influence of the initial learning gating factor, the microprocessor scales the policy gradient derived from the composite reward signal and then feeds the scaled gradient into the adaptive moment estimator optimizer to limit the policy parameter update magnitude when the latent features of the target domain have not yet been aligned. The soft output of the fault classifier continues to provide diagnostic guidance components at this stage, and the domain fitting components and policy entropy still participate in the calculation of the composite reward signal.

[0076] Once the edge computing node has continuously run for 1000 sampling windows and the sliding variance of the multi-channel time-series monitoring data has stabilized within the preset tolerance range, the initial accumulation of online data in the target domain is completed. Subsequently, the course learning gating factor monotonically increases with the training rounds of the target domain diagnostic strategy network, gradually increasing to 1.0 during the training process. In each training round, the policy gradient is scaled using this gating factor, and the parameters of the policy network and value network are updated using the near-end policy optimization algorithm. After reaching the preset convergence condition, the model parameters of the target domain diagnostic strategy network are locked, the real-time constructed state vector is input, and the diagnostic action command composed of discrete fault types and continuous severity evaluation coefficients is output, completing the state switching of the high-voltage current transformer from power-on calibration and initial data accumulation to adaptive online diagnosis.

[0077] Example 6: After the smart transformer in the distribution network is connected to the field diagnostic system, a pre-calibration is performed first. The system continuously captures a preset number of power frequency voltage and current waveform sampling points according to the preset acquisition frequency to form an initial state sequence. The power frequency high voltage standard source applies the rated voltage and rated current to the primary side of the smart transformer. The analog-to-digital conversion acquisition card synchronously acquires the secondary side response signal. The ambient temperature sensor and humidity sensor acquire the ambient temperature value and ambient humidity value, respectively.

[0078] The temperature correction factor and humidity correction factor are determined based on standard calibration data under different temperature and humidity conditions at the substation site. Within the ambient temperature range of -20℃ to 70℃, least squares linear fitting is performed on 2000 sets of measured data to obtain a proportional drift coefficient of 0.0003 corresponding to the deviation of the secondary transformer ratio for every 1℃ change in temperature. Within the relative humidity range of 30% to 90%, the dielectric loss tangent is measured to be 0.12% relative shift for every 1% change in relative humidity using a high-voltage dielectric loss tester. The system calculates the secondary transformer ratio deviation correction factor and the dielectric loss tangent correction factor based on the above calibration relationship, and writes the corrected baseline response feature vector into the storage unit as the initial reference for subsequent multi-channel time-series monitoring data processing.

[0079] After completing the pre-calibration, the system generates a current transformer secondary side ratio error sequence, a partial discharge ultra-high frequency pulse frequency sequence, and a dielectric loss factor change rate sequence based on the secondary side response signal. The three types of sequences are standardized and normalized, and a three-channel monitoring matrix is ​​formed according to the sliding time window. Then, the three-channel monitoring matrix at the current sampling time is spliced ​​with the time domain differential matrix of multiple consecutive windows along the channel dimension to assemble a dynamic spatiotemporal state tensor.

[0080] For the laboratory source domain transformer fault dataset and the field target domain transformer operation dataset, the feature extractor forms a high-dimensional feature matrix of the source domain and a high-dimensional feature matrix of the target domain, respectively. The input mapping is completed through the source domain feature transformation matrix and the target domain feature transformation matrix. During training, the known label samples of the source domain and the unlabeled dynamic spatiotemporal state tensor of the target domain are input into the domain adversarial feature extraction network in parallel. The classification cross-entropy loss is calculated by the fault classifier, and the domain discriminator is calculated by the domain discriminator. The feature extractor parameters are updated using the gradient reversal layer. At the same time, the maximum mean difference between the latent features of the source domain and the latent features of the target domain in the regenerating kernel Hilbert space is calculated by the kernel function. When the cross-domain distribution distance is greater than the preset domain adaptation convergence threshold, the source domain feature transformation matrix, the target domain feature transformation matrix and the feature extractor parameters are updated until the cross-domain distribution distance is lower than the threshold. The low-dimensional domain-aligned latent feature vector is obtained, and the source domain class center is obtained according to the source domain label category.

[0081] During the pre-calibration and pre-screening process, the edge lightweight evaluation node also extracts a four-dimensional low-order derived feature vector from the dynamic spatiotemporal state tensor. This vector includes excitation impedance, zero-sequence current polarity, partial discharge pulse frequency, and corrected baseline response characteristics. It is only used for rapid state screening before the full target domain diagnostic strategy network is put into operation and is not used as the state vector of the Markov decision process. The pre-screening node uses four types of fault level indication signals: normal, overheating aging, partial discharge, and inter-turn short circuit. It also sets false alarm penalty factors, missed alarm penalty factors, and early fault state transition detection benefit factors. After screening the transient signals within the preset sampling time, it calculates the instant reward value based on the false alarm indication variable, missed alarm indication variable, and effective detection indication variable. The state transition probability matrix and the strategy network weights of the pre-screening node are determined through offline multi-round iterations.

[0082] During online diagnosis, the system extracts the low-dimensional domain-aligned latent feature vector at the current sampling time and calculates the first-order difference mean vector based on the latent feature vectors of continuous historical sampling windows. The current low-dimensional domain-aligned latent feature vector is concatenated with the first-order difference mean vector to construct the state vector of the Markov decision process. The sliding weighted mean of the latent feature difference vectors of continuous historical windows is obtained according to the time decay weight to obtain the historical tangent direction vector. The cosine of the projection angle between the first-order difference mean vector and the historical tangent direction vector is calculated.

[0083] The target domain diagnostic strategy network receives the state vector and outputs a diagnostic action command composed of discrete fault types and continuous severity evaluation coefficients. Discrete fault types include transformer normal state, core magnetic saturation state, insulation partial discharge state, and ratio error drift state. The soft output of the fault classifier is used to calculate the diagnostic guidance component. The reciprocal of the normalized distance from the current low-dimensional domain aligned latent feature vector to the source domain class center is used to calculate the domain fitting component. The action distribution of the target domain diagnostic strategy network is used to calculate the strategy entropy. When the cosine of the projection angle is negative, the weighting coefficient of the domain fitting component is reduced in the same direction as the sum of the cosine and 1. Then, the diagnostic guidance component, the domain fitting component, and the strategy entropy are combined into a composite reward signal. The state transition probability matrix obtained in the pre-screening stage is only used as the initial constraint for offline training. The online strategy parameters are updated according to the strategy gradient derived from the composite reward signal.

[0084] During the on-site online monitoring process, the system hardware execution unit polls the secondary side sampling data of the intelligent current transformer according to a preset cycle. When the secondary side current distortion rate exceeds the preset threshold for a consecutive preset number of cycles, and the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, the system converts the real-time multi-channel time-series monitoring data into a dynamic spatiotemporal state tensor. After passing through the target domain feature transformation matrix and feature extractor, a low-dimensional domain aligned latent feature vector is generated, and then the state vector is constructed and input into the target domain diagnostic strategy network.

[0085] The target domain diagnostic strategy network output determines the fault category, continuous severity assessment coefficient, and corresponding confidence level. When the confidence level is greater than the preset confidence threshold, the edge computing node sends an alarm message to the distribution automation master station, which includes timestamp, phase information, fault category determination, and quantitative degradation coefficient. For diagnostic results that meet the field control conditions, the signal output interface simultaneously sends an alarm circuit breaker trip preparation command, completing physical signal acquisition, dynamic spatiotemporal state tensor assembly, cross-domain feature extraction, state vector construction, intelligent diagnosis, and decision command output.

[0086] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for intelligent diagnosis of mutual inductors based on the fusion of reinforcement learning and transfer learning, characterized in that, Includes the following steps: Step S1: Obtain multi-channel time-series monitoring data of the current transformer to be diagnosed, and assemble a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window; Step S2: Input the known label dataset of the source domain and the unlabeled dynamic spatiotemporal state tensor of the target domain into the adversarial feature extraction network, map to generate a low-dimensional domain-aligned latent feature vector, and extract the source domain class center of the known label dataset of the source domain in the latent space. Step S3: Concatenate the low-dimensional domain aligned latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. Step S4: Input the state vector into the target domain diagnostic strategy network and output a diagnostic action command containing the fault type and degradation degree evaluation coefficient. Step S5: Determine the diagnostic guidance component based on the soft output of the fault classifier; determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain-aligned latent feature vector to the source domain class center; combine the policy entropy of the target domain diagnostic policy network output distribution; calculate the weighted composite reward signal; and decrease the weighting coefficient of the domain fitting component in the same direction when the cosine of the projection angle is negative; update the target domain diagnostic policy network parameters by scaling the composite reward signal using the course learning gating factor that is monotonically increasing based on the training rounds; and output the instrument transformer diagnostic result.

2. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S1, acquiring multi-channel time-series monitoring data includes: collecting the current transformer secondary side ratio error sequence, the partial discharge ultra-high frequency pulse frequency sequence, and the dielectric loss factor change rate sequence; assembling the dynamic spatiotemporal state tensor includes: standardizing and normalizing the multi-channel time-series monitoring data according to the preset sampling frequency of 100Hz, extracting sampling points according to the set sliding time window to form a three-channel monitoring matrix, and splicing the three-channel monitoring matrix at the current sampling time with the time-domain differential matrix of multiple consecutive windows along the channel dimension to generate a dynamic spatiotemporal state tensor.

3. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S2, the domain adversarial feature extraction network includes a feature extractor, a fault classifier, and a domain discriminator. The process of mapping and generating low-dimensional domain-aligned latent feature vectors and extracting source domain class centers includes: inputting known-label samples from the source domain and unlabeled dynamic spatiotemporal state tensors from the target domain into the feature extractor in parallel to obtain latent feature vectors from the source and target domains; calculating the mean of the latent feature vectors from the source domains according to the label categories to generate source domain class centers; inputting the latent feature vectors from the source domains into the fault classifier to calculate the classification cross-entropy loss, and simultaneously inputting the latent feature vectors from the source and target domains into the domain discriminator to calculate the domain discriminant binary classification loss; using a gradient reversal layer to apply a negative adversarial coefficient to the gradient of the domain discriminator during backpropagation, alternately updating the parameters of the feature extractor and the domain discriminator, and outputting low-dimensional domain-aligned latent feature vectors upon convergence.

4. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 3, characterized in that, The parameter update process of the feature extractor also includes: on the basis of classification cross-entropy loss and domain discriminant binary classification loss with negative adversarial coefficient, a kernel-based maximum mean difference loss term is introduced; the maximum mean difference loss term is weighted by a preset penalty weight coefficient to construct a composite transfer loss function. By minimizing the composite transfer loss function, the distribution distance of the latent features of the source domain and the target domain in the regenerating kernel Hilbert space is reduced, so that the generated low-dimensional domain-aligned latent feature vector has domain independence while preserving the separability between fault classes.

5. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S3, the process of constructing the state vector and calculating the cosine value of the projection angle includes: Step S31, obtaining the low-dimensional domain aligned latent feature vector at the current sampling time, extracting the latent feature vectors under multiple consecutive previous historical sampling windows, calculating the difference between the latent feature vectors of adjacent historical windows, and taking the arithmetic mean of the obtained difference vectors to obtain the first-order difference mean vector; Step S32, performing vector concatenation between the low-dimensional domain aligned latent feature vector at the current sampling time and the first-order difference mean vector to generate the state vector; Step S33, taking the sliding weighted mean of the latent feature difference vectors of multiple consecutive previous historical sampling windows according to the time decay weight, generating the historical tangent direction vector, calculating the dot product of the first-order difference mean vector and the historical tangent direction vector, dividing by the product of their magnitudes to obtain the cosine value of the projection angle.

6. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S5, the process of weighted calculation of the composite reward signal includes: under the unlabeled sampling condition in the target domain, the state vector is input into the fault classifier to obtain the classification probability distribution of each type of fault, and the logarithm of the output class with the highest probability is multiplied by the severity evaluation coefficient in the diagnostic action instruction as the diagnostic guidance component; the inverse of the Euclidean distance between the state vector at the current sampling time and the class center of the source domain is calculated and normalized as the domain fitting component; the policy entropy of the output action distribution of the target domain diagnostic policy network is calculated as the entropy regularization term; when the cosine value of the projection angle is greater than or equal to 0, the diagnostic guidance component, the domain fitting component and the policy entropy are combined according to the preset weights; when the cosine value of the projection angle is negative, the cosine value of the projection angle is added to 1 as the attenuation ratio, and the preset weights of the domain fitting component are proportionally reduced in the same direction and then summed to generate the composite reward signal.

7. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S4, the diagnostic action command is composed of a combination of discrete fault types and continuous severity evaluation coefficients; wherein, discrete fault types include transformer normal state, core magnetic saturation state, insulation partial discharge state, and ratio error drift state, and continuous severity evaluation coefficients are continuous real numbers between 0.0 and 1.0; the target domain diagnostic strategy network adopts an Actor-Critic architecture, which includes a strategy network for outputting action distribution and a value network for estimating state value.

8. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 7, characterized in that, The parameter update process of the target domain diagnostic policy network adopts the near-end policy optimization algorithm. By calculating the ratio of the new policy probability to the old policy probability, the magnitude is constrained by the preset upper and lower limit comparison values ​​of pruning to prevent the single-step policy update magnitude from being too large. The product of the pruned ratio and the generalized advantage estimate is used as the update target. Combined with the gradient value after scaling the gating factor learned in the course, the adaptive moment estimation optimizer is used to iteratively update the weights of the policy network and the value network.

9. The intelligent diagnostic method for mutual inductors based on the fusion of reinforcement learning and transfer learning according to claim 1, characterized in that, In step S5, the process of outputting the instrument transformer diagnosis results includes: when the course learning gating factor reaches the preset cutoff threshold and the smooth mean fluctuation range of the composite reward signal is less than the preset convergence threshold in multiple consecutive training rounds, the network parameter iteration is stopped and the model parameters of the target domain diagnosis strategy network are locked; the instrument transformer monitoring data collected in real time at the substation site is converted into a state vector input to the locked target domain diagnosis strategy network, and the diagnosis results containing the fault category and quantitative degradation coefficient are output to the edge computing node in real time.

10. A current transformer intelligent diagnostic system based on the fusion of reinforcement learning and transfer learning, used to implement the current transformer intelligent diagnostic method based on the fusion of reinforcement learning and transfer learning as described in claim 1, characterized in that, include: The data acquisition and tensor assembly module is used to acquire multi-channel time-series monitoring data of the current transformer to be diagnosed and assemble a dynamic spatiotemporal state tensor containing the current sampling window and the historical evolution window. The feature extraction and class center construction module is used to map the source domain known label dataset and the target domain unlabeled dynamic spatiotemporal state tensor input domain adversarial feature extraction network to generate low-dimensional domain aligned latent feature vectors, and extract the source domain class centers of the source domain known label dataset in the latent space. The state vector construction and tangent cosine calculation module is used to align the low-dimensional domain of the latent feature vector at the current sampling time with the first-order difference mean vector of the latent feature vector under a preset number of historical sampling windows to construct the state vector of the Markov decision process, and calculate the cosine value of the projection angle between the first-order difference mean vector and the historical tangent direction vector. The diagnostic strategy output module is used to input the state vector into the target domain diagnostic strategy network and output diagnostic action instructions containing fault type and degradation degree evaluation coefficients. The composite reward calculation and policy update module is used to determine the diagnostic guidance component based on the soft output of the fault classifier, determine the domain fitting component based on the reciprocal of the normalized distance from the low-dimensional domain aligned latent feature vector to the source domain class center, combine the policy entropy of the target domain diagnostic policy network output distribution, calculate the composite reward signal in a weighted manner, and attenuate the weighting coefficients of the domain fitting component in the same direction when the cosine of the projection angle is negative, use the policy gradient derived from the composite reward signal by scaling the course learning gating factor based on the monotonically increasing training rounds, update the parameters of the target domain diagnostic policy network, and output the instrument transformer diagnostic results.

Images