Intelligent monitoring method and system for high and low voltage switch cabinet based on internet of things

Abstract

The application discloses a high-low voltage switch cabinet intelligent monitoring method and system based on the Internet of Things, and belongs to the technical field of switch cabinet monitoring. The technical scheme points of the application are as follows: time sequence data and infrared image data of the high-low voltage switch cabinet at the current time are acquired; a first fault feature vector is obtained according to the time sequence data, the infrared image data and a preset model, and the preset model is obtained through training; prediction data of the next time is obtained according to the time sequence data and a digital twin model; a confidence degree corresponding to a current fault state is obtained according to the fault feature vector and the prediction data of the next time; a control action is obtained according to the confidence degree and a deep reinforcement learning model, and the deep reinforcement learning model is obtained through training. According to the domain adaptation and fusion of the transfer learning module, the generated fault feature vector can accurately reflect the fault grade, and the control action is selected through the confidence degree of the data and the deep learning model, so that rapid and accurate fault control is realized.

Description

Technical Field

[0001] This invention relates to the field of switchgear monitoring technology, and more specifically to an intelligent monitoring method and system for high and low voltage switchgear based on the Internet of Things. Background Technology

[0002] In power systems, switchgear, as a key piece of equipment for power transmission and distribution, directly affects the safety and stability of the power system. With the expansion of power system scale and the increasing demands for intelligence, higher challenges are posed to the accuracy, real-time performance, and intelligence of switchgear fault monitoring and control. Traditional switchgear fault monitoring methods often rely on single sensor data or simple model analysis, which struggles to cope with complex and changing operating environments and individual differences between different switchgear units. These methods often suffer from low fault identification accuracy, poor cross-device generalization ability, untimely decision-making, and a lack of dynamic optimization, seriously affecting the reliable operation of the power system. Therefore, existing technologies have shortcomings. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent monitoring method and system for high and low voltage switchgear based on the Internet of Things. By using a transfer learning module for domain adaptation and fusion, the problem of differences in feature distribution among different switchgear is solved, enhancing the cross-domain generalization ability of the model and thereby improving the accuracy of fault feature extraction.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides an intelligent monitoring method for high and low voltage switchgear based on the Internet of Things, comprising:

[0006] Acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment;

[0007] Based on the time-series data, the infrared image data, and the preset model, a first fault feature vector is obtained. The preset model includes a convolutional long short-term memory network module, an attention mechanism module, and a transfer learning module. The preset model is obtained through training.

[0008] Based on the time series data and the digital twin model, the predicted data for the next moment is obtained;

[0009] Based on the first fault feature vector and the predicted data for the next moment, the confidence level corresponding to the current fault state is obtained;

[0010] The control action is obtained based on the confidence level and the deep reinforcement learning model, which is obtained through training.

[0011] As a further improvement of the present invention, the preset model is obtained through training, including:

[0012] Obtain source domain data and target domain data;

[0013] An iterative operation is performed based on the source domain data and the target domain data. The iterative operation includes obtaining a first feature vector and a second feature vector based on the source domain data, the target domain data, and the feature extractor; obtaining a fault classification loss and a domain classification loss based on the first feature vector and the second feature vector; updating the parameters in the transfer learning module based on the fault classification loss and the domain classification loss, until a termination condition is reached to obtain a first preset model; the transfer learning module includes the feature extractor.

[0014] Based on the target domain data and Bayesian inference, the parameters in the feature extractor are updated to obtain a second preset model;

[0015] The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the target domain data and the second preset model to obtain the preset model.

[0016] As a further improvement of the present invention, the transfer learning module further includes a label predictor, a gradient reversal layer, and a domain classifier, which obtains fault classification loss and domain classification loss based on the first feature vector and the second feature vector, including:

[0017] Based on the first feature vector and the label predictor, a second fault feature vector is obtained;

[0018] The fault classification loss is obtained based on the second fault feature vector;

[0019] The domain classification loss is obtained based on the second feature vector, the gradient inversion layer, and the domain classifier.

[0020] As a further improvement of the present invention, updating the parameters in the transfer learning module based on the fault classification loss and the domain classification loss includes:

[0021] The parameters in the label predictor and the feature extractor are updated based on the fault classification loss.

[0022] The parameters in the domain classifier and the feature extractor are updated based on the domain classification loss.

[0023] As a further improvement of the present invention, the parameters in the feature extractor are updated based on the target domain data and Bayesian inference to obtain a second preset model, including:

[0024] Based on the target domain data and Bayesian inference, the posterior probability corresponding to the target domain data is obtained;

[0025] Based on the fault classification loss and the domain classification loss, the gradient of the total loss with respect to the parameters in the feature extractor is obtained;

[0026] Based on the posterior probability and the gradient, the parameters in the feature extractor are updated in a weighted manner to obtain the second preset model.

[0027] As a further improvement of the present invention, the parameters in the convolutional long short-term memory network module and the attention mechanism module are updated according to the target domain data and the second preset model to obtain the preset model, including:

[0028] Based on the target domain data and the second preset model, the fault identification result is obtained;

[0029] The deviation vector is obtained based on the fault identification results and the preset results;

[0030] The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the deviation vector to obtain the preset model.

[0031] As a further improvement of the present invention, the confidence level corresponding to the current fault state is obtained based on the first fault feature vector and the prediction data at the next moment, including:

[0032] Based on the first fault feature vector and the predicted data at the next moment, the first basic allocation probability and the second basic allocation probability are obtained.

[0033] The confidence level is obtained based on the first basic allocation probability and the second basic allocation probability.

[0034] As a further improvement of the present invention, the deep reinforcement learning model is obtained through training, including:

[0035] Based on each sample in the source domain data, the basic allocation probability of each fault level is obtained;

[0036] The confidence level of each fault level is obtained based on the basic allocation probability of each fault level.

[0037] Based on each sample and the digital twin model, the predicted value corresponding to each sample is obtained;

[0038] Based on the predicted values ​​and the current deep reinforcement learning model, the reward function corresponding to each control action is obtained;

[0039] The deep reinforcement learning model is obtained by updating the current deep reinforcement learning model according to the reward function.

[0040] As a further improvement of the present invention, the basic allocation probability of each fault level is obtained based on each sample in the source domain data, including:

[0041] Cluster analysis is performed on each sample in the source domain data to obtain the category corresponding to each fault level;

[0042] For each category, the correlation between the samples included in it and each fault level is calculated separately;

[0043] The basic allocation probability for each fault level is obtained based on the correlation.

[0044] This invention provides an intelligent monitoring system for high and low voltage switchgear based on the Internet of Things, comprising:

[0045] The monitoring module is used to acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment;

[0046] The fault identification module is used to obtain a first fault feature vector based on the time series data, the infrared image data, and the preset model. The preset model includes a convolutional long short-term memory network module, an attention mechanism module, and a transfer learning module. The preset model is obtained through training.

[0047] The prediction module is used to obtain the prediction data for the next moment based on the time series data and the digital twin model;

[0048] The state recognition module is used to obtain the confidence level corresponding to the current fault state based on the first fault feature vector and the prediction data at the next moment.

[0049] The control module is used to obtain control actions based on the confidence level and the deep reinforcement learning model.

[0050] This invention extracts temporal and spatial features through a convolutional long short-term memory network module and an attention mechanism module, respectively. Then, it performs domain adaptation and fusion through a transfer learning module, which solves the problem of feature distribution differences among different switch cabinets, enhances the model's cross-domain generalization ability, and thus improves the accuracy of fault feature extraction. Furthermore, it obtains the confidence level corresponding to the current fault state through the first fault feature vector and the predicted data at the next time step. Finally, the reinforcement learning module selects the optimal control action based on the confidence level to solve the fault problem. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0052] Figure 2 This is a schematic diagram of the pre-defined model.

[0053] Figure 3This is a schematic diagram illustrating the steps of an iterative operation.

[0054] Figure 4 A schematic diagram illustrating some steps involved in training a deep reinforcement learning model. Detailed Implementation

[0055] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0056] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0057] like Figure 1 As shown in the figure, this application provides an intelligent monitoring method for high and low voltage switchgear based on the Internet of Things, including:

[0058] Acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment;

[0059] Based on time-series data, infrared image data, and a pre-set model, the first fault feature vector is obtained. The pre-set model includes a convolutional long short-term memory network module, an attention mechanism module, and a transfer learning module. The pre-set model is obtained through training.

[0060] Based on time series data and digital twin models, predictive data for the next moment is obtained;

[0061] Based on the first fault feature vector and the predicted data for the next time step, the confidence level corresponding to the current fault state is obtained;

[0062] The control action is obtained based on the confidence level and the deep reinforcement learning model, which is obtained through training.

[0063] This embodiment can be used to monitor faults in multiple components of high and low voltage switchgear. For example, it can be used to monitor contact faults. In this case, the timing data and infrared image data of the high and low voltage switchgear are the timing data and infrared image data corresponding to the contact. The data corresponding to the current moment includes data corresponding to multiple sampling points between the previous moment and the current moment. This embodiment does not limit the number of sampling points or the interval between two adjacent moments. The timing data can include temperature timing data, current timing data, humidity timing data, and contact resistance timing data, which can be collected by temperature sensors, current sensors, humidity sensors, and online monitoring devices based on electromagnetic induction. The infrared image data is the infrared image of the contact corresponding to the current moment, collected by a thermal imager. Since the acquisition frequency of each device or sensor is different, the number of values ​​included in each timing data may be different. An interpolation algorithm can be used to ensure that each timing data includes the same number of values.

[0064] Next, the data of each type in the time series is arranged in chronological order, resulting in multiple vectors. Each vector includes four elements, corresponding to temperature, current, humidity, and contact resistance, and has already been normalized. These vectors are then input into a convolutional long short-term memory network module in chronological order to obtain the temporal feature vector corresponding to the current time step. The infrared image is then preprocessed to fit its size to the model's input requirements, and then input into the attention mechanism module to obtain the spatial feature vector corresponding to the current time step.

[0065] The convolutional long short-term memory (LSTM) network module and attention mechanism module are trained, and model training is a technique that can be implemented by those skilled in the art, so it will not be described in detail here. Unlike the exemplary techniques, the LSM network module in this embodiment includes a feature extraction branch and a reconstruction branch. The feature extraction branch is used to extract a temporal feature vector from the hidden state at the final time step, and the reconstruction branch is used to predict the vector for the next time step based on the hidden state output. This vector is used to calculate the loss function and train the LSM network module. The loss function used during training is obtained based on the temperature time-series reconstruction loss, the temperature change rate and contact resistance correlation loss, and the contact resistance time-series reconstruction loss. The temperature time-series reconstruction loss is the mean square error between the predicted temperature in the vector and the actual temperature in the vector; similarly, the contact resistance time-series reconstruction loss is the mean square error between the predicted contact resistance in the vector and the actual contact resistance in the vector. To calculate the loss relating temperature change rate to contact resistance, first, the rate of temperature change is calculated based on the predicted temperature in the vector. Then, the correlation coefficient between this rate of change and the contact resistance in the predicted vector is calculated. Next, the correlation coefficient between the rate of change and the contact resistance in the predicted vector is calculated again based on the actual data. The absolute value of the difference between the two correlation coefficients is used as the loss relating temperature change rate to contact resistance. Finally, the three losses are weighted to obtain the final loss function value, and the parameters in the convolutional long short-term memory network module are updated based on the loss function value. It should be noted that the reconstruction branch is only used when training the model.

[0066] This embodiment found that contact oxidation, loose bolts, and other factors can cause changes in contact resistance, which can easily lead to overheating and temperature rise in the contacts. Therefore, this embodiment designs a loss function based on the characteristics of contact faults, and guides the model to learn the changing patterns of features and the correlations between features, so that the extracted features can better reflect the true state of the contacts, providing high-quality input for subsequent fault monitoring and decision-making.

[0067] Furthermore, the attention mechanism module in this embodiment includes a CBAM (Convolutional Block Attention Module) unit. The CBAM unit first performs channel attention calculation, learns the weights in the channel dimension, and filters out important channels related to temperature and hotspots. Then, it performs spatial attention calculation, learns the weights of pixel positions, and focuses on the specific location of the touch in the image, further enhancing the feature representation of the hotspot region. Finally, it outputs a spatial attention feature map, where each pixel value represents the probability that the location is a hotspot. The spatial feature vector output branch transforms the feature map into a one-dimensional vector by performing global average pooling, thus obtaining the spatial feature vector. The hotspot morphology feature branch is used to output a morphology description vector, which is specifically obtained by concatenating values ​​such as the area, roundness, and major axis / minor axis ratio of the hotspot. During training, based on the spatial attention feature map and the manually labeled real mask region, a local hotspot attention loss is obtained using the formula of binary cross-entropy loss. A hotspot morphology loss is obtained based on the mean square error between the output morphology description vector and the real morphology description vector. Finally, the loss function corresponding to the attention mechanism module is obtained by weighting the local hotspot attention loss and the hotspot morphology loss. The parameters in the attention mechanism module are updated through this loss function. The hotspot morphology feature branch is only used when training the model.

[0068] This embodiment guides the CBAM unit of the model to accurately focus on the real hotspot area through local hotspot attention loss, enhances the model's ability to perceive hotspot locations, makes the extracted features more focused on key areas related to faults, reduces interference from irrelevant background information, and enables the model to learn the morphological features of hotspots through hotspot morphology loss, enriching the hotspot information dimension contained in the feature vector, making the features more discriminative, and finally obtaining accurate spatial feature vectors from all parties.

[0069] After obtaining the time feature vector and spatial feature vector output at the current moment, they are input into the transfer learning module, which outputs the first fault feature vector at the current moment. The elements in the first fault feature vector represent the probability corresponding to each fault level and the probability of no fault. For example, if the first fault feature vector includes 4 elements, the first element represents the probability of no fault, the second element represents the probability of a level 1 fault, the third element represents the probability of a level 2 fault, and so on. Therefore, based on the first fault feature vector, it can be determined whether the contact has a fault and the severity level of the fault. This embodiment does not limit the way the fault level is divided, and those skilled in the art can determine it according to the actual situation. Then, based on the time series data and the digital twin model, the predicted data for the next moment is obtained. The digital twin model is trained and can accurately draw the cabinet, contacts, cooling fans and other components according to the actual size of the high and low voltage switchgear being monitored. At the same time, relevant physical field parameters, such as rated voltage and cooling fan speed, are set to accurately simulate the operation of the high and low voltage switchgear. When the time series data corresponding to the current moment is input, the digital twin model can output the predicted data of the high and low voltage switchgear for the next moment. In addition to the data types in the time series data, the predicted data can also include current distribution and electric field intensity distribution. Finally, the optimal fault control action is accurately obtained based on the predicted data and the deep learning model.

[0070] This embodiment extracts temporal and spatial features through a convolutional long short-term memory network module and an attention mechanism module, respectively. Then, it performs domain adaptation and fusion through a transfer learning module, which solves the problem of feature distribution differences among different switch cabinets, enhances the model's cross-domain generalization ability, and thus improves the accuracy of fault feature extraction. Furthermore, it obtains the confidence level corresponding to the current fault state through the first fault feature vector and the predicted data at the next time step. Finally, the reinforcement learning module selects the optimal control action based on the confidence level to solve the fault problem.

[0071] Furthermore, this embodiment provides a step for training a preset model, including:

[0072] Obtain source domain data and target domain data;

[0073] Iterative operations are performed based on source domain data and target domain data. The iterative operations include obtaining a first feature vector and a second feature vector based on the source domain data, target domain data, and feature extractor; obtaining fault classification loss and domain classification loss based on the first feature vector and the second feature vector; updating the parameters in the transfer learning module based on the fault classification loss and domain classification loss, until the termination condition is met, and obtaining a first preset model; the transfer learning module includes a feature extractor.

[0074] Based on the target domain data and Bayesian inference, the parameters in the feature extractor are updated to obtain the second preset model;

[0075] The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the target domain data and the second preset model to obtain the preset model.

[0076] The source domain data consists of historical data from other types of switchgear, including time feature vectors, spatial feature vectors, and the actual fault level at each moment. The target domain data consists of historical data from currently monitored high and low voltage switchgear, also including time and spatial feature vectors at each moment, but without specifying the actual fault level. The source domain data is divided into multiple data groups, each containing multiple samples. Similarly, the target domain data also includes multiple data groups. Because the model training process involves multiple iterations, to ensure accuracy, each iteration uses one data group from the source domain and one data group from the target domain, and these data groups are different each time.

[0077] Specifically, firstly, a dataset from the source domain is input into the feature extractor to obtain the first feature vector for each sample in the dataset. The feature extractor concatenates the temporal and spatial feature vectors of each sample and performs dimensionality reduction to extract more representative key features. Next, a dataset from the target domain is input into the feature extractor to obtain the second feature vector for each sample in the dataset. Then, based on the first and second feature vectors, the fault classification loss and domain classification loss are obtained. The parameters in the transfer learning module are then updated based on these losses to finally obtain the preset model.

[0078] In this embodiment, when the sample size of high and low voltage switchgear is insufficient and it is difficult to accurately determine the fault situation, the transfer learning model effectively reduces the difference in feature distribution between different domains, improves cross-domain classification performance, and corrects the parameters of the transfer learning model based on the posterior probability of matching source domain features with target domain features through Bayesian inference. This enables the model to better adapt to the target domain data, solves the problem of insufficient cross-domain generalization ability, and thus quickly builds a high-performance fault identification model even with limited labeled data in the target domain. Furthermore, the parameters in the convolutional long short-term memory network module and the attention mechanism module are fine-tuned through the bias vector to improve the accuracy of feature extraction, ensuring the stability and superiority of the preset model in cross-domain scenarios.

[0079] Furthermore, this embodiment provides a step for obtaining fault classification loss and domain classification loss based on a first feature vector and a second feature vector, including:

[0080] Based on the first feature vector and the label predictor, the second fault feature vector is obtained;

[0081] The fault classification loss is obtained based on the second fault feature vector;

[0082] Based on the second feature vector, the gradient inversion layer, and the domain classifier, the domain classification loss is obtained.

[0083] Among them, such as Figure 2 As shown, the transfer learning module includes two branches: a label prediction branch, which includes a label predictor to output fault feature vectors, and a domain classification branch, which includes a gradient reversal layer and a domain classifier. Specifically, after obtaining the first feature vector, it is input into the label predictor. The label predictor outputs a second fault feature vector through a fully connected layer and an activation function. After obtaining the second feature vector, it and the first feature vector are input into the gradient reversal layer. Since the gradient reversal layer reverses the gradient sign during backpropagation, the output vector is not processed during forward propagation. The input second and first feature vectors are directly output to the domain classifier. The domain classifier outputs the probability corresponding to each second and first feature vector through a fully connected layer and an activation function. The closer the probability is to 0, the easier it is for the feature vector to be identified as coming from the source domain; the closer the probability is to 1, the easier it is for the feature vector to be identified as coming from the target domain.

[0084] Based on the second fault feature vector corresponding to each first feature vector output by the label predictor, the fault condition and fault level corresponding to each first feature vector can be obtained. Then, the true fault level corresponding to the first feature vector is determined based on historical data. The fault classification loss corresponding to the feature vector can be calculated based on the formula of cross-entropy loss. Finally, the mean of the fault classification loss corresponding to each first feature vector is used as the final fault classification loss used to update the parameters.

[0085] After obtaining the probabilities corresponding to each first feature vector and each second feature vector through the domain classifier, the domain classification loss corresponding to each first feature vector and each second feature vector is obtained based on the case that it truly belongs to the source domain or the target domain, according to the binary cross-entropy loss. Then, the mean of the domain classification loss of each feature vector is used as the final domain classification loss used to update the parameters.

[0086] This embodiment uses fault classification loss to measure the degree to which the model's prediction of fault level matches the actual fault situation, thereby encouraging the model to focus on learning effective fault features and improving the accuracy of fault classification. Furthermore, by using domain classification loss, it enhances the model's generalization ability across different domains, ensuring that the model can maintain good fault monitoring performance when facing switchgear data from different operating environments and individual devices, ultimately making the trained model more accurate.

[0087] Furthermore, this embodiment provides a step for updating the parameters in the transfer learning module based on the fault classification loss and the domain classification loss, including:

[0088] Update the parameters in the label predictor and feature extractor based on the fault classification loss;

[0089] Update the parameters in the domain classifier and feature extractor based on the domain classification loss.

[0090] Specifically, after obtaining the fault classification loss, the parameters in the domain classification branch are first fixed, and the parameters of the label predictor and feature extractor are updated. This can be done through backpropagation and by using gradient descent to update the parameters of the label predictor and feature extractor. Similarly, this method can be used to update the parameters of the domain classifier using the domain classification loss. When updating the parameters of the domain classifier, all other parameters need to be fixed. This parameter update method is conventional for those skilled in the art and will not be elaborated here. Finally, all parameters except those of the feature extractor are fixed, and the parameters of the feature extractor are updated again using the domain classification loss. The difference is that when the gradient inversion layer receives the gradient, the sign of the gradient needs to be inverted before propagating it to the feature extractor, so that the feature extractor updates in the direction of increasing domain classification loss.

[0091] After the update is complete, as follows Figure 3 As shown, a data set from the source domain and a data set from the target domain are obtained again. The above steps are repeated to obtain multiple first feature vectors and second feature vectors. The feature distribution of the source domain data can be fitted based on each first feature vector, and the feature distribution of the target domain data can be fitted based on each second feature vector. This embodiment does not limit the fitting method; for example, kernel density estimation can be used. Then, based on the feature distributions of the source and target domain data, the probability distributions corresponding to the source and target domain data can be obtained. The difference between the two probability distributions is then calculated using KL divergence. Next, the actual fault situation of the sample corresponding to the second feature vector is obtained. Since the historical data does not include the actual fault situation of the target domain data, it can be manually labeled. Then, the proportion of correctly classified faults is calculated based on the actual fault situation and the fault situation obtained from the fault feature vector. If the difference calculated by KL divergence is less than the preset difference, and the proportion of correctly classified faults is greater than the preset proportion, the termination condition is met, and the first preset model is obtained.

[0092] If the termination condition is not met, the parameters in the label predictor, feature extractor, and domain classifier are updated again according to the above steps until the termination condition is met, resulting in the first preset model. The domain classification branch is only used during the training process.

[0093] This embodiment first adjusts the feature extractor and label predictor using fault classification loss to ensure the accuracy of fault classification. Then, it updates the parameters in the domain classifier and feature extractor using domain classification loss. Although the domain classifier is not used for real-time data processing, by continuously updating the parameters of the domain classifier, its ability to discriminate domain labels can be continuously optimized. That is, it can more accurately distinguish whether a sample comes from the source domain or the target domain. The optimized discriminative ability will act in reverse on the feature extractor through the gradient reversal layer, so that the feature extractor can strip away redundant information related to domain differences in the features and retain only the core features related to fault judgment. Thus, it can improve the fault classification performance and achieve effective cross-domain feature transfer.

[0094] Furthermore, this embodiment provides a step of updating the parameters in the feature extractor based on target domain data and Bayesian inference to obtain a second preset model, including:

[0095] Based on the target domain data and Bayesian inference, the posterior probability corresponding to the target domain data is obtained;

[0096] Based on the fault classification loss and the domain classification loss, the gradient of the total loss with respect to the parameters in the feature extractor is obtained;

[0097] Based on the posterior probability and gradient, the parameters in the feature extractor are updated in a weighted manner to obtain the second preset model.

[0098] Specifically, after obtaining the first preset model, the proportion of samples belonging to each fault level in the source domain data is first calculated. For example, assuming the fault feature vector includes four elements, it includes four levels: no fault, level 1 fault, level 2 fault, and level 3 fault. Based on the actual fault situation of each sample in the source domain data, the number of samples belonging to each level can be obtained, and thus the proportion of samples for each fault level can be obtained, denoted as . , , and These correspond to no fault, Level 1 fault, Level 2 fault, and Level 3 fault, respectively.

[0099] The target domain data is then divided into training and testing sets. The similarity between the second feature vector corresponding to each sample in the training set and each sample in the fault-free source domain data is calculated. The mean of the similarity is used as the likelihood probability of the target domain data with respect to fault-free conditions. Similarly, the likelihood probability of the target domain data for other levels can be calculated. , and Then, using the law of total probability, the probability of evidence can be calculated. Then, using Bayesian probability, the posterior probability of the target domain data for each fault level can be obtained. For example, the posterior probability of the target domain data for no fault is... Similarly, we can obtain , and .

[0100] Next, the fault classification loss and domain classification loss corresponding to each first feature vector used to update the model parameters during the training process of the first preset model are obtained. Then, the two losses are weighted and summed to obtain the total loss corresponding to each first feature vector. This embodiment does not restrict the weights; for example, the weights of the two losses can be set to be equal. Then, the gradient of each total loss with respect to the parameters in the feature extractor is calculated, and the mean of these gradients is obtained. For any parameter in the feature extractor The updated parameters are obtained based on the gradient and posterior probability. for:

[0101]

[0102] in The learning rate is indicated, and this embodiment does not limit the specific value of the learning rate. The first preset model after updating the parameters in the feature extractor is denoted as the second preset model. Furthermore, this embodiment only provides the steps for updating once. If a more accurate model is needed, the source domain data and target domain data can be divided into multiple data groups, and the steps of calculating the posterior probability and updating the parameters can be repeated until the learning rate drops to the preset standard. Then, through multiple iterations, a more accurate second preset model can be obtained.

[0103] This embodiment uses posterior probabilities as weights to adjust the parameters of the feature extractor. It fully utilizes the knowledge of the matching degree between the target domain and the source domain categories contained in the posterior probabilities, making the feature extractor more inclined to extract features that are more discriminative for cross-domain fault classification during the learning process. For example, for samples with a high proportion of a certain class in the posterior probability, their corresponding gradients will be given greater weight when updating parameters, thereby strengthening the model's learning of features related to that class, effectively alleviating the problem of distribution differences between the source domain and the target domain, and further improving the model's fault classification accuracy and cross-domain generalization ability in the target domain.

[0104] Furthermore, this embodiment provides a step of updating the parameters in the convolutional long short-term memory network module and the attention mechanism module based on the target domain data and the second preset model to obtain the preset model, including:

[0105] Based on the target domain data and the second preset model, the fault identification results are obtained;

[0106] The deviation vector is obtained based on the fault identification results and the preset results;

[0107] The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the bias vector to obtain the preset model.

[0108] Specifically, the test set in the target domain data is input into the second preset model to obtain the fault identification vector corresponding to each sample, thereby determining the fault level corresponding to each sample. Then, samples with incorrect fault level identification are manually identified, and the time feature vectors and spatial feature vectors corresponding to these samples are obtained. Next, the time feature vector and spatial feature vector corresponding to a fault-free sample are obtained. For the time feature vector corresponding to each incorrectly identified sample, the difference between it and the time feature vector corresponding to the fault-free sample is calculated, and the mean of these differences is calculated to obtain the time deviation vector. Similarly, the spatial deviation vector can be obtained.

[0109] Next, the parameters in the convolutional long short-term memory network module are updated based on the time bias vector to obtain the updated parameters. for:

[0110]

[0111] in, The parameters before the update. For learning rate, This is the temporal bias vector. Specifically, since the hidden layers of a convolutional long short-term memory network contain convolutional kernel weights, these weights are used to perform convolution operations on the input data to extract temporal features. Therefore, the convolutional kernel is essentially a multidimensional array, i.e. The dimension is ,in Indicates the spatial size of the convolution kernel. This represents the number of feature channels, therefore Dimensions and The dimensions are different, so the calculation is performed differently. At this time, dimensionality compatibility can be achieved through broadcasting operations. Similarly, for the parameters in the attention mechanism module, the updated parameters can also be obtained based on the spatial deviation vector and the above formula. This embodiment will not elaborate on this further. Finally, the second preset model with updated parameters is recorded as the preset model. Furthermore, this embodiment only provides the steps for updating once. If a more accurate model is required, the test set can be divided into multiple data groups to update the parameters multiple times.

[0112] This embodiment first updates the parameters using fault classification loss and domain classification loss, enabling the model to initially possess cross-domain and fault classification capabilities. Then, based on Bayesian inference and using posterior probability, the model can more accurately focus on information closely related to the source domain category during parameter updates, further optimizing the model's adaptability to target domain data. Through these two updates, the model can adapt to target domain data to a certain extent, reducing overall misclassification. Finally, it focuses on the weakest link of the model, namely misidentified samples, and updates the parameters again by generating a bias vector. This effectively corrects the model's biases in temporal feature extraction, feature attention allocation, etc., thereby further improving the model's fault identification accuracy in target domain data.

[0113] Furthermore, this embodiment provides a step for obtaining the confidence level corresponding to the current fault state based on the first fault feature vector and the prediction data at the next time step, including:

[0114] Based on the first fault feature vector and the predicted data at the next moment, the first basic allocation probability and the second basic allocation probability are obtained.

[0115] The confidence level is obtained based on the first and second basic allocation probabilities.

[0116] Specifically, based on the first fault feature vector at the current moment, the probability corresponding to each fault level can be obtained. Furthermore, based on the difference between this probability and 1, the probability that each fault level will not occur can be obtained. Therefore, for each fault level, there are two probability values, denoted as the first basic allocation probability corresponding to that fault level. Similarly, based on the predicted data for the next moment and the preset model, the first fault feature vector corresponding to the next moment can be obtained, leading to the probability of each fault level and the probability that each fault level will not occur in the next moment, thus obtaining the second basic allocation probability for each fault level. For each fault level, based on its corresponding first and second basic allocation probabilities, its corresponding confidence level can be obtained by fusing them using the Dempster combination rule. Based on the confidence level, the fault level at the current moment can be finally determined. If the fault level is no fault, no further steps are performed; otherwise, the confidence level is input into the trained deep reinforcement learning model, causing the model to output the optimal control action for handling the fault.

[0117] This embodiment makes fault control more forward-looking and comprehensive by using the fault feature vector corresponding to the current time and the fault feature vector corresponding to the next time. The confidence level is a quantification of the comprehensive trust in the fault level. The higher the confidence level, the higher the probability of being in that fault level. The confidence level is input into the trained deep learning model. The deep learning model will output the optimal control action that conforms to the confidence level according to the mapping relationship between the trained confidence level and the control action, so as to reasonably solve the fault problem.

[0118] Furthermore, such as Figure 4 As shown, this embodiment provides a step for training a deep reinforcement learning model, including:

[0119] Based on each sample in the source domain data, the basic allocation probability of each fault level is obtained;

[0120] The confidence level of each fault level is obtained based on the basic allocation probability of each fault level.

[0121] Based on each sample and the digital twin model, the predicted value for each sample is obtained;

[0122] Based on the predicted values ​​and the current deep reinforcement learning model, the reward function corresponding to each control action is obtained;

[0123] The deep reinforcement learning model is obtained by updating the current deep reinforcement learning model according to the reward function.

[0124] Specifically, firstly, each sample in the source domain data is clustered into multiple categories based on its corresponding true fault level, with each category corresponding to a fault level. Then, for each category, time-series data and infrared image data for each sample are acquired. The time-series data can include four types: temperature time-series data, current time-series data, humidity time-series data, and contact resistance time-series data. Next, the correlation coefficient between each type and the fault level corresponding to that category is calculated, and the temperature gradient of each infrared image is calculated to obtain the correlation coefficient between the temperature gradient and the fault level corresponding to that category. This correlation coefficient is then used as the correlation coefficient between the infrared image data and the fault level corresponding to that category. This process then yields the correlation coefficients between temperature, current, humidity, contact resistance, and infrared image data and the corresponding fault level for each category. Next, for each fault level, the support for each data type is calculated. For example, to obtain the support for temperature in a level 2 fault, multiple temperature ranges can be initially obtained based on historical data, each corresponding to a fault level. The upper and lower limits of the ranges corresponding to level 2 faults are then obtained, and the mean of the difference between the current temperature and the upper and lower limits is calculated. The ratio of these two means is recorded as the support for temperature in the level 2 fault, with the support ranging from 0 to 1. Similarly, the support for each type within each fault level can be obtained. Therefore, for each type within each fault level, based on its corresponding correlation coefficient, support, and Dempster's combination rule, its corresponding basic allocation probability can be obtained. Thus, for each category, five basic allocation probabilities are obtained, corresponding to temperature, current, humidity, contact resistance, and infrared image data, respectively. These five basic allocation probabilities are summed to obtain the basic allocation probability for each category, i.e., the basic allocation probability for each fault level. The confidence level for each fault level can be obtained based on the basic allocation probability and Dempster's combination rule corresponding to each fault level.

[0125] Subsequently, for each fault category, samples from each source domain are sequentially input into the deep learning model for multiple iterations. The deep learning model can obtain the predicted value (i.e., the predicted data for the next moment) for each sample under different control actions. Based on the predicted data for the next moment and the data collected at the current moment, the value of the reward function can be calculated. The reward function is used to measure the mitigation of the fault problem after executing the control action. Through multiple iterations, the deep learning model learns which control action corresponds to a higher reward function for each fault level, thereby establishing a mapping relationship between confidence and control action. The establishment of the reward function and the updating of the deep learning model based on the reward function are technical means that can be implemented by those skilled in the art, and will not be elaborated here. Different control actions can also be set by those skilled in the art based on experience, such as transferring part of the load (e.g., 30% / 50%) of the current high and low voltage switchgear to other standby switchgear in the same area, reducing the maximum allowable power of the high and low voltage switchgear, and starting / stopping / adjusting the cooling fan, etc., which are not limited in this embodiment.

[0126] Based on the above analysis, we can summarize as follows: Figure 4 As shown, the steps provided in this embodiment for obtaining the basic allocation probability of each fault level based on each sample in the source domain data include:

[0127] Cluster analysis is performed on each sample in the source domain data to obtain the category corresponding to each fault level;

[0128] For each category, the correlation between the samples it includes and each fault level is calculated separately;

[0129] The basic allocation probability of each fault level is obtained based on the correlation.

[0130] This embodiment trains a deep learning model based on confidence and reward functions, and constructs a mapping relationship between confidence and control actions. In scenarios such as remote substations where it is difficult to cover manually, it can automatically complete fault control actions, reducing labor costs and human error rates.

[0131] Furthermore, embodiments of this application provide an intelligent monitoring system for high and low voltage switchgear based on the Internet of Things, comprising:

[0132] The monitoring module is used to acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment;

[0133] The fault identification module is used to obtain the first fault feature vector based on time series data, infrared image data and a preset model. The preset model includes a convolutional long short-term memory network module, an attention mechanism module and a transfer learning module. The preset model is obtained through training.

[0134] The prediction module is used to obtain the prediction data for the next time step based on time series data and digital twin models;

[0135] The state recognition module is used to obtain the confidence level corresponding to the current fault state based on the fault feature vector and the prediction data for the next moment.

[0136] The control module is used to obtain control actions based on confidence and deep reinforcement learning models.

[0137] The intelligent monitoring method and system for high and low voltage switchgear based on the Internet of Things provided in this application extracts temporal and spatial features through a convolutional long short-term memory network module and an attention mechanism module, respectively. Then, it performs domain adaptation and fusion through a transfer learning module, which solves the problem of feature distribution differences among different switchgears, enhances the model's cross-domain generalization ability, and thus improves the accuracy of fault feature extraction. Furthermore, it obtains the confidence level corresponding to the current fault state through the first fault feature vector and the predicted data at the next moment. Finally, the reinforcement learning module selects the optimal control action based on the confidence level to solve the fault problem.

[0138] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0140] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0141] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of high and low voltage switchgear based on the Internet of Things, characterized in that, include: Acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment; Based on the time-series data, the infrared image data, and the preset model, a first fault feature vector is obtained. The preset model includes a convolutional long short-term memory network module, an attention mechanism module, and a transfer learning module. The preset model is obtained through training. Based on the time series data and the digital twin model, the predicted data for the next moment is obtained; Based on the first fault feature vector and the predicted data for the next moment, the confidence level corresponding to the current fault state is obtained; The control action is obtained based on the confidence level and the deep reinforcement learning model, which is obtained through training.

2. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 1, characterized in that, The preset model is obtained through training and includes: Obtain source domain data and target domain data; An iterative operation is performed based on the source domain data and the target domain data. The iterative operation includes obtaining a first feature vector and a second feature vector based on the source domain data, the target domain data, and the feature extractor; obtaining a fault classification loss and a domain classification loss based on the first feature vector and the second feature vector; updating the parameters in the transfer learning module based on the fault classification loss and the domain classification loss, until a termination condition is reached to obtain a first preset model; the transfer learning module includes the feature extractor. Based on the target domain data and Bayesian inference, the parameters in the feature extractor are updated to obtain a second preset model; The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the target domain data and the second preset model to obtain the preset model.

3. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 2, characterized in that, The transfer learning module further includes a label predictor, a gradient reversal layer, and a domain classifier, which obtains fault classification loss and domain classification loss based on the first feature vector and the second feature vector, including: Based on the first feature vector and the label predictor, a second fault feature vector is obtained; The fault classification loss is obtained based on the second fault feature vector; The domain classification loss is obtained based on the second feature vector, the gradient inversion layer, and the domain classifier.

4. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 3, characterized in that, The parameters in the transfer learning module are updated based on the fault classification loss and the domain classification loss, including: The parameters in the label predictor and the feature extractor are updated based on the fault classification loss. The parameters in the domain classifier and the feature extractor are updated based on the domain classification loss.

5. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 2, characterized in that, Based on the target domain data and Bayesian inference, the parameters in the feature extractor are updated to obtain a second preset model, including: Based on the target domain data and Bayesian inference, the posterior probability corresponding to the target domain data is obtained; Based on the fault classification loss and the domain classification loss, the gradient of the total loss with respect to the parameters in the feature extractor is obtained; Based on the posterior probability and the gradient, the parameters in the feature extractor are updated in a weighted manner to obtain the second preset model.

6. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 2, characterized in that, The preset model is obtained by updating the parameters in the convolutional long short-term memory network module and the attention mechanism module based on the target domain data and the second preset model, including: Based on the target domain data and the second preset model, the fault identification result is obtained; The deviation vector is obtained based on the fault identification results and the preset results; The parameters in the convolutional long short-term memory network module and the attention mechanism module are updated based on the deviation vector to obtain the preset model.

7. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 2, characterized in that, Based on the first fault feature vector and the predicted data for the next time step, the confidence level corresponding to the current fault state is obtained, including: Based on the first fault feature vector and the predicted data at the next moment, the first basic allocation probability and the second basic allocation probability are obtained. The confidence level is obtained based on the first basic allocation probability and the second basic allocation probability.

8. The intelligent monitoring method for high and low voltage switchgear based on the Internet of Things according to claim 7, characterized in that, The deep reinforcement learning model is obtained through training and includes: Based on each sample in the source domain data, the basic allocation probability of each fault level is obtained; The confidence level of each fault level is obtained based on the basic allocation probability of each fault level. Based on each sample and the digital twin model, the predicted value corresponding to each sample is obtained; Based on the predicted values ​​and the current deep reinforcement learning model, the reward function corresponding to each control action is obtained; The deep reinforcement learning model is obtained by updating the current deep reinforcement learning model according to the reward function.

9. A method for intelligent monitoring of high and low voltage switchgear based on the Internet of Things according to claim 8, characterized in that, Based on each sample in the source domain data, the basic allocation probability for each fault level is obtained, including: Cluster analysis is performed on each sample in the source domain data to obtain the category corresponding to each fault level; For each category, the correlation between the samples included in it and each fault level is calculated separately; The basic allocation probability for each fault level is obtained based on the correlation.

10. An intelligent monitoring system for high and low voltage switchgear based on the Internet of Things, characterized in that, include: The monitoring module is used to acquire the timing data and infrared image data of the high and low voltage switchgear at the current moment; The fault identification module is used to obtain a first fault feature vector based on the time series data, the infrared image data, and the preset model. The preset model includes a convolutional long short-term memory network module, an attention mechanism module, and a transfer learning module. The preset model is obtained through training. The prediction module is used to obtain the prediction data for the next moment based on the time series data and the digital twin model; The state recognition module is used to obtain the confidence level corresponding to the current fault state based on the first fault feature vector and the prediction data at the next moment. The control module is used to obtain control actions based on the confidence level and the deep reinforcement learning model.

Images