An electrical-non-electrical feature quantity fusion power transformer inter-turn short circuit fault recognition method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD
- Filing Date
- 2026-03-19
- Publication Date
- 2026-07-21
Smart Images
Figure CN122432855A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer fault identification technology, and in particular to a method and system for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristic quantities. Background Technology
[0002] Oil-immersed transformers are core equipment in power systems, and their operational reliability directly affects the safety and stability of the power grid. Statistics show that more than 90% of transformer faults originate from winding faults. Among these, inter-turn short circuits are difficult to detect effectively because the short-circuit loop current is large while the main circuit current does not change significantly. This is because traditional differential protection and overcurrent protection are unable to effectively identify them.
[0003] In existing technologies, inter-turn short-circuit fault identification methods based on a single electrical characteristic (such as the detection method based on the rate of change of current disclosed in Chinese patent CN201910123456.7) judge faults by monitoring electrical parameters such as winding current. However, this method has low sensitivity to early, minor faults and is easily affected by power grid fluctuations and harmonic interference, resulting in a high false positive and false negative rate in practical applications. The reason for this is that a single electrical quantity can only reflect the electromagnetic transient characteristics caused by the fault and cannot capture the physical changes such as thermal effects that accompany the fault development process. The information dimension is limited, making it difficult to comprehensively characterize the fault state. When the fault is minor (such as 1%~2% inter-turn short circuit), the changes in electrical quantities are weak and easily submerged in normal fluctuations, making it difficult for existing methods to achieve reliable early warning. Summary of the Invention
[0004] To address the difficulty in identifying minor faults in existing inter-turn short-circuit fault identification technologies using only electrical quantities, this invention provides a method and system for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristic quantities. This method achieves complementary enhancement of rapid response from electrical quantities and spatial location using non-electrical quantities, significantly improving the identification accuracy of minor inter-turn short circuits. The specific technical solution is as follows: This application provides a method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics, including: S1. Obtain the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing. S2. Preprocess the electrical and non-electrical characteristic quantities respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform, converting the active power loss at the input terminal into a loss waveform, and using the infrared temperature distribution image as a temperature distribution map. S3. Input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. S4. The current feature vector, loss feature vector, and temperature feature vector are spliced and fused in the feature layer to obtain a fused feature vector. S5. Input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.
[0005] Preferably, the acquisition of the electrical and non-electrical characteristics of the target oil-immersed power transformer specifically includes: The three-phase instantaneous current, the active power loss at the input terminal, and the infrared temperature distribution image of the transformer casing are synchronously acquired according to the same time base, and the three-phase instantaneous current and the active power loss at the input terminal are timestamped to form a time sequence correspondence with the acquisition time of the infrared temperature distribution image.
[0006] Preferably, the step of converting the three-phase instantaneous current into a current waveform and converting the active power loss at the input terminal into a loss waveform further includes: Based on the acquisition time of the infrared temperature distribution image, current waveform segments and loss waveform segments of a preset duration are extracted. The preset duration is determined according to the transformer thermal time constant.
[0007] Preferably, the deep convolutional neural network is a ResNet-18 network, whose structure includes, in sequence, a 7×7 convolutional layer, a batch normalization layer, a 3×3 max pooling layer, 8 residual blocks, an activation function layer, a global average pooling layer, and a fully connected layer. The residual blocks include two types: those with 1×1 convolutional layers and those without 1×1 convolutional layers; the activation function is the modified linear unit ReLU.
[0008] Preferably, a method for identifying inter-turn short-circuit faults in power transformers based on the fusion of electrical and non-electrical characteristics further includes: Dilated convolutions are introduced into the residual blocks of the ResNet-18 network. By setting different expansion rates, receptive fields of different scales are used in the branches of the current waveform, loss waveform, and temperature distribution diagram, respectively. The branches of the current waveform and loss waveform are processed using a small expansion ratio to extract local transient features; The branches of the temperature distribution map are processed using a large expansion ratio to extract global spatial distribution features.
[0009] Preferably, the classification module includes a fully connected layer and a Softmax classifier; the number of neurons in the fully connected layer matches the number of fault categories to be identified, and the Softmax classifier is used to convert the output of the fully connected layer into a probability distribution corresponding to each category, and the category corresponding to the maximum probability is taken as the final fault identification result.
[0010] Preferably, the output probability calculation formula of the Softmax classifier is as follows: in, X This represents the input to the fully connected layer, which is a fused feature vector. W j For the first j Class weight vector, P ( y = j | X ) indicates that the sample belongs to the first j The probability of a class.
[0011] Preferably, the inter-turn short-circuit fault identification results include a normal state and multiple categories corresponding to different degrees of inter-turn short-circuit fault states; the different degrees of inter-turn short-circuit fault states at least include inter-turn short-circuit fault states with a short-circuit turn ratio in the range of 1% to 5%.
[0012] Preferably, the step of taking the category corresponding to the maximum probability as the final fault identification result output further includes: The difference between the maximum probability and the second-highest probability is calculated as a confidence index. When the confidence index is lower than a preset threshold, a supplementary sampling instruction is triggered to re-collect data and perform secondary identification.
[0013] This application also provides a power transformer inter-turn short-circuit fault identification system that integrates electrical and non-electrical characteristic quantities, including: The data acquisition unit is used to acquire the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing; The data preprocessing unit is used to preprocess the electrical characteristic quantity and the non-electrical characteristic quantity respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform diagram, converting the active power loss at the input terminal into a loss waveform diagram, and using the infrared temperature distribution image as a temperature distribution diagram. The feature extraction unit is used to input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. The feature fusion unit is used to splice and fuse the current feature vector, loss feature vector and temperature feature vector in the feature layer to obtain a fused feature vector. The classification and identification unit is used to input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses a method and system for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical features. This method deeply fuses electrical features reflecting electromagnetic transients (current waveforms and loss waveforms) with non-electrical features reflecting thermal effects (infrared temperature distribution images). The current waveform is sensitive to the moment of fault occurrence, the loss waveform is strongly correlated with the severity of the fault, and the temperature distribution image has spatial localization capabilities. These three types of features are complementary in terms of time scale, physical nature, and response mechanism, comprehensively characterizing the fault state from multiple dimensions. This effectively overcomes the shortcomings of single parameters, such as susceptibility to interference and incomplete information. Furthermore, three parallel deep convolutional neural network branches are used to extract the three types of image features respectively, and these features are then concatenated and fused at the feature layer. The powerful nonlinear representation capability of deep convolutional neural networks can automatically uncover deep correlation patterns between electrical and non-electrical quantities. Even with slight changes in electrical quantities, the model can still capture fault signs through subtle changes in temperature distribution, significantly improving the detection rate of minor faults. Attached Figure Description
[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0016] Figure 1 A flowchart of a power transformer inter-turn short-circuit fault identification method based on the fusion of electrical and non-electrical characteristics is provided for an embodiment of the present invention.
[0017] Figure 2 This is an example diagram of the high-voltage winding current waveform provided in an embodiment of the present invention.
[0018] Figure 3 This is an example diagram of the fault winding loss waveform provided in an embodiment of the present invention.
[0019] Figure 4 An example diagram of the infrared temperature distribution of a transformer casing provided in an embodiment of the present invention.
[0020] Figure 5 This is a schematic diagram of the ResNet18 model structure provided in an embodiment of the present invention.
[0021] Figure 6 This is a schematic diagram of the training curve of the model provided in an embodiment of the present invention.
[0022] Figure 7 This is a schematic diagram of the extraction of the confusion matrix provided in an embodiment of the present invention.
[0023] Figure 8 This is a schematic diagram illustrating the test results of the model recognition accuracy provided in an embodiment of the present invention.
[0024] Figure 9 This is a schematic diagram of a power transformer inter-turn short-circuit fault identification system that integrates electrical and non-electrical characteristics, provided as an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] It should be understood that, when used in this specification, the terms “comprising” and “including” indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0027] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0028] It should also be further understood that the term "and / or" as used in this specification refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.
[0029] Please refer to the following examples. Figures 1 to 9 .
[0030] Please see Figure 1This application provides a method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics, including: S1. Obtain the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing. In specific implementation, the acquisition methods for electrical and non-electrical characteristic quantities are as follows: a current transformer is installed at the high-voltage side bushing of each phase to acquire the instantaneous value of the three-phase current; the instantaneous value of the three-phase voltage at the transformer input terminal is obtained through the voltage acquisition unit, and the active power loss data is calculated based on the instantaneous values of the three-phase voltage and current; the infrared temperature distribution data of the transformer shell is acquired through an infrared thermal imager fixed at a preset position on the transformer.
[0031] For oil-immersed power transformers, electrical and non-electrical characteristics such as three-phase instantaneous current, incoming terminal voltage, and casing temperature are collected. Current sensors (CTs) are installed on each phase at the high-voltage side bushing of the transformer, with a sampling rate set to ≥10kHz, to collect the instantaneous values of the three-phase current. i A (t) , i B (t) , i C (t) The active power loss is calculated by measuring the voltage and current at the input terminal using a power analyzer, and the calculation formula is as follows: in, p(t) This refers to the instantaneous active power of the three phases; u A (t) , u B (t) , u C (t) This represents the instantaneous value of the three-phase voltage at the incoming line.
[0032] S2. Preprocess the electrical and non-electrical characteristic quantities respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform, converting the active power loss at the input terminal into a loss waveform, and using the infrared temperature distribution image as a temperature distribution map. Specifically, the preprocessing includes performing time-series alignment and noise reduction preprocessing on the three-phase instantaneous current data and active power loss data to generate corresponding two-dimensional waveform images; and performing normalization and size standardization preprocessing on the infrared temperature distribution data to generate a two-dimensional infrared temperature distribution image that matches the size of the waveform image.
[0033] Model training is performed using a multi-source dataset, with training samples primarily consisting of transformer operating history data, experimental data, and simulation data. The electrical and non-electrical quantity data from the training samples are converted into image formats suitable for deep convolutional neural network input, thereby constructing three parallel image subsets: current waveform dataset, loss waveform dataset, and casing temperature dataset. In this embodiment, each subset is configured with samples corresponding to six states. Examples of images generated for each subset are shown below. Figures 2-4 As shown.
[0034] For each image, it is manually labeled according to its corresponding fault state, such as normal, 1% inter-turn short circuit, 2% inter-turn short circuit, 3% inter-turn short circuit, 4% inter-turn short circuit, and 5% inter-turn short circuit, establishing a mapping relationship between images and fault categories. Then, the three labeled subsets are randomly shuffled and divided into training, validation, and test sets in a 7:1:1 ratio. The training set is used for model parameter learning, the validation set for hyperparameter tuning and early stopping detection, and the test set for final model performance evaluation.
[0035] S3. Input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. In this embodiment, a deep convolutional neural network using a deep residual network (ResNet-18) as the basic architecture network is described below. Figure 5 As shown, the input first passes through a 7×7 convolutional layer, then a batch normalization layer, and finally a 3×3 max pooling layer before entering the first residual block. There are two types of residual blocks: one without a 1×1 convolutional layer and the other with a 1×1 convolutional layer. Each residual block consists of a convolutional layer, a batch normalization layer, and an activation function layer. After passing through 8 residual blocks, the output classification is obtained after passing through an activation function layer and a fully connected layer. The Corrected Linear Unit (ReLU) is used as the primary activation function, and its function expression is shown below: The cross-entropy loss function is used to measure the difference between the model's predicted distribution and the true distribution, and this guides parameter optimization. Assuming probability distribution *a* represents the desired output and probability distribution *b* represents the actual output, the cross-entropy is defined as: Three parallel branches of a deep residual network (ResNet-18) with identical structure but independent weights are used to analyze the current waveform, loss waveform, and infrared temperature distribution, respectively. Figure 3Feature extraction is performed on the input images. The current waveform mainly reflects the electromagnetic transient characteristics at the moment of the fault and is sensitive to the time of the fault occurrence; the loss waveform reflects the energy characteristics of the fault and is strongly correlated with the severity of the fault; while the shell temperature distribution map reflects the thermal effects after the fault, has thermal inertia but strong spatial positioning ability. The three types of features are complementary in terms of time scale, physical nature and fault response mechanism.
[0036] S4. The current feature vector, loss feature vector, and temperature feature vector are spliced and fused in the feature layer to obtain a fused feature vector. To achieve complementary enhancement of electrical and non-electrical features, this embodiment fuses the feature vectors output from the three branches at the feature layer. Specifically, the extracted current network features, loss network features, and temperature network features, each with a dimension of 512, are concatenated to form a 1536-dimensional fused feature vector. Through feature layer concatenation and fusion, the rapid response characteristics of electrical quantities and the spatial positioning capabilities of non-electrical quantities can be utilized simultaneously to achieve a comprehensive characterization of inter-turn short-circuit faults.
[0037] S5. Input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.
[0038] Specifically, the classification module includes a fully connected layer and a Softmax classifier; the number of neurons in the fully connected layer matches the number of fault categories to be identified, and the Softmax classifier is used to convert the output of the fully connected layer into a probability distribution corresponding to each category, and the category corresponding to the maximum probability is taken as the final fault identification result.
[0039] In practice, the fused feature vector is input into a fully connected layer, which contains 6 neurons, corresponding to 6 fault categories (normal state, 1% inter-turn short circuit, 2% inter-turn short circuit, 3% inter-turn short circuit, 4% inter-turn short circuit, and 5% inter-turn short circuit).
[0040] Specifically, a Softmax classifier is used to convert the output of the fully connected layer into probability distributions for each category. The calculation formula is as follows: in, X This represents the input to the fully connected layer, which is a fused feature vector. W j For the first j Class weight vector, P ( y = j | X ) indicates that the sample belongs to the first j The probability of each class is calculated. The class with the highest probability is then selected as the model's recognition and prediction result.
[0041] This invention discloses a method and system for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical features. This method deeply fuses electrical features reflecting electromagnetic transients (current waveforms and loss waveforms) with non-electrical features reflecting thermal effects (infrared temperature distribution images). The current waveform is sensitive to the moment of fault occurrence, the loss waveform is strongly correlated with fault severity, and the temperature distribution image has spatial localization capabilities. These three types of features are complementary in terms of time scale, physical nature, and response mechanism, comprehensively characterizing the fault state from multiple dimensions. This effectively overcomes the shortcomings of single parameters being susceptible to interference and lacking complete information. Furthermore, three parallel deep convolutional neural network branches are used to extract the three types of image features respectively, and these features are then concatenated and fused at the feature layer. The powerful nonlinear representation capability of deep convolutional neural networks can automatically uncover deep correlation patterns between electrical and non-electrical quantities. Even with slight changes in electrical quantities, the model can still capture fault signs through subtle changes in temperature distribution, significantly improving the detection rate of minor faults.
[0042] To better illustrate the advantages of the proposed method for identifying inter-turn short-circuit faults in power transformers based on the fusion of electrical and non-electrical characteristics, specific examples are provided below: The dataset samples were divided into six categories: normal samples and samples with 1-5% of the windings experiencing inter-turn short-circuit faults. 300 samples were collected for each category, totaling 1800 samples per dataset and 900 samples per category. These samples were then evenly divided into training, validation, and test sets in a 7:1:1 ratio. All classification models underwent 4000 iterations, using a step-decreasing learning rate scheduler and 14 linear warm-ups. During each iteration, the batch size was 100, with an initial learning rate of 0.01 and a weight decay of 0.01, multiplied by 0.1 every 30 epochs. The final model accuracy was determined by the optimal performance after the multi-parameter dataset fusion identification results stabilized.
[0043] The training curve of the model is as follows Figure 6 As shown, the model began to converge after approximately 1000 iterations, achieving 75% accuracy on the training data. Simultaneously, the accuracy on the validation set also showed a steady improvement, eventually stabilizing at 87.33%. The loss function values on both the training and validation sets showed a decreasing trend, with the validation loss value dropping to approximately 0.3849. Overall, the model's training and validation accuracy steadily improved, while the loss gradually decreased, indicating that the model can effectively learn and generalize to unseen data.
[0044] To examine the detection performance of each model in more detail, extract the confusion matrix as follows: Figure 7As shown in the figure, the distribution of detection results indicates that the model has high accuracy. However, due to the similarity of features between different categories, the model is often misled into making incorrect judgments, which is particularly evident on independent datasets. The final recognition accuracy of the model after stabilization on each class of data is shown below. Figure 8 As shown, the multi-parameter inter-turn fault identification algorithm based on the ResNet18 model has achieved success, both in comparison with similar algorithms and in terms of overall recognition performance improvement. Based on this algorithm, an average classification accuracy of 94.5% was achieved in identifying the degree of inter-turn short-circuit faults in oil-immersed transformer windings.
[0045] As demonstrated by the examples, the method in this embodiment abandons simple feature weighting or early decision fusion strategies, and innovatively adopts a multi-branch deep feature fusion architecture based on an improved ResNet-18. Leveraging the powerful nonlinear representation learning capabilities of deep convolutional neural networks, it automatically and efficiently mines deep and complex correlation patterns between electrical and non-electrical features. The model accuracy is calculated as follows: Ultimately, the model based on this example achieved a stable accuracy of 94.5%, demonstrating its practicality.
[0046] Specifically, in a preferred embodiment of this application, obtaining the electrical and non-electrical characteristics of the target oil-immersed power transformer specifically includes: The three-phase instantaneous current, the active power loss at the input terminal, and the infrared temperature distribution image of the transformer casing are synchronously acquired according to the same time base, and the three-phase instantaneous current and the active power loss at the input terminal are timestamped to form a time sequence correspondence with the acquisition time of the infrared temperature distribution image.
[0047] In practical applications, a satellite positioning system or BeiDou timing module can be configured on the transformer monitoring terminal to provide a unified clock source for the current sensor, power analyzer, and infrared thermal imager. Timestamps are added to the collected three-phase instantaneous current and active power loss data at the input end. The infrared thermal imager also records the corresponding acquisition time when acquiring each frame of temperature distribution image. The timestamps establish a temporal correspondence between electrical and non-electrical characteristics, providing a temporal alignment basis for subsequent multimodal feature fusion.
[0048] This synchronous acquisition and timing alignment mechanism solves the problem of time scale mismatch between electrical and non-electrical quantities, enabling deep neural networks to learn the causal relationship between electrical transients and thermal field distribution evolution, and avoiding feature confusion caused by time misalignment.
[0049] Specifically, the process of converting the three-phase instantaneous current into a current waveform and converting the active power loss at the input terminal into a loss waveform further includes: Based on the acquisition time of the infrared temperature distribution image, current waveform segments and loss waveform segments of a preset duration are extracted. The preset duration is determined according to the transformer thermal time constant.
[0050] The acquisition time of the infrared temperature distribution image Based on the baseline, truncate forward from to The current waveform segments and loss waveform segments within the time interval are used to convert the extracted time-series data into fixed-size current waveform diagrams and loss waveform diagrams, respectively.
[0051] Among them, the preset duration Based on the thermal time constant of the transformer Determined. Thermal time constant. This characterizes the time required for heat generated in the transformer windings to be conducted to the outer casing and reach a stable temperature. Its value is related to factors such as the transformer's capacity, structure, and heat dissipation conditions. In this embodiment, The value range is 0.5 Up to 2 Preferred For a typical oil-immersed power transformer, The value ranges from 1 second to 60 seconds.
[0052] In the interception mechanism of this preferred embodiment, when an inter-turn short circuit occurs in the transformer, the electrical quantities (current, loss) change instantaneously, while the change in the casing temperature has thermal inertia and requires a certain time delay before it can be captured by the infrared thermal imager. If the electrical quantity waveforms at the same moment as the temperature map are used directly, the electrical quantities reflect the steady-state characteristics after the fault occurs, while the temperature map has not yet fully responded, resulting in a time misalignment between the two in terms of physical mechanism. By intercepting the waveform segment corresponding to the thermal time constant, the electrical quantity waveforms input to the model correspond to the heat accumulation process that leads to the current temperature distribution, thereby establishing a physical causal relationship between electrical quantity anomalies and thermal effects, and improving the model's learning efficiency of the fault mechanism.
[0053] Specifically, in a preferred embodiment of this application, a method for identifying inter-turn short-circuit faults in power transformers based on the fusion of electrical and non-electrical characteristics further includes: Dilated convolutions are introduced into the residual blocks of the ResNet-18 network. By setting different expansion rates, receptive fields of different scales are used in the branches of the current waveform, loss waveform, and temperature distribution diagram, respectively. The branches of the current waveform and loss waveform are processed using a small expansion ratio to extract local transient features; The branches of the temperature distribution map are processed using a large expansion ratio to extract global spatial distribution features.
[0054] Dilated convolutions are introduced into the residual blocks of the ResNet-18 network. By setting different dilation rates, the current waveform, loss waveform, and temperature distribution are analyzed. Figure 3 Different receptive fields of different scales are used in each parallel branch to achieve adaptive feature extraction guided by physical mechanisms.
[0055] Dilated convolution expands the receptive field by inserting holes (i.e., zero-filling) between the elements of the convolution kernel. The formula for calculating its equivalent receptive field size is as follows: in, This is the original kernel size. For expansion rate, This represents the equivalent receptive field size. While keeping the number of parameters constant, dilated convolution can exponentially expand the receptive field, making it suitable for extracting features at different scales.
[0056] In this implementation, the specific configuration of the three parallel branches is as follows: (1) Current waveform branch and loss waveform branch: Local transient features are extracted using a small expansion ratio. Current and loss waveforms are derived from one-dimensional time-series signals. Their image characteristics mainly manifest as local variations in the waveform along the time axis, such as spikes, distortions, and abrupt amplitude changes. These local transient features are highly sensitive to the timing and severity of the fault, requiring a small receptive field to maintain time resolution.
[0057] Therefore, in the two branches processing the current waveform and the loss waveform, the dilation rate of the dilated convolution is set to a small value, preferably r=1 or r=2. When r=1, the dilated convolution degenerates into a standard convolution, with the receptive field being the same size as the convolution kernel; when r=2, the equivalent receptive field of the 3×3 convolution kernel expands to 5×5, while still maintaining the ability to finely perceive local waveforms. By setting a small dilation rate, these two branches can focus on local morphological changes in the waveform, accurately capturing the electromagnetic transient characteristics caused by faults.
[0058] (2) Temperature distribution map branch: global spatial distribution features are extracted using a large expansion ratio. The infrared temperature distribution map of the transformer casing reflects the thermal effects after a fault. Its characteristics are manifested in the temperature field distribution throughout the transformer area, including hot spot locations, temperature gradients, and isotherm shapes. These spatial distribution features are global and require a large sensing field to perceive the temperature correlation between different spatial locations.
[0059] Therefore, in the branch processing the temperature distribution map, the dilation rate of the dilated convolution is set to a large value, preferably r=4 or r=8. Taking a 3×3 convolution kernel as an example, when r=4, the equivalent receptive field expands to 9×9; when r=8, the equivalent receptive field expands to 17×17. By setting a large dilation rate, this branch can obtain a receptive field covering the entire transformer area at a relatively shallow network layer, effectively extracting the global spatial distribution characteristics of the thermal field, and achieving precise hotspot location and accurate temperature gradient perception.
[0060] (3) Gradual adjustment strategy of expansion rate To further optimize feature extraction, this implementation can also employ a progressive dilation rate adjustment strategy at different stages of the network: a smaller dilation rate is used in shallow residual blocks to extract local detail features; as the network deepens, the dilation rate is gradually increased to expand the receptive field and extract semantic features. Specifically, for the temperature distribution map branch, r=2 can be set in the first two residual blocks, r=4 in the middle three residual blocks, and r=8 in the last three residual blocks, achieving progressive feature extraction from local to global.
[0061] This preferred embodiment introduces dilated convolution and sets up multi-scale receptive fields. Addressing the inherent physical differences among the three types of input data, it uses a small expansion rate in the electrical quantity branch to extract local transient features and identify waveform abrupt changes; and a large expansion rate in the temperature branch to extract global spatial distribution features and perceive thermal field evolution. This allows the network structure to actively adapt to data characteristics, improving the targeting of feature extraction. Furthermore, by using dilated convolution to flexibly adjust the receptive field while maintaining constant parameter values, the three branches extract features at their respective optimal scales, avoiding information loss due to receptive field mismatch. The electrical quantity branch maintains high-resolution perception, while the temperature branch achieves global contextual understanding. Because the multi-scale receptive field design enables the electrical quantity branch to capture subtle changes in the waveform, and the temperature branch to accurately identify hotspot locations, both types of features are fully extracted at their optimal scales, significantly improving the accuracy of identifying minor inter-turn short circuits and the precision of hotspot location.
[0062] Specifically, in a preferred embodiment of this application, the step of taking the category corresponding to the maximum probability as the final fault identification result output further includes: The difference between the maximum probability and the second-highest probability is calculated as a confidence index. When the confidence index is lower than a preset threshold, a supplementary sampling instruction is triggered to re-collect data and perform secondary identification.
[0063] First, the Softmax classifier outputs the probability distribution of samples belonging to each category. These correspond to six fault states: normal state and short circuit in the 1%~5% range. Let the maximum probability value be... Its corresponding category is The second most likely value is Its corresponding category is Calculate the confidence index. The difference between the highest probability and the second highest probability: Set confidence threshold In this embodiment, the threshold value ranges from 0.2 to 0.4, preferably 0.3. The calculated confidence index... With preset threshold Comparison: like This indicates that the model has a high degree of confidence in the current recognition result, and at this point, it can be directly used... As the final fault identification result output; like This indicates that the model has low confidence in the current recognition result, meaning that the features of the current sample are ambiguous among multiple categories. In this case, the recognition result is not output directly, but a supplementary sampling instruction is triggered.
[0064] After triggering the supplementary sampling command, the system controls the data acquisition module to re-acquire the electrical and non-electrical characteristics of the current transformer, repeating steps S1 to S5 to perform secondary identification on the re-acquired data. After secondary identification, the confidence index is recalculated. If the confidence index of the second identification is still lower than the threshold, the time-series cumulative decision strategy is adopted: the results of the three consecutive identifications are voted on, and the category with the most occurrences is taken as the final output; If the confidence index of the secondary recognition reaches the threshold requirement, the corresponding recognition result will be output directly.
[0065] The aforementioned confidence assessment and supplementary sampling mechanism effectively avoids misjudgments in cases of feature ambiguity or data disturbance, thus improving the reliability of the fault identification system. Furthermore, when the confidence level falls below a threshold, supplementary sampling is automatically triggered to verify the initial results through secondary identification, ensuring both accuracy and avoiding data redundancy caused by a fixed sampling frequency.
[0066] This application also provides a power transformer inter-turn short-circuit fault identification system that integrates electrical and non-electrical characteristic quantities, including: The data acquisition unit is used to acquire the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing; The data preprocessing unit is used to preprocess the electrical characteristic quantity and the non-electrical characteristic quantity respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform diagram, converting the active power loss at the input terminal into a loss waveform diagram, and using the infrared temperature distribution image as a temperature distribution diagram. The feature extraction unit is used to input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. The feature fusion unit is used to splice and fuse the current feature vector, loss feature vector and temperature feature vector in the feature layer to obtain a fused feature vector. The classification and identification unit is used to input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.
[0067] The functional explanation of each unit in this embodiment is the same as that of a power transformer inter-turn short-circuit fault identification method that integrates electrical and non-electrical characteristic quantities, and the technical effect is the same, so it will not be repeated here.
[0068] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.
[0069] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.
[0070] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0071] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the specification of the present invention.
Claims
1. A method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics, characterized in that, include: S1. Obtain the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing. S2. Preprocess the electrical and non-electrical characteristic quantities respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform, converting the active power loss at the input terminal into a loss waveform, and using the infrared temperature distribution image as a temperature distribution map. S3. Input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. S4. The current feature vector, loss feature vector, and temperature feature vector are spliced and fused in the feature layer to obtain a fused feature vector. S5. Input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.
2. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, The acquisition of the electrical and non-electrical characteristics of the target oil-immersed power transformer specifically includes: The three-phase instantaneous current, the active power loss at the input terminal, and the infrared temperature distribution image of the transformer casing are synchronously acquired according to the same time base, and the three-phase instantaneous current and the active power loss at the input terminal are timestamped to form a time sequence correspondence with the acquisition time of the infrared temperature distribution image.
3. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, The process of converting the three-phase instantaneous current into a current waveform and converting the active power loss at the input end into a loss waveform further includes: Based on the acquisition time of the infrared temperature distribution image, current waveform segments and loss waveform segments of a preset duration are extracted. The preset duration is determined according to the transformer thermal time constant.
4. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, The deep convolutional neural network is a ResNet-18 network, whose structure includes, in sequence, a 7×7 convolutional layer, a batch normalization layer, a 3×3 max pooling layer, 8 residual blocks, an activation function layer, a global average pooling layer, and a fully connected layer. The residual blocks include two types: those with 1×1 convolutional layers and those without 1×1 convolutional layers; the activation function is the modified linear unit ReLU.
5. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, Also includes: Dilated convolutions are introduced into the residual blocks of the ResNet-18 network. By setting different expansion rates, receptive fields of different scales are used in the branches of the current waveform, loss waveform, and temperature distribution diagram, respectively. The branches of the current waveform and loss waveform are processed using a small expansion ratio to extract local transient features; The branches of the temperature distribution map are processed using a large expansion ratio to extract global spatial distribution features.
6. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, The classification module includes a fully connected layer and a Softmax classifier. The number of neurons in the fully connected layer matches the number of fault categories to be identified. The Softmax classifier is used to convert the output of the fully connected layer into a probability distribution corresponding to each category, and the category with the highest probability is taken as the final fault identification result.
7. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 6, characterized in that, The formula for calculating the output probability of the Softmax classifier is: in, X This represents the input to the fully connected layer, which is a fused feature vector. W j For the first j Class weight vector, P ( y = j | X ) indicates that the sample belongs to the first j The probability of a class.
8. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 6, characterized in that, The inter-turn short-circuit fault identification results include a normal state and multiple categories corresponding to different degrees of inter-turn short-circuit fault states; the different degrees of inter-turn short-circuit fault states include at least inter-turn short-circuit fault states with a short-circuit turn ratio in the range of 1% to 5%.
9. The method for identifying inter-turn short-circuit faults in power transformers by fusing electrical and non-electrical characteristics according to claim 1, characterized in that, The step of taking the category corresponding to the maximum probability as the final fault identification result output further includes: The difference between the maximum probability and the second-highest probability is calculated as a confidence index. When the confidence index is lower than a preset threshold, a supplementary sampling instruction is triggered to re-collect data and perform secondary identification.
10. A power transformer inter-turn short-circuit fault identification system based on the fusion of electrical and non-electrical characteristic quantities, characterized in that, The method described by any one of claims 1-9 comprises: The data acquisition unit is used to acquire the electrical and non-electrical characteristics of the target oil-immersed power transformer; the electrical characteristics include at least the three-phase instantaneous current data and active power loss data of the transformer windings, and the non-electrical characteristics include at least the infrared temperature distribution data of the transformer casing; The data preprocessing unit is used to preprocess the electrical characteristic quantity and the non-electrical characteristic quantity respectively. The preprocessing includes: converting the three-phase instantaneous current into a current waveform diagram, converting the active power loss at the input terminal into a loss waveform diagram, and using the infrared temperature distribution image as a temperature distribution diagram. The feature extraction unit is used to input the current waveform, loss waveform, and infrared temperature distribution map into three independent branches of a deep convolutional neural network, and extract the corresponding current feature vector, loss feature vector, and temperature feature vector through each branch. The feature fusion unit is used to splice and fuse the current feature vector, loss feature vector and temperature feature vector in the feature layer to obtain a fused feature vector. The classification and identification unit is used to input the fused feature vector into the classification module, and output the inter-turn short-circuit fault identification result of the target oil-immersed power transformer through the classification module.