Intelligent early warning system for security threats of power grid system
By employing multi-source vibration signal acquisition and fusion technology and deep learning models, the problem of low diagnostic accuracy in transformer condition monitoring has been solved, enabling efficient and intelligent identification and timely early warning of transformer faults.
Patent Information
- Application Number
- CN202511143553.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-15
- Publication Date
- 2025-11-25
AI Technical Summary
Existing transformer condition monitoring technologies suffer from low diagnostic accuracy and poor signal quality. In particular, traditional methods are prone to lag, and signal quality deterioration is caused by environmental noise interference, making it difficult to detect early faults in a timely manner.
Multi-source vibration signal acquisition and fusion technology is adopted, combined with deep learning model for fault diagnosis. It includes modules for data acquisition, signal denoising, signal fusion, two-dimensional representation and fault diagnosis. Multiple vibration signals of the transformer are acquired by sensor array, and through signal denoising and feature extraction, they are converted into image form for intelligent recognition.
It improves signal integrity and quality, enhances the accuracy and timeliness of transformer fault diagnosis, provides more reliable intelligent early warning protection, and avoids the lag of traditional methods and the limitations of human experience.
Smart Images
Figure CN121009501A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of power grid safety, and in particular to an intelligent early warning system for security threats of a power grid system. BACKGROUND
[0002] With the increasing demand for reliability and safety of power supply in society, the operating environment of the power grid is becoming increasingly complex, and the security threats faced by the power grid are also showing a diversified development trend. Among them, equipment failure, especially the operation anomaly of key node equipment such as transformers, has become one of the important sources of security threats of the power grid. As the core equipment for power transmission and transformation, once a transformer fails, such as winding short circuit, overload heating, insulation breakdown, etc., not only will the equipment itself be damaged, but also may cause cascading trip, voltage fluctuation, large-area power outage and other serious power grid safety incidents, constituting a substantial security threat.
[0003] The existing transformer state monitoring technology faces many challenges in actual application. First, traditional monitoring methods mostly rely on temperature, current or oil quality detection, but such methods often have a lag, making it difficult to discover early or potential failures in a timely manner. Second, although some transformer failure warning methods based on vibration signals have been proposed, the collection process is disturbed by environmental noise and operating noise, resulting in a decline in signal quality, and the feature extraction relies on human experience, which is highly subjective and difficult to fully capture fault information, resulting in low diagnostic accuracy. SUMMARY
[0004] The purpose of the present application is to solve the problem of low diagnostic accuracy mentioned in the background art, and to propose an intelligent early warning system for security threats of a power grid system.
[0005] In the first aspect of the present application, an intelligent early warning system for security threats of a power grid system is provided, which comprises:
[0006] a data acquisition module for acquiring a plurality of vibration signals of a target transformer;
[0007] a signal denoising module for denoising the plurality of vibration signals to obtain a plurality of first target signals;
[0008] a signal fusion module for fusing the plurality of first target signals to obtain a second target signal;
[0009] a two-dimensional representation module for two-dimensional representation of the second target signal to obtain a target image;
[0010] a fault diagnosis module for taking the second target signal and the target image as inputs of a pre-trained state recognition model to obtain the operating state of the target transformer;
[0011] An early warning module is configured to generate and output early warning information if the running state is abnormal.
[0012] Optionally, the data acquisition module synchronously acquires the vibration signals through the sensor array arranged on the surface of the target transformer tank; and the plurality of vibration signals include vibration signals at at least three positions of the top, the bottom, the front and the side.
[0013] Optionally, the signal denoising module includes:
[0014] A signal decomposition module is configured to perform complete ensemble empirical mode decomposition on the target original signal with adaptive noise to obtain a plurality of modal components;
[0015] An entropy calculation module is configured to calculate sample entropy of each modal component;
[0016] A component classification module is configured to divide the plurality of modal components into a high-frequency noise component set, a signal-noise mixed component set and a low-frequency effective component set according to the sample entropy of each modal component;
[0017] A component denoising module is configured to perform wavelet threshold denoising on the modal components in the signal-noise mixed component set to obtain a denoised signal-noise mixed component set;
[0018] A signal reconstruction module is configured to perform signal reconstruction according to the low-frequency effective component set and the denoised signal-noise mixed component set to obtain a denoised vibration signal.
[0019] Optionally, the component classification module includes:
[0020] A primary classification module is configured to divide the plurality of modal components into a noise candidate set, a signal-noise mixed component set and a low-frequency effective component set by using a clustering algorithm;
[0021] A threshold calculation module is configured to calculate a clustering center of the noise candidate set as a noise threshold;
[0022] A secondary division module is configured to include a modal component with sample entropy greater than the noise threshold in the noise candidate set into the high-frequency noise component set, and include a modal component with sample entropy not greater than the noise threshold into the signal-noise mixed component set to obtain a final classification result.
[0023] Optionally, the component denoising module uses an improved threshold function to denoise the modal components, and the improved threshold function is:
[0024] ;
[0025] wherein, is an original wavelet coefficient; is a wavelet coefficient processed by a threshold function; and sgn is a sign function. is the threshold value of the i-th layer; P is the normalized sample entropy; e is a natural constant.
[0026] Optionally, the signal fusion module comprises:
[0027] A similarity strength calculation module is configured to calculate the similarity and correlation strength between any two vibration signals using a cross-correlation function.
[0028] ;
[0029] wherein, is the similarity between signal i and signal j; is the correlation strength between signal i and signal j; L is the signal length; x i [n] and x j [n] are the values of signal i and signal j at the n-th sampling point.
[0030] A weight distribution module is configured to distribute weights according to the correlation strength of each vibration signal with other vibration signals.
[0031] ;
[0032] wherein, E i is the total correlation strength of signal i with other signals; w i is the fusion weight of signal i; M is the number of signals.
[0033] A weighted fusion module is configured to fuse according to the fusion weights of the signals to obtain a second target signal.
[0034] Optionally, the two-dimensional representation module converts the second target signal into a two-dimensional target image through a Markov transition field.
[0035] Optionally, the state recognition model comprises:
[0036] A first branch is configured to extract features of the second target signal using a one-dimensional convolutional neural network to obtain a first feature vector.
[0037] A second branch is configured to extract features of the target image using a residual network to obtain a second feature vector.
[0038] A fusion layer is configured to fuse the first feature vector and the second feature vector using an attention mechanism to obtain a third feature vector.
[0039] A classification layer is configured to output a classification result of the operating state of the transformer using a fully connected network.
[0040] Optionally, the first branch comprises a plurality of cascaded multi-scale convolution modules; the operation process of any one multi-scale convolution module comprises:
[0041] The input feature map is convoluted by multiple parallel convolution layers to obtain feature information under different receptive field scales; the convolution kernel sizes of the multiple parallel convolution layers are 3x1, 5x1 and 7x1 respectively;
[0042] The output feature maps of the multiple convolution layers are spliced in the channel dimension to form a feature map of multi-scale feature fusion;
[0043] The spliced feature map is transformed and compressed in the channel dimension by a 1x1 convolution layer to obtain an output feature map of the multi-scale convolution module.
[0044] Optionally, the second branch includes a spatial compression layer, a residual backbone network and a reshaping layer; wherein:
[0045] The spatial compression layer is configured to perform spatial compression on the input image by using multi-layer stride convolution to obtain a first feature map;
[0046] The residual backbone network is configured to perform feature extraction on the first feature map by using multiple cascaded residual blocks to obtain a second feature map;
[0047] The reshaping layer is configured to convert the second feature map into a one-dimensional feature vector.
[0048] The beneficial effects of the present application are:
[0049] Through the collection and fusion of multi-source vibration signals, the completeness and representativeness of the signals are improved, the defect that a single signal is greatly affected by noise is overcome, and the signal quality is improved through denoising processing. The time domain signal is converted into an image form, combined with a deep learning model, automatic extraction and intelligent identification of complex fault features are realized, the limitations of traditional manual experience are avoided, the accuracy and timeliness of transformer fault diagnosis are improved, and more reliable intelligent early warning protection is provided for power grid safety. BRIEF DESCRIPTION OF DRAWINGS
[0050] Figure 1 An architecture diagram of a power grid system security threat intelligent early warning system is provided for the embodiments of the present application;
[0051] Figure 2 A network architecture diagram of a state recognition model is provided for the embodiments of the present application. DETAILED DESCRIPTION
[0052] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purposes, the specific embodiments, structures, features and effects according to the present application are described in detail below in combination with the drawings and preferred embodiments.
[0053] The embodiment of the present application provides a power grid system security threat intelligent early warning system. Figure 1 , Figure 1 The embodiment of the present application provides a power grid system security threat intelligent early warning system. The system comprises:
[0054] The data acquisition module is used for acquiring a plurality of vibration signals of the target transformer.
[0055] The signal denoising module is used for denoising the plurality of vibration signals to obtain a plurality of first target signals.
[0056] The signal fusion module is used for fusing the plurality of first target signals to obtain a second target signal.
[0057] The two-dimensional representation module is used for two-dimensionally representing the second target signal to obtain a target image.
[0058] The fault diagnosis module is used for taking the second target signal and the target image as inputs of a pre-trained state recognition model (DBMSR, Dual-Branch Multimodal State Recognition Model) to obtain an operation state of the target transformer.
[0059] The early warning module is used for generating and outputting early warning information if the operation state is abnormal.
[0060] The data acquisition module synchronously acquires the vibration signals through a sensor array arranged on the surface of the target transformer box, and the plurality of vibration signals include vibration signals at at least three positions of the top, the bottom, the front and the side. The operation state includes a normal state and abnormal states such as winding loosening, core loosening and winding deformation.
[0061] Based on the power grid system security threat intelligent early warning system provided by the embodiment of the present application, the acquisition and fusion of the multi-source vibration signals improve the integrity and representativeness of the signals, overcome the defect that a single signal is greatly affected by noise, and the signal is denoised to improve the signal quality. The time domain signal is converted into an image form, combined with a deep learning model, automatic extraction and intelligent identification of complex fault characteristics are realized, the limitations of traditional manual experience are avoided, the accuracy and timeliness of the transformer fault diagnosis are improved, and more reliable intelligent early warning protection is provided for the power grid security.
[0062] In one embodiment, three vibration signals are collected at the top, bottom and front of the transformer. The top and bottom are in close mechanical connection with the winding and core, and follow the principle of the shortest path of vibration transmission, so the signals contain direct fault characteristics. The vibration transmission path of the front is relatively indirect, but it can receive the vibration transmitted by the insulating oil, and can supplement key information. The combination of the three can reduce information omission and ensure that the collected signals can fully reflect the mechanical state of the transformer.
[0063] In one implementation, the system further includes a preprocessing module configured to perform missing value supplementing and normalization on the vibration signal.
[0064] In one embodiment, the signal denoising module includes:
[0065] A signal decomposition module configured to perform complete ensemble empirical mode decomposition with adaptive noise on the target original signal to obtain a plurality of modal components.
[0066] An entropy calculation module configured to calculate the sample entropy of each modal component.
[0067] A component classification module configured to divide the plurality of modal components into a high-frequency noise component set, a signal-noise mixed component set and a low-frequency effective component set according to the sample entropy of each modal component.
[0068] A component denoising module configured to perform wavelet threshold denoising on the modal components in the signal-noise mixed component set to obtain a denoised signal-noise mixed component set.
[0069] A signal reconstruction module configured to perform signal reconstruction according to the low-frequency effective component set and the denoised signal-noise mixed component set to obtain a denoised vibration signal.
[0070] This embodiment realizes fine denoising of the vibration signal through complete ensemble empirical mode decomposition (CEEMDAN) decomposition and sample entropy classification. Only the signal-noise mixed components are subjected to wavelet threshold processing, which effectively suppresses high-frequency noise and maximizes the retention of key fault characteristics, thereby improving signal fidelity and diagnostic accuracy.
[0071] In one implementation, the component classification module includes:
[0072] A preliminary classification module configured to divide the plurality of modal components into a noise candidate set, a signal-noise mixed component set and a low-frequency effective component set using a clustering algorithm. Specifically, a K-means clustering algorithm can be used.
[0073] A threshold calculation module configured to calculate the clustering center of the noise candidate set as a noise threshold.
[0074] The secondary division module is configured to include a modal component with sample entropy greater than a noise threshold in a high-frequency noise component set, and include a modal component with sample entropy not greater than the noise threshold in a signal-to-noise mixed component set, to obtain a final classification result.
[0075] The implementation mode realizes initial classification of the modal component by K-means clustering, adaptively calculates a noise threshold, and effectively distinguishes the high-frequency noise and the signal-to-noise mixed component by combining the secondary division of the sample entropy, so as to reduce misclassification and information loss, enhance the denoising effect, and ensure the feature integrity and diagnostic reliability of the vibration signal.
[0076] In one implementation mode, the component denoising module comprises:
[0077] The normalization module is configured to perform normalization processing on sample entropy of the target modal component to obtain normalized sample entropy. The target modal component is any modal component in the signal-to-noise mixed component set.
[0078] The wavelet decomposition module is configured to perform wavelet decomposition on the target modal component to obtain an approximation coefficient of a last layer and detail coefficients of each layer. Specifically, a db4 wavelet basis function can be used, and the number of decomposition layers is 3.
[0079] The hierarchical threshold calculation module is configured to calculate a threshold of each layer of the detail coefficients. Specifically, the formula can be used: to perform threshold calculation. The formula is a general threshold calculation formula, and details are not described herein.
[0080] The threshold processing module is configured to use an improved threshold function to process the detail coefficients to obtain processed detail coefficients. Specifically, the improved threshold function is:
[0081] ;
[0082] wherein, is an original wavelet coefficient; is a wavelet coefficient processed by the threshold function; sgn is a sign function; is a threshold of the i-th layer; P is a normalized sample entropy; and e is a natural constant.
[0083] The modal reconstruction module is configured to perform inverse wavelet transform according to the approximation coefficient and the processed detail coefficient to obtain a denoised modal component.
[0084] The improved threshold function solves the discontinuity of the hard threshold and the deviation of the soft threshold, and realizes adaptive adjustment of the threshold. When there is more noise, the sample entropy P is larger, the function is close to the strong shrinkage soft threshold, and the noise is effectively compressed; when there is more effective information, the sample entropy P is smaller, the function is close to the low-deviation hard threshold, and the effective signal is retained.
[0085] In order to verify the effectiveness of the denoising method proposed in the embodiment, different threshold functions (hard threshold function, soft threshold function, improved threshold function) are used for denoising processing, and the signal-to-noise ratio (SNR) and the root mean square error (RMSE) are used as evaluation indexes, and the comparison results are shown in Table 1. As can be seen from Table 1, the denoising method of the present application has higher signal-to-noise ratio and lower root mean square error. The improvement of the signal-to-noise ratio shows that the improved threshold function can more effectively improve the proportion of effective components in the signal and reduce the influence of noise. The reduction of the root mean square error shows that the deviation between the signal after denoising and the original signal is smaller, and the denoising effect is better.
[0086] Table 1
[0087] Threshold function Signal-to-noise ratio (dB) Root mean square error Hard threshold function 18.575 0.0508 Soft threshold function 19.261 0.0346 Improved threshold function 22.658 0.0147
[0088] In one embodiment, the signal fusion module comprises:
[0089] A similarity strength calculation module is configured to calculate the similarity and correlation strength between any two vibration signals using a cross-correlation function. Specifically, the calculation formula is:
[0090] ;
[0091] Wherein, is the similarity of signal i and signal j; is the correlation strength of signal i and signal j; L is the signal length; x i [n] and x j [n] are the values of signal i and signal j at the nth sampling point;
[0092] A weight distribution module is configured to distribute weights according to the correlation strength of each vibration signal and other vibration signals. Specifically, the calculation formula is:
[0093] ;
[0094] Wherein, E i is the total correlation strength of signal i and other signals; w i is the fusion weight of signal i; M is the number of signals;
[0095] A weighted fusion module is configured to fuse according to the fusion weight of each signal to obtain a second target signal. Specifically, the calculation formula is: .
[0096] The cross-correlation function quantifies the similarity of two signals in time and shape. The embodiment utilizes the characteristic that the signal is not correlated with the noise (the higher the correlation strength, the more obvious the characteristics shared by the two signals, and the more reliable the information), and gives high weight to high correlation signals and low weight to low correlation signals through cross-correlation function calculation. By effectively integrating multi-source information through weighted fusion, the dominant signal characteristics are highlighted, weakly correlated noise is suppressed, the representativeness and stability of the fused signal are improved, and the accuracy and robustness of subsequent fault diagnosis are enhanced.
[0097] In one embodiment, the two-dimensional characterization module can convert the second target signal into a two-dimensional target image through a Markov transition field.
[0098] The embodiment maps the one-dimensional vibration signal into a two-dimensional image using a Markov transition field (MTF), which fully retains the time dependence and dynamic change characteristics of the signal. The MTF can capture the transition probability between different states of the signal, enhance the expression of time series information and highlight local structures, which is conducive to deep learning models to extract rich spatiotemporal features, thereby improving the accuracy and robustness of fault diagnosis.
[0099] In one embodiment, referring to Figure 2 , Figure 2 A network architecture diagram of a state recognition model provided by an embodiment of the present application. The state recognition model comprises:
[0100] The first branch is configured to extract features of the second target signal using a one-dimensional convolutional neural network to obtain a first feature vector. Specifically, the first branch comprises a plurality of cascaded multi-scale convolution modules and a shaping layer. The plurality of multi-scale convolution modules have the same structure, and the shaping layer converts the output of the last multi-scale convolution module into a one-dimensional first feature vector through global average pooling.
[0101] The second branch is configured to extract features of the target image using a residual network to obtain a second feature vector. Specifically, the second branch comprises a spatial compression layer, a residual backbone network, and a shaping layer. The spatial compression layer is configured to compress the input image in space using a plurality of layers of stride convolution to obtain a first feature map; the residual backbone network is configured to extract features of the first feature map using a plurality of cascaded residual blocks to obtain a second feature map; and the shaping layer converts the second feature map into a one-dimensional second feature vector through global average pooling.
[0102] The fusion layer is configured to fuse the first feature vector and the second feature vector using an attention mechanism to obtain a third feature vector. Specifically, the first feature vector and the second feature vector are spliced, and an attention weight is output using a fully connected layer The third feature vector is obtained by fusing the two vectors using attention weights The expression of the calculation process is:
[0103] wherein W and b are learnable parameters of the full connection layer, and Sigmoid is an activation function.
[0104] The classification layer is used to output a classification result of the running state of the transformer using the full connection network.
[0105] In an implementation manner, the operation process of any one multi-scale convolution module comprises:
[0106] The input feature map is subjected to convolution operation using a plurality of parallel convolution layers to obtain feature information under different receptive field scales. The convolution kernel sizes of the plurality of parallel convolution layers are 3x1, 5x1 and 7x1 respectively, and the convolution step lengths are all 2.
[0107] The output feature maps of the plurality of convolution layers are spliced in the channel dimension to form a multi-scale feature fusion feature map.
[0108] The spliced feature map is subjected to channel dimension transformation and compression through a 1x1 convolution layer to obtain an output feature map of the multi-scale convolution module.
[0109] In an implementation manner, the spatial compression layer adopts two layers of 7x7 convolution layers with a step length of 2 for feature compression. It should be noted that when the step length of the convolution layer is not particularly mentioned, the default step length is 1. The residual backbone network adopts a backbone network of ResNet18, and there are a total of 4 residual blocks. ResNet18 is a mature residual network, and thus no more details are given herein.
[0110] The embodiment extracts multi-scale and deep features of one-dimensional vibration signals and two-dimensional target images through the double-branch structure respectively, and fully mines the time sequence and spatial information of the signals. The one-dimensional convolution network multi-scale module captures rich time features, and the residual network effectively extracts deep spatial features of the image, thereby avoiding gradient disappearance. The fusion layer based on the attention mechanism adaptively adjusts the feature weights of the two types, maximizes the complementary advantages, improves the accuracy and discriminability of the feature expression, and enhances the robustness and diagnostic precision of the state recognition model.
[0111] To verify the effectiveness of the state recognition model DBMSR proposed in the embodiment, the state recognition model is compared with the single-modal input model (1D-CNN and ResNet18) in performance, and the accuracy is used as an evaluation index. The specific comparison results are shown in Table 2.
[0112] Table 2
[0113] Model Accuracy (%) 1D-CNN 90.33 ResNet18 93.27 DBMSR 98.65
[0114] From table 2, it can be seen that the state recognition model DBMSR proposed in the application has higher accuracy, which indicates that the overall prediction result is closer to the real situation, reduces false positives and false negatives, and improves the safety guarantee capability of the system.
[0115] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any simple modification, equivalent change and modification of the above embodiments made according to the technical essence of the present application, as long as it does not deviate from the technical solution of the present application, belongs to the scope of the technical solution of the present application.
Claims
1. An intelligent early warning system for security threats to a power grid system, characterized in that, The system includes: The data acquisition module is used to acquire multiple vibration signals from the target transformer. The signal denoising module is used to denoise multiple vibration signals to obtain multiple first target signals; The signal fusion module is used to fuse multiple first target signals to obtain a second target signal; A two-dimensional representation module is used to perform two-dimensional representation on the second target signal to obtain a target image; The fault diagnosis module is used to take the second target signal and the target image as input to a pre-trained state recognition model to obtain the operating state of the target transformer; The early warning module is used to generate and output early warning information if the operating state is abnormal.
2. The intelligent early warning system for power grid system security threats according to claim 1, characterized in that, The data acquisition module synchronously acquires vibration signals through a sensor array deployed on the surface of the target transformer housing; the multiple vibration signals include vibration signals from at least three locations: top, bottom, front, and side.
3. The intelligent early warning system for power grid system security threats according to claim 1, characterized in that, The signal denoising module includes: The signal decomposition module is used to perform full set empirical mode decomposition with adaptive noise on the original target signal to obtain multiple mode components; The entropy calculation module is used to calculate the sample entropy of each modal component; The component classification module is used to divide multiple modal components into a high-frequency noise component set, a signal-noise mixture component set, and a low-frequency effective component set based on the sample entropy of each modal component. The component denoising module is used to perform wavelet threshold denoising on the modal components in the signal-noise mixture component set to obtain a denoised signal-noise mixture component set. The signal reconstruction module is used to reconstruct the signal based on the low-frequency effective component set and the denoised signal-noise mixture component set to obtain the denoised vibration signal.
4. The intelligent early warning system for power grid system security threats according to claim 3, characterized in that, The component classification module includes: The initial classification module is used to divide multiple modal components into a noise candidate set, a signal-noise mixture component set, and a low-frequency effective component set using a clustering algorithm; A threshold calculation module is used to calculate the cluster centers of the noise candidate set as the noise threshold. The secondary partitioning module is used to include modal components whose sample entropy is greater than the noise threshold in the noise candidate set into the high-frequency noise component set, and to include modal components whose sample entropy is not greater than the noise threshold into the signal-noise mixture component set, so as to obtain the final classification result.
5. The intelligent early warning system for power grid system security threats according to claim 3, characterized in that, The component denoising module uses an improved threshold function to denoise the modal components. The improved threshold function is as follows: ； in, These are the original wavelet coefficients; These are the wavelet coefficients after thresholding; sgn is the sign function. is the threshold of the i-th layer; P is the normalized sample entropy; e is the natural constant.
6. The intelligent early warning system for power grid system security threats according to claim 1, characterized in that, The signal fusion module includes: The similarity intensity calculation module is used to calculate the similarity and correlation intensity between any two vibration signals using a cross-correlation function. ； in, It is the similarity between signal i and signal j; It is the correlation strength between signal i and signal j; L is the signal length; x i [n] and x j [n] represents the values of signal i and signal j at the nth sampling point; The weighting module is used to assign weights to each vibration signal based on the correlation intensity with other vibration signals. ； Among them, E i It is the total correlation strength between signal i and other signals; w i is the fusion weight of signal i; M is the number of signals; The weighted fusion module is used to fuse the signals according to their respective fusion weights to obtain the second target signal.
7. The intelligent early warning system for power grid system security threats according to claim 1, characterized in that, The two-dimensional representation module converts the second target signal into a two-dimensional target image through a Markov transformation field.
8. The intelligent early warning system for power grid system security threats according to claim 1, characterized in that, The state recognition model includes: The first branch is used to extract features from the second target signal using a one-dimensional convolutional neural network to obtain a first feature vector. The second branch is used to extract features from the target image using a residual network to obtain a second feature vector. The fusion layer is used to fuse the first feature vector and the second feature vector using an attention mechanism to obtain a third feature vector; The classification layer is used to classify the operating status of transformers using a fully connected network.
9. The intelligent early warning system for power grid system security threats according to claim 8, characterized in that, The first branch includes multiple cascaded multi-scale convolutional modules; the operation process of any one of the multi-scale convolutional modules includes: Multiple parallel convolutional layers are used to perform convolution operations on the input feature map to obtain feature information at different receptive field scales; the kernel sizes of the multiple parallel convolutional layers are 3×1, 5×1, and 7×1, respectively. The output feature maps of multiple convolutional layers are concatenated along the channel dimension to form a feature map with multi-scale feature fusion. The output feature map of this multi-scale convolutional module is obtained by transforming and compressing the channel dimension of the concatenated feature map through a 1×1 convolutional layer.
10. The intelligent early warning system for power grid system security threats according to claim 8, characterized in that, The second branch includes a spatial compression layer, a residual backbone network, and a shaping layer; wherein: The spatial compression layer is used to spatially compress the input image using multi-layer stride convolution to obtain a first feature map. The residual backbone network is used to extract features from the first feature map using multiple cascaded residual blocks to obtain a second feature map; The shaping layer is used to convert the second feature map into a one-dimensional feature vector.