Power transmission and distribution equipment fault detection method and system based on deep learning

Abstract

The invention provides a power transmission and distribution equipment fault detection method and system based on deep learning, and relates to the technical field of deep learning, and the method comprises the steps: synchronously collecting operation electrical parameters, physical state parameters and appearance image data of power transmission and distribution equipment through a multi-source data collection system, and obtaining a standardized data set after preprocessing; a multi-modal deep learning fault detection model comprising a data fusion module, a feature extraction module and a fault classification module is constructed, the data fusion module adopts an attention mechanism to realize cross-modal data fusion, and the feature extraction module combines CNN and LSTM to extract space and time sequence features; then training the model by using a data set containing normal samples and various fault samples to ensure that the classification accuracy of the model reaches the standard; and finally, inputting data acquired in real time into the trained model to realize fault type identification, probability output and alarm, generating a detection report containing a fault position, and periodically updating the model to ensure generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a power transmission and distribution equipment fault detection method and system based on deep learning. BACKGROUND

[0002] As a core component of the power system, the operation state of power transmission and distribution equipment directly affects the safety and stability of power supply. With the development of the power system towards intelligence and scale, the structure of power transmission and distribution equipment is becoming more and more complex, and the operation environment is also showing diversification characteristics. Short circuit, insulation aging, poor contact and other faults occur frequently. If they are not detected and processed in time, they may cause equipment damage, power outage accidents, even threaten personnel safety, and cause huge economic losses. At present, the power transmission and distribution equipment fault detection mainly adopts traditional detection methods and conventional intelligent detection methods. The traditional detection methods include manual inspection, periodic offline test, etc. Manual inspection relies on the experience of operation and maintenance personnel, has strong subjectivity, low efficiency and high missed detection rate, and is difficult to meet the detection demand of large-scale equipment; although periodic offline test can obtain relatively accurate equipment state information, it needs to interrupt the equipment operation, affects the continuity of power supply, and cannot realize real-time fault monitoring. The conventional intelligent detection method is mostly based on single type data (such as electric parameters or vibration data) to construct detection model, for example, fault classification model based on traditional machine learning algorithms such as support vector machine (SVM) and decision tree. This kind of method has obvious limitations: on the one hand, single data cannot fully reflect the equipment operation state, and is easy to cause fault misjudgment or omission due to incomplete data information; on the other hand, the feature extraction ability of traditional machine learning algorithm for complex nonlinear data is limited, and it is difficult to capture the dynamic correlation and time sequence change rule between multiple parameters when the fault of power transmission and distribution equipment occurs. In the scene with various fault types and complex operation environment, the detection accuracy and generalization ability are difficult to meet the actual demand. In addition, although some existing detection methods based on deep learning introduce multi-modal data, they mostly use simple splicing or weighted summation method in data fusion link, and do not fully consider the contribution difference of different modal data to fault detection, resulting in low quality of fused features; at the same time, after the model training is completed, there is no effective updating mechanism, which cannot adapt to the change of equipment operation state and the emergence of new fault types, and the detection accuracy will decrease significantly after long-term use. Therefore, it is urgent to develop a power transmission and distribution equipment fault detection method which can realize efficient fusion of multi-source data, accurately extract fault features, has real-time detection ability and can be dynamically optimized, so as to solve the problems of low detection accuracy, poor real-time performance and weak generalization ability in the prior art, and ensure the safe and stable operation of the power system. SUMMARY

[0003] To solve the above problems, the application provides a power transmission and distribution equipment fault detection method and system based on deep learning, which realizes fault type identification, probability output and alarm, and generates a detection report containing the fault location, while periodically updating the model to ensure the generalization ability. The application can improve the accuracy and real-time performance of power transmission and distribution equipment fault detection, reduce the missed detection and false detection rate, and provide accurate support for equipment operation and maintenance.

[0004] To achieve the above-mentioned purpose, the technical scheme adopted by the application is: In the first aspect of the application, a power transmission and distribution equipment fault detection method based on deep learning is provided, comprising the following steps: A sensor group, a data transmission component and a data preprocessing unit deployed in the power transmission and distribution equipment body and the surrounding environment are used to collect and preprocess multi-source data of the power transmission and distribution equipment. The sensor group includes current sensors, voltage sensors, temperature sensors, vibration sensors and image acquisition devices, which are used to synchronously collect operating electrical parameters, physical state parameters and appearance image data of the power transmission and distribution equipment. The data preprocessing unit performs outlier rejection, missing value completion and data standardization processing on the collected multi-source data to obtain a standardized data set. A fault detection model based on multi-modal deep learning is constructed to fuse multi-source data, extract multi-source data features and classify faults. An attention mechanism is used to fuse cross-modal data of electrical parameter sequence data, physical state parameter sequence data and image data in the standardized data set, and output a fusion feature matrix. CNN sub-modules and LSTM sub-modules are used to extract spatial features and time sequence correlation features in the fusion feature matrix, respectively, and output a multi-dimensional fault feature vector. A fully connected neural network is used to input the multi-dimensional fault feature vector into a Softmax classifier to output a fault type probability distribution of the power transmission and distribution equipment. By real-time collection of multi-source operating data of the target power transmission and distribution equipment, the data is preprocessed and input into the trained fault detection model. The model outputs the fault type and corresponding probability of the target equipment. When the probability of a certain fault type exceeds a preset threshold, a fault alarm signal is triggered, and a fault detection report is generated. The fault detection report at least contains fault type, fault occurrence time, fault probability and fault location positioning information.

[0005] As a preferred, a power transmission and distribution equipment fault sample data set is constructed, which includes normal operation sample data and labeled sample data of five typical faults of the power transmission and distribution equipment, including short circuit fault, insulation aging fault, poor contact fault, mechanical wear fault and overheating fault. The labeled sample data is attached with fault type label and fault severity level label. The fault detection model is trained and optimized, the fault sample data set is divided into a training set, a validation set and a test set according to a preset proportion, a cross-entropy loss function is used as a model loss function, an Adam optimizer is used to iteratively update parameters of the fault detection model until the fault classification accuracy of the fault detection model on the validation set is stabilized at more than 95%, and a trained fault detection model is obtained.

[0006] Preferably, in the preprocessing step, the 3 sigma rule is used to identify outliers in the collected multi-source data, and data outside the range [mu-3sigma, mu+3sigma] is determined as an outlier and removed, where mu is the data mean and sigma is the data standard deviation. For missing values in the data set after removing outliers, if the missing rate is less than 5%, linear interpolation is used to complete it; if the missing rate is between 5%-15%, an interpolation method based on K-neighbor algorithm is used to complete it; if the missing rate is greater than 15%, the data of the corresponding time period is re-collected.

[0007] Preferably, in the attention mechanism, the physical state parameter sequence data and the image data are feature mapped to obtain initial feature vectors of each modality data, to obtain the electrical parameter feature vector V1, the physical state parameter feature vector V2 and the image feature vector V3. An attention weight calculation layer is constructed, V1, V2 and V3 are input into the weight calculation layer, and the attention weights of the feature vectors of each modality are calculated by the following formula: Wherein, Wq is an attention weight matrix, which is adaptively learned through a model training process, W1, W2 and W3 are the attention weights of V1, V2 and V3 respectively. The fused feature matrix F is calculated according to the following formula: ; The weighted fusion of multi-modal data is realized.

[0008] Preferably, the CNN sub-module adopts an improved LeNet-5 network structure, including 3 convolutional layers, 2 pooling layers and 1 flattening layer; wherein the number of convolutional kernels of the first convolutional layer is 32, the size of the convolutional kernel is 3x3, the step is 1, and the padding mode is same; the number of convolutional kernels of the second convolutional layer is 64, the size of the convolutional kernel is 3x3, the step is 1, and the padding mode is same; the number of convolutional kernels of the third convolutional layer is 128, the size of the convolutional kernel is 5x5, the step is 2, and the padding mode is valid; the pooling layer adopts maximum pooling, and the pooling kernel size is 2x2 and the step is 2; the flattening layer converts the feature map output by the pooling layer into a one-dimensional feature vector as the output of the CNN sub-module.

[0009] As preferred, the LSTM submodule comprises 2 hidden layers, the number of neurons of each hidden layer is 256, the Dropout regularization technique is used to inhibit model overfitting, and the Dropout proportion is set to 0.3; the input of the LSTM submodule is the one-dimensional feature vector output by the CNN submodule, after time sequence feature extraction through two hidden layers, the output time sequence correlation feature vector is spliced with the spatial feature vector output by the CNN submodule to form a multi-dimensional fault feature vector.

[0010] More preferably, the cross-entropy loss function is calculated as follows: ; Wherein N is the batch data amount, is the true label of the i th sample, is the prediction probability of the i th sample.

[0011] As preferred, the fault location positioning information is obtained by the following way: inputting the appearance image of the power transmission and distribution equipment collected by the image acquisition device into the CNN submodule of the trained fault detection model, generating a heat map of the image through the Grad-CAM (Gradient-weighted Class Activation Mapping) technology, and the area with the highest brightness in the heat map is the fault location, and the coordinate information of the area and the corresponding equipment component name are included in the fault detection report.

[0012] As preferred, new power transmission and distribution equipment fault sample data is periodically collected, the fault sample data set is expanded, the training process of the model fault detection model is repeated through a preset update period, the parameters of the trained fault detection model are updated to adapt to the changes of the running state of the power transmission and distribution equipment and the occurrence of new fault types, and the detection accuracy and generalization ability of the model are ensured.

[0013] In the second aspect of the present application, a deep learning-based power transmission and distribution equipment fault detection system is also provided, comprising sequentially connected: A data acquisition and processing module is used for acquiring and preprocessing multi-source data of the power transmission and distribution equipment through a sensor group, a data transmission component and a data preprocessing unit arranged in the power transmission and distribution equipment body and the surrounding environment, the sensor group comprises a current sensor, a voltage sensor, a temperature sensor, a vibration sensor and an image acquisition device, and is used for synchronously acquiring running electric parameters, physical state parameters and appearance image data of the power transmission and distribution equipment; the data preprocessing unit performs outlier rejection, missing value completion and data standardization processing on the collected multi-source data to obtain a standardized data set; The fault detection model generation module is used for constructing a multi-modal deep learning-based fault detection model for fusing multi-source data, extracting multi-source data features and classifying faults, adopting an attention mechanism to perform cross-modal data fusion on electrical parameter sequence data, physical state parameter sequence data and image data in a standardized data set, and output a fusion feature matrix; spatial features and time sequence correlation features in the fusion feature matrix are extracted through a CNN sub-module and an LSTM sub-module respectively, and a multi-dimensional fault feature vector is output; the multi-dimensional fault feature vector is input into a Softmax classifier through a fully connected neural network, and a fault type probability distribution of the power transmission and distribution equipment is output. The power transmission and distribution equipment fault detection module is used for collecting multi-source operation data of a target power transmission and distribution equipment in real time, inputting the data after data preprocessing into the trained fault detection model, outputting the fault type and corresponding probability of the target equipment from the model, triggering a fault alarm signal when the probability of a certain fault type exceeds a preset threshold, and generating a fault detection report, wherein the fault detection report at least includes the fault type, fault occurrence time, fault probability and fault location positioning information.

[0014] The present application has the following advantages: The present application adopts a multi-source data acquisition method to synchronously acquire electrical parameters, physical state parameters and appearance image data of the power transmission and distribution equipment, covers multi-dimensional information of equipment operation, avoids detection deviation caused by incomplete single data information, introduces an attention mechanism into the data fusion module to adaptively assign weights to different modal data, highlights features with high contribution to fault detection, and combines the CNN and LSTM constructed feature extraction module to accurately capture the spatial distribution and time sequence correlation of fault features, so that the classification accuracy of the model in multiple fault scenarios is stable at more than 95%, and the false detection rate and the missed detection rate are significantly reduced. The present application realizes real-time fault detection and rapid response by constructing a data transmission module through 5G communication technology combined with an edge computing node, the edge computing node can preliminarily preprocess the collected data, reduce the data transmission amount, reduce the transmission delay, ensure that the real-time fault detection response time is less than 1 second, and when the fault probability exceeds the preset threshold, the local audible and visual alarm and the remote push alarm can be triggered synchronously, the fault information is transmitted to the operation and maintenance personnel in time, and the fault work order is automatically generated to assign tasks, which greatly shortens the fault response and processing time and avoids accident expansion. The present application has the following advantages: The present application uses the Grad-CAM technology to process the equipment appearance image obtained by the image acquisition device, generates a fault area heat map, accurately locates the fault occurrence position, and includes the coordinate information and the corresponding equipment component name in the detection report, which provides clear maintenance guidance for the operation and maintenance personnel, avoids blind troubleshooting, significantly improves the operation and maintenance efficiency, and reduces the operation and maintenance cost. The model can be dynamically optimized to ensure long-term detection performance: This invention sets up a periodic model update step, collecting new fault sample data every 3 months to expand the dataset, retraining the model to update parameters, so that the model can adapt to changes in the operating status of power transmission and distribution equipment and the emergence of new fault types, effectively avoiding the problem of declining generalization ability of the model due to long-term use, ensuring that the detection model has high-precision detection performance for a long time, and extending the practical application life of the method. Wide applicability and compatibility with various types of power transmission and distribution equipment: The multi-source data acquisition system of this invention can flexibly adjust the sensor deployment scheme and data acquisition parameters according to the operating characteristics of different types of power transmission and distribution equipment (such as transformers, circuit breakers, cables, etc.); the fault detection model is trained through multiple types of fault samples and can cover the typical fault types of most power transmission and distribution equipment. There is no need to redesign the detection scheme for a single equipment type. It has wide applicability and can be widely promoted and applied to power system power transmission and distribution equipment fault diagnosis scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of a deep learning-based fault detection method for power transmission and distribution equipment according to the present invention.

[0016] Figure 2 This is a framework diagram of a deep learning-based fault detection system for power transmission and distribution equipment according to the present invention. Detailed Implementation

[0017] Please see Figure 1 As shown, in a first aspect of the present invention, a fault detection method for power transmission and distribution equipment based on deep learning is provided, comprising the following steps: S1. A sensor group, data transmission component, and data preprocessing unit deployed on the power transmission and distribution equipment and its surrounding environment are used to collect and preprocess multi-source data from the power transmission and distribution equipment. The sensor group includes current sensors, voltage sensors, temperature sensors, vibration sensors, and image acquisition devices, used to simultaneously collect the operating electrical parameters, physical state parameters, and appearance image data of the power transmission and distribution equipment. The data preprocessing unit performs outlier removal, missing value completion, and data standardization on the collected multi-source data to obtain a standardized dataset. In this embodiment, a Hall effect current sensor, model CSM010M, with a measurement range of 0~1000A and an accuracy of ±0.5%, is used as the current sensor and is deployed at the output terminals of the high and low voltage windings of the transformer. A voltage transformer, model JDZ10-110, with a measurement range of 0~110kV and an accuracy of 0.2, is used as the voltage sensor and is deployed synchronously with the current sensor. A fiber optic temperature sensor, model FOT-200, with a measurement range of -50℃~200℃ and an accuracy of ±0.3℃, is used and is embedded in the transformer windings, core, and top of the tank. A piezoelectric vibration sensor, model VS-080, with a measurement range of 0~500Hz and a sensitivity of 100mV / g, is used and is installed on the side wall of the transformer tank. An industrial high-definition camera, model MV-CA050-10GM, with a resolution of 5 megapixels and a frame rate of 25fps, is used as the image acquisition device. It supports low-light shooting and is fixed 3m in front of the transformer by a bracket, covering key components such as the transformer tank, bushings, and radiators.

[0018] Through the aforementioned sensor group and data transmission module, data was continuously collected from the transformer for 720 hours (30 days). The collection frequency was set as follows: current and voltage data were collected once every 100ms, temperature data was collected once every 1 minute, vibration data was collected once every 500ms, and image data was collected once every 5 minutes. The collection timestamp and sensor deployment location information were recorded synchronously. A total of approximately 86GB of raw data was obtained, which included normal operation data of the transformer, artificially simulated fault data (short circuit, insulation aging, core overheating, etc.), and historical fault data.

[0019] The collected current, voltage, temperature, and vibration data were processed using the 3σ criterion. Taking the low-voltage side current data as an example, its mean μ=450A and standard deviation σ=25A were calculated. Data less than 375A (μ-3σ) or greater than 525A (μ+3σ) were identified as outliers. A total of 128 outlier data were removed, mainly due to jump data caused by temporary poor contact of the sensor.

[0020] Missing value completion: After statistical analysis of the preprocessed data, it was found that the missing rate of temperature data was 3.2% (due to a brief power outage of the fiber optic sensor), and linear interpolation was used to complete it; the missing rate of vibration data was 8.5%, and the KNN algorithm (K=5) was used for interpolation completion; the missing rate of the remaining data was less than 1%, and it was directly completed by averaging the data before and after completion. After completion, the completeness of the dataset reached 100%. Data standardization: The Z-Score method is used to standardize the completed electrical parameters and physical state parameters. Taking the high-voltage side voltage data as an example, the original data range is 108kV~112kV. After standardization, the mean of the data is 0 and the standard deviation is 1. The image data is normalized by converting the pixel values ​​from [0,255] to [0,1]. At the same time, the image size is unified to 224×224 pixels by the Resize operation to facilitate subsequent model input.

[0021] S2. Construct a fault detection model based on multimodal deep learning to fuse multi-source data, extract features from multi-source data, and classify faults. Employ an attention mechanism to perform cross-modal data fusion of electrical parameter sequence data, physical state parameter sequence data, and image data in a standardized dataset, and output a fused feature matrix. Extract spatial features and temporal correlation features from the fused feature matrix through CNN and LSTM sub-modules, respectively, and output a multi-dimensional fault feature vector. Input the multi-dimensional fault feature vector into a Softmax classifier through a fully connected neural network to output the probability distribution of fault types for power transmission and distribution equipment. Model training and detection are based on Python 3.8 environment, using TensorFlow 2.8 deep learning framework, data processing relies on Pandas 1.5 and NumPy 1.23 libraries, image visualization uses Matplotlib 3.6 library, and it is deployed on a server with Ubuntu 20.04 operating system (CPU: Intel Xeon Gold 6348, GPU: NVIDIA A100, memory 64GB).

[0022] Feature mapping was performed on the standardized electrical parameter sequences (current, voltage), physical state parameter sequences (temperature, vibration), and image data respectively: the electrical parameter sequences were converted into 256-dimensional feature vectors V1 through a one-layer fully connected network, the physical state parameter sequences were converted into 256-dimensional feature vectors V2 through the same structure, and the image data were used to extract 256-dimensional feature vectors V3 through a pre-trained ResNet50 network (with the last 3 layers truncated). An attention weight calculation layer was constructed, and the attention weight matrix Wq was initialized as a 256×256 random matrix. The weights of each modality were calculated using the attention weight formula of each modality feature vector. During training, Wq was adaptively updated. When the system was stable, the average weights of V1, V2, and V3 were 0.35, 0.42, and 0.23, respectively, indicating that the physical state parameters (temperature, vibration) contributed the most to transformer fault detection, which is consistent with the actual situation that transformer faults are often accompanied by temperature rise and abnormal vibration.

[0023] The CNN submodule uses an improved LeNet-5 architecture. The input is a 224×224×3 fused feature matrix, which is reconstructed from the fused feature vector output by the attention mechanism. The first convolutional layer, with 32 3×3 convolutional kernels, outputs a feature map of size 224×224×32. After ReLU activation and 2×2 max pooling, the size becomes 112×112×32. The second convolutional layer, with 64 3×3 convolutional kernels, outputs a 112×112×64 feature map, which becomes 56×56×64 after pooling. The third convolutional layer, with 128 5×5 convolutional kernels, outputs a 52×52×128 feature map, which becomes 26×26×128 after pooling. A flattening layer transforms the feature map into an 86528-dimensional spatial feature vector. LSTM submodule: Sets up 2 hidden layers, each containing 256 neurons, with a Dropout ratio of 0.3. The input is the spatial feature vector output by the CNN submodule. Constructs time-series data with one time step per hour. Extracts time-series correlation features through the LSTM layer and outputs a 256-dimensional time-series feature vector. The spatial feature vector and the time-series feature vector are concatenated to form an 86784-dimensional multi-dimensional fault feature vector.

[0024] A three-layer fully connected neural network is used as the classifier. The first layer has 1024 neurons, the second layer has 512 neurons, and the third layer has 6 neurons (corresponding to 5 types of faults + 1 type of normal state). The activation functions are ReLU, ReLU, and Softmax, respectively, and the probability distribution of each category is output.

[0025] The preprocessed 720 hours of data were divided into samples according to the ratio of "normal operation: short circuit fault: insulation aging fault: poor contact fault: mechanical wear fault: overheating fault = 4:1:1:1:1:1", generating a total of 1440 samples (each sample contains 1 hour of multi-source data). Among them, there were 576 normal samples and 144 samples of each type of fault. Each sample was labeled with a fault type label (such as "normal" or "winding short circuit") and a fault severity label (level 1 to 5, with level 1 being the mildest and level 5 being the most severe).

[0026] The 1440 samples were divided into a training set (1008 samples), a validation set (288 samples), and a test set (144 samples) in a ratio of 7:2:1 to ensure that the fault type distribution of the training set and the test set is consistent. Training parameter settings: The model parameters are initialized using He normality, the batch size is 64, the initial learning rate is 0.001, the learning rate decays by 0.9 times every 10 epochs, the loss function is cross-entropy loss function, the optimizer is Adam, and the number of training epochs is set to 50. Training process monitoring: By the 8th epoch, the classification accuracy on the validation set reached 95.1%; training continued to the 22nd epoch, and the accuracy on the validation set stabilized at 96.8%, with no improvement for 5 consecutive epochs. Training was then stopped, and the current model parameters were saved. At this point, the training set loss value decreased from the initial 1.82 to 0.08, and the validation set loss value decreased from 1.79 to 0.12, with no obvious overfitting. The test set data was input into the trained model, and the test results showed that the model's accuracy in identifying various faults was above 95%. The accuracy in identifying "winding short circuit" faults was the highest (98.6%), while the accuracy in identifying "insulation aging" faults was the lowest (95.2%). The average accuracy was 96.5%, which meets the actual detection requirements.

[0027] First, define a reference value for tracking the evolution trend of sensor behavior scores. This enhances the system's ability to accurately assess score changes. The purpose of this reference value is to mitigate the impact of historical scores by incorporating their influence, thus preventing score fluctuations caused by changes in current at a particular moment. The drastic fluctuations caused a sudden change in the fault assessment, affecting the initial baseline value. Set to 1.

[0028] However, due to the existence of not only short-term fluctuations but also complex fault modes such as "chronic drift," "periodic fluctuations," or "response lag" in the operation of power transmission and distribution systems, relying solely on the data (including average values) input into the fault detection model may still result in missed faults. Therefore, this invention designs a dynamic stability adjustment term based on current derivative. This is used to measure the intensity of current fluctuations in power transmission and distribution equipment over a recent period. It is defined as follows: ; in Differential window length (recommended) ), indicating recent The average absolute value of the first-order difference of current in power transmission and distribution equipment within one second. This is a cached value of historical ratings, where k is a time point. This item reflects whether the rating trend is stable; a higher value indicates more significant equipment fluctuations and a higher probability of failure.

[0029] The current turbulence rating for power transmission and distribution equipment is as follows: ; in, This is the regularization weight factor, with a recommended range of 0.1 to 0.3, set during system deployment. This structure achieves two objectives: firstly, through... Ensure consistency with the system's expected behavior, and then through... Suppress potential collective drift misjudgments. For example, during the start-up phase of power transmission and distribution equipment, if the start-up temperatures of all power transmission and distribution equipment are inconsistent, It may be misjudged as normal at this time. It will be significantly magnified. The score was therefore lowered.

[0030] Ultimately, the above three factors are considered together: the current turbulence score of power transmission and distribution equipment. Smoothing reference Scoring stability adjustment item and group differences The health score is constructed as follows: in: The long-term trend and current performance weights of the current rating of power distribution equipment are controlled separately; Penalties for unstable behavior of other power distribution equipment; Introducing score fluctuation adjustment to suppress score oscillations; all weight coefficients satisfy... The recommended initial value is , , , .

[0031] For example, if power transmission and distribution equipment The behavioral score dropped from 0.8 to 0.4 in the past 5 seconds, and showed significant differences compared to other probes. ), strong fluctuations ( If the instantaneous score is 0.6, the overall health score may be below 0.5, entering the "low confidence" range. In the next step of output control, its weighting coefficient will be significantly reduced, thereby reducing the impact of abnormal data on control commands.

[0032] The trained model is deployed to the substation server, and transformer operation data is collected in real time through a multi-source data acquisition system. The latest pre-processed data is input into the model every 5 minutes, and the model completes the prediction and outputs the result within 1 second.

[0033] S3. By collecting multi-source operating data of the target power transmission and distribution equipment in real time, the data is preprocessed and then input into the trained fault detection model. The model outputs the fault type and corresponding probability of the target equipment. When the probability of a certain fault type exceeds a preset threshold, a fault alarm signal is triggered and a fault detection report is generated. The fault detection report includes at least the fault type, fault occurrence time, fault probability and fault location information. At a certain moment, the data input to the model showed a current of 448A, a voltage of 110.5kV, a winding temperature of 65℃, and a vibration amplitude of 0.12g. The image showed no obvious abnormalities, and the model output a "normal operation" probability of 99.2%. The probabilities of other fault types were all less than 0.5%. The system did not alarm and only recorded the operating data. Fault Scenario: Simulating a transformer "core overheating" fault (the core temperature is raised to 95℃ by heating elements). After the data is input into the model, the output "core overheating" probability is 92.3% (exceeding the preset threshold of 85%). The system immediately triggers a local audible and visual alarm (a red strobe light illuminates and the alarm emits a "beep beep" sound). At the same time, alarm information is pushed to 3 maintenance personnel via SMS and the maintenance APP (content: "Core overheating fault detected in transformer No. 1 of the substation, probability 92.3%, occurrence time 202X-X-XXX:XX, fault location: lower middle part of the core"). A fault work order is generated on the monitoring platform, and maintenance tasks are automatically assigned. The transformer image acquired at that moment was processed using Grad-CAM technology to generate a heat map. The heat map showed that the brightness was highest in the lower part of the iron core (pixel value 0.92, full score 1). Combined with the sensor deployment location (the temperature sensor in the lower part of the iron core collected a temperature of 95℃), the fault location was determined to be in the lower part of the iron core, which was completely consistent with the actual simulated fault location. Based on this information, the maintenance personnel directly carried out maintenance and completed the fault handling in only 40 minutes, saving more than 2 hours compared with the traditional troubleshooting method.

[0034] Three months after implementation, new data (including data on a new type of fault, "bushing leakage") was collected from the transformer and three other transformers of the same type. The sample dataset was expanded to 2,000 samples, and the model training steps were re-executed. The updated model achieved an accuracy of 94.8% in identifying "bushing leakage" faults, and the overall average accuracy improved to 97.2%, verifying the effectiveness of the model update mechanism and ensuring that the model can adapt to new fault types and changes in equipment operating status.

[0035] Please see Figure 2 As shown, in a second aspect of the present invention, a fault detection system for power transmission and distribution equipment based on deep learning is also provided, comprising: The data acquisition and processing module is used to collect and preprocess multi-source data from the power transmission and distribution equipment through sensor groups, data transmission components, and a data preprocessing unit deployed on the equipment itself and its surrounding environment. The sensor groups include current sensors, voltage sensors, temperature sensors, vibration sensors, and image acquisition devices, which are used to simultaneously collect the operating electrical parameters, physical state parameters, and appearance image data of the power transmission and distribution equipment. The data preprocessing unit performs outlier removal, missing value completion, and data standardization on the collected multi-source data to obtain a standardized dataset. The fault detection model generation module is used to construct a fault detection model based on multimodal deep learning. It is used for fusing multi-source data, extracting features from multi-source data, and classifying faults. An attention mechanism is employed to perform cross-modal data fusion of electrical parameter sequence data, physical state parameter sequence data, and image data in a standardized dataset, outputting a fused feature matrix. Spatial and temporal correlation features are extracted from the fused feature matrix through CNN and LSTM sub-modules, respectively, outputting a multi-dimensional fault feature vector. Finally, a fully connected neural network is used to input the multi-dimensional fault feature vector into a Softmax classifier, outputting the probability distribution of fault types for the power transmission and distribution equipment. The power transmission and distribution equipment fault detection module is used to collect multi-source operating data of the target power transmission and distribution equipment in real time. After data preprocessing, the data is input into a trained fault detection model. The model outputs the fault type and corresponding probability of the target equipment. When the probability of a certain fault type exceeds a preset threshold, a fault alarm signal is triggered, and a fault detection report is generated. The fault detection report includes at least the fault type, fault occurrence time, fault probability, and fault location information. In summary, this invention provides an efficient, accurate, and dynamically adjustable method and system for attracting live streaming traffic, which effectively promotes the development of traffic operation on video live streaming platforms and enhances the platform's core competitiveness and market influence.

[0036] This invention can improve the accuracy and real-time performance of fault detection in power transmission and distribution equipment, reduce the rate of missed and false detections, and provide precise support for equipment operation and maintenance.

[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault detection method for power transmission and distribution equipment based on deep learning, characterized in that, Includes the following steps: Sensor arrays, data transmission components, and data preprocessing units deployed on and around the power transmission and distribution equipment are used to collect and preprocess multi-source data from the equipment. The sensor arrays include current sensors, voltage sensors, temperature sensors, vibration sensors, and image acquisition devices, used to simultaneously collect operating electrical parameters, physical state parameters, and appearance image data of the power transmission and distribution equipment. The data preprocessing unit performs outlier removal, missing value completion, and data standardization on the collected multi-source data to obtain a standardized dataset. A fault detection model based on multimodal deep learning is constructed to fuse multi-source data, extract features from multi-source data, and classify faults. An attention mechanism is used to perform cross-modal data fusion of electrical parameter sequence data, physical state parameter sequence data, and image data in a standardized dataset, and output a fusion feature matrix. Spatial features and temporal correlation features are extracted from the fused feature matrix by CNN and LSTM sub-modules respectively, and a multi-dimensional fault feature vector is output. The multi-dimensional fault feature vector is then input into a Softmax classifier through a fully connected neural network to output the probability distribution of fault types of power transmission and distribution equipment. By collecting multi-source operating data of the target power transmission and distribution equipment in real time, the data is preprocessed and then input into the trained fault detection model. The model outputs the fault type and corresponding probability of the target equipment. When the probability of a certain fault type exceeds a preset threshold, a fault alarm signal is triggered and a fault detection report is generated. The fault detection report includes at least the fault type, fault occurrence time, fault probability, and fault location information.

2. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, By constructing a fault sample dataset for power transmission and distribution equipment, the fault sample dataset includes normal operation sample data of power transmission and distribution equipment and labeled sample data of 5 typical faults, including short circuit faults, insulation aging faults, poor contact faults, mechanical wear faults and overheating faults. The labeled sample data is accompanied by fault type labels and fault severity level labels. The fault detection model is trained and optimized by dividing the fault sample dataset into training, validation, and test sets according to a preset ratio. The cross-entropy loss function is used as the model loss function, and the Adam optimizer is used to iteratively update the parameters of the fault detection model until the fault classification accuracy of the fault detection model on the validation set is stable at over 95%, thus obtaining the trained fault detection model.

3. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, In the preprocessing step, the 3σ criterion is used to identify outliers in the collected multi-source data. Data that exceeds the range of [μ-3σ,μ+3σ] are identified as outliers and removed, where μ is the data mean and σ is the data standard deviation. For missing values ​​in the dataset after outlier removal, if the missing rate is less than 5%, linear interpolation is used for completion; if the missing rate is between 5% and 15%, K-nearest neighbor-based interpolation is used for completion; if the missing rate is greater than 15%, data for the corresponding time period is re-collected.

4. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, In the attention mechanism, physical state parameter sequence data and image data are feature-mapped to obtain the initial feature vectors of each modality data, resulting in electrical parameter feature vector V1, physical state parameter feature vector V2, and image feature vector V3. Construct an attention weight calculation layer, inputting V1, V2, and V3 into the weight calculation layer, and calculate the attention weights of each modality feature vector using the following formula: Where Wq is the attention weight matrix, which is obtained through adaptive learning during the model training process, and W1, W2, and W3 are the attention weights of V1, V2, and V3, respectively. The fusion feature matrix F is calculated using the following formula: ； Achieve weighted fusion of multimodal data.

5. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The CNN submodule adopts an improved LeNet-5 network structure, including 3 convolutional layers, 2 pooling layers, and 1 flattening layer. The first convolutional layer has 32 kernels, a kernel size of 3×3, a stride of 1, and uses the same padding method. The second convolutional layer has 64 kernels, a kernel size of 3×3, a stride of 1, and uses the same padding method. The third convolutional layer has 128 kernels, a kernel size of 5×5, a stride of 2, and uses the valid padding method. All pooling layers use max pooling, with a kernel size of 2×2 and a stride of 2. The flattening layer converts the feature maps output by the pooling layers into one-dimensional feature vectors, which are then used as the output of the CNN submodule.

6. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The LSTM submodule includes two hidden layers, each with 256 neurons. Dropout regularization is used to suppress model overfitting, and the Dropout ratio is set to 0.

3. The input of the LSTM submodule is the one-dimensional feature vector output by the CNN submodule. After temporal feature extraction through the two hidden layers, the output is a temporal correlation feature vector, which is concatenated with the spatial feature vector output by the CNN submodule to form a multi-dimensional fault feature vector.

7. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 2, characterized in that, The cross-entropy loss function is calculated as follows: ； Where N is the batch data size. Let i be the true label of the i-th sample. Let be the predicted probability of the i-th sample.

8. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, The fault location information is obtained in the following way: the appearance image of the power transmission and distribution equipment acquired by the image acquisition device is input into the CNN sub-module of the trained fault detection model, and a heat map of the image is generated by gradient weighted class activation mapping technology. The area with the highest brightness in the heat map is the fault location, and the coordinate information of the area and the corresponding equipment component name are included in the fault detection report.

9. The method for fault detection of power transmission and distribution equipment based on deep learning according to claim 1, characterized in that, Regularly collect new fault sample data of power transmission and distribution equipment, expand the fault sample dataset, and repeat the training process of the fault detection model through a preset update cycle to update the parameters of the trained fault detection model, so as to adapt to the changes in the operating status of power transmission and distribution equipment and the emergence of new fault types, and ensure the detection accuracy and generalization ability of the model.

10. A fault detection system for power transmission and distribution equipment based on deep learning, characterized in that, Including interconnected: The data acquisition and processing module is used to collect and preprocess multi-source data from the power transmission and distribution equipment through sensor groups, data transmission components, and a data preprocessing unit deployed on the equipment itself and its surrounding environment. The sensor groups include current sensors, voltage sensors, temperature sensors, vibration sensors, and image acquisition devices, which are used to simultaneously collect the operating electrical parameters, physical state parameters, and appearance image data of the power transmission and distribution equipment. The data preprocessing unit performs outlier removal, missing value completion, and data standardization on the collected multi-source data to obtain a standardized dataset. The fault detection model generation module is used to build a fault detection model based on multimodal deep learning. It is used to fuse multi-source data, extract features from multi-source data, and classify faults. It uses an attention mechanism to perform cross-modal data fusion of electrical parameter sequence data, physical state parameter sequence data, and image data in the standardized dataset, and outputs a fusion feature matrix. Spatial features and temporal correlation features are extracted from the fused feature matrix by CNN and LSTM sub-modules respectively, and a multi-dimensional fault feature vector is output. The multi-dimensional fault feature vector is then input into a Softmax classifier through a fully connected neural network to output the probability distribution of fault types of power transmission and distribution equipment. The power transmission and distribution equipment fault detection module is used to collect multi-source operating data of the target power transmission and distribution equipment in real time, and input the data into the trained fault detection model after data preprocessing. The model outputs the fault type and corresponding probability of the target equipment. When the probability of a certain fault type exceeds a preset threshold, a fault alarm signal is triggered and a fault detection report is generated. The fault detection report includes at least the fault type, fault occurrence time, fault probability and fault location information.

Images