Transformer area fault early warning method based on multi-source data fusion

Abstract

The invention relates to the technical field of low-voltage distribution area fault early warning, in particular to an area fault early warning method based on multi-source data fusion. The method comprises the following steps: acquiring transformer area operation data from a power distribution automation system, a power utilization information acquisition system and a meteorological system, and preprocessing the transformer area operation data by adopting a linear interpolation method and a normalization method; constructing a deep belief network model to perform layer-by-layer feature learning on the multi-source data, and screening out an optimal feature subset; the optimal feature subset is input into an improved dendritic neural network model for fusion prediction, a Leaky ReLU activation function is adopted in a synaptic layer of the model, a Tanh activation function is adopted in a dendritic layer, and summation is carried out on output of the dendritic layer by a film layer to obtain a fault early warning result. Through multi-source data fusion and the improved neural network model, the accuracy and timeliness of transformer area fault early warning are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of low-voltage distribution transformer area fault early warning technology, specifically a transformer area fault early warning method based on multi-source data fusion. Background Technology

[0002] The large-scale grid connection of renewable energy places higher demands on grid stability. As a key link at the end of the grid, the fault early warning of low-voltage distribution substations is crucial to ensuring power supply reliability.

[0003] Existing methods for early warning of transformer substation faults mainly suffer from the following problems: First, they often rely on a single data source and fixed thresholds, making it difficult to comprehensively reflect the operating status of the substation, resulting in low accuracy under multi-factor interference. Second, traditional machine learning methods lack the ability to learn the nonlinear evolution of faults, making it difficult to capture early warning features. Third, deep learning models generally suffer from the vanishing gradient problem during training, especially the low gradient propagation efficiency of the sigmoid activation function in deep networks, leading to difficulties in model convergence. With the increasing complexity of distributed energy access and load characteristics, traditional methods can no longer meet the needs of new power systems in terms of early warning accuracy, real-time performance, and generalization ability. Summary of the Invention

[0004] This invention provides a method for early warning of transformer area faults based on multi-source data fusion, so as to improve the accuracy and timeliness of early warning.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for early warning of transformer area faults based on multi-source data fusion, comprising: S100: Collect transformer operation data from three data sources: transformer area distribution automation system, user electricity consumption information collection system and meteorological system; fill missing values ​​in the transformer operation data using linear interpolation method and normalize the data using bias normalization method to obtain preprocessed multi-source transformer area data; S200: Construct a deep belief network model composed of multiple layers of restricted Boltzmann machines stacked together. Input the preprocessed multi-source transformer area data into the corresponding deep belief network model. Extract deep feature representations through layer-by-layer feature learning. Select the optimal feature subset of each data source based on the deep feature representations. S300: The optimal feature subset is input into the dendritic neural network model for fusion prediction; the dendritic neural network model includes a synaptic layer, a dendritic layer and a membrane layer. The synaptic layer processes the optimal feature subset using the Leaky ReLU activation function and outputs it to the dendritic layer. The dendritic layer performs a multiplication operation on the output of the synaptic layer and processes it using the Tanh activation function before outputting it to the membrane layer. The membrane layer sums the output of the dendritic layer to obtain the transformer area fault warning result.

[0006] As a preferred embodiment of the present invention, the transformer area operation data includes: The three-phase voltage, three-phase current, three-phase power, and power factor of transformers and lines are collected from the distribution automation system.

[0007] The user's three-phase voltage, three-phase current, user electricity consumption, and historical power outage information are collected from the user electricity consumption information collection system. Temperature, humidity, light intensity, and air pressure are collected from the meteorological system.

[0008] As a preferred technical solution of the present invention, the construction steps of the deep belief network model include: constructing corresponding deep belief network models for the substation operation data collected by the substation distribution automation system, the user electricity consumption information collection system and the meteorological system, respectively, wherein each deep belief network model is composed of multiple layers of restricted Boltzmann machines stacked together.

[0009] As a preferred embodiment of the present invention, the construction steps of the deep belief network model further include: stacking multiple restricted Boltzmann machines layer by layer, and connecting a backpropagation neural network to the top layer of the multiple restricted Boltzmann machines.

[0010] As a preferred embodiment of the present invention, the training steps of the deep belief network model include: unsupervised pre-training using the multiple restricted Boltzmann machines, and supervised fine-tuning using the backpropagation neural network.

[0011] As a preferred embodiment of the present invention, the restricted Boltzmann machine includes a visible layer and a hidden layer, and the visible layer and the hidden layer are connected by a bidirectional full connection; the weight matrix and bias vector between the visible layer and the hidden layer are updated using a contrastive divergence algorithm.

[0012] As a preferred technical solution of the present invention, the step of selecting the optimal feature subset of each data source includes: inputting the preprocessed multi-source transformer area data into the trained deep belief network model to obtain deep feature representations, and selecting the optimal feature subset in the feature space that achieves a balance between prediction performance and feature simplification based on the deep feature representations.

[0013] As a preferred technical solution of the present invention, the step of processing the optimal feature subset using the Leaky ReLU activation function in the synaptic layer includes: performing weighted concatenation and thresholding on the input features in the optimal feature subset, performing nonlinear transformation through the Leaky ReLU activation function and outputting it to the dendritic layer; the Leaky ReLU activation function performs linear propagation on positive inputs, introduces positive slope coefficients on negative inputs, and retains gradient information in the negative input region.

[0014] As a preferred technical solution of the present invention, the step of multiplying the output of the dendritic layer to the synaptic layer and processing it with the Tanh activation function includes: obtaining multiple synaptic layer outputs on the same dendritic branch, performing element-wise multiplication on the multiple synaptic layer outputs to obtain a product result, performing a nonlinear transformation on the product result through the Tanh activation function and outputting it to the membrane layer.

[0015] As a preferred embodiment of the present invention, the dendritic neural network model includes multiple dendritic branches, each dendritic branch contains multiple synaptic layers and one dendritic layer, and the membrane layer receives the outputs of all dendritic layers and sums them to obtain the fault warning result of the transformer area.

[0016] The beneficial effects of this invention are: 1. This invention integrates multi-source heterogeneous data from three systems—distribution automation, electricity consumption information, and meteorology—to construct a comprehensive system for sensing the operational status of power distribution areas, overcoming the limitations of insufficient information dimensions from traditional single data sources. Combined with the layer-by-layer feature extraction capabilities of deep belief networks, it can deeply mine fault precursor information contained in multi-source data, significantly improving the comprehensiveness and accuracy of fault identification.

[0017] 2. This invention employs an improved dendritic neural network model. The synaptic layer introduces the Leaky ReLU activation function to preserve gradients in the negative input region, while the dendritic layer uses the Tanh activation function to achieve output centering and efficient gradient propagation. This combination completely solves the gradient vanishing and gradient shift problems of the traditional Sigmoid function in deep network training, enhancing the model's ability to capture high-order interactions of multi-source features and improving the stability and accuracy of fusion prediction.

[0018] 3. This invention uses deep belief networks for feature dimensionality reduction and optimized selection, simplifying model input and improving computational efficiency while maintaining feature representation capabilities. The improved dendritic neural network achieves efficient fusion of multi-source data through parallel processing of dendritic branches and the integration mechanism of membrane layers, significantly outperforming traditional methods in terms of early warning accuracy, timeliness, and computational efficiency. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for early warning of transformer area faults based on multi-source data fusion according to the present invention. Figure 2 This is a diagram of the fault early warning framework for transformer substations provided by the present invention; Figure 3 This is the prediction flowchart provided by the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1: This example provides a method for early warning of transformer area faults based on multi-source data fusion, such as... Figure 1 As shown, the overall technical framework is as follows: Figure 2 As shown. The specific steps are as follows: S100: Collect transformer operation data from three data sources: transformer area distribution automation system, user electricity consumption information collection system and meteorological system; fill missing values ​​in the transformer operation data using linear interpolation method and normalize the data using bias normalization method to obtain preprocessed multi-source transformer area data; Furthermore, the operating data of the transformer area includes: The three-phase voltage, three-phase current, three-phase power, and power factor of transformers and lines are collected from the distribution automation system. The user's three-phase voltage, three-phase current, user electricity consumption, and historical power outage information are collected from the user electricity consumption information collection system. Temperature, humidity, light intensity, and air pressure are collected from the meteorological system.

[0022] Specifically, this step involves collecting transformer operation data from three data sources: the transformer distribution automation system, the user electricity consumption information collection system, and the meteorological system, and then preprocessing the data to obtain preprocessed multi-source transformer data.

[0023] The specific operational data of the transformer substations collected from the three data sources includes: Data collected from the power distribution automation system includes electrical parameters such as three-phase voltage, three-phase current, three-phase power, power factor, and three-phase forward / reverse power of transformers and lines. Data collected from the user electricity consumption information collection system includes: user three-phase voltage, three-phase current, user electricity consumption, power factor, historical power outage time and number of power outages, and other electricity consumption information. Data collected from the meteorological system includes meteorological parameters such as temperature, humidity, light intensity, sunshine hours, and air pressure.

[0024] Build datasets for the three data sources respectively Each dataset consists of vectors labeled with transformer area faults. (Example: Distribution Automation System Dataset) For example, the fault type label includes , respectively, correspond to pole-mounted transformer fault, line 1 to line 5 fault, user 1 to user 5 fault, and normal state, and t represents the time since the current state occurred. For example, if the vector formed by the fault labels is (1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2), it means that the current state occurred 2 hours after the pole-mounted transformer fault.

[0025] Linear interpolation is used to impute missing values ​​in the operational data of power distribution areas. This method is based on the linear relationship between two adjacent known data points and estimates intermediate missing values ​​through linear combination, which can restore the integrity of the original sequence without significantly distorting the data distribution.

[0026] The imputed data is normalized using bias normalization. The maximum value in the original data is selected. and minimum value According to the formula: ; Obtain the normalized value ,in These are the original data values. Normalization maps features with different dimensions and numerical ranges to the [0, 1] interval, eliminating the influence of different feature scales on model training and promoting parameter optimization under the gradient descent algorithm.

[0027] After the above preprocessing steps, the preprocessed multi-source transformer area data is obtained, which serves as the input for the subsequent deep belief network model.

[0028] S200: Construct a deep belief network model composed of multiple layers of restricted Boltzmann machines stacked together. Input the preprocessed multi-source transformer area data into the corresponding deep belief network model. Extract deep feature representations through layer-by-layer feature learning. Select the optimal feature subset of each data source based on the deep feature representations. Furthermore, the construction steps of the deep belief network model include: constructing corresponding deep belief network models for the distribution automation system, user electricity information collection system and meteorological system collected by the distribution area operation data, and each deep belief network model is composed of multiple layers of restricted Boltzmann machines stacked together.

[0029] Furthermore, the construction steps of the deep belief network model also include: stacking multiple restricted Boltzmann machines layer by layer, and connecting a backpropagation neural network to the top layer of the multiple restricted Boltzmann machines.

[0030] Specifically, deep belief network (DBN) models are constructed for the distribution automation system, user electricity consumption information collection system, and meteorological system collecting distribution area operation data. Each DBN model consists of multiple layers of restricted Boltzmann machines (RBMs) stacked together, with a backpropagation (BP) neural network connected to the top layer of multiple RBMs.

[0031] Restricted Boltzmann machines include the visible layer and hidden layers Place them separately Individual and There are 10 neurons. A bidirectional fully connected layer exists between the visible and hidden layers, but there are no connections between neurons within the same layer. Energy is jointly distributed across the two layers. The formula is as follows: ; in, These are the bias vectors for neurons in the visible and hidden layers, respectively. : Visual layer neurons; These are hidden layer neurons; This is the weight matrix between any two neurons in the visible layer and the hidden layer.

[0032] When a visible or hidden layer is activated, the activation probability formula for another layer is as follows: ; ; in, This is the Sigmoid activation function.

[0033] Furthermore, the training steps of the deep belief network model include: unsupervised pre-training using the multiple restricted Boltzmann machines, and supervised fine-tuning using the backpropagation neural network.

[0034] Furthermore, the restricted Boltzmann machine includes a visible layer and a hidden layer, and the visible layer and the hidden layer are connected by a bidirectional full connection; the weight matrix and bias vector between the visible layer and the hidden layer are updated using a contrastive divergence algorithm.

[0035] Specifically, the training of the DBN model is divided into two stages: unsupervised pre-training and supervised fine-tuning.

[0036] Unsupervised pre-training phase: Pre-training is performed layer by layer using a bottom-up approach with multiple layers of Restricted Block Models (RBMs) to ensure optimal mapping at each layer. Specifically, the model inputs preprocessed data into the visible layer of the first RBM in batches, calculates the hidden layer states through forward propagation, and then reconstructs the visible layer data in reverse. The contrastive divergence (CD) algorithm is used to update the weight matrix based on the difference between the original and reconstructed data. and bias vector After the first RBM layer is trained, its hidden layer output is used as the input to the second RBM layer, and so on, to train all RBM layers. The training process for each layer iterates through all data batches, completing one epoch, and repeats this process multiple times until a preset number of iterations is reached. In this embodiment, the preferred number of iterations is 50–100. After training, the weight matrix and bias vector obtained from each layer are output as model parameters.

[0037] Supervised fine-tuning stage: Supervised learning is performed using the top-level BP neural network. The BP neural network propagates error information from top to bottom back to each layer of the RBM, and performs global fine-tuning and optimization of the weight matrix and bias vector of the entire network to reduce prediction error and avoid the model getting trapped in local optima.

[0038] Furthermore, the step of selecting the optimal feature subset from each data source includes: inputting the preprocessed multi-source transformer area data into the trained deep belief network model to obtain deep feature representations, and selecting the optimal feature subset in the feature space that achieves a balance between prediction performance and feature simplification based on the deep feature representations.

[0039] Specifically, the preprocessed multi-source transformer area data is input into the trained DBN model for feature extraction. The model extracts feature information of transformer area fault data layer by layer through its multi-layer hidden structure. Each layer performs a nonlinear transformation on the input based on the weight matrix and bias vector obtained from training, and finally forms a deep feature representation in the highest hidden layer.

[0040] The optimal feature subset is selected in the feature space based on deep feature representation. This process analyzes the contribution of deep features to fault prediction and selects feature combinations that achieve a good balance between prediction performance and feature simplification. Potential fault modes are deeply mined from data in distribution automation systems, user electricity consumption information collection systems, and meteorological systems, ultimately selecting the optimal feature sets for each system and outputting the corresponding prediction data for each system, providing high-dimensional and representative feature input for the subsequent fusion prediction stage. After feature extraction, the fusion prediction stage begins, and the overall prediction process is as follows: Figure 3 As shown.

[0041] S300: The optimal feature subset is input into the dendritic neural network model for fusion prediction; the dendritic neural network model includes a synaptic layer, a dendritic layer and a membrane layer. The synaptic layer processes the optimal feature subset using the Leaky ReLU activation function and outputs it to the dendritic layer. The dendritic layer performs a multiplication operation on the output of the synaptic layer and processes it using the Tanh activation function before outputting it to the membrane layer. The membrane layer sums the output of the dendritic layer to obtain the transformer area fault warning result.

[0042] Furthermore, the step of processing the optimal feature subset using the Leaky ReLU activation function in the synaptic layer includes: performing weighted concatenation and thresholding on the input features in the optimal feature subset, performing nonlinear transformation through the Leaky ReLU activation function, and outputting to the dendritic layer; the Leaky ReLU activation function performs linear propagation on positive inputs, introduces positive slope coefficients on negative inputs, and retains gradient information in the negative input region.

[0043] Specifically, the synaptic layer is responsible for receiving input signals and performing nonlinear transformations. The optimal feature subset output by the DBN model is input to the synaptic layer, where weighted connections and thresholding are applied. The synaptic layer uses the Leaky ReLU activation function for nonlinear transformation and outputs it to the dendritic layer; the calculation formula is as follows: ; in, It is the first input synaptic layer A characteristic signal, Representing the The input feature to the first The weight of each dendritic branch, These are preset gain parameters. It is a threshold. Indicates the first The input feature at the th ... The synaptic layer outputs on each dendritic branch.

[0044] The expression for the Leaky ReLU activation function is: ; in, The positive slope coefficient for the negative input region is preferred in this embodiment. , Indicates "if", Indicates "otherwise", and when When the input is zero, the output is the input itself. When the input is less than or equal to zero, the output is the product of the input and the positive slope coefficient. The Leaky ReLU activation function linearly propagates positive inputs and introduces a positive slope coefficient for negative inputs, ensuring that there is still a non-zero gradient for negative inputs. This allows neurons to participate in learning across the entire input range, effectively avoiding the gradient vanishing problem.

[0045] Furthermore, the step of multiplying the output of the dendritic layer to the synaptic layer and processing it with the Tanh activation function includes: obtaining multiple synaptic layer outputs on the same dendritic branch, performing element-wise multiplication on the multiple synaptic layer outputs to obtain a product result, performing a nonlinear transformation on the product result through the Tanh activation function, and outputting it to the membrane layer.

[0046] Specifically, the dendritic layer integrates the outputs of multiple synaptic layers on the same dendritic branch. (The remaining text appears to be incomplete and requires further context.) On a dendritic branch Each synaptic layer output The outputs from multiple synaptic layers are multiplied element-wise to obtain a product, which is then nonlinearly transformed using the Tanh activation function and output to the membrane layer. The calculation formula for the dendritic layer is as follows: ; in, It is the first The output of the dendritic layer with each dendritic branch, Indicates multiplication operation. This represents the number of synaptic layers on that dendritic branch.

[0047] The expression for the Tanh activation function is: ; The Tanh function has an output range of [-1, 1], which centers the output. Under the same input conditions, its absolute gradient value is greater than that of the Sigmoid function, which can propagate gradients more efficiently and reduce gradient decay in deep networks.

[0048] By combining the synaptic layer Leaky ReLU and the dendritic layer Tanh, Leaky ReLU is responsible for the sparsity of positive input and the preservation of negative input gradients, while Tanh is responsible for output centering and nonlinear fitting, which significantly enhances the adaptability and robustness of the model.

[0049] Furthermore, the dendritic neural network model includes multiple dendritic branches, each dendritic branch containing multiple synaptic layers and one dendritic layer. The membrane layer receives the outputs of all dendritic layers and sums them to obtain the fault warning result for the transformer area.

[0050] Specifically, the membrane layer receives the dendritic layer output from all dendritic branches, and... Output of each dendritic branch Summing yields the final output of the improved dendritic neural network model. The formula for calculating the membrane layer is as follows: ; in, It is the number of dendritic branches. The summation result of the membrane layers represents the final predicted value for the fault warning in the transformer area.

[0051] The improved dendritic neural network model described above integrates the optimal features from three data sources—distribution automation system, user electricity consumption information collection system, and meteorological system—in a high-order manner, fully exploring the inherent correlations among these multi-source features to achieve accurate early warning and location of power outage faults in transformer substations. The prediction results can be converted into actual fault type and time information through inverse normalization.

[0052] Example 2: This example uses actual operating data of a low-voltage distribution transformer area in a certain region as an example to verify the application effect of the transformer area fault early warning method based on multi-source data fusion in transformer area fault early warning.

[0053] The entire 2024 operational data of a residential power distribution area in a certain city was selected as the research object. This area is equipped with one 315kVA transformer, supplying electricity to 98 users. Data collection included electrical parameters of the distribution automation system, user electricity consumption information, and local meteorological data, with a sampling interval of 15 minutes. Approximately 35,000 valid data points were collected throughout the year, including 32,100 data points related to normal operation and 2,900 data points related to faults and anomalies.

[0054] Data preprocessing and feature extraction were performed according to the method described in Example 1. In the DBN feature extraction stage, DBN models were constructed for the three data sources, with the number of hidden layer neurons set to 60, 30, and 15 respectively. After training, the original features were reduced from a total of 42 dimensions to 21 dimensions, simplifying the model complexity while maintaining feature expressive power.

[0055] In the DNM fusion prediction stage, a dendritic neural network model with 5 dendritic branches was constructed. The dataset was divided into training and test sets in a 7:3 ratio for validation. To evaluate the performance of the method of this invention, it was compared with traditional threshold judgment methods, support vector machines (SVM), and long short-term memory networks (LSTM). The results are shown in Table 1.

[0056] surface Comparison of early warning performance of different methods

[0057] As shown in Table 1, the accuracy of the method of this invention reaches 91.2%, and the recall rate reaches 89.5%, both of which are superior to the comparative methods. Compared with the traditional threshold judgment method, the accuracy is improved by 16.7 percentage points; compared with the deep learning method LSTM, the accuracy is improved by 4.4 percentage points. The average early warning time reaches 2.3 hours, providing maintenance personnel with sufficient time for fault handling.

[0058] After applying the method of this invention, the distribution area issued 28 accurate early warnings and 2 false alarms throughout the year, reducing the average power outage time from 3.8 hours to 1.5 hours and improving maintenance efficiency by approximately 30%. This embodiment demonstrates that, through multi-source data fusion and an improved DNM network, the method of this invention can effectively improve the accuracy and timeliness of fault early warning in distribution areas, providing reliable technical support for the safe operation of low-voltage distribution areas.

[0059] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning of transformer area faults based on multi-source data fusion, characterized in that, include: S100: Collects transformer area operation data from three data sources: the transformer area distribution automation system, the user electricity consumption information collection system, and the meteorological system. Missing values ​​in the operation data of the transformer substations are filled using linear interpolation, and normalization is performed using bias normalization to obtain preprocessed multi-source transformer substation data. S200: Construct a deep belief network model composed of multiple layers of restricted Boltzmann machines stacked together. Input the preprocessed multi-source transformer area data into the corresponding deep belief network model. Extract deep feature representations through layer-by-layer feature learning. Select the optimal feature subset of each data source based on the deep feature representations. S300: The optimal feature subset is input into the dendritic neural network model for fusion prediction; the dendritic neural network model includes a synaptic layer, a dendritic layer and a membrane layer. The synaptic layer processes the optimal feature subset using the Leaky ReLU activation function and outputs it to the dendritic layer. The dendritic layer performs a multiplication operation on the output of the synaptic layer and processes it using the Tanh activation function before outputting it to the membrane layer. The membrane layer sums the output of the dendritic layer to obtain the transformer area fault warning result.

2. The method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The operational data of the transformer area includes: The three-phase voltage, three-phase current, three-phase power, and power factor of transformers and lines are collected from the distribution automation system. The user's three-phase voltage, three-phase current, user electricity consumption, and historical power outage information are collected from the user electricity consumption information collection system. Temperature, humidity, light intensity, and air pressure are collected from the meteorological system.

3. The method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The construction steps of the deep belief network model include: constructing corresponding deep belief network models for the operation data of the distribution automation system, user electricity information collection system and meteorological system, respectively. Each deep belief network model is composed of multiple layers of restricted Boltzmann machines stacked together.

4. The method for early warning of transformer area faults based on multi-source data fusion according to claim 3, characterized in that, The construction steps of the deep belief network model further include: stacking multiple restricted Boltzmann machines layer by layer, and connecting a backpropagation neural network to the top layer of the multiple restricted Boltzmann machines.

5. The method for early warning of transformer area faults based on multi-source data fusion according to claim 4, characterized in that, The training steps of the deep belief network model include: unsupervised pre-training using the multiple restricted Boltzmann machines, and supervised fine-tuning using the backpropagation neural network.

6. The method for early warning of transformer area faults based on multi-source data fusion according to claim 5, characterized in that, The restricted Boltzmann machine includes a visible layer and a hidden layer, and the visible layer and the hidden layer are connected by a bidirectional full connection; the weight matrix and bias vector between the visible layer and the hidden layer are updated using a contrastive divergence algorithm.

7. The method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The step of selecting the optimal feature subset from each data source includes: inputting the preprocessed multi-source transformer area data into the trained deep belief network model to obtain deep feature representations, and selecting the optimal feature subset in the feature space that achieves a balance between prediction performance and feature simplification based on the deep feature representations.

8. The method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The step of processing the optimal feature subset using the Leaky ReLU activation function in the synaptic layer includes: performing weighted concatenation and thresholding on the input features in the optimal feature subset, performing nonlinear transformation through the Leaky ReLU activation function, and outputting the result to the dendritic layer; the Leaky ReLU activation function performs linear propagation on positive inputs, introduces positive slope coefficients on negative inputs, and retains gradient information in the negative input region.

9. A method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The steps of multiplying the output of the dendritic layer to the synaptic layer and processing it with the Tanh activation function include: obtaining multiple synaptic layer outputs on the same dendritic branch, performing element-wise multiplication on the multiple synaptic layer outputs to obtain the product result, performing a nonlinear transformation on the product result through the Tanh activation function, and outputting it to the membrane layer.

10. A method for early warning of transformer area faults based on multi-source data fusion according to claim 1, characterized in that, The dendritic neural network model includes multiple dendritic branches, each dendritic branch contains multiple synaptic layers and one dendritic layer, and the membrane layer receives the outputs of all dendritic layers and sums them to obtain the fault warning result of the transformer area.

Images