Rapid intelligent ring main unit partial discharge detection system

Abstract

The invention provides a rapid intelligent ring main unit partial discharge detection system, and belongs to the technical field of ring main unit partial discharge detection, and the rapid intelligent ring main unit partial discharge detection system comprises a partial discharge generator, a data acquisition module and a processor. The partial discharge generator is connected with 10kV voltage and is used for simulating the real working condition when partial discharge of the ring main unit occurs; the data acquisition module is electrically connected with the partial discharge generator, the processor is electrically connected with the data acquisition module, a deep separable convolutional neural network algorithm is contained in the processor, and the model is lightened, so that a neural network runs on embedded equipment and is used for quickly judging whether partial discharge occurs or not; intelligent diagnosis of partial discharge is facilitated, and healthy and stable operation of the ring main unit is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of partial discharge detection of ring main units, and particularly relates to a fast intelligent partial discharge detection system for ring main units. Background Art

[0002] The ring main unit is an important part of the power distribution system. At present, among various faults that occur in the ring main unit, partial discharge is one of the main causes leading to ring main unit failures. Partial discharge is an incomplete breakdown discharge phenomenon caused by internal insulation materials or surface defects during the use of equipment. When there are defects in the insulator, the defective part has a lower breakdown voltage than the surrounding area. When the electric field applied to the insulator reaches a certain intensity, the electric field first causes breakdown of the insulation defect part. Since the duration of the partial discharge signal is only a few nanoseconds to a few microseconds and the generated energy is small, it will not cause complete breakdown of the insulation between the electrodes. However, electrons, ions, etc. generated during the occurrence of the partial discharge phenomenon will continuously damage the insulating medium under the action of the applied electric field, causing the insulating material to slowly decompose. At the same time, chemical substances such as ozone and nitrogen oxides generated by the insulation decomposition accelerate the oxidation and corrosion of the insulating material, and ultimately cause the internal insulation of the ring main unit to completely fail.

[0003] In China, the ring main units are mostly maintained by pre-fault preventive detection and post-fault maintenance detection methods. Due to the scattered installation locations and large numbers of ring main units, the workload of regular preventive detection is extremely large, resulting in a waste of a large amount of time and human cost resources. The preventive detection itself has a long detection cycle, and due to the limitations of detection means, it is difficult to completely detect faults, which has great constraints. However, existing CNN models still have some deficiencies in the task of partial discharge diagnosis, especially from the perspectives of model complexity, portability, and calculation time, mainly including:

[0004] 1. The model is large and difficult to transplant:

[0005] Many CNN models for partial discharge diagnosis often contain a large number of convolutional layers and parameters, especially deep convolutional neural networks and some more complex architectures (such as hybrid convolutional networks, networks with attention mechanisms, etc.). These models usually have millions or even tens of millions of parameters, and the training and inference processes require a large amount of computing resources. Large models require a large amount of memory and storage space. For edge devices or embedded systems with limited hardware resources, it may be very difficult to transplant these models. These complex CNN models usually rely on high-performance servers or dedicated computing resources for training and inference, and for devices that need to be deployed on-site and far from the data center, they may not have enough computing power to support these huge neural networks.

[0006] 2. Long calculation time and poor real-time performance:

[0007] Large CNN models usually require a long computing time for inference, which cannot meet the requirements for application scenarios that need real-time diagnosis of partial discharge. If the computing speed of the CNN model is too slow, some key discharge events may be missed, reducing the timeliness of fault detection. The detection of partial discharge requires strong real-time performance because the partial discharge phenomenon is usually an early signal of equipment insulation deterioration or fault. If it cannot be detected in real time, it may delay the fault warning of the equipment and cause more serious consequences.

[0008] 3. Over-reliance on training data and lack of generalization ability:

[0009] Strong data dependence: The training of large-scale convolutional neural networks usually relies on a large amount of labeled data. For the diagnosis of partial discharge, especially in some specific environments, the collected partial discharge signals often have strong heterogeneity. Therefore, the trained CNN model may only perform well in certain specific data sets or equipment environments, but has poor adaptability to other equipment or environments. This problem of lack of generalization ability limits the practical application of the CNN model, especially in the detection of partial discharge in different power equipment or different environments, with poor portability and applicability. Summary of the Invention

[0010] The purpose of the present invention is to provide a fast intelligent ring main unit partial discharge detection system, aiming to solve the problems of large model size, difficult transplantation, long computing time, poor real-time performance, over-reliance on training data, and lack of generalization ability in the prior art.

[0011] To achieve the above purpose, the present invention provides the following technical solutions:

[0012] A fast intelligent ring main unit partial discharge detection system includes a partial discharge generator, a data acquisition module, and a processor; the partial discharge generator is connected to a 10 kV voltage and is used to simulate the real working conditions when partial discharge occurs in the ring main unit; the data acquisition module is electrically connected to the partial discharge generator, the processor is electrically connected to the data acquisition module, and the processor contains a depthwise separable convolutional neural network algorithm. By making the model lightweight, it realizes the operation of the neural network on embedded devices and is used to quickly judge whether partial discharge occurs.

[0013] As a preferred solution of the present invention, the data acquisition module includes:

[0014] A current acquisition circuit for measuring alternating current;

[0015] An ADC module for the processor to read current data;

[0016] The current acquisition circuit is designed with a clamp structure, wrapped with a copper shielding tape on the outside, and the interface is a waterproof connector. Using the Rogowski coil principle, it can sense signals in the frequency band of 500k - 100MHz in real time, realize the coupling of partial discharge signals of cables in the ring main unit, and output a voltage signal proportional to the current through the internal conversion circuit.

[0017] As a preferred solution of the present invention, the voltage signal needs to be transmitted to the high-speed ADC module through a conditioning circuit. During the transmission process, short cables and shielded wires are used to reduce noise and signal attenuation.

[0018] As a preferred solution of the present invention, the preprocessing of the current data set includes the following steps:

[0019] S1. The preprocessing includes analyzing missing values and outliers in the data set, analyzing the correlation between features, and balancing the data set;

[0020] S2. Input the preprocessed data into the depthwise separable convolution model. Through the convolution method, rich feature information is extracted. Spatial information is extracted through the depth convolution layer, and channel information is fused through the pointwise convolution layer. The output of the pointwise convolution layer will pass through an activation function;

[0021] S3. After batch normalization, pooling layer, residual connection, fully connected layer, and output layer, the output result is obtained.

[0022] As a preferred solution of the present invention, the depthwise separable convolution model uses depth convolution with a convolution kernel size of 3×3. This convolution kernel is a weight matrix, specifically as follows:

[0023]

[0024] As a preferred solution of the present invention, after depthwise separable convolution, BatchNorm and activation function ReLU processing are performed. Among them,

[0025] The calculation process of Batch Norm is specifically as follows:

[0026] S1. For the input data β = {x 1 , …, x m}, a total of m data, the output is y i = BN γ,β (x i ), where β, γ, and β are the scaling parameter and translation parameter respectively;

[0027] S2. First, calculate the mean μ and variance σ 2 of this batch of data x, and the calculation method is as follows:

[0028]

[0029] S3. Then, perform normalization on x to obtain The calculation method is as follows:

[0030]

[0031] S4. Finally, introduce scaling and translation variables γ and β, and after normalization, obtain

[0032] As a preferred solution of the present invention, after the depthwise separable convolution, batch normalization BatchNorm and activation function ReLU processing are performed, where

[0033] The specific calculation process of Batch Norm is as follows:

[0034] S1. The input data is β = {x 1 , …, x m}, a total of m data, and the output is y i = BN γ,β (x i ), and β, γ, and β are the scaling parameter and translation parameter respectively;

[0035] S2. First, calculate the mean μ and variance σ of this batch of data x 2 , and the calculation method is as follows:

[0036]

[0037]

[0038] S3. Then, perform normalization on x to obtain The calculation method is as follows:

[0039]

[0040] S4. Finally, introduce scaling and translation variables γ and β , and after normalization, obtain

[0041] As a preferred solution of the present invention, after the output of MaxPooling, it is input to the fully connected layer, and the real number vector is converted into a probability distribution through the softmax function, so that the probability value of each category is between 0 and 1, and the sum of the probabilities of two categories is 1. The specific calculation method is as follows:

[0042]

[0043] Among them, the partial discharge current is marked with 0, and the normal current is marked with 1.

[0044] As a preferred embodiment of the present invention, the depthwise separable convolution model includes depthwise convolution and pointwise convolution. The depthwise convolution independently performs convolution operations on each input channel to extract the spatial features within the channels. The pointwise convolution uses a 1×1 convolution kernel to perform a linear combination between the channels on the output of the depthwise convolution to integrate the information between the channels. And the model divides the training set and the test set in an 8:2 ratio for training and validation.

[0045] As a preferred embodiment of the present invention, in the case of no GPU support, the processor relies on the efficient computing characteristics of the depthwise separable convolution to achieve fast inference on the embedded device. And when detecting partial discharge in the ring main unit, the correct prediction ratio of the fault samples is not less than 95%, the correct prediction ratio of the normal samples is not less than 95%, the ratio of undetected fault samples is not higher than 5%, and the false alarm ratio of the normal samples is not higher than 5%.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. The present invention uses depthwise separable convolution as an alternative to the standard convolution, which can solve the deficiencies of the existing convolutional neural network in partial discharge diagnosis in multiple aspects. The depthwise separable convolution mainly reduces the computational complexity and the number of model parameters by decomposing the convolution operation into two independent steps: depthwise convolution and pointwise convolution.

[0048] 2. The present invention can reduce the amount of calculation and the size of the model. The standard convolution operation convolves each channel of the input, resulting in a large amount of calculation. While the depthwise separable convolution decomposes the convolution operation into depthwise convolution and pointwise convolution. The depthwise convolution independently convolves each input channel, and each output channel corresponds to a convolution kernel, so that each channel only undergoes one convolution operation. This not only reduces the storage pressure but also improves the calculation efficiency and reduces the consumption of hardware resources.

[0049] 3. The present invention can improve the real-time performance. Since the depthwise separable convolution greatly reduces the amount of calculation, the inference speed is significantly improved, which is particularly important for partial discharge detection because the discharge phenomenon may occur quickly, and the inference delay must be kept at the lowest level. Reducing the amount of calculation not only improves the real-time performance but also enables the device to respond to the partial discharge signal faster.

[0050] 4. The present invention can improve portability and edge computing capabilities. The lightweight feature of depthwise separable convolution enables the model to be more easily deployed to devices with limited computing resources. For power equipment such as ring main units, there are often limitations in storage, computing, and power consumption. Standard convolutional neural networks are too large, while depthwise separable convolution can reduce the model size, making it easier to transplant the model to edge devices or embedded devices to achieve real-time detection of partial discharge. Depthwise separable convolution has high computing efficiency on embedded platforms. Especially in the case without GPU support, due to its high computing efficiency, depthwise separable convolution can achieve efficient inference on hardware better. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the accompanying drawings:

[0052] Figure 1 is a system block diagram of the method for detecting partial discharge in a fast intelligent ring main unit of the present invention;

[0053] Figure 2 is a schematic diagram of the current sensor of the present invention;

[0054] Figure 3 is a schematic diagram of standard convolution of the present invention;

[0055] Figure 4 is a schematic diagram of depthwise separable convolution of the present invention;

[0056] Figure 5 is a schematic diagram of the design of the depthwise separable convolution model of the present invention;

[0057] Figure 6 is the confusion matrix diagram obtained by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] Embodiment 1

[0060] Please refer to Figures 1-6 , the present invention provides the following technical solutions:

[0061] A rapid intelligent partial discharge detection system for ring main units, characterized by comprising a partial discharge generator, a data acquisition module, and a processor; the partial discharge generator is connected to a 10 kV voltage and is used to simulate the real working conditions during the occurrence of partial discharge in the ring main unit; the data acquisition module is electrically connected to the partial discharge generator, and the processor is electrically connected to the data acquisition module. The processor contains a depthwise separable convolutional neural network algorithm, which realizes the operation of the neural network on embedded devices by making the model lightweight and is used to quickly determine whether partial discharge occurs.

[0062] In a specific embodiment of the present invention, the data acquisition module includes a current sensor and an ADC, and after collecting data, it transmits the data to the processor through the ADC. The current sensor is mainly used for cable joints with grounding wires. Using the Rogowski coil principle, it can sense signals in the 500 k - 100 MHz frequency band in real time, realizing the coupling of partial discharge signals of the cables in the ring main unit. The current sensor outputs a voltage signal proportional to the current through an internal conversion circuit, and the output voltage V out is caused by the partial discharge current I in and is obtained through RC integration

[0063]

[0064] where C is the capacitance value, ∫I in (t)dt is the integral of the input current over time, which is actually the charge Q, and R is the equivalent integration resistance.

[0065] Specifically, the data acquisition module includes:

[0066] A current acquisition circuit for measuring alternating current;

[0067] An ADC module for the processor to read the current data;

[0068] The current acquisition circuit is designed with a clamp structure, externally wrapped with a copper shielding tape, and the interface is a waterproof connector. Using the Rogowski coil principle, it can sense signals in the 500 k - 100 MHz frequency band in real time, realizing the coupling of partial discharge signals of the cables in the ring main unit, and outputs a voltage signal proportional to the current through an internal conversion circuit.

[0069] Specifically, the voltage signal needs to be transmitted to the high - speed ADC module through a conditioning circuit. During the transmission process, short cables and shielded wires are used to reduce noise and signal attenuation.

[0070] The depthwise separable convolution model and point convolution are transplanted inside the processor. The depth convolution independently performs convolution operations on each input channel to extract the spatial features within the channel; the point convolution uses a 1×1 convolution kernel to perform a linear combination between channels on the output of the depth convolution and integrate the information between channels. The training set and test set are divided according to 8:2, and it can quickly determine whether partial discharge occurs;

[0071] As Figure 5 shown, the depthwise separable convolution model is designed as follows:

[0072] 1. Input: It means the input part, representing the original data received by the neural network;

[0073] 2. Input Embedding: It means mapping the input data to a high-dimensional feature space, performing preliminary feature extraction on the input data, and is implemented through a fully connected layer;

[0074] 3. DSC Model (depthwise separable convolution model): This is the core part of the neural network. It adopts a combination of DSCblock and Maxpooling for depth feature extraction and dimensionality reduction processing, extracts the local features of the input data, and reduces the computational amount at the same time;

[0075] The core structure includes:

[0076] Depthwise Convolution: Independently perform convolution on each channel to extract the spatial features within the channel;

[0077] Pointwise Convolution: Use 1×1 convolution to fuse channel information and extract cross-channel features;

[0078] Maxpooling: Perform downsampling on the features, compress the features, prevent the data from being too large to reduce the computational amount, and extract local significant features;

[0079] The overall process includes:

[0080] The data sequentially passes through several DSC Blocks for depth feature extraction;

[0081] Add Maxpooling (maximum pooling operation) at specific positions to gradually reduce the spatial dimension of the feature map;

[0082] 4. Fully Connected Layer:

[0083] Flatten the depth features extracted by the DSC Model and perform feature integration through a fully connected layer, which can compress the feature vector and establish the association between the final features and the output classes;

[0084] 5.Output (Final output part):

[0085] Finally, the result is output through Softmax, and the recognition results of normal and partial discharge in the data can be obtained. The real number vector is transformed into a probability distribution through the Softmax function. Mark the partial discharge current with 0 and the normal current with 1, so as to obtain the recognition results of normal and partial discharge in the data, and realize the judgment and output of the partial discharge situation of the ring main unit;

[0086] In the application process of the depthwise separable convolution model, the DSC Block module is used, and the calculation amount is small, which is suitable for lightweight deployment and is more suitable for application in embedded devices compared with standard convolution;

[0087] By attaching Figure 6 From the confusion matrix of the training results shown, it can be seen that the recognition accuracy has reached a good effect. True Positive (TP): Actually a positive class, and the model prediction is also a positive class.

[0088] False Negative (FN): Actually a positive class, but the model predicts it as a negative class (missed detection).

[0089] False Positive (FP): Actually a negative class, but the model predicts it as a positive class (false alarm).

[0090] True Negative (TN): Actually a negative class, and the model prediction is also a negative class.

[0091] TP (upper left) = 0.972: The correct prediction ratio of the model for fault samples is 97.2%.

[0092] FN (upper right) = 0.028: The proportion of fault samples missed by the model is 2.8%.

[0093] FP (lower left) = 0.020: The false alarm ratio of the model for normal samples is 2.0%.

[0094] TN (lower right) = 0.980: The correct prediction ratio of the model for normal samples is 98.0%

[0095] Specifically, the preprocessing of the current dataset includes the following steps:

[0096] S1. The preprocessing includes analyzing the missing values and outliers in the dataset, analyzing the correlation between features, and balancing the dataset;

[0097] S2. Input the preprocessed data into the depthwise separable convolution model, extract rich feature information through convolution, extract spatial information through the depthwise convolution layer, fuse channel information through the pointwise convolution layer, and the output of the pointwise convolution layer will pass through an activation function;

[0098] S3. Obtain the output result after batch normalization, pooling layer, residual connection, fully connected layer, and output layer.

[0099] Specifically, the depthwise separable convolution model uses a depthwise convolution with a convolution kernel size of 3×3. This convolution kernel is a weight matrix, specifically as follows:

[0100] In this embodiment:

[0101] Specifically, after the depthwise separable convolution, batch normalization (Batch Norm) and the activation function ReLU are performed. Among them,

[0102] The calculation process of Batch Norm is specifically as follows:

[0103] S1. For the input data β = {x 1 , …, x m}, a total of m data, the output is y i = BN γ,β (x i ), where β, γ, and β are the scaling parameter and translation parameter respectively;

[0104] S2. First, calculate the mean μ and variance σ of this batch of data x 2 , and the calculation method is as follows:

[0105]

[0106] S3. Then, normalize x to obtain , and the calculation method is as follows:

[0107]

[0108] S4. Finally, introduce the scaling and translation variables γ and β , and after normalization, obtain

[0109] Specifically, during the process of using multiple depthwise separable convolution blocks to extract features, max pooling (MaxPooling) operation is performed to reduce the computational amount. The max pooling operation downsamples the features at specific positions, compresses the feature size, prevents the data from being too large, and extracts local significant features.

[0110] Specifically, after the output of MaxPooling, it is input into the fully connected layer, and the real number vector is converted into a probability distribution through the softmax function, so that the probability value of each category is between 0 and 1, and the sum of the probabilities of two categories is 1. The specific calculation method is as follows:

[0111]

[0112] Among them, the partial discharge current is marked with 0, and the normal current is marked with 1.

[0113] Specifically, in the case of no GPU support, the processor relies on the efficient computing characteristics of depthwise separable convolution to achieve fast inference on embedded devices. When detecting partial discharge in the ring main unit, the correct prediction ratio of fault samples is not less than 95%, the correct prediction ratio of normal samples is not less than 95%, the ratio of missed detection of fault samples is not higher than 5%, and the false alarm ratio of normal samples is not higher than 5%.

[0114] The standard convolution operation will perform convolution on each channel of the input, with a large amount of calculation. While the depthwise separable convolution decomposes the convolution operation into depth convolution and point convolution. The depth convolution performs convolution operations independently on each input channel, and each output channel corresponds to a convolution kernel, so that each channel only performs one convolution operation. The point convolution performs channel mixing through 1x1 convolution to combine the output of the depth convolution. For the same feature map with height H, width W, and C in channels, and an input image with size H×W×C in the calculation amount of the depth convolution part of the depthwise separable convolution is H×W×C in ×K×K, the calculation amount of the point convolution part is, and the total calculation amount is H×W×C in ×(K×K + C out ). At the same time, the standard convolution needs to perform convolution calculations on all input channels, and the calculation amount is H×W×C in ×C out ×K×K; especially in the case of a large size or a large convolution kernel, the advantage of the depthwise separable convolution is more obvious. Similarly, in the application process, the larger the data volume, the more significant the effect of the depthwise separable convolution. Therefore, the depthwise separable convolution reduces the number of parameters in the model. This not only reduces the storage pressure, but also can improve the computing efficiency and reduce the consumption of hardware resources.

[0115] The working principle and usage process of the present invention:

[0116] Step 1: System assembly and connection

[0117] Select a partial discharge generator that can withstand a 10kV voltage and accurately connect it to a 10kV voltage source to ensure that it can stably simulate the real electrical environment of partial discharge in the ring main unit and provide a reliable signal source for subsequent detection;

[0118] The data acquisition module uses a combination of high-precision current sensor and adapted ADC; the current sensor is designed as a clamp-type structure, covered with copper shielding tape to enhance anti-interference ability, and the interface adopts waterproof design to ensure safety and stability; it is based on the principle of Rogowski coil, can keenly capture partial discharge signals in the 500k-100MHz frequency band, and convert the current signal into a proportional voltage signal through the internal conversion circuit; the voltage signal is transmitted to the high-speed ADC module through a carefully designed conditioning circuit, and the transmission cable is short and shielded to minimize noise interference and signal attenuation, ensuring the accuracy and integrity of data acquisition;

[0119] The processor is reliably connected to the data acquisition module, and a deep separable convolutional neural network algorithm model is pre-implanted in the processor. The model is optimized and can efficiently process the collected data to achieve rapid and accurate judgment of partial discharge.

[0120] Data processing and model training steps:

[0121] Step 2: Data preprocessing and model configuration

[0122] The collected current data set is first subjected to comprehensive preprocessing analysis. Professional data processing software or algorithms are used to carefully identify and process missing values ​​and outliers in the data set to ensure data reliability; the correlation between data features is deeply analyzed to mine potential information; and the data set is balanced through appropriate data enhancement or sampling techniques to avoid the adverse effects of data bias on model training.

[0123] The core structure of the depthwise separable convolution model uses a 3×3 convolution kernel for depthwise convolution operations, which serves as a weight matrix:

[0124] The deep convolution layer focuses on extracting the spatial features within each channel of the input data, independently processing the data of each channel, and effectively reducing computational redundancy; the point-by-point convolution layer uses a 1×1 convolution kernel to fuse channel information to achieve cross-channel feature extraction and enhance the model's comprehensive understanding of data; the output of the point-by-point convolution layer is connected to the activation function for nonlinear transformation, and then goes through batch normalization, pooling layer, residual connection, full connection layer and other processing processes in sequence. Batch normalization is based on the following formula for the input data β={x 1 ,...,x m}Processing: Calculate the mean first With variance Normalize x to get (ε is a minimum value to prevent the denominator from being zero), and then introduce the scaling parameter γ and the translation parameter β to obtain This stabilizes the training process and accelerates convergence; the pooling layer uses the max-pooling operation to downsample the feature map at specific positions, compresses the data scale, reduces the computational amount, and highlights local significant features; the residual connection helps to alleviate the problem of gradient disappearance and improve the model training effect; the fully connected layer integrates deep features, establishes a close association with the output categories, and finally converts the real number vector into a probability distribution through the Softmax function. Mark the partial discharge current with 0 and the normal current with 1 to clearly define the discharge state;

[0125] Step 3: Model training and optimization

[0126] Divide the training set and the test set according to the scientific ratio of 8:2, and use optimization algorithms such as mini-batch gradient descent to train the depthwise separable convolution model; during the training process, closely monitor the performance indicators of the model on the validation set, such as accuracy, recall rate, F1 value, etc., and dynamically adjust the model hyperparameters according to the changes in the indicators, including the learning rate, the number of convolutional kernels, the number of network layers, etc.; through multiple iterative trainings, continuously optimize the model weights to enable it to accurately identify partial discharge signals and ensure high reliability and stability in actual detection.

[0127] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A fast intelligent ring main unit partial discharge detection system, characterized in that: It includes a partial discharge generator, a data acquisition module, and a processor; the partial discharge generator is connected to a 10kV voltage to simulate the actual working condition of a ring main unit when partial discharge occurs; the data acquisition module is electrically connected to the partial discharge generator, and the processor is electrically connected to the data acquisition module. The processor contains a deep separable convolutional neural network algorithm, which enables the neural network to run on an embedded device by making the model lightweight, so as to quickly determine whether partial discharge occurs.

2. A fast intelligent ring main unit partial discharge detection system according to claim 1, characterized in that: The data acquisition module comprises: A current acquisition circuit is used for measuring AC current; ADC module, used by the processor to read current data; The current acquisition circuit adopts a clamp-type structure design, is wrapped with a copper shielding tape on the outside, and has a waterproof connector interface. It uses the Rogowski coil principle to sense signals in the 500k-100MHz frequency band in real time, achieves coupling of partial discharge signals of cables in the ring network cabinet, and outputs a voltage signal proportional to the current through an internal conversion circuit.

3. A fast intelligent ring main unit partial discharge detection system according to claim 2, characterized in that: The voltage signal needs to be transmitted to the high-speed ADC module through a conditioning circuit. During the transmission process, short cables and shielded wires are used to reduce noise and signal attenuation.

4. A fast intelligent ring main unit partial discharge detection system according to claim 3, characterized in that: The preprocessing of the current data set comprises the following steps: S1. Preprocessing includes analyzing missing values ​​and outliers in the data set, analyzing the correlation between features, and balancing the data set; S2. Input the preprocessed data into the deep separable convolutional model, extract rich feature information through the convolution method, extract spatial information through the deep convolution layer, fuse channel information through the point-by-point convolution layer, and the output of the point-by-point convolution layer will pass through the activation function; S3, the output result is obtained after batch normalization, pooling layer, residual connection, fully connected layer and output layer.

5. A fast intelligent ring main unit partial discharge detection system according to claim 4, characterized in that: The depthwise separable convolution model uses a depthwise convolution with a 3×3 convolution kernel size, which is a weight matrix as follows:

6. A fast intelligent ring main unit partial discharge detection system according to claim 5, characterized in that: The batch normalization Batch Norm and activation function ReLU processing are performed after the depth-separable convolution, wherein, The calculation process of Batch Norm is as follows: S1, for the input data is β={x1,…,x m }There are m data in total, and the output is y i =BN γ,β (x i ), β, γ and β are scaling parameters and translation parameters respectively, y i is the i-th parameter; S2. First find the mean μ and variance σ of the batch data x 2 , calculated as follows: S3, then normalize x to get The calculation is as follows: S4. Finally, we introduce the scaling and translation variables γ and β and normalize them to get Where: x i represents the i-th data sample in the batch data, Indicates the summation operation on these m data samples.

7. A fast intelligent ring main unit partial discharge detection system according to claim 6, characterized in that: In the process of extracting features using multiple depth-separable convolution blocks, a maximum pooling operation is performed to reduce the amount of calculation. The maximum pooling operation downsamples the features at specific locations, compresses the feature size, prevents the data from being too large, and extracts local significant features.

8. A fast intelligent ring main unit partial discharge detection system according to claim 7, characterized in that: After the MaxPooling output, it is input to the fully connected layer, and the real number vector is converted into a probability distribution through the softmax function, so that the probability value of each category is between 0 and 1, and the sum of the probabilities of the two categories is 1. The specific calculation method is as follows: The local discharge current is marked with 0, and the normal current is marked with 1. i is the i-th element in the input real vector, Indicates x i Perform exponential operations, It is the sum of all elements of the input vector after exponential operation.

9. A fast intelligent ring main unit partial discharge detection system according to claim 8, characterized in that: The depthwise separable convolution model includes depthwise convolution and point convolution. The depthwise convolution performs convolution operations on each input channel independently to extract the spatial features within the channel. The point convolution uses a 1×1 convolution kernel to perform linear combinations of the outputs of the depthwise convolution between channels to integrate the information between channels. The model is trained and verified by dividing the training set and the test set into an 8:2 ratio.

10. A fast intelligent ring main unit partial discharge detection system according to claim 9, characterized in that: The processor achieves fast reasoning on embedded devices without GPU support by relying on the efficient computing characteristics of deep separable convolution. When performing partial discharge detection on ring main units, the correct prediction ratio of fault samples is not less than 95%, the correct prediction ratio of normal samples is not less than 95%, the ratio of missed fault samples is not more than 5%, and the false alarm ratio of normal samples is not more than 5%.

Images