A high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors

Abstract

The application relates to the field of switch cabinet arc fault diagnosis, in particular to a high-voltage switch cabinet arc fault diagnosis system based on an intelligent sensor, which comprises the following modules: a sensor data coding fusion module, which generates a coding matrix through multi-scale wavelet transformation of time sequence collection data of an arc light sensor, a current sensor and an ultrasonic sensor of a high-voltage switch cabinet; an arc feature extraction module, which generates an arc feature vector through an arc feature extraction backbone network based on a conditional parameter convolution and an attention mechanism architecture; a fault cause classification module, which generates an arc fault cause through a fault cause classification head; a fault trend prediction module, which generates an arc fault trend prediction value through a fault trend regression head; and a fault diagnosis and early warning module, which performs arc fault diagnosis management. The application can adaptively focus on arc fault features and output diagnosed faults and predicted fault trends, thereby guaranteeing the diagnosis stability of high-voltage switch cabinet arc faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of switch cabinet arc fault diagnosis, and particularly relates to a high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors. BACKGROUND

[0002] As a core equipment of power system transmission and distribution links, the high-voltage switch cabinet is widely used in power plants, substations, industrial plants and urban power distribution networks, and undertakes key functions such as circuit on-off control, power distribution and fault isolation, so its operation stability directly determines the power supply reliability and safety of the power system.

[0003] Among them, the arc fault of the high-voltage switch cabinet is the most destructive fault type. When the insulation structure in the cabinet fails, such as insulation aging, partial discharge, external foreign matter invasion, etc., it is easy to cause interphase or ground arc, and further damage the equipment parts in the cabinet. Therefore, the early and accurate diagnosis and trend warning of the arc fault of the high-voltage switch cabinet have become the core research direction in the field of power equipment state monitoring.

[0004] At present, with the development of intelligent sensing technology and multi-source information fusion method, combined with multiple sensors such as ultrasonic wave, temperature, ground electric wave, etc., joint diagnosis has become an effective way to improve fault recognition accuracy. However, the existing system still has the following problems in the fusion processing of multi-source heterogeneous sensing data: first, the sampling frequency and signal characteristics of different sensors are quite different, and it is difficult to realize time alignment and feature coordination; second, the simple convolution mapping model has limited expression ability for arc features under complex working conditions; third, most diagnosis models only focus on fault classification, lack of quantitative prediction of fault evolution trend, and cannot fully support state maintenance decision.

[0005] Therefore, how to realize the comprehensive prediction of the cause of arc insulation failure and the development trend of the fault based on multi-modal sensors to provide prospective warning for the high-voltage switch cabinet is a technical problem to be solved at present. SUMMARY

[0006] Therefore, the present application provides a high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors, which adopts multi-scale wavelet transform to effectively extract key details and transient features in time series data of different sensors, and an arc feature extraction backbone network based on conditional parameter convolution and attention mechanism, which can adaptively focus on the most relevant signal features of arc fault, and output the diagnosed fault and predicted fault trend, ensuring the stability of the diagnosis of the arc fault of the high-voltage switch cabinet.

[0007] To achieve the above purpose, the present application provides a high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors, which comprises:

[0008] The sensor data encoding fusion module is configured to perform multi-scale wavelet transform on time-series data collected by an arc sensor, a current sensor, and an ultrasonic sensor of the high-voltage switch cabinet to generate an encoding matrix.

[0009] The arc feature extraction module is configured to perform arc feature extraction on the encoding matrix by using an arc feature extraction backbone network based on a conditional parameter convolution and an attention mechanism architecture to generate an arc feature vector.

[0010] The fault cause classification module is configured to perform fault cause classification on the arc feature vector by using a fault cause classification head based on a Softmax function to generate an arc fault cause.

[0011] The fault trend prediction module is configured to perform fault trend prediction on the arc feature vector by using a fault trend regression head based on a linear activation function to generate an arc fault trend prediction value.

[0012] The fault diagnosis and early warning module is configured to perform arc fault diagnosis and management based on the arc fault cause and the arc fault trend prediction value.

[0013] Further, the arc feature extraction backbone network includes a conditional parameter convolution subnetwork, an attention subnetwork, and a feature aggregation layer, and the arc feature extraction module includes:

[0014] The multi-stage conditional convolution calculation unit is configured to extract multi-resolution and semantic-level arc features by using a conditional parameter convolution subnetwork with multiple stages to generate a multi-semantic resolution fusion feature matrix.

[0015] The fault feature focusing unit is configured to focus on the correlation features of arc faults by using an attention subnetwork with channel attention and spatial attention to generate a focused arc fault feature matrix.

[0016] The global feature aggregation unit is configured to generate the arc feature vector by using a pooling operation of a feature aggregation layer based on multiple channels of the focused arc fault feature matrix.

[0017] Further, the conditional parameter convolution subnetwork includes a first-stage convolution block, a second-stage convolution block, a third-stage convolution block, a fourth-stage convolution block, and a feature fusion layer, and the multi-stage conditional convolution calculation unit includes:

[0018] The high-resolution low-semantic feature extraction subunit is configured to generate a high-resolution low-semantic feature matrix by using a first-stage convolution block with a set number of basic-stage convolution blocks.

[0019] an intermediate semantic extraction subunit configured to generate an intermediate feature matrix by passing the high-resolution low-semantic feature matrix through a second stage convolutional block with a second number of basic stage convolutional blocks, and generate a low-resolution high-semantic feature matrix by passing the intermediate feature matrix through a third stage convolutional block with a third number of multi-layer parameter convolutional blocks;

[0020] a high-level semantic extraction subunit configured to generate a high-level semantic feature matrix by passing the low-resolution high-semantic feature matrix through a fourth stage convolutional block with a fourth number of basic stage convolutional blocks;

[0021] a multi-stage feature aggregation subunit configured to generate the multi-semantic resolution fusion feature matrix by adding the intermediate feature matrix and the low-resolution high-semantic feature matrix after being passed through an up-sampling layer, respectively, to the high-level semantic feature matrix.

[0022] the high-resolution low-semantic feature extraction subunit is configured to generate a high-resolution low-semantic feature matrix by passing the encoding matrix through an initial down-sampling and through the first stage convolutional block;

[0023] the intermediate semantic extraction subunit is configured to generate an intermediate feature matrix by passing the high-resolution low-semantic feature matrix through a first standard down-sampling and through the second stage convolutional block, and generate a low-resolution high-semantic feature matrix by passing the intermediate feature matrix through a second standard down-sampling and through the third stage convolutional block;

[0024] the high-level semantic extraction subunit is configured to generate a high-level semantic feature matrix by passing the low-resolution high-semantic feature matrix through a third standard down-sampling and through the fourth stage convolutional block;

[0025] wherein the sampling step lengths of the first, second, and third standard down-samplings are equal and smaller than the sampling step length of the initial down-sampling.

[0026] Further, the fault feature focusing unit comprises:

[0027] a channel attention subunit configured to generate a channel attention weighting map by passing the multi-semantic resolution fusion feature matrix through channel attention with global average pooling and global maximum pooling;

[0028] a spatial attention subunit configured to generate the focused electric arc fault feature matrix by passing the channel attention weighting map through spatial attention with channel dimension average pooling and channel dimension maximum pooling.

[0029] Further, the sensor data encoding fusion module comprises:

[0030] an adaptive wavelet transform unit, configured to respectively perform multi-scale wavelet transform on the time-series collected data of the arc light sensor, the current sensor and the ultrasonic sensor of the high-voltage switch cabinet to generate arc light wavelet coefficients, current wavelet coefficients and ultrasonic wavelet coefficients;

[0031] a mapping transform unit, configured to respectively map the arc light wavelet coefficients, the current wavelet coefficients and the ultrasonic wavelet coefficients to a first channel, a second channel and a third channel to generate the encoding matrix.

[0032] Further, the adaptive wavelet transform unit comprises:

[0033] an ultrasonic wavelet basis function calculation subunit, configured to calculate a Gaussian window function weighted ultrasonic wavelet basis for a resonance frequency of the time-series collected data of the ultrasonic sensor;

[0034] an ultrasonic conversion subunit, configured to perform multi-scale wavelet transform on the time-series collected data of the ultrasonic sensor by setting the ultrasonic wavelet basis to generate the ultrasonic wavelet coefficients.

[0035] Further, the fault cause classification module is configured to sequentially pass the arc feature vector through a classification mapping layer of the fault cause classification head and a Softmax function to generate the arc fault cause.

[0036] Further, the fault trend prediction module is configured to sequentially pass the arc feature vector through a regression mapping layer of the fault trend regression head, a Sigmoid function and a scalar degree to generate the arc fault trend prediction value.

[0037] Further, the high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors further comprises:

[0038] a training optimization module, configured to construct a loss function based on a cross-entropy loss term and a Huber loss term, and perform collaborative training optimization on the arc feature extraction backbone network, the fault cause classification head and the fault trend regression head based on the loss function.

[0039] Compared with the prior art, the present application has the beneficial effects that the present application adopts multi-scale wavelet transform to effectively extract key details and transient features in time-series data of different sensors, and the arc feature extraction backbone network based on conditional parameter convolution and attention mechanism can adaptively focus on the most relevant signal features of arc fault, and output the diagnosed fault and predicted fault trend, thereby ensuring the stability of the diagnosis of high-voltage switch cabinet arc fault.

[0040] Especially, the present application ensures the richness of features through a multi-resolution feature extraction architecture, shallow stages retain the details of the high-frequency signal of the initial arc of the arc, and deep stages extract highly abstract semantic features of insulation breakdown and poor contact. The introduction of conditional parameter convolution enables the network to dynamically adapt to the characteristics of different sensor signals and different fault types, thereby extracting more discriminative features.

[0041] Especially, the present application calculates the sensor resonance frequency and constructs a customized wavelet basis weighted by a Gaussian window function, so that the generated wavelet basis function is optimally matched with the input ultrasonic signal in shape and frequency center, making the ultrasonic wave and the arc light, current coefficient more matched in information density and quality, making the fusion coding matrix of the three more consistent, facilitating efficient learning of the neural network. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A structure schematic diagram of the high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors according to an embodiment of the present application;

[0043] Figure 2 A flowchart of the high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors according to an embodiment of the present application;

[0044] Figure 3 A flowchart of the arc feature extraction backbone network of the high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors according to an embodiment of the present application;

[0045] Figure 4 A flowchart of the attention subnetwork of the high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors according to an embodiment of the present application. DETAILED DESCRIPTION

[0046] In order to make the objects and advantages of the present application clearer, the present application will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0047] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application and are not intended to limit the protection scope of the present application.

[0048] It should be noted that in the description of the present application, the terms "upper", "lower", "left", "right", "inner", "outer" and the like indicate the direction or positional relationship terms based on the direction or positional relationship shown in the drawings, which are only for the convenience of description and do not indicate or imply that the device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0049] In addition, it also needs to be explained that, in the description of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0050] As shown in Figures 1 to 4 The present application provides a high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors, which adopts multi-scale wavelet transform to effectively extract key details and transient characteristics in different sensor time series data, and a arc feature extraction backbone network based on conditional parameter convolution and attention mechanism can adaptively focus on the most relevant signal features of arc fault, and output the diagnosed fault and predicted fault trend, ensuring the stability of high-voltage switch cabinet arc fault diagnosis.

[0051] As shown in Figure 1 and 2 The present embodiment provides a high-voltage switch cabinet arc fault diagnosis system based on intelligent sensors, which includes:

[0052] The sensor data encoding fusion module is used to generate an encoding matrix by multi-scale wavelet transform on the time series data collected by the arc light sensor, current sensor and ultrasonic sensor of the high-voltage switch cabinet;

[0053] The arc feature extraction module is used to generate an arc feature vector by an arc feature extraction backbone network based on conditional parameter convolution and attention mechanism architecture on the encoding matrix;

[0054] The fault cause classification module is used to generate an arc fault cause by a fault cause classification head based on Softmax function on the arc feature vector;

[0055] The fault trend prediction module is used to generate an arc fault trend prediction value by a fault trend regression head based on linear activation function on the arc feature vector;

[0056] The fault diagnosis and early warning module is used to perform arc fault diagnosis and management based on the arc fault cause and the arc fault trend prediction value.

[0057] Especially, the output arc fault cause includes no fault corresponding to a prediction probability of 0.02, poor contact connection corresponding to a prediction probability of 0.85, insulation aging breakdown corresponding to a prediction probability of 0.1, and foreign matter invasion corresponding to a prediction probability of 0.03. The output arc fault trend prediction value includes 0.4 representing an early fault sign, 0.7 representing a medium energy arc, and 0.9 representing a high energy arc. The fault diagnosis early warning module analyzes that there is a 85% possibility that the fault is caused by poor contact, and the severity index of the current fault is 0.9, which belongs to a high-risk state, and thus the early warning suggests immediate treatment.

[0058] As shown in Figure 3 Further, the arc feature extraction backbone network includes a conditional parameter convolution subnetwork, an attention subnetwork, and a feature aggregation layer, and the arc feature extraction module includes:

[0059] A multi-stage conditional convolution calculation unit is configured to extract multi-resolution and semantic level arc features by passing the encoding matrix through the conditional parameter convolution subnetwork with multiple stages to generate a multi-semantic resolution fusion feature matrix;

[0060] A fault feature focusing unit is configured to focus on the associated features of the arc fault by passing the multi-semantic resolution fusion feature matrix through the attention subnetwork with channel attention and spatial attention to generate a focused arc fault feature matrix;

[0061] A global feature aggregation unit is configured to generate the arc feature vector based on multiple channels of the focused arc fault feature matrix through the pooling operation of the feature aggregation layer.

[0062] Further, the conditional parameter convolution subnetwork includes a first stage convolution block, a second stage convolution block, a third stage convolution block, a fourth stage convolution block, and a feature fusion layer, and the multi-stage conditional convolution calculation unit includes:

[0063] A high-resolution low-semantic feature extraction subunit is configured to generate a high-resolution low-semantic feature matrix by passing the encoding matrix through the first stage convolution block with a set number of basic stage convolution blocks;

[0064] An intermediate semantic extraction subunit is configured to generate an intermediate feature matrix by passing the high-resolution low-semantic feature matrix through the second stage convolution block with a set number of basic stage convolution blocks, and generate a low-resolution high-semantic feature matrix by passing the intermediate feature matrix through the third stage convolution block with a set number of multi-layer parameter convolution blocks;

[0065] A high-level semantic extraction subunit is configured to generate a high-level semantic feature matrix by passing the low-resolution high-semantic feature matrix through the fourth stage convolution block with a set number of basic stage convolution blocks;

[0066] A multi-stage feature aggregation subunit is configured to add the intermediate feature matrix and the low-resolution high semantic feature matrix, after being respectively passed through an up-sampling layer, to the high-level semantic feature matrix to generate the multi-semantic resolution fusion feature matrix.

[0067] In particular, the multi-resolution feature extraction architecture ensures the richness of the features. The shallow stage retains the detailed features of the high-resolution micro high-frequency signals of the initial arcing of the arc, and the deep stage extracts highly abstracted semantic features of insulation breakdown and poor contact. The introduction of the conditional parameter convolution enables the network to dynamically adapt to the characteristics of different sensor signals and different fault types, thereby extracting more discriminative features.

[0068] Further, the high-resolution low semantic feature extraction subunit is configured to pass the encoding matrix through an initial down-sampling, and then through the first stage convolution block to generate a high-resolution low semantic feature matrix.

[0069] The intermediate semantic extraction subunit is configured to pass the high-resolution low semantic feature matrix through a first standard down-sampling, and then through the second stage convolution block to generate an intermediate feature matrix, and pass the intermediate feature matrix through a second standard down-sampling, and then through the third stage convolution block to generate a low-resolution high semantic feature matrix.

[0070] The high-level semantic extraction subunit is configured to pass the low-resolution high semantic feature matrix through a third standard down-sampling, and then through the fourth stage convolution block to generate a high-level semantic feature matrix.

[0071] The sampling step lengths of the first, second, and third standard down-samplings are equal and smaller than the sampling step length of the initial down-sampling.

[0072] In particular, the progressive design of high-resolution detail retention, multi-semantic level extraction, and cross-scale feature aggregation solves the problem that the traditional convolutional network cannot simultaneously achieve high resolution and high semantic, and weak fault features are easily masked by semantic enhancement.

[0073] Specifically, the multi-stage conditional convolution calculation unit can be represented as:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] In the formula, I represents the encoding matrix. These represent the initial sampling features of the initial downsampling, the first sampling features of the first standard downsampling, the second sampling features of the second standard downsampling, and the third sampling features of the third standard downsampling, respectively. This indicates an initial downsampling with a 4x4 kernel and a sampling stride of 4. These represent the first, second, and third standard downsampling with a convolution kernel of 2 and a sampling stride of 2, respectively. , This indicates the number of convolutional blocks in the base stage. The first-stage convolutional block, the second-stage convolutional block, and the fourth-stage convolutional block. The preferred value is 3, which means repeating the conditional parameterized convolution three times. Indicates the number of multi-layer parameters and convolutional blocks. The third stage of convolutional block, The optimal value is 9, which means repeating the conditional parameterized convolution nine times. This represents a single-stage Conditional Parameterized Convolution (CPConv). , , , These represent the high-resolution low-semantic feature matrix, the intermediate feature matrix, the low-resolution high-semantic feature matrix, and the high-level semantic feature matrix, respectively. This indicates an upsampling layer, preferably using a 1x1 convolutional block. This indicates element-wise addition. This represents the multi-semantic resolution fused feature matrix.

[0084] Preferably, a first-order conditionally parameterized convolution It can be represented as:

[0085]

[0086]

[0087] In the formula, Represents the weight vector, by , , composition, denotes activation function, denotes a full connection layer, denotes a global average pooling, denotes input data, , , denote a depthwise separable convolution with convolution kernel, convolution kernel, convolution kernel, denotes a Hadamard product, denotes output data.

[0088] In particular, the first-stage convolution block rapidly reduces the resolution by a large step size, increases the number of channels, realizes fast down-sampling, extracts low-level, high-resolution arc detail features, the second-stage convolution block further down-samples, fuses more extensive context information, forms middle-level features, the third-stage convolution block continues to down-sample, the receptive field is further increased, and more semantic features are extracted, and the fourth-stage convolution block generates the most abstract, low-resolution feature map with the richest semantic information. The output features of the second-stage convolution block, the third-stage convolution block and the fourth-stage convolution block are fused by up-sampling and addition to realize a comprehensive output feature that is a feature map containing both detail and semantic information and enhanced.

[0089] As shown in Figure 4 , further, the fault feature focusing unit comprises:

[0090] a channel attention sub-unit configured to generate a channel attention weighting map by channel attention with global average pooling and global maximum pooling on the multi-semantic resolution fusion feature matrix;

[0091] a spatial attention sub-unit configured to generate the focused arc fault feature matrix by spatial attention with channel dimension average pooling and channel dimension maximum pooling on the channel attention weighting map.

[0092] In particular, the channel attention accurately captures all series of fault features from subtle signs to violent outbreaks, greatly improving the ability to distinguish different fault modes. The spatial attention can automatically focus on the key features most relevant to the fault, effectively suppress the interference of background noise and irrelevant information, and greatly reduce false positives and false negatives.

[0093] Specifically, the process of generating the focused arc fault feature matrix by the attention sub-network with channel attention and spatial attention can be represented as:

[0094]

[0095]

[0096]

[0097]

[0098] wherein, denotes a channel attention map, denotes a Sigmoid activation function, denotes a multi-layer perceptron, denotes a global average pooling and a global max pooling, respectively, denotes a multi-semantics resolution fusion feature matrix, denotes a Hadamard product, denotes a channel attention weighting map, denotes a spatial attention map, denotes a 7x7 convolution operation, , denotes a channel dimension average pooling and a channel dimension max pooling, respectively, denotes a focused arc fault feature matrix.

[0099] Specifically, the feature aggregation layer is a generalized average pooling (GeM Pooling), and the process of generating the arc feature vector by the global feature aggregation unit can be represented as:

[0100]

[0101]

[0102] wherein, denotes a k-th channel pooled feature, denotes a k-th channel input feature set, denotes a number of elements of the k-th channel input feature set, denotes a k-th channel input feature set, denotes a k-th channel input feature set, denotes a k-th channel learnable parameter, denotes an arc feature vector, and Dropout denotes a dropout layer, denotes an activation function, denotes layer normalization, denotes a learnable weight matrix and a bias term.

[0103] Further, the fault cause classification module is configured to sequentially pass the arc feature vector through a classification mapping layer of the fault cause classification head and a Softmax function to generate the arc fault cause, specifically:

[0104]

[0105] wherein, represents the arc fault reason, represents the arc feature vector, , respectively represent the weight matrix and the bias term of the classification mapping layer.

[0106] Further, the fault trend prediction module is used to sequentially pass the arc feature vector through the regression mapping layer of the fault trend regression head, the Sigmoid function and the scalar degree, to generate the arc fault trend prediction value, specifically:

[0107]

[0108] wherein, represents the arc fault trend prediction value, respectively represent the learnable weight matrix and the bias term, and scale represents the scalar degree, represents the arc feature vector.

[0109] Especially, the complex fault state is converted into a range-fixed, meaning clear quantitative index, so that the operation and maintenance personnel can quickly and intuitively grasp the severity of the fault.

[0110] Further, the high-voltage switch cabinet arc fault diagnosis system further comprises:

[0111] The training optimization module is used to construct a loss function based on a cross-entropy loss term and a Huber loss term, and to perform collaborative training optimization on the arc feature extraction backbone network, the fault reason classification head and the fault trend regression head based on the loss function.

[0112] Specifically, the loss function can be represented as:

[0113]

[0114]

[0115]

[0116] wherein, , respectively represent the cross-entropy loss term and the Huber loss term, represents a label smoothing coefficient, preferably 0.1, represents the fault reason category, represents an indicator function, represents the predicted arc fault reason, represents the true arc fault value of the sample, represents an arc fault trend prediction value, represents a threshold parameter of Huber loss, preferably 0.5, is a learnable hyperparameter.

[0117] Further, the sensor data encoding fusion module comprises:

[0118] an adaptive wavelet transform unit, configured to respectively perform multi-scale wavelet transform on the time-series collected data of the arc sensor, the current sensor and the ultrasonic sensor of the high-voltage switch cabinet, to generate arc light wavelet coefficients, current wavelet coefficients and ultrasonic wavelet coefficients;

[0119] a mapping transform unit, configured to respectively map the arc light wavelet coefficients, the current wavelet coefficients and the ultrasonic wavelet coefficients to a first channel, a second channel and a third channel, to generate the encoding matrix.

[0120] Further, the adaptive wavelet transform unit comprises:

[0121] an ultrasonic wavelet basis function calculation sub-unit, configured to calculate a Gaussian window function weighted ultrasonic wavelet basis for a resonance frequency of the time-series collected data of the ultrasonic sensor;

[0122] an ultrasonic conversion sub-unit, configured to perform multi-scale wavelet transform on the time-series collected data of the ultrasonic sensor by setting the ultrasonic wavelet basis, to generate the ultrasonic wavelet coefficients.

[0123] Specifically, the multi-scale wavelet transform can be represented as:

[0124]

[0125] wherein, represents a wavelet transform coefficient, represents a normalization factor, represents a shift amount, represents the time-series collected data of the arc sensor, the current sensor and the ultrasonic sensor, represents a wavelet basis.

[0126] wherein, the ultrasonic sensor corresponds to a∈[2, 10], and the ultrasonic wavelet basis is wherein, represents a resonance frequency, represents an ultrasonic signal.

[0127] wherein, the arc sensor corresponds to a∈[1, 5], and the wavelet basis is a Db4 wavelet support segment. The current sensor corresponds to a∈[2, 10], and ψ is a Db5 / Symlet wavelet.

[0128] In particular, the optimized ultrasonic features can effectively help the model to distinguish different fault types that exhibit similar current and arc light signals, for example, surface discharge of insulating materials can generate a wide frequency, burst-like acoustic wave, and mechanical component loose vibration can generate a relatively fixed frequency, persistent acoustic wave, so that the generated wavelet basis function is optimally matched with the input ultrasonic signal in shape and frequency center, making the ultrasonic wave more matched with the arc light and current coefficient in information density and quality, and making the fusion encoding matrix of the three more consistent, facilitating efficient learning of the neural network.

[0129] In this embodiment, multi-scale wavelet transform is used to effectively extract key details and transient features in different sensor time series data, and the arc feature extraction backbone network based on conditional parameter convolution and attention mechanism can adaptively focus on the signal features most relevant to arc faults and output the diagnosed faults and predicted fault trends, ensuring the stability of the diagnosis of high-voltage switch cabinet arc faults. Through the multi-resolution feature extraction architecture, the richness of the features is ensured, the details of the high-resolution arc initial arc micro-high-frequency signals are retained in the shallow stage, and the highly abstract semantic features of insulation breakdown and poor contact are extracted in the deep stage. The introduction of conditional parameter convolution enables the network to dynamically adapt to the characteristics of different sensor signals and different fault types, thereby extracting features with stronger discrimination. By calculating the sensor resonance frequency and constructing a customized wavelet basis weighted by a Gaussian window function, the generated wavelet basis function is optimally matched with the input ultrasonic signal in shape and frequency center, making the ultrasonic wave more matched with the arc light and current coefficient in information density and quality, making the fusion encoding matrix of the three more consistent, facilitating efficient learning of the neural network.

[0130] Those skilled in the art can appreciate that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0131] So far, the technical solutions of the present application have been described in conjunction with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the present application, and the technical solutions after such changes or replacements will fall within the protection scope of the present application.

[0132] The above merely illustrates the preferred embodiments of the present application, and is not used to limit the present application; for those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A high-voltage switchgear arc fault diagnosis system based on intelligent sensors, characterized in that, include: The sensor data encoding and fusion module is used to generate an encoding matrix by performing multi-scale wavelet transform on the time-series acquisition data of arc sensor, current sensor and ultrasonic sensor of high voltage switchgear. An arc feature extraction module is used to generate an arc feature vector by passing the encoding matrix through an arc feature extraction backbone network based on a conditional parameter convolution and attention mechanism architecture. The fault cause classification module is used to pass the arc feature vector through a fault cause classification head based on the Softmax function to generate the arc fault cause; The fault trend prediction module is used to pass the arc feature vector through a fault trend regression head based on a linear activation function to generate arc fault trend prediction values. The fault diagnosis and early warning module is used to perform arc fault diagnosis and management based on the cause of the arc fault and the arc fault trend prediction value. The arc feature extraction backbone network includes a conditional parameter convolutional subnetwork, an attention subnetwork, and a feature aggregation layer. The conditional parameter convolutional subnetwork includes a first-stage convolutional block, a second-stage convolutional block, a third-stage convolutional block, a fourth-stage convolutional block, and a feature fusion layer. The arc feature extraction module includes: A multi-stage conditional convolutional computation unit is used to pass the encoding matrix through a conditional parameter convolutional sub-network with multiple stages to extract arc features at multiple resolutions and semantic levels, so as to generate a multi-semantic resolution fusion feature matrix. The multi-stage conditional convolution computation unit includes: A high-resolution low-semantic feature extraction subunit is used to generate a high-resolution low-semantic feature matrix by passing the encoding matrix through a first-stage convolutional block that sets the number of base-stage convolutional blocks. The intermediate semantic extraction subunit is used to generate an intermediate feature matrix by passing the high-resolution low semantic feature matrix through a second-stage convolutional block with a set number of base-stage convolutional blocks, and then passing the intermediate feature matrix through a third-stage convolutional block with a set number of multi-layer parameter convolutional blocks to generate a low-resolution high semantic feature matrix. The high-level semantic extraction subunit is used to pass the low-resolution high-semantic feature matrix through a fourth-stage convolutional block, which is the same number of convolutional blocks as the base stage, to generate a high-level semantic feature matrix. A multi-stage feature aggregation subunit is used to pass the intermediate feature matrix and the low-resolution high semantic feature matrix through an upsampling layer, and then add them to the high-level semantic feature matrix to generate the multi-semantic resolution fusion feature matrix. The fault feature focusing unit is used to focus the associated features of the arc fault through the multi-semantic resolution fusion feature matrix and an attention sub-network with channel attention and spatial attention to generate a focused arc fault feature matrix. A global feature aggregation unit is used to generate the arc feature vector by pooling operations of the feature aggregation layer based on multiple channels of the focused arc fault feature matrix.

2. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 1, characterized in that, The high-resolution low-semantic feature extraction subunit is used to generate a high-resolution low-semantic feature matrix by passing the encoding matrix through the first-stage convolutional block after initial downsampling. The intermediate semantic extraction subunit is used to downsample the high-resolution low semantic feature matrix through a first standard and then pass it through a second-stage convolutional block to generate an intermediate feature matrix, and then downsample the intermediate feature matrix through a second standard and pass it through a third-stage convolutional block to generate a low-resolution high semantic feature matrix. The high-level semantic extraction subunit is used to downsample the low-resolution high-semantic feature matrix using a third standard, and then pass it through the fourth-stage convolutional block to generate a high-level semantic feature matrix. The sampling step sizes of the first standard downsampling, the second standard downsampling, and the third standard downsampling are all equal and smaller than the sampling step size of the initial downsampling.

3. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 1, characterized in that, The fault feature focusing unit includes: The channel attention subunit is used to pass the multi-semantic resolution fusion feature matrix through channel attention with global average pooling and global max pooling to generate a channel attention weighted map; A spatial attention subunit is used to generate the focused arc fault feature matrix by passing the channel attention weighted map through spatial attention with channel-dimensional average pooling and channel-dimensional max pooling.

4. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 1, characterized in that, The sensor data encoding and fusion module includes: An adaptive wavelet transform unit is used to process the time-series acquisition data of the arc light sensor, current sensor and ultrasonic sensor of the high-voltage switchgear through multi-scale wavelet transform to generate arc light wavelet coefficients, current wavelet coefficients and ultrasonic wavelet coefficients respectively. The mapping transformation unit is used to map the arc wavelet coefficients, current wavelet coefficients and ultrasonic wavelet coefficients to the first channel, the second channel and the third channel, respectively, to generate the encoding matrix.

5. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 4, characterized in that, The adaptive wavelet transform unit includes: The ultrasonic wavelet basis function calculation subunit is used to calculate the ultrasonic wavelet basis weighted by Gaussian window function from the resonant frequencies of the time-series acquisition data of the ultrasonic sensor. An ultrasonic transduction subunit is used to generate ultrasonic wavelet coefficients by performing a multi-scale wavelet transform on the ultrasonic wavelet basis using the time-series data acquired by the ultrasonic sensor.

6. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 1, characterized in that, The fault cause classification module is used to sequentially pass the arc feature vector through the classification mapping layer and the Softmax function of the fault cause classification head to generate the arc fault cause.

7. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to claim 1, characterized in that, The fault trend prediction module is used to sequentially pass the arc feature vector through the regression mapping layer, sigmoid function and scalar degree of the fault trend regression head to generate the arc fault trend prediction value.

8. The high-voltage switchgear arc fault diagnosis system based on intelligent sensors according to any one of claims 1 to 7, characterized in that, Also includes: The training optimization module is used to construct a loss function based on cross-entropy loss and Huber loss, and to perform collaborative training and optimization on the arc feature extraction backbone network, fault cause classification head and fault trend regression head based on the loss function.

Images