Fault detection method for direct-current power transmission system

Abstract

The invention discloses a DC power transmission system fault detection method, and belongs to the technical field of DC power transmission system fault detection, and the DC power transmission system fault detection method comprises the following steps: collecting DC power transmission system fault information in real time; preprocessing the original data to realize time-space association matching of the multi-source monitoring data; an intelligent diagnosis model is constructed, key features of fault data are rapidly extracted by adopting multiple algorithms, and fault types and confidence coefficients are output; the GPS position information of the robot and the UWB indoor positioning label in the inspection area are combined to realize meter-scale precision positioning of a fault point; constructing a direct-current power transmission system fault database, and continuously optimizing the diagnosis model by adopting transfer learning and incremental training technologies; the method is relatively comprehensive in fault detection of the direct-current power transmission system, is relatively high in fault diagnosis capability, can continuously learn, and realizes autonomous judgment of fault types and positioning of fault points.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault detection of direct current transmission systems, in particular to a fault detection method of a direct current transmission system. BACKGROUND

[0002] With the expansion of power equipment scale and the complication of operation environment, the limitations of manual inspection are becoming more and more prominent, especially in extreme environments, safety and efficiency become important challenges, and power intelligent inspection robots emerge as the times require, which can autonomously navigate in complex substation environments, perform efficient and accurate equipment inspection tasks, cover key equipment such as transformers and switch cabinets, and ensure the stable operation of the power system.

[0003] For example, the patent with publication number CN112003183B discloses a power transmission network fault processing system and method. It includes a substation, a power transmission line and a tower, multiple towers are arranged between adjacent two substations, the power transmission line is connected between the substation and the tower and between adjacent two towers, a patrol robot for detecting fault position is arranged on the power transmission line between adjacent two substations, the patrol robot is slidingly fitted on the power transmission line, a background controller, a display module, a wireless communication module and a storage module are arranged in the substation, the display module, the storage module and the wireless communication module are signal connected with the background controller, and the wireless communication module is signal connected with the patrol robot. The present application combines the background controller, the patrol robot and the insulator contamination state detection system, so that timely response can be realized when the power transmission network line fails.

[0004] For example, the patent with publication number CN205679725U discloses a mobile detection device for fault diagnosis of damage and broken strands of power transmission lines, which comprises a magnetic flux leakage detection module, a direct current power supply and a line patrol robot. The magnetic flux leakage detection module comprises an electromagnet, a coil, a Hall sensor, an isolation layer, a data acquisition card and an embedded system. The electromagnet forms a stable electromagnetic field after magnetization. The coil is wrapped around the surface of the electromagnet, and the Hall sensor receives the magnetic flux leakage field signal around the power transmission conductor. The isolation layer is installed between the electromagnet, the Hall sensor and the power transmission conductor. The data acquisition card processes and digitizes the magnetic flux leakage field signal output by the Hall sensor. The embedded system compares the collected magnetic flux leakage field signals. The line patrol robot moves on the power transmission conductor through the moving wheels and grippers. The direct current power supply provides power for each component. The utility model has the advantages of simple structure, convenient use and reliability, can quickly detect the damage and broken strand faults of the power transmission line, is not limited by the environment such as terrain, and is suitable for long-distance and complex environment power transmission line fault detection.

[0005] In the multi-scene deployment of power intelligent inspection robots, most of the existing inspection equipment can only meet single detection needs, and the diagnosis capability for direct current transmission systems is limited, making it difficult to quickly analyze and locate fault points.

[0006] In view of the above problems, it is urgent to make innovative design on the basis of the existing fault detection method of the direct current transmission system. SUMMARY

[0007] The purpose of the present application is to provide a fault detection method of a direct current transmission system to solve the problem that the existing inspection equipment can only meet the single detection requirement and has limited diagnosis capability for the direct current transmission system, which makes it difficult to quickly analyze and locate the fault point in the multi-scene deployment of the power intelligent inspection robot proposed in the background art.

[0008] To achieve the above purpose, the present application provides the following technical solution: a fault detection method of a direct current transmission system, the fault detection method of the direct current transmission system comprising the following steps:

[0009] A camera, a direct current sensor, an infrared thermal imaging module, an ultraviolet imager and a partial discharge sensor are added to the key detection nodes of the power inspection robot to construct a multi-dimensional fault perception network, capture characteristic signals such as current mutation, local overheating, corona discharge and insulation defects of the direct current transmission system in the fault state, and collect fault information of the direct current transmission system in real time.

[0010] Based on the edge computing module carried by the robot, the original data collected by each sensor are preprocessed, 5G and optical fiber double link transmission are adopted, and the time stamp marking and position label coding technology of data are combined to realize the spatio-temporal correlation matching of multi-source monitoring data and provide complete context for fault analysis.

[0011] An intelligent diagnosis model is constructed in combination with the multi-source data collected by the robot, deep learning algorithm, convolutional neural network, image recognition algorithm, weighted fusion algorithm, random forest algorithm and LSTM algorithm are adopted for complex faults to quickly extract key features of fault data, realize preliminary classification of multiple fault types, and take image data, current, temperature, ultraviolet and PD signals collected by the robot as input and output fault type and confidence.

[0012] In combination with the GPS position information of the robot and the UWB indoor positioning label in the inspection area, a space coordinate mapping model is established, a time sequence-space correlation matrix is constructed based on the time sequence of the occurrence of fault features, and meter-level precision positioning of the fault point is realized.

[0013] A fault database of the direct current transmission system is constructed, monitoring data, fault types, treatment schemes and other information in historical fault cases are integrated, transfer learning and incremental training technology are adopted to continuously optimize the diagnosis model, and end-to-end automatic fault detection is achieved.

[0014] Preferably, the direct current sensor is used to collect the direct current of the line and equipment in real time, capture the current mutation characteristics at the time of failure; the infrared thermal imaging module is used to identify the heat of the power line and equipment, and the failure of the direct current equipment is often accompanied by local overheating, which requires to improve the resolution and temperature measurement accuracy of thermal imaging to identify temperature difference anomalies as low as 0.5℃; the ultraviolet imager is used to capture the ultraviolet spectrum of 300-400nm, quantify the number of discharge photons, and distinguish normal corona from fault discharge; the partial discharge sensor is aimed at the converter valve, DC bushing and other key equipment of the converter station, and detects the partial discharge signal through an ultrahigh frequency or very high frequency sensor to locate the insulation defect.

[0015] Preferably, the pre-processing of the edge computing module can specifically realize data dimension reduction, eliminate invalid data, compress images, and reduce the transmission data volume by more than 70%; real-time screening, early identification of obvious abnormal data, priority transmission of key data, and guarantee of fault response timeliness.

[0016] Preferably, the deep learning algorithm is based on the infrastructure of a convolutional neural network, and three types of core task algorithms are derived for different fault detection requirements of the direct current transmission system, specifically including:

[0017] Image classification algorithm, the fault features of the direct current transmission system are hidden in the deep network, and the residual network solves the gradient disappearance problem of the deep convolutional neural network through residual blocks, as follows:

[0018]

[0019] wherein x is the residual block, is the feature after passing through 2 convolutional layers and ReLU activation, if the input and output channel numbers or sizes are inconsistent, the dimension of x needs to be adjusted through 1x1 convolution, that is:

[0020]

[0021] wherein, is the convolution kernel for adjusting the dimension.

[0022] Target detection algorithm, judges whether there is a fault in the image, and outputs the bounding box of the fault, adapts to the multi-component panoramic image collected by the inspection robot, sets the top-left corner coordinates of the grid as , the offset of the bounding box predicted by the model is , the width-height ratio is , and the absolute coordinates and size of the bounding box are:

[0023]

[0024]

[0025]

[0026]

[0027] wherein, is a Sigmoid function, , is the width and height of the predefined anchor box.

[0028] The semantic segmentation algorithm classifies each pixel in the image, accurately segments the contour of the fault, and is suitable for the scene where the fault area is fine and irregular in the DC power transmission system. It is composed of a transpose convolution layer, a convolution layer and a ReLU, which gradually restores the feature map size. Let the input feature map , the transpose convolution kernel , the output feature map The size is:

[0029]

[0030]

[0031] wherein, S is the transpose convolution step length, and P is the padding number.

[0032] Preferably, the weighted fusion algorithm fuses multi-source data, and the data collected by sensors A, B and C is respectively , , The corresponding weights are respectively , , The calculation method of the fused data X is:

[0033]

[0034]

[0035] For data with high reliability and strong association with faults, high numerical weight is given, and vice versa.

[0036] Preferably, the random forest algorithm deeply analyzes the complex data set and realizes the preliminary classification of multiple fault types. The classification task specifically includes:

[0037] There are T decision trees in the random forest. For a sample x, the tth tree predicts its label as The class set is K, and the probability formula of sample x being classified into class k is:

[0038]

[0039] wherein, is an indicator function, when =k, =1, otherwise 0.

[0040] Preferably, the LSTM algorithm enables accurate diagnosis and prediction of power faults. Fault data in the power system exhibits obvious time-series characteristics. By utilizing these time-series characteristics and learning from historical fault data, the algorithm determines and predicts the type and development trend of faults. The algorithm formula is as follows:

[0041]

[0042] in, The output of the forget gate at time t. For the forget gate, input at the current time The weight matrix for linear transformation. The state hidden in the previous moment is for forgetting the gate. The weight matrix for linear transformation. For the bias term of the forget gate;

[0043]

[0044]

[0045] in, The input gate output at time t determines the current new information. How much is input into the cell state? For activation function, In the input gate, the input at the current time is... The weight matrix for linear transformation. Let be the candidate cell state at time t, which is the newly generated candidate information based on the current input. It is the hyperbolic tangent activation function;

[0046]

[0047] in, Let t represent the cell state at time t. The cell state at time t-1;

[0048]

[0049]

[0050] in, The output of the gate at time t determines the cell state. How many of them are output as hidden states? , In the output gate, the input at the current time... The weight matrix for linear transformation. for the output gate, the hidden state of the previous time is linearly transformed by a weight matrix, is a bias term for the output gate.

[0051] Preferably, the fault types include short-circuit fault, overload fault, insulation fault, and specifically:

[0052] The short-circuit fault is subdivided into three-phase short-circuit, two-phase short-circuit, single-phase ground short-circuit, and two-phase ground short-circuit.

[0053] The overload fault is usually caused by the long-time operation of the power equipment in a state exceeding its rated load. When the load exceeds the rated capacity of the equipment, the current increases, and the temperature of the equipment rises.

[0054] The insulation fault is manifested as a decrease in the insulation performance of the power equipment. When the insulation performance decreases to a certain extent, partial discharge occurs, producing corona and ozone, which accelerates the damage to the insulation.

[0055] Preferably, the spatial coordinate mapping model integrates data from different positioning technologies into a unified coordinate system, achieving a comprehensive and accurate description of the location of the power fault point. The method of coordinate transformation matrix is adopted, and the coordinates obtained by GPS measurement are After coordinate conversion and fusion with UWB data, the coordinates in the unified coordinate system are The conversion relationship is represented as:

[0056]

[0057] wherein, is a rotation and translation parameter, , , is a translation parameter.

[0058] Preferably, the step of constructing the time-space correlation matrix specifically includes:

[0059] Timestamping the fault information collected by the power inspection robot to record the accurate time of occurrence of each fault feature ; obtaining the spatial coordinates of the fault point using the spatial coordinate mapping model ; quantifying the time and space information, dividing the time into m time intervals , , …… , dividing the space into n spatial regions , , …… , the matrix element represents the number of times or the probability-related measure that the jth spatial region exhibits the failure feature in the ith time interval; the position of the failure point is determined by analyzing the distribution of the elements in the matrix.

[0060] Compared with the prior art, the present application has the beneficial effects that:

[0061] The deep learning algorithm can realize the whole-process processing of recognition, positioning and segmentation of the image failure of the DC power transmission system through the feature extraction capability of the convolutional neural network and the combination of the three major task models of classification, detection and segmentation.

[0062] The key parameters such as current, temperature, ultraviolet intensity and partial discharge amplitude are taken as model inputs, and after algorithm operation, the model outputs the specific failure type, providing clear failure judgment basis for operation and maintenance personnel. The deep learning algorithm is relatively comprehensive for failure detection of the DC power transmission system, and has strong failure diagnosis capability.

[0063] By fusing GPS position information, indoor positioning label information and failure feature time series analysis, the positioning accuracy of the failure point is controlled within 1-5 meters, realizing meter-level positioning of the failure point, and greatly reducing the failure searching time and workload of the operation and maintenance personnel.

[0064] In actual application process, the model can receive new failure data in real time, autonomously learn new failure features, continuously optimize model structure and parameters, and gradually realize end-to-end failure detection, i.e. full-process automation from multi-source data acquisition, preprocessing, analysis to failure type identification and failure point positioning, without manual intervention, greatly improving failure detection efficiency and accuracy, and adapting to the changing operation environment and failure mode of the DC power transmission system. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 The DC power transmission system failure detection flowchart of the present application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0067] EMBODIMENT

[0068] Please refer to Figure 1 The present application provides the following technical solutions: a DC power transmission system failure detection method, the DC power transmission system failure detection method comprising the following steps:

[0069] S1, add cameras, DC current sensors, infrared thermal imaging modules, ultraviolet imagers, and partial discharge sensors to the key detection nodes of the power inspection robot, construct a multi-dimensional fault perception network, capture characteristic signals such as current mutation, local overheating, corona discharge, and insulation defects of the DC power transmission system in a fault state, and collect fault information of the DC power transmission system in real time.

[0070] In this embodiment, the DC current sensor is used to collect the DC current of the line and equipment in real time and capture the current mutation characteristics during a fault. The infrared thermal imaging module is used to identify the heat of the power line and equipment. Faults of DC equipment are often accompanied by local overheating, so it is necessary to improve the resolution and temperature measurement accuracy of thermal imaging to identify temperature difference anomalies as low as 0.5°C. The ultraviolet imager is used to capture the ultraviolet spectrum of 300-400 nm, quantify the number of discharge photons, and distinguish between normal corona and fault discharge. The partial discharge sensor is aimed at key equipment such as converter valves and DC bushings in the converter station. It detects partial discharge signals through ultra-high frequency or very high frequency sensors and locates insulation defects.

[0071] S2, based on the edge computing module carried by the robot, pre-process the raw data collected by each sensor, use 5G and optical fiber dual-link transmission, combine timestamp marking and location label encoding technology, realize spatio-temporal correlation matching of multi-source monitoring data, and provide complete context for fault analysis.

[0072] In this embodiment, the pre-processing of the edge computing module can achieve: data dimensionality reduction, elimination of invalid data, compression of images, and reduction of transmission data volume by more than 70%; real-time screening, early identification of obvious abnormal data, priority transmission of key data, and guarantee of fault response timeliness.

[0073] S3, combine the multi-source data collected by the robot to construct an intelligent diagnosis model. For complex faults, use deep learning algorithms, convolutional neural networks, image recognition algorithms, weighted fusion algorithms, random forest algorithms, and LSTM algorithms to quickly extract key features of fault data, realize preliminary classification of multiple fault types, and use image data, current, temperature, ultraviolet, and PD signals collected by the robot as input to output fault types and confidence.

[0074] In this embodiment, the deep learning algorithm is based on the basic architecture of the convolutional neural network. For different fault detection needs of the DC power transmission system image, three types of core task algorithms are derived, including:

[0075] Image classification algorithm, the fault features of the DC power transmission system are hidden in the deep network, and the residual network solves the gradient disappearance problem of the deep convolutional neural network through residual blocks, as follows:

[0076]

[0077] wherein x is a residual block, is a feature after 2 convolutional layers and ReLU activation, if the input and output channel numbers or sizes are inconsistent, the dimension of x needs to be adjusted by 1x1 convolution, that is:

[0078]

[0079] wherein, is the convolution kernel for adjusting the dimension.

[0080] The target detection algorithm judges whether there is a fault in the image, and outputs the bounding box of the fault, which is suitable for the multi-component panoramic image collected by the inspection robot, and the left upper corner coordinates of the grid are , the bounding box offset predicted by the model is , the width-height ratio is , and the absolute coordinates and size of the bounding box are:

[0081]

[0082]

[0083]

[0084]

[0085] wherein, is a Sigmoid function, , is the width and height of the predefined anchor box.

[0086] The semantic segmentation algorithm classifies each pixel in the image and accurately segments the outline of the fault, which is suitable for the scene where the fault area is fine and irregular in the DC power transmission system, which is composed of a transpose convolutional layer, a convolutional layer and a ReLU, gradually recovering the feature map size, wherein the input feature map is , the transpose convolution kernel is , and the output feature map is The size of the output feature map is:

[0087]

[0088]

[0089] wherein S is the transpose convolution step, and P is the padding number.

[0090] In this embodiment, the weighted fusion algorithm fuses multi-source data, and the data collected by sensors A, B and C is respectively , , , and the corresponding weights are respectively , , The calculation method of the fused data X is as follows:

[0091]

[0092]

[0093] For data with high reliability and strong fault correlation, a high numerical weight is given, and vice versa.

[0094] In this embodiment, the random forest algorithm deeply analyzes the complex data set, realizes the preliminary classification of multiple fault types, and the classification task specifically includes:

[0095] There are T decision trees in the random forest, and for a sample x, the prediction label of the tth tree is The category set is K, and the probability formula of classifying the sample x into category k is:

[0096]

[0097] Wherein, is an indicator function, when =k, =1, otherwise 0.

[0098] In this embodiment, the LSTM algorithm realizes accurate diagnosis and prediction of power failure. The fault data in the power system presents obvious time sequence characteristics. By learning the historical fault data, the type and development trend of the fault are judged and predicted, and the algorithm formula is:

[0099]

[0100] Wherein, is the output of the forgetting gate at time t, is the weight matrix of the linear transformation of the input of the current time in the forgetting gate to the hidden state of the previous time, is the weight matrix of the linear transformation of the previous time hidden state to the hidden state of the current time in the forgetting gate, is the bias term of the forgetting gate;

[0101]

[0102]

[0103] Wherein, is the input gate output at time t, which decides how much new information is input to the cell state, is an activation function, is the input gate, which is the linear transformation of the input The weight matrix for linear transformation. Let be the candidate cell state at time t, which is the newly generated candidate information based on the current input. It is the hyperbolic tangent activation function;

[0104]

[0105] in, Let t represent the cell state at time t. The cell state at time t-1;

[0106]

[0107]

[0108] in, The output of the gate at time t determines the cell state. How many of them are output as hidden states? , In the output gate, the input at the current time... The weight matrix for linear transformation. In the output gate, the hidden state of the previous moment is... The weight matrix for linear transformation. This is the bias term for the output gate.

[0109] S4. By combining the robot's GPS location information with the UWB indoor positioning tags in the inspection area, a spatial coordinate mapping model is established. Based on the chronological order of the occurrence of fault characteristics, a time-space correlation matrix is ​​constructed to achieve meter-level accuracy in locating fault points.

[0110] It should be noted that the timestamp and location tag encoding technology can specifically achieve: unified identification, with each piece of data carrying both time and location keys, and the algorithm can automatically match data with the same key; historical tracing, by combining the topology information of power lines, it is possible to associate upstream and downstream data of the same line and analyze the fault propagation path.

[0111] S5. Construct a fault database for DC transmission systems, integrate monitoring data, fault types, and handling solutions from historical fault cases, and use transfer learning and incremental training techniques to continuously optimize the diagnostic model to achieve end-to-end fully automated fault detection.

[0112] In this embodiment, the fault types include short-circuit faults, overload faults, and insulation faults, specifically:

[0113] Short circuit faults are classified into three-phase short circuits, two-phase short circuits, single-phase-to-ground short circuits, and two-phase-to-ground short circuits.

[0114] Overload failure is usually caused by long-term operation of power equipment in a state exceeding its rated load. When the load exceeds the rated capacity of the equipment, the current increases, and the temperature of the equipment rises;

[0115] Insulation failure is manifested as a decrease in the insulation performance of power equipment. When the insulation performance decreases to a certain extent, partial discharge occurs, producing corona and ozone, which accelerates the damage to the insulation.

[0116] In this embodiment, the spatial coordinate mapping model integrates data from different positioning technologies into a unified coordinate system, achieving comprehensive and accurate description of the location of the power failure point. The method of coordinate transformation matrix is adopted, and the coordinates obtained by GPS measurement are After coordinate conversion and fusion with UWB data, the coordinates in the unified coordinate system are , and the conversion relationship is represented as:

[0117]

[0118] wherein, is the rotation and translation parameter, , , is the translation parameter.

[0119] In this embodiment, the step of constructing the time-space correlation matrix specifically includes:

[0120] The time stamp of the failure information collected by the power inspection robot is marked, and the accurate time of occurrence of each failure feature is recorded ;

[0121] The spatial coordinates of the failure point obtained by the spatial coordinate mapping model are ;

[0122] Quantify the time and space information, divide the time into m time intervals , , …… Divide the space into n space regions , , …… , and the matrix element represents the number of times or probability related measure of the failure feature in the jth space region in the ith time interval.

[0123] The position of the failure point is determined by analyzing the distribution of the elements in the matrix.

[0124] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "connected", "connection" should be interpreted broadly, for example, can be fixed connection, can also be detachable connection, or integrally connected; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium. 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.

[0125] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A fault detection method for a DC transmission system, characterized in that, The fault detection method for DC transmission systems includes the following steps: Cameras, DC current sensors, infrared thermal imaging modules, ultraviolet imagers, and partial discharge sensors are added to key detection nodes of power inspection robots to build a multi-dimensional fault perception network. This network captures characteristic signals of DC transmission systems under fault conditions, such as sudden current changes, local overheating, corona discharge, and insulation defects, and collects fault information of DC transmission systems in real time. Based on the edge computing module on the robot, the raw data collected by each sensor is preprocessed and transmitted using 5G and fiber optic dual links. Combined with the timestamp and location tag encoding technology of the data, the spatiotemporal correlation matching of multi-source monitoring data is realized, providing a complete context for fault analysis. By combining multi-source data collected by the robot, an intelligent diagnostic model is constructed. For complex faults, deep learning algorithms, convolutional neural networks, image recognition algorithms, weighted fusion algorithms, random forest algorithms, and LSTM algorithms are used to quickly extract key features of fault data and achieve preliminary classification of multiple fault types. The model takes image data, current, temperature, ultraviolet light, and PD signals collected by the robot as inputs and outputs fault type and confidence level. By combining the robot's GPS location information with the UWB indoor positioning tags in the inspection area, a spatial coordinate mapping model is established. Based on the chronological order of the occurrence of fault characteristics, a time-space correlation matrix is ​​constructed to achieve meter-level accuracy in locating fault points. A fault database for DC transmission systems is constructed, integrating monitoring data, fault types, and handling solutions from historical fault cases. Transfer learning and incremental training techniques are used to continuously optimize the diagnostic model, achieving end-to-end fully automated fault detection.

2. The fault detection method for a DC transmission system according to claim 1, characterized in that: The DC current sensor is used to collect the DC current of the line and equipment in real time and capture the characteristics of sudden current changes during faults. Infrared thermal imaging modules are used to identify the heat of power lines and equipment. Faults in DC equipment are often accompanied by local overheating, so it is necessary to improve the thermal imaging resolution and temperature measurement accuracy to identify temperature abnormalities as low as 0.5℃. Ultraviolet imagers are used to capture the ultraviolet spectrum of 300-400nm, quantify the number of discharge photons, and distinguish between normal corona discharge and fault discharge. Partial discharge sensors are used to detect partial discharge signals in key equipment such as converter valves and DC bushings in converter stations, and to locate insulation defects by using ultra-high frequency or extra-high frequency sensors.

3. The fault detection method for a DC transmission system according to claim 1, characterized in that: The preprocessing of the edge computing module can be specifically implemented as follows: Data dimensionality reduction, eliminating invalid data, and image compression reduce the amount of data transmitted by more than 70%; Real-time filtering allows for early identification of obviously abnormal data, prioritizing the transmission of critical data and ensuring timely fault response.

4. The fault detection method for a DC transmission system according to claim 1, characterized in that: The deep learning algorithm is based on the convolutional neural network architecture and, for different fault detection requirements of DC transmission system images, derives three core task algorithms, specifically including: Image classification algorithms, fault features in DC transmission systems are hidden in deep networks, and residual networks solve the gradient vanishing problem in deep convolutional neural networks through residual blocks, as shown in the following formula: Where x is the residual block, For features obtained after two convolutional layers and ReLU activation, if the number or size of the input and output channels is inconsistent, the dimension of x needs to be adjusted by a 1×1 convolution, i.e.: in, To adjust the convolution kernel for different dimensions; The target detection algorithm determines whether a fault exists in an image and outputs the fault's bounding box. It is adapted to multi-component panoramic images collected by inspection robots. The coordinates of the top-left corner of the grid are set as follows: The model predicts the bounding box offset as follows: The aspect ratio is The absolute coordinates and dimensions of the bounding box are: in, For the Sigmoid function, , Define the width and height of the predefined anchor frame; Semantic segmentation algorithms classify each pixel in an image, accurately segmenting the contours of faults. They are suitable for scenarios with fine and irregular fault regions in DC transmission systems. The algorithm consists of transposed convolutional layers, convolutional layers, and ReLU, progressively restoring the feature map size. Let the input feature map be... transposed convolution kernel Output feature map The dimensions are: Where S is the stride of the transposed convolution and P is the padding number.

5. The fault detection method for a DC transmission system according to claim 1, characterized in that: The weighted fusion algorithm fuses multi-source data, with data collected by sensors A, B, and C respectively. , , The corresponding weights are respectively , , The calculation method for the merged data X is as follows: Data with high reliability and strong correlation with failures are assigned high numerical weights, while data with low reliability and weak correlation with failures are assigned low numerical weights.

6. The fault detection method for a DC transmission system according to claim 1, characterized in that: The random forest algorithm performs in-depth analysis of complex datasets to achieve preliminary classification of multiple fault types. The classification task specifically includes: In a random forest, there are T decision trees. For a sample x, the t-th tree predicts the label as follows: Let the set of categories be K, and the probability formula for a sample x being classified into category k is: in, For indicator functions, when When =k, =1, otherwise 0.

7. The fault detection method for a DC transmission system according to claim 1, characterized in that: The LSTM algorithm described above enables accurate diagnosis and prediction of power system faults. Fault data in power systems exhibits obvious time-series characteristics. By utilizing these time-series characteristics and learning from historical fault data, the algorithm determines and predicts the type and development trend of faults. The algorithm formula is as follows: in, The output of the forget gate at time t. For the forget gate, input at the current time The weight matrix for linear transformation. The state hidden in the previous moment is for forgetting the gate. The weight matrix for linear transformation. For the bias term of the forget gate; in, The input gate output at time t determines the current new information. How much is input into the cell state? For activation function, In the input gate, the input at the current time is... The weight matrix for linear transformation. Let be the candidate cell state at time t, which is the newly generated candidate information based on the current input. It is the hyperbolic tangent activation function; in, Let t represent the cell state at time t. The cell state at time t-1; in, The output of the gate at time t determines the cell state. How many of them are output as hidden states? , In the output gate, the input at the current time... The weight matrix for linear transformation. In the output gate, the hidden state of the previous moment is... The weight matrix for linear transformation. This is the bias term for the output gate.

8. The fault detection method for a DC transmission system according to claim 1, characterized in that: The fault types include short-circuit faults, overload faults, and insulation faults, specifically: Short circuit faults are classified into three-phase short circuits, two-phase short circuits, single-phase-to-ground short circuits, and two-phase-to-ground short circuits. Overload faults are usually caused by electrical equipment operating for a long time in a state exceeding its rated load. When the load exceeds the rated capacity of the equipment, the current will increase and the equipment temperature will rise. Insulation faults manifest as a decline in the insulation performance of electrical equipment. When the insulation performance declines to a certain extent, partial discharge occurs, generating corona and ozone, which accelerates the damage to the insulation.

9. A fault detection method for a DC transmission system according to claim 1, characterized in that: The spatial coordinate mapping model integrates data from different positioning technologies and unifies them into a single coordinate system, enabling a comprehensive and accurate description of the location of power fault points. It employs a coordinate transformation matrix method, assuming the coordinates obtained from GPS measurements are... After coordinate transformation and fusion with UWB data, the coordinates in the unified coordinate system are: The transformation relationship is expressed as: in, For rotation and translation parameters, , , These are the translation parameters.

10. A fault detection method for a DC transmission system according to claim 9, characterized in that: The steps for constructing the time-space correlation matrix specifically include: The fault information collected by the power inspection robot is timestamped to record the exact time when each fault feature occurs. ; Spatial coordinates of the fault point obtained using the spatial coordinate mapping model ; Quantify time and space information, and divide time into m time intervals. , , ... Divide the space into n spatial regions , , ..., Matrix elements This represents the number of times or probability-related measures of fault characteristics appearing in the j-th spatial region within the i-th time interval; The location of the fault point is determined by analyzing the distribution of elements in the matrix.

Images