Intelligent line inspection management system and method based on deep learning

Abstract

The application provides a kind of intelligent line intelligent inspection management system and method based on deep learning.The system comprises: data acquisition device arranged in the area to be inspected, for collecting monitoring data and environmental data on different power lines;Server, for receiving monitoring data and environmental data on different power lines reported by data acquisition device in real time;According to the monitoring data and environmental data of each power line in the preset time period, the power inspection model is used to inspect the power line, and the inspection result is obtained, and the inspection result of each power line is pushed to the related terminal equipment;Wherein, whether the fault occurs and the fault type when the fault occurs are included in the inspection result;The power inspection model is a model capable of predicting and identifying power line faults obtained by optimizing and training long short-term memory network using improved population optimization algorithm.The system and method are used to reduce the dependence of power grid safety inspection process on artificial effect.

Description

Technical Field

[0001] This application relates to the field of intelligent inspection technology for robots, and in particular to a smart line intelligent inspection management system and method based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of electricity demand, the scale of urban power grids is expanding and their structure is becoming increasingly complex. Any fault or accident in the power grid can have serious consequences; therefore, safety inspection is particularly important in the application of urban power grids.

[0003] Traditional power grid safety inspections rely primarily on manual labor. Inspectors must conduct on-site checks along power lines, which is not only inefficient but also difficult to comprehensively cover all areas, especially remote and complex terrain. Furthermore, manual inspections pose safety hazards, with inspectors potentially facing risks such as electric shock and falls from heights. Summary of the Invention

[0004] This application provides a deep learning-based intelligent inspection management system and method for smart lines, which aims to reduce the reliance on manual labor in the safety inspection process.

[0005] In a first aspect, embodiments of this application provide a smart line intelligent inspection and management system based on deep learning, including: a data acquisition device set in the area to be inspected, the data acquisition device being used to collect monitoring data and environmental data on different power lines;

[0006] The server is connected to the data acquisition device and is used to receive real-time monitoring data and environmental data from different power lines reported by the data acquisition device.

[0007] The server is also used to inspect the power lines using a power inspection model based on the monitoring data and environmental data of each power line within a preset time period, and obtain inspection results. The inspection results include whether a fault will occur and the type of fault if a fault occurs. The power inspection model is a model that can predict and identify power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm.

[0008] The server is also used to push the inspection results of each power line to the relevant terminal devices.

[0009] In one possible implementation, the data acquisition device includes multiple monitoring devices for each power line, such as temperature sensors, humidity sensors, and image acquisition devices, for detecting current, voltage, and load information.

[0010] The multiple monitoring devices are used to detect the current, voltage, and load information of the power lines, and the monitoring data of each power line includes current, voltage, and load information.

[0011] The temperature sensor is used to detect the temperature of the environment where the power line is located, and the humidity sensor is used to detect the humidity of the environment where the power line is located.

[0012] The image acquisition device is used to acquire image data of the power line.

[0013] In one possible implementation, the server is specifically used for:

[0014] For each power line, based on the current, voltage, load information, temperature, humidity and image data reported by the data acquisition device, a feature data matrix of the power line is obtained. The feature data matrix includes current, voltage, load information, temperature, humidity and image recognition results at multiple time points. The image recognition results are obtained by anomaly identification of the image data.

[0015] The feature data matrix of the power line is input into the power inspection model. The power inspection model extracts the features of each dimension in a forward sequence according to time order to obtain a forward feature sequence. It also extracts the features of each dimension in a reverse sequence according to time order to obtain a reverse feature sequence. The forward and reverse feature sequences of each dimension are then fused to obtain a fused feature sequence.

[0016] Based on the feature weights of each dimension, the fused feature sequences of multiple dimensions are fused again to obtain time series features;

[0017] Based on the time series characteristics, the power lines are analyzed and processed to obtain the real-time inspection results of the power lines.

[0018] In one possible implementation, the server is further configured to:

[0019] Based on historical data from multiple pre-acquired power lines, a sample set is obtained. The sample set includes multiple samples, and each sample includes current, voltage, load information, temperature, humidity, image recognition results, and fault information for a power line at multiple time points within a set time period. The fault information includes whether a fault occurred within a preset time period after the set time period, and the type of fault if a fault occurred. The duration of the set time period is consistent with the duration of the preset time period.

[0020] Construct multiple initial models with identical structures. Each initial model includes a bidirectional LSTM network, a self-attention layer, a fully connected layer, and a classification output layer. The bidirectional LSTM network includes two LSTM layers for forward and backward feature extraction.

[0021] Based on the sample set and the multiple initial models, model training is performed, and during the training process, the improved Sand Cat population optimization algorithm is used to optimize multiple hyperparameters of the model to obtain the power inspection model.

[0022] In one possible implementation, the server is specifically used for:

[0023] Step a: Initialize the maximum number of iterations of the improved sand cat swarm optimization algorithm, and multiple sand cat individuals. The position of each sand cat individual is used to represent a set of parameter configuration schemes. Each set of parameter configuration schemes includes the learning rate, the number of hidden layer neurons, and the regularization coefficient.

[0024] Step b: Select a set of parameter configuration schemes for each initial model and configure the parameters to obtain multiple initial inspection models;

[0025] Step c: Select multiple samples from the sample set and input them into different initial inspection models to perform line fault inspection processing, obtain the inspection results output by each initial inspection model, and determine the fitness value of each set of parameter configuration schemes based on the output results of each initial inspection model and the fault information in the input samples using the accuracy function of the Dune Cat swarm optimization algorithm.

[0026] Step d: Compare the highest fitness value in this round with the highest fitness value in the previous round, and select the parameter configuration scheme with the highest fitness value as the target parameter configuration scheme for this round;

[0027] Step e: If the preset iteration termination condition is not met, the sand cat group optimization algorithm calculates the R value of each individual sand cat according to the target parameter configuration scheme, and updates the position of the multiple sand cat individuals using the position update formula to obtain new multiple sets of parameter configuration schemes. Steps b to e are repeated until the iteration termination condition is met, and the target parameter configuration scheme obtained in the last iteration is taken as the final parameter configuration scheme.

[0028] Step f: Configure the final parameter configuration scheme in the initial model to obtain the power inspection model.

[0029] In one possible implementation, the server is further specifically used for:

[0030] Based on the position initialization formula, the initial position of each individual sand cat is obtained;

[0031] The position initialization formula is as follows:

[0032]

[0033] Indicates the first The first individual Sand Cat The value of the dimension variable; Indicates the first Upper bound of a dimensional variable; Indicates the first The lower bound of a dimensional variable; A value of 1 indicates a learning rate. The value of 2 indicates the number of neurons in the hidden layer. The value of 3 indicates the regularization coefficient.

[0034] In one possible implementation, the server is further specifically used for:

[0035] Based on the target parameter configuration scheme, determine the optimal position for this iteration;

[0036] Based on the R-value of the sand cat swarm optimization algorithm, the behavior of each individual sand cat is determined. The behavior of each individual sand cat includes attack behavior and search behavior. The R-value is calculated using the following formula:

[0037]

[0038]

[0039] In the above formula, The sensitivity of the pre-defined individual sand cat is defined by T; T represents the maximum number of iterations; t represents the current iteration number. This indicates the sensitivity range of the individual Sand Cat in the current iteration;

[0040] Based on the behavior of each individual sand cat, the position is updated using the position update formula corresponding to the behavior, and the updated position of each individual sand cat is obtained.

[0041] In one possible implementation, the system further includes: multiple terminal devices for managing power lines in different areas and multiple inspection robots;

[0042] The server is also specifically used to generate regional power inspection reports based on the inspection results of each power line. Each regional power inspection report includes the inspection results of all power lines within the corresponding region, as well as the alarms corresponding to the power lines that may fail.

[0043] The server is also used to push regional power inspection reports from different areas to the corresponding terminal devices for display.

[0044] The server is also used to send fault warning commands to inspection robots within a preset range of power lines where faults may occur.

[0045] Upon receiving a fault warning command, the inspection robot will disconnect the power to the faulty power line according to the command.

[0046] In one possible implementation, the server is further configured to:

[0047] Based on the inspection results of each power line and the inspection feedback data returned by the terminal equipment corresponding to each power line, the power inspection model is optimized.

[0048] Secondly, embodiments of this application provide a deep learning-based intelligent inspection and management method for smart lines, applied to a server in the deep learning-based intelligent inspection and management system for smart lines according to the first aspect and / or various possible implementations of the first aspect. The method includes:

[0049] Receives real-time monitoring data and environmental data from different power lines reported by the data acquisition device;

[0050] For each power line, based on the monitoring data and environmental data of the power line within a preset time period, a power line inspection model is used to inspect the power line and obtain inspection results. The inspection results include whether a fault will occur and the type of fault if a fault occurs. The power line inspection model is a model that can predict and identify power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm.

[0051] The inspection results of each circuit line are pushed to the relevant terminal equipment.

[0052] This application provides a deep learning-based intelligent power line inspection management system and method. The system includes: a data acquisition device installed within the inspection area to collect monitoring data and environmental data from different power lines; a server connected to the data acquisition device to receive real-time reports of monitoring data and environmental data from different power lines; based on the monitoring data and environmental data of each power line within a preset time period, a power line inspection model is used to inspect the power lines, obtain inspection results, and push the inspection results of each power line to relevant terminal devices; wherein, the inspection results include whether a fault will occur and the type of fault if a fault occurs; the power line inspection model is a model capable of predicting and identifying power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm. This system aims to reduce the reliance on manual labor in the safety inspection process. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0054] Figure 1 A schematic diagram of the structure of the intelligent inspection management system for smart lines based on deep learning provided in Embodiment 1 of this application;

[0055] Figure 2 A flowchart illustrating the intelligent inspection management system and method for smart lines based on deep learning provided in Embodiment 1 of this application;

[0056] Figure 3 This is a flowchart illustrating the intelligent inspection management system and method for smart lines based on deep learning provided in Embodiment 2 of this application.

[0057] Figure 4 This is a schematic diagram of the structure of a smart line intelligent inspection management system based on deep learning, provided in Embodiment 3 of this application.

[0058] Figure 5 This is a flowchart illustrating a deep learning-based intelligent inspection management system and method for smart lines, as provided in Embodiment 3 of this application.

[0059] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0061] To clearly understand the technical solution of this application, the solutions of the prior art will be described in detail below.

[0062] Power grid safety inspections are a crucial link in ensuring the safe and stable operation of the power system. Through these inspections, potential faults and hidden dangers can be identified and addressed promptly, preventing large-scale power outages and ensuring the continuity and reliability of power supply. Typically, inspection personnel perform routine maintenance and repairs on line equipment using methods such as on-site operation, manual data recording, and reciprocating visits.

[0063] However, relying on manual inspection methods has many drawbacks. Inspectors need to conduct on-site inspections along power lines, which is not only inefficient but also difficult to cover all areas, especially remote areas and areas with complex terrain. Safety inspections of the power grid also need to be carried out at night, which will violate the inspectors' physiological cycles and cause them harm. In addition, the harsh inspection environment can not only cause physical harm to inspectors (such as coal conveying corridors, water treatment rooms, and gas-insulated substations (GIS) rooms), but also expose inspectors to risks such as electric shock and falls from heights (such as cliffs, mountains, and other harsh terrain; extreme weather such as rain, snow, and freezing).

[0064] Besides the aforementioned manual power grid safety inspection methods, there are also manned helicopter inspections, robot inspections, and drone inspections. Among these, manned helicopter inspections have not been widely adopted due to high investment costs, large manpower requirements, complex management, and extensive technical reserves. Robotic and drone inspections, on the other hand, suffer from data silos because the collected data is relatively independent and lacks data exchange and sharing mechanisms. This makes it impossible to centrally analyze and manage data from multiple inspection areas, coordinate work across multiple inspection areas, or make overall decisions.

[0065] Therefore, there is a lack of existing technologies that can replace manual inspections and break down the "data silos" phenomenon in power grid safety inspections.

[0066] Based on the above background technology description, during the research of inspection methods, the inventors discovered that a smart line intelligent inspection management system based on deep learning algorithms can be designed to acquire inspection data in real time and centrally process and analyze the inspection data. This system can achieve unattended, manageable, and controllable intelligent inspection even in complex line environments and harsh working environments.

[0067] Based on the inventor's technical concept, this application provides a deep learning-based intelligent power line inspection system and method. The system includes a data acquisition device and a server installed within the inspection area. The data acquisition device collects monitoring data and environmental data from different power lines. The server, connected to the data acquisition device, receives real-time reports of monitoring and environmental data from different power lines. Based on the monitoring and environmental data for each power line within a preset time period, a power line inspection model is used to inspect the power lines, obtaining inspection results. The inspection results for each power line are then pushed to relevant terminal devices. The inspection results include whether a fault will occur and the type of fault if it does. The power line inspection model is a model capable of predicting and identifying power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm. The inspection method based on this system not only effectively avoids the drawbacks of traditional manual inspections but also breaks down the "data silos" barrier present in robot and drone inspections.

[0068] The intelligent power line inspection and management system based on deep learning provided in this application can perform inspection operations in various power grid environments, such as large urban power grid environments with construction congestion, complex line distribution, and wide-ranging fault impacts. This application does not impose any restrictions on the operating environment of this management system.

[0069] Figure 1 This is a schematic diagram of the structure of the intelligent inspection and management system for smart railway lines based on deep learning provided in Embodiment 1 of this application, as shown below. Figure 1 As shown, the intelligent power line inspection management system 10 based on deep learning provided in this application includes a data acquisition device 101 and a server 102. The acquisition device 101 is connected to the server 102. The acquisition device 101 is used to collect monitoring data and environmental data on different power lines. The server 102 is used to receive the monitoring data and environmental data on different power lines reported in real time by the data acquisition device 101, and to inspect the power lines using a power line inspection model based on the monitoring data and environmental data of each power line within a preset time period, obtain the inspection results, and finally push the inspection results of each power line to the relevant terminal device 103.

[0070] The data acquisition device 101 can be various monitoring devices, sensor devices, image acquisition devices, etc., such as high-definition cameras, infrared sensors, temperature sensors, gas sensors, wind speed sensors, smart surveillance cameras, and IoT smart sensors, etc. This application does not limit them.

[0071] The data acquisition device 101 and the server 102 transmit data wirelessly, such as 5G technology, Zigbee technology, wireless Fidelity (Wifi), Bluetooth, etc.

[0072] Terminal equipment 103 can be a portable mobile communication terminal (such as a smartphone or tablet), computer equipment, cockpit management terminal, and large display screen, etc.

[0073] Furthermore, managers can also use applications and mini-programs on terminal devices to conduct real-time, remote, and intelligent inspections and management of multiple power lines.

[0074] Figure 2 This is a flowchart illustrating the intelligent inspection management system and method for smart railway lines based on deep learning provided in Embodiment 1 of this application. Figure 2 As shown in this embodiment, the intelligent inspection management system and method for smart railway lines based on deep learning includes:

[0075] S201. The data acquisition device collects monitoring data and environmental data from different power lines;

[0076] In this solution, in order to achieve intelligent inspection of power lines, the intelligent power line inspection management system based on deep learning needs to use the data acquisition device in the system to collect relevant data on different power lines in real time, and send the acquired data to the server in the system based on wireless communication technology; the server then analyzes and processes the received data in real time based on the power inspection model to obtain the inspection results of the power lines.

[0077] In this step, to avoid relying on manual inspections, multiple data acquisition devices need to be configured in the system to collect real-time monitoring data and environmental data from different power lines. Monitoring data refers to electrical parameters related to the power lines, such as line voltage, current, and power, as well as line status, such as fire, snow cover, or disconnection. Environmental data refers to information about the environment in which the power lines are located, such as wind speed, temperature, humidity, and pressure.

[0078] In one specific implementation, the data acquisition device can be electrical parameter monitoring equipment, environmental monitoring sensors, and image acquisition equipment installed at locations such as substations or transmission lines within the power grid. The electrical parameter monitoring equipment can collect monitoring data from power lines; the environmental monitoring sensors, for example, can be voltage transformers, current transformers, power sensors, and partial discharge sensors; the environmental monitoring sensors can be meteorological sensors, temperature sensors, humidity sensors, infrared sensors, pressure sensors, etc.

[0079] In one specific implementation, the data acquisition device can be an image acquisition device, such as a high-definition camera or a smart monitoring camera. By using the image acquisition device to collect image data of key parts of the power line (such as line channels, tower bases, conductors, insulators, etc.), the status information of the line can be obtained based on the acquired image data, such as the line being on fire, the line being covered by snow, or the line being disconnected.

[0080] In one specific implementation, the data acquisition device can be an Internet of Things (IoT) smart sensor. An IoT smart sensor is a new type of sensor that integrates electrical parameter monitoring equipment with environmental monitoring sensors, enabling it to collect monitoring data such as electrical parameters and operating status of power equipment, as well as environmental data.

[0081] In one specific implementation, the data acquisition device includes multiple monitoring devices for each power line, such as temperature sensors, humidity sensors, and image acquisition devices, for detecting current, voltage, and load information.

[0082] In this implementation, multiple monitoring devices are used to detect the current, voltage, and load information of the power lines. Correspondingly, the monitoring data collected by the multiple monitoring devices for each power line includes current, voltage, and load information. Temperature sensors are used to detect the temperature of the environment where the power lines are located, and humidity sensors are used to detect the humidity of the environment where the power lines are located. Image acquisition devices are used to acquire image data of the power lines.

[0083] Specifically, multiple monitoring devices can be voltage transformers, current transformers, power sensors, etc.

[0084] Optionally, one image acquisition device can monitor one power line, or one image acquisition device can monitor multiple power lines simultaneously.

[0085] S202, The data acquisition device sends the collected monitoring data and environmental data from different power lines to the server.

[0086] S202. The server receives real-time monitoring data and environmental data from different power lines reported by the data acquisition device.

[0087] In the above steps, the data acquisition device needs to transmit the real-time collected monitoring data and environmental data to the server via wireless communication, so that the server can process and analyze the data. Correspondingly, the server needs to receive the real-time monitoring data and environmental data from different power lines reported by the data acquisition device.

[0088] In one specific implementation, wireless communication methods such as 5G, Zigbee, Wi-Fi, and Bluetooth can be used.

[0089] Optionally, a smart gateway can be used to authenticate the data acquisition devices set up in the system and encrypt the data reported by the data acquisition devices to ensure the stability, reliability, and security of the data received by the server. The smart gateway can be an Acrel series smart gateway. Specifically, an authentication mechanism can be used to determine the authorization status of each data acquisition device set up in the system, ensuring that only authorized data acquisition devices can report data to the server. This not only ensures data reliability and improves the accuracy of server data analysis and processing but also effectively prevents unauthorized access and attacks by external malicious devices, reducing the security risks of system attacks, virus intrusions, etc. Encryption algorithms can be used to encrypt the data, transmitting it in encrypted form to effectively prevent data leakage and tampering, which could lead to inaccurate analysis of data from various power lines in the power grid by the subsequent server analysis.

[0090] S203. For each power line, the server uses a power inspection model to inspect the power line based on the monitoring data and environmental data of the power line within a preset time period, and obtains the inspection results.

[0091] In this step, after receiving the monitoring data and environmental data of the power lines within a preset time period, the server predicts and judges the monitoring data and environmental data of each power line based on a pre-built and trained power inspection model to obtain inspection results. For a given power line, after inputting the monitoring data and environmental data of that power line within the preset time period into the power inspection model, it will determine whether a line fault is about to occur on the power line, and if so, determine the type of fault, thus obtaining inspection results for regulatory personnel to review.

[0092] Specifically, the inspection results include whether a fault will occur and the type of fault if it does occur; the power line inspection model is a model obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm, which can predict and identify power line faults.

[0093] In one specific implementation, if the data acquisition device includes multiple monitoring devices for detecting current, voltage, and load information for each power line, as well as temperature sensors, humidity sensors, and image acquisition devices, and correspondingly, the monitoring data and environmental data for each power line include the power line's current, voltage, and load information, the humidity and temperature of the environment where the power line is located, and image data of the power line; then, the server performs inspection on the acquired data in the following manner, which includes:

[0094] S2031. For each power line, based on the current, voltage, load information, temperature, humidity and image data reported by the data acquisition device, obtain the feature data matrix of the power line.

[0095] In this step, the feature data matrix includes current, voltage, load information, temperature, humidity, and image recognition results at multiple time points. The image recognition results are obtained by anomaly identification of the image data.

[0096] The number of time points in each feature matrix is ​​determined by the preset data collection frequency and the preset time period duration. For example, if the model is preset to process multidimensional time series within 10 minutes at a time, and the sample collection frequency is 1 time / second, then the number of time points in each time feature matrix is ​​600.

[0097] In one possible implementation, in order to reduce the amount of data processed by the server and improve the server's data processing speed, the server may no longer receive data reported by data acquisition devices whose data values ​​remain constant as the power lines operate.

[0098] In one possible implementation, the image recognition result includes the state type of the power line, such as normal state, and abnormal states such as fire state and line disconnection state. This application does not limit the types of abnormal states that can be included in the image recognition result. The image recognition result can be obtained based on a pre-trained image recognition model.

[0099] Specifically, based on a pre-trained image recognition model, the image data reported by the data acquisition device at the current moment is processed to obtain the image features extracted from the image data at the current moment; the extracted image features at the current moment are compared with the pre-extracted image features of the power line when it is in normal condition to obtain the difference features between the two image features; based on the difference features between the two images, the image data at the current moment is classified to obtain the image recognition result of the power line; wherein, the image recognition result includes the status information of the power line.

[0100] Furthermore, the image recognition model can be a Convolutional Neural Network (CNN) model. The following steps can be used to train a CNN model:

[0101] Step 1: Collect image data of power lines when they are normal and image data of power lines when they are abnormal, and label the state type of each image data. For example, normal state is labeled as 0; fire state is labeled as 1; disconnection state is labeled as 2, etc., to obtain the training sample set of the image recognition model.

[0102] Step 2: Based on the constructed initial convolutional neural network model, the model is trained using the training sample set of the image recognition model to obtain the image recognition model. The initial convolutional neural network model includes convolutional layers, pooling layers, and fully connected layers.

[0103] In this implementation, a CNN model is used to perform image recognition processing on real-time image data of each power line captured by multiple cameras to obtain image recognition results that can indicate the status of each power line. This ensures the accuracy of image data recognition. Furthermore, the image recognition results are used as a feature of one dimension in the subsequent power inspection model, which can help the power inspection model to more quickly determine whether a power line is about to (or has already) experienced a fault. At the same time, based on the power line recognition results, the accuracy of fault type judgment can be further improved.

[0104] Optionally, the image recognition results may include location information of at least one power line covered in each image data. Correspondingly, the inspection results may include location information of power lines that are about to fail (or have already failed), to indicate the location of the power lines to be repaired, so that the corresponding power lines can be quickly addressed.

[0105] S2032. Input the feature data matrix of the power line into the power inspection model. The power inspection model extracts the features of each dimension in a forward sequence according to the time order to obtain a forward feature sequence. It also extracts the features of each dimension in a reverse sequence according to the time order to obtain a reverse feature sequence. Finally, it fuses the forward and reverse feature sequences of each dimension to obtain a fused feature sequence.

[0106] In this step, the core of the power line inspection model is a bidirectional Short Term Memory (LSTM) network. For a feature of one dimension within a time step, after inputting into the power line inspection model, firstly, in forward chronological order, the forward layer extracts the positive feature vector for that dimension within that time step, storing it in the hidden layer of the forward layer. Simultaneously, in reverse chronological order, the backward layer extracts the negative feature vector for that dimension within that time step, storing it in the hidden layer of the backward layer. Based on this, the positive and negative feature vectors of a feature of one dimension within a time step are obtained. Further, this method is used to obtain the positive and negative feature vectors of each dimension across multiple time steps, resulting in the positive and negative feature sequences for each dimension. Finally, the positive and negative feature sequences stored in the hidden layers for each dimension are integrated to obtain a fused feature sequence across multiple time steps for each dimension, which is then used as the output of the bidirectional LSTM network. The fused feature sequence output by the bidirectional LSTM network contains fused feature vectors corresponding to the features of each dimension within multiple time steps.

[0107] It should be understood that, compared to LSTM networks, bidirectional LSTM networks extract features from both directions of the time series simultaneously, building upon the LSTM network. This allows for a more comprehensive extraction of time series features, enabling accurate prediction of whether power lines will experience faults within a predetermined timeframe.

[0108] S2033. Based on the feature weights of each dimension, the fused feature sequences of multiple dimensions are fused again to obtain time series features.

[0109] In this step, applied to the power inspection model, in order to obtain more accurate and reasonable prediction results, in the output layer of the bidirectional LSTM network, the fused feature sequence of each dimension is fused again according to the fused feature sequence output by the bidirectional LSTM network and the feature weight of each dimension, so as to obtain the final time series features.

[0110] In one specific implementation, the output layer of the bidirectional LSTM network is a self-attention layer. For each dimension's fused feature vectors within a time step, the self-attention layer assigns each vector in the input fused feature sequence to the weight matrix W of the linear transformation learned during model training. Q W K W VMultiply them to obtain the query vector Q, matching vector K, and target vector V for each vector in the sequence of fused features;

[0111] Using formula The self-attention score of each fused feature vector in the fused feature sequence is obtained. It represents the structure composed of all vectors in the fused feature sequence. 3D matrix; This indicates the number of vectors in the fused feature sequence;

[0112] The self-attention score of each fused feature is normalized using the Softmax function to obtain the self-attention weight of each fused feature vector in the fused feature sequence; based on the self-attention weight of each fused feature vector and the target vector, the final time series features are obtained.

[0113] In this implementation, since the contribution of features of different dimensions to the final result often varies, the self-attention layer dynamically assigns different weights to each dimension of features based on the internal structure of the data of each dimension in the input data learned during the training process and the correlation between the data of different dimensions. This is to distinguish the importance of each dimension of features. At the same time, through the self-attention layer, more valuable information can be better filtered out from the data, which can improve the efficiency of power inspection model in analyzing and processing data.

[0114] S2034. Based on the time series characteristics, the power lines are analyzed and processed to obtain the real-time inspection results of the power lines.

[0115] In this step, the real-time acquired time-series features are input into the subsequent feature classification layer in the power inspection model. The time-series features of each power line in the power grid are classified and processed in multiple dimensions, and the inspection results corresponding to the time-series features of each power line in multiple dimensions can be obtained in real time.

[0116] In one possible implementation, in the power line inspection model, the output layer of the attention layer is a fully connected layer, and the output layer of the fully connected layer is a classification output layer. For each power line, the fully connected layer and the classification output layer analyze and process the time series features output from the attention layer, and the classification result output by the classification output layer is the inspection result of that power line.

[0117] S204. The server pushes the inspection results of each circuit line to the relevant terminal devices.

[0118] In this step, in order to enable supervisory personnel to respond to the inspection results in a timely manner and to carry out timely maintenance and repair of faulty lines, the inspection results of each power line need to be pushed to the terminal equipment corresponding to each power line for supervisory personnel to view.

[0119] Optionally, only the relevant information of the faulty power line can be pushed to the terminal equipment corresponding to the faulty line to reduce unnecessary data transmission and reduce the processing burden on the terminal equipment.

[0120] In one specific implementation, the terminal device 20 can be a portable mobile communication terminal (such as a smartphone or tablet), a computer device, a cockpit management terminal, or an indicator screen, etc.

[0121] Optionally, managers can also use applications, mini-programs, etc. on terminal devices to conduct real-time, remote, and intelligent inspections and management of multiple power lines.

[0122] The intelligent power line inspection management system and method based on deep learning provided in this application embodiment receive real-time monitoring data and environmental data collected by data acquisition devices on different power lines through a server. For each power line within a preset time period, a power line inspection model is used to inspect the monitoring data and environmental data, obtaining inspection results including whether each power line will experience a fault and the type of fault if it does. The power line inspection model is a model that can predict and identify power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm. Finally, the inspection results of each power line are pushed to relevant terminal devices to achieve intelligent safety inspection of the power grid, reducing reliance on manual inspection and effectively avoiding many drawbacks of traditional manual inspection. Meanwhile, the method provided in this embodiment, through the setting of a deep learning-based intelligent line inspection management system, simultaneously acquires relevant data of power lines in different areas, and performs unified processing of the data based on the power inspection model in the server, breaking down the "data silo" barrier existing in robot inspection and drone inspection, realizing centralized management of relevant data of power lines in multiple areas, which is conducive to the overall planning of the power grid.

[0123] Figure 3 This is a flowchart illustrating the intelligent inspection management system and method for smart railway lines based on deep learning provided in Embodiment 2 of this application, as shown below. Figure 3 As shown, based on the above embodiments, the intelligent power line inspection management system and method based on deep learning provided in this application further includes a method for obtaining a power line inspection model. This method is applied to the system's server and includes:

[0124] S301. Obtain a sample set based on historical data of multiple power lines acquired in advance.

[0125] In this step, the historical data for multiple power lines includes monitoring data and environmental data collected by the data acquisition device from different power lines. The sample set obtained from the historical data includes multiple samples, each containing current, voltage, load information, temperature, humidity, image recognition results, and fault information for a power line at multiple time points within a set time period. The fault information includes whether a fault occurred within a preset time period after the set time period, and the type of fault if it occurred. The duration of the set time period is consistent with the preset time period.

[0126] Specifically, in the sample set, the current, voltage, load information, temperature, humidity, and image recognition results of a power line in each sample at multiple time points within a set time period are used as the input to the model, and the fault information corresponding to the current, voltage, load information, temperature, humidity, and image recognition results at the aforementioned multiple time points is used as the output of the model training.

[0127] The fault type in the fault information is represented by a tag code. For example, the tag code for overload fault is 1, the tag code for short circuit fault is 2, the tag code for ground fault is 3, the tag code for open circuit fault is 4, the tag code for high resistance fault is 5, and the tag code for lightning strike fault is 6.

[0128] In one specific implementation, the sample set is divided into a training set, a test set, and a validation set for model training. For example, 80% of the sample set is used as the training set, 20% as the test set, and 20% of the training set is used as the validation set.

[0129] In one possible implementation, to ensure the reliability of the data in the sample set, the data needs to be preprocessed. This preprocessing includes: removing data with obvious errors; and performing mean-reduction and variance normalization on the data for each feature corresponding to each power line, as shown in the following formula:

[0130]

[0131] In the above formula, Original sample data representing a one-dimensional feature corresponding to a certain power line; This represents the mean of all sample data for a specific feature corresponding to a particular power line. The standard deviation of all sample data representing a certain dimension of a feature corresponding to a certain power line; This represents the normalized sample data corresponding to a certain feature of a power line.

[0132] S302. Construct multiple initial models with identical structures. Each initial model includes a bidirectional LSTM network, a self-attention layer, a fully connected layer, and a classification output layer. The bidirectional LSTM network includes two LSTM layers for forward and reverse feature extraction.

[0133] In this step, multiple identical initial models need to be built simultaneously. These models are used to select the optimal parameter configuration from multiple options during the subsequent hyperparameter optimization process. The initial models include a bidirectional LSTM network, a self-attention layer, a fully connected layer, and a classification output layer.

[0134] Optionally, the classification output layer can be a Softmax classifier, which can perform classification processing tasks with multiple labels and multiple classifications. For example, a power line can have multiple fault types.

[0135] Specifically, each LSTM network's main structure includes a gating unit. This gating unit comprises a forget gate, an input gate, and an output gate.

[0136] S303. Based on the sample set and multiple initial models, the model is trained, and during the training process, the improved Sand Cat population optimization algorithm is used to optimize multiple hyperparameters of the model to obtain the power inspection model.

[0137] In this step, for an initial model, a hyperparameter combination needs to be assigned to obtain an initial inspection model. During its training process, it is trained according to pre-set training parameters. The hyperparameter combination is a combination of the learning rate, the number of hidden layer neurons, and the regularization coefficient. The pre-set training parameters include the batch size and the number of training iterations. This application does not limit the settings for the batch size and the number of training iterations. For example, the pre-set training parameters can be: a batch size of 64 and a number of training iterations of 100.

[0138] Specifically, based on the improved Sand Cat population optimization algorithm, the optimal combination of hyperparameters is obtained during the model training process, and the training model corresponding to this hyperparameter combination is used as the final power line inspection model.

[0139] In one specific implementation, the following method is used to optimize multiple hyperparameters of the model during training to obtain a reasonable parameter configuration scheme for the learning rate, the number of hidden layer nodes, and the regularization coefficient. This method includes:

[0140] Step a: Initialize the maximum number of iterations of the improved sand cat swarm optimization algorithm, and multiple sand cat individuals. The position of each sand cat individual is used to represent a set of parameter configuration schemes. Each set of parameter configuration schemes includes the learning rate, the number of hidden layer neurons, and the regularization coefficient.

[0141] Step b: Select a set of parameter configuration schemes for each initial model and configure the parameters to obtain multiple initial inspection models;

[0142] In this step, during the initial iteration, the initial position of each individual sand cat can be obtained based on the position initialization formula, and multiple sets of parameter configuration schemes corresponding to multiple initial positions can be configured in multiple initial models to obtain multiple initial inspection models. In each subsequent iteration, based on the position update method in step e, the new position of each individual sand cat can be obtained, and multiple sets of parameter configuration schemes corresponding to the new positions of multiple individuals sand cat can be configured in multiple initial inspection models to obtain multiple initial inspection models.

[0143] In one specific implementation, the initial position of each individual sand cat can be obtained in the following way:

[0144] Based on the position initialization formula, the initial position of each individual sand cat is obtained;

[0145] The position initialization formula is as follows:

[0146]

[0147] Indicates the first The first individual Sand Cat The value of the dimension variable; Indicates the first Upper bound of a dimensional variable; Indicates the first The lower bound of a dimensional variable; A value of 1 indicates a learning rate. The value of 2 indicates the number of neurons in the hidden layer. The value of 3 indicates the regularization coefficient.

[0148] Specifically, the upper and lower bounds of the learning rate, the number of hidden layer neurons, and the regularization coefficient are preset, and the magnitude of these values ​​is not limited in this application. For example, the learning rate is preset to a range of [0.001, 0.99], the number of hidden layer neurons to a range of [3, 5], and the regularization coefficient to a range of [-1, 1].

[0149] In one possible implementation, to make the initial position distribution of multiple sand cat individuals more uniform and thus expand the coverage of the search space, multiple three-dimensional vectors can be obtained using the following method to replace the initial position in the position initialization formula. The method includes:

[0150] Based on the formula for generating random sequences, and a randomly obtained three-dimensional vector. Obtain N three-dimensional vectors; where N is the pre-defined number of individual sand cats. All are taken from Random numbers between intervals; where the formula for generating a random sequence is:

[0151]

[0152] In the above formula, Indicates the first The first individual Sand Cat The random term of the dimensional variable, ; For control parameters; This represents the value of the Tth iteration;

[0153] Correspondingly, the position initialization formula is: .

[0154] In this implementation, the random sequence generation formula described above is used to obtain a sequence that is more complex than the ordinary one. The more evenly distributed random terms in the algorithm result in a more uniform initial position distribution for individual sand cats, thereby expanding the coverage of the search space and helping to accelerate the efficiency of the hyperparameter optimization process.

[0155] Step c: Select multiple samples from the sample set and input them into different initial inspection models to process line fault inspections. Obtain the inspection results output by each initial inspection model. Based on the output results of each initial inspection model and the fault information in the input samples, use the accuracy function of the Dune Cat swarm optimization algorithm to determine the fitness value of each set of parameter configuration schemes.

[0156] In this step, a subset of sample data needs to be selected from the sample set to train and validate different initial inspection models. The precision of each initial inspection model is determined based on the number of true positives and false positives in the validation results. This, in turn, determines the fitness value of the parameter configuration scheme corresponding to each initial inspection model. The precision function is: , Indicates the number of true cases predicted; This indicates the number of predicted false positives.

[0157] Step d: Compare the highest fitness value in this round with the highest fitness value in the previous round, and select the parameter configuration scheme with the highest fitness value as the target parameter configuration scheme for this round;

[0158] Step e: If the preset iteration termination condition is not met, then according to the target parameter configuration scheme, the R value of each individual sand cat in the sand cat group optimization algorithm is used to update the position of the multiple sand cat individuals using the position update formula to obtain new multiple sets of parameter configuration schemes. Repeat steps b to e until the iteration termination condition is met, and take the target parameter configuration scheme obtained in the last iteration as the final parameter configuration scheme.

[0159] In this step, to determine the optimal hyperparameter combination more quickly, the positions of the individual Sand Cats need to be continuously adjusted so that multiple Sand Cats continuously search for the optimal solution (the behavior corresponding to the Sand Cat is the search behavior) or move towards the optimal solution (the behavior corresponding to the Sand Cat is the attack behavior). If the current iteration count is less than the maximum iteration count, it is determined that the preset iteration condition is not met.

[0160] Optionally, the location of each individual sand cat can be updated in the following way:

[0161] Step e.1: Determine the optimal position for this iteration based on the target parameter configuration scheme;

[0162] Step e.2: Calculate the R-value for each individual sand cat using the sand cat swarm optimization algorithm, and determine the behavior of each individual sand cat. The behavior of an individual sand cat includes attack behavior and search behavior, and the R-value is calculated using the following formula:

[0163]

[0164]

[0165] In the above formula, The sensitivity of the pre-defined individual sand cat is defined by T; T represents the maximum number of iterations; t represents the current iteration number. This indicates the sensitivity range of the individual Sand Cat in the current iteration;

[0166] In this step, the random number selected when calculating the R value for each individual Sand Cat is... They may all be different. When | When |≤1, the Sand Cat individual's behavior is aggressive; when | When |>1, the behavior of the individual Sand Cat is a search behavior.

[0167] Step e.3: Based on the behavior of each individual Sand Cat, update the position using the position update formula corresponding to the behavior to obtain the updated position of each individual Sand Cat.

[0168] In this step, for each individual Sand Cat, if the Sand Cat's behavior is aggressive, the position of the Sand Cat in the next iteration is calculated using the following formula:

[0169]

[0170]

[0171] in, Indicates the position of the individual Sand Cat in the next iteration; Indicates the optimal position in this iteration; Represents any random number between (0,1); This indicates the random location of an individual sand cat; Indicates the current location of the individual Sand Cat; This represents the predation angle of a randomly selected individual sand cat, and ;

[0172] If the sand cat's behavior is a search behavior, then the new location of the individual sand cat is calculated using the following formula:

[0173]

[0174] In the above formula, Random numbers that follow a normal distribution; It is a randomly generated 1 A 3D matrix, where all values ​​are between (0,1).

[0175] Specifically, based on the above position update formula, the position of each individual Sand Cat is dynamically adjusted (i.e., multiple parameter configuration schemes are continuously updated), so that multiple Sand Cat individuals continuously search for the optimal solution or move towards the optimal solution, quickly determining the best hyperparameter combination to optimize the hyperparameters of the power line inspection model.

[0176] In one possible implementation, in order to improve the iteration speed of the sand cat swarm optimization algorithm, before updating the position of the sand cat individuals, the sand cat individual farthest from the optimal position of the current iteration can be determined based on the distance between each sand cat individual and the optimal position of the current iteration. For that sand cat individual, the position of the sand cat individual is directly updated to the optimal position.

[0177] Step f: Configure the final parameter configuration scheme in the initial model to obtain the power inspection model.

[0178] In this step, the hyperparameter combination in the final parameter configuration scheme is the optimal combination. Therefore, this optimal hyperparameter combination needs to be configured in the initial model, and the model needs to be trained using a complete sample set to obtain the final power line inspection model.

[0179] In this implementation, the improved Dune Cat swarm optimization algorithm, with its powerful global search capability, comprehensively and rapidly explores the hyperparameter search space to optimize the hyperparameters of the power line inspection model. This method effectively avoids getting trapped in local optima, ensuring that the hyperparameter combination used in the power line inspection model is optimal. Furthermore, a reasonable learning rate parameter can reduce slow convergence or convergence failure during training; reasonable hidden layer node parameters not only enhance the model's analytical capabilities but also prevent overfitting or underfitting; and a reasonable regularization coefficient helps improve the model's generalization ability. Therefore, when using the power line inspection model with the optimal hyperparameter combination for data analysis and processing, it can make more accurate predictions and judgments about data from different power lines, improving the accuracy and speed of data analysis and processing by the power line inspection model.

[0180] In one specific implementation, in the practical application of the power model, the power inspection model can be optimized based on the inspection results of each power line and the inspection feedback data returned by the terminal equipment corresponding to each power line.

[0181] In this step, to further optimize and adjust the power line inspection model and improve its prediction accuracy, the terminal equipment corresponding to each power line can evaluate the acquired power line inspection model using four indicators: accuracy, precision, recall, and F-score. Based on the evaluation results, inspection feedback data is returned to the server. The server then optimizes the power line inspection model based on the inspection feedback data returned by the terminal equipment. Specifically, the evaluation indicators can be calculated using the following formula:

[0182]

[0183]

[0184]

[0185]

[0186] In the above formula, Indicates accuracy; Indicates accuracy; Indicates recall rate; Indicates the number of true cases predicted; Indicates the number of predicted false positives; Indicates the number of true negatives predicted; This indicates the number of predicted false negatives.

[0187] Specifically, the hyperparameters of the model can be optimized based on the performance evaluation data fed back during the actual application of the power inspection model.

[0188] In this implementation, the power inspection model is optimized and adjusted based on the inspection results in actual applications and the corresponding inspection feedback data, so as to continuously improve the prediction accuracy and stability of the model and reduce the error rate of inspection results.

[0189] The deep learning-based intelligent power line inspection management system and method provided in this application, based on the above embodiments, also provides a method for constructing and training a power line inspection model. During model training, the optimal hyperparameter combination is obtained based on the global search strategy and algorithmic advantages of the improved population optimization algorithm, thereby acquiring the power line inspection model with optimal performance. Based on this power line inspection model, the data reported by the data acquisition device is analyzed and processed, and the inspection results of each power line in the power grid are obtained in real time. The performance advantages of the power line inspection model itself ensure the accuracy and stability of data analysis.

[0190] Based on the above embodiments, Figure 4 This is a schematic diagram of the structure of a deep learning-based intelligent railway line inspection and management system provided in Embodiment 3 of this application. Figure 4 As shown, the intelligent line inspection management system 10 based on deep learning also includes: multiple terminal devices 103 for managing power lines in different areas and multiple inspection robots 104.

[0191] Among them, the terminal device 103 can be a portable mobile communication terminal (such as a smartphone or tablet), computer equipment, cockpit management terminal, and large display screen; the inspection robot 104 can include indoor inspection robots and outdoor inspection robots.

[0192] Figure 5 This is a flowchart illustrating a deep learning-based intelligent inspection management system and method for smart railway lines, as shown in Embodiment 3 of this application. Figure 5 As shown, based on the above embodiments, the intelligent inspection management system for smart lines based on deep learning provided in this application can also take the following measures to alarm and process abnormal inspection results, the method including:

[0193] S401. The server generates a regional power inspection report based on the inspection results of each power line. Each regional power inspection report includes the inspection results of all power lines within the corresponding region, as well as the alarms corresponding to the power lines that may fail.

[0194] In this step, fault information for each power line within a given area can be obtained based on the power line inspection model. Furthermore, if any power line is about to fail, the server will also add the alarm information corresponding to that line to the inspection report to instruct supervisory personnel to carry out appropriate maintenance.

[0195] S402. The server pushes regional power inspection reports for different areas to the corresponding terminal devices for display.

[0196] S403, The terminal device displays the regional power inspection report within its corresponding area.

[0197] In this solution, supervisory personnel within the corresponding area can promptly repair and handle power lines based on alarm information displayed in the power inspection report on the terminal device, thereby achieving supervision of all power lines within their area.

[0198] Alternatively, inspection robots can quickly respond to fault information in the inspection results, as detailed below:

[0199] S501: The server sends a fault warning command to the inspection robot within a preset range of the power line where a fault may occur.

[0200] In this step, each power line is equipped with an inspection robot within a preset range. When the server detects a power line that is about to fail based on the power line inspection results, it sends a fault warning command to the inspection robots within the preset range of the circuit that is about to fail.

[0201] The preset range for each power line can be determined based on the actual robot configuration and the size of the inspection area; this application does not impose any restrictions on this. For example, when the inspection area is small, or when the inspection area is large but contains a large number of inspection robots, the preset range for the power lines within that inspection area can be small, such as 100 meters.

[0202] S502. Upon receiving a fault warning command, the inspection robot will disconnect the power to the faulty power line according to the fault warning command.

[0203] In this solution, if the inspection robot receives a fault warning command, it will use the power line identification information in the command message to cut off the power to the power line that is about to fail, effectively preventing the adverse effects of the line fault and realizing rapid supervision of the power inspection model in its area.

[0204] In one specific implementation, after receiving a fault warning command, the inspection robot can send a trip command to the intelligent circuit breaker installed on the power line where the fault is about to occur, as indicated by the command information, via wireless communication; the intelligent circuit breaker then controls the relay protection device or circuit breaker of the power line to trigger a trip operation based on the trip command.

[0205] The intelligent power line inspection management system and method based on deep learning provided in this application pushes alarm information from power inspection results to the terminal equipment corresponding to the line that is about to fail. Based on the alarm information, supervisory personnel can promptly repair the line; or send it to the inspection robot corresponding to the faulty line. Based on the alarm information, the inspection robot can promptly cut off the power to the line that is about to fail. This method realizes rapid early warning and maintenance of lines that are about to experience abnormalities. This solution not only reduces the reliance on manual power inspection, but also completes the inspection task faster and more accurately than manual inspection. Moreover, it can predict and warn whether a line will fail before the line fails, reducing the maintenance cost of line failures and reducing unnecessary losses caused by line failures to a certain extent.

[0206] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0207] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A smart railway line intelligent inspection and management system based on deep learning, characterized in that, include: A data acquisition device is installed within the area to be inspected. This device collects monitoring data and environmental data from different power lines. The data acquisition device includes multiple monitoring devices for each power line to detect current, voltage, and load information, as well as temperature sensors, humidity sensors, and image acquisition devices. The multiple monitoring devices detect the current, voltage, and load information of the power lines, and the monitoring data for each power line includes current, voltage, and load information. The temperature sensor detects the temperature of the environment surrounding the power line, and the humidity sensor detects the humidity of the environment surrounding the power line. The image acquisition device collects image data of the power lines. The server is connected to the data acquisition device and is used to receive real-time monitoring data and environmental data from different power lines reported by the data acquisition device. The server is also used to inspect the power lines using a power inspection model based on the monitoring data and environmental data of each power line within a preset time period, and obtain inspection results. The inspection results include whether a fault will occur and the type of fault if a fault occurs. The power inspection model is a model that can predict and identify power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm. The server is also used to push the inspection results of each power line to the relevant terminal devices. Specifically, the server is used for: For each power line, based on the current, voltage, load information, temperature, humidity and image data reported by the data acquisition device, a feature data matrix of the power line is obtained. The feature data matrix includes current, voltage, load information, temperature, humidity and image recognition results at multiple time points. The image recognition results are obtained by anomaly identification of the image data. The feature data matrix of the power line is input into the power inspection model. The power inspection model extracts the features of each dimension in a forward sequence according to time order to obtain a forward feature sequence. It also extracts the features of each dimension in a reverse sequence according to time order to obtain a reverse feature sequence. The forward and reverse feature sequences of each dimension are then fused to obtain a fused feature sequence. Based on the feature weights of each dimension, the fused feature sequences of multiple dimensions are fused again to obtain time series features; Based on the time series characteristics, the power lines are analyzed and processed to obtain the real-time inspection results of the power lines.

2. The system according to claim 1, characterized in that, The server is also used for: Based on historical data from multiple pre-acquired power lines, a sample set is obtained. The sample set includes multiple samples, and each sample includes current, voltage, load information, temperature, humidity, image recognition results, and fault information for a power line at multiple time points within a set time period. The fault information includes whether a fault occurred within a preset time period after the set time period, and the type of fault if a fault occurred. The duration of the set time period is consistent with the duration of the preset time period. Construct multiple initial models with identical structures. Each initial model includes a bidirectional LSTM network, a self-attention layer, a fully connected layer, and a classification output layer. The bidirectional LSTM network includes two LSTM layers for forward and backward feature extraction. Based on the sample set and the multiple initial models, model training is performed, and during the training process, the improved Sand Cat population optimization algorithm is used to optimize multiple hyperparameters of the model to obtain the power inspection model.

3. The system according to claim 2, characterized in that, The server is specifically used for: Step a: Initialize the maximum number of iterations of the improved sand cat swarm optimization algorithm, and multiple sand cat individuals. The position of each sand cat individual is used to represent a set of parameter configuration schemes. Each set of parameter configuration schemes includes the learning rate, the number of hidden layer neurons, and the regularization coefficient. Step b: Select a set of parameter configuration schemes for each initial model and configure the parameters to obtain multiple initial inspection models; Step c: Select multiple samples from the sample set and input them into different initial inspection models to perform line fault inspection processing, obtain the inspection results output by each initial inspection model, and determine the fitness value of each set of parameter configuration schemes based on the output results of each initial inspection model and the fault information in the input samples using the accuracy function of the Dune Cat swarm optimization algorithm. Step d: Compare the highest fitness value in this round with the highest fitness value in the previous round, and select the parameter configuration scheme with the highest fitness value as the target parameter configuration scheme for this round; Step e: If the preset iteration termination condition is not met, then according to the target parameter configuration scheme, the R value of each individual sand cat in the sand cat group optimization algorithm is used to update the position of the multiple sand cat individuals using the position update formula to obtain new multiple sets of parameter configuration schemes. Repeat steps b to e until the iteration termination condition is met, and take the target parameter configuration scheme obtained in the last iteration as the final parameter configuration scheme. Step f: Configure the final parameter configuration scheme in the initial model to obtain the power inspection model.

4. The system according to claim 3, characterized in that, The server is also specifically used for: Based on the position initialization formula, the initial position of each individual sand cat is obtained; The position initialization formula is as follows: ， Indicates the first The first individual Sand Cat The value of the dimension variable; Indicates the first Upper bound of a dimensional variable; Indicates the first The lower bound of a dimensional variable; A value of 1 indicates a learning rate. The value of 2 indicates the number of neurons in the hidden layer. The value of 3 indicates the regularization coefficient.

5. The system according to claim 3, characterized in that, The server is also specifically used for: Based on the target parameter configuration scheme, determine the optimal position for this iteration; The R-value of each individual sand cat is calculated based on the sand cat swarm optimization algorithm to determine the behavior of each individual sand cat. The behavior of each individual sand cat includes attack behavior and search behavior. The R-value is calculated using the following formula: ， ， In the above formula, The sensitivity of the pre-defined individual sand cat is defined by T; T represents the maximum number of iterations; t represents the current iteration number. This indicates the sensitivity range of the individual Sand Cat in the current iteration; Based on the behavior of each individual sand cat, the position is updated using the position update formula corresponding to the behavior, and the updated position of each individual sand cat is obtained.

6. The system according to claim 1, characterized in that, The system also includes: multiple terminal devices for managing power lines in different areas and multiple inspection robots; The server is also specifically used to generate regional power inspection reports based on the inspection results of each power line. Each regional power inspection report includes the inspection results of all power lines within the corresponding region, as well as the alarms corresponding to the power lines that may fail. The server is also used to push regional power inspection reports from different areas to the corresponding terminal devices for display. The server is also used to send fault warning commands to inspection robots within a preset range of power lines where faults may occur. Upon receiving a fault warning command, the inspection robot will disconnect the power to the faulty power line according to the command.

7. The system according to claim 1, characterized in that, The server is also used for: Based on the inspection results of each power line and the inspection feedback data returned by the terminal equipment corresponding to each power line, the power inspection model is optimized.

8. A method for intelligent inspection and management of smart railway lines based on deep learning, characterized in that, The method, applied to a server in the deep learning-based intelligent railway line inspection management system according to any one of claims 1 to 7, comprises: Receives real-time monitoring data and environmental data from different power lines reported by the data acquisition device; For each power line, based on the monitoring data and environmental data of the power line within a preset time period, a power line inspection model is used to inspect the power line and obtain inspection results. The inspection results include whether a fault will occur and the type of fault if a fault occurs. The power line inspection model is a model that can predict and identify power line faults, obtained by optimizing and training a long short-term memory network using an improved population optimization algorithm. The inspection results of each circuit line are pushed to the relevant terminal equipment.

Images