Power transmission equipment state monitoring method based on internet of things

Abstract

The application discloses a power transmission equipment state monitoring method based on an Internet of Things, and relates to the technical field of power system detection, comprising the following steps: collecting state data and environmental parameters of power transmission equipment through an Internet of Things sensor network, and calculating risk indexes of each equipment; when the maximum value of the risk indexes of the equipment is greater than a preset risk threshold, marking the equipment as a risk equipment; injecting a characteristic electric signal at the risk equipment; collecting signal response values of downstream equipment after receiving the characteristic electric signal, and determining a next risk equipment according to the signal response values; determining a risk propagation path according to the position of the risk equipment, and collecting images and identifying defects along the risk propagation path; and outputting an equipment state monitoring report according to the defect identification result. The application solves the technical problem of low detection efficiency of existing power transmission equipment.

Description

Technical Field

[0001] This invention belongs to the field of power system detection and relates to Internet of Things (IoT) sensor technology, specifically an IoT-based method for monitoring the status of power transmission equipment. Background Technology

[0002] In existing methods for monitoring the condition of power transmission equipment, traditional approaches often rely on manual inspections or data collection from single sensors, resulting in low monitoring efficiency and limited coverage. Manual inspections are limited by labor costs and environmental factors, making it difficult to achieve real-time monitoring of large-scale power transmission networks, and they lack the ability to identify hidden defects. Single-sensor monitoring can only acquire local parameters of the equipment and cannot comprehensively reflect the correlation between the equipment's operating status and the environment.

[0003] Crucially, existing methods generally lack correlation analysis of equipment status: they cannot establish the status impact relationships between different devices, nor can they trace the propagation path of risks in the transmission network. For example, when a device malfunctions, correlation analysis cannot determine whether it will have a cascading impact on downstream devices, nor can changes in environmental parameters be used to assess the possibility of multiple devices experiencing simultaneous defects. This limitation means that monitoring results can only reflect the status of individual devices, making it difficult to control overall risks at the system level. This can easily lead to missed defects or delayed fault warnings, affecting the safe and stable operation of the transmission network. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a power transmission equipment status monitoring method based on the Internet of Things to solve the technical problem of low efficiency in existing power transmission equipment status detection methods.

[0005] To achieve the above objectives, a first aspect of the present invention provides a method for monitoring the status of power transmission equipment based on the Internet of Things, comprising:

[0006] The status data and environmental parameters of power transmission equipment are collected through an Internet of Things (IoT) sensor network to calculate the risk index of each piece of equipment.

[0007] When the maximum value of the risk index of a device is greater than the preset risk threshold, it is marked as a risky device.

[0008] A characteristic electrical signal is injected at the risk device; the characteristic electrical signal includes a power frequency fundamental wave and a preset characteristic frequency band;

[0009] Collect the signal response value of downstream equipment after receiving the characteristic electrical signal, and determine the next risky equipment based on the signal response value;

[0010] Determine the risk propagation path based on the location of the risky equipment, and perform image acquisition and defect identification along the risk propagation path;

[0011] Output equipment status monitoring reports based on defect identification results.

[0012] Furthermore, the status data of the power transmission equipment includes conductor temperature, three-phase current range, frequency domain energy of tower vibration, and fitting temperature; the environmental parameters include temperature T, humidity H, and wind speed W.

[0013] Furthermore, the formula for calculating the risk index ERI is: Among them, T max Indicates the maximum heat resistance temperature of the conductor, W crit The preset critical wind speed is represented by α, β, and γ, which represent the weighting coefficients of each component, determined by the entropy weighting method, and α+β+γ=1.

[0014] Furthermore, determining the next risky device based on the signal response value includes:

[0015] Calculate the response deviation ΔD of downstream device i based on the signal response value. i ;

[0016] When the response deviation ΔD i When the preset threshold is exceeded, downstream device i will be marked as the next risk device;

[0017] Inject a characteristic electrical signal at the next risk device;

[0018] Collect the signal response value of the downstream device of the next risk device after receiving the characteristic electrical signal, and determine the next risk device of the next risk device based on the signal response value;

[0019] Repeat the above steps until there are no downstream devices or the response deviation of all downstream devices is less than or equal to the preset threshold.

[0020] Furthermore, the formula for calculating the response deviation is: ; where A i A represents the signal amplitude of device i. i0 The reference amplitude of device i is represented by η, the phase deviation weighting coefficient is represented by φ. i φ represents the phase angle of device i. i0 The reference phase angle is represented by device i.

[0021] Furthermore, the defect identification process is as follows:

[0022] The acquired visible light image is input into the first model, which outputs the device type and location mask; the first model is built based on a deep learning algorithm and is used to identify the type of power transmission equipment.

[0023] The acquired visible light images, infrared thermal images, and ultraviolet corona discharge images are cropped and magnified based on the location mask to obtain image groups from different devices;

[0024] The device defect identification model is obtained by loading the pre-trained weights of the second model according to the device type; the second model is built based on a deep learning algorithm and is used to identify device defects, and the pre-trained weights of the second model include pre-trained weights obtained by training on historical device images of different device types.

[0025] Input images from different devices into the corresponding device defect recognition model, and output the device defect type.

[0026] Furthermore, the input data for the second model is a multi-channel image formed by fusing visible light images, infrared thermal imaging images, and ultraviolet corona discharge images, and the magnified image has the same size as the original visible light image.

[0027] Furthermore, the second model includes a multimodal feature extraction module, a cross-modal attention fusion module, a feature enhancement module, and a defect localization module; wherein,

[0028] The multimodal feature extraction module is used to extract features from the input image group to obtain high-dimensional feature maps corresponding to each modality;

[0029] The cross-modal attention fusion module is used to perform cross-modal information interaction and fusion on the feature maps of each modality, and output the fused comprehensive feature map;

[0030] The feature enhancement module is used to perform multi-scale feature enhancement and context information supplementation on the fused comprehensive feature map, and output an enhanced feature map;

[0031] The defect localization module is used to determine the defect type and locate the defect in the enhanced feature map, and output the defect type and the defect location mask of the device.

[0032] Furthermore, the multimodal feature extraction module includes a visible light branch, an infrared branch, and an ultraviolet branch. Each branch contains multiple convolutional layers, multiple residual blocks, and multiple dilated convolutional layers, and the connection method within the branch is as follows:

[0033] The output of the convolutional layer is connected to the input of the residual block, and the output of the residual block is connected to the input of the dilated convolutional layer after being reduced in dimensionality by a 1×1 convolutional layer.

[0034] The outputs of the three branches are processed by a convolutional layer to achieve channel alignment and size unification, and then the outputs are high-dimensional feature maps corresponding to each modality.

[0035] The cross-modal attention fusion module includes a cross-attention matrix, a channel attention module, a spatial attention module, and a feature reconstruction convolutional layer. The feature reconstruction convolutional layer comprises multiple convolutional layers, and the internal calculation process of the cross-modal attention fusion module is as follows:

[0036] The high-dimensional feature maps corresponding to each modality are used to calculate the cross-attention matrix in pairs to obtain the cross-fused modality guidance features. The modality guidance matrix and the high-dimensional feature map corresponding to each modality are added element-wise to obtain each enhanced single-modality feature. The three enhanced single-modality features are concatenated by channels to form a high-dimensional feature set. The high-dimensional feature set is then passed through the channel attention module and the spatial attention module in sequence to obtain the attention-enhanced features. Finally, the attention-enhanced features are input into the feature reorganization module to obtain the fused comprehensive feature map.

[0037] The feature enhancement module includes a multi-scale parallel dilated convolution group, a feature pyramid module, and a hybrid connection module. The multi-scale parallel dilated convolution group is calculated as follows: the fused comprehensive features are divided into three parallel branches, which are then passed through 3×3 dilated convolution layers with dilation rates of 1, 3, and 5, respectively, to obtain features with different receptive fields. The feature pyramid module is calculated as follows: the output features of the three branches are upsampled sequentially in ascending order of size to obtain a feature set fused with multi-scale context. Then, this feature set is passed through three parallel 1×1... The convolutional layer yields scale-aligned first feature map F1, second feature map F2, and third feature map F3. The calculation process of the hybrid connection module is as follows: F1, F2, and F3 are concatenated along the channel dimension and compressed through a 1×1 convolutional layer to obtain the intermediate feature Fmid. The fused comprehensive feature is then adjusted through a 1×1 convolutional layer and residually connected with the intermediate feature Fmid. Finally, the residually connected feature is input into the channel attention module to generate a weight vector. The weight vector is then multiplied channel-by-channel with the residually connected feature to output the enhanced feature map.

[0038] The defect localization module includes a classification head and a segmentation head. The classification head contains a fully connected layer and a Softmax function layer, and the segmentation head contains a transposed convolutional layer and a convolutional layer. The internal calculation process of the defect localization module is as follows: the enhanced feature map is global average pooled and then input into the classification head to output the defect type probability of the device. At the same time, the enhanced feature map is input into the segmentation head to output the location mask of the defect.

[0039] Furthermore, the training process of the second model includes:

[0040] Historical multimodal image data of different types of power transmission equipment were collected; the multimodal images include visible light images, infrared thermal imaging images, and ultraviolet corona discharge images.

[0041] Label the corresponding defect type and defect location mask in historical multimodal image data, and construct training datasets categorized by device type;

[0042] The second model parameters are initialized based on the second model architecture, and the training process is started for different types of devices:

[0043] Before training, the output dimension of the classification head in the defect localization module is adjusted by the preset mapping relationship between device type name and number of categories, and the training dataset of the corresponding device type is loaded.

[0044] During training, a multi-task loss function is used, taking the enhanced feature map as input and optimizing the model parameters through backpropagation until the loss function converges; the multi-task loss function is composed of a cross-entropy loss function and a Dice loss function weighted in a 1:1 ratio.

[0045] After training is completed for each type of device, the corresponding model parameters are saved as pre-training weights for each device type, resulting in a pre-training weight library for each device type.

[0046] Furthermore, the step of outputting a device status monitoring report based on the defect identification results includes...

[0047] Defect identification results are summarized by equipment type; the defect identification results include equipment number, location, defect type, and risk index;

[0048] The risk propagation path is presented using a topology diagram;

[0049] Input the status data and environmental parameters of each device along the risk propagation path into the artificial intelligence model, and output the defect correlation factors and handling suggestions of the device.

[0050] The defect identification results, risk propagation paths, defect-related factors, and handling suggestions are compiled into a standardized report to obtain the equipment condition monitoring report.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] Compared to traditional methods that rely on a single sensor or manual inspection, this invention first accurately locates the initial risk equipment by calculating the risk index, then tracks the risk propagation path by using the response deviation of characteristic electrical signals, and finally combines multimodal images and dual-model defect identification. This not only solves the problem of insufficient identification of hidden defects by traditional methods, but also achieves global control from single-point defects to system risks through risk propagation path analysis, greatly improving the accuracy and comprehensiveness of monitoring.

[0053] Meanwhile, this invention designs a modular defect identification model architecture for different equipment types. By dynamically adjusting the classification head parameters and loading dedicated pre-trained weights, it balances model versatility with equipment specificity, solving the problem of insufficient identification accuracy of existing general-purpose models on diverse power transmission equipment. Furthermore, the combination of data augmentation and multi-task loss functions further enhances the model's adaptability to complex environments, reducing misjudgments caused by insufficient data samples or environmental interference. Overall, the fusion of technologies—"IoT sensor monitoring + feature electrical signal localization + multimodal imaging + deep learning recognition"—significantly reduces labor costs and provides more reliable technical support for the efficient operation and maintenance of power transmission networks. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating the IoT-based power transmission equipment status monitoring method provided by the present invention.

[0056] Figure 2 A flowchart illustrating another IoT-based method for monitoring the status of power transmission equipment provided by this invention;

[0057] Figure 3 A flowchart illustrating another IoT-based method for monitoring the status of power transmission equipment provided by this invention;

[0058] Figure 4 This is a flowchart illustrating another IoT-based method for monitoring the status of power transmission equipment provided by the present invention. Detailed Implementation

[0059] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, to address the technical problems of insufficient identification of hidden defects and difficulty in achieving global control of system risks caused by relying on a single sensor or manual inspection in existing technologies, this application provides an Internet of Things-based method for monitoring the status of power transmission equipment. This method includes:

[0061] The status data and environmental parameters of power transmission equipment are collected through an Internet of Things (IoT) sensor network to calculate the risk index of each piece of equipment.

[0062] When the maximum value of the risk index of a device is greater than the preset risk threshold, it is marked as a risky device.

[0063] A characteristic electrical signal is injected at the risk device; the characteristic electrical signal includes a power frequency fundamental wave and a preset characteristic frequency band;

[0064] Collect the signal response value of downstream equipment after receiving the characteristic electrical signal, and determine the next risky equipment based on the signal response value;

[0065] Determine the risk propagation path based on the location of the risky equipment, and perform image acquisition and defect identification along the risk propagation path;

[0066] Output equipment status monitoring reports based on defect identification results.

[0067] Based on this, the present invention integrates IoT monitoring, characteristic electrical signal localization and defect identification technologies to achieve comprehensive monitoring from single-point defects to system risks, thereby improving the accuracy and efficiency of power transmission equipment status monitoring.

[0068] like Figure 1 As shown in the embodiments of this application, the method for monitoring the status of power transmission equipment based on the Internet of Things includes:

[0069] S1. Collect status data and environmental parameters of power transmission equipment through IoT sensor networks, and calculate the risk index of each device.

[0070] Among them, the Internet of Things (IoT) sensor network is a network system composed of various sensor nodes, data transmission modules, and gateways deployed in and around power transmission equipment; status data can include the operating parameters of the power transmission equipment (such as equipment temperature, operating current, vibration amplitude, etc.) and physical parameters (such as equipment appearance integrity, connection tightness, etc.); environmental parameters typically include the temperature, humidity, air pressure, precipitation, light intensity, etc. of the environment in which the power transmission equipment is located; the risk index is used to quantitatively assess the probability of power transmission equipment failure or defects, providing a quantitative basis for the subsequent identification of risky equipment.

[0071] In some implementations, the risk index can be calculated using a weighted comprehensive score based on the Analytic Hierarchy Process (AHP), which involves first determining the weights of each state data and environmental parameter, then standardizing each parameter and summing them by weight to obtain the risk index; or it can be calculated using an index based on the fuzzy comprehensive evaluation method, which involves constructing a fuzzy evaluation matrix and combining it with a membership function to determine the index value corresponding to the risk level.

[0072] For example, if conductor temperature, ambient humidity, and wind speed are key influencing factors for a transmission line, the weights of the three factors can be determined to be 0.5, 0.3, and 0.2 respectively using the analytic hierarchy process. After standardizing the ratio of conductor temperature to heat resistance limit, humidity percentage, and wind speed to critical wind speed, the risk index of the conductor can be obtained by summing the weights.

[0073] S2. When the maximum value of the risk index of the equipment is greater than the preset risk threshold, it is marked as a risky equipment and a characteristic electrical signal is injected at the risky equipment.

[0074] Among them, the characteristic electrical signal refers to the electrical signal with a specific identifier injected in order to locate the risk propagation path. It is used to determine whether there is an associated risk by the signal response of downstream equipment. It can usually include the power frequency fundamental wave, such as the 50Hz or 60Hz fundamental wave signal during normal operation of the power system, and the preset characteristic frequency band, such as high-frequency signals or modulated signals in a specific frequency range.

[0075] In some implementation methods, the risk threshold can be determined by statistical analysis of historical failure data, that is, by analyzing the risk index of equipment that has failed in the past before the failure, and taking the average value or 90th percentile value as the threshold; or it can be determined by expert review in combination with the actual working conditions on site, and dynamically adjusted by comprehensively considering factors such as the importance of the equipment and its years of operation.

[0076] It should be noted that the injection of characteristic electrical signals must ensure that it does not affect the normal operation of the power transmission equipment. The signal strength and frequency should be within the tolerance range of the equipment and be identifiable to avoid confusion with interference signals in the power grid.

[0077] For example, for a 220kV transmission line, according to historical data statistics, the risk index of equipment that has overheated is mostly above 0.7, so the preset risk threshold is set to 0.7; the injected characteristic electrical signal can be a 50Hz power frequency fundamental wave superimposed with a high frequency characteristic band of 10kHz-100kHz, which is convenient for downstream equipment to identify.

[0078] S3. Collect the signal response value of the downstream equipment after receiving the characteristic electrical signal, and determine the next risky equipment based on the signal response value.

[0079] Among them, the signal response value represents the performance parameters of the signal in terms of amplitude, phase, frequency, etc. after the downstream equipment receives the characteristic electrical signal. It is used to reflect the transmission or response status of the equipment to the characteristic electrical signal and to determine whether there is an abnormality in the equipment.

[0080] In some implementations, the next risky device can be determined by comparing the difference between the signal response value and the reference value. For example, the deviation rate between the response value and the reference value can be calculated. When the deviation rate exceeds a preset ratio (such as 10%), the downstream device is marked as the next risky device. Alternatively, a normal range for the response value can be set. When the response value exceeds the range, the corresponding downstream device is marked as the next risky device.

[0081] It should be noted that the division of downstream equipment should be determined according to the topology of the transmission network. For example, adjacent equipment can be determined as downstream equipment in sequence according to the line direction to ensure that the tracking logic of risk propagation path is clear.

[0082] For example, taking a transformer as the current risk device and its downstream device as the adjacent circuit breaker, after injecting a characteristic electrical signal, the amplitude and phase of the signal received by the circuit breaker are collected. If the amplitude deviates from the reference value by 15%, exceeding the preset 10% threshold, then the circuit breaker is marked as the next risk device.

[0083] S4. Determine the risk propagation path based on the location of the risky equipment, and perform image acquisition and defect identification along the risk propagation path.

[0084] Among them, the risk propagation path is used to reflect the path of risk spreading from the initial risk device to other devices in the power transmission network. Therefore, it is necessary to perform close-range, multi-angle image acquisition and defect identification on the devices along the path to improve the identification accuracy of problems such as appearance defects and local damage of the devices.

[0085] In some implementations, defect identification can be achieved by manually analyzing the collected images, with professionals judging whether the equipment has defects such as cracks, rust, or deformation based on the image features; alternatively, traditional machine vision algorithms (such as edge detection and template matching) can be used to process the images and identify possible defect areas.

[0086] For example, based on the location of the marked risk equipment (such as tower A, conductor segment B, and tower C), the risk propagation path is determined to be tower A → conductor segment B → tower C. Visible light images of the tower body of tower A, the surface of conductor segment B, and the hardware of tower C are collected along this path. The operation and maintenance personnel analyze the images to determine whether there are defects such as tower tilting, broken conductor strands, or loose hardware.

[0087] S5. Output an equipment status monitoring report based on the defect identification results.

[0088] The defect identification results may include information such as the risk equipment number, specific location, defect type (such as cracks, overheating, corrosion, etc.), and defect severity.

[0089] In some implementations, equipment status monitoring reports can be categorized and organized according to equipment type or risk level, clearly presenting the status of each risky equipment, a diagram of the risk propagation path, and the distribution characteristics of defects. The reports can also include a preliminary analysis of the causes of defects and corresponding maintenance recommendations, such as the priority of maintenance and the recommended handling measures.

[0090] It should be noted that the report output should adopt a standardized format to facilitate understanding and use by different departments or personnel. At the same time, it is necessary to ensure the accuracy and objectivity of the report content and generate it based on actual monitoring and identification results.

[0091] For example, in the equipment condition monitoring report, the defect identification results are summarized by equipment type such as poles, conductors, and fittings. The risk propagation path from pole A to pole C is displayed in the form of a topology map, indicating that pole A has foundation cracks (moderate severity) and conductor segment B has local corrosion (minor severity). It is recommended to prioritize foundation reinforcement treatment for pole A.

[0092] Based on the above technical solutions, the IoT-based power transmission equipment status monitoring method provided in this application achieves comprehensive perception of the power transmission equipment status and environment through an IoT sensor network, accurately pinpoints initial risk equipment by combining risk indices, tracks risk propagation paths using characteristic electrical signals to clarify the scope of risk spread, and then obtains specific defect information of the equipment through image acquisition and defect identification, ultimately generating a comprehensive monitoring report. This method integrates multiple existing technical means, expands the monitoring coverage and applicability, reduces reliance on single sensors or manual inspections, and realizes full-process management from risk identification to defect location, providing strong support for the safe operation and efficient maintenance of power transmission equipment.

[0093] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below:

[0094] S101. Collect status data and environmental parameters of power transmission equipment through an Internet of Things (IoT) sensor network.

[0095] The status data of power transmission equipment includes conductor temperature, three-phase current range, frequency domain energy of tower vibration, and fitting temperature; environmental parameters include temperature (T), humidity (H), and wind speed (W). The Internet of Things (IoT) sensor network can consist of temperature sensors, current sensors, vibration sensors, and environmental monitoring sensors deployed on conductors, towers, fittings, and other equipment to achieve real-time acquisition and transmission of various parameters.

[0096] In some implementations, the deployment density of sensors can be adjusted according to the importance of the equipment and the complexity of the environment. For critical power transmission nodes, the number of sensors can be increased to improve the accuracy and redundancy of data acquisition. The acquisition frequency can also be dynamically set, increasing the acquisition frequency to capture more details when the equipment is under high load or the environment changes significantly.

[0097] It should be noted that during the data collection process, the synchronization of sensor data must be ensured to avoid affecting the accuracy of subsequent risk index calculations due to excessive time differences in data collection from different sensors. At the same time, the collected data should be preliminarily filtered to remove obvious outliers, such as data that exceeds the reasonable range due to sensor malfunctions.

[0098] For example, in a section of high-voltage transmission line, temperature sensors are installed on each tower to monitor the temperature of the fittings, a temperature sensor and a current sensor are installed every 500 meters on the conductor, a vibration sensor is installed at the bottom of the tower, and an environmental sensor is installed at a monitoring station near the line to collect temperature, humidity, and wind speed data in real time, and the data is sent to the data processing center via a wireless transmission module.

[0099] S102. Determine the parameters and weighting coefficients in the risk index calculation formula.

[0100] The risk index calculation formula involves parameters including the actual conductor temperature T and the conductor's heat resistance limit temperature T. max Ambient humidity H, ambient wind speed W, preset critical wind speed W crit And the various weighting coefficients α, β, and γ. The conductor's maximum heat resistance temperature T. max The heat resistance limit can be determined based on the material and model of the conductor. Different models of conductors have their corresponding national or industry standards specifying the heat resistance limit; preset critical wind speed W crit It can be determined based on meteorological data of the area where the transmission line is located and the line design standards, and is usually the wind speed threshold that may affect the safe operation of the line.

[0101] In some implementations, the weighting coefficients α, β, and γ are determined using the entropy weighting method. The entropy weighting method is an objective weighting method based on the information entropy of the data itself, avoiding the subjective influence of human factors. Specifically, historical state data and historical environmental parameters over a certain period are first collected. After standardizing the data, the information entropy of each indicator is calculated. Then, the weighting coefficient of each indicator is calculated based on the information entropy. The smaller the information entropy, the greater the degree of variation of the indicator, the greater its impact on decision-making, and thus the larger the weighting coefficient.

[0102] It should be noted that, in determining T max and W critWhen calculating weight coefficients using the entropy weight method, it is necessary to refer to the latest standards and specifications to ensure the accuracy and timeliness of the parameters. If the amount of data is small or there are outliers, the weight coefficient calculation results may have large deviations. In this case, the weight coefficients can be adjusted appropriately based on expert experience.

[0103] For example, the heat resistance limit temperature T of a certain type of wire max The temperature is set at 70℃ according to national standards; based on meteorological statistics of the area where the line is located, the maximum possible wind speed in the area is 25 m / s. In conjunction with the line design requirements, the preset critical wind speed W is... crit The value was determined to be 20 m / s; 365 sets of data on conductor temperature, ambient humidity, and wind speed were collected over the past year, and α=0.4, β=0.2, and γ=0.4 were calculated using the entropy weight method.

[0104] S103. Calculate the risk index of each device based on the collected status data, environmental parameters, and determined parameters and weighting coefficients.

[0105] The formula for calculating the risk index ERI is as follows: This formula comprehensively reflects the risk status of the equipment by standardizing and weighting the various parameters, and the values ​​of each parameter have been normalized.

[0106] For example, a power transmission device collects data on the conductor temperature T = 50℃, the ambient humidity H = 60%, and the wind speed W = 10m / s. Given that T... max =70℃, W crit =20m / s, α=0.4, β=0.2, γ=0.4, substituting into the calculation formula, we get: ERI=0.4×(50 / 70)²+0.2×60 / 100+0.4×10 / 20≈0.4×0.51+0.2×0.6+0.4×0.5=0.204+0.12+0.2=0.524, that is, the risk index of the equipment is 0.524.

[0107] Based on the above technical solution, the conductor temperature term in the risk index calculation formula adopts a square form, which strengthens the nonlinear impact of high temperature on equipment—as the temperature approaches the heat resistance limit, its accelerating effect on equipment aging and insulation performance degradation shows an increasing trend, reflecting the physical characteristics of power transmission equipment; the humidity and wind speed terms are standardized by being compared with the benchmark value, reflecting both the degree to which environmental parameters deviate from the safe range and ensuring the comparability of different parameters on the same order of magnitude; and the weighting coefficients are determined by the entropy weighting method, objectively allocating weights based on the information entropy of the data itself, avoiding the bias of subjective experience, and making the contribution of each factor to risk more consistent with the statistical laws of actual operating data. The risk index provides an operable standard for risk classification through clear mathematical logic, making the marking of risky equipment have an objective basis, laying a precise initial positioning foundation for subsequent steps such as characteristic electrical signal injection and risk propagation path tracking, thereby improving the scientificity and reliability of the entire monitoring method.

[0108] In one possible implementation of this application embodiment, the above-mentioned S2 can be specifically implemented by the following S201 and S202, which are described in detail below:

[0109] S201. Set a preset risk threshold and compare the maximum risk index of each device with the threshold to mark the risky devices.

[0110] The preset risk threshold is a critical value used to determine whether a device has a risk. When the maximum risk index of a device exceeds this threshold, it indicates that the device has a high risk of failure or defect and needs to be marked as a risk device for further monitoring.

[0111] In some implementation methods, the determination of the preset risk threshold can be combined with historical fault data and equipment safety operation requirements. By statistically analyzing the risk index distribution of equipment that has experienced past faults before the fault, the 95th percentile value is taken as the initial threshold. Then, it is dynamically adjusted according to the equipment type (such as key transmission nodes, ordinary line sections) and the number of years of operation. For important equipment or equipment with a long operating period, the threshold can be appropriately reduced to improve monitoring sensitivity.

[0112] It should be noted that the preset risk threshold is not fixed and needs to be calibrated regularly based on new operating data and failure cases to ensure that it can effectively identify risky equipment while avoiding excessively high false positive rates due to excessively low thresholds, which would increase unnecessary subsequent monitoring costs.

[0113] For example, by analyzing the fault records of power transmission equipment in a certain region over 5 years, it was found that 95% of the faulty equipment had a risk index exceeding 0.65 before the fault. Therefore, the initial preset risk threshold was set to 0.65. For 220kV transmission line equipment that bears important loads, the threshold was adjusted to 0.6 to identify potential risks earlier.

[0114] S202. Inject a characteristic electrical signal at the marked risk device, the characteristic electrical signal including a power frequency fundamental wave and a preset characteristic frequency band.

[0115] Among them, the power frequency fundamental wave refers to the rated frequency fundamental wave when the power system is operating normally, such as 50Hz in my country. The preset characteristic frequency band refers to a specific frequency range set for easy identification, such as high-frequency signals of 1kHz-10kHz. The combination of the two can make the characteristic electrical signal not only conform to the operating characteristics of the power grid, but also have unique identifiability, making it easy for downstream equipment to distinguish the signal from other interference signals in the power grid.

[0116] In some implementations, the injection of characteristic electrical signals can be accomplished by connecting a dedicated signal generator to the terminal of the risky equipment. During injection, the amplitude of the signal must be controlled to not exceed 5% of the rated voltage of the equipment to avoid interference with the normal operation of the equipment. The selection of the preset characteristic frequency band should avoid the main harmonic frequencies in the power grid to reduce signal attenuation and distortion. After determining the harmonic distribution of the power grid through spectrum analysis, the frequency band with lower harmonic energy can be selected as the characteristic frequency band.

[0117] It should be noted that before injecting characteristic electrical signals, it is necessary to confirm that the risk equipment is in a energized operating state (except in special circumstances) to ensure that the signal can be transmitted normally to the downstream equipment along the transmission line; at the same time, the injection time and signal parameters need to be recorded to provide a reference benchmark for the subsequent signal response analysis of the downstream equipment.

[0118] For example, a characteristic electrical signal is injected into the fittings of a certain risky tower. The signal uses a 50Hz power frequency fundamental wave as the carrier and is superimposed with a preset characteristic frequency band of 2kHz-5kHz. The signal amplitude is 3% of the rated voltage of the equipment. The signal is connected to the monitoring interface of the fitting through the output terminal of the signal generator. The injection is stopped after 10 seconds, and the system waits for the signal response from the downstream equipment.

[0119] Based on the above technical solution, by using characteristic electrical signals as "marking probes", the compatibility between the power frequency fundamental wave and the power grid operation characteristics is used to ensure stable signal transmission. At the same time, by pre-setting characteristic frequency bands, the signals are given unique identification, making the response of downstream equipment a direct basis for judging risk transmission. This not only realizes the leap from single-point risk identification to chain-like risk tracking, but also accurately locks abnormal equipment on the risk diffusion path through the transmission characteristics of electrical signals, providing a brand-new technical path for global risk control of power transmission networks.

[0120] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below:

[0121] S301. Acquire the signal response value after the downstream device receives the characteristic electrical signal, wherein the signal response value includes the signal amplitude and phase angle.

[0122] Among them, the signal amplitude A i This refers to the strength and phase angle φ of the characteristic electrical signal received by downstream device i. i This refers to the phase shift of a signal during transmission, and together they reflect the changes in the signal's transmission characteristics after passing through downstream equipment. Data acquisition can be achieved through signal detection sensors (such as current transformers and voltage sensors) deployed on downstream equipment, acquiring the signal's amplitude and phase data in real time and uploading it to the processing system.

[0123] In some implementations, to improve data accuracy, multiple sets of sensors can be used to collect data simultaneously and average it, or filtering algorithms can be used to remove noise interference from the signal. For the same downstream device, signal response values ​​can be collected multiple times at different time points to reduce the impact of instantaneous fluctuations.

[0124] It should be noted that before data acquisition, it is necessary to ensure that the frequency range of the signal detection sensor matches that of the characteristic electrical signal to avoid signal distortion due to insufficient sensor bandwidth. At the same time, environmental parameters (such as temperature and humidity) during data acquisition should be recorded to facilitate subsequent analysis of the impact of the environment on signal transmission.

[0125] For example, a signal detection sensor is installed on tower B, 1 kilometer downstream of a certain risky device (tower A). When a characteristic electrical signal is injected into tower A, the sensor collects the signal in real time with an amplitude of 5V and a phase angle of 30°, and uploads the data to the monitoring center.

[0126] S302. Calculate the response deviation ΔD of the downstream equipment based on the signal response value. i When ΔD i When the threshold is exceeded, the device is marked as the next risky device.

[0127] The formula for calculating the response deviation ΔDi is: In the formula A i0 φ is the reference amplitude of device i, that is, the amplitude of the characteristic electrical signal received when the device is operating normally. i0 Let be the reference phase angle of device i, i.e., the phase angle of the characteristic electrical signal received during normal operation of the device. η is the phase deviation weighting coefficient, used to balance the influence of amplitude deviation and phase deviation. The preset threshold is the critical value for judging whether the device has a risk. When ΔD i If this value is exceeded, it indicates that the device is transmitting signals abnormally and may have a defect.

[0128] In some implementations, the reference amplitude A i0 and reference phase angle φ i0The phase deviation weighting coefficient η can be determined by averaging multiple measurements taken under normal operating conditions. The coefficient can be adjusted according to the equipment type; for equipment sensitive to phase changes (such as transformers), η can be 0.8-1.0, while for equipment more sensitive to amplitude changes (such as circuit breakers), η can be 0.3-0.5. The preset threshold can be determined by statistically analyzing the ΔD values ​​of historical faulty equipment. i The distribution is determined by using the 90th percentile as a threshold value; for example, by analyzing the ΔD of 100 faulty devices. i If its 90th percentile value is 0.25, then the preset threshold is set to 0.25.

[0129] It should be noted that calculating ΔD i It is necessary to ensure A i With A i0 φ i With φ i0 The measurement conditions must be consistent, including the same injected signal parameters and environmental conditions; otherwise, the deviation calculation will be distorted. If the equipment has no historical benchmark data, the signal response value of a health device of the same model and the same number of years of operation can be selected as the benchmark value.

[0130] For example, the reference amplitude A of a downstream device i i0 =4V, reference phase angle φ i0 =25°, η=0.5, the collected A i =5V, φ i =35°, then ΔD i =√[(5-4) / 4)²+0.5×|35-25|]=√[0.0625+0.5×10]=√5.0625≈2.25. If the preset threshold is 1.5, then the device is marked as the next risk device.

[0131] S303. Repeatedly inject the characteristic electrical signal at the marked next risk device, collect the signal response value of its downstream device and determine the new next risk device, until there is no downstream device or the response deviation of all downstream devices is less than or equal to the preset threshold.

[0132] The purpose of repeatedly injecting characteristic electrical signals is to trace the propagation path of risks in the transmission network. By detecting the signal responses of downstream equipment step by step, the scope and direction of risk spread can be clarified. The definition of downstream equipment is based on the transmission network topology. For example, towers, transformers, circuit breakers, etc., connected sequentially along the line are all considered downstream equipment of the current equipment.

[0133] In some implementations, the upstream and downstream relationships of each device can be pre-determined using a GIS (Geographic Information System) map to form a device connection topology map. When a risk device is marked, the system automatically identifies its downstream device list and starts the data collection task. If there are many downstream devices, signals can be injected in batches, prioritizing the detection of the downstream devices closest to the current risk device.

[0134] It should be noted that the characteristic electrical signal parameters (such as the power frequency fundamental frequency and the preset characteristic frequency band) of the repeated injection must be consistent with the initial injection signal to ensure the comparability of the signal response values. When the response deviation of a downstream device is less than or equal to the threshold for multiple consecutive measurements, the detection of the downstream device can be stopped to reduce unnecessary operations.

[0135] For example, tower B has been marked as the next risky device, and its downstream devices include towers C and D. The same characteristic electrical signal as that of tower A is injected at tower B. The ΔDi of tower C is measured to be 0.3 (less than the threshold of 0.25), and the ΔD of tower D is measured to be... i =0.3 (greater than the threshold of 0.25), then tower D is marked as the next risk device; continue to inject signals at tower D, there are no other devices downstream, so the detection stops, and the risk propagation path is tower A → tower B → tower D.

[0136] Based on the above technical solution, by collecting signal response values, calculating response deviation, and dynamically tracking downstream equipment, the risk propagation path is located step by step. This not only ensures the accuracy of identifying the next risky equipment, but also covers the entire potential risk range through a repeated detection mechanism, providing a clear target path for subsequent inspections and defect identification, and improving the systematicness and pertinence of power transmission network status monitoring.

[0137] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the above S4 specifically includes the following S401 to S403:

[0138] S401. Determine the risk propagation path based on the location of the risky equipment.

[0139] The precise coordinates of the risky equipment can be obtained through positioning modules (such as GPS or BeiDou positioning) in the IoT sensor network, and the topology information of the power transmission network (such as equipment connection relationships and line routes) is pre-stored in the system. The determination of the risk propagation path is based on the location of the marked risky equipment and the next risky equipment, forming a chain path by connecting them sequentially according to their upstream and downstream connections in the power transmission network. This path is used to visually reflect the trajectory of the risk spreading from the initial equipment to other equipment.

[0140] In some implementations, the location coordinates of risky devices can be overlaid with the power transmission network topology map using a Geographic Information System (GIS) to automatically generate a visualized risk propagation path. The path is marked with the number and risk index of each risky device. For power transmission networks with branches, if multiple downstream devices are marked as risky devices, multiple parallel risk propagation paths can be generated and sorted according to the sum of risk indices, with high-risk paths being inspected first.

[0141] It should be noted that when determining the risk propagation path, the accuracy of the equipment location coordinates must be ensured. If the positioning signal of a certain equipment is weak, its coordinates can be estimated by combining the location of adjacent equipment and the line length. At the same time, the path must cover all marked risk equipment to avoid missing key nodes.

[0142] For example, the initial risk device is pole A (coordinates: X1, Y1), and the next risk devices are pole B (coordinates: X2, Y2) and pole C (coordinates: X3, Y3). Pole B and pole C are downstream branches of pole A. After overlaying through the GIS system, two risk propagation paths are generated: pole A → pole B and pole A → pole C. The risk index of pole A is marked as 0.8, the risk index of pole B as 0.7, and the risk index of pole C as 0.6 on the path map.

[0143] S402. Acquire images along the risk propagation path to obtain multimodal images and process them into image groups from different devices.

[0144] The multimodal images can be acquired using visible light cameras, infrared thermal imaging cameras, and ultraviolet corona discharge cameras mounted on drones or robots. Visible light images reflect the equipment's appearance, infrared thermal imaging images reflect the equipment's temperature distribution and can identify overheating defects, and ultraviolet corona discharge images reflect the equipment's partial discharge and can identify insulation defects. Image processing includes cropping the original image to preserve the equipment area based on the position mask output by the first model, and then enlarging the cropped image to improve detail clarity. The final result is a combination of visible light, infrared, and ultraviolet images for each device, i.e., an image set.

[0145] It should be noted that when cropping and enlarging images, spatial alignment of each modality must be maintained to ensure that visible light, infrared, and ultraviolet images from the same device correspond to the same area.

[0146] For example, along the path "tower A → tower B", a visible light image (capturing the tower body and fittings), an infrared thermal image (detecting the temperature of the fittings), and an ultraviolet corona discharge image (detecting whether there is partial discharge) are acquired at tower A. The images are input into the first model, and the position mask of tower A is output (delineating the area of ​​the tower body and fittings in the image). The image of the fittings area is cropped according to the mask and enlarged to 512×512 pixels to obtain the image group of tower A.

[0147] S403: Defect identification is performed through dual-model collaboration, and the defect type and location mask of the output device are output.

[0148] Defect identification relies on a first model and a second model: The first model is built based on a deep learning algorithm (such as the improved YOLOv5), takes a visible light image as input, and outputs a device type and a location mask; the device type includes, but is not limited to, poles, wires, and hardware, and the location mask is used to mark the pixel region of the device in the image; the second model is also built based on a deep learning algorithm, takes a multimodal image group of the device as input (fused into a multi-channel image), and outputs a defect type and a defect location mask; the defect type includes, but is not limited to, cracks, overheating, and corona discharge, and the defect location mask is used to mark the specific location of the defect in the device.

[0149] In one embodiment provided in this application, the model structure may be as follows:

[0150] The first model consists of a backbone network (such as CSPDarknet), a neck network (such as PANet), and a head network. The head network outputs the bounding box coordinates of the device class probability and the location mask, and the optimal prediction results are selected through non-maximum suppression.

[0151] The second model includes a multimodal feature extraction module, a cross-modal attention fusion module, a feature enhancement module, and a defect localization module. Among them:

[0152] The multimodal feature extraction module is responsible for extracting features from the input visible light, infrared thermal imaging, and ultraviolet corona discharge image sets to obtain high-dimensional feature maps corresponding to each mode.

[0153] This module comprises three independent branches: visible light, infrared, and ultraviolet. Each branch consists of multiple convolutional layers, residual blocks, and dilated convolutional layers. Internally, each branch employs a structure where the output of a convolutional layer is connected to the input of a residual block, and the output of the residual block, after 1×1 convolutional dimensionality reduction, is then connected to the input of a dilated convolutional layer. This approach extracts modality-specific features (such as texture details in visible light, temperature distribution in infrared, and discharge intensity in ultraviolet) through convolutional layers and residual blocks, while also enabling feature reuse within the branch through skip connections, preventing the loss of deep features. Furthermore, the dilated convolutional layers in different branches use differentiated dilation rates to adapt to the characteristics of each modality's data. For example, the infrared branch, which focuses on capturing a wide temperature field, uses a larger dilation rate, laying a high-quality feature foundation for subsequent fusion. The outputs of the three branches are aligned and sized using convolutional layer channels to form a standardized high-dimensional feature map, ensuring the feasibility of cross-modal fusion.

[0154] Specifically, the calculation process within each branch can be described as follows:

[0155] (1) Convolutional layer operation: For the input feature map After convolution kernel and bias The convolution operation outputs a feature map. The formula is: Where k is the kernel size, H' and W' are the output feature map sizes, and C in H represents the number of channels in the input feature map, H and W represent the length and width of the input feature map, and C represents the number of channels in the input feature map. out This indicates the number of channels in the output feature map.

[0156] (2) Residual block operation: The residual block output F(X) satisfies ,in This is a combination operation of convolutional layers within residual blocks, which alleviates the gradient vanishing problem in deep networks through skip connections.

[0157] (3) Dilated convolutional layer operation: For dilated convolution with dilation rate d, the effective kernel size is k+(k-1)(d-1), and the formula for calculating the output feature map Y is: d is the dilation rate of the dilated convolution, which expands the receptive field by introducing interval sampling.

[0158] The cross-modal attention fusion module is used to perform cross-modal information interaction and fusion on feature maps of various modalities, and output a fused comprehensive feature map. Internally, it first calculates the cross-attention matrix between each modality to generate modality-guided features, and then adds the guided features to the original features element-wise to obtain enhanced single-modal features. Subsequently, the three enhanced feature channels are concatenated, and then passed through the channel attention module to strengthen key modal features and the spatial attention module to focus on key areas of the device. Finally, the features are compressed in dimensionality by the feature recombination convolutional layer.

[0159] This module achieves intermodal information complementarity through bidirectional attention interaction, allowing the fused features to retain the advantages of each modality while highlighting key information related to the defect, thereby improving the ability to represent complex defects.

[0160] The pairwise calculation of the cross-attention matrix refers to the process of calculating the cross-attention matrix pairwise for the high-dimensional feature maps corresponding to each mode output by the multimodal feature extraction module. This includes calculating the cross-attention matrix between the high-dimensional feature maps of the visible light mode and the infrared mode, the cross-attention matrix between the high-dimensional feature maps of the visible light mode and the ultraviolet mode, and the cross-attention matrix between the high-dimensional feature maps of the infrared mode and the ultraviolet mode, thereby achieving information interaction and guidance between different modal features. In this way, the texture structure information of the visible light mode can guide the infrared mode to focus on the spatial location of temperature anomaly areas, the temperature distribution information of the infrared mode can assist the visible light mode in strengthening its focus on overheat-related appearance defects, and the corona discharge information of the ultraviolet mode can be combined with the visible light and infrared modes to accurately locate the specific location of insulation defects. Thus, each modal feature retains its unique advantages while absorbing effective information from other modes, improving the comprehensive characterization capability of complex defects in power transmission equipment.

[0161] Specifically, taking the visible light mode and infrared mode as examples, the calculation process of the cross-modal attention fusion module can be described as follows:

[0162] (1) Cross-attention matrix calculation: For visible light modal features and infrared modal characteristics Cross-attention matrix The formula for calculation is: , where p and q are the spatial location indices of the feature map, and the similarity is calculated and normalized by scaling the dot product.

[0163] (2) Modality-guided feature generation: Infrared-guided visible light features based on attention matrix. pass Each enhanced modal feature is calculated to achieve intermodal knowledge transfer.

[0164] (3) Channel attention weighting: The three enhanced single-modal features are concatenated by channels to form a high-dimensional feature set. Then perform channel attention weighting. Calculate and obtain the channel attention-weighted feature F. cam : Where GAP stands for global average pooling, FC stands for fully connected layer, and σ stands for the sigmoid function. This is for element-wise multiplication.

[0165] (4) Spatial attention weighting: Then, according to the formula Calculate spatial attention weights Then through The fused integrated feature map F is obtained sam GMP stands for Global Max Pooling, and Conv stands for Convolutional Layer.

[0166] The feature enhancement module is used to perform multi-scale feature enhancement and contextual information supplementation on the fused comprehensive feature map, and output an enhanced feature map.

[0167] This module divides the comprehensive features into three parallel branches. It obtains different receptive field features through 3×3 dilated convolutions with dilation rates of 1, 3, and 5 to adapt to defects of different sizes, such as small cracks and large-area corrosion. Then, the feature pyramid module upsamples the branch outputs by size and adds them element by element to fuse multi-scale contextual information. Finally, the hybrid connection module connects the fused features with the residuals of the original comprehensive features and obtains the enhanced feature map through channel attention weighting.

[0168] By using multi-scale parallelism and cross-scale fusion, the problem of single-scale features being sensitive to defect size is effectively compensated. At the same time, residual connections avoid feature degradation, allowing the enhanced features to more comprehensively cover the local details and global context of defects, thus providing support for accurate identification.

[0169] Specifically, the calculation process of the feature enhancement module can be described as follows:

[0170] (1) Multi-scale dilated convolution: for input features The algorithm is divided into three parallel branches, each passing through a 3×3 dilated convolutional layer with dilation rates of 1, 3, and 5, respectively, to obtain features of different receptive fields. , , Where AtrousConv represents dilated convolution;

[0171] (2) Feature pyramid fusion: Bilinear interpolation is used for upsampling. Small-sized features are upsampled sequentially to the size of large-sized features and then summed, i.e., F up1 =F sam2 +Upsample(F sam1 size=F sam2 .size), F up2 =F sam3 +Upsample(F sam2 size=F sam3 .size), F up3 =F up2 +Upsample(F up1 size=F sam3.size) enables cross-scale feature complementarity;

[0172] Then the feature set of the multi-scale context = {F up2 F up3 F sam3 The first feature map F1, the second feature map F2, and the third feature map F3 are obtained by passing three parallel 1×1 convolutional layers.

[0173] (3) Hybrid connection: F1, F2 and F3 are spliced ​​together according to the channel dimension to obtain F cat The intermediate feature F is obtained by channel compression through a 1×1 convolutional layer. mid The integrated features F after fusion sam Channel adjustment is performed using a 1×1 convolutional layer, and then combined with the intermediate feature F. mid Residual connections are performed, and finally, the residual-connected features are input into the channel attention module to generate a weight vector. The weight vector is then multiplied channel by channel with the residual-connected features to output an enhanced feature map.

[0174] The defect localization module is used to determine the defect type and location on the enhanced feature map, outputting the defect type and the defect location mask. This module includes a classification head (fully connected layer + Softmax function) and a segmentation head (transposed convolutional layer + convolutional layer).

[0175] The classification head outputs the defect type probability through the feature vector after global average pooling, while the segmentation head outputs a pixel-level defect location mask by restoring the feature map size through transposed convolution. Simultaneously, a feature interaction gate is designed between the 256-dimensional vector of the classification branch and the 128-channel feature map of the segmentation branch—expanding the classification vector into a 128-channel feature map through a 1×1 convolution, multiplying it point-by-point with the segmentation branch features, and then feeding it into the final convolutional layer. This allows the classification semantics to guide the segmentation and localization, ensuring both the accuracy of defect type judgment and precise defect location labeling. This solves the misjudgment problem caused by the separation of type and location in traditional methods, and improves the overall accuracy of defect recognition.

[0176] In one embodiment provided in this application, the training process of the first model may include:

[0177] Collect historical visible light image data of different types of power transmission equipment. These images should cover the equipment status under different operating conditions and environmental conditions, such as towers, conductors, and fittings under different seasons, weather, and light intensities.

[0178] The collected historical image data is labeled, including equipment type labels (such as towers, conductors, and fittings) and corresponding location masks (marking the specific pixel area of ​​the equipment in the image), to construct the training dataset for the first model;

[0179] Based on deep learning algorithms, such as the YOLO series and Faster R-CNN, the initial architecture of the first model is constructed. This architecture includes a backbone network for feature extraction, a neck network for feature fusion, and a head network for output device type and location masking.

[0180] Before training, data augmentation processing is performed on the training dataset, such as random cropping, flipping, rotating, and adding noise, to improve the model's generalization ability and adaptability to different scenarios.

[0181] The processed training dataset is input into the first model, and a loss function is defined, such as bounding box regression loss or classification loss. The model parameters are optimized through the backpropagation algorithm. During the training process, the model performance is evaluated periodically using the validation set. When the model's accuracy on the validation set no longer improves or reaches a preset threshold, training is stopped, and the model parameters at this time are saved as the final parameters of the first model.

[0182] In one embodiment provided in this application, the training process of the second model may include:

[0183] Collect multiple sets of historical multimodal image data of different types of power transmission equipment. Each set of multimodal image data includes visible light images, infrared thermal imaging images, and ultraviolet corona discharge images.

[0184] Label the corresponding defect type and defect location mask in historical multimodal image data, and construct training datasets categorized by device type;

[0185] The second model parameters are initialized based on the second model architecture, and the training process is started for different types of devices:

[0186] Before training, the output dimension of the classification head in the defect localization module is adjusted by the preset mapping relationship between device type name and number of categories, and the training dataset of the corresponding device type is loaded.

[0187] During training, a multi-task loss function is used, taking the enhanced feature map as input and optimizing the model parameters through backpropagation until the loss function converges. The multi-task loss function is composed of a cross-entropy loss function and a Dice loss function weighted in a 1:1 ratio.

[0188] After training is completed for each type of device, the corresponding model parameters are saved as pre-training weights for each device type, resulting in a pre-training weight library for each device type.

[0189] In some implementations, the position mask of the first model can help the second model focus on the device area and reduce background interference. When fusing the input images of the second model, the influence of brightness differences can be eliminated through normalization processing (scaling the pixel values ​​of each modality image to 0 and 1). In terms of preset parameters, the dilation rate of the dilated convolution of the second model can be adjusted according to the device size, such as using a dilation rate of 1 for small-sized fittings and a dilation rate of 3 for large-sized towers.

[0190] For example, the image group of tower B (fused into a 3-channel image) is input into the second model, and the pre-trained weights of tower-type equipment are loaded; the multimodal feature extraction module outputs high-dimensional features of visible light, infrared and ultraviolet light; after cross-modal fusion, the feature enhancement module strengthens the abnormal features of the fitting area; the defect localization module outputs "fitting overheating" (defect type) and the location mask of the overheating area.

[0191] Based on the above technical solutions, by determining the risk propagation path, the inspection can be ensured to be clearly targeted, avoiding the inefficiency caused by blind inspections. Multimodal image acquisition includes the multi-dimensional status of the equipment, with visible light images reflecting appearance, infrared thermal imaging images showing temperature distribution, and ultraviolet corona discharge images showing local discharge, compensating for the lack of information from a single image modality. Furthermore, the dual-model collaboration mechanism demonstrates significant advantages: the first model first completes equipment type identification and outputs a location mask, based on which the pre-trained weights of the corresponding equipment type in the second model are called. This enables defect identification of different equipment using a unified model structure, ensuring the universality and convenience of model application through a unified architecture, while fully adapting to the characteristics of different equipment through pre-trained weights for each device, improving identification accuracy in complex scenarios. Ultimately, the dual-model collaboration achieves accurate defect identification and location, providing a reliable basis for subsequent equipment status assessment.

[0192] In one possible implementation of this application embodiment, the above-mentioned S5 specifically includes the following S501 to S503:

[0193] S501. Summarize the defect identification results by equipment type and clarify the key information of each equipment.

[0194] The defect identification results include equipment number, location (latitude and longitude coordinates of the equipment or line segment number), defect type, and risk index. The results are summarized by equipment type, categorized according to types such as poles, conductors, and fittings, facilitating rapid location of defect distribution across different equipment types.

[0195] In some implementations, the aggregation process can be achieved through database queries, storing the defect identification results in a structured data table, and grouping and statistically analyzing them using the "equipment type" field to generate a statistical table containing the number of defects and average risk index for each type of equipment; for cases where the same equipment has multiple defects, they are sorted from high to low risk index to highlight the main risks.

[0196] For example, the summary results show that: 3 pole and tower equipment have defects, of which pole and tower G1023 has a defect type of "foundation crack" with a risk index of 0.78; 2 conductor equipment have defects, of which conductor L205 has a defect type of "local overheating" with a risk index of 0.65. A preliminary summary table is formed according to this classification.

[0197] S502. Present the risk propagation path with a topology diagram and analyze the defect correlation factors and handling suggestions through artificial intelligence models.

[0198] The topology map is drawn based on the physical connections of the power transmission network. Nodes represent risky equipment, arrows indicate the direction of risk propagation, and equipment numbers and risk indices are labeled next to nodes, visually demonstrating the trajectory of risk spread from the initial equipment to downstream equipment. The artificial intelligence model is an existing large language model that can analyze input data and instructions, combined with professional knowledge of power transmission equipment operation and maintenance and historical cases, to generate analysis results that conform to the actual scenario. For example, after inputting the status data and environmental parameters of each device along the risk propagation path, the model is guided to focus on the correlation between equipment defects and status data and environmental parameters, analyze the possible causes of defects, and output targeted handling suggestions. For example, the output reasons and suggestions are: "Abnormal rise in conductor temperature may be related to excessively high ambient temperature and three-phase current imbalance," and "It is recommended to adjust the three-phase load to reduce the current range and simultaneously perform heat dissipation treatment on the conductors."

[0199] In some implementations, the topology diagram can be automatically generated using visualization tools (such as the NetworkX library in MATLAB or Python), supporting interactive operations such as zooming and clicking to view device details.

[0200] It should be noted that the topology map of the risk propagation path must be consistent with the actual route to avoid analytical bias caused by incorrect topological relationships; the input data of the artificial intelligence model must be standardized (such as normalizing parameters such as temperature and humidity to the range of 0-1) to ensure the stability of the model output.

[0201] For example, in the topology diagram, the initial risk device "tower G1023" points to the downstream risk devices "conductor L205" and "tower G1024" via arrows, with risk indices of 0.78, 0.65, and 0.62 marked next to the nodes, respectively. The conductor temperatures (55℃, 62℃), ambient temperatures (38℃), and three-phase current ranges (8A, 12A) of these devices are input into the model, and the output correlation factor "overheating of conductor L205 is significantly correlated with ambient temperature 38℃ and current range of 12A" is provided, with the handling suggestion "prioritize the repair of the joints of conductor L205 and adjust the three-phase current to a range ≤5A".

[0202] S503. Organize various types of information into standardized reports and output equipment status monitoring reports.

[0203] The standardized report includes a cover (report number, generation time, monitoring scope), table of contents, summary table of defect identification results, risk propagation path topology diagram, defect correlation factor analysis, list of handling suggestions, and appendices (equipment parameter description, model parameter description, etc.).

[0204] For example, in the final output of the equipment condition monitoring report, the summary table lists the defect details by tower and conductor, the topology diagram clearly marks the direction of risk propagation, the analysis section points out that "the high temperature environment exacerbates the spread of conductor overheating defects", and the recommendation list is sorted by priority, among which "tower G1023 foundation crack repair" and "conductor L205 joint treatment" are listed as the highest priority.

[0205] Based on the above technical solution, by summarizing defect information by type, visualizing risk propagation paths, and combining artificial intelligence to analyze related factors and generate standardized reports, the monitoring results are presented systematically and applied efficiently. This process ensures the integrity and accuracy of the information, and provides a clear basis for the operation and maintenance decisions of power transmission equipment through structured reports, thereby improving the efficiency of the process from monitoring to handling.

[0206] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0207] Working principle of the invention:

[0208] By collecting status data of power transmission equipment and relevant parameters of its environment through an Internet of Things (IoT) sensor network, the risk level of each piece of equipment is assessed, and potentially risky equipment is identified. For these risky devices, electrical signals containing the fundamental power frequency and specific characteristic frequency bands are injected. By collecting the responses of downstream equipment after receiving the signals, the propagation path of the risk in the power transmission network is traced and determined. Subsequently, multimodal images of the equipment are collected along this path, and the images are processed and analyzed using a deep learning model to identify the types and locations of defects in the equipment. Finally, based on the defect identification results, relevant information is summarized, the risk propagation path is presented, the associated factors of the defects are analyzed, and treatment suggestions are proposed to form a standardized equipment status monitoring report, achieving comprehensive and accurate monitoring of the status of power transmission equipment.

[0209] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for monitoring the status of power transmission equipment based on the Internet of Things, characterized in that, include: The status data and environmental parameters of power transmission equipment are collected through an Internet of Things (IoT) sensor network to calculate the risk index of each piece of equipment. When the maximum value of the risk index of a device is greater than the preset risk threshold, it is marked as a risky device. Inject a characteristic electrical signal at the risk device; The characteristic electrical signal includes a power frequency fundamental wave and a preset characteristic frequency band; Collect the signal response value of downstream equipment after receiving the characteristic electrical signal, and determine the next risky equipment based on the signal response value; Determine the risk propagation path based on the location of the risky equipment, and perform image acquisition and defect identification along the risk propagation path; Based on the defect identification results, an equipment status monitoring report is output; among which... The defect identification process is as follows: The acquired visible light image is input into the first model, which outputs the device type and location mask; the first model is built based on a deep learning algorithm and is used to identify the type of power transmission equipment. The acquired visible light images, infrared thermal images, and ultraviolet corona discharge images are cropped and magnified based on the location mask to obtain image groups from different devices; The device defect identification model is obtained by loading the pre-trained weights of the second model according to the device type; the second model is built based on a deep learning algorithm and is used to identify device defects, and the pre-trained weights of the second model include pre-trained weights obtained by training on historical device images of different device types. Input images from different devices into the corresponding device defect recognition model, and output the device defect type.

2. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The formula for calculating the risk index ERI is: Among them, T max Indicates the maximum heat resistance temperature of the conductor, W crit The preset critical wind speed is represented by α, β, and γ, which are weighting coefficients determined by the entropy weighting method, and α+β+γ=1; T represents temperature, H represents humidity, and W represents wind speed.

3. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The step of determining the next risky device based on the signal response value includes: Calculate the response deviation ΔD of downstream device i based on the signal response value. i ; When the response deviation ΔD i When the preset threshold is exceeded, downstream device i will be marked as the next risk device; Inject a characteristic electrical signal at the next risk device; Collect the signal response value of the downstream device of the next risk device after receiving the characteristic electrical signal, and determine the next risk device of the next risk device based on the signal response value; Repeat the above steps until there are no downstream devices or the response deviation of all downstream devices is less than or equal to the preset threshold.

4. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 3, characterized in that, The formula for calculating the response deviation is: ; where A i A represents the signal amplitude of device i. i0 The reference amplitude of device i is represented by η, the phase deviation weighting coefficient is represented by φ. i φ represents the phase angle of device i. i0 The reference phase angle is represented by device i.

5. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The input data for the second model is a multi-channel image formed by fusing visible light images, infrared thermal imaging images, and ultraviolet corona discharge images, and the magnified image has the same size as the original visible light image.

6. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The second model includes a multimodal feature extraction module, a cross-modal attention fusion module, a feature enhancement module, and a defect localization module; wherein, The multimodal feature extraction module is used to extract features from the input image group to obtain high-dimensional feature maps corresponding to each modality; The cross-modal attention fusion module is used to perform cross-modal information interaction and fusion on the feature maps of each modality, and output the fused comprehensive feature map; The feature enhancement module is used to perform multi-scale feature enhancement and context information supplementation on the fused comprehensive feature map, and output an enhanced feature map; The defect localization module is used to determine the defect type and locate the defect in the enhanced feature map, and output the defect type and the defect location mask of the device.

7. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 6, characterized in that, The multimodal feature extraction module includes visible light branch, infrared branch, and ultraviolet branch. Each branch contains multiple convolutional layers, multiple residual blocks, and multiple dilated convolutional layers, and the connection method within the branch is as follows: The output of the convolutional layer is connected to the input of the residual block, and the output of the residual block is connected to the input of the dilated convolutional layer after being reduced in dimensionality by a 1×1 convolutional layer. The outputs of the three branches are processed by a convolutional layer to achieve channel alignment and size unification, and then the outputs are high-dimensional feature maps corresponding to each modality. The cross-modal attention fusion module includes a cross-attention matrix, a channel attention module, a spatial attention module, and a feature reconstruction convolutional layer. The feature reconstruction convolutional layer comprises multiple convolutional layers, and the internal calculation process of the cross-modal attention fusion module is as follows: The high-dimensional feature maps corresponding to each modality are used to calculate the cross-attention matrix in pairs to obtain the cross-fused modality guidance features. The modality guidance matrix and the high-dimensional feature map corresponding to each modality are added element-wise to obtain each enhanced single-modality feature. The three enhanced single-modality features are concatenated by channels to form a high-dimensional feature set. The high-dimensional feature set is then passed through the channel attention module and the spatial attention module in sequence to obtain the attention-enhanced features. Finally, the attention-enhanced features are input into the feature reorganization module to obtain the fused comprehensive feature map. The feature enhancement module includes a multi-scale parallel dilated convolution group, a feature pyramid module, and a hybrid connection module. The multi-scale parallel dilated convolution group is calculated as follows: the fused comprehensive features are divided into three parallel branches, which are then passed through 3×3 dilated convolution layers with dilation rates of 1, 3, and 5, respectively, to obtain features with different receptive fields. The feature pyramid module is calculated as follows: the output features of the three branches are upsampled sequentially in ascending order of size to obtain a feature set fused with multi-scale context. Then, this feature set is passed through three parallel 1×1... The convolutional layer yields scale-aligned first feature map F1, second feature map F2, and third feature map F3. The calculation process of the hybrid connection module is as follows: F1, F2, and F3 are concatenated along the channel dimension and compressed through a 1×1 convolutional layer to obtain the intermediate feature Fmid. The fused comprehensive feature is then adjusted through a 1×1 convolutional layer and residually connected with the intermediate feature Fmid. Finally, the residually connected feature is input into the channel attention module to generate a weight vector. The weight vector is then multiplied channel-by-channel with the residually connected feature to output the enhanced feature map. The defect localization module includes a classification head and a segmentation head. The classification head contains a fully connected layer and a Softmax function layer, and the segmentation head contains a transposed convolutional layer and a convolutional layer. The internal calculation process of the defect localization module is as follows: the enhanced feature map is global average pooled and then input into the classification head to output the defect type probability of the device. At the same time, the enhanced feature map is input into the segmentation head to output the location mask of the defect.

8. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The training process of the second model includes: Historical multimodal image data of different types of power transmission equipment were collected; the multimodal images include visible light images, infrared thermal imaging images, and ultraviolet corona discharge images. Label the corresponding defect type and defect location mask in historical multimodal image data, and construct training datasets categorized by device type; The second model parameters are initialized based on the second model architecture, and the training process is started for different types of devices: Before training, the output dimension of the classification head in the defect localization module is adjusted by the preset mapping relationship between device type name and number of categories, and the training dataset of the corresponding device type is loaded. During training, a multi-task loss function is used, taking the enhanced feature map as input and optimizing the model parameters through backpropagation until the loss function converges; the multi-task loss function is composed of a cross-entropy loss function and a Dice loss function weighted in a 1:1 ratio. After training is completed for each type of device, the corresponding model parameters are saved as pre-training weights for each device type, resulting in a pre-training weight library for each device type.

9. The method for monitoring the status of power transmission equipment based on the Internet of Things according to claim 1, characterized in that, The status data of the power transmission equipment includes conductor temperature, three-phase current range, frequency domain energy of tower vibration, and fitting temperature; the environmental parameters include temperature T, humidity H, and wind speed W.

Images