Defect detection method and device for power transmission equipment and electronic equipment

Abstract

The embodiment of the invention provides a defect detection method and device for power transmission equipment and electronic equipment. The method comprises the following steps: simulating a plurality of defect scenes by using a first twin model of first power transmission equipment, and generating a simulation defect image and simulation operation data of each defect scene; the simulation defect images and the simulation operation data of all the defect scenes are processed through a bidirectional mapping model, target mapping data are obtained, and the target mapping data are used for representing the corresponding relation between the defect semantic features and the numerical value interval features under the different defect scenes; generating multiple groups of first training samples based on the target mapping data; utilizing multiple groups of first training samples to train the initial federal sub-model to obtain a first federal sub-model; and determining target defect risk information corresponding to the first power transmission equipment according to the first federal sub-model, the current first image data of the first power transmission equipment and the first actual operation data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission systems, and in particular to a defect detection method and device for a power transmission device and an electronic device. BACKGROUND

[0002] In the operation of a power transmission system, its operation data, such as tower state, line current or voltage, and inspection data, such as point cloud and image, collected by an inspection device, are collected, and the operation data and inspection data of the power transmission system are combined to predict the risks of the power transmission system.

[0003] However, the existing risk prediction method only aligns the operation data and inspection data in the time and space dimensions, without considering their internal correlation, resulting in poor prediction accuracy and low warning credibility. SUMMARY

[0004] The embodiments of the present application provide a defect detection method and device for a power transmission device and an electronic device, which can accurately identify early hidden risks such as small cracks in insulators and low-frequency fluctuations in leakage current, and significantly improve the accuracy and reliability of risk inspection.

[0005] In a first aspect, the embodiments of the present application provide a defect detection method for a power transmission device, which includes: simulating a plurality of defect scenarios using a first twin model of a first power transmission device, and generating simulation defect images and simulation operation data for each defect scenario; processing the simulation defect images and simulation operation data for each defect scenario through a bidirectional mapping model to obtain target mapping data, the target mapping data being used to represent the corresponding relationship between defect semantic features and numerical interval features in different defect scenarios; generating a plurality of first training samples based on the target mapping data, each first training sample including a training defect image, training operation data, and actual defect risk information; training an initial federal sub-model using the plurality of first training samples to obtain a first federal sub-model; and determining target defect risk information corresponding to the first power transmission device according to the first federal sub-model, first image data, and first actual operation data of the first power transmission device.

[0006] In a second aspect, the embodiments of the present application provide a defect detection device for a power transmission device, which includes:

[0007] a simulation module configured to simulate a plurality of defect scenarios using a first twin model of a first power transmission device, and generate simulation defect images and simulation operation data for each defect scenario;

[0008] The data processing module is configured to process the simulation defect images and the simulation operation data of each defect scene by using the bidirectional mapping model to obtain target mapping data, the target mapping data being used to represent a corresponding relationship between defect semantic features and numerical interval features under different defect scenes.

[0009] The generation module is configured to generate a plurality of groups of first training samples based on the target mapping data, each group of the first training samples including a training defect image, training operation data, and actual defect risk information.

[0010] The training module is configured to train the initial federated sub-model by using the plurality of groups of first training samples to obtain a first federated sub-model.

[0011] The determination module is configured to determine target defect risk information corresponding to the first power transmission device according to the first federated sub-model, first image data of the first power transmission device, and first actual operation data of the first power transmission device.

[0012] In a third aspect, an electronic device is provided, including a memory and a processor.

[0013] The memory stores computer-executable instructions.

[0014] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the first aspect and / or various possible implementation manners of the first aspect.

[0015] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are executed by a processor, the computer-executable instructions are used to implement the first aspect and / or various possible implementation manners of the first aspect.

[0016] In a fifth aspect, a computer program product is provided, and the computer program product includes a computer program. When the computer program is executed by a processor, the computer program implements the first aspect and / or various possible implementation manners of the first aspect.

[0017] In the embodiment of the present application, the first twin body model corresponding to the first power transmission equipment is used to simulate multiple defect scenarios and generate simulation defect images and simulation running data under each scenario. Further, the simulation defect images and simulation running data are processed by the bidirectional mapping model to obtain target mapping data representing the corresponding relationship between the defect semantic features and the numerical interval features under different defect scenarios, thereby realizing the quantitative association of the defect semantic features extracted from the inspection image and the numerical interval features in the running data, and physically verifying the risk authenticity based on the first twin body model. On this basis, a plurality of first training samples are generated based on the target mapping data, and the initial federated sub-model is trained using the plurality of first training samples to obtain the first federated sub-model. Then, according to the first federated sub-model obtained by training, the target defect risk information corresponding to the first power transmission equipment is determined in combination with the current first image data and the first actual running data of the equipment. Through the above-mentioned manner, the embodiment not only solves the semantic gap problem caused by the dependence on the spatio-temporal alignment in the traditional method, but also avoids the misjudgment caused by the lack of physical mechanism (such as misjudging the sag contraction caused by temperature drop as icing), thereby accurately identifying early hidden risks such as small cracks in insulators and low-frequency fluctuations in leakage current, and significantly improving the accuracy and reliability of risk inspection. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and serve to explain the principles of the present application together with the specification.

[0019] Figure 1 A schematic diagram of an application scenario of a defect detection method of a power transmission equipment according to an embodiment of the present application;

[0020] Figure 2 A flowchart of a defect detection method of a power transmission equipment according to an embodiment of the present application;

[0021] Figure 3 A flowchart of another defect detection method of a power transmission equipment according to an embodiment of the present application;

[0022] Figure 4 A structural diagram of a defect detection device of a power transmission equipment according to an embodiment of the present application;

[0023] Figure 5 A structural diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] Embodiments will be described in detail below with reference to examples illustrated in the accompanying drawings. When the description below refers to the drawings, unless otherwise noted, like numbers in different figures refer to like or similar elements. The embodiments described in the following examples do not represent all the implementations consistent with the present application. They are merely examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0025] It should be noted that the brief description of terms in the present application is only for the convenience of understanding the following description of embodiments, and is not intended to limit the embodiments of the present application. Unless otherwise specified, these terms should be understood according to their ordinary and general meanings.

[0026] The terms "first", "second", "third" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar or like objects or entities, and do not necessarily mean a specific order or sequence, unless otherwise noted. It should be understood that the terms used in this way can be interchanged under appropriate circumstances.

[0027] The terms "include" and "have" and any variations thereof are intended to cover but not exclusive inclusion, for example, a product or device including a series of components does not necessarily limit to all components clearly listed, but can include other components not clearly listed or inherent to these products or devices.

[0028] The term "module" refers to any known or later developed hardware, software, firmware, artificial intelligence, fuzzy logic, or combination of hardware or / and software code capable of performing a function associated with that element.

[0029] During the operation of the power transmission system, the operation data of the power transmission system, such as the state of the tower, the line current or voltage, etc., and the inspection data of the power transmission system, such as the point cloud and image collected by the inspection equipment, are actually collected. By combining the operation data and the inspection data of the power transmission system, the risk of the power transmission system can be predicted.

[0030] However, the existing risk prediction method only aligns the operation data and the inspection data in the time and space dimensions, without considering the internal correlation, resulting in poor prediction accuracy and low warning credibility. For example, the change in the conductor sag value caused by the sudden temperature drop is easily mistaken for a conductor icing risk, resulting in a false alarm.

[0031] Therefore, an embodiment of the present application provides a defect detection method of a power transmission device, which simulates multiple defect scenarios and generates simulation defect images and simulation running data under each scenario through a first twin body model corresponding to a first power transmission device; further processes the simulation defect images and the simulation running data through a bidirectional mapping model to obtain target mapping data for representing the corresponding relationship between defect semantic features and numerical interval features under different defect scenarios, so as to realize the quantitative association of the defect semantic features extracted from the inspection image and the numerical interval features in the running data, and physically verify the risk authenticity based on the first twin body model. On this basis, multiple sets of first training samples are generated based on the target mapping data, and the initial federated sub-model is trained using the multiple sets of first training samples to obtain a first federated sub-model; then, the target defect risk information corresponding to the first power transmission device is determined according to the first federated sub-model obtained by training, in combination with the current first image data and the first actual running data of the device. In this way, the embodiment not only solves the semantic gap problem caused by the dependence on time and space alignment in the traditional method, but also avoids the misjudgment caused by the lack of physical mechanism (such as misjudging the sag contraction caused by temperature drop as icing), so as to accurately identify early hidden risks such as small cracks in insulators and low-frequency fluctuations of leakage current, and significantly improve the accuracy and reliability of risk inspection.

[0032] Before introducing the defect detection method of the power transmission device provided by the embodiment of the present application, the application scenario of the defect detection method of the power transmission device is described.

[0033] The embodiment of the present application is applied to a power transmission system. As shown in Figure 1 The power transmission system includes multiple power transmission devices, such as power transmission device 1, power transmission device 2, …, and power transmission device n. Each power transmission device is an edge end and is in communication connection with a server (i.e., cloud). For each power transmission device (i.e., edge end), detection data of the power transmission device is collected by an inspection device, and a twin body model corresponding to the power transmission device is constructed based on the detection data of the power transmission device. Multiple defect scenarios are simulated through the twin body model to generate simulation defect data and simulation running data, and the simulation defect data and the simulation running data can be used to train a federated sub-model, which can be used to identify defect risk information of the power transmission device. In addition, the server (i.e., cloud) can aggregate and process the model parameters of the federated sub-model of each edge end and deliver them to each power transmission device (i.e., edge end) to enable each power transmission device to iteratively update the federated sub-model.

[0034] Further, at each power transmission device (i.e., edge end), the simulated defect images and the simulated operation data of each defect scenario are processed by a bidirectional mapping model to obtain target mapping data, which is used to represent the corresponding relationship between the defect semantic features and the numerical interval features under different defect scenarios. The bidirectional mapping model provides the target mapping data to provide data support for the basic risk identification of the federal sub-model.

[0035] The defect detection method of the power transmission device provided by the embodiments of the present application will be described in detail below in combination with the drawings and application scenarios.

[0036] Figure 2 A flowchart of a defect detection method of a power transmission device provided by the embodiments of the present application. The defect detection method of the power transmission device is applied to a first power transmission device. As Figure 2 The defect detection method of the power transmission device can include the following steps:

[0037] S201, simulating a plurality of defect scenarios by using a first twin model of the first power transmission device, and generating simulated defect images and simulated operation data of each defect scenario.

[0038] The first twin model refers to a calculated virtual entity that is mapped in geometry, physical properties, dynamic behavior, and functional logic with the first power transmission device.

[0039] In some embodiments, before performing S201, the method further includes the step of constructing the first twin model of the first power transmission device based on second detection data of the first power transmission device. Specifically, the construction process of the first twin model is as follows:

[0040] S301, obtaining second detection data of the first power transmission device by a patrol device, the second detection data including three-dimensional point cloud data and image data.

[0041] The three-dimensional point cloud data is used to describe the contour and spatial structure of the first power transmission device. For example, the three-dimensional point cloud data can include the three-dimensional point cloud data of the tower and the conductor of the first power transmission device. The three-dimensional point cloud data can be collected by a radar carried on the patrol device (such as a patrol robot). For example, the collection resolution of the three-dimensional point cloud data is 0.05 mm.

[0042] The image data refers to the image data collected by a camera carried on the patrol device (such as a patrol robot). For example, the image data of the detection points of the first power transmission device. The image data can represent the surface features of the tower and the conductor of the first power transmission device, such as color, texture, and other visual information. For example, the image data can have a resolution of 1920x1080.

[0043] S302, construct a first twin-body model corresponding to the first power transmission device according to the second detection data.

[0044] Specifically, the first twin-body model is generated in combination with the three-dimensional point cloud data, the image data and the original device data (i.e., design drawings) of the first power transmission device. The scale of the first twin-body model to the first power transmission device is 1:1.

[0045] After the first twin-body model of the first power transmission device is constructed, the second actual operation data of the first power transmission device is obtained. The second actual operation data is mapped to the first twin-body model.

[0046] The second actual operation data is used to represent the actual operation state of the first power transmission device. For example, the second actual operation data can include leakage current, temperature (such as conductor temperature), etc. The second actual operation data can be collected by sensors arranged on the first power transmission device or inspection equipment.

[0047] Mapping the second actual operation data to the first twin-body model means updating the physical state of the first twin-body model according to the second actual operation data. For example, the color of the conductor of the first twin-body model is updated according to the conductor temperature. For example, when the conductor temperature is less than or equal to the first preset conductor temperature (such as 40°C), the conductor of the first twin-body model is rendered green. When the conductor temperature is greater than the first preset conductor temperature (such as 40°C) and less than or equal to the second preset conductor temperature (such as 60°C), the conductor of the first twin-body model is rendered yellow. When the conductor temperature is greater than the second preset conductor temperature (such as 60°C), the conductor of the first twin-body model is rendered red.

[0048] For example, according to the leakage current, the current load is marked on the conductor. Specifically, the mapping rule of "100A current corresponds to an increase of 0.1mm in the diameter of the conductor in the twin-body model" is used to mark the current load on the conductor. For example, when the leakage current is 300A, the display diameter of the conductor in the first twin-body model will increase by 0.3mm based on the basic value, and the current load value is marked by a dynamic label.

[0049] In some embodiments, the method can further include virtually and physically calibrating the first twin-body model. For example, virtually and physically calibrating the first twin-body model can include the following steps:

[0050] S401, obtaining the second actual operation data and the actual conductor sag value of the first power transmission device.

[0051] S402, determining the predicted conductor sag value corresponding to the second actual operation data by using the first twin-body model.

[0052] S403 compares the predicted conductor sag value with the actual conductor sag value to obtain the comparison result.

[0053] S404. Based on the comparison results, the first twin model is modified to obtain the modified first twin model.

[0054] For example, a virtual-real calibration algorithm is used to correct the first twin model based on the comparison results, resulting in a corrected first twin model.

[0055] In some examples, a virtual-real calibration formula is used to correct the first twin model, resulting in a corrected first twin model. The virtual-real calibration formula is as follows:

[0056] Virtual / real calibration formula:

[0057]

[0058] This represents the physical parameters of the modified first twin model; Represents the physical parameters of the first twin model; Indicates the parameter correction coefficient; This represents measured data; Represents parameters based on time t Twin simulation calculation data. and They are the same physical quantity and have the same unit.

[0059] In other examples, the virtual-real calibration algorithm can be an extended Kalman filter (EKF).

[0060] Specifically, the physical parameters of the first twin model are used as the state vector X, and the measured data are used as the observation vector Z. Virtual-real calibration is achieved through extended Kalman filter (EKF) prediction-update. The physical parameters of the first twin model can include conductor elasticity coefficients, tower stiffness, etc. The measured data can include actual conductor sag values, conductor temperatures, etc.

[0061] The calibration process is as follows: Prediction step: Xpred = A × Xprev + B × u, where A is the state transition matrix, which is the identity matrix; B is the control matrix; and u is the environmental disturbance term, such as the wind speed influence coefficient.

[0062] The update step is K = Ppred x HT x (H x Ppred x HT + R)-1, wherein K is a Kalman gain, Ppred is a prediction error covariance, H is an observation matrix, R is an observation noise covariance, and the value of R is 0.01, and the final corrected parameter Xcorr is Xpred + K x (Z - H x Xpred); compared with the traditional least square method, the calibration error is reduced by 30%, and the convergence speed is increased by 25%.

[0063] In the embodiments of the application, the extended Kalman filter is used to update the physical parameters of the first twin model, the parameters are dynamically corrected through the "prediction-update" iteration logic, the parameter change trend is derived based on the twin model in the prediction stage, and the parameter deviation is adjusted in combination with the measured data in the update stage, so that the long-term consistency of the simulation state of the twin model and the actual state in the field is ensured.

[0064] In some embodiments, after the first twin model is constructed and calibrated, a plurality of defect scenarios are simulated by using the first twin model, and simulation defect images and simulation running data of each defect scenario are generated. The defect scenarios may include, for example, gradual change of insulator point erosion area, hardware corrosion, wire breakage, etc. The simulation running data may include leakage current and wire temperature.

[0065] For example, by using the first twin model, the gradual change of the insulator point erosion area is simulated, and the leakage current and the wire temperature of the first twin model under different insulator point erosion areas are calculated, that is, the simulation defect images and the simulation running data are obtained.

[0066] In S202, the simulation defect images and the simulation running data of each defect scenario are processed by using the bidirectional mapping model to obtain target mapping data, and the target mapping data are used to represent the corresponding relationship between the defect semantic features and the numerical interval features under different defect scenarios.

[0067] The defect semantic features are used to describe the defect types. The numerical interval features are numerical intervals corresponding to a defect type. For example, the defect semantic features are the insulator point erosion area of 0-1 mm², and the numerical interval features are the leakage current of 5-8 mA.

[0068] The target mapping data are data representing the corresponding relationship between the defect semantic features and the numerical interval features of a plurality of defects.

[0069] For example, the target mapping data include:

[0070] The insulator point erosion area of 0-1 mm²→ the leakage current of 5-8 mA;

[0071] The insulator point erosion area of 1-3 mm²→ the leakage current of 8-15 mA;

[0072] Insulator point erosion area 3-5 mm² → leakage current 15-20 mA.

[0073] The bidirectional mapping model is an attention mechanism driven model.

[0074] In some embodiments, the simulation defect images and the simulation running data of each defect scene are processed by the bidirectional mapping model to obtain the target mapping data, which can include the following steps:

[0075] S501, the simulation defect images and the simulation running data of each defect scene are processed by the bidirectional mapping model to obtain the target mapping data, which can include the following steps:

[0076] Specifically, the simulation defect image is input into a lightweight convolutional neural network to extract high-level and abstract visual features, and output a fixed-length 128-dimensional semantic feature vector. The semantic feature vector encodes key semantic information such as erosion edge texture and defect morphology. In addition, the simulation running data is processed to obtain a numerical feature vector. For example, a 4-dimensional numerical feature vector composed of leakage current, conductor temperature, humidity, and wind speed.

[0077] S502, the simulation defect images and the simulation running data of each defect scene are processed by the bidirectional mapping model to obtain the target mapping data, which can include the following steps:

[0078] The 128-dimensional semantic feature vector and the 4-dimensional numerical feature vector are input into a cross-attention layer. The layer uses a Scaled Dot-Product mechanism to calculate attention weights, so that the numerical feature vector can focus on the most relevant part of the semantic feature, thereby realizing effective fusion of two different modal features, and outputting a fused context feature representation, i.e. fusion feature. The above fusion feature is processed through a hidden module containing 2 fully connected layers, each layer containing 64 neurons and using a ReLU activation function for nonlinear transformation, outputting a complex and high-order mapping relationship in the output interval, i.e. target mapping data.

[0079] In some embodiments, after the first twin model is constructed, the first twin model can be adjusted according to the environmental parameters of the environment where the first power transmission equipment is located, to obtain an adjusted first twin model. The adjusted first twin model is used to simulate multiple defect scenarios and generate simulation defect images and simulation running data for each defect scenario, and the semantic feature vectors corresponding to each simulation defect image and the numerical feature vectors corresponding to each simulation running data are processed to generate target mapping data.

[0080] The environmental parameters can include at least one of a wind speed, an environmental humidity, and an environmental temperature.

[0081] Of course, the first twin body model after the virtual-real calibration and the correction can also be adjusted to obtain an adjusted first twin body model.

[0082] The process of adjusting the first twin body model according to the environmental parameters of the environment in which the first power transmission device is located to obtain an adjusted first twin body model is described below.

[0083] In some embodiments, adjusting the first twin body model according to the environmental parameters to obtain an adjusted first twin body model includes: if the environmental parameters meet a triggering condition, adjusting the first twin body model to obtain an adjusted first twin body model.

[0084] The environmental parameters meeting the triggering condition include at least one of:

[0085] a rate of change of the environmental temperature is less than a preset temperature change rate, and the preset temperature change rate is a negative value (i.e., a first triggering condition);

[0086] a wind speed is greater than a preset wind speed threshold (i.e., a second triggering condition);

[0087] an environmental humidity is greater than a preset humidity threshold (i.e., a third triggering condition).

[0088] It should be noted that the monitoring frequency of the environmental parameters is 1 time per minute, and when the change range of the environmental parameters is greater than a preset range (such as 20%), the instant correction is triggered, otherwise, the correction is performed every preset time interval (such as 30 minutes).

[0089] Specifically, in some examples, if the wind speed is greater than the preset wind speed threshold, the conductor wind deflection force calculation model in the first twin body model is updated.

[0090] For example, the preset wind speed threshold is, for example, 10 m / s.

[0091] The conductor wind deflection force calculation model is as follows: F = 0.613 x p x v2 x d x L.

[0092] Wherein, p is the air density kg / m3, v is the wind speed m / s, d is the conductor diameter, and L is the span.

[0093] Specifically, when the wind speed is greater than the preset wind speed threshold (such as 10 m / s), the first twin body model calls and updates the conductor wind deflection force calculation model.

[0094] and, after updating the conductor windage force calculation model in the first twin model, the target mapping data is adjusted synchronously, i.e. the target mapping data is reacquired. For example, under the same erosion area of the insulator, the leakage current threshold at a wind speed of 15 m / s is increased by 12% compared to that at a wind speed of 5 m / s.

[0095] In some examples, if the ambient humidity is greater than a preset humidity threshold, the insulator leakage current correction model in the first twin model is updated.

[0096] For example, the preset humidity threshold is 85%.

[0097] The insulator leakage current correction model is shown in the following formula:

[0098] Icorr = Imeas x (1 + 0.02 x (RH - 85%)).

[0099] where RH is the actual measured relative humidity (i.e. ambient humidity), Imeas is the measured leakage current, and Icorr is the corrected leakage current.

[0100] and, after updating the insulator leakage current correction model in the first twin model, the target mapping data is adjusted synchronously, i.e. the target mapping data is reacquired. For example, the insulator erosion area-leakage current mapping relationship is adjusted synchronously, such as that the current threshold for the same erosion area at a humidity of 90% is increased by 8% compared to that at a humidity of 80%.

[0101] In some examples, if the rate of change of the ambient temperature is less than a preset temperature change rate, the conductor thermal expansion and contraction model in the first twin model is updated, and the preset temperature change rate is a negative value.

[0102] For example, the preset temperature change rate is -2℃ / h.

[0103] The conductor thermal expansion and contraction model is shown in the following formula:

[0104] Lcorr = L0 x (1 + aT x AT).

[0105] where L0 is the conductor span at room temperature, aT is the thermal expansion coefficient of the conductor, the thermal expansion coefficient of aluminum strand is 23 x 10-6 / ℃, and AT is the temperature change amount.

[0106] and, after updating the conductor thermal expansion and contraction model in the first twin model, the target mapping data is adjusted synchronously, i.e. the target mapping data is reacquired, i.e. the defect semantic features and the corresponding numerical interval features are updated.

[0107] In some embodiments, if multiple parameters in the environmental parameters satisfy respective trigger conditions, the corresponding physical model in the first twin model is updated in order according to the priority of the trigger conditions corresponding to the various parameters in the environmental parameters.

[0108] The trigger condition corresponding to the ambient temperature (i.e., the first trigger condition) is that the rate of change of the ambient temperature is less than a preset temperature change rate. The trigger condition corresponding to the wind speed (i.e., the second trigger condition) is that the wind speed is greater than a preset wind speed threshold. The trigger condition corresponding to the ambient humidity (i.e., the third trigger condition) is that the ambient humidity is greater than a preset humidity threshold. The priority of the first trigger condition is higher than the priority of the second trigger condition, and the priority of the second trigger condition is higher than the priority of the third trigger condition.

[0109] That is, the temperature drops sharply (priority 1, immediate correction) > wind speed > 8 m / s (priority 2, corrected within 5 seconds) > humidity > 85% (priority 3, corrected within 30 seconds).

[0110] In this embodiment, the bidirectional mapping model introduces an attention mechanism, and pre-simulates multiple defect scenarios such as gradual change of insulator point erosion area and deepening of hardware corrosion degree through the twin model, to obtain the change law of operating data (such as leakage current, conductor temperature, etc.) under each defect scenario. In this way, when training the bidirectional mapping model, the key features such as defect edge texture and numerical pulse type fluctuation peak are automatically focused, and interference information such as background vegetation and normal numerical fluctuation is filtered, and a dynamic association table of defect semantic features and numerical interval features, i.e., target mapping data, is output.

[0111] Furthermore, the first twin model is adjusted based on the environmental parameters, which improves the accuracy of the first twin model in simulating operating data under different scenarios, and thus more accurate and reasonable target mapping data is obtained.

[0112] S203, based on the target mapping data, generate a plurality of first training samples, each first training sample including a training defect image, training operating data, and actual defect risk information.

[0113] In each first training sample, the training defect image and the training operating data satisfy the corresponding relationship between the defect semantic features and the numerical interval features represented by the target mapping data. The training defect image can be simulation image data. For example, an insulator image containing a point erosion of a specific area. The training operating data can include leakage current and conductor temperature. The training operating data can also be simulation operating data.

[0114] It should be noted that the training defect images and the training operation data in each group of samples strictly follow the corresponding relationship between the defect semantic features and the numerical interval features represented by the target mapping data. For example, an image labeled as "insulator point erosion area 1-3 mm²" must have corresponding leakage current simulation data falling within the "8-15 mA" interval specified by the target mapping data.

[0115] The actual defect risk information (i.e., actual defect risk label) is used to indicate the real defect category represented by each group of training samples. For example, the actual defect risk label includes the following four types: no defect, insulator obvious defect, current exceeding standard, and temperature anomaly.

[0116] S204, training the initial federated sub-model using the plurality of groups of first training samples to obtain a first federated sub-model.

[0117] The first federated sub-model is used to identify the target defect risk information of the first power transmission equipment.

[0118] Specifically, the training defect images and the training operation data are input into the initial federated sub-model to output a predicted defect risk label, and the actual defect risk label is used as supervision information to iteratively train the initial federated sub-model to obtain the first federated sub-model.

[0119] S205, determining the target defect risk information corresponding to the first power transmission equipment according to the first federated sub-model, first image data of the first power transmission equipment, and first actual operation data.

[0120] The first image data is an image of a detection point of the first power transmission equipment. For example, the first image data is an image of an insulator containing a point erosion of a specific area.

[0121] The first actual operation data is used to represent the current actual operation state of the first power transmission equipment. For example, the first actual operation data can include leakage current, temperature (such as conductor temperature), etc. The first actual operation data can be collected by sensors arranged on the first power transmission equipment or inspection equipment.

[0122] The target defect risk information (i.e., target defect risk label) is used to indicate the current defect category of the first power transmission equipment. For example, the target defect risk label includes the following four types: no defect, insulator obvious defect, current exceeding standard, and temperature anomaly.

[0123] Exemplarily, the first federated sub-model adopts a MobileNetV2 lightweight network. The input of the first federated sub-model is first image data, such as a 640*640 resolution inspection image, and first actual operation data, such as 4 numerical data including leakage current, conductor temperature, wind speed, and humidity. The input of the first federated sub-model is a target defect risk label, such as no defect, obvious defect of insulator, current exceeding standard, and temperature anomaly.

[0124] In some embodiments, after determining the target defect risk information corresponding to the first power transmission device, the method further includes: determining a target defect level of the first power transmission device according to the first image data, the first actual operation data, and the target defect risk information. According to the target defect level of the first power transmission device, different strategies are adopted to generate defect data of the first power transmission device.

[0125] In some examples, determining the target defect level of the first power transmission device according to the first image data, the first actual operation data, and the target defect risk information includes: determining a current fluctuation range corresponding to the leakage current according to the leakage current, and determining a temperature deviation according to the conductor temperature; determining a defect area according to the first image data; and determining a target risk level according to the current fluctuation range, the temperature deviation, the defect area, and the target defect risk information.

[0126] For example, level 1 risk (no risk): current fluctuation range < 5%, temperature deviation < 3°C, and target defect risk label is no defect;

[0127] Level 2 risk (low risk): current fluctuation range is 5%-10% or temperature deviation is 3%-5%, and target defect risk label is no defect;

[0128] Level 3 risk (medium risk): current fluctuation range is 10%-20%, or temperature deviation is 5%-8%, or target defect risk label is obvious defect of insulator and defect area < 2mm²;

[0129] Level 4 risk (higher risk): current fluctuation range is 20%-30%, or temperature deviation is 8%-12%, or target defect risk label is obvious defect of insulator and defect area is 2-5mm²;

[0130] Level 5 risk (high risk): current fluctuation range > 30%, or temperature deviation > 12%, or target defect risk label is obvious defect of insulator and defect area > 5mm².

[0131] In some embodiments, generating defect data of the first power transmission device according to the target defect level of the first power transmission device can include:

[0132] S601, if the target defect level is less than a first preset defect level, identifying the first image data to determine a defect region and a non-defect region.

[0133] The first preset defect level is 3. That is, when the target defect level is 1 (no risk) or 2 (low risk), the first image data is identified to determine the defect region (i.e., the region of interest, ROI) and the non-defect region.

[0134] S602, based on the first image data, determining semantic defect labels and position information of the defect region (i.e., coordinate information of the boundary of the defect region).

[0135] The semantic defect labels can be generated based on the defect semantic features.

[0136] S603, compressing the non-defect region in the first image data to obtain compressed first image data.

[0137] The non-defect region in the first image data is compressed according to a preset compression ratio (e.g., 10:1) to obtain the compressed first image data.

[0138] S604, generating defect data of the first power transmission equipment according to the semantic defect labels, the position information of the defect region, and the compressed first image data.

[0139] In some embodiments, according to the target defect level of the first power transmission equipment, different strategies are adopted to generate the defect data of the first power transmission equipment, which can include: if the target defect level is greater than or equal to the first preset defect level, a first twin model is used to simulate the first power transmission equipment based on the first actual operation data to generate defect anomaly data.

[0140] The defect anomaly data can include, for example, a wire icing twin cross-section diagram, an insulator electric field distribution cloud diagram, and a key component stress distribution diagram.

[0141] In this embodiment, by introducing an edge intelligent hierarchical inference mechanism, the dynamic adaptation of computing and communication resources according to the risk level is realized. For low-risk scenarios (such as current fluctuation < 10%, temperature deviation < 5℃, no semantic defects), the edge end only calls lightweight models such as MobileNetV2 to realize fast recognition, and adopts the "semantic priority strategy" to compress data, greatly reducing the data transmission amount to the cloud and improving the overall efficiency of the system. For high-risk scenarios (such as current fluctuation > 20% or temperature deviation > 8% or semantic defect area > 2mm²), the edge end triggers the collaborative inference of "federal sub-model and twin local simulation". In this mode, the system preferentially collects and transmits the key abnormal data (such as icing cross-section graph) generated by the twin, which has small data volume but high information density. After receiving the high-risk alarm, the cloud synchronously schedules the dedicated GPU computing cluster for reception and accelerated analysis, thereby greatly shortening the overall response and decision-making time of high-risk events and ensuring the safe and stable operation of the power transmission line.

[0142] In some embodiments, after the first federal sub-model of the first power transmission device is trained, the model parameter update gradient of the first federal sub-model is uploaded to the server, so that the server aggregates the model parameter update gradient of the first federal sub-model and the model parameter update gradient of the federal sub-model of other power transmission devices to obtain an aggregated model parameter update gradient, and updates the global model using the aggregated model parameter update gradient to obtain a global model weight. The global model weight (such as the model weight) is sent to each first power transmission device (i.e., the edge end). The edge end updates the federal sub-model based on the global model weight. For example, the first power transmission device updates the first federal sub-model based on the global model weight.

[0143] For example, the gradient is encrypted using the RSA asymmetric encryption algorithm. The public key required for encryption is dynamically generated by the cloud federal server and updated every 24 hours, effectively preventing the gradient from being stolen and reverse-engineered during transmission.

[0144] For example, the server aggregates the model parameter update gradient of the first federal sub-model and the model parameter update gradient of the federal sub-model of other power transmission devices to obtain an aggregated model parameter update gradient, including:

[0145] S701, according to the number of training samples of each federal sub-model and the total amount of training samples of each federal sub-model, determine the edge node weight of each federal sub-model.

[0146] For example, the edge node weight of the federal sub-model is determined according to the following formula:

[0147]

[0148] wherein, is an edge node weight of the federated sub-model; is a training sample number of the federated sub-model; is a total training sample number of each federated sub-model.

[0149] For example, the edge node weight of the first federated sub-model is a ratio of the training sample number of the first federated sub-model to the total training sample number of each federated sub-model.

[0150] S702, according to the model weight of each federated sub-model and the corresponding edge node weight, determining the aggregated global model weight.

[0151] For example, the global model weight is determined according to the following formula:

[0152]

[0153] wherein, denotes the global model weight, is an edge node weight of the federated sub-model; is the model weight of the i-th edge node (i.e. federated sub-model) updated based on the model parameter update gradient;

[0154] wherein, when the gradient variance (model parameter update gradient) of a certain edge node is greater than a preset update gradient (such as 0.1), it indicates that the gradient difference with other nodes is too large, and the edge node weight of the edge node (federated sub-model) is reduced, so as to determine the global model weight based on the reduced edge node weight.

[0155] For example, when the gradient variance of a certain edge node is greater than 0.1 (the gradient difference with other nodes is too large), its weight is temporarily reduced to ωi×0.5 to prevent abnormal gradient from causing global model shock; the aggregation frequency matches the edge iteration number, the edge uploads the gradient once every 5 rounds (100 samples per round) of training, and the cloud completes the aggregation and issues the update within 10 seconds after receiving the gradient of all nodes.

[0156] In some embodiments, if the target defect level is greater than or equal to a first preset defect level, the fuzzy risk probability of the multi-source data is determined according to the first image data and the first actual running data (i.e. multi-source data).

[0157] wherein, the fuzzy risk probability is a probability that the multi-source data is real defect data.

[0158] In some examples, determining the fuzzy risk probability of the multi-source data can include: performing feature extraction on the first image data to obtain a semantic feature vector, determining a semantic feature confidence of the semantic feature vector; performing feature extraction on the first actual operation data (such as leakage current and / or conductor temperature) to obtain a numerical feature vector, and determining a numerical feature fluctuation degree (i.e. current fluctuation range) according to the numerical feature vector. The fuzzy risk probability of the multi-source data is determined by using a preset ANFIS model, the semantic feature confidence, and the numerical feature fluctuation degree (i.e. current fluctuation range).

[0159] The ANFIS model has input variables of a semantic feature confidence ( ) and a numerical feature fluctuation degree ( ), wherein each input variable can be divided into three fuzzy subsets (i.e. low, medium, and high), and each fuzzy subset is generated by using a Gaussian function. For example, the high fuzzy subset uses a Gaussian function as:

[0160] .

[0161] The fuzzy risk probability is determined according to a preset fuzzy rule.

[0162] For example, the preset fuzzy rule can include:

[0163] Pre-set fuzzy rule 1: If = low And = low Then = low (the output probability range is 0-0.2);

[0164] That is, if the semantic feature confidence ( ) belongs to the low fuzzy subset, and the numerical feature fluctuation degree ( ) belongs to the low fuzzy subset, then the fuzzy risk probability ( ) is a low fuzzy risk probability, and the fuzzy risk probability range is 0-0.2.

[0165] Pre-set fuzzy rule 2: If = low And = medium Then = low-medium (the output probability range is 0.1-0.3);

[0166] That is, if the semantic feature confidence ( ) belongs to the low fuzzy subset, and the numerical feature fluctuation degree ( ) belongs to the medium fuzzy subset, then the fuzzy risk probability ( ) is a low-medium fuzzy risk probability, and the fuzzy risk probability range is 0.1-0.3.

[0167] Pre-set fuzzy rule 3: If = Low And = High Then = Medium (output probability range 0.2-0.4);

[0168] i.e. if semantic feature confidence ( ) belongs to low ambiguity subset and numerical feature volatility ( ) belongs to high ambiguity subset, then ambiguity risk probability ( ) is medium ambiguity risk probability, ambiguity risk probability range is 0.2-0.4.

[0169] Pre-set ambiguity rule 4: If = Medium And = Low Then = Low-Medium (output probability range 0.1-0.3);

[0170] i.e. if semantic feature confidence ( ) belongs to medium ambiguity subset and numerical feature volatility ( ) belongs to low ambiguity subset, then ambiguity risk probability ( ) is low-medium ambiguity risk probability, ambiguity risk probability range is 0.1-0.3.

[0171] Pre-set ambiguity rule 5: If = Medium And = Medium Then = Medium (output probability range 0.3-0.6);

[0172] i.e. if semantic feature confidence ( ) belongs to medium ambiguity subset and numerical feature volatility ( ) belongs to medium ambiguity subset, then ambiguity risk probability ( ) is medium ambiguity risk probability, ambiguity risk probability range is 0.3-0.6.

[0173] Pre-set ambiguity rule 6: If = Medium And = High Then = Medium-High (output probability range 0.5-0.8);

[0174] i.e. if semantic feature confidence ( ) belongs to medium ambiguity subset and numerical feature volatility ( ) belongs to high ambiguity subset, then ambiguity risk probability ( ) is medium-high ambiguity risk probability, ambiguity risk probability range is 0.5-0.8.

[0175] Pre-set ambiguity rule 7: If = High And = Medium Then = Medium (output probability range is 0.3-0.6);

[0176] That is, if the semantic feature confidence ( ) belongs to the high fuzzy subset, and the numerical feature fluctuation degree ( ) belongs to the low fuzzy subset, then the fuzzy risk probability ( ) is a medium fuzzy risk probability, and the fuzzy risk probability range is 0.3-0.6.

[0177] Pre-set fuzzy rule 8: If = High And = Medium Then = Medium-High (output probability range is 0.5-0.8);

[0178] That is, if the semantic feature confidence ( ) belongs to the high fuzzy subset, and the numerical feature fluctuation degree ( ) belongs to the medium fuzzy subset, then the fuzzy risk probability ( ) is a medium-high fuzzy risk probability, and the fuzzy risk probability range is 0.5-0.8.

[0179] Pre-set fuzzy rule 9: If = High And = High Then = High (output probability range is 0.7-1.0);

[0180] That is, if the semantic feature confidence ( ) belongs to the high fuzzy subset, and the numerical feature fluctuation degree ( ) belongs to the high fuzzy subset, then the fuzzy risk probability ( ) is a high fuzzy risk probability, and the fuzzy risk probability range is 0.7-1.0.

[0181] Among them, the specific value of the output probability is determined by the ANFIS model through 1000 sets of labeled samples (containing semantic confidence, numerical fluctuation degree and true risk label) to determine the initial value, and then dynamically optimized through twin body simulation results.

[0182] For example, the fuzzy risk probability can be determined according to the following formula.

[0183]

[0184] Among them, is the fuzzy risk probability of multi-source data, is the weight coefficient of the semantic feature confidence, semantic feature confidence of image data (i.e. inspection data), weight coefficient of numerical feature fluctuation degree, numerical feature fluctuation degree, fuzzy inference correction coefficient, fuzzy inference deviation correction term.

[0185] It should be noted that the value range of the fuzzy risk probability is between 0 and 1, where 0 represents no risk and 1 represents a certain risk, which is the core quantitative index for judging whether the data has risk.

[0186] The weight coefficient of the semantic feature confidence is obtained by training the ANFIS model based on historical labeled data (such as insulator erosion image-real risk corresponding sample), and the value is between 0 and 1. For key semantic features in the inspection data (such as edge texture of insulator defects and morphological features of conductor broken strands), the coefficient will give a higher value to highlight the influence of key features on risk determination.

[0187] The semantic feature confidence represents the credibility of the model recognizing a certain type of defect semantics (such as insulator point erosion and hardware corrosion), and the value range is between 0 and 1. The higher the credibility, the closer the value is to 1.

[0188] The weight coefficient of the numerical feature fluctuation degree is obtained by training the ANFIS model, and the value is between 0 and 1. For core numerical indicators in the transmission operation data (such as insulator leakage current and conductor stress), the coefficient will give a higher value to emphasize the role of core operation data in risk determination.

[0189] The specific calculation method of the numerical feature fluctuation degree is "(current value-historical mean value) / historical standard deviation", and the value range is between 0 and 1. The larger the numerical fluctuation, the closer the value is to 1, which can intuitively reflect the degree of abnormality of the operation data.

[0190] The fuzzy inference correction coefficient is dynamically adjusted based on the simulation verification results of the first twin model (i.e. predicted operation data). When the simulation verification results of the first twin model have small deviation from the measured data (actual operation data), the fuzzy inference correction coefficient takes a value between 0.05 and 0.1. When the simulation verification results of the first twin model have large deviation from the measured data (actual operation data), the fuzzy inference correction coefficient takes a value between 0.2 and 0.3.

[0191] The fuzzy inference deviation correction term is output by the fuzzy rule base of the ANFIS model, and the specific calculation method is "absolute value difference between twin simulation risk trend and actual data trend", and the value range is between 0 and 1. The more consistent the two trends, the closer the value is to 0, which can effectively compensate for the inference deviation caused by the uncertainty of multi-source data.

[0192] In some examples, if the ambiguity risk probability is greater than a preset probability, a risk parameter is determined according to the first image data and the first actual operation data (i.e., multi-source data), wherein the risk parameter can include a defect position and a numerical anomaly amplitude. The risk parameter is input into the first twin model to simulate by the first twin model and output predicted operation data. If the predicted operation data matches the first actual operation data, it is determined that the multi-source data indicates a real risk and the first power transmission device has a real risk. Otherwise, there is no real risk.

[0193] For example, the predicted operation data is a simulated current (Isim) and a simulated temperature (Tsim), and the first actual operation data is an actual current (Imeas) and an actual temperature (Tmeas). If the relative error of the predicted operation data and the first actual operation data is less than a preset relative error (e.g., 5%) and the absolute error of the predicted operation data and the first actual operation data is less than a preset absolute error (e.g., 1%), it is determined that the multi-source data is a real risk, i.e., the first power transmission device has a real risk.

[0194] For example, the relative error: |(Isim - Imeas) / Imeas| < 5%, and the absolute error: |Tsim - Tmeas| < 1°C, it is determined that the risk is a real risk, otherwise, there is no real risk, i.e., it is considered as a false alarm.

[0195] In some embodiments, the matching degree of the ambiguity risk probability and the real risk is counted, and if the matching degree is less than a preset matching degree (e.g., 90%), the mean of the Gaussian function of the ambiguity subset in the membership function parameter is adjusted. For example, from 0.5 to 0.45.

[0196] In some embodiments, after it is determined that the multi-source data indicates a real risk and the first power transmission device has a real risk, the cloud calls the first twin model to perform risk diffusion simulation, i.e., inputs the risk parameter (such as ice thickness, wind speed, etc.) into the first twin model, simulates the risk change trend of the first power transmission device in a preset time period, and outputs a risk disposal strategy and a predicted implementation effect of the risk disposal strategy. According to the risk disposal strategy, a control instruction (i.e., an emergency instruction) is sent to the inspection device to dispose the risk.

[0197] For example, the risk parameter is ice thickness, wind speed, etc. The risk change trend of the first power transmission device in the preset time period is the change of the conductor stress when the ice thickness increases and the force condition of the adjacent tower. The risk disposal strategy is load adjustment, unmanned aerial vehicle ice melting, and inspection robot reinforcement.

[0198] Among them, the simulation parameter setting of risk diffusion pre-play: the time step is 10 minutes, and the simulation duration is 2 hours; the simulation formula of icing risk: icing thickness growth model H(t) = H0 x exp(kt), wherein H0 is the initial icing thickness mm, k is the growth coefficient (k = 0.02 when the wind speed is 10 m / s, k = 0.05 when the wind speed is 15 m / s), and t is the time min.

[0199] The evaluation index of risk disposal strategy includes: disposal cost (yuan), disposal time (min), risk elimination rate (%), and the optimal scheme is selected by using weighted scoring method (weight ratio 4:3:3). Example: scheme 1 (unmanned aerial vehicle deicing): score 85 points (cost 2000 yuan, time 15 min, elimination rate 98%); scheme 2 (load adjustment): score 78 points (cost 500 yuan, time 30 min, elimination rate 85%), finally scheme 1 is selected.

[0200] For example, according to the risk disposal strategy, sending a control instruction to the inspection equipment to dispose the risk can include: in a high-risk scenario (high-risk level), the cloud sends a control instruction to the inspection equipment (such as an inspection robot unmanned aerial vehicle and / or a regional power grid regulation terminal), and the inspection equipment (such as an inspection robot and / or an unmanned aerial vehicle) executes the control instruction.

[0201] For example, the inspection robot responds to the control instruction and lays monitoring sensors at the risk point. The unmanned aerial vehicle responds to the control instruction and performs deicing operation. The regional power grid regulation terminal responds to the instruction and issues disposal instructions for composite adjustment. And the pre-play of the first twin body model "deicing temperature requirement" and other key disposal nodes are pushed to each power transmission equipment in real time.

[0202] In this application, the "ANFIS+ digital twin simulation" double-layer inspection system, the core innovation is: the first layer of intelligent screening: through the ANFIS model to process data uncertainty, the semantic feature confidence and the numerical feature fluctuation are converted into fuzzy risk probability, realizing preliminary risk screening, overcoming the limitation of single threshold judgment. The second layer of physical verification and closed-loop optimization: using digital twin to simulate and verify the risk, comparing the simulation data with the measured data (such as current error <5%) to determine the authenticity of the risk. At the same time, the ANFIS model can automatically learn the simulation results, dynamically optimize its membership function and fuzzy rules, so as to continuously improve the processing accuracy of uncertain data.

[0203] In addition, the blockchain technology is used to record the emergency disposal results of the risk disposal strategy (i.e. emergency instruction). The risk inspection results (whether it is a real risk) and verification data (measured operation data and detection report) are input into the first twin body model for risk reproduction, and the risk full-process simulation data is output to verify the accuracy of the first twin body model.

[0204] After completing the risk treatment, the system uses the blockchain to record the emergency treatment results, and starts the "virtual-real closed loop verification and reverse feature mining" mechanism: first, the inspection results and on-site secondary verification data are input into the digital twin to reproduce the risk whole process evolution (such as the time sequence process of erosion area from 0.5mm² to 5mm²), to cross-verify the accuracy of ANFIS reasoning and twin simulation, and to generate complete risk time sequence samples; further, based on convolutional autoencoder (CAE), the time sequence data is mined reversely, the deep features are extracted by the encoder, the data is reconstructed by the decoder, and the features with reconstruction error exceeding the threshold are determined as implicit features that are difficult to capture by traditional methods, such as surface micro temperature difference (>0.8℃) or leakage current low frequency fluctuation (<0.2Hz) in the early stage of insulator erosion; finally, the implicit features mined are supplemented to the input layer of the semantic-physical mapping model, and the ANFIS fuzzy rules are updated (such as adding the rule of "low frequency fluctuation + micro temperature difference → early erosion"), to realize the continuous evolution of the model and the accurate identification of early implicit risks.

[0205] In some embodiments, the cloud builds a power transmission risk knowledge graph, which includes power transmission equipment, features (such as semantic feature vectors, numerical feature vectors, defect semantic features, numerical interval features), and schemes (risk treatment strategies). Based on the power transmission risk knowledge graph, the cloud edge scheduling, ANFIS rules, and twin parameters (i.e. physical model parameters of the twin model) are adjusted.

[0206] The cloud supports the overall optimization of the system by building a power transmission risk knowledge graph, which integrates multi-dimensional structured knowledge such as power transmission equipment, risk features (such as semantic feature vectors, numerical feature vectors, defect semantic features, numerical interval features), and treatment schemes (such as risk treatment strategies). The nodes cover equipment parameters such as tower type and conductor material, risk features such as semantics, numbers, and physics, treatment schemes such as unmanned aerial vehicle deicing and load adjustment, historical emergency cases, and model verification results. The edges between the nodes clearly express the association between elements (such as "composite material tower → icing risk → unmanned aerial vehicle deicing scheme"), providing unified knowledge support for system self-optimization. Based on this, the system introduces a reinforcement learning agent, aiming to optimize the accuracy of inspection, the shortest emergency response time, and the lowest cloud edge resource consumption. Through interaction with the knowledge graph, the system dynamically adjusts the operation mechanism of the three core modules: including optimizing the cloud edge resource scheduling strategy based on the historical risk frequency in the region (such as increasing the reasoning depth of the edge in high-risk areas), optimizing the ANFIS fuzzy rules based on new equipment defect cases, and real-time correcting the parameters of the digital twin based on equipment aging data. The system automatically completes reinforcement learning iteration at a preset period without human intervention, continuously adapts to new power transmission equipment and extreme working conditions, and feeds back the iteration results to the cloud twin database to form a closed loop evolution capability.

[0207] The defect detection method of the power transmission equipment provided by the embodiment of the application is schematically described below with a specific example.

[0208] As shown in Figure 3 The defect detection method of the power transmission equipment includes the following steps:

[0209] S801, a digital twin (i.e., the twin model above, or simply the twin) of the power transmission equipment (or physical equipment, such as the first power transmission equipment above) is constructed, and the digital twin is calibrated virtually and actually.

[0210] Specifically, the detection data of the power transmission equipment is obtained through the inspection equipment, and the detection data includes three-dimensional point cloud data and image data. The twin model is generated in combination with the three-dimensional point cloud data, image data and original equipment data (i.e., design drawings) of the power transmission equipment. The scale of the twin model to the power transmission equipment is 1:1.

[0211] The edge virtual-actual calibration algorithm is deployed to compare the simulation state (such as the predicted conductor sag value) of the digital twin (i.e., the twin model, or simply the twin) with the field actual data (such as the actual conductor sag value) in real time. For example, the conductor sag value calculated by the twin is compared with the LiDAR measured value, and the physical parameters of the twin are dynamically corrected, such as adjusting the conductor elastic coefficient, to ensure that the twin is highly consistent with the state of the physical equipment (i.e., the power transmission equipment).

[0212] In some embodiments, a twin identification (i.e., twin ID) is set for each twin, each detection data and actual operation data, and simulation operation data (i.e., multi-source data) output by the twin are obtained, a time-space stamp of the multi-source data is obtained, and the twin identification of the twin, the time-space stamp of the multi-source data and the multi-source data are stored in association to a distributed twin database. That is, through the association mode of "twin ID+time-space stamp", the multi-source data is integrated and uploaded to the cloud distributed twin database.

[0213] Among them, the twin ID coding rule: adopt 'device type-area code-device number' three-part structure, example: 'Tower-025-03' (Tower represents the tower, 025 is the area code, 03 is the 3rd base tower in this area), 'Wire-025-03-01' (Wire represents the wire, 025 is the area code, 03 is the associated tower number, 01 is the first wire in this tower); The space-time stamp accuracy is millisecond level (format: YYYY-MM-DDHH: MM: SS.fff), the association logic: through the edge data gateway, bind the point cloud data (including LiDAR device ID), running data (including sensor ID) and twin ID under the same space-time stamp (error ≤100ms), generate an index table of 'ID-space-time stamp-data type', ensure that the cloud distributed twin database can locate all associated data of a twin within 1 second through the index table.

[0214] S802, simulate different defect scenarios using digital twins, train attention mechanism driven bidirectional mapping model, and use bidirectional mapping model (semantic-physical mapping system) to obtain semantic-numeric integrated data (i.e. target mapping data).

[0215] When implementing S802, the following steps are included:

[0216] S8021, based on the cloud distributed twin database, build a dynamic semantic-physical collaborative mapping system, i.e. bidirectional mapping model (i.e. attention mechanism driven bidirectional mapping model).

[0217] For example, using digital twins, simulate different defect scenarios such as gradual change of insulator point erosion area, and simultaneously calculate the change law of numerical values such as leakage current and temperature (i.e. simulation running data) under each scenario. Using the simulation defect image and simulation running data, train the bidirectional mapping model, which can automatically focus on key semantic features such as erosion edge texture and numerical indicators such as pulse type fluctuation peak, and output a mapping table that updates in real time with the state of the twin, i.e. target mapping data.

[0218] The specific structure of the attention mechanism-driven bidirectional mapping model is as follows: the input layer includes a semantic feature branch (128 dimensions, including the eroded edge texture and defect morphology features extracted by a lightweight CNN) and a numerical feature branch (4 dimensions, including the current, temperature, humidity, and wind speed), the feature interaction is realized through a cross-attention layer (the attention weight calculation adopts a Scaled Dot-Product mechanism), the hidden layer is provided with two full connection layers (each layer has 64 neurons, and the activation function is ReLU), and the output layer is a 10-dimensional mapping vector corresponding to the semantic-numerical correlation interval of five defect scenarios, i.e., "insulator point erosion area 0-1 mm²→leakage current 5-8 mA", "insulator point erosion area 1-3 mm²→leakage current 8-15 mA", "insulator point erosion area 3-5 mm²→leakage current 15-20 mA", and the like.

[0219] The dynamic sample library (i.e., the distributed twin database) is updated once a day, and 50 groups of defect data simulated by the twins are added each time (including the three core defects of insulator erosion, hardware corrosion, and conductor breakage).

[0220] S8022, the physical mechanism real-time correction layer is embedded in the twin, and when the environmental parameters of the power transmission equipment change, the physical model parameters of the twin are updated.

[0221] For example, when the environmental parameters such as wind speed change, the twin automatically updates the mechanical model of conductor wind deflection stress calculation and synchronously adjusts the mapping relationship between semantics and values.

[0222] The mechanical model update logic of the physical mechanism real-time correction layer is as follows: when the wind speed sensor detects that the wind speed is greater than 10 m / s, the twin automatically calls the conductor wind deflection stress calculation model. The conductor wind deflection stress calculation model can be expressed as: F=0.613×ρ×v²×d×L, where ρ is the air density kg / m³, v is the wind speed m / s, d is the conductor diameter m, and L is the span m.

[0223] The semantic-numerical mapping relationship (i.e., the semantic-physical mapping system) is synchronously adjusted. For example, under the same insulator erosion area, the leakage current threshold at a wind speed of 15 m / s is increased by 12% compared with that at a wind speed of 5 m / s; the environmental parameter monitoring frequency is once every minute, and when the parameter change amplitude is greater than 20%, the real-time correction is triggered, otherwise the batch correction is performed every 30 minutes.

[0224] The correction logic of non-wind speed environmental parameters is as follows:

[0225] 1. Humidity correction: When the humidity sensor detects that the relative humidity > 85%, trigger the insulator leakage current correction model, the correction formula: Icorr = Imeas x (1 + 0.02 x (RH - 85%)), where RH is the relative humidity, Imeas is the measured leakage current, Icorr is the corrected current. Adjust the "insulator erosion area-leakage current" mapping relationship synchronously, such as when the humidity is 90%, the current threshold of the same erosion area is increased by 8% compared to when the humidity is 80%.

[0226] 2. Temperature drop correction: When the temperature change rate < -2℃ / h, trigger the wire thermal expansion and contraction model, the correction formula: Lcorr = L0 x (1 + αT x ΔT), where L0 is the span of the wire at room temperature, αT is the thermal expansion coefficient of the wire, the aluminum strand is 23 x 10^-6 / ℃, and ΔT is the temperature change. And update the mapping threshold of semantic features (such as wire sag) and numerical features (such as span); the correction priority of various environmental parameters: temperature drop (priority 1, immediate correction) > wind speed > 8m / s (priority 2, corrected within 5 seconds) > humidity > 85% (priority 3, corrected within 30 seconds).

[0227] S803, using the federated sub-model, semantic-numerical integrated data, and the current detection data (such as image data, i.e. video data) and actual operation data (such as leakage current, wire temperature, humidity, wind speed) of the first power transmission equipment to perform basic risk identification, and obtain the basic risk label of the first power transmission equipment (i.e. the target defect risk information in the foregoing).

[0228] Specifically, a "federated learning driven cloud-edge hierarchical reasoning architecture" is designed, and a localized federated sub-model is deployed at the edge of the inspection robot (i.e. the inspection equipment), the regional sensor gateway, etc. The edge can independently complete the basic risk label of insulator obvious defects, current exceeding standard, etc. based on local data.

[0229] Among them, the specific configuration of the localized federated sub-model: the sub-model deployed at the edge is a MobileNetV2 lightweight network, the input is a 640x640 resolution inspection video and 4-dimensional numerical data, and the output is 4 types of basic risk labels: no defect, insulator obvious defect, current exceeding standard, and temperature anomaly.

[0230] The model parameter update gradient (not the original data) of the federated sub-model is uploaded to the cloud to avoid the problem of power transmission data privacy leakage. After the cloud federated server (i.e. the server in the foregoing) receives the model parameter update gradient of each edge node (i.e. the edge or power transmission equipment), it aggregates and optimizes the global risk identification model, and then distributes the updated model parameters (i.e. the global model parameters in the foregoing) to the edge to complete iteration.

[0231] The transmission rule of the model parameter update gradient is: only the weight gradient of the convolution layer and the fully connected layer (about 200 KB / time) is uploaded, the original image and numerical data are not transmitted, the gradient encryption adopts the RSA asymmetric encryption algorithm, the key is dynamically generated by the cloud federal server and updated once every 24 hours, and the original data is prevented from being inversely calculated from the gradient.

[0232] The cloud federal parameter aggregation adopts the weighted federal average (FedAvg-W) algorithm, and the core logic is:

[0233] The weight distribution rule is that the edge node weight ωi=ni / N (ni is the local training sample amount of the ith edge node, and N is the sum of the total sample amount of all edge nodes), so as to avoid that the node with small data amount (such as the inspection robot with sample amount <100) excessively affects the global model.

[0234] The aggregation calculation step is GlobalW=Σ(ωi×LocalWi) (GlobalW is the updated global model weight, and LocalWi is the parameter gradient updated weight of the ith edge node).

[0235] The aggregation constraint is that when the gradient variance of a certain edge node is greater than 0.1 (the gradient difference with other nodes is too large), the weight of the edge node is temporarily reduced to ωi×0.5, so as to prevent the global model from being shaken by abnormal gradient. The aggregation frequency matches the iteration number of the edge end, the edge end uploads the gradient once every 5 rounds (100 samples per round), and the cloud receives all node gradients and completes aggregation and delivery within 10 seconds.

[0236] Based on the semantic-numerical integrated data, a federal learning cloud edge architecture is designed: the edge end uses the deployed localized federal sub-model to perform basic risk identification, and transmits the parameter gradient instead of the original data to protect privacy; the cloud federal server aggregates the parameter gradient, optimizes the global risk identification model, introduces an edge intelligent hierarchical reasoning mechanism, uses a lightweight model for low risk, and triggers cooperative reasoning to preferentially transmit twin abnormal data for high risk.

[0237] S804, using the ANFIS model, determining a fuzzy risk probability of the multi-source data according to the basic risk label, the current detection data and the actual operation data (i.e. multi-source data) of the first power transmission equipment, and performing real risk verification (i.e. risk inspection result) based on the fuzzy risk probability.

[0238] The detection data is, for example, image data, i.e. image data. The actual operation data is, for example, leakage current, conductor temperature, humidity and wind speed.

[0239] In some examples, determining the fuzzy risk probability of the multi-source data can include: performing feature extraction on the first image data to obtain a semantic feature vector, determining a semantic feature confidence of the semantic feature vector; performing feature extraction on the first actual operation data (such as leakage current and / or conductor temperature) to obtain a numerical feature vector, and determining a numerical feature fluctuation degree (i.e. current fluctuation range) according to the numerical feature vector. The fuzzy risk probability of the multi-source data is determined by using a preset ANFIS model, the semantic feature confidence, and the numerical feature fluctuation degree (i.e. current fluctuation range).

[0240] The ANFIS model has input variables of a semantic feature confidence ( ) and a numerical feature fluctuation degree ( ), wherein each input variable can be divided into three fuzzy subsets (i.e. low, medium, and high), and each fuzzy subset is generated by using a Gaussian function. For example, the high fuzzy subset uses a Gaussian function as:

[0241] .

[0242] The fuzzy risk probability is determined according to a preset fuzzy rule.

[0243] For example, the preset fuzzy rule is as described above, and will not be described here again.

[0244] S805, after verifying the existence of a real risk, the cloud calls the twin to perform risk diffusion pre-performance, that is, inputs risk parameters (such as ice thickness, wind speed, etc.) into the twin to simulate the risk change trend of the power transmission equipment within a preset time period, outputs a risk disposal strategy, and predicts the implementation effect of the risk disposal strategy, and according to the risk disposal strategy, sends a control instruction (i.e. emergency instruction) to the inspection equipment to perform risk disposal.

[0245] Specifically, based on the real risk determination result, the “twin pre-performance-unmanned cooperation” closed-loop emergency scheme is started in a high-risk scenario. The cloud first calls the digital twin to start “risk diffusion pre-performance”: inputs current risk parameters such as ice thickness and wind speed, simulates the risk development trend in the future short time, such as conductor stress change when ice thickness increases, adjacent tower stress condition, and then outputs multiple optimal disposal schemes (such as load adjustment, unmanned aerial vehicle ice melting, and inspection robot reinforcement), and synchronously predicts the implementation effect of each scheme.

[0246] Simulation parameter settings of risk diffusion pre-emption: time step is 10 minutes, simulation duration is 2 hours; simulation formula of icing risk: icing thickness growth model H(t) = H0 x exp(kt), wherein H0 is initial icing thickness mm, k is growth coefficient (k = 0.02 when wind speed is 10 m / s, k = 0.05 when wind speed is 15 m / s), t is time min; evaluation indexes of treatment scheme: including treatment cost (yuan), treatment time-consuming (min), risk elimination rate (%), the optimal scheme is selected by using weighted scoring method (weight ratio 4:3:3), example: scheme 1 (unmanned aerial vehicle de-icing) scores 85 points (cost 2000 yuan, time-consuming 15 min, elimination rate 98%), scheme 2 (load adjustment) scores 78 points (cost 500 yuan, time-consuming 30 min, elimination rate 85%), finally scheme 1 is selected.

[0247] S806, using a blockchain technology, record the emergency treatment result of the risk treatment strategy (i.e. emergency instruction). The risk inspection result (verify whether it is a real risk), verification data (measured operation data and detection report) are input into the twin to reproduce the risk, and the risk full-process simulation data is output to verify the accuracy of the twin model.

[0248] Specifically, the cloud end synchronously issues treatment instructions to the automatic driving inspection robot (instructs it to go to the risk point to lay monitoring sensors), the unmanned aerial vehicle (instructs it to perform de-icing operation), and the regional power grid regulation terminal (instructs it to perform load adjustment) through the "unmanned equipment cooperative scheduling protocol", and pushes the key treatment nodes such as "de-icing temperature requirement" pre-empted by the twin to each device in real time. At the same time, the blockchain technology is used to record the emergency instruction and the execution result of the equipment, so as to ensure that the instruction is not tamperable, the process is traceable, and the response efficiency and treatment accuracy are optimized.

[0249] Blockchain data query permission division: cloud end master node (administrator permission): can query all on-chain data (including instruction plaintext, execution result, node signature), supports data export and audit; regional scheduling node (operation permission): can only query the instructions and execution results in the jurisdiction area (hide the instruction encryption field); unmanned equipment control node (read-only permission): can only query the instructions and historical execution results of its own equipment, and cannot query other equipment data; the permission verification adopts a public key-private key mechanism, and the node needs to obtain a special key through cloud end CA authentication; data storage period: emergency instruction and execution result: permanent storage (for historical case tracing and knowledge graph updating); temporary interaction data (such as equipment status heartbeat packet): stored for 3 months and then automatically archived to the cloud end cold backup database (in compressed format, compression ratio 20:1), after archiving, the query can be restored by administrator permission.

[0250] S806 can specifically include:

[0251] S8061, Virtual-real interaction verification model accuracy: Based on the emergency disposal results, a virtual-real interaction verification mechanism is designed. The inspection results (i.e. risk inspection results) and on-site verification data (such as robot secondary inspection images, manual detection reports) are input into the digital twin. Through the twin, the whole process of risk from occurrence to development (such as the process of gradual expansion of insulator erosion area) is reproduced, so as to verify the reasoning accuracy of the ANFIS model and the simulation accuracy of the digital twin, and to check the deviation problems in the model or simulation.

[0252] The specific process of virtual-real interaction verification: input data: inspection results (such as insulator 25#-C phase point erosion, Prisk=0.92), on-site verification data (robot secondary inspection image (resolution 1920x1080), manual detection report (erosion area 3.2mm²)); twin reproduction process: load 25# tower initial twin model→ input erosion parameters verified on site→ simulate risk development for 0-48 hours (erosion area increases from 3.2mm² to 5.1mm²); model accuracy verification index: ANFIS reasoning accuracy = (number of correctly identified risks / total number of risks) x 100%, twin simulation accuracy = (1- |simulation parameter- measured parameter| / measured parameter) x 100%, both indicators ≥ 90%.

[0253] S8062, Reverse mining of implicit features and model updating: Start the reverse feature mining algorithm, based on the complete risk data reproduced by the digital twin, reverse mine the early risk implicit features that are difficult to capture by traditional methods, such as the small temperature difference on the surface of the insulator in the early stage of erosion, and the low frequency fluctuation of leakage current in the corresponding power transmission operation data. These implicit features are supplemented to the semantic-physical mapping model (i.e. semantic-physical mapping system), and the ANFIS fuzzy rules are updated (such as adding the reasoning rule of "low frequency current fluctuation + small temperature difference → early erosion risk"), which extends the risk identification lead time.

[0254] Specific implementation of reverse feature mining algorithm: Convolutional Autoencoder (CAE) is used, the input is the time series data (temperature, current, sampling frequency 1Hz, time length 24 hours) and image data (1 frame per hour, a total of 24 frames) reproduced by the twin, the encoder contains 3 convolutional layers (convolution kernel 3x3, step 1), and the decoder contains 3 deconvolutional layers. Implicit features are mined by minimizing reconstruction error (MSE); Quantitative threshold of implicit features: when the insulator surface temperature difference > 0.8℃ and the leakage current low frequency fluctuation < 0.2Hz, it is determined as early erosion implicit feature; Model updating steps: convert the implicit features into a 16-dimensional feature vector and supplement them to the input layer of the semantic-physical mapping model, and add a new rule to the ANFIS rule base: If low frequency current fluctuation = low And surface temperature difference = medium Then early erosion risk = medium.

[0255] Model training logic after implicit feature supplement: semantic-physical mapping model: fine-tuning training, only freeze the bottom 30% of the convolutional layer parameters of the model, train the upper fully connected layer (including the newly added 16-dimensional implicit feature input layer), the learning rate is set to 1 / 10 of the initial training (i.e. 0.0001), to avoid parameter shock; Training data: use a mixed set of old samples + new implicit feature samples (8000 old samples, 200 new implicit feature samples, ratio 40:1); Termination condition: stop training when the model's mean square error (MSE) < 0.005 and the MSE change amount is < 0.0001 for 5 consecutive iterations; ANFIS rule update training: after adding new rules, use 500 labeled data containing implicit features (such as "low frequency current fluctuation + slight temperature difference" corresponding to early corrosion samples) to incrementally train ANFIS, train for 20 rounds, and terminate when the model prediction accuracy is > 92%, ensuring that the new rules are effectively integrated into the model.

[0256] S807, the cloud builds a power transmission risk knowledge graph, which includes power transmission equipment, features (such as semantic feature vectors, numerical feature vectors, defect semantic features, numerical interval features), and solutions (risk disposal strategies). Based on the power transmission risk knowledge graph, cloud-edge scheduling, ANFIS rules, and twin parameters (i.e. physical model parameters of the twin model).

[0257] S8071, power transmission risk knowledge graph construction: based on implicit features and model update, build a full-link self-optimization system of knowledge graph and reinforcement learning. The cloud first builds a power transmission risk knowledge graph, which includes device parameters such as tower type and conductor material, risk features such as semantics, numbers, and physics, various disposal solutions, and historical cases such as emergency cases and verification results. The edges between nodes represent the association between elements. For example, "a specific type of conductor → icing risk → corresponding de-icing solution", forming a structured knowledge reserve.

[0258] Node attribute details of power transmission risk knowledge graph: Equipment parameter node: tower model (such as ZG200), conductor material (such as aluminum conductor), installation time (such as 2020-01-15), design life (such as 30 years) and other attributes. Risk feature node: feature type (semantic or numerical or physical), feature threshold (such as current fluctuation > 30%), associated defects (such as icing) and other attributes. Disposal scheme node: scheme name (such as unmanned aerial vehicle ice melting), applicable scene (icing thickness 5-10 mm), execution condition (wind speed < 8 m / s) and other attributes. Historical case node: case time (such as 2024-03-20), risk level (5 levels), disposal result (risk elimination rate 98%) and other attributes. Node association relationship example: ZG200 tower → icing risk → unmanned aerial vehicle ice melting scheme, aluminum conductor → strand breakage risk → inspection robot reinforcement scheme.

[0259] Knowledge graph storage and update mechanism: Storage database: Use open source Neo4j graph database, deployment architecture is 1 master 2 slave cluster (master node is responsible for reading and writing, slave node is responsible for backup and query shunting), single node storage capacity supports 1 million + nodes and 5 million + associated edges, query response time < 500 ms.

[0260] Real-time update trigger condition: Update the graph immediately when the following scenarios occur: a) Add new emergency cases (within 10 minutes after disposal is completed), automatically create historical case nodes, and associate corresponding equipment parameter-risk feature-disposal scheme nodes; b) Discover new defects (such as composite material tower cracking), add risk feature nodes and associated equipment parameter-disposal scheme edges; c) Update verification: After real-time update, the cloud knowledge graph verification module (based on rule engine) checks the node attribute integrity (such as case nodes need to include risk level and disposal result) and association rationality (such as aluminum conductor cannot be associated with "composite material tower exclusive disposal scheme"), and triggers manual audit process if verification fails.

[0261] S8072, reinforcement learning driven full link optimization: Introduce reinforcement learning agent, with "highest inspection accuracy, shortest emergency response time, and lowest cloud edge resource consumption" as reward function, agent interacts with knowledge graph to dynamically optimize three core modules. Among them, one is to optimize cloud edge resource scheduling strategy (such as adjusting edge model reasoning depth according to historical risk frequency of a certain area in knowledge graph); two is to optimize ANFIS model fuzzy rules (such as optimizing membership function combined with new insulator defect cases in knowledge graph); three is to optimize digital twin parameters (such as correcting twin life decay model according to equipment aging data in knowledge graph); the system automatically completes reinforcement learning iteration regularly without human intervention, which can adapt to new equipment and extreme scenarios, and the iteration results are fed back to the twin database.

[0262] Training details of the reinforcement learning agent: State space: contains cloud-edge resource utilization (CPU / memory), inspection accuracy, response time 3 types of state variables; Action space: contains adjusting cloud-edge scheduling frequency (5-30 min), ANFIS membership function parameters, twin body calibration period (5-15 min) 3 types of actions; Training process: PPO algorithm is adopted, training batch is 100 batches, sample size of each batch is 500 historical cases, learning rate is 0.001; Iteration period and feedback logic: iteration is automatically started at 2 o'clock every Monday morning, after iteration, the optimized cloud-edge scheduling strategy (such as the area with historical risk frequency > 5 times / month, the edge-end reasoning depth is increased by 30%), ANFIS rule, and twin body parameter are written into the cloud twin database, and the old parameters are overwritten.

[0263] In the embodiment of the application, cross-modal risk accurate identification based on physical mechanism is realized. Through digital twin driven cross-modal semantic-physical dynamic mapping, the defect semantic features extracted in the inspection image are quantitatively associated with the numerical interval features in the operation data, and the first twin body model is used to physically verify the risk authenticity. This method effectively overcomes the misjudgment problem caused by the semantic gap and the lack of physical mechanism in the traditional scheme, can accurately identify early hidden risks such as small cracks in insulators and low-frequency fluctuations of leakage current, and significantly improves the accuracy and reliability of the risk inspection results.

[0264] A responsive and privacy-safe hierarchical reasoning architecture is constructed. Through the collaborative design of federated learning and edge hierarchical reasoning, the system can dynamically adjust the calculation strategy according to the risk level: in the low-risk scenario, light-weight feature extraction is performed on the edge, and in the high-risk scenario, the collaborative reasoning mechanism of "federated model + twin body local simulation" is triggered. This architecture not only solves the response delay and resource waste problem under the traditional fixed calculation allocation mode, but also realizes the unity of real-time response, accurate analysis and data privacy protection through federated learning, which only interacts with gradient parameters in the model training process without uploading original data.

[0265] A full-link self-optimization system with continuous evolution capability is established. The system integrates emergency cases and implicit features through reverse mining based on a knowledge graph, and introduces a reinforcement learning agent to dynamically optimize cloud-edge scheduling strategies, ANFIS fuzzy rules and digital twin parameters. This self-optimization mechanism enables the system to periodically iterate the model without human intervention, adapt to the inspection needs of new power transmission equipment and extreme working conditions, and fundamentally solves the pain points of traditional schemes that rely on manual updates and are difficult to adapt to dynamic changes, significantly reducing operation and maintenance costs and continuously improving system performance.

[0266] The embodiment of the application also provides a defect detection device for a power transmission equipment. As shown in Figure 4As shown, the defect detection device 400 of the power transmission equipment includes an analog module 401, a data processing module 402, a generation module 403, a training module 404, and a determination module 405. Among them, the analog module 401 is configured to simulate a plurality of defect scenarios by using a first twin model of a first power transmission equipment, and generate a simulation defect image and simulation running data of each defect scenario. The data processing module 402 is configured to process the simulation defect image and the simulation running data of each defect scenario by using a bidirectional mapping model, to obtain target mapping data, the target mapping data being used to represent a corresponding relationship between defect semantic features and numerical interval features under different defect scenarios. The generation module 403 is configured to generate a plurality of first training samples based on the target mapping data, each first training sample including a training defect image, training running data, and actual defect risk information. The training module 404 is configured to train an initial federal sub-model by using the plurality of first training samples, to obtain a first federal sub-model. The determination module 405 is configured to determine target defect risk information corresponding to the first power transmission equipment according to the first federal sub-model, first image data of the first power transmission equipment, and first actual running data.

[0267] The electronic device provided in this embodiment can execute the method provided in the method embodiment, and has similar implementation principles and technical effects. Details are not described herein again.

[0268] Figure 5 A structural schematic diagram of an electronic device provided in this application is shown in FIG. 1. Figure 5 As shown, the electronic device 500 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the electronic device 500 further includes a communication component 503. The processor 501, the memory 502, and the communication component 503 are connected through a bus 504.

[0269] In the specific implementation process, the processor 501 executes the computer execution instructions stored in the memory 502, so that the processor 501 executes the method described above.

[0270] The specific implementation process of the processor 501 can refer to the method embodiment described above, and has similar implementation principles and technical effects. Details are not described herein again.

[0271] This application also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the method described above is implemented.

[0272] This application also provides a computer readable storage medium, which stores computer execution instructions. When the processor executes the computer execution instructions, the method described above is implemented.

Claims

1. A method for detecting defects of a power transmission device, characterized by, The method applied to a first power transmission device comprises: simulating multiple defect scenarios by using a first twin body model of the first power transmission device, and generating simulation defect images and simulation operation data of each defect scenario; processing the simulation defect images and the simulation operation data of each defect scenario by a bidirectional mapping model to obtain target mapping data, the target mapping data being used to represent a corresponding relationship between defect semantic features and numerical interval features under different defect scenarios; generating multiple groups of first training samples based on the target mapping data, each group of first training samples comprising a training defect image, training operation data and actual defect risk information; training an initial federal sub-model by using the multiple groups of first training samples to obtain a first federal sub-model; determining target defect risk information corresponding to the first power transmission device according to the first federal sub-model, current first image data and first actual operation data of the first power transmission device.

2. The method of claim 1, wherein, The processing of the simulation defect images and the simulation operation data of each defect scenario by the bidirectional mapping model to obtain the target mapping data comprises: extracting features of the simulation defect images and the simulation operation data of each defect scenario by the bidirectional mapping model to obtain semantic feature vectors corresponding to each simulation defect image and numerical feature vectors corresponding to each simulation operation data; performing feature interaction processing on the semantic feature vectors corresponding to each simulation defect image and the numerical feature vectors corresponding to each simulation operation data to generate the target mapping data.

3. The method of claim 1, wherein, The method further comprises: obtaining environmental parameters of an environment where the first power transmission device is located, wherein the environmental parameters comprise at least one of wind speed, environmental humidity and environmental temperature; adjusting the first twin body model according to the environmental parameters to obtain an adjusted first twin body model; The simulation of multiple defect scenarios by using the first twin body model of the first power transmission device and the generation of simulation defect images and simulation operation data of each defect scenario comprise: simulating multiple defect scenarios by using the adjusted first twin body model and generating simulation defect images and simulation operation data of each defect scenario.

4. The method of claim 3, wherein, The adjustment of the first twin body model according to the environmental parameters to obtain the adjusted first twin body model comprises: if the wind speed is greater than a preset wind speed threshold, updating a conductor wind deflection force calculation model in the first twin body model; if the environmental humidity is greater than a preset humidity threshold, updating an insulator leakage current correction model in the first twin body model; if a change rate of the environmental temperature is less than a preset temperature change rate, updating a conductor thermal expansion and contraction model in the first twin body model, the preset temperature change rate being a negative value.

5. The method of claim 4, wherein, The adjustment of the first twin body model according to the environmental parameters to obtain the adjusted first twin body model comprises: if multiple parameters in the environmental parameters satisfy respective trigger conditions, updating corresponding physical models in the first twin body model in sequence according to priorities of the respective trigger conditions corresponding to the multiple parameters in the environmental parameters. The triggering condition corresponding to the ambient temperature is that a change rate of the ambient temperature is less than a preset temperature change rate, the triggering condition corresponding to the wind speed is that the wind speed is greater than a preset wind speed threshold, and the triggering condition corresponding to the ambient humidity is that the ambient humidity is greater than a preset humidity threshold. The priority of the triggering condition corresponding to the ambient temperature is higher than the priority of the triggering condition corresponding to the wind speed, and the priority of the triggering condition corresponding to the wind speed is higher than the priority of the triggering condition corresponding to the ambient humidity.

6. The method of claim 3, wherein, The method further comprises: obtaining second actual operation data and an actual conductor sag value of the first power transmission equipment at present; determining a predicted conductor sag value corresponding to the second actual operation data by using the first twin model; comparing the predicted conductor sag value with the actual conductor sag value to obtain a comparison result; based on the comparison result, correcting the first twin model to obtain a corrected first twin model.

7. The method of claim 1, wherein, The method further comprises: determining a target defect level of the first power transmission equipment according to the first image data, the first actual operation data and the target defect risk information; generating defect data of the first power transmission equipment by using different strategies according to the target defect level.

8. The method of claim 7, wherein, The generating of the defect data of the first power transmission equipment by using different strategies according to the target defect level comprises: if the target defect level is less than a first preset defect level, identifying the first image data to determine a defect region and a non-defect region, determining a semantic defect label and position information of the defect region based on the first image data, compressing the non-defect region in the first image data to obtain compressed first image data, and generating the defect data of the first power transmission equipment according to the semantic defect label, the position information of the defect region and the compressed first image data; if the target defect level is greater than or equal to the first preset defect level, performing local simulation on the first power transmission equipment based on the first actual operation data by using the first twin model to generate defect anomaly data.

9. A defect detection apparatus of a power transmission device, characterized by comprising: comprise: a simulation module configured to simulate a plurality of defect scenarios by using a first twin model of a first power transmission equipment, and generate simulation defect images and simulation operation data of each defect scenario; a data processing module configured to process the simulation defect images and the simulation operation data of each defect scenario by using a bidirectional mapping model to obtain target mapping data, the target mapping data being used to represent a corresponding relationship between defect semantic features and numerical interval features in different defect scenarios; a generation module configured to generate a plurality of first training samples based on the target mapping data, each first training sample comprising a training defect image, training operation data and actual defect risk information; a training module configured to train an initial federated sub-model by using the plurality of first training samples to obtain a first federated sub-model. A determining module is configured to determine target defect risk information corresponding to the first power transmission device according to the first federal sub-model, the first image data and the first actual operation data of the first power transmission device.

10. An electronic device, comprising: Comprise: A memory, a processor; The memory stores computer execution instructions; The processor executes the computer execution instructions stored in the memory, so that the processor executes the method in any one of claims 1-8.

Images