Single-phase earth fault unmanned aerial vehicle detection method, system, device and medium

Abstract

The invention discloses a single-phase earth fault unmanned aerial vehicle detection method, system and device and a medium, and belongs to the technical field of power system fault detection, and the method comprises the steps: obtaining multi-spectral visual data and positioning association data of a target inspection region, and constructing a multi-dimensional environment state space model; double-flow heterogeneous feature extraction is executed through the multi-dimensional environment state space model, and multi-scale fusion visual representation is obtained; and obtaining a space-time environment semantic mapping vector for first feature dimension calibration of the feature extraction network by combining multi-scale fusion visual representation with environment information analyzed by positioning associated data. According to the method, the internal coupling relation between the space-time environment semantic vector and the multi-scale visual representation is deeply mined by means of cross-modal information fusion, fault related characteristics such as electric arcs and ablation traces are more distinct in the data level, and the reliability degree of fault preliminary screening in complex scenes such as mountainous areas and thunderstorm weather is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system fault detection technology, specifically to a method, system, equipment, and medium for detecting single-phase grounding faults using a drone. Background Technology

[0002] In the entire cycle of normalized operation of the power system, single-phase grounding faults are a typical fault with a relatively high frequency of occurrence. The ability to detect and accurately locate the fault point in a timely and effective manner is crucial. Timely and effective detection means finding the fault as soon as possible, and accurate location means finding the exact location of the fault. This is of irreplaceable value for maintaining the stability of power supply and reducing various economic and functional losses caused by power outages.

[0003] Currently, the mainstream method used in the industry for detecting single-phase grounding faults is still the traditional manual detection mode. The specific operation process is as follows: Relevant personnel first need to make a preliminary judgment on the approximate range of the possible fault based on the information data transmitted by the dispatch center or the alarm signal triggered by the fault. The information data transmitted by the dispatch center refers to the relevant operation messages issued by the power dispatch center, and the alarm signal triggered by the fault is the alarm prompt automatically issued by the system after the fault occurs. Then, they drive a vehicle carrying detection instruments and equipment to the designated area. After arriving at the site, they need to set up a temporary operation platform based on the actual conditions of the transmission line's erection height and the site's terrain, or directly use an extended insulated operating rod to gradually approach the transmission line to carry out manual detection work. The extended insulated operating rod is a longer operating rod that can isolate high-voltage current. The manual detection work relies on the operator to operate the equipment by hand to carry out the inspection.

[0004] This traditional manual inspection method has many significant limitations: the entire inspection process involves multiple related steps, such as troubleshooting and sorting out the fault area, transporting and transferring inspection equipment, and setting up and preparing on-site working conditions. The operation steps are complicated and the overall time consumption is relatively long. Transporting and transferring inspection equipment means moving the inspection instruments from the warehouse or vehicle to the work site. Setting up and preparing on-site working conditions means setting up the platform, tools, etc. needed for the work on site. Especially in areas with complex terrain conditions such as high mountains, deep valleys and ravines, the work vehicles often have difficulty passing smoothly, and the staff have to travel back and forth on foot. This undoubtedly further prolongs the time cycle of the inspection operation, ultimately resulting in a low overall efficiency of fault detection and difficulty in quickly locating the fault point.

[0005] In the current technological application scenarios, the detection of single-phase grounding faults still mainly relies on manual operation. This situation not only results in consistently low efficiency of the detection work, but also presents a prominent problem of difficulty in effectively guaranteeing the accuracy of the detection results. The inability to effectively guarantee the accuracy of the detection results means that the judgment result cannot be guaranteed to be correct. This situation can no longer meet the actual application needs of the power system for rapid response and accurate location of faults. Summary of the Invention

[0006] To address the aforementioned technical challenges, a method for detecting single-phase grounding faults using a UAV is proposed. This method includes: acquiring multi-spectral visual data and location-related data of the target inspection area, and constructing a multi-dimensional environmental state space model; performing dual-stream heterogeneous feature extraction through the multi-dimensional environmental state space model to obtain a multi-scale fused visual representation; combining the multi-scale fused visual representation with environmental information parsed from the location-related data to obtain a spatiotemporal environmental semantic mapping vector, which is used for the first feature dimension calibration of the feature extraction network; analyzing the coupling relationship between the spatiotemporal environmental semantic mapping vector and the multi-scale fused visual representation based on the first feature dimension calibration to obtain a semantic-scale dual-drive attention threshold, and adjusting the second feature dimension of feature fusion; obtaining a target fault probability prediction model based on the adjusted second feature dimension distribution and a channel adaptive weighting mechanism; and obtaining a current feedback-based truth verification model using the target fault probability prediction model and contact signal feedback to lock in the conductor fault.

[0007] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the method includes: performing dual-stream heterogeneous feature extraction through a multi-dimensional environmental state space model to obtain a multi-scale fused visual representation, including: extracting heterogeneous spectral features based on the multi-dimensional state space and splicing them to generate a hybrid data stream; extracting multi-level local and global features in parallel through a convolutional processing unit and a self-attention processing unit based on the hybrid data stream; and combining the multi-level local and global features to splice the feature sequences to obtain a multi-scale fused visual representation.

[0008] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the method includes: obtaining a spatiotemporal environmental semantic mapping vector by combining multi-scale fusion visual representation with environmental information parsed from positioning association data, including: obtaining geographic and weather text descriptions based on the UAV's real-time positioning coordinates; constructing an environmental text set information flow by performing high-dimensional encoding through text descriptions combined with word vector embedding algorithms; extracting local and global text vectors representing micro and macro states based on the environmental text set information flow; and generating a spatiotemporal environmental semantic mapping vector by concatenating the local and global text vectors.

[0009] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the following steps are included: analyzing the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fusion visual representation to obtain the semantic-scale dual-drive attention threshold, and adjusting the second feature dimension of feature fusion: parsing multi-level image feature tensors based on multi-scale fusion visual representation; mapping the spatiotemporal environment semantic mapping vector to the corresponding feature space based on the multi-level image feature tensor to construct a cross-modal interaction basis; calculating the semantic attention weight to the feature tensor through the cross-modal interaction basis; and performing weighted fusion based on the attention weight to generate fused features constrained by the semantic-scale dual-drive attention threshold.

[0010] The beneficial effects of this preferred technical solution are as follows: it aligns with the semantic relationship between the spatiotemporal environment and the multi-level features of the image, highlighting the relevant features of single-phase grounding faults, reducing interference from environmentally irrelevant information, and making subsequent detection more aligned with the feature extraction needs of actual scenarios; by integrating semantic and visual information through cross-modal interaction, it enriches the representational dimensions of the fused features, making it easier to capture subtle features of faults in complex scenarios and improving the feature's ability to distinguish faults; relying on semantic-scale dual-drive constraints to regulate the feature fusion process, it reduces the mixing of invalid features, making the fused features more targeted and effective, and helping fault detection to align with actual working conditions.

[0011] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the target fault probability prediction model obtained based on the adjusted second feature dimension distribution and combined with the channel adaptive weighting mechanism includes: projecting multi-scale text image fusion features to a unified scale space through convolutional mapping operation, and merging them according to the channel dimension to generate new fusion features; performing global average pooling operation based on the new fusion features to extract global context descriptors of the channel dimension; performing nonlinear transformation on the global context descriptors through a multilayer perceptron network, and generating channel importance weight coefficients using an activation function; performing element-wise multiplication operation on the channel importance weight coefficients and the new fusion features to generate the final target features; and parsing the final target features through a classifier to output probability values ​​representing the possibility of wire faults, thereby constructing the target fault probability prediction model.

[0012] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the method includes: obtaining a truth verification model based on current feedback by combining a target fault probability prediction model with contact signal feedback; extracting a high-dimensional state transition vector of fault features from the target fault probability prediction model to generate a confidence score for fault determination; constructing a fault initial screening decision logic based on the confidence score; generating a physical verification trigger command when the score exceeds a preset benchmark; constructing a contact detection path planning for suspected fault points by combining the physical verification trigger command with UAV flight status data; and establishing a logical framework for final truth determination based on the contact detection path planning and the real-time data stream of contact signal feedback to obtain the truth verification model based on current feedback.

[0013] The beneficial effects of this preferred technical solution are as follows: relying on the high-dimensional state transition vector transformed from the fault characteristics related to single-phase grounding faults, a confidence score for fault determination is gradually generated. (The high-dimensional state transition vector here is actually the transformation of various fault characteristics into multi-dimensional change data, and the confidence score is the reliability score for judging whether it is a fault.) This establishes a targeted initial screening decision logic, which can quickly locate suspected fault points among many potential locations, greatly reducing unnecessary detection steps, giving the fault investigation work a clear direction, and effectively saving time and cost in the overall detection process.

[0014] As a preferred embodiment of the single-phase grounding fault UAV detection method of the present invention, the following is included: Based on contact-type detection path planning and combined with real-time data streams from contact-type signal feedback, a logical framework for final truth determination is established, comprising: driving the robotic arm mounted on the UAV to perform grasping and locking actions based on a physical verification trigger command, attaching the physical detector to the conductor or tension clamp under test; injecting a detection signal of a specific frequency into the line under test through the physical detector to form a closed-loop test circuit; collecting real-time current amplitude data of the line based on the loop feedback of the detection signal; performing secondary confirmation of the fault nature by combining the difference between the real-time current amplitude data and a preset fault current threshold; updating the output state of the truth verification model based on the secondary confirmation result to complete the locking of the single-phase grounding fault.

[0015] As a preferred embodiment of the single-phase grounding fault UAV detection system of the present invention, it is characterized by comprising: a multi-dimensional environmental space construction module, used to acquire multi-spectral visual data and positioning association data of the target inspection area, and construct a multi-dimensional environmental state space model; a dual-stream heterogeneous feature extraction module, used to perform dual-stream heterogeneous feature extraction through the multi-dimensional environmental state space model to obtain a multi-scale fused visual representation; a semantic dimension calibration module, used to obtain a spatiotemporal environmental semantic mapping vector by combining the environmental information parsed from the multi-scale fused visual representation with the positioning association data, used for the first feature dimension calibration of the feature extraction network; a dual-drive attention adjustment module, used to combine the first feature dimension calibration, analyze the coupling relationship between the spatiotemporal environmental semantic mapping vector and the multi-scale fused visual representation, obtain the semantic-scale dual-drive attention threshold, and perform the second feature dimension adjustment of feature fusion; a probability prediction model construction module, used to obtain a target fault probability prediction model based on the adjusted multi-dimensional feature distribution and combined with a channel adaptive weighting mechanism; and a physical truth verification module, used to obtain a truth verification model based on current feedback through the target fault probability prediction model and combined with contact signal feedback, to lock the conductor fault.

[0016] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for detecting a single-phase grounding fault in a drone.

[0017] A computer-readable storage medium having a computer program stored thereon, the steps of a single-phase grounding fault detection method for unmanned aerial vehicles (UAVs) implemented when the computer program is executed by a processor.

[0018] The beneficial effects of this invention are as follows: It systematically integrates multi-spectral visual acquisition data and high-precision positioning association information of the target inspection area to construct a multi-dimensional environmental state space model covering geographical environment, meteorological conditions and line visual features; it leverages cross-modal information fusion to deeply mine the inherent coupling relationship between spatiotemporal environmental semantic vectors and multi-scale visual representations. Cross-modal fusion combines different types of information such as images and text for analysis, weakens the interference of irrelevant factors such as complex weather and terrain, and makes fault-related features such as electric arcs and ablation marks more distinct at the data level, significantly improving the reliability of fault screening in complex scenarios such as mountainous areas and thunderstorms.

[0019] It captures local detail features and global contextual information in image data to form a hierarchical multi-scale visual representation; the collaborative channel adaptive weighting mechanism dynamically allocates the importance of different feature channels, strengthens the response intensity of fault-related features, and extracts local and global image features using two different technical paths, making the fault probability prediction results more consistent with the actual operating conditions of the line, and can quickly focus on high-suspection fault locations, greatly reducing the time and resource consumption caused by invalid detection links. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The above is a flowchart of a method for detecting single-phase grounding faults in a UAV, provided as an embodiment of the present invention.

[0022] Figure 2 The present invention provides a flowchart of the feature fusion structure of a UAV detection method for single-phase grounding faults according to an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1-2 This is the first embodiment of the present invention, which provides a method for detecting a single-phase ground fault in a UAV, including: S1: Acquire multispectral visual data and location-related data of the target inspection area, and construct a multidimensional environmental state space model.

[0025] S2: Perform dual-stream heterogeneous feature extraction through a multi-dimensional environmental state space model to obtain a multi-scale fused visual representation.

[0026] S3: By combining multi-scale fusion visual representations with the environmental information parsed from location-related data, a spatiotemporal environmental semantic mapping vector is obtained, which is used for the calibration of the first feature dimension of the feature extraction network.

[0027] S4: Combine the calibration of the first feature dimension, analyze the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fusion visual representation, obtain the semantic scale dual-drive attention threshold, and adjust the second feature dimension of feature fusion.

[0028] S5: Based on the adjusted second feature dimension distribution, a target fault probability prediction model is obtained by combining the channel adaptive weighting mechanism.

[0029] S6: By combining the target fault probability prediction model with contact signal feedback, a truth verification model based on current feedback is obtained to lock in the conductor fault.

[0030] It should be noted that existing single-phase grounding fault detection mainly relies on manual operation, requiring staff to carry equipment to the site and manually detect faults using long insulating rods. This relies on personal experience and cannot guarantee the objectivity of fault identification. Simply relying on manual vision or basic equipment ignores the influence of environmental factors (such as weather and geographical location) on fault morphology (such as electric arc, corona, and ultraviolet spots). It is difficult to conduct effective non-contact initial screening under complex weather conditions. The lack of effective linkage between physical contact verification and visual judgment can easily lead to false detection or missed detection.

[0031] Therefore, to address the aforementioned issues, this invention constructs a multi-dimensional environmental state space model containing multi-spectral visual and positioning data through steps S1 to S6, providing a foundation for comprehensive perception; it obtains multi-scale fused visual representations through dual-stream heterogeneous feature extraction, capturing the micro and macro features of faults; furthermore, it introduces a spatiotemporal environmental semantic mapping vector, using environmental information to calibrate the feature network in the first dimension, and combines semantic scale dual-drive attention thresholds for the second dimension adjustment, effectively solving the problem of poor robustness of single visual features in complex environments; channel adaptive weighting constructs a high-precision fault probability prediction model for initial screening, significantly reducing invalid operations; and it combines contact signal feedback to construct a truth verification model, forming a closed-loop detection logic of "visual environment initial screening plus physical contact verification," thereby achieving the identification of single-phase grounding faults.

[0032] Example 2, refer to Figure 1-2 This is the second embodiment of the present invention, which differs from the first embodiment in that: a method for detecting a single-phase grounding fault in a UAV further includes: In this embodiment, step S1 involves acquiring multispectral visual data and location-related data of the target inspection area and constructing a multidimensional environmental state space model. This can be achieved by the collaborative operation of a high-definition visible light camera and an infrared thermal imaging camera mounted on a UAV, which collect RGB images and thermal imaging data of the conductor and its surrounding environment, while simultaneously recording the UAV's real-time positioning coordinates. Specifically, this process establishes a comprehensive dataset that includes physical representations and environmental background. It correlates fault features such as arcs, coronas, and burn marks captured by multispectral images with geographical environment and real-time weather information (such as rainfall and air humidity) obtained based on coordinate back lookup, constructing a multidimensional environmental state space model that reflects both the microscopic line condition and the macroscopic environmental background. This addresses the problem of single visual judgment being greatly affected by environmental interference and having unstable accuracy.

[0033] In one optional implementation, step S1 involves acquiring multi-spectral visual data and location association data of the target inspection area and constructing a multi-dimensional environmental state space model. It can also perform multi-dimensional perception of specific physical phenomena of single-phase grounding faults, using visible light data to capture obvious physical structural damage such as insulator bursting, conductor melting, or surface ablation, while using infrared thermal imaging data to capture abnormal temperature rises or ultraviolet spot features that are difficult to detect with the naked eye, thereby forming a complementary visual evidence chain to ensure the complete capture of fault features under different lighting conditions.

[0034] In another optional implementation, the acquisition of multispectral visual data and location-related data of the target inspection area in step S1, and the construction of a multidimensional environmental state space model can also be achieved through in-depth mining of the location-related data. High-precision positioning information (such as GPS or RTK coordinates) obtained by UAVs can be used as index keys to query and associate the geographical attributes (such as mountainous areas and plains) and meteorological text descriptions (such as thunderstorms and strong winds) of the current inspection area in real time. This unstructured environmental context information is incorporated into the state space, providing the necessary data foundation for the subsequent construction of semantic mapping vectors through word vector embedding algorithms, thereby assisting in judging the impact weight of weather factors on line faults.

[0035] Specifically, in step S2, dual-stream heterogeneous feature extraction is performed using a multi-dimensional environmental state space model to obtain a multi-scale fused visual representation, including the following steps A1-A3: A1: Extract heterogeneous spectral features based on multidimensional state space and splice them to generate a hybrid data stream; A2: Based on a hybrid data stream, multi-level local and global features are extracted in parallel through convolutional processing units and self-attention processing units; A3: Combining multi-level local and global features, the feature sequences are spliced ​​together to obtain a multi-scale fused visual representation.

[0036] In this embodiment of the application, step A2, which involves extracting features through a convolutional processing unit, employs a convolutional block containing multiple processing layers for local feature parsing, including the following steps A211-A213: A211: The concatenated preliminary features are input into the first branch, which is mainly composed of multiple CNN convolutional blocks.

[0037] A212: Data processing is performed sequentially within the convolutional block. The processing unit includes a convolutional layer, a batch normalization (BN) layer, and an activation layer, which perform calculations on the local texture details of the image data.

[0038] For example, features of local visual morphology such as electric arc, corona, ultraviolet spots, insulator bursting, conductor melting, and burn marks that exist during single-phase grounding faults can be extracted through convolutional layers.

[0039] A213: After processing by convolutional blocks, features at different scales are extracted through the FPN (Feature Pyramid Network) module to complete the hierarchical construction of local features.

[0040] In an optional embodiment, for feature extraction via a self-attention processing unit, the extraction method can also be global feature parsing using a transformer network architecture, including the following steps A221-A223: A221: The initial features after splicing are synchronously input into the second branch, which is designed to process global context information.

[0041] A222: It uses the Swing Transformer (moving window layer Transformer) as the backbone network to process the input features and captures long-distance dependencies in the image through a self-attention mechanism.

[0042] A223: After the backbone network is processed, it also goes through the FPN layer to obtain features at different scales, and outputs a feature hierarchy containing global information.

[0043] In another alternative embodiment, for parallel extraction of multi-level local and global features, the extraction method can also be to construct a dual-stream heterogeneous network and integrate the feature streams, including the following steps A231-A234: A231: Construct a parallel network structure containing CNN branches and Swin Transformer branches, which are used to process local details and global semantics, respectively.

[0044] A232: In the CNN branch, the physical morphological features of the wires and environment are extracted through convolution operations, corresponding to the specific appearance traces when the fault occurs.

[0045] A233: In the Swing Transformer branch, the overall features of the environmental background are extracted through the attention mechanism, which corresponds to the macroscopic visual representation of the area where the conductor is located.

[0046] A234: Concat the features output from the two branches to generate fused image features of four different scales, C2, C3, C4, and C5, which serve as the basis for subsequent semantic vector interaction.

[0047] It should be noted that multi-scale feature representation allows subsequent semantic vectors to assign different levels of attention to image features at different scales. This enables a preliminary judgment of the current conductor's condition through multiple dimensions such as image, weather, and location without directly relying on the detector, thereby reducing the frequency of detector use and improving detection efficiency.

[0048] Furthermore, in step S3, the spatiotemporal environment semantic mapping vector is obtained by combining the environmental information parsed from the multi-scale fusion visual representation with the localization association data. This vector is used for the first feature dimension calibration of the feature extraction network, including the following steps B1-B4: B1: Obtaining geographic and weather text descriptions based on real-time UAV positioning coordinate association; B2: Construct an information flow of environmental text set by performing high-dimensional encoding through text description combined with word vector embedding algorithm; B3: Extract local and global text vectors representing micro and macro states respectively based on the information flow of the environmental text set; B4: Generate a spatiotemporal environment semantic mapping vector by concatenating local and global text vectors.

[0049] In this embodiment of the application, step B3 involves extracting local and global text vectors representing micro and macro states based on the environmental text set information stream. The extraction method involves distinguishing texts describing environmental states at different scales from the environmental text set information stream and performing vectorization processing on each, including the following steps B211-B213: B211: In the environmental text set information stream constructed through step B2, the geographical and weather text descriptions obtained by associating the real-time positioning coordinates of the UAV are divided into two categories: one category is used to describe the local environmental state of the area where the current guide is located, and the other category is used to describe the overall environmental background of the area where the guide is located.

[0050] B212: Input the two types of text into the word vector embedding algorithm to perform high-dimensional encoding. In the same vector space, a set of local text vectors and a set of global text vectors are obtained respectively. The local text vectors focus on reflecting the specific weather and location descriptions near the conductor, while the global text vectors focus on reflecting the weather and geographical information over a larger area.

[0051] For example, in the information flow of the environmental text set, one part of the text is used to describe the weather information and geographical location (e.g., location coordinates) of the area where the current conductor is located, and another part of the text is used to describe the overall meteorological and geographical environment of the area where the conductor is located. By performing high-dimensional vectorization and aggregation processing on these two parts of the text respectively, we can obtain local text vectors representing the micro state and global text vectors representing the macro state, which provides a basis for the subsequent generation of spatiotemporal environmental semantic mapping vectors (semantic vectors).

[0052] B213: Aggregate the obtained local text vector set and global text vector set respectively to form a single local text vector for representing the micro state and a single global text vector for representing the macro state, providing input for vector concatenation in step B4.

[0053] In an optional embodiment, for extracting local and global text vectors representing micro and macro states respectively from the environmental text set information stream, the extraction method can also be to classify the environmental text set information stream based on the text's temporal update characteristics, including the following steps B221-B223: B221: During UAV inspections, based on the time attributes of the text descriptions, the environmental text collection information flow is divided into a text collection reflecting real-time weather changes and a text collection reflecting relatively stable geographical information and long-term meteorological characteristics. The former focuses on describing short-term weather conditions, while the latter focuses on describing geographical location and long-term climate.

[0054] B222: Perform high-dimensional encoding of word vector embedding algorithm on the two types of text sets mentioned above, map real-time weather text to a set of local text vectors, map geographical and long-term meteorological text to a set of global text vectors, and perform vector aggregation within their respective sets to obtain local text vectors representing local short-term environmental states and global text vectors representing long-term macro-environmental states.

[0055] B223: The obtained local text vector and global text vector are used as micro-state vector and macro-state vector respectively, providing local and global text feature representations with time dimension distinction for constructing spatiotemporal environment semantic mapping vectors.

[0056] In another optional embodiment, for extracting local and global text vectors representing micro and macro states respectively based on the information flow of the environmental text set, the extraction method can also be to construct local and global text vectors based on the descriptive granularity of the text in spatial scope, including the following steps B231-B234: B231: Based on the description of spatial range in the information flow of the environmental text set, texts related to the small-scale environment near the conductor are classified as local environmental texts, and texts describing the geographical distribution and overall meteorological conditions of a larger area are classified as macro-environmental texts.

[0057] B232: Perform high-dimensional encoding of word vector embedding algorithm on local environment text and macro environment text respectively, so that local environment text forms a set of local text vectors in vector space, and macro environment text forms a set of global text vectors in the same vector space.

[0058] B233: Perform aggregation operations on the local text vector set to obtain local text vectors that characterize the environmental state of the spatial range near the conductor; perform aggregation operations on the global text vector set to obtain global text vectors that characterize the environmental state of a larger spatial range.

[0059] B234: The obtained local text vector and global text vector are used as text representations of micro and macro states, respectively. These are combined with the image features of different scales corresponding to the multi-scale fusion visual representation in subsequent steps to construct a spatiotemporal environment semantic mapping vector input that matches the spatial range.

[0060] It should be noted that since single-phase grounding faults are often closely related to image phenomena such as electric arcs, coronas, and conductor burn marks, as well as weather and geographical location, by mapping text information of different scales into local and global text vectors and participating in the construction of spatiotemporal environmental semantic mapping vectors, the influence of weather information and geographical location on fault judgment can be more fully incorporated when analyzing the coupling relationship between spatiotemporal environmental semantic mapping vectors and multi-scale fusion visual representations. Thus, based on the fusion of image and environmental information, a more comprehensive judgment can be made on whether a single-phase grounding fault exists in the conductor.

[0061] Furthermore, in step S4, the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fused visual representation is analyzed by combining the first feature dimension calibration, to obtain the semantic scale dual-drive attention threshold, and the second feature dimension of feature fusion is adjusted, including the following steps C1-C4: C1: Analysis of multi-level image feature tensors based on multi-scale fusion visual representation; C2: Based on multi-level image feature tensors, the spatiotemporal environment semantic mapping vector is mapped to the corresponding feature space to build the foundation for cross-modal interaction; C3: Weights of attention to feature tensors based on fundamental computational semantics through cross-modal interaction; C4: Weighted fusion based on attention weights generates fused features constrained by semantic-scale dual-drive attention thresholds.

[0062] In this embodiment of the application, step C3, regarding the attention weight of the feature tensor calculated based on cross-modal interaction, is calculated by establishing cross-attention interaction channels at different scales based on the correspondence between multi-scale image features and spatiotemporal environment semantic mapping vectors, including the following steps C211-C213: C211: In the multi-scale fusion visual representation obtained in step S2, image features of four different scales, C2, C3, C4 and C5, are extracted so that each scale feature corresponds to the image representation of the conductor and its surrounding environment at different spatial resolutions.

[0063] C212: The spatiotemporal environment semantic mapping vector obtained in step S3 is used as a unified semantic vector and input into multiple parallel cross-attention interaction modules, so that each cross-attention interaction module interacts with image features at a certain scale, thereby calculating the attention of the semantic vector to the image features at each scale.

[0064] For example, in the fusion module, Cross Attention1, Cross Attention2, Cross Attention3, and Cross Attention4 are set up. The input of Cross Attention1 is the semantic mapping vector of C2 and the spatiotemporal environment, the input of Cross Attention2 is the semantic mapping vector of C3 and the spatiotemporal environment, the input of Cross Attention3 is the semantic mapping vector of C4 and the spatiotemporal environment, and the input of Cross Attention4 is the semantic mapping vector of C5 and the spatiotemporal environment. Through the above correspondence, the attention of semantic vectors at different scales is made to show differences, so as to reflect the different importance of image features at different scales for fault judgment.

[0065] C213: Within each cross-attention interaction module, based on the degree of matching between the spatiotemporal environment semantic mapping vector and the corresponding scale image features, the attention weight of semantics to the image feature tensor at that scale is obtained, providing basic weight information for subsequent weighted fusion in step C4 to generate fused features constrained by semantic-scale dual-drive attention threshold.

[0066] In an optional embodiment, the calculation method for the attention weight of the feature tensor based on the computational semantics of cross-modal interaction can also be as follows: emphasizing the differential expression of the attention degree of the spatiotemporal environment semantic mapping vector to image features at different scales, including the following steps C221-C223: C221: In multi-scale fusion visual representation, the four scale features C2, C3, C4, and C5 are clearly defined to correspond to the information of the wires and environment images from finer to coarser granularity, respectively. These features are input as multi-level image feature tensors into a parallel cross-attention interaction structure to establish a semantic interaction channel that does not interfere with each other at different scales.

[0067] C222: The spatiotemporal environment semantic mapping vector is reused in various cross-attention interaction modules, so that the same semantic vector interacts with C2, C3, C4 and C5 respectively. The cross-attention mechanism calculates the degree of semantic response to image features at each scale, thus reflecting the different attention of the semantic vector to images at different scales, that is, the different importance of image features at different scales to semantic information.

[0068] C223: Integrate the attention weights output by each cross-attention interaction module, and use these weights as one of the bases for the semantic scale dual-drive attention threshold. This provides a prerequisite for subsequent weighted processing according to different scale weights when fusing image and semantic features in a unified scale space.

[0069] In another alternative embodiment, the calculation of attention weights on feature tensors based on cross-modal interaction semantics can also be performed by linking parallel cross-attention interaction with the subsequent multi-scale text-image fusion feature generation process, including the following steps C231-C234: C231: For image features at four scales, C2, C3, C4, and C5, the spatiotemporal environment semantic mapping vector is input into the corresponding Cross Attention1, Cross Attention2, Cross Attention3, and Cross Attention4 respectively, so that each cross attention module outputs text-image fusion features that match the scale, and implicitly calculates the semantic attention to image features at that scale within the module.

[0070] C232: After obtaining four different sizes of text-image fusion features, the internal weights reflecting semantic attention in the fusion features at each scale are used as the quantitative results of the semantic-image feature coupling relationship at that scale. These results can be understood as the attention size of the semantic vector to the feature tensor at different scales.

[0071] C233: Based on text-image fusion features, fusion features of different scales are mapped to a unified scale space through 1×1 convolution, so that the attention weights formed in each cross-attention module can participate in the subsequent channel-dimension merging operation under a unified scale, thereby forming a brand-new fusion feature to integrate multi-scale and semantic information.

[0072] C234: In the new fusion feature, the semantic attention difference to image feature tensors at each scale reflected by the parallel cross-attention interaction module is retained. This allows the feature to reflect the constraint effect of the semantic-scale dual-drive attention threshold on the final target feature formation process when combined with the channel adaptive weighting mechanism, global pooling, channel weight calculation and classifier parsing.

[0073] It should be noted that by mapping the semantic vector to the corresponding feature space and constructing a cross-modal interaction basis, and then using the parallel cross-attention interaction module to calculate the attention weight of semantics to image feature tensors at different scales, the differences in the coupling relationship between image and environmental information can be reflected during the fusion process.

[0074] Furthermore, in step S5, based on the adjusted second feature dimension distribution and combined with the channel adaptive weighting mechanism, a target fault probability prediction model is obtained, including the following steps D1-D5: D1: Multi-scale text image fusion features are projected onto a unified scale space through convolutional mapping operations, and then merged according to channel dimensions to generate new fusion features; D2: Perform global average pooling operation based on the new fusion features to extract the global context descriptor of the channel dimension; D3: The global context descriptor is transformed nonlinearly through a multilayer perceptron network, and the channel importance weight coefficients are generated using an activation function; D4: Perform element-wise multiplication based on the channel importance weight coefficients and the new fusion features to generate the final target features; D5: The final target features are analyzed by a classifier, and the probability values ​​representing the possibility of conductor failure are output to construct a target failure probability prediction model.

[0075] In this embodiment of the application, step D4 involves performing element-wise multiplication on the new fused features based on the channel importance weight coefficients to generate the final target features. The processing method is to scale each channel of the new fused features element-wise using the channel importance weight coefficients, thereby highlighting the fault-related channel responses at different scales. This includes the following steps D211-D213: D211: After completing step D3, the channel importance weight coefficients for each channel of the new fusion feature are obtained. Each weight corresponds to a channel and is used to reflect the importance of the channel in characterizing the fault state of the conductor.

[0076] D212: The channel importance weight coefficients are multiplied element-wise with the new fusion features along the channel dimension, which amplifies the channels with larger weight coefficients and relatively suppresses the channels with smaller weight coefficients, thereby highlighting the channel responses that better reflect fault information in the fusion features.

[0077] For example, when processing multi-scale text image fusion features, if some channels focus more on characterizing phenomena such as electric arcs, conductor burn marks, and insulator bursts, after the element-wise multiplication operation of the channel importance weight coefficients and the new fusion features, the response amplitude of these channels in the final target features will be significantly improved, enabling the subsequent classifier to make more concentrated use of these feature information related to single-phase grounding faults.

[0078] D213: After completing the element-wise multiplication operation, the results of all channels are recombined to form the final target feature that has been weighted and adjusted in the channel dimension, providing input for the analysis by the classifier in step D5.

[0079] In an optional embodiment, element-wise multiplication is performed on the channel importance weight coefficients and the novel fusion features to generate the final target features. The processing can also be further enhanced by highlighting the interaction between multiple scales to make a preliminary image-level judgment of conductor faults, including the following steps D221-D223: D221: In the novel fusion feature, different channels correspond to the channel representation formed by fusing image features and semantic vectors at different scales. The channel importance weight coefficients obtained by global pooling and multilayer perceptron network reflect the relative importance of each scale fusion feature in describing the fault state.

[0080] D222: By using the channel importance weighting coefficient, the new fused features are multiplied channel by channel, so that the scale channel more related to the fault phenomenon receives a higher response value, and the scale channel related to the environmental background or interference information receives a lower response value, thereby achieving effective differentiation of multi-scale channels in the final target features.

[0081] D223: After obtaining the final target feature, the feature is input into the classifier, which outputs the probability that the target feature is a fault point. This determines whether the current conductor corresponding to the target feature is a faulty conductor, enabling a preliminary judgment of the current conductor status based on multi-dimensional information such as images, weather, and location without using physical detectors. This provides a basis for reducing the frequency of detector use and improving detection efficiency.

[0082] In another optional embodiment, an element-wise multiplication operation is performed on the channel importance weight coefficient and the new fused feature to generate the final target feature. The processing can also be carried out by connecting the final target feature with the subsequent fault probability output and truth value verification process, including the following steps D231-D234: D231: After performing global average pooling and nonlinear transformation of the multilayer perceptron network on the new fused features, the channel importance weight coefficients of each channel are obtained. These weight coefficients are then used to perform element-wise multiplication on the channel dimensions of the new fused features to obtain the final target features that can centrally reflect the fault features.

[0083] D232: Input the final target feature into the classifier, and use the classifier to analyze the feature to obtain the probability value that represents the possibility of conductor fault. This will build a target fault probability prediction model and realize the probabilistic description of whether a single-phase ground fault exists in the conductor.

[0084] D233: When using the probability value output by the target fault probability prediction model, the probability value can be corresponding to the confidence score of fault determination. When the score exceeds the preset benchmark, according to the technical solution in the claim, a physical verification trigger command is further generated to drive the robotic arm mounting device on the UAV to perform grasping and locking actions, and to attach the physical detector to the suspected faulty wire or tension clamp.

[0085] D234: After the physical detector is connected, a detection signal of a specific frequency is injected into the line under test and the real-time current amplitude data of the line is collected. The nature of the fault is then confirmed by combining the difference between the real-time current amplitude data and the preset fault current threshold. The output state of the truth verification model is updated based on the result of the second confirmation. Thus, based on the target fault probability prediction model established by the image and environmental information, the final lock of the single-phase grounding fault is achieved through current feedback.

[0086] It should be noted that, through the above feature fusion and channel adaptive weighting mechanism, on the one hand, a target fault probability prediction model is formed at the level of multimodal information of images and text, realizing the probabilistic judgment of the fault state of the conductor; on the other hand, it is connected with the truth verification model based on current feedback formed by contact signal feedback, so that the detection process of single-phase grounding fault can be extended from the preliminary judgment of multidimensional information such as images, weather, and location to the closed-loop verification of physical current.

[0087] Furthermore, in step S6, the target fault probability prediction model is combined with the contact signal feedback to obtain a truth verification model based on current feedback, thereby locking in the conductor fault, including the following steps E1-E8: E1: Extract high-dimensional state transition vectors of fault features through the target fault probability prediction model to generate a confidence score for fault determination; E2: Construct a fault screening decision logic based on confidence score. When the score exceeds the preset benchmark, a physical verification trigger instruction is generated. E3: Combine physical verification trigger commands with UAV flight status data to construct contact-based detection path planning for suspected fault points; E4: Based on contact detection path planning and combined with the real-time data stream of contact signal feedback, a logical framework for final truth determination is established, resulting in a truth verification model based on current feedback. E5: Based on the physical verification trigger command, the robotic arm mounted on the drone performs grasping and locking actions, attaching the physical detector to the conductor or tension clamp to be tested. E6: A specific frequency detection signal is injected into the circuit under test through a physical detector to form a closed-loop test circuit; Based on the loop feedback of the detection signal, the real-time current amplitude data of the line is collected; E7: Combine the difference between real-time current amplitude data and preset fault current threshold to perform secondary confirmation of the fault nature; E8: Update the output state of the truth verification model based on the secondary confirmation results to complete the locking of single-phase grounding faults.

[0088] In this embodiment of the application, in step E5, the robotic arm mounting device on the UAV is driven to perform grasping and locking actions based on the physical verification trigger command, and the physical detector is attached to the conductor to be tested or the tension clamp. The attachment method is to complete the UAV pose adjustment and detector attachment operation under the constraints of contact detection path planning, including the following steps E211-E213: E211: After generating the physical verification trigger command in step E2, combined with the contact detection path planning for the suspected fault point constructed in step E3, the UAV is guided to fly along the planned path to the vicinity of the suspected fault point based on the flight status data, so that the UAV body and the robotic arm mounting device are in a position suitable for performing the grasping action relative to the wire to be tested or the tension clamp.

[0089] E212: When the UAV reaches the position of the conductor or tension clamp to be tested and reaches the predetermined flight attitude, it controls the robotic arm mounting device to perform gripping and locking actions in sequence according to the physical verification trigger command, so as to firmly attach the physical detector to the bare conductor or tension clamp, so that a stable physical connection is formed between the physical detector and the line under test, which meets the requirements of subsequent injection of specific frequency detection signals and acquisition of real-time current amplitude data.

[0090] For example, when the fault judgment score output by the target fault probability prediction model exceeds the preset benchmark, a physical verification trigger command is generated according to the fault screening decision logic. The UAV flies to the suspected faulty wire location according to the contact detection path planning. At the location, the robotic arm mounting device is controlled to complete the gripping and locking operation of the wire, and the physical detector is attached to the wire. This provides the physical basis for injecting a specific frequency detection signal into the line under test, forming a closed-loop test circuit, and collecting real-time current amplitude data of the line in step E6.

[0091] E213: After the physical detector is attached, maintain the stability of the UAV attitude and the locking state of the robotic arm to ensure that the physical detector maintains reliable contact throughout the entire closed-loop test circuit operation. This ensures that the real-time current amplitude data collected based on the feedback from the detection signal circuit can accurately reflect the current operating status of the line, providing reliable input data for the truth value verification model based on current feedback.

[0092] In an optional embodiment, for physical verification trigger commands, the robotic arm mounting device on the UAV is driven to perform grasping and locking actions, attaching the physical detector to the conductor to be tested or the tension clamp. The attachment method can also be to reuse the robotic arm mounting action in multiple fault checks, in combination with the continuous operation requirements of the inspection line, including the following steps E221-E223: E221: In a fault inspection, when the target fault probability prediction model determines that the current conductor is a faulty conductor, and after completing the physical detector connection, injecting a specific frequency detection signal into the line under test, collecting real-time current amplitude data, and performing secondary confirmation of the fault nature according to steps E5-E7, the fault point detection ends according to the output state of the truth verification model, and the robot arm mounting device is controlled to release the locking connection of the current conductor or tension clamp.

[0093] E222: After the current attachment is removed, the UAV continues to fly along the predetermined inspection route. When a new fault determination confidence score is obtained again through the target fault probability prediction model during subsequent inspections and exceeds the preset benchmark, a physical verification trigger command is generated again to guide the UAV to fly to the vicinity of a new suspected fault point. The robotic arm attachment device is then repeatedly executed to grab and lock the physical detector onto the new conductor or tension clamp to be tested.

[0094] E223: Through the repeated attaching and unattaching process described above, the UAV can sequentially perform physical detector attachment, detection signal injection, real-time current amplitude acquisition, and secondary confirmation of the fault nature at multiple suspected fault points in a single inspection mission. Each fault inspection is carried out in the same manner, thereby completing multi-point truth verification based on current feedback without replacing the equipment. This is consistent with the operation method of "Furthermore, if it is determined that the current conductor corresponding to the target feature is a faulty conductor, the UAV continues to inspect along the inspection route, and each fault inspection is carried out in the manner described above."

[0095] In another optional embodiment, for physical verification trigger commands, the robotic arm mounted on the UAV performs grasping and locking actions to attach the physical detector to the conductor under test or tension clamp. The attachment method can also be to integrate the robotic arm attachment action with the contact detection path planning and closed-loop test circuit construction process, including the following steps E231-E234: E231: When constructing contact detection path planning for suspected fault points, the flight status data of the UAV, the spatial position of the wire to be tested or the tension clamp, and the operating space requirements of the robotic arm mounting device are comprehensively considered to ensure that the endpoint of the path planning meets the spatial conditions for the robotic arm to perform grasping and locking actions smoothly, thus providing constraints for the subsequent attachment of physical detectors.

[0096] E232: After the UAV flies along the planned contact detection path to the destination, it automatically switches to a flight mode suitable for the operation of the robotic arm according to the physical verification trigger command. In this mode, it maintains stable hovering or makes small-range attitude adjustments and drives the robotic arm mounting device to perform grasping and locking actions, reliably attaching the physical detector to the conductor to be tested or the tension clamp, so that the physical connection is naturally formed at the planned destination.

[0097] E233: After the physical detector is attached, a detection signal of a specific frequency is injected into the circuit under test through the physical detector in step E6 to form a closed-loop test circuit. The real-time current amplitude data of the circuit is collected based on the loop feedback of the detection signal, so that the robot arm mounting action, the detection signal injection and the current acquisition process are continuously connected in time.

[0098] E234: The real-time current amplitude data collected under the closed-loop test circuit is compared with the preset fault current threshold. The nature of the fault is confirmed for the second time based on the difference between the two. The output state of the truth verification model is updated according to steps E7-E8. Finally, the single-phase grounding fault is locked, and the physical detector connected by the robotic arm mounting device becomes the link between the target fault probability prediction model and the current feedback truth verification model.

[0099] It should be noted that by organically combining the fault determination results output by the target fault probability prediction model with contact detection path planning, robotic arm mounting device operation, closed-loop test circuit construction, and truth verification model based on current feedback, the UAV can make a preliminary judgment on multi-dimensional information such as images, weather, and location, and then use physical detectors to make a secondary confirmation of suspected faulty wires, thereby avoiding complete reliance on manual carrying of equipment and the use of long insulating rods for on-site detection.

[0100] Example 3 is the third embodiment of the present invention. It differs from the previous two embodiments in that: a method for detecting single-phase grounding faults in a UAV further includes, in order to verify and explain the technical effects used in this method, comparing the experimental results with scientific demonstration methods to verify the real effect of this method.

[0101] The experimental design is divided into three comparison dimensions, corresponding to single visual feature detection, multi-scale visual and environmental semantic fusion detection, and the full-process detection including contact verification proposed in this application. In the first round of verification, the research focused on comparing the feature extraction mechanism of traditional convolutional neural networks (CNN) with the dual-stream heterogeneous feature extraction architecture innovatively adopted in this application. Conventional technical paths rely on stacking single convolutional layers to complete feature capture. In scenarios with complex background information, this method is prone to loss of information on subtle fault textures (such as tiny cracks on the surface of wires), which limits the accuracy of subsequent detection.

[0102] In the second round of verification, a spatiotemporal environment semantic mapping vector was introduced to compensate for the inherent defects of detection that rely solely on image visual features. In adverse weather conditions such as fog, strong light reflection, and heavy rain, simple visual detection models often mistake environmental interference for fault features, leading to frequent false alarms and seriously affecting the practicality and reliability of the detection system.

[0103] The third round of verification focuses on the performance evaluation of the truth verification model based on current feedback. For high-confidence suspected fault points output by the visual and semantic fusion model, the system will automatically drive the robotic arm on the UAV to carry physical detectors, inject a detection signal of a specific frequency into the transmission line under test, and collect feedback current amplitude data in the circuit in real time.

[0104] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: a single-phase grounding fault UAV detection system includes a multi-dimensional environmental space construction module for acquiring multi-spectral visual data and positioning association data of the target inspection area and constructing a multi-dimensional environmental state space model; a dual-stream heterogeneous feature extraction module for performing dual-stream heterogeneous feature extraction through the multi-dimensional environmental state space model to obtain a multi-scale fused visual representation; a semantic dimension calibration module for obtaining a spatiotemporal environmental semantic mapping vector by combining the environmental information parsed from the multi-scale fused visual representation with the positioning association data, used for the first feature dimension calibration of the feature extraction network; a dual-drive attention adjustment module for combining the first feature dimension calibration, analyzing the coupling relationship between the spatiotemporal environmental semantic mapping vector and the multi-scale fused visual representation, obtaining a semantic-scale dual-drive attention threshold, and adjusting the second feature dimension of feature fusion; a probability prediction model construction module for obtaining a target fault probability prediction model based on the adjusted multi-dimensional feature distribution and a channel adaptive weighting mechanism; and a physical truth verification module for obtaining a current feedback-based truth verification model through the target fault probability prediction model and contact signal feedback, and locking the conductor fault.

[0105] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0107] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0108] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented in combination with any of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting single-phase grounding faults in unmanned aerial vehicles (UAVs), characterized in that: include, Acquire multispectral visual data and location-related data of the target inspection area, and construct a multidimensional environmental state space model; The multi-dimensional environmental state space model is used to perform dual-stream heterogeneous feature extraction to obtain a multi-scale fused visual representation. By combining multi-scale fusion visual representations with environmental information parsed from location-related data, a spatiotemporal environmental semantic mapping vector is obtained, which is used for the calibration of the first feature dimension of the feature extraction network. By combining the first feature dimension calibration, the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fusion visual representation is analyzed to obtain the semantic-scale dual-drive attention threshold, and the second feature dimension of feature fusion is adjusted. Based on the adjusted second feature dimension distribution, combined with the channel adaptive weighting mechanism, a target fault probability prediction model is obtained. By combining the target fault probability prediction model with contact signal feedback to obtain a truth verification model based on current feedback, the fault in the conductor can be identified.

2. The method for detecting a single-phase grounding fault in a UAV as described in claim 1, characterized in that: By performing dual-stream heterogeneous feature extraction using the aforementioned multi-dimensional environmental state-space model, a multi-scale fused visual representation is obtained, including... Heterogeneous spectral features are extracted based on multidimensional state space and spliced ​​to generate a hybrid data stream; Based on the hybrid data stream, multi-level local and global features are extracted in parallel through convolutional processing units and self-attention processing units. By combining multi-level local and global features, the feature sequences are spliced ​​together to obtain a multi-scale fused visual representation.

3. The method for detecting a single-phase grounding fault in a UAV as described in claim 2, characterized in that: The spatiotemporal environmental semantic mapping vector obtained by combining the multi-scale fused visual representation with the location-related data includes the following: Geographic and weather text descriptions are obtained by associating real-time positioning coordinates of drones. By combining text description with word vector embedding algorithm to perform high-dimensional encoding, an information flow of environmental text set is constructed. Based on the information flow of the environmental text set, local and global text vectors representing micro and macro states are extracted respectively; Spatiotemporal environment semantic mapping vectors are generated by concatenating local and global text vectors.

4. The method for detecting a single-phase grounding fault in a UAV as described in claim 3, characterized in that: Analyzing the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fused visual representation, a semantic-scale dual-drive attention threshold is obtained. The second feature dimension adjustment for feature fusion includes... Analysis of multi-level image feature tensors based on multi-scale fusion visual representation; Based on multi-level image feature tensors, spatiotemporal environment semantic mapping vectors are mapped to corresponding feature spaces to build a foundation for cross-modal interaction; The attention weights of feature tensors are calculated based on the semantics of cross-modal interaction. Weighted fusion based on attention weights generates fused features constrained by semantic-scale dual-drive attention thresholds.

5. The method for detecting a single-phase grounding fault in a UAV as described in claim 4, characterized in that: Based on the adjusted distribution of the second feature dimension, and combined with the channel adaptive weighting mechanism, the target fault probability prediction model includes: Multi-scale text image fusion features are projected onto a unified scale space through convolutional mapping operations and merged according to channel dimensions to generate new fusion features; Global average pooling is performed based on novel fusion features to extract global context descriptors at the channel dimension; The global context descriptor is nonlinearly transformed using a multilayer perceptron network, and channel importance weight coefficients are generated using activation functions. Element-wise multiplication is performed based on the channel importance weight coefficients and the novel fusion features to generate the final target features; The final target features are analyzed by a classifier, and the probability value representing the possibility of wire failure is output to construct a target failure probability prediction model.

6. The method for detecting a single-phase grounding fault in a UAV as described in claim 5, characterized in that: The target fault probability prediction model, combined with contact signal feedback, yields a truth verification model based on current feedback, including: High-dimensional state transition vectors of fault features are extracted by the target fault probability prediction model to generate a confidence score for fault determination. The initial fault screening decision logic is constructed based on confidence scores. When the score exceeds the preset benchmark, a physical verification trigger instruction is generated. By combining physical verification trigger commands with UAV flight status data, a contact-based detection path planning system is constructed for suspected fault points. Based on contact detection path planning and combined with the real-time data stream of contact signal feedback, a logical framework for final truth determination is established, resulting in a truth verification model based on current feedback.

7. The method for detecting a single-phase grounding fault in a UAV as described in claim 6, characterized in that: Based on the aforementioned contact-based detection path planning, and combined with the real-time data stream from contact-based signal feedback, a logical framework for final truth determination is established, including: Based on the physical verification trigger command, the robotic arm mounted on the drone is driven to perform grasping and locking actions, and the physical detector is attached to the conductor or tension clamp to be tested. A closed-loop test circuit is formed by injecting a detection signal of a specific frequency into the circuit under test through a physical detector. Based on the loop feedback of the detection signal, the real-time current amplitude data of the line is collected; By combining the difference between real-time current amplitude data and preset fault current threshold, the nature of the fault is confirmed a second time. The output state of the truth verification model is updated based on the secondary confirmation results to complete the locking of single-phase grounding faults.

8. A single-phase ground fault UAV detection system, employing a single-phase ground fault UAV detection method as described in any one of claims 1 to 7, characterized in that, include: The multi-dimensional environment space construction module is used to acquire multi-spectral visual data and positioning correlation data of the target inspection area, and to construct a multi-dimensional environment state space model. The dual-stream heterogeneous feature extraction module is used to perform dual-stream heterogeneous feature extraction through the multi-dimensional environmental state space model to obtain a multi-scale fused visual representation. The semantic dimension calibration module is used to obtain the spatiotemporal environment semantic mapping vector by combining the environmental information parsed from the multi-scale fusion visual representation with the localization association data, which is used for the first feature dimension calibration of the feature extraction network. The dual-drive attention adjustment module is used to combine the first feature dimension calibration, analyze the coupling relationship between the spatiotemporal environment semantic mapping vector and the multi-scale fusion visual representation, obtain the semantic-scale dual-drive attention threshold, and perform the second feature dimension adjustment for feature fusion. The probability prediction model building module is used to obtain the target fault probability prediction model based on the adjusted multidimensional feature distribution and combined with the channel adaptive weighting mechanism. The physical truth verification module is used to obtain a current feedback-based truth verification model by combining the target fault probability prediction model with contact signal feedback, and to lock in the conductor fault.

9. A computer device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a single-phase grounding fault detection method for unmanned aerial vehicles according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a single-phase grounding fault detection method for unmanned aerial vehicles according to any one of claims 1 to 7.

Images