Substation unmanned aerial vehicle inspection positioning method and system based on multi-element information fusion

Abstract

The application discloses a kind of based on multi-element information fusion's substation unmanned aerial vehicle inspection positioning method and system, the method of the present application includes the spectral reflectance data of substation equipment surface and the electric field phase data in surrounding space are acquired synchronously;Optimal pose of current time k of unmanned aerial vehicle is fused and solved;The abnormality of the physical state of substation equipment is evaluated;Spectrum-phase feature mutual information entropy is calculated;Whether spectrum-phase feature mutual information entropy exceeds preset diagnostic threshold is judged, if exceeds preset diagnostic threshold, then it is judged that the substation equipment of optimal pose of current time k of unmanned aerial vehicle exists equipment failure.The present application aims at solving the singleness of positioning information and environmental vulnerability problem and the problem of the fragmentation of observation data and the lack of causal association existing in traditional substation unmanned aerial vehicle inspection positioning method, improve the accuracy, reliability and explainability of substation unmanned aerial vehicle inspection positioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power transmission line inspection of power systems, and particularly relates to a substation unmanned aerial vehicle (UAV) inspection positioning method and system based on multi-element information fusion. BACKGROUND

[0002] With the deep penetration of automation and intelligent technology in the field of key infrastructure operation and maintenance, using unmanned aerial vehicles (UAVs) for large-scale and high-frequency autonomous inspection has become an important development direction to improve efficiency and ensure safety. In particular, in the complex structure and special environment of a substation, how to ensure that the UAV has high-precision autonomous navigation capability in all-weather and all-scenarios, and can surpass traditional surface detection to realize deep insight and predictive diagnosis of the internal health status of the equipment, has become the core issue to promote the technological progress in this field. Although the existing technology has made certain progress in this regard, there are still several technical challenges to be solved in practice: 1. The problem of singularity and environmental vulnerability of positioning information. Existing UAV positioning solutions usually rely on a single information source, such as the Global Navigation Satellite System (GNSS) or the Simultaneous Localization and Mapping (SLAM) technology based on vision / laser radar. However, in the interior of a substation, strong electromagnetic interference often causes GNSS signals to be severely attenuated or even completely interrupted; at the same time, a large number of appearanceally similar insulators, frameworks, and other equipment make pure vision or laser SLAM methods prone to perception degradation and matching ambiguity, resulting in positioning drift or failure. This sensitivity to specific environmental conditions limits the ability of the UAV to achieve continuous and robust positioning in the entire substation. 2. The problem of fragmentation of observation data and the lack of causal correlation. Current inspection technology usually processes different types of sensor data as independent information channels. For example, a multispectral camera is used to identify physical state changes on the surface of the equipment (such as rust, dirt), while other sensors may be used to monitor the electrical characteristics of the environment. There is a lack of effective fusion mechanism between these observation data to reveal their inherent physical correlation. Therefore, the system can only report two isolated phenomena: "abnormality on the surface of the equipment" and "disturbance in the electric field", but cannot answer the key causal question: "whether the abnormality of the equipment at this location caused the disturbance in the electric field at this location", which makes the diagnosis remain at the surface description and makes it difficult to conduct deep fault tracing and prediction. SUMMARY

[0003] The technical problem to be solved by the present application: In view of the above problems of the prior art, the present application provides a substation unmanned aerial vehicle (UAV) inspection positioning method and system based on multi-element information fusion, which aims to solve the problems of singularity and environmental vulnerability of positioning information and fragmentation of observation data and lack of causal correlation in the traditional substation UAV inspection positioning method, and improve the accuracy, reliability, and explainability of the substation UAV inspection positioning.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is:

[0005] A power substation unmanned aerial vehicle inspection positioning method based on multi-element information fusion, comprising the following steps:

[0006] S1, synchronously acquiring multi-element information of the power substation equipment, the multi-element information including spectral reflectance data of the surface of the power substation equipment and electric field phase data in the surrounding space;

[0007] S2, respectively based on the spectral reflectance data and the electric field phase data, acquiring respective unmanned aerial vehicle pose estimates, and fusing the two unmanned aerial vehicle poses to solve the optimal pose of the unmanned aerial vehicle at the current time k ;

[0008] S3, based on the optimal pose of the current time k and the spectral reflectance data, evaluating the physical state abnormality of the power substation equipment;

[0009] S4, using a pre-trained equipment state-electric field coupling disturbance model VAE to predict the posterior probability distribution of the physical state abnormality of the power substation equipment under the condition of the spectral reflectance data, based on the posterior probability distribution of the physical state abnormality of the power substation equipment under the condition of the spectral reflectance data, calculating the spectral-phase feature mutual information entropy;

[0010] S5, judging whether the spectral-phase feature mutual information entropy exceeds a preset diagnosis threshold, if it exceeds the preset diagnosis threshold, determining that the power substation equipment at the optimal pose of the unmanned aerial vehicle at the current time k exists equipment failure.

[0011] Optionally, when synchronously acquiring the spectral reflectance data of the surface of the power substation equipment and the electric field phase data in the surrounding space in step S1, the constraint condition for the spectral reflectance data of the surface of the power substation equipment and the electric field phase data in the surrounding space to be synchronous is that the following three Boolean conditions ~ are all true:

[0012] ;

[0013] ;

[0014] ;

[0015] Among them, is the collection timestamp of the spectral reflectance data, is the collection timestamp of the electric field phase data, is the maximum allowed timestamp skew. synchronization quality factor at the index of the spectral reflectance data, the spectral synchronization quality factor being the synchronization accuracy between the terminal clock and the master clock when the spectral reflectance data is collected, synchronization quality factor at the index of the electric field phase data, the electric field synchronization quality factor being the synchronization accuracy between the terminal clock and the master clock when the electric field phase data is collected.

[0016] Optionally, step S2 comprises:

[0017] S2.1, matching the spectral reflectance data as landmark observations with a preset geometric-spectral map to obtain a first pose estimate , and calculating a spectral landmark re-projection error of each matched feature point , the geometric-spectral map comprising mapping relationships between different poses of different substation equipment and corresponding spectral reflectance data thereof; matching the electric field phase data as field observations with a preset electric field phase map to obtain a second pose estimate , and calculating an electric field measurement residual , the electric field phase map comprising mapping relationships between different poses and corresponding electric field phase data thereof; calculating a spectral localization confidence of each matched point by a Huber cost function ; calculating an electric field localization confidence of each matched point by a Huber cost function ;

[0018] S2.2, constructing a factor graph G: adding an optimal pose of a previous time instant k-1 , a real-time pose to be solved at a current time instant k two variable nodes, adding an IMU factor as a factor node between the two variable nodes, the IMU factor taking a pre-integration value of an IMU sensor, if the spectral landmark re-projection error of each matched feature point is lower than a preset minimum confidence threshold in the current period, adding a spectral factor connected with the real-time pose at the current time instant k as a factor node for each matched feature point, the spectral factor taking the spectral landmark re-projection error of the matched feature point ; if the electric field localization confidence is lower than a preset minimum confidence threshold in the current period, adding an electric field factor connected with the real-time pose at the current time instant k The connected electric field factor is taken as a factor node, and the value of the electric field factor is an electric field positioning reliability Thus, a factor graph G composed of variable nodes and factor nodes is obtained.

[0019] S2.3, calling a nonlinear optimizer to solve the constructed factor graph G to obtain a real-time pose of the current time k that minimizes the sum of the weighted errors of all factor nodes The real-time pose of the current time k is taken as the optimal pose of the current time k The real-time pose of the current time k is taken as the optimal pose of the current time k Output.

[0020] Optionally, step S3 includes:

[0021] S3.1, indexing the corresponding device identifier of the substation equipment from the pre-set geometric-spectral map based on the real-time pose of the unmanned aerial vehicle, the geometric-spectral map including a mapping relationship between different poses of different substation equipment and corresponding spectral reflectivity data thereof;

[0022] S3.2, querying the corresponding equipment health state spectral database according to the device identifier of the substation equipment to obtain reference spectral reflectivity data of the substation equipment in a healthy state;

[0023] S3.3, taking the statistical distance between the spectral reflectivity data of the substation equipment and the reference spectral reflectivity data as the physical state abnormality degree of the substation equipment.

[0024] Optionally, step S4 includes:

[0025] S4.1, judging whether the physical state abnormality degree of the substation equipment is less than a preset minimum abnormality threshold, if yes, taking the spectral-phase feature mutual information entropy of the substation equipment as 0; jumping to step S5; otherwise, jumping to step S4.2.

[0026] ​​S4.2, calculate the overall probability distribution of the physical state anomaly of the substation equipment and its marginal entropy based on the historical observation sample database of the physical state anomaly of the substation equipment; use the pre-trained Equipment State-Electric Field Coupled Disturbance Model (VAE) to predict the posterior probability distribution of the physical state anomaly of the substation equipment under the given spectral reflectance data, and calculate the conditional entropy of the physical state anomaly with respect to the spectral reflectance data; subtract the conditional entropy from the marginal entropy to obtain the spectral-phase feature mutual information entropy, and proceed to step S5; the pre-trained Equipment State-Electric Field Coupled Disturbance Model (VAE) establishes the mapping relationship between the input physical state anomaly of the substation equipment and the spectral reflectance data, and the output posterior probability distribution of the physical state anomaly of the substation equipment under the given spectral reflectance data.

[0027] Optionally, in step S5, the optimal pose of the UAV at the current time k is determined. After a substation equipment failure is detected, the process also includes: using a pre-trained gradient boosting decision tree model to obtain the corresponding physical root cause identifier of the failure, based on the physical state anomaly degree and spectral reflectance data of the substation equipment. The physical root cause identifier of the fault This is a unique code that identifies the specific physical cause of a fault; the gradient boosting decision tree model pre-trains the input substation equipment's physical state anomaly degree and spectral reflectance data, and outputs a fault physical root cause identifier. The mapping relationship between the fault and its category probabilities is established, and the physical root cause of the fault with the highest category probability is identified. As the final physical root cause identifier of the fault The spectral reflectance data is used to obtain the corresponding electric field perturbation feature labels using a pre-trained k1-nearest neighbor classifier. The electric field disturbance feature identifier To uniquely identify fault-induced electric field disturbance patterns, the k1-nearest neighbor classifier extracts features from spectral reflectance data, calculates the distance between the extracted features and basic features in a feature library of various fault-induced electric field disturbance patterns, identifies the k1 nearest neighbor basic features, and votes on the fault-induced electric field disturbance patterns corresponding to the k1 nearest neighbor basic features to obtain the final tree-thick electric field disturbance feature identifier. .

[0028] Optionally, in step S5, the optimal pose of the UAV at the current time k is determined. After a substation equipment failure occurs, the process also includes: inputting the spectral-phase feature mutual information entropy into a logical sigmoid function model to obtain the corresponding failure evolution probability. The function expression of the logic S-shaped function model is:

[0029] ;

[0030] Wherein, The spectrum-phase feature mutual information entropy.

[0031] The application further provides a substation unmanned aerial vehicle inspection positioning system based on multi-element information fusion, comprising a microprocessor and a memory connected with each other, and the microprocessor is programmed or configured to execute the substation unmanned aerial vehicle inspection positioning method based on multi-element information fusion.

[0032] The application further provides a computer readable storage medium, which stores a computer program or instructions, and the computer program or instructions are programmed or configured to execute the substation unmanned aerial vehicle inspection positioning method based on multi-element information fusion by means of a processor.

[0033] The application further provides a computer program product, which comprises a computer program or instructions, and the computer program or instructions are programmed or configured to execute the substation unmanned aerial vehicle inspection positioning method based on multi-element information fusion by means of a processor.

[0034] Compared with the prior art, the application mainly has the following beneficial effects: 1. For the singularity of positioning information, an adaptive factor graph optimizer is used to weight and fuse the visual positioning result based on spectral features and the new positioning result based on spatial electric field phase features. The optimizer can evaluate the confidence of each positioning source in real time and dynamically adjust the contribution weight of each positioning source in the final pose solution, so that a continuous, accurate and long-term drift-free unified pose can be ensured in the challenging environment of GNSS failure and visual feature repetition. 2. For the fragmentation of observation data, a device state-electric field coupling disturbance model is introduced as a “fusion bridge”, and a new fusion parameter, spectrum-phase feature mutual information entropy, is calculated based on the model. The parameter can quantitatively measure the statistical correlation between the device physical state anomaly reflected by the spectral data and the local electric field disturbance perceived by the electric field data at a certain accurate space-time point. This design enables the system to establish a causal chain from “physical appearance” to “electrical features”, and improves the diagnosis from isolated phenomenon description to deep mechanism tracing. Therefore, the application can solve the problems of singularity of positioning information and environmental vulnerability and the problems of fragmentation of observation data and lack of causal relationship in the traditional substation unmanned aerial vehicle inspection positioning method, and improve the accuracy, reliability and explainability of the substation unmanned aerial vehicle inspection positioning. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a basic flowchart of the embodiment method of the application.

[0036] Figure 2 The module logic division and work flow diagram of the embodiment method of the present application.

[0037] Figure 3 The work flow diagram of the cause-effect correlation analysis module in the embodiment of the present application.

[0038] Figure 4 The work flow diagram of the positioning diagnosis generation module in the embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the person skilled in the art better understand the technical solutions of the present application, the technical solutions of the present application will be further described in detail below in combination with the drawings in the embodiments of the present application.

[0040] As shown in the figure, Figure 1 The substation unmanned aerial vehicle inspection positioning method based on multi-element information fusion in the embodiment includes the following steps:

[0041] S1, synchronously acquiring multi-element information of the substation equipment, the multi-element information including spectral reflectance data of the surface of the substation equipment and electric field phase data in the surrounding space;

[0042] S2, respectively acquiring respective unmanned aerial vehicle pose estimates based on the spectral reflectance data and the electric field phase data, and fusing the two unmanned aerial vehicle poses to solve the optimal pose of the unmanned aerial vehicle at the current time k ;

[0043] S3, evaluating the physical state abnormality degree of the substation equipment based on the optimal pose of the current time k and the spectral reflectance data;

[0044] S4, using the pre-trained equipment state-electric field coupling disturbance model VAE to predict the posterior probability distribution of the physical state abnormality degree of the substation equipment under the condition of the spectral reflectance data, and calculating the spectral-phase feature mutual information entropy based on the posterior probability distribution of the physical state abnormality degree of the substation equipment under the condition of the spectral reflectance data;

[0045] S5, judging whether the spectral-phase feature mutual information entropy exceeds a preset diagnosis threshold, if it exceeds the preset diagnosis threshold, it is determined that the substation equipment at the optimal pose of the unmanned aerial vehicle at the current time k exists equipment failure.

[0046] As an optional implementation manner, as shown in the figure, Figure 2As shown, step S1 in the embodiment is implemented by using a data synchronization acquisition module, step S2 is implemented by using a hybrid positioning module, step S3 is implemented by using a physical state evaluation module, step S4 is implemented by using a causal correlation analysis module, and step S5 is implemented by using a positioning diagnosis generation module.

[0047] When the spectral reflectance data of the substation equipment surface and the electric field phase data in the surrounding space are acquired synchronously in step S1 of the embodiment, the constraint condition for the spectral reflectance data of the substation equipment surface and the electric field phase data in the surrounding space to be synchronous is to satisfy the following three Boolean conditions True at the same time:

[0048] ;

[0049] ;

[0050] ;

[0051] wherein, is the collection timestamp of the spectral reflectance data, is the collection timestamp of the electric field phase data, is the maximum allowed timestamp skew, is the spectral synchronization quality factor at index in the spectral reflectance data, which is the synchronization accuracy between the terminal clock and the master clock when the spectral reflectance data is collected, is the electric field synchronization quality factor at index in the electric field phase data, which is the synchronization accuracy between the terminal clock and the master clock when the electric field phase data is collected.

[0052] In step S1 of the embodiment, the collection timestamps of the spectral reflectance data and the electric field phase data are synchronized to nanosecond-level accuracy by a high-precision clock synchronization protocol to generate a synchronous data pair containing the same timestamp. The spectral reflectance data is processed into a normalized multispectral reflectance feature vector ; and the electric field phase data is processed into a multi-point induced electric field synchronous phase difference vector The data synchronization acquisition module broadcasts a unified time reference to the multispectral camera and the electric field sensor array through a precision time protocol (PTP). Each sensor attaches a high-precision timestamp generated by the unified time reference to its data frame at the moment of data collection, with a resolution of nanoseconds . Subsequent modules only process data pairs with the same or within a preset time window timestamp.​

[0053] It is further explained that the execution of this embodiment relies on the following defined key parameters, specifically: the master clock timestamp is denoted as T master , which is a nanosecond resolution timestamp generated by the system master clock as a global unified time reference; it is produced by a Grandmaster-Clock (GMC) compliant with IEEE-1588v2 (PTP) protocol deployed on the UAV or ground station. This clock is usually disciplined by GPS signals to ensure long-term stability and absolute time accuracy. The sensor acquisition timestamp is denoted as T sensor , which is the timestamp acquired from the local PTP slave clock of the multispectral camera or electric field sensor at the moment of completing a data acquisition exposure or sampling, synchronized with T master ; it is provided by a PTP slave clock module (Slave-Clock) integrated on the sensor hardware interface or data acquisition card (DAQ). This module continuously synchronizes with the Grandmaster-Clock and attaches the current timestamp to each output data frame; the normalized multispectral reflectance feature vector is denoted as , which is an N-dimensional vector whose dimensions correspond to the corrected and normalized reflectance values of the device surface under different spectral bands; it is an N-dimensional vector representing the spectral characteristics of the current observation point surface, which is collected and processed by the UAV at time k. Its specific acquisition approach and determination method are as follows: the original multispectral image containing the target device surface is collected by the onboard multispectral camera (MicaSense-RedEdge-P model); secondly, the center region of the image is processed by a pre-processing algorithm, which includes radiometric calibration (converts the original DN value to radiance value) and reflectance calculation (converts the radiance value to reflectance value independent of illumination using illumination sensor data), to obtain the reflectance value of each band; finally, the reflectance values are normalized. As a numerical calculation example: a four-band camera measures the reflectance values of each band of a point as [0.2, 0.3, 0.4, 0.1], and since the sum is exactly 1.0, the processed , which is the vector [0.2, 0.3, 0.4, 0.1]. The multi-point induced electric field synchronous phase difference vector is denoted as , which is an M-dimensional vector whose dimensions correspond to the phase difference of the induced electric field signal measured by M different positions of the electric field probe on the UAV relative to a high-stability local reference clock; the original induced voltage signal is collected by M high-impedance electric field probes installed on the UAV wings, fuselage, etc. Processing: the signal is processed by a phase-locked amplifier, which uses a master clock timestamp T masterThe synchronized local reference frequency (50 Hz), the phase difference between each probe signal and the reference frequency is calculated, forming a vector. The maximum allowed timestamp skew is denoted as , which is the maximum tolerable time difference for judging whether two data points from different sensors belong to the same transient event; it is an empirical parameter determined according to the physical target. To capture partial discharge events in the order of microseconds, it needs to be set to a value significantly smaller than 1 microsecond, specifically, it is set to 100 nanoseconds in the present embodiment. This value ensures that the temporal correlation is much higher than the rate of change of the observed physical phenomenon. The synchronization quality factor is denoted as Q sync ; it is a dimensionless value quantifying the synchronization accuracy between the PTP slave clock and the master clock at a certain time; it is calculated in real time by the PTP protocol stack software. Its calculation takes into account two core indicators: the offset from master and the mean path delay. The minimum quality factor threshold is denoted as Q min , which is the minimum synchronization quality factor required to judge whether a synchronization is valid; it is an engineering threshold set according to the system accuracy requirement. If the timestamp error is required to be less than 50 nanoseconds, the corresponding Q sync value needs to be deduced according to the calculation formula of Q min . Specifically, in the present embodiment, Q min is set to a value corresponding to an offset from master less than 50 nanoseconds; the valid synchronization data pair is denoted as P sync , which is a structured data object containing the timestamp, spectral feature vector, electric field phase vector, and quality factor, and is the final output of the module; it is generated after pairing and processing the original data that meets all synchronization and quality check conditions. The input of the data synchronization acquisition module is the following two parallel and asynchronous data streams: spectral data stream: composed of a series of data packets, each containing {raw multispectral image data, synchronization timestamp , spectral synchronization quality factor } at the moment the spectral reflectance data frame is collected. Among them, spec is the data packet index mark of the spectral data stream; electric field data stream: composed of a series of data packets, each containing {raw multi-point induced voltage signal, synchronization timestamp , electric field synchronization quality factor } at the moment the electric field phase data sample is collected. phase is the data packet index mark of the electric field data stream; data buffering: the system maintains two independent, time-based first-in-first-out (FIFO) buffers: spectral buffer and electric field buffer. Each buffer stores the data packets that arrive within the last W milliseconds (W=100 ms in the present embodiment). Reference data selection and iteration: the system takes the spectral buffer as the reference, and sequentially takes out the earliest data packet and records it as Candidate pair search: For the retrieved data packets The system searches the electric field buffer for all data packets that satisfy the following time proximity condition, and denotes them as follows: The temporal proximity condition is characterized as follows:

[0054] ;

[0055] in, Indicates obtaining the name The data packet contains the collection timestamp. The values ​​of the fields below; Indicates obtaining the name The data packet contains the collection timestamp. The value of the following field; Synchronization quality check: One or more of the conditions for time proximity were found. The system for each Perform the following synchronous quality gating judgment:

[0056] and ;

[0057] Choose the timestamp closest to of As the best match, and denoted as .

[0058] and The description of the occurrence and The same applies, so I won't elaborate further; if the synchronization quality of all candidates fails to meet the standard, the match is considered a failure, and the current candidate is discarded. The process returns to the baseline data selection and iterative processing of the next spectral data packet. If no temporally neighboring candidate is found in the candidate pair search, it is considered a match failure, and the current pair is discarded. The system then returns the reference data selection and iterative processing for the next spectral data packet. Data pair generation and encapsulation: The system generates and encapsulates successfully matched data pairs. and The original data in the dataset is processed to calculate the normalized multispectral reflectance eigenvectors, which are denoted as follows: ; and the synchronous phase difference vector of the multi-point induced electric field, denoted as Generate the final valid synchronized data pair and denote it as... Its data structure is as follows:

[0059] ;

[0060] in, This serves as a representative timestamp for the synchronization event. Output to the subsequent causal relationship analysis module.

[0061] One key of the method of the embodiment is to fuse the information of the two dimensions of time proximity and timestamp reliability to make a binary decision. A deterministic logic AND-gate gating method based on multiple thresholds is adopted. The reason for selecting this method is that the effectiveness of data synchronization is a problem of the rule of the weakest link, and any dimension that does not meet (the time skew is too large, and any sensor clock is out of lock) will lead to the invalidity of the entire data pair. The deterministic logic gate can most directly and unambiguously implement this one-vote veto strong constraint, avoiding the risk that an essentially invalid data pair may be misjudged as valid due to improper weight setting in linear weighting and other methods.

[0062] Maximum allowed timestamp skew Deterministic logic: the value is determined by the characteristic time scale of the physical phenomenon to be observed To ensure that the same event is captured, must be much smaller than Its determination principle is: For example, for partial discharge with a characteristic time of 1 microsecond ( ), is set to 100 nanoseconds ( ).

[0063] Minimum quality factor threshold Q min Deterministic logic: the value is determined by the requirement of the system for the absolute accuracy of the timestamp Q min must be set to a level that ensures that the master-slave clock offset reported by the PTP protocol is less than In this embodiment, if the system requires that the timestamp error does not exceed 50 nanoseconds, Q min is set to the quality evaluation value corresponding to in the PTP protocol stack. Among them, is the time difference between the local clock of each sensor (multispectral camera, electric field sensor) on the UAV and the system master clock (such as the PTP master clock), in nanoseconds (ns).

[0064] There is a non-compensatory, serial logical relationship between the input parameters and the final output (whether the data pair is valid). The timestamp difference is negatively correlated with validity, and has a hard cutoff threshold . The spectral synchronization quality factor is positively correlated with validity, and has a hard lower threshold . The electric field synchronization quality factor is positively correlated with validity, and has a hard lower threshold These three conditions must be satisfied simultaneously, there is no case that one parameter performs well can compensate another parameter performs poor. Valid synchronized data pair is generated if and only if the following three Boolean conditions are true simultaneously: , which represents the absolute value of the difference between the timestamps of the spectral data and the electric field data, must be less than or equal to a preset maximum allowed timestamp skew. , which represents the quality factor of the spectral data, must be no less than a minimum quality factor threshold. The spectral synchronization quality factor is obtained by calculating the sharpness (e.g. Laplacian operator gradient value) or signal-to-noise ratio of the spectral image. , which represents the quality factor of the electric field data, must be no less than a minimum quality factor threshold. The electric field synchronization quality factor is obtained by calculating the signal-to-noise ratio (SNR) of the electric field signal or its initial deviation from a theoretical benchmark model. In this embodiment, the is set to 0.7; the sufficient and necessary condition for the generation of a valid synchronized data pair is that the absolute difference between its spectral data acquisition timestamp and electric field data acquisition timestamp is no more than a preset maximum allowed timestamp skew; and the synchronization quality factor corresponding to the spectral data acquisition timestamp is no less than a preset minimum quality factor threshold; and the synchronization quality factor corresponding to the electric field data acquisition timestamp is also no less than the minimum quality factor threshold.

[0065] In this embodiment, the spectral localization confidence is defined as ; and the electric field localization confidence is defined as . Two localization factors are constructed in the factor graph optimizer: spectral factor and electric field factor. The hybrid localization module evaluates the quality of the two localization methods in real time: the spectral localization confidence is determined by calculating the re-projection error of the spectral matching; and the electric field localization confidence is determined by evaluating the signal-to-noise ratio of the electric field signal. In the factor graph optimization solving process, the hybrid localization module determines the localization confidence based on the spectral localization confidence and the electric field localization confidence.The two factors are dynamically weighted to achieve optimal fusion of complementary advantages. The spectral positioning is accurate in the feature-rich area but is susceptible to occlusion, while the electric field positioning is not affected by occlusion but has lower accuracy at a distance. The dynamically weighted factor graph fusion framework can automatically rely on more reliable sources in different scenarios, ensuring the continuity and robustness of positioning. It should be further noted that the embodiment details a hybrid positioning method for robust pose estimation of a UAV in a strong electromagnetic interference and visual feature repetition environment. The method specifically involves a hybrid positioning module for performing step S2, which processes spectral landmark observations and electric field intensity observations in parallel, and performs weighted fusion based on dynamically evaluated confidence of each observation source within a factor graph optimization framework, thereby generating a continuous, accurate, and drift-free real-time pose. Specifically, step S2 in the embodiment includes:

[0066] S2.1, the spectral reflectance data is taken as landmark observations, and a pre-set geometric-spectral map is matched with the feature points to obtain a first pose estimate , and the spectral landmark re-projection error of each matched feature point is calculated , the geometric-spectral map includes the mapping relationship between different poses of different substation equipment and their corresponding spectral reflectance data; the electric field phase data is taken as field observations, and a pre-set electric field phase map is matched with the feature points to obtain a second pose estimate , and the electric field measurement residual is calculated , the electric field phase map includes the mapping relationship between different poses and their corresponding electric field phase data; the spectral landmark re-projection error of each matched point is calculated by a Huber cost function to obtain the confidence of spectral positioning ; the electric field measurement residual is calculated by a Huber cost function to obtain the confidence of electric field positioning ;

[0067] S2.2, constructing a factor graph G: adding the optimal pose of the previous time k-1 , the real-time pose to be solved at the current time k two variable nodes, adding a connected IMU factor as a factor node between the two variable nodes, and the IMU factor takes the pre-integration value of the IMU sensor; if the spectral landmark re-projection error of the matched feature point in the current period is lower than the preset minimum confidence threshold, adding a factor node for the real-time pose at the current time k The connected spectral factor as a factor node, the value of the spectral factor is the spectral landmark re-projection error of the matching feature point ; if there is no electric field positioning confidence in the current period below the preset minimum confidence threshold, add a real-time pose of the current time k The connected electric field factor as a factor node, the value of the electric field factor is the electric field positioning confidence , so as to obtain a factor graph G composed of variable nodes and factor nodes;

[0068] S2.3, calling a nonlinear optimizer to solve the constructed factor graph G to obtain the real-time pose of the current time k which minimizes the sum of the weighted errors of all factor nodes , the real-time pose of the current time k as the optimal pose of the current time k Output.

[0069] Denote the geometry-spectrum map as , which is a pre-constructed map containing three-dimensional space points and their associated spectral features. By offline scanning of the substation, using a UAV platform combined with Li-DAR and multispectral camera, running a multi-modal SLAM algorithm (such as an extended version of VINS-Fusion) to generate. Each three-dimensional point in the map is assigned a normalized multispectral reflectance feature vector of dimension N .

[0070] Denote the electric field phase map as , which is a pre-constructed map describing the mapping relationship between three-dimensional coordinates and theoretical electric field phase vectors in the substation space. Based on the CAD design diagram of the substation, using finite element analysis software (COMSOL-Multiphysics) to perform electromagnetic field simulation, obtaining the multi-point induced electric field synchronous phase difference vector of each point in space relative to the reference power source phase under healthy operating conditions. Acquisition method: discretize the simulation results and store them as a three-dimensional grid or K-D tree structure for fast querying.

[0071] Denote the first pose estimate as , the current pose of the UAV calculated only by matching spectral data with the geometry-spectrum map; by extracting feature points in the current multispectral image and matching them with the spectral features in , then using a robust Perspective-n-Point (PnP) algorithm (EPnP combined with RANSAC) to solve.

[0072] Denote the second pose estimate as , the current pose of the UAV only solved by matching the electric field phase data with the electric field phase map. By comparing the real-time measured multi-point induced electric field synchronous phase difference vector ( ) with , using particle filtering or nonlinear optimization algorithm, searching and determining the map position that best matches the measured value, which is the .

[0073] Let the spectral landmark re-projection error be , representing the difference between the position of the spectral landmark in the map projected onto the image and the actual observed position under the current pose estimate; Specifically, a three-dimensional landmark point in the is projected back to the current image plane using , and the Mahalanobis-distance between the spectral features of the projected position and the spectral features stored in the landmark point is calculated. This distance is . Let the electric field measurement residual be , representing the difference between the theoretical electric field phase queried from the map and the actual measured electric field phase under the current pose estimate. Using , the theoretical electric field phase difference vector of this position is queried in , and the Euclidean distance between it and the real-time measured is calculated; this Euclidean distance is . Let the spectral localization confidence be , a dimensionless value, quantifying the reliability of the current spectral localization result, the higher the value the more reliable; the calculation model is based on robust statistical theory. Specifically, by inputting the spectral landmark re-projection error into a Huber robust cost function. This function gives higher weight to smaller errors (inliers) and lower and constant weight to larger errors (outliers), so that is inversely proportional to , and can effectively suppress the influence of outliers. Let the electric field localization confidence be , a dimensionless value, quantifying the reliability of the current electric field localization result, the higher the value the more reliable. The calculation model is also based on robust statistical theory. By inputting the electric field measurement residual ( ) into a corresponding Huber robust cost function. This way makes be inversely proportional to , while being insensitive to large residuals caused by sudden strong electromagnetic interference; Let the factor graph be G, a probabilistic graph model, used to represent variables (poses) and constraints (observations) in the pose optimization problem; dynamically constructed by the system at each time step, containing variable nodes (the UAV pose to be optimized ) and factor nodes (constraints from IMU, spectral observations and electric field observations). Let the final real-time pose be denoted as , the final pose estimate that fuses all information sources after optimization of the factor graph G; the factor graph G is solved by a non-linear least squares optimizer (Levenberg-Marquardt algorithm) to find the UAV pose that minimizes the weighted sum of all factor errors. The flowchart of the process to compute the final real-time pose is decomposed as follows: At time k, the inputs of the hybrid localization module include the valid synchronized data pairs , containing the normalized multi-spectral reflectance feature vector and the multi-point induced electric field synchronous phase difference vector . The pre-computed geometric-spectral map and the electric field phase map . The raw readings of the inertial measurement unit (IMU) between time k-1 and k. The hybrid localization module initiates two parallel processing branches simultaneously: Spectral branch: feature matching between the current frame and to solve the first pose estimate and to compute the spectral landmark re-projection error for each matched point . Electric field branch: matching between the current measured and to solve the second pose estimate and to compute the electric field measurement residual . Confidence dynamic assessment: based on all computed by the spectral branch, the integrated spectral localization confidence is computed by Huber cost function. Based on computed by the electric field branch, the electric field localization confidence is computed by Huber cost function. In the factor graph G, the following nodes are constructed for time k: (1) variable nodes: add a variable node representing the current UAV pose to be optimized, and a variable node representing the pose at time k-1 . (2) IMU factor: add an IMU pre-integration factor connecting the pose at time k-1 and the current UAV pose , which represents the high-frequency motion constraint. (3) spectral factor: for each successfully matched spectral landmark, add a factor connecting the UAV pose and the landmark. The strength of this factor is determined by the spectral localization confidence . (4) electric field factor: add a factor connecting the UAV pose The electric field factor represents the constraint of the current electric field measurement on the pose. Its strength is determined by the electric field positioning confidence level. Decision. In the factor graph construction, a minimum confidence threshold is set; in this example, it is 0.1. If... If the confidence level is below this minimum threshold, no spectral factor will be added for this period. This corresponds to situations where the drone enters a completely featureless area or is severely obstructed. If the value is below the minimum confidence threshold, no electric field factor is added in this period. This corresponds to the UAV entering areas with extremely weak electric field signals or encountering extreme electromagnetic interference. The gating logic in this example ensures that only reliable observation data is used for the final optimization, preventing low-quality data from corrupting the positioning results. In this embodiment, the constructed factor map G is finally solved by calling a nonlinear optimizer to obtain the pose state that minimizes the sum of the weighted errors of all factors. The optimized pose is then used as the final real-time pose. Output it and use it as the next cycle. This embodiment employs probabilistic optimization fusion based on factor graphs. Compared to traditional Kalman filters, the factor graph framework: supports nonlinear models: it can directly handle inherent nonlinear observation models such as camera projection and electric field distribution without forced linearization; supports asynchronous multi-sensor operation: it naturally supports sensor data input with different frequencies and delays; and achieves global optimization: it can perform batch optimization (or sliding window optimization) on multiple poses and constraints within a time window, effectively suppressing error accumulation and long-term drift, and providing better global consistency than recursive filtering methods. Measurement residuals or The relationship between the strength of the corresponding information matrix and the cost is non-linear and inversely proportional, defined by the Huber robust cost function. The contributions of all input sources (IMU, spectrum, electric field) are cooperative and non-competitive within the factor graph framework. They collectively define a unified total cost function that needs to be minimized. The system seeks a pose solution that reconciles all sensor observations with minimal cost. The final real-time pose is calculated by finding a pose state that minimizes the value of a total cost function. This total cost function is the sum of multiple cost terms. These cost terms include: a motion cost term generated from inertial measurement unit (IMU) data; a spectral observation cost term generated from spectral landmark observations and weighted by their corresponding spectral positioning confidence; and an electric field observation cost term generated from electric field strength observations and weighted by their corresponding electric field positioning confidence.

[0074] Further explanation regarding hybrid positioning: In a controlled indoor or small-scale outdoor verification environment, a separate optical 3D motion capture system (Vicon™ or OptiTrack™ system) with accuracy far exceeding that of the system under test is employed; a) System deployment: High-frequency infrared cameras are deployed at multiple locations in the verification site to construct a high-precision 3D motion capture space covering the entire flight area. b) Marker installation: Multiple reflective markers are fixed on the rigid body of the UAV; c) Coordinate system one: Through a calibration procedure, the coordinate system of the optical motion capture system is precisely aligned with the geometric-spectral map used in this invention. Phase map of electric field Coordinate system one. d) Synchronous acquisition: During the UAV's flight mission, the optical motion capture system calculates the three-dimensional spatial position of the reflective markers in real time at a frequency of ≥100Hz, thereby obtaining the UAV's true ground pose at each time point. Simultaneously, the hybrid positioning module of this invention outputs its estimated final real-time pose. The two data streams are strictly timestamped using the Network Time Protocol (NTP). In a large-scale outdoor verification environment: a standalone, measurement-grade differential GPS system combined with a high-precision inertial navigation system (INS) is employed. System deployment: A combined navigation system including an RTK-GPS receiver and a tactical-grade or navigation-grade IMU is mounted on the UAV. Simultaneously, RTK base stations are set up at known precise coordinates on the ground. Data acquisition: The RTK system provides centimeter-level real-time position information by receiving differential correction signals from the base stations. The high-precision INS provides high-frequency, low-drift attitude information. Data post-processing: To obtain the highest accuracy ground truth, the raw RTK and INS data acquired during flight are offline post-processed and tightly coupled to generate a smooth and accurate ground truth trajectory, which serves as the ground truth trajectory. Similarly, the timestamp of this trajectory needs to be consistent with... Alignment is performed. The final real-time pose is then determined under the same timestamp. The included position components and the ground true pose The three-dimensional Euclidean distance between the included position components is used as the final pose error. Final pose error Used to characterize the degree of consistency between the final real-time pose output by the hybrid positioning module and the physical truth of the UAV. Final pose error Its unit is centimeters (cm), and its theoretical range is... When the final pose error When the value approaches 0, it indicates that the final real-time pose output by the hybrid positioning module of this invention is highly consistent with the physical truth value of the UAV, resulting in high positioning accuracy. This enables the UAV to safely perform close-range, precise inspection tasks. When the final pose error... As the numerical value increases, the deviation in the positioning result increases, and the reliability of the system decreases. When this error exceeds a preset safety threshold, it means that the positioning function may fail. The preset safety threshold is determined by an expert group based on experimental data; in this embodiment, the initial safety threshold is set at 30cm; this affects the final pose error. The core intermediate parameter is the spectral location reliability. and electric field location confidence The aforementioned With final pose error There is a negative correlation between them. The value is derived from the spectral landmark reprojection error. The robust cost function is calculated when the drone is in an area with rich visual features and no obstructions. The value is low, resulting in The value approaches 1. In the factor graph optimization framework, a high... This generates a strong information matrix, providing robust geometric constraints for pose optimization. Therefore, The increase will lead to the final pose error. Monotonically decreasing. This design accurately maps to the physical reality that high-quality visual observations can improve positioning accuracy. With final pose error There is also a negative correlation between them. The value is derived from the electric field measurement residual. Calculations show that when the drone is in an area with a stable electric field distribution and far from strong interference sources, The value is low, resulting in The value approaches 1. A high This will also provide a strong constraint in factor graph optimization. Therefore, The increase in [something] will also lead to the final pose error. Monotonically decreasing. This design accurately maps the physical logic that a stable electric field signal matched to the map can improve positioning accuracy. The core of the factor graph optimizer lies in its ability to adjust based on the... and A dynamically generated information matrix is ​​used to allocate the contribution of different observation sources. When Gao Er When the spectral density is low, the optimizer primarily relies on the spectral factor to correct the pose; conversely, when it is high, the optimizer does not. This dynamic weighted fusion mechanism ensures the final... The most reliable source is determined by the current, rather than being contaminated by unreliable sources. In scene one: ideal environment, scene two: visual degradation (thick fog), scene three: strong electromagnetic interference, scene four: double degradation, scene five: partial visual occlusion and scene six: visual feature repetition zone, the technical performance verification experiment of the optimizer is verified, and the technical performance verification data and results are shown in Table 1.

[0075] Table 1: Technical performance verification data and results of the optimizer

[0076]

[0077] As can be seen from Table 1, compared with the data of scene one, scene two and scene three, in the ideal scene one, and are very high, and the system realizes the highest accuracy positioning of 2.1 cm. In scene two, due to visual degradation caused by thick fog, drops to 0.10, but still remains at a high level of 0.93, the system automatically relies on the electric field constraint, and the final pose error only increases slightly to 4.8 cm, still maintaining centimeter-level accuracy. Similarly, in scene three, when the electric field encounters strong interference, causing to drop to 0.15, the system automatically relies on the high-confidence spectral constraint ( ), and is controlled at 5.2 cm. These three groups of data powerfully prove the core advantage of the fusion mechanism described in the embodiment: compared with any single source positioning technology (which will directly fail in scene two or scene three), the embodiment realizes system-level robustness in the case of failure of a single sensing modality through dynamic evaluation and fusion, ensuring high-precision positioning capability in all scenes. The data of scene four shows that when two sensing modalities are severely degraded at the same time ( and are less than 0.2), the final pose error increases sharply and exceeds the available range, and the system can recognize the positioning loss state. This proves that the embodiment not only provides accurate results when available, but also provides explicit failure warnings that can be used for decision-making when unavailable, avoiding the risk of incorrect navigation. Scenes five and six simulate more common complex situations. Even when the visual is partially occluded or the features are repeated, causing to be medium or low, as long as the electric field positioning remains reliable, the system can still output high-precision positioning results (4.5 cm and 6.7 cm). This proves the practical application value of the embodiment in complex structures such as substations and environments with variable visual conditions.

[0078] Based on a large amount of technical performance verification data analysis, combined with the specific requirements of substation inspection tasks for safety and accuracy, the final pose error Three clear, operational, and directly related to the UAV flight control system application intervals are set: (1) High-precision inspection interval, range [0cm, 10cm], positioning results are highly reliable, meeting the accuracy requirements of all fine inspection tasks. All pre-set tasks are allowed to be executed: including high-speed autonomous flight, close to equipment (such as insulator strings, transformer bushings) for less than 1.5 meters distance of fixed-point hovering observation. (2) Cautionary flight interval, range (10cm, 30cm] positioning results are basically reliable, but there is a certain deviation, which does not meet the close observation requirements, but is sufficient to support safe navigation. Limit tasks and adjust flight parameters: automatically stop close observation tasks, increase the lower limit of the safe distance from the equipment to 3 meters, and limit the maximum flight speed to 50% of the pre-set value, while trying to reacquire higher-precision positioning through small-range maneuvering. (3) Positioning warning interval, range >30cm positioning results are unreliable, there is a risk of collision with obstacles. Trigger emergency avoidance and relocation procedures: immediately stop the current task, autonomously climb to an absolute safe height (such as 5 meters above all equipment) in the region, and perform a hovering relocation procedure. If it cannot recover to the high-precision inspection interval within the pre-set time (such as 30 seconds), send an alarm to the ground station and request manual takeover.

[0079] Step S3 of the embodiment is implemented based on a physical state evaluation module. The physical state evaluation module is specifically configured to: based on the final real-time pose, index the corresponding equipment identifier from the pre-set geometric-spectral map; retrieve the baseline spectral feature vector corresponding to the equipment identifier from the equipment health state spectral database; and calculate the statistical distance between the current real-time collected normalized multi-spectral reflectance feature vector and the baseline spectral feature vector as the physical state abnormality. Specifically, step S3 in the embodiment includes:

[0080] S3.1, based on the real-time pose of the UAV, index the corresponding equipment identifier of the substation equipment from the pre-set geometric-spectral map The geometric-spectral map includes the mapping relationship between different poses of different substation equipment and their corresponding spectral reflectance data;

[0081] S3.2, according to the equipment identifier of the substation equipment, query the corresponding equipment health state spectral database to obtain the baseline spectral reflectance data of the substation equipment in the healthy state;

[0082] S3.3, according to the statistical distance between the spectral reflectance data and the baseline spectral reflectance data of the substation equipment, as the physical state abnormality of the substation equipment.

[0083] For the physical state assessment module, it should be noted that the execution of this embodiment depends on the following key parameters, specifically: uniquely identifying a specific device on the map relies on the device identifier, denoted as... It is a string or integer used to uniquely identify substation equipment (including but not limited to the high-voltage side bushing of the A-phase main transformer) in the map. Its specific method of acquisition at runtime is: through the geometry-spectral map. Execute a function based on the final real-time pose. The nearest neighbor search algorithm. Specifically, from the final real-time pose... Starting from the position component, a virtual ray is emitted along the line of sight indicated by its attitude component. The first device model that this ray intersects within a preset detection distance (10 meters) is identified by its pre-associated identifier as the currently observed device identifier. .

[0084] The criteria for judging the health status of substation equipment are stored in the equipment health status spectral database, and are recorded as follows: This is a database storing the reference spectral feature vectors of various substation equipment under known health conditions. The database was established as follows: when equipment is newly commissioned or confirmed by professionals to be in a fault-free and healthy state, a multispectral camera of the same model as the onboard sensor of a UAV is used to take multiple images of the key observation surfaces of the equipment under standard lighting conditions. The collected data is processed to extract its normalized multispectral reflectance feature vector, and the results of multiple measurements are averaged to eliminate random errors, ultimately obtaining a stable reference vector, which is then correlated with the corresponding equipment identifier. After association, the data is stored in the device's health status spectral database. .

[0085] From the above equipment health status spectral database The health benchmark retrieved from the database is the benchmark spectral feature vector, denoted as... It represents the identifier of the current observation equipment. The corresponding N-dimensional normalized multispectral reflectance vector characterizing its health status. This vector is specifically obtained using device identification. As the primary key, in the device health status spectral database This is obtained through a direct query operation. Each element of the vector represents the normalized reflectance of a specific spectral band, and the sum of all elements is always 1.

[0086] With reference spectral eigenvectors The real-time data used for comparison is the normalized multispectral reflectance eigenvector. To quantify the similarity between two spectral vectors, an intermediate measure, the Barthel coefficient, denoted as BC, is introduced. This is a standard statistical coefficient that measures the similarity between two probability distributions, with a range of [0,1]. A value closer to 1 indicates a higher similarity. Its calculation model is derived from the standard definition in probability theory and statistics. The calculation process is expressed as: For... and Given two vectors, multiply their corresponding elements pairwise, sum the products across all dimensions, and finally take the square root of the sum. As a numerical example: If... And from the database:

[0087] .

[0088] The calculation of BC is as follows:

[0089] .

[0090] The final quantization result output by this module is the physical state anomaly degree, denoted as... This is a dimensionless numerical value used to quantify the degree to which a device's current physical state deviates from its health baseline; the larger the value, the higher the degree of anomaly. Its calculation model is derived from the definition of the Bach distance in information theory. The calculation process is expressed as: calculating the negative natural logarithm of the previously obtained Bach coefficient BC. As a numerical calculation example: based on the BC value of 0.974 obtained in the previous step, then... The result is approximately 0.026. Physical state anomaly degree. The calculation process is broken down as follows: At time k, the initial input of the physical state assessment module is the final real-time pose provided by the hybrid positioning module. Provided by the data synchronization acquisition module, and... Normalized multispectral reflectance feature vector of timestamp synchronization Preloaded geometry-spectral map and equipment health status spectral database Spatial indexing and device identification, based on Position and posture, in The process involves performing a raycast query. Logical judgment is made regarding target validity: if the raycast does not intersect with any device model within a preset detection distance (10 meters), the current field of view is considered empty, the evaluation process for this cycle is terminated, and an empty result is output. If the raycast intersects with a device model, the device identifier of that model is extracted. Proceed to the next step. Benchmark data retrieval: Using logical judgment, data obtained from the target validity process. As a search keyword, in Retrieve the corresponding reference spectral feature vector Logical judgment, benchmark existence: if the search fails (including but not limited to the device being a newly added device, not yet having a health record), mark the device as no benchmark, stop calculation, output If the acquisition is successful , proceed to the following steps. Abnormality quantification calculation: based on the current measured and the retrieved , perform the calculation of the Bhattacharyya distance. First, calculate the intermediate value Bhattacharyya coefficient BC, and then based on BC, calculate the final physical state abnormality degree . The final output of the process is a data pair . The output clearly indicates the specific device being evaluated, uniquely identified by , and gives the degree of deviation of its current physical state from the health benchmark.

[0091] The core of the calculation of this embodiment lies in how to effectively compare two spectral feature vectors, rather than the traditional multi-parameter fusion; the choice of Bhattacharyya distance as the core calculation method, rather than the more common Euclidean distance or cosine similarity, is based on the following technical considerations: the normalized multi-spectral reflectance vector can be mathematically regarded as describing the discrete probability distribution of photon energy at different wavebands. The Bhattacharyya distance is a statistical quantity specifically used to measure the degree of overlap or similarity between two probability distributions, and its selection has a solid theoretical basis. Compared with the Euclidean distance, which is sensitive to the absolute value difference of each component of the vector, the Bhattacharyya distance is more sensitive to the similarity of the overall distribution shape. This means that it can better tolerate small fluctuations in overall reflectance caused by slight changes in illumination or sensor noise, while showing high sensitivity to changes in spectral distribution shape caused by physical state changes (such as rust, dirt accumulation, and heat-induced discoloration) (i.e. significant increases in reflectance at some wavebands and decreases at others). When the real-time measured and the are more similar, the sum of their vector inner products will be greater, resulting in the Bhattacharyya coefficient BC being more close to 1. Since is the negative natural logarithm of BC, at this time will be more close to 0, indicating lower abnormality. Conversely, when the device surface is abnormal (such as a patch of rust causing a sharp increase in reflectance at the red light waveband, while other wavebands decrease), and the distribution shape of will produce a more significant difference, resulting in their BC being more close to 0, while will tend to be a larger positive number, thereby quantitatively indicating higher abnormality. This relationship conforms to the intuitive logic that the greater the difference, the higher the abnormality.

[0092] The physical state abnormality degree The final calculation method is as follows: calculate the eigenvector used to measure the normalized multispectral reflectance. With the reference spectral characteristics The similarity is calculated using the Bach coefficient (BC). The calculation process involves multiplying the elements of the two vectors at the same position along the same dimension pairwise, summing all these products, and finally calculating the square root of the sum. Subsequently, the anomaly degree of the physical state is obtained by calculating the negative natural logarithm of the Bach coefficient (BC). Physical state anomaly degree The theoretical range of values ​​is [0, +∞). This range has a clear and crucial technical meaning: when Approaching 0: This state means that the Bartholomew's coefficient BC approaches 1. This physically characterizes the real-time measurement. Reference spectral eigenvectors of equipment health status The distributions almost completely overlap. Technically, this directly indicates that there is no observable statistical difference between the physical state of the tested device's surface and a preset health baseline, meaning the device is in a healthy or normal state. As it increases: This state means that the Bartholomew's coefficient BC approaches 0. This is physically characterized... and The greater the difference in distribution, the more fundamental the changes in spectral reflectance characteristics can be caused by factors such as rust on the equipment surface, oil contamination, or aging and deterioration of insulating materials. Technically, The greater the increase, the higher the degree of anomaly present on the surface of the tested equipment, directly and quantitatively indicating this. If the physical state of the equipment surface changes, from smooth ceramic to a corroded metal surface, its ability to reflect different wavelengths of light will inevitably change, thus leading to its... The distribution shape deviates from the original health baseline The Bach coefficient (BC) is a classic statistical tool used to quantify such differences in distribution patterns. When the difference between two vectors increases, their BC value inevitably decreases. By using the negative natural logarithm -ln(·) for transformation, this design cleverly achieves two objectives: firstly, it ensures the output value... The positive correlation with the vector difference degree, the BC decreases, and the -ln(BC) increases. Secondly, by using the characteristics of the logarithmic function, the area where the BC is close to 1 is stretched, so that the algorithm has higher sensitivity to the initial small changes of the device state, that is, the initial failure. At the same time, the area where the BC is close to 0 is compressed, preventing the output value from exploding in meaningless growth in extreme abnormal situations. This design accurately maps the state changes in the physical world to mathematical measures with clear diagnostic significance. Table 2 shows the quantitative evaluation results of the method described in this embodiment for the same device (transformer high-voltage side bushing) under different surface physical states. The feature vector here is the normalized reflectivity of four wavebands (blue, green, red, and near-infrared).

[0093] Table 2: Quantitative evaluation results of different objects of the same device

[0094]

[0095] In this embodiment, Values below 0.01 are considered as system or environmental noise and can be ignored. This logic is realized in practice by setting a minimum threshold, ensuring stable judgment of the true health state. The above data strongly prove the effectiveness and accuracy of the method described in this embodiment and verify the theoretical analysis of this embodiment. Comparing the health state and the mild dust accumulation columns, although there is a slight fluctuation, the calculated value (0.001) is extremely small and at a negligible noise level. This proves that the method can accurately identify the health state and has good robustness to slight surface contamination that does not affect the core function of the device, avoiding false positives. Comparing the severe industrial dirt accumulation ( ) and the early point rust ( ), both are identified as abnormal, but the latter gets a higher abnormality score. This clearly shows that the method not only detects abnormalities, but also quantitatively distinguishes according to the severity of spectral feature changes. Rust changes the spectral distribution more significantly than dirt (most significantly in the red light band), and the algorithm accurately captures and quantifies this difference, which is crucial for determining the type and severity of the failure. The value (0.260) of the severe corrosion column is much higher than that of all other abnormal states. This proves that this embodiment can give a clear and significant quantitative indicator for severe failures that pose a direct safety threat. Combining this quantitative result with the device identification , the system can issue a severe corrosion warning for the transformer high-voltage side bushing, and the precise and operable diagnostic output of abnormality 0.260 constitutes a significant technical progress compared to the existing technology of ambiguous and non-point state evaluation.

[0096] In order to quantify the The value is converted into specific instructions for guiding operation and maintenance. In this embodiment, a combination of statistical analysis and expert experience is used to develop the division standard. First, by measuring a large number of known state (covering health, various fault types and severity) equipment samples, the statistical distribution of the value is obtained Then, combining the experience of power equipment maintenance experts, the key inflection points in the statistical distribution are associated with the actual operation and maintenance requirements (such as needing attention, suggesting inspection, and immediate treatment), and the following division intervals are finally determined: (1) Normal interval, the value range is [0, 0.05). The surface state of the equipment has no significant difference with the health benchmark, or only has negligible small changes. Only record: store the equipment identification, pose, timestamp and physical state abnormality value of the equipment into the historical database, and do not trigger any alarm or notification. (2) Attention interval, the value range is [0.05, 0.15). A moderate degree of spectral deviation is detected, which may correspond to heavy fouling, initial discoloration or early corrosion, etc. It does not constitute an emergency threat, but needs to be included in the maintenance plan. Generate a low-priority work order: automatically generate a low-priority inspection work order for the corresponding equipment identification in the operation and maintenance management system, and suggest that manual verification be performed during the next planned maintenance period. (3) Alarm interval, the value range is [0.15, +∞). A severe spectral deviation is detected, which is highly likely to correspond to serious corrosion, oil leakage, insulation damage, etc. It needs to be immediately intervened. Trigger immediate alarm: immediately push high-priority alarm information to the operation and maintenance center, and the information content includes equipment identification, accurate pose, abnormality value of physical state abnormality and real-time image, and suggests that on-site disposal be performed immediately.

[0097] As shown in Figure 3 Step S4 of this embodiment is realized through a causal correlation analysis module. The equipment state-electric field coupling disturbance model in the causal correlation analysis module is a pre-trained generative probabilistic model representing the conditional probability relationship between the physical state abnormality and the local electric field phase disturbance spectrum. This embodiment details that the causal correlation analysis module calls a pre-trained equipment state-electric field coupling disturbance model, based on a physical state abnormality associated with a specific spatiotemporal point and synchronously collected electric field phase data, calculates the spectral-phase feature mutual information entropy between the two, and thereby generates a confidence index representing that the physical state abnormality is the direct cause of the electric field disturbance; specifically, step S4 includes:

[0098] S4.1, determining whether the physical state abnormality of the substation equipment is less than the preset minimum abnormality threshold value, if yes, taking the spectral-phase feature mutual information entropy of the substation equipment as 0; jumping to step S5; otherwise, jumping to step S4.2;

[0099] S4.2, calculate the overall probability distribution of the physical state anomaly degree of the substation equipment and its marginal entropy according to the historical observation sample database of the physical state anomaly degree of the substation equipment; use the pre-trained equipment state-electric field coupling disturbance model VAE to predict the posterior probability distribution of the physical state anomaly degree of the substation equipment under the condition of the spectral reflectivity data, and calculate the conditional entropy of the physical state anomaly degree with respect to the spectral reflectivity data; subtract the conditional entropy from the marginal entropy to obtain the spectral-physical state anomaly degree feature mutual information entropy, and jump to step S5; the equipment state-electric field coupling disturbance model VAE is pre-trained to establish a mapping relationship between the input physical state anomaly degree of the substation equipment and the spectral reflectivity data and the output posterior probability distribution of the physical state anomaly degree of the substation equipment under the condition of the spectral reflectivity data.

[0100] The execution of step S4 of the embodiment depends on a series of defined key parameters. The physical state anomaly degree is a prerequisite for starting the causal correlation analysis; the electric field phase disturbance spectrum is denoted as , and its physical meaning is a normalized vector representing the energy distribution of the electric field phase fluctuation in a specific frequency domain; the specific acquisition approach and determination method of the parameter are as follows: from a series of valid synchronous data pairs associated with a specific pose output by the data synchronous acquisition module described in the embodiment, extract the time series of the multi-point induced electric field synchronous phase difference vector in a continuous time window (100 milliseconds); secondly, for each dimension of the time series, each dimension represents a corresponding electric field probe, and then a fast Fourier transform (FFT) pre-processing algorithm is applied to convert it from the time domain to the frequency domain to obtain the power of each frequency component; then, the powers of the frequency components are summed and normalized to form a probability density vector representing the energy distribution of the electric field phase disturbance in the time window, and this vector is the electric field phase disturbance spectrum . As a numerical calculation example: after FFT of the phase time series of a certain probe, the powers in three key frequency bands are [0.0, 5.2, 1.3], and the normalized disturbance spectrum of this dimension is [0.0 / (5.2+1.3), 5.2 / 6.5, 1.3 / 6.5], i.e. [0.0, 0.8, 0.2]. The conditional probability distribution is denoted as , and its logical meaning is that under the condition of a specific physical state anomaly degree , a specific electric field phase disturbance spectrum The probability of this conditional probability is calculated using a pre-trained variational autoencoder (VAE) that serves as a state-electric field coupled perturbation model. This VAE model is obtained from a simulation database containing a large number of equipment fault samples (using COMSOL-Multiphysics® to simulate electromagnetic fields of different degrees of insulator flashover, equipment corrosion, etc.) and historical measured data, constructing a database containing tens of thousands of pairs of data. The dataset of samples is used to train the VAE offline until it can respond to input. Its decoder generates an electric field phase perturbation spectrum that closely matches the real situation. Distribution. The marginal entropy of the anomaly degree of a physical state, denoted as... Its physical meaning is a measure of the uncertainty regarding the degree of anomaly in the physical state of a device when no electric field information is known. Specifically, it is obtained by: based on all historical datasets... The statistical histogram of values ​​is calculated using the standard Shannon entropy formula. The conditional entropy of the physical state anomaly degree is denoted as... Its physical meaning is that when a specific electric field phase perturbation spectrum is observed... Next, the remaining uncertainty regarding the degree of physical state anomaly of the equipment is measured. Specifically, this is obtained using the conditional probability distribution derived from the pre-trained VAE model described above. The calculation is performed using Bayes' theorem and the standard conditional entropy formula. The spectral-phase characteristic mutual information entropy, denoted as... Its physical meaning is that by observing the phase perturbation spectrum of the electric field... Eliminable anomalies related to physical state The uncertainty is a quantity that directly quantifies the statistical correlation or information gain between the two. Its specific acquisition method is from... Subtract As an example of numerical calculation: if calculated based on historical data... marginal entropy It is 1.5 bits, and when a specific value is observed... Then, the conditional entropy was calculated using the VAE model. If the value is reduced to 0.2 bits, then the mutual information entropy between the two will be... It is 1.5 - 0.2 = 1.3 bits.

[0101] The calculation process of spectral-phase feature mutual information entropy is decomposed as follows: from the final real-time pose Physical state anomaly and electric field phase perturbation spectrum Composition; the system first processes the input Perform the judgment. Set a minimum anomaly threshold. , the value example is 0.05. If , it is considered that the current device state is normal or the abnormality degree is negligible, and there is no need to perform causal analysis. The flow is directly terminated, and the mutual information entropy is output as 0. If , it is considered that there is a significant physical state abnormality, and subsequent causal correlation calculation is started; the following steps are entered: for marginal entropy calculation: based on a database containing a large number of historical observation samples, the overall probability distribution of is calculated, and the marginal entropy is calculated therefrom. For conditional entropy calculation: a pre-trained device state-electric field coupling disturbance model (VAE) is called. The current data is input into the model, the posterior probability distribution of under the condition of the current is calculated, and the conditional entropy is calculated therefrom. For spectrum-phase feature mutual information entropy solving: the obtained conditional entropy is subtracted from the obtained marginal entropy , and the final difference is obtained, which is the spectrum-phase feature mutual information entropy. The final output of the present process is a scalar value, i.e. the spectrum-phase feature mutual information entropy . The value is in bits, which quantitatively represents the causal correlation strength between the detected physical state abnormality and the electric field phase disturbance at the current observation position. The input parameters and have a nonlinear positive correlation with the final output . Specifically, when a specific physical state abnormality pattern (a distribution of ) and a specific electric field phase disturbance spectrum pattern (a distribution of ) show stronger consistency and co-occurrence in historical data, the pre-trained VAE model can more accurately infer one from the other, resulting in a significant decrease in the conditional entropy , and thus an increase in the final mutual information entropy . This relationship ensures that only when there is a clear and repeatable pattern correlation between the physical abnormality and the electric field disturbance, the system will output a high causal correlation degree.

[0102] As shown in Figure 4 , the step S5 of the present embodiment is implemented by a positioning diagnosis generation module. The predictive diagnosis result generated by the positioning diagnosis generation module in the present embodiment is a structured data object, which at least includes: the three-dimensional coordinates of the final real-time pose, the identified fault physical root source identifier, the electric field disturbance feature identifier associated therewith, and the fault evolution probability calculated based on the spectrum-phase feature mutual information entropy. Specifically, in step S5, the optimal pose of the unmanned aerial vehicle at the current time k The method further comprises: obtaining, by using a pre-trained gradient boosting decision tree model, a fault physical root cause identifier corresponding to the physical state abnormality degree and the spectral reflectance data of the substation equipment , the fault physical root cause identifier being an encoding uniquely identifying a specific fault physical cause; the gradient boosting decision tree model pre-trains a mapping relationship between input physical state abnormality degree and spectral reflectance data of the substation equipment and output fault physical root cause identifier and a category probability thereof, and the fault physical root cause identifier with the maximum category probability is taken as the final obtained fault physical root cause identifier ; obtaining, by using a pre-trained k1-nearest neighbor classifier, an electric field disturbance feature identifier corresponding to the spectral reflectance data , the electric field disturbance feature identifier being an encoding uniquely identifying an electric field disturbance mode caused by the fault, the k1-nearest neighbor classifier being used to extract features from the spectral reflectance data, calculate distances between the extracted features and basic features in a feature library of electric field disturbance modes caused by multiple faults, find out k1 nearest neighbor basic features, and vote for electric field disturbance modes corresponding to the k1 nearest neighbor basic features to obtain a final electric field disturbance feature identifier .

[0103] The optimal pose of the unmanned aerial vehicle at the current time k is determined in step S5 of the embodiment The method further comprises: inputting the spectral-phase feature mutual information entropy into a logistic S-shaped function model to obtain a fault evolution probability corresponding to the spectral-phase feature mutual information entropy , and a function expression of the logistic S-shaped function model being:

[0104] ;

[0105] wherein, is the spectral-phase feature mutual information entropy.

[0106] The structured data object obtained in step S5 of the embodiment has a format of: {coordinate (X, Y, Z), root cause ID, disturbance ID, evolution probability}. The root cause ID is the fault physical root cause identifier , and the disturbance ID is the electric field disturbance feature identifier , the coordinates (X, Y, Z) are directly derived from the output of the hybrid localization module. This embodiment elaborates the localization diagnosis generation module, which, when the spectral-phase feature mutual information entropy exceeds the preset diagnosis threshold, generates a structured data object containing the accurate three-dimensional position, the fault physical root source identification, the electric field disturbance feature identification, and the fault evolution probability by invoking a fault mode classifier and a probability mapping function. The execution of this embodiment relies on a series of precisely defined key parameters. The spectral-phase feature mutual information entropy is a dimensionless value quantifying the strength of the causal association between the equipment physical state anomaly and the local electric field disturbance. The preset diagnosis threshold is denoted as , the logical meaning of which is a preset numerical threshold for determining whether the strength of the causal association is sufficient to trigger predictive diagnosis; the specific determination method of this parameter is: based on statistical analysis of a historical database containing a large number of confirmed fault cases, calculate the probability distribution of the values corresponding to all fault cases, and set the 95th percentile of this distribution as . This approach ensures that only statistically highly significant strong association events will trigger diagnosis, thereby ensuring high detection rate while effectively controlling false positive rate. As a numerical calculation example: if the value domain distribution of the spectral-phase feature mutual information entropy in the historical fault database is extensive, and the statistical analysis shows that the 95th percentile value is 1.1 bits, then is set to 1.1. The fault physical root source identification is denoted as , the logical meaning of which is a code uniquely identifying the specific fault physical cause (including but not limited to heavy pollution on the surface of insulators, severe rust on metal components). The calculation model of this parameter is a pre-trained multi-class classifier, specifically using the gradient boosting decision tree model. The source of this model is: using the same simulation and measurement data set as the causal association analysis module, this data set has been pre-labeled by experts with the physical root source class of each sample. Using this labeled data set to train the GBDT model offline, so that it can output a probability highest fault class identification according to the combined features of the input physical state anomaly degree and the electric field phase disturbance spectrum . As a numerical calculation example: after the model receives a set of inputs, its output class probabilities are {pollution: 0.92, rust: 0.05, damage: 0.03}, then is determined as pollution. The electric field disturbance feature identification is denoted as , the logical meaning of which is a code uniquely identifying the specific pattern of the electric field disturbance caused by the fault (including but not limited to high-frequency corona discharge, power frequency harmonic distortion). The calculation model of this parameter is a pre-trained k1-Nearest-Neighbors (k1-NN) classifier. The source of this model is: extracting all the electric field phase disturbance spectra vectors from the historical data set, and clustering and labeling them by experts to form a feature library. The k1-NN model finds the nearest k1 neighbors by calculating the distance between the current input and all the vectors in the feature library, and determines the feature identification of the current disturbance by voting according to the categories of the k1 neighbors. The fault evolution probability is denoted as , the physical meaning of which is a value between 0 and 1, quantifying the possibility of the currently identified fault further deteriorating or leading to failure within a pre-set time window in the future. The pre-set time window is set to 3 months in this embodiment. The calculation model of this parameter is a calibrated Logistic-Sigmoid function. The source of this model is: collecting a large number of historical fault cases, each case containing the value at the initial diagnosis and the binary label (1 for deterioration, 0 for no deterioration) of whether it deteriorates within the pre-set time window. The center point and steepness parameters of the Logistic-Sigmoid function are fitted using these data, thereby establishing an accurate mapping relationship from the spectrum-phase feature mutual information entropy to the fault evolution probability. As a numerical calculation example: assuming the calibrated Sigmoid function is , when the measured is 1.8 bits, the calculation process is 2.5 x (1.8-1.5) = 0.75, e -0.75 ≈0.472, then . The predictive diagnosis result is denoted as , the logical meaning of which is a structured data object encapsulating all the information of this positioning diagnosis. Its acquisition approach is to integrate all the key parameters calculated by this module into a pre-defined data structure. The calculation process of the predictive diagnosis result is decomposed as follows: the starting input of this process is a set of data strongly associated with a single spatiotemporal point: spectrum-phase feature mutual information entropy , final real-time pose , physical state abnormality , and electric field phase disturbance spectrum . The system compares the input with the pre-set diagnosis threshold . If ; it is determined that the causal association strength observed at present has reached the standard for triggering diagnosis, and the process continues. If ; it is considered that the correlation strength is not enough to constitute a clear failure precursor, the process terminates, and no diagnostic result is generated. A pre-trained gradient boosting decision tree (GBDT) model is called to obtain the failure physical root identification as input. At the same time, a pre-trained k1-nearest neighbor (k1-NN) model is called to obtain the electric field disturbance feature identification as input. .

[0107] The input is substituted into the calibrated logistic S-shaped function model to calculate the corresponding failure evolution probability . Then a structured data object is created and the content is filled in the predetermined format.

[0108] The final output of the process is a structured predictive diagnostic result . The output predictive diagnostic result is passed to the system top-level application or operation and maintenance platform as a direct basis for decision support. The core calculation of this embodiment is to convert the single spectrum-phase feature mutual information entropy into a failure evolution probability with clear physical meaning and foresight, which is a kind of statistical calibration-based nonlinear mapping. The evolution of failure is usually not linear. When the correlation is low, the evolution probability grows slowly; when the correlation exceeds a certain critical point, the probability rises rapidly; when the correlation is very high, the probability tends to be saturated (close to 1). The S-shaped function perfectly simulates this critical-mutation-saturation nonlinear process. There is a monotonically increasing nonlinear positive correlation between the input parameter and the final output . This means that the higher the value of the spectrum-phase feature mutual information entropy, the higher the calculated failure evolution probability, which completely conforms to the physical logic that stronger causal correlation points to more dangerous failure precursors.

[0109] Further explanation: the core algorithm output value of this embodiment is the failure evolution probability . Due to the mathematical properties of the logistic S-shaped function, its output value domain is strictly limited to the [0, 1] interval. When tends to 0: this state means that the input is lower than the center point parameter of the S-shaped function. It means that the causal correlation strength between the currently observed physical state anomaly and the electric field disturbance is lower, and it is less necessary to constitute a statistically significant failure precursor. The system judges that the current device state even if there is an anomaly, the possibility of its evolution into a serious failure within the preset time window is lower. When tends to 1: this state means that the input The higher the value is than the center point parameter of the S-shaped function, the stronger the causal relationship between the identified physical state anomaly and the electric field disturbance at the current pose, and the greater the resulting fault symptom. The higher the probability that the system determines the fault will worsen further within a preset time window or cause equipment failure, the higher the priority of intervention measures required. This embodiment affects the final output value. The only direct key input parameter is the spectral-phase characteristic mutual information entropy. Final output With input parameters There exists a monotonically increasing non-linear positive correlation between them. A logistic sigmoid function is used for mapping to ensure... Follow The sigmoid function increases monotonically with increasing sigmoid value, which aligns with the fundamental principle that stronger correlations correspond to higher risks. The nonlinear nature of the sigmoid function makes it possible for... When the value crosses the critical region, A rapid and significant jump will occur. This accurately simulates the abrupt changes in many physical faults from stable defects to unstable deterioration, making this embodiment more sensitive to impending dangers and constituting a significant improvement over simple linear models.

[0110] In this embodiment, a preset diagnostic threshold is used. The bit depth is set to 1.1 bits, and the logic S-shaped function parameters are set as follows: center point 1.5, steepness 2.5. Taking scenarios A3: light dust accumulation on insulator strings, scenario B3: early pitting corrosion on bushing surfaces, scenario C: severe salt spray contamination on insulator strings, scenario D: severe corrosion of disconnector contacts, scenario E: slight abnormality in oil chromatography due to local overheating of the main transformer, and scenario F: continuous partial discharge caused by severe moisture in cable terminals as examples, the predictive diagnostic results generated by the method in this embodiment are demonstrated for different substation equipment fault scenarios.

[0111] Table 3: Predictive diagnostic results generated by the method in this embodiment for different scenarios

[0112]

[0113] As shown in Table 3, in scenario A3 where no diagnosis was triggered, all subsequent parameters were empty or default values, and the fault evolution probability was recorded as 0. The above data strongly demonstrates the effectiveness, accuracy, and practical value of the predictive diagnostic results generated in this embodiment. Comparing scenario A3 and scenario C, both occurred on insulator strings at the same physical location. However, due to the severe salt spray contamination in scenario C leading to stronger spectral-electric field coupling, its spectral-phase characteristic mutual information entropy was 1.60, far higher than the 0.85 of the light dust accumulation in scenario A3. Therefore, the system successfully triggered the diagnosis and provided a moderately high fault evolution probability. The value is 0.56. This proves that this embodiment can penetrate surface phenomena and accurately quantify the true risk of a fault based on the strength of causal correlation rather than a single degree of anomaly, effectively distinguishing between the presence of anomalies and the existence of a threat. Comparing scenario D (corrosion of the disconnecting switch) and scenario F (moisture discharge at the cable terminal), both were identified as high-risk events. (The values ​​are 0.75 and 0.92 respectively). However, the structured data object output in this embodiment is obtained through... and The report clearly identifies the physical root cause and specific manifestation of the fault. Maintenance personnel not only know there is a danger from the report, but also immediately understand what is wrong (switch corrosion / cable dampness) and the extent of the damage (DC offset / pulse discharge). This comprehensive diagnostic approach, compared to existing technologies that only output a general risk value, is a decisive step towards achieving accurate and efficient maintenance, constituting a significant technological advancement. The mutual information entropy of scenarios B3 (early corrosion) and E (localized overheating) is similar, at 1.25 and 1.30 respectively, but the system provides a clear, non-zero probability of fault evolution in both scenarios. The values ​​are 0.34 and 0.40, respectively. This indicates that this embodiment can not only diagnose the current state but also predict the likelihood of future deterioration of these nascent faults based on a model calibrated from historical data. This forward-looking information allows maintenance strategies to shift from passive response to proactive prevention, intervening before faults cause greater losses, which has significant economic and safety value.

[0114] To quantify the probability of fault evolution This translates into specific, executable operational instructions to guide operations and maintenance. This embodiment employs a joint decision-making logic based on a risk matrix and operations and maintenance costs to formulate the classification criteria. First, historical data is analyzed to establish... The correlation with the severity level of actual failure consequences (e.g., general, severe, catastrophic); then, considering the costs of different intervention measures (human resources, materials, power outage losses), determining at what probability level which response measure can achieve the risk-cost optimum, in this embodiment, is the failure evolution probability. The defined level intervals are as follows: (1) monitoring observation level, the value range is (0, 0.3], there is identifiable failure precursor, but the causal correlation is weak, and the possibility of deterioration in the short term is low. An observation work order is generated: an observation level work order is automatically generated for the device in the operation and maintenance system, including complete predictive diagnosis result data. It is suggested to pay special attention to the point in the next routine inspection, without the need to send personnel immediately. (2) early warning treatment level, the value range is (0.3, 0.7], there is a clear failure precursor, the causal correlation is moderate, and there is a higher possibility of deterioration in the future cycle. Trigger early warning and plan maintenance: immediately push the early warning information to the operation supervisor. Automatically mark the point as high priority in the task planning system, and suggest arranging a special survey or planned maintenance within 72 hours. (3) emergency intervention level, the value range is (0.7, 1.0], there is a strong failure precursor, the causal correlation is clear and strong, and it is extremely likely to cause device failure or safety accidents in the short term. Trigger emergency alarm and generate a rush repair work order: immediately send the highest level alarm to all relevant personnel through multiple ways such as SMS, APP push, etc. An emergency rush repair work order is automatically generated, requiring personnel to go to the scene immediately with the corresponding tools, and considering taking temporary isolation or load reduction measures.

[0115] The core linkage effect of the system constructed in this embodiment is reflected on the intelligent enhanced closed loop of positioning-diagnosis-adaptive positioning: driving the adaptive correction of the positioning map: when the system generates a high mutual information entropy diagnosis result (for example, confirming that the insulator at coordinates (X, Y, Z) has caused high-frequency corona discharge due to heavy pollution), the diagnosis result is fed back to the hybrid positioning module. The hybrid positioning module automatically updates the map it relies on accordingly: ① update the geometric-spectral map, update the spectral feature vector of the equipment at coordinates (X, Y, Z) to its current abnormal state; ② update the electric field phase map, superimpose the identified electric field disturbance features in the reference model of the area around the coordinates. The internal relationship is that the output of the diagnosis is the input of the correction and enhancement of the positioning system environment model. This makes the unmanned aerial vehicle more accurate and robust in positioning when passing through the area in future inspection because the map is more matched with the reality. Driving adaptive planning of inspection tasks: the same high-confidence diagnosis result is passed to the task planning system of the unmanned aerial vehicle. The system will autonomously decide to mark the coordinates (X, Y, Z) of the diagnosis result as a high-priority point of attention. In subsequent inspection tasks, the system will automatically increase the access frequency to the point or automatically perform more detailed reconnaissance flying actions when flying to the point. The internal relationship is that the results of positioning and diagnosis change from analysis of history to intelligent guidance of future behavior, realizing the transition from fixed inspection mode to dynamic reconnaissance mode, and improving the pertinence and efficiency of inspection. It should be noted that all the calculation formulas in this application file use regression analysis including but not limited to machine learning algorithms to deeply analyze the collected relevant parameters, identify their natural trends and mutual relationships. Professional software such as Python's Scikit-learn library or R language is used to automatically generate mathematical models that match the data. Then, the model performance is objectively evaluated through cross-validation and other methods, and combined with continuous feedback and optimization to ensure that the created formula truly reflects the internal law of the data, thereby ensuring its effectiveness and accuracy. In all the calculation formulas in this application, the parameters in each formula are processed by consistent range of dimensionless to ensure that different physical quantities are compared on the same scale; the dimensionless technology includes but is not limited to maximum and minimum normalization, Z-Score standardization. The method of this embodiment solves the two core technical problems of "robust positioning" and "deep diagnosis" in unmanned aerial vehicle inspection in substation environment through the construction of an innovative multi-element information fusion framework, not only significantly improves the positioning accuracy and robustness of unmanned aerial vehicles in strong electromagnetic interference and visual feature repetitive environment, but more importantly, it gives the system a predictive diagnostic ability based on causal tracing, so that it can discover and accurately locate equipment failure precursors in advance, greatly improving the intelligent level and safety of substation operation and maintenance.

Claims

1. A method for unmanned aerial vehicle (UAV) inspection and positioning of substations based on multi-source information fusion, characterized in that, Includes the following steps: S1, synchronously acquire multi-dimensional information of the substation equipment, including spectral reflectance data of the surface of the substation equipment and electric field phase data in the surrounding space; S2, obtains the UAV pose estimates based on spectral reflectance data and electric field phase data respectively, and fuses the two UAV pose estimates to calculate the optimal UAV pose at the current time k. ; S3, the optimal pose based on the current time k. Using spectral reflectance data to assess the degree of physical condition anomaly of substation equipment; S4. The physical state anomaly degree and spectral reflectance data of the substation equipment are used to predict the posterior probability distribution of the physical state anomaly degree of the substation equipment under the condition of obtaining the spectral reflectance data using the pre-trained equipment state-electric field coupled perturbation model VAE. Based on the posterior probability distribution of the physical state anomaly degree of the substation equipment under the condition of obtaining the spectral reflectance data, the spectral-phase feature mutual information entropy is calculated. S5, determine whether the spectral-phase feature mutual information entropy exceeds a preset diagnostic threshold. If it exceeds the preset diagnostic threshold, determine the optimal pose of the UAV at the current time k. The substation equipment is malfunctioning.

2. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 1, characterized in that, In step S1, when synchronously acquiring the spectral reflectance data of the substation equipment surface and the electric field phase data in the surrounding space, the constraint condition for synchronizing the spectral reflectance data of the substation equipment surface and the electric field phase data in the surrounding space is to satisfy the following three Boolean conditions. ~ Both are true: ； ； ； in, This is the timestamp for the acquisition of spectral reflectance data. This is the timestamp for the acquisition of electric field phase data. To determine the maximum allowed timestamp skew, Indexing in spectral reflectance data The spectral synchronization quality factor refers to the synchronization accuracy between the terminal's slave clock and the master clock when acquiring spectral reflectance data. Indexing the electric field phase data The electric field synchronization quality factor is the synchronization accuracy between the terminal's slave clock and master clock when acquiring electric field phase data.

3. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 1, characterized in that, Step S2 includes: S2.1, The spectral reflectance data is used as a landmark observation and compared with a preset geometric-spectral map. Perform feature point matching to obtain the first pose estimate. And calculate the spectral landmark reprojection error for each matched feature point. The geometric-spectral map This includes the mapping relationship between different poses of equipment in different substations and their corresponding spectral reflectance data; the electric field phase data is used as field observation and compared with a preset electric field phase map. Perform feature matching to obtain a second pose estimate And calculate the electric field measurement residual. The electric field phase map This includes the mapping relationship between different poses and their corresponding electric field phase data; and the reprojection error of the spectral landmarks at each matching point. The spectral location reliability was calculated using the Huber cost function. ; the residual of electric field measurement The electric field location confidence level was calculated using the Huber cost function. ; S2.2, Constructing the factor graph G: Adding the optimal pose from the previous time step k-1. The real-time pose of the object at time k needs to be determined. Two variable nodes are connected by an IMU factor as a factor node. The IMU factor is the pre-integrated value of the IMU sensor. If there is no matching feature point in the current period, the spectral landmark reprojection error is considered. If the confidence level is below the preset minimum confidence threshold, add a real-time pose with the current time k to each matched feature point. Connected spectral factors serve as factor nodes, and the value of this spectral factor is the spectral landmark reprojection error of the matching feature points. If there is no electric field location confidence level in this period; If the confidence level is below the preset minimum confidence threshold, then add a real-time pose with the current time k. Connected electric field factors serve as factor nodes, and the value of this electric field factor is the electric field location confidence level. Thus, a factor graph G consisting of variable nodes and factor nodes is obtained; S2.3, invoke the nonlinear optimizer to solve the constructed factor graph G, and obtain the real-time pose at the current time k that minimizes the sum of the weighted errors of all factor nodes. The real-time pose of k at the current time. As the optimal pose at the current time k Output.

4. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 3, characterized in that, Step S3 includes: S3.1, Based on the real-time pose of the UAV from a pre-set geometric-spectral map The geometric-spectral map indexes out the corresponding equipment identifiers of the substation equipment. This includes the mapping relationship between the different orientations of different substation equipment and their corresponding spectral reflectance data; S3.2, Query the corresponding equipment health status spectral database based on the equipment identifier of the substation equipment to obtain the reference spectral reflectance data of the substation equipment in a healthy state; S3.3, The statistical distance between the spectral reflectance data of the substation equipment and the reference spectral reflectance data is used as the degree of physical state anomaly of the substation equipment.

5. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 1, characterized in that, Step S4 includes: S4.1 Determine whether the physical state anomaly degree of the substation equipment is less than the preset minimum anomaly degree threshold. If it is, set the spectral-phase feature mutual information entropy of the substation equipment to 0; then proceed to step S5; otherwise, proceed to step S4.

2. S4.2, calculate the overall probability distribution of the physical state anomaly of the substation equipment and its marginal entropy based on the historical observation sample database of the physical state anomaly of the substation equipment; use the pre-trained Equipment State-Electric Field Coupled Disturbance Model (VAE) to predict the posterior probability distribution of the physical state anomaly of the substation equipment under the given spectral reflectance data, and calculate the conditional entropy of the physical state anomaly with respect to the spectral reflectance data; subtract the conditional entropy from the marginal entropy to obtain the spectral-phase feature mutual information entropy, and proceed to step S5; the pre-trained Equipment State-Electric Field Coupled Disturbance Model (VAE) establishes the mapping relationship between the input physical state anomaly of the substation equipment and the spectral reflectance data, and the output posterior probability distribution of the physical state anomaly of the substation equipment under the given spectral reflectance data.

6. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 1, characterized in that, In step S5, the optimal pose of the UAV at the current time k is determined. After a substation equipment failure is detected, the process also includes: using a pre-trained gradient boosting decision tree model to obtain the corresponding physical root cause identifier of the failure, based on the physical state anomaly degree and spectral reflectance data of the substation equipment. The physical root cause identifier of the fault This is a unique code that identifies the specific physical cause of a fault; the gradient boosting decision tree model pre-trains the input substation equipment's physical state anomaly degree and spectral reflectance data, and outputs a fault physical root cause identifier. The mapping relationship between the fault and its category probabilities is established, and the physical root cause of the fault with the highest category probability is identified. As the final physical root cause identifier of the fault The spectral reflectance data is used to obtain the corresponding electric field perturbation feature labels using a pre-trained k1-nearest neighbor classifier. The electric field disturbance feature identifier To uniquely identify fault-induced electric field disturbance patterns, the k1-nearest neighbor classifier extracts features from spectral reflectance data, calculates the distance between the extracted features and basic features in a feature library of various fault-induced electric field disturbance patterns, identifies the k1 nearest neighbor basic features, and votes on the fault-induced electric field disturbance patterns corresponding to the k1 nearest neighbor basic features to obtain the final tree-thick electric field disturbance feature identifier. .

7. The substation UAV inspection and positioning method based on multi-source information fusion according to claim 6, characterized in that, In step S5, the optimal pose of the UAV at the current time k is determined. After a substation equipment failure occurs, the process also includes: inputting the spectral-phase feature mutual information entropy into a logical sigmoid function model to obtain the corresponding failure evolution probability. The functional expression of the logical sigmoid function model is: ； in, The mutual information entropy of spectral-phase characteristics.

8. A substation unmanned aerial vehicle (UAV) inspection and positioning system based on multi-source information fusion, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the substation UAV inspection and positioning method based on multi-source information fusion as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the substation UAV inspection and positioning method based on multi-source information fusion as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, The computer program or instructions are programmed or configured to execute, via a processor, the substation UAV inspection and positioning method based on multi-source information fusion as described in any one of claims 1 to 7.

Images