Power transmission line connection fitting fault detection method based on acoustics and electromagnetic induction

Abstract

The invention discloses a power transmission line connection fitting fault detection method based on acoustics and electromagnetic induction, and belongs to the technical field of intelligent inspection and nondestructive detection of power transmission lines. A multi-mode sensor is carried by an unmanned aerial vehicle, ultrasonic wave, vibration, images and power frequency electromagnetic field data are synchronously collected, and a multi-dimensional original data set is constructed. Signal quality is improved by adopting wavelet noise reduction, beam forming and sound image fusion technologies, and flight vibration and electromagnetic interference are effectively suppressed by combining an adaptive filtering algorithm and a physical shielding structure. And acoustic, image and electromagnetic characteristics are extracted and normalized and fused, a dynamic threshold reference library is established, and intelligent grading discrimination of abnormity, defects and faults is realized through multi-stage early warning logic. A detection result automatically generates a report and is mapped to a three-dimensional line model, and operation and maintenance system linkage is supported. According to the method, synchronous identification of surface and internal defects is realized, the anti-interference capability is high, the detection accuracy is high, the inspection efficiency and safety are remarkably improved, and the method is suitable for intelligent operation and maintenance of the high-voltage transmission line.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent inspection and non-destructive testing of power transmission lines, and particularly relates to a power transmission line connecting hardware fault detection method based on acoustics and electromagnetic induction. BACKGROUND

[0002] In high-voltage and ultra-high-voltage power transmission systems, connecting hardware such as compression pipes, strain clamps, suspension clamps, and connecting bolts are key components for reliable connection and fixation of conductors. They are prone to cracks, looseness, and breakage due to factors such as mechanical fatigue, electrical corrosion, thermal cycle stress, and installation defects, as they operate in high-voltage, high-current, strong electromagnetic fields, and complex weather conditions. Once a fault occurs, it may cause conductor shedding, short-circuit tripping, or even large-scale power outage accidents, seriously threatening the safe and stable operation of the power grid.

[0003] Currently, the detection of connecting hardware for power transmission lines mainly relies on manual inspection, regular replacement, or single non-destructive testing technology. Manual inspection requires workers to climb towers for close-range observation or use handheld devices for detection, which has high operation risk, high labor intensity, and low efficiency, especially in complex terrain areas such as mountains and rivers, making it difficult to achieve comprehensive coverage. Existing automated detection methods mostly use single physical principles, such as ultrasonic detection for internal defects, infrared thermal imaging for contact failure based on temperature difference, ultraviolet imaging for capturing partial discharge phenomena, and X-ray for internal structure imaging but with radiation safety issues. These methods generally have the defects of single detection dimension, weak anti-interference ability, and strong subjective criteria, making it difficult to simultaneously identify complex faults such as surface cracks, internal damage, and connection looseness, and prone to missed detection or misjudgment.

[0004] In addition, with the popularization and application of unmanned aerial vehicle inspection technology, although the inspection efficiency has been significantly improved, there are still two major technical challenges when carrying detection equipment: first, the mechanical vibration generated during the flight of the unmanned aerial vehicle will introduce strong acoustic noise, seriously interfering with the collection of weak acoustic emission signals from the hardware; second, the power frequency electromagnetic field generated by the power transmission line itself will cause coupling interference to the sensor, affecting the measurement accuracy of electromagnetic and acoustic signals. Existing technologies lack effective dynamic noise suppression and multi-source interference shielding mechanisms, resulting in a decrease in the reliability of detection data.

[0005] More critically, current detection systems generally lack multi-modal data fusion capability and intelligent discrimination models, making it difficult to perform collaborative analysis on acoustic, electromagnetic, and image data, and unable to achieve qualitative identification of fault types and quantitative evaluation of risk levels. The early warning mechanism is lagging and relies on manual interpretation, which restricts the development of intelligent operation and maintenance.

[0006] Therefore, there is an urgent need for a new non-destructive testing method that can overcome environmental interference, achieve multi-physical field collaborative perception, and have automatic fusion analysis and hierarchical warning capability, to solve the key technical problems of single detection dimension, poor anti-interference ability, incomplete fault identification, and untimely warning in the prior art, and to improve the intelligent, precise and efficient level of state monitoring of transmission line connecting hardware. SUMMARY

[0007] To solve the above problems, the present application provides a transmission line connecting hardware fault detection method based on acoustics and electromagnetic induction, which constructs a comprehensive solution of "data acquisition-intelligent processing-accurate detection-real-time warning" through systematic technical architecture and innovative algorithm.

[0008] The technical scheme adopted by the present application is:

[0009] The transmission line connecting hardware fault detection method based on acoustics and electromagnetic induction includes the following steps:

[0010] Step 1: Multi-modal data collaborative acquisition: a multi-modal sensor system is carried by a drone to synchronously collect ultrasonic signals, vibration frequency spectrum, high-definition images and power frequency electromagnetic field data of the transmission line connecting hardware, and to construct a full-dimensional original data set to provide a data basis for subsequent intelligent analysis;

[0011] Step 2: Data preprocessing and feature extraction: based on the full-dimensional original data set, the collected multi-source signals are subjected to noise reduction processing and feature extraction, and wavelet analysis, beam forming and image recognition are used to convert the original data into quantifiable acoustic, image and electromagnetic feature parameters, realizing the conversion from data to features;

[0012] Step 3: Anti-interference processing: based on the full-dimensional original data set, combined with adaptive filtering algorithm and insulation-metal-shock-absorbing physical shielding structure, environmental noise is suppressed to ensure the accuracy and signal fidelity of feature extraction;

[0013] Step 4: Multi-modal data fusion and fault identification: based on the high-quality features after anti-interference, through multi-modal fusion model and dynamic threshold comparison, collaborative analysis of acoustic, electric and image features is realized, and the leap from single-source judgment to comprehensive diagnosis is completed;

[0014] Step 5: Multi-level warning and result output: finally, multi-level warning logic is executed based on the fusion results, visual alarm information is generated and pushed to the operation and maintenance system, forming a closed loop of detection-identification-decision-response, realizing accurate grading and efficient disposal of faults.

[0015] Further, in step 1, the multi-modal data collaborative acquisition is based on a deployment detection platform; the deployment detection platform is equipped with a detection system integrating a high-frequency ultrasonic sensor array, a multi-channel microphone array, a high-definition imaging module, and an electromagnetic induction probe, and is carried on a UAV nacelle to fly to the vicinity of a connection fitting of a power transmission line to be inspected and keep stable hovering.

[0016] Further, in step 1, the four types of original data are synchronously acquired;

[0017] In the same time window, the following multi-modal signals are synchronously acquired:

[0018] The high-frequency ultrasonic signal is obtained by emitting a 100 kHz-1 MHz pulse wave through the ultrasonic probe and receiving the internal reflection echo of the fitting, and obtaining internal structural abnormal information;

[0019] The vibration and acoustic emission signals are obtained by using the multi-channel microphone array to collect acoustic emission signals and vibration frequency spectra caused by micro-crack propagation or loosening of the fitting under running conditions;

[0020] The high-definition camera captures the surface image of the fitting, i.e., the high-definition visible light image;

[0021] The electromagnetic induction probe is used to measure the 50 Hz / 60 Hz electric field and magnetic field strength and their spatial distribution around the fitting, and to establish an electromagnetic characteristic reference, i.e., power frequency electromagnetic field data.

[0022] Further, in step 1, the process of constructing the multi-modal original data set is as follows: the ultrasonic signal, vibration frequency spectrum, high-definition image, and power frequency electromagnetic field data are aligned according to the time stamp and spatial coordinates to form a multi-modal original data set containing ultrasonic waveform, vibration frequency spectrum, visible light image, and electromagnetic field parameters.

[0023] Further, in step 2, the process of using wavelet analysis, beamforming, and image recognition is as follows:

[0024] Wavelet analysis: band-pass filtering is used on the acoustic signal to remove environmental noise; Db4 wavelet basis is selected to perform 6-layer wavelet packet decomposition on the ultrasonic signal to separate noise and effective signal; Birgé-Massart hard threshold is used on the low-frequency approximation coefficient A6 to remove non-stationary noise; improved SURE soft threshold function is used on the high-frequency detail coefficients D1-D6 to retain defect features; a cross-scale correlation correction algorithm is introduced to repair edge distortion in wavelet reconstruction and improve the signal-to-noise ratio to more than 38 dB;

[0025] Beamforming: the beamforming algorithm is used to calculate the phase difference and amplitude difference of the signals received by each microphone, construct a spatial filter, and realize directional focusing of the sound source; the line array direction finding algorithm is used to estimate the direction of arrival DOA of the sound wave to complete the direction positioning; the sound source thermodynamic map is superimposed on the high-definition image in the form of a rainbow-colored spot to generate a visual sound-image fusion image to intuitively identify the abnormal area;

[0026] Image recognition: extract parameters such as magnetic field gradient rate of change, magnetic field distortion coefficient, etc., and compare with historical normal data to identify electromagnetic field anomaly distribution; use image processing algorithm to extract crack pixel ratio, edge blur, and rust area as surface defect characteristics.

[0027] Further, in step 2, the energy distribution and peak frequency in the acoustic feature parameters, the crack ratio and spot intensity of the image feature parameters, and the distortion coefficient and gradient change of the electromagnetic feature parameters are normalized and fused to form a unified dimension fault feature vector library.

[0028] Further, in step 3, the adaptive filtering algorithm process is as follows:

[0029] The body vibration noise of the unmanned aerial vehicle during stable flight is used as the reference signal x(n);

[0030] An adaptive filter is constructed based on the least mean square error LMS algorithm to generate a reverse compensation signal Y(n) in real time, which satisfies Y(n) = -T(n) and dynamically cancels the flight vibration noise;

[0031] The filter weight iterative update formula is:

[0032] α(n+1) = α(n) + 2μe(n)x(n);

[0033] In the formula, T(n) is the compensation signal output by the adaptive filter; α(n) is the weight vector of the adaptive filter; μ is the step size parameter; e(n) is the error signal, e(n) = z(n) - s(n), z(n) = s(n) + T(n) + Y(n), z(n) is the true output signal, s(n) is the target signal from the acoustic emission or ultrasonic response of the fitting; x(n) is the reference input signal; n is the discrete time sequence number.

[0034] Further, in step 3, the insulation-metal-shock-absorbing physical shielding structure is: encapsulating the sensor module in an insulation-metal-shock-absorbing three-layer structure; the inner layer is a polytetrafluoroethylene insulation layer that blocks the line leakage current; the middle layer is an aluminum alloy shielding layer that attenuates power frequency electromagnetic interference; the outer layer is a silica gel shock-absorbing layer that reduces external mechanical vibration transmission.

[0035] Further, in step 4, a multi-layer perception MLP or attention mechanism network deep learning fusion model is constructed, which inputs acoustic spot intensity, electromagnetic distortion coefficient, and image crack features as multi-source data, and realizes adaptive allocation of feature weights through joint training to improve the accuracy of composite fault recognition;

[0036] Based on historical inspection data, a baseline feature database of the normal state of the fitting is established; the initial threshold of each parameter is set, and machine learning is adapted according to the working conditions such as voltage grade, temperature and humidity, and wind speed.

[0037] Further, in step 5, the multi-level early warning logic triggers a hierarchical early warning according to the parameter over-limit condition and the multi-modal cross-validation result:

[0038] The first level of early warning, i.e. anomaly, triggers the condition that a single parameter exceeds the threshold;

[0039] The second level of early warning, i.e. defect, triggers the condition that a single parameter is significantly over-limit, or multiple parameters are slightly over-limit and the echocardiogram appears abnormal light spots;

[0040] The third level of early warning, i.e. fault, triggers the condition that multiple parameters are continuously over-limit, and the electromagnetic distortion coefficient is greater than 15% or the light spot intensity increases suddenly.

[0041] The beneficial effects of the present application are:

[0042] The power transmission line connecting fitting fault detection method based on acoustics and electromagnetic induction proposed by the present application has the following significant beneficial effects compared with the prior art:

[0043] 1. Realize comprehensive and synchronous detection of multiple types of faults, significantly improve the detection integrity:

[0044] The present application realizes the synchronous identification of multiple fault types such as surface cracks, internal defects, bolt loosening, poor contact and corrosion aging of connecting fittings through multi-physical field collaborative collection of high-frequency ultrasonic perception, multi-channel acoustic emission, high-definition imaging and electromagnetic induction technology. It breaks through the limitations of traditional single detection means, effectively avoids missed detection and misjudgment, has high detection coverage, and comprehensively improves the comprehensiveness and reliability of fault identification.

[0045] 2. Construct a multi-modal data fusion analysis system to greatly improve the diagnosis accuracy:

[0046] Through wavelet denoising, beam forming, cross-modal feature extraction and deep learning fusion model, the acoustic, image and electromagnetic three types of heterogeneous information are quantized, normalized and intelligently weighted and fused to construct a unified dimension fault feature vector library. Compared with single parameter criterion, the fault recognition accuracy of the present application is high, and the discrimination ability for complex hidden defects is significantly enhanced.

[0047] 3. Dual composite anti-interference mechanism to ensure detection accuracy in complex environment:

[0048] In view of the two problems of flight vibration and strong electromagnetic interference of the line in the unmanned aerial vehicle inspection, the adaptive filtering algorithm + physical shielding structure double anti-interference scheme is innovatively used; the composite anti-interference mechanism ensures that high-quality detection signals can still be stably obtained in harsh environments such as high wind speed and strong electric field.

[0049] 4. Dynamic threshold and multi-level early warning system are introduced to realize intelligent hierarchical response:

[0050] The baseline feature database based on historical data is established, and combined with the working condition parameters such as voltage level, temperature and humidity, and wind speed, the threshold is self-adaptively calibrated through the machine learning model. The three-level early warning logic is designed, and the response is automatically triggered according to the parameter overrun quantity, amplitude and multi-modal cross verification result.

[0051] 5. Non-contact unmanned aerial vehicle inspection greatly improves operation efficiency and safety:

[0052] The unmanned aerial vehicle is used as a mobile carrier, and an integrated detection pod is carried to realize non-contact rapid inspection of high-altitude fittings. The detection time of a single base tower is shortened, and the inspection efficiency is improved. At the same time, the safety risks such as electric shock and falling during high-altitude operation are completely avoided, the labor cost is reduced, and the development of transmission line inspection towards intelligence and unmanned is promoted.

[0053] 6. Support closed-loop verification and system integration, have strong engineering application value:

[0054] The detection results can automatically generate electronic reports, and are mapped to the three-dimensional line model through the visualization engine to realize integrated display of "sound source positioning-image labeling-risk level". It supports docking with power PMS, OMS and other operation and maintenance management systems, and automatically pushes alarm information. After maintenance, the detection effect can be detected and verified again to form a complete closed loop of "detection-warning-disposal-verification", and has high generalizability and industrialization prospect.

[0055] In summary, through the technical collaborative innovation of multi-modal perception fusion, intelligent algorithm processing, anti-interference optimization and hierarchical early warning decision, the present application solves the key technical problems of single detection dimension, weak anti-interference ability, unclear criterion and low efficiency in existing transmission line fitting detection, significantly improves the comprehensiveness, accuracy, robustness and intelligence level of detection, provides efficient and reliable technical support for safe and stable operation of power system, and has important social benefits and broad market application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0057] Figure 1 The schematic diagram of the framework of the transmission line connecting hardware fault detection method based on acoustic and electromagnetic induction;

[0058] Figure 2 The schematic diagram of the data set construction;

[0059] Figure 3 The schematic diagram of the beam forming result generated by the array receiving signal through weighting processing;

[0060] Figure 4 The schematic diagram of the process of realizing acoustic imaging through acousto-optic image processing;

[0061] Figure 5 The schematic diagram of the process of wavelet analysis;

[0062] Figure 6 The schematic diagram of the process of adaptive filtering algorithm;

[0063] Figure 7 The schematic diagram of the insulation-metal-damping physical shielding structure;

[0064] Figure 8 The hardware integration framework diagram of the transmission line connecting hardware fault nondestructive detection method based on the fusion of acoustic and electromagnetic induction technology;

[0065] Figure 9 The schematic diagram of the whole-process closed-loop control of signal processing-feature recognition-early warning pushing;

[0066] In the figure: 1-insulation shell, 2-grid-shaped metal shielding layer, 3-internal insulation shielding layer. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, but not all the embodiments. All other embodiments obtained by those skilled in the art without creative labor based on the embodiments in the present application belong to the protection scope of the present application.

[0068] To address the shortcomings of existing detection methods, such as limited dimensionality, weak anti-interference capabilities, and difficulty in simultaneously identifying surface and internal composite faults in transmission line connection hardware, and to achieve high-precision, robust, and non-destructive intelligent detection, this embodiment provides a fault detection method for transmission line connection hardware based on acoustics and electromagnetic induction.

[0069] like Figure 1 As shown, the method for detecting faults in transmission line connection hardware includes the following steps:

[0070] Step 1: Multimodal Data Collaborative Acquisition

[0071] By using a drone equipped with a multimodal sensor system, ultrasonic signals, vibration spectra, high-definition images, and power frequency electromagnetic field data of power transmission line connection fittings are collected simultaneously to construct a full-dimensional raw dataset, providing a data foundation for subsequent intelligent analysis.

[0072] like Figure 2 As shown, the specific steps are as follows:

[0073] Step 1.1 Deploy the detection platform:

[0074] The detection system, which integrates a high-frequency ultrasonic sensor array, a multi-channel microphone array, a high-definition imaging module, and an electromagnetic induction probe, is mounted on a drone pod and flies to the vicinity of the connection hardware of the power transmission line to be inspected, at a distance d∈[0.5m,3m], and hovers stably.

[0075] The current position and attitude information are obtained using the RTK high-precision positioning module and IMU attitude sensor in the UAV flight control system: position coordinates P drone :P drone =(x d ,y d ,z d ); where x d ,y d ,z d The coordinates are x-axis, y-axis, and z-axis, respectively; attitude angle θ': Where φ is the roll angle and θ is the pitch angle. This refers to the yaw angle.

[0076] By combining the distance r measured by the laser rangefinder to the center of the fitting, the spatial position P of the fitting is calculated. fitting :

[0077] P fitting =P drone +r·R(θ')·e z ;

[0078] In the formula, R(θ') is the rotation matrix; e z e is the unit vector in the forward-looking direction. z= [0, 0, 1] T .

[0079] The position information is used for subsequent sound source positioning, image registration and data labeling.

[0080] Step 1.2 Synchronous acquisition of four types of raw data:

[0081] In the same time window, the following multi-modal signals are acquired synchronously: high-frequency ultrasonic signal: 100 kHz-1 MHz pulse wave is transmitted through the ultrasonic probe, and the internal reflection echo of the gold fitting is received to obtain internal structural abnormal information;

[0082] The ultrasonic probe transmits a modulated pulse signal with a center frequency fc∈[100 kHz, 1 MHz]:

[0083]

[0084] In the formula, s tx (t) is the transmitted ultrasonic pulse signal, representing the excitation voltage signal output by the ultrasonic probe at time t; A0 is the amplitude of the transmitted signal; f c is the center frequency of the ultrasonic signal; t is the time variable, representing the position of the signal on the time axis; τ is the pulse width, indicating the duration of the ultrasonic pulse; is a rectangular window function that defines the time range of the pulse.

[0085] When the ultrasonic wave propagates inside the gold fitting, it encounters interfaces such as cracks and pores and is reflected, and the received signal is a superposition of multiple path echoes:

[0086]

[0087] In the formula, s rx (t) is the received ultrasonic echo signal, representing the voltage signal received by the ultrasonic probe; K is the total number of reflection interfaces, representing the number of effective reflection surfaces encountered by the ultrasonic wave in the propagation path; α k is the reflection coefficient of the kth reflection surface; s tx (t-τ k ) is a time-delayed version of the transmitted signal, representing the original transmitted signal s tx (t) after a time delay of τ k ; τ k is the round-trip propagation time of the kth echo, d k is the depth, v is the speed of sound in the material; * is the convolution operator, representing the linear superposition process of the signal and the system response; h systcm (t) is the system impulse response, which includes the vibration response of the ultrasonic transducer, the influence of the coupling layer, the frequency characteristics of the preamplifier, etc.; n(t) is the environmental noise and electronic noise.

[0088] By analyzing the echo time and amplitude changes, the internal defect location and severity can be identified.

[0089] Vibration and acoustic emission signals: use M-element microphone array to collect wideband acoustic emission signals generated by the expansion of micro-cracks, bolt loosening discharge, etc. in the fittings;

[0090] The signal received by the mth microphone is:

[0091]

[0092] In the formula, x m (t) is the time-domain signal received by the mth microphone; m is the microphone array element number; a p (t) is the emission signal of the pth sound source; p is the sound source number; P is the total number of sound sources; δ() is the Dirac delta function; d is the array element spacing, indicating the physical distance between adjacent microphones; θ p is the incidence angle of the pth sound source, indicating the angle between the sound wave arrival direction and the array normal; c is the sound speed, indicating the propagation speed of sound waves in air; is the propagation time delay; * is the convolution operator; h m (t) is the transfer function of the mth channel; v m (t) is the noise of the mth channel.

[0093] High-definition visible light image: take the image of the fitting surface by a high-definition camera, which is used for subsequent visual analysis;

[0094] The imaging process can be modeled as a pinhole camera model:

[0095]

[0096] In the formula, (u, v) is the image pixel coordinate, indicating the corresponding two-dimensional pixel position of the three-dimensional point on the fitting surface in the imaging plane; K is the intrinsic matrix, containing the normalized focal length f x , f y , principal point coordinates c x , c y ; R, C are extrinsic, i.e. rotation matrix and camera center position; [I|-C] is the extrinsic transformation matrix, i.e. rigid body transformation of the world coordinate system point to the camera coordinate system; P world is the three-dimensional point in the world coordinate system, i.e. the three-dimensional coordinates of a certain point on the fitting surface in the global space.

[0097] Power frequency electromagnetic field data: use electromagnetic induction probe to measure the 50Hz / 60Hz electric field and magnetic field intensity and their spatial distribution around the fitting, establish electromagnetic characteristic reference; use three-axis electromagnetic induction probe to measure the 50Hz / 60Hz electric field E(t) and magnetic field B(t) around the fitting;

[0098] The magnetic field strength is determined by the Ampere's law:

[0099]

[0100] For the fittings near the current-carrying conductor, the magnetic field amplitude is approximately:

[0101]

[0102] The actual measured value is the time-domain signal:

[0103]

[0104] The fundamental amplitude and phase are extracted by FFT as electromagnetic feature inputs.

[0105] In the formula, is the closed path integral, which represents the integral of the magnetic field along a closed loop; B is the magnetic induction intensity, which describes the magnetic field vector at a certain point in space; dI is the path microelement vector, which is a small line segment along the closed loop, and the direction is tangential; μ0 is the vacuum permeability, which represents the response ability of vacuum to the magnetic field; I enc is the total current enclosed by the loop, which is the algebraic sum of the conductor currents passing through the area enclosed by the closed loop; I is the line current effective value, which is the working current flowing through the transmission conductor, determining the strength of the magnetic field; r is the vertical distance from the probe to the conductor, i.e., the spatial distance between the electromagnetic induction probe and the charged conductor; B(t) is the time-varying magnetic field signal, which represents the magnetic induction intensity at the fitting position varying with time; B0 is the peak value of the magnetic field amplitude, which represents the maximum strength of the sinusoidal signal, reflecting the relationship between current size and position; f power is the power frequency, i.e., the operating frequency of the power system; t is the time variable; is the initial phase, which represents the phase shift of the magnetic field signal relative to the reference time.

[0106] Step 1.3 Constructing a multi-modal raw data set:

[0107] Align the above four types of data by timestamp and spatial coordinates to form a multi-modal raw data set containing ultrasound waveforms, vibration spectra, visible light images, and electromagnetic field parameters, which serves as the basis input for subsequent processing.

[0108] To ensure the accuracy of subsequent fusion analysis, time synchronization and spatial registration must be achieved;

[0109] Time synchronization mechanism: all sensors start acquisition through a unified trigger signal and are marked with GPS time stamp t i :

[0110] T = {(t i , D i , S i} i = 1, 2,..., N} ;

[0111] In the formula, T is a time synchronization data set, indicating a data set collected synchronously by all sensors in the whole sampling process; t i is a time stamp; D i is a multi-modal data packet at the i-th moment; S i is a state label, such as the attitude of a UAV and environmental parameters.

[0112] Unified spatial coordinates: a unified spatial coordinate system is established, such as WGS-84+local northeast ENU coordinate system, and each sensor collection point is mapped to the coordinate system of the fitting body:

[0113]

[0114] In the formula, X common is the spatial point coordinates in the unified coordinate system; is a rigid body transformation matrix from the sensor coordinate system to the fitting coordinate system; X local is the original coordinates in the local coordinate system of the sensor.

[0115] Step 2: data preprocessing and feature extraction:

[0116] In data processing, the overall data processing procedure includes data noise reduction, feature extraction, multi-modal fusion and feature library construction, etc. First, the collected signal data at each level is preprocessed, and a band-pass filter technology is used to reduce noise in the environment, retaining an effective frequency band of 20 kHz-1 MHz.

[0117] For acoustic signals, a beamforming algorithm is used for preprocessing. This algorithm constructs a spatial directivity filter by accurately calculating the phase difference and amplitude difference of the signals received by each microphone, and realizes spatial filtering of the received signals. After beamforming processing, the output signals of each subarray are sent to the backend signal processing unit, and the array received signals are processed by weighting to generate the beamforming result, the principle of which is shown in Figure 3 .

[0118] The line array direction finding algorithm is used to complete the direction of arrival estimation, and the signal is sent back to the ground station after AD digital-to-analog conversion for visual display. The processed data are deeply fused with acoustic imaging technology, combined with image data collected by a high-definition imaging module, and acoustic imaging is realized through acousto-optic image processing, so that the sound source point is finally presented in the form of a rainbow spot on the ground station interface. The higher the frequency of the sound source, the more significant the visual effect of the light spot, and the basic implementation process is as follows Figure 4 .

[0119] For ultrasonic frequency signals, wavelet analysis method is used for noise reduction preprocessing, and the wavelet noise reduction technology used in the embodiment realizes high-fidelity signal enhancement through multi-level decomposition and adaptive threshold processing.

[0120] As shown in the specific implementation process as follows: Figure 5

[0121] Step 2.1 Ambient noise suppression and signal noise reduction:

[0122] In order to improve the signal-to-noise ratio and feature fidelity of acoustic signals in the fault detection of transmission line connecting hardware, the embodiment implements multi-level collaborative noise reduction processing on the collected high-frequency ultrasonic signals and acoustic emission signals, specifically:

[0123] Band-pass filter preprocessing: first, the original acoustic signal x(t) is subjected to band-pass filtering, and the effective fault feature frequency band 20kHz-1MHz is retained, and low-frequency environmental noise such as wind noise, mechanical vibration, and high-frequency electromagnetic interference is suppressed. Let the filtered signal be y(t), and its frequency domain expression is:

[0124] Y(f)=X(f)·H(f);

[0125]

[0126] In the formula, Y(f) is the frequency domain representation of the filtered signal; X(f) is the Fourier transform of the original signal; H(f) is the frequency response function of the band-pass filter; f is the frequency variable, representing the frequency component of the signal.

[0127] This step preliminarily improves the signal-to-noise ratio SNR and provides high-quality input signals for subsequent wavelet analysis.

[0128] Wavelet packet decomposition: the filtered signal is subjected to 6-layer wavelet packet decomposition using Daubechies wavelet basis Db4, and the signal is decomposed into 2 6 =32 equal-bandwidth sub-bands, realizing more fine time-frequency localization analysis.

[0129] Let the wavelet packet coefficient of the jth layer and the kth node be d j,k (n), and its decomposition process is realized recursively by the following formula:

[0130] Approximation coefficient:

[0131] Detail coefficient:

[0132] In the formula, j is the decomposition layer number, i.e. the scale index, representing the level of wavelet packet decomposition; k is the node index, i.e. the frequency band number; a j,k ​(n) is the approximation coefficient of the kth node in the jth layer, reflecting the low-frequency component of the signal at this scale, representing the overall trend or structural information of the signal;d j,k (n) is the detail coefficient of the kth node in the jth layer, reflecting the high-frequency component of the signal at this scale, containing local features such as mutations, edges, impacts, etc., such as transient acoustic emission caused by crack propagation; h(m) is the low-pass filter coefficient corresponding to the scaling function; g(m) is the high-pass filter coefficient corresponding to the wavelet function; n is the discrete time index, representing the sample point number, used for discrete signal processing; m is the summation variable, realizing the convolution operation of the filter h or g with the signal a j,k 2n is the downsampling operation, which is performed on the output after each decomposition.

[0133] Hierarchical mixed threshold processing: For different frequency band coefficient characteristics, a hierarchical mixed threshold strategy is used for nonlinear denoising:

[0134] Low-frequency approximation coefficient A6 uses hard threshold processing: The 6th layer low-frequency approximation coefficient a6(n) mainly contains signal trend and structural noise, and the hard threshold method based on Birgé-Massart criterion is used:

[0135]

[0136] Based on Birgé-Massart criterion, further set the compression ratio M j = N / j 2 for multi-layer coefficients to improve the efficiency of low-frequency denoising.

[0137] In the formula, a6(n) is the denoised estimate value of the 6th layer low-frequency approximation coefficient; a6(n) is the approximation coefficient of the 6th layer wavelet decomposition; n is the discrete time index or coefficient position; λ H is the hard threshold value, which is the critical value for judging whether to retain the coefficient. Coefficients below this value are considered noise and are set to zero; σ is the noise standard deviation estimate value, which represents the parameter of the noise intensity in the signal, estimated by the first layer detail coefficient, σ = median(|d1|) / 0.6745, d1 is the first layer detail coefficient; N is the signal length, i.e. the total number of sampling points; lnN is the natural logarithm, which reflects the trend that the threshold slowly rises with the increase of the signal length; M j is the maximum number of non-zero coefficients retained in the jth layer; j is the number of wavelet decomposition layers.

[0138] High-frequency detail coefficient D1-D6 uses improved SURE soft threshold processing: In order to retain weak defect characteristics, such as transient response caused by 0.1mm level cracks, the high-frequency detail coefficient uses the improved Stein unbiased risk estimate SURE soft threshold function:

[0139]

[0140] where d j,k (n) is the wavelet detail coefficient after threshold processing; d j,k (n) is the original wavelet detail coefficient; λ s is the SURE adaptive threshold; a is a dynamic attenuation factor, a = 0.3.

[0141] The SURE soft thresholding process maximizes the preservation of the transient impact characteristics caused by defects while suppressing noise.

[0142] Wavelet packet reconstruction and edge distortion correction: After thresholding, wavelet packet reconstruction is performed to obtain the denoised signal x(t). However, due to the thresholding operation, the phase continuity of the signal edge may be damaged, introducing cross-scale correlation correction algorithm for repair.

[0143] Define the mutual information I j,j+1 between the wavelet coefficients of adjacent scales j and j+1 as:

[0144]

[0145] where I j,j+1 is the mutual information, which measures the statistical correlation between the jth layer and the j+1th layer wavelet coefficients; d j (n) is the jth layer wavelet detail coefficient; d j+1 (n) is the j+1th layer wavelet detail coefficient; p(d j (n), d j+1 (n)) is the joint probability density function; p(d j (n)) is the edge probability density function; p(d j+1 (n)) is the edge probability density function.

[0146] According to the trend of mutual information, the weight relationship of adjacent layer coefficients is dynamically adjusted in the reconstruction process to compensate for the energy loss and edge oscillation caused by threshold truncation, ensuring that the rise time error of the defect signal is controlled within <0.1 μs.

[0147] Step 2.2 Sound source positioning and acoustic imaging:

[0148] To achieve high-precision spatial positioning and visual expression of abnormal sound sources of transmission line connecting hardware, this embodiment uses a multi-channel microphone array combined with a beamforming algorithm to complete sound source direction estimation, and generates intuitive detection result graphs through sound image fusion technology. Specifically as follows:

[0149] Microwave linear array model establishment: Let a uniform linear array ULA composed of M equally spaced microphones be arranged along the x-axis with a spacing d, satisfying λ min is the wavelength corresponding to the highest frequency to avoid spatial aliasing.

[0150] When a plane wave arrives at the array normal direction with an incident angle θ, the propagation delay of the mth microphone relative to the reference element is:

[0151]

[0152] Where T m (m, θ) is the sound propagation delay of the mth microphone relative to the reference element; m is the serial number of the microphone element; d is the element spacing, i.e. the distance between adjacent microphones; θ is the sound incident angle; sinθ is the sine value of the incident angle; (m-1)dsinθ is the sound path difference between the mth element and the reference element; c is the sound propagation speed in the medium.

[0153] Beamforming: using the delay-and-sum beamforming method, the channel signals are phase-aligned and superimposed to enhance the target direction signal and suppress other direction interference.

[0154] Let the time-domain signal received by the mth microphone be x m (t), then the beamformer output is:

[0155]

[0156] Where y(θ, t) is the output signal of the beamformer at direction θ and time t; t is the time variable; M is the total number of the microphone array; x m (t) is the time-domain signal received by the mth microphone; ω m is the weighting coefficient of the mth channel.

[0157] It is more efficient to express in the frequency domain, let X m (f) be the Fourier transform of x m (t), then the frequency-domain beam response is:

[0158] Y(θ, f) = W H X(f, θ);

[0159] Where Y(θ, f) is the frequency-domain beam response; X(f, θ) is the array receiving vector, which contains the complex number representation of the phase-aligned channel signals, embodying the spatial direction information; W is the weight vector, which controls the weight vector of the array channel gain and phase, determines the beam shape;() H is the conjugate transpose operator.

[0160] Sound source spatial spectrum calculation and DOA estimation: by scanning all possible angles θ, the output power in each direction is calculated to construct the beamforming spatial spectrum:

[0161] P BF(θ) = |Y(θ, f)| 2 ;

[0162] where P BF (θ) is the beamforming output power, representing the output energy of the array assuming the sound source comes from direction θ;

[0163] or using the more robust Capon minimum variance distortion response (MVDR) beamformer:

[0164]

[0165] where P MVDR (θ) is the MVDR spatial spectrum value, reflecting the signal energy estimation at direction θ, with higher resolution and anti-interference ability; a(θ) is the array manifold vector, describing the phase relationship between each array element when a plane wave is incident at angle θ; R is the array covariance matrix, R = E[X(f)X H (f)], reflecting the statistical correlation between each channel signal;

[0166] By finding the peak position of P(θ), the sound wave arrival direction can be estimated:

[0167]

[0168] where θ is the estimated sound wave arrival direction, i.e., the azimuth angle of the sound source relative to the array normal; argmax is the angle corresponding to the maximum value.

[0169] Multi-band fusion and two-dimensional sound source thermal map generation: Since the acoustic emission signal of the fitting fault has wide frequency characteristics, this embodiment adopts a multi-band fusion strategy to improve positioning accuracy: in key frequency bands, such as 40kHz-100kHz, the spatial spectrum P(θ, f k ) is calculated respectively;

[0170] The results for each frequency are weighted and fused:

[0171]

[0172] where P total (θ) is the total spatial spectrum after fusion, representing the sound source intensity response in the direction of angle θ, which is used for final DOA estimation; f k is the center frequency of the kth analysis frequency band; P(θ, f k ) is the beamforming spatial spectrum at frequency f k ; ω k is the weighting coefficient of the kth frequency band, which is automatically adjusted according to the signal-to-noise ratio (SNR) or the fault feature energy of the frequency band, with higher signal-to-noise ratio frequency bands being given greater weight.

[0173] The weight ωk The automatic adjustment can be based on a signal-to-noise ratio or a fault characteristic frequency band energy.

[0174] The one-dimensional angle information θ is combined with the unmanned aerial vehicle pose data, the GPS+attitude sensor, and is mapped to a three-dimensional space coordinate system to obtain the spatial position (x s ,y s ,z s ) of the sound source on the power transmission line.

[0175] A two-dimensional sound source heat map H(x, y) is further generated, and a pixel value of the heat map represents a sound energy density at the position:

[0176]

[0177] In the formula, H(x, y) is an intensity value of the sound source heat map at a pixel point (x, y); (x, y) is a pixel coordinate in an image plane; max(P) is a maximum value of P total (θ) in all directions; I0 is a normalized intensity coefficient, which controls the overall display brightness of the heat map, and facilitates fusion display with a visible light image.

[0178] Sound image fusion visualization: after spatial registration of the sound source heat map H(x, y) and the high-definition visible light image I vis (x, y), a sound image fusion image is generated by using a pseudo-color mapping+transparency fusion method:

[0179] I fusion (x, y) = a.ColorMap(H(x, y)) + (1-a).I vis (x, y);

[0180] In the formula, I fusion (x, y) is a pixel value of the fusion output image, which represents a pixel intensity or color value of the sound image fusion image at the coordinate (x, y), and the sound image fusion image is fused with the acoustic heat information and the visible light background; ColorMap() is a pseudo-color mapping function, which converts the gray heat map H(x, y) into a color image palette function; I vis (x, y) is a pixel value of the high-definition visible light image, that is, a pixel value of the original visible light image at (x, y); a is an image fusion weight coefficient, which controls a proportion of the acoustic heat map in the fusion result.

[0181] The finally generated visual sound image fusion image can intuitively identify abnormal sound points such as fixture cracks and looseness, and facilitates rapid positioning by operation and maintenance personnel.

[0182] By establishing a microphone linear array model, applying beamforming algorithm and DOA estimation technology to realize accurate positioning of sound source direction, combining multi-frequency fusion and space mapping to generate two-dimensional sound energy heat map, and finally through pseudo-color superposition and high-definition image fusion to form an intuitive "acoustic image". This method breaks through the limitations of traditional auditory detection, realizes the visualization, quantification and spatialization expression of fault sound source, and provides reliable basis for subsequent intelligent diagnosis and early warning.

[0183] Step 2.3 Electromagnetic feature extraction:

[0184] In order to identify the electromagnetic field anomalies caused by loosening, poor contact or corrosion of the connecting hardware of the power transmission line, the electromagnetic induction probe is used to collect the power frequency 50Hz / 60Hz magnetic field distribution around it, and the following two core feature parameters are extracted:

[0185] Magnetic field gradient change rate: defined as the change rate of the spatial magnetic field intensity along the hardware, used to detect local current distribution unevenness or increased contact resistance anomalies.

[0186] Suppose N magnetic field intensity values B i = B(x i ), i = 1, 2,... N are collected at equal intervals in the axial direction of the hardware, then the magnetic field gradient is:

[0187]

[0188] Define the average gradient change rate as:

[0189]

[0190] Further calculate the relative gradient change rate, that is, compared with the historical normal data:

[0191]

[0192] In the formula, G i is the magnetic field gradient at the i-th point, representing the local change rate of the magnetic field intensity in the axial direction of the hardware, used to identify the current distribution abnormal area; B i = B(x i ) is the magnetic induction intensity measured at position x i ; x i is the spatial coordinate of the i-th sampling point along the axial direction of the hardware; Δx is the spatial distance between adjacent sampling points; i is the sampling point number; is the average magnetic field gradient, representing the overall trend of the magnetic field change in the entire hardware region, which is used as a state evaluation index; N is the total number of sampling points; N-2 is the effective point number for calculating the gradient; η G is the relative gradient change rate, representing the change degree of the current state relative to the historical normal state, which is one of the fault criteria; This represents the average magnetic field gradient currently detected. This is the baseline average gradient under historical normal conditions.

[0193] When |η G When the current exceeds 10%, an abnormal current distribution is identified, which may be due to loose bolts or oxidation of the contact surface.

[0194] Magnetic field distortion coefficient: reflects the degree of degradation of the symmetry and smoothness of the spatial distribution of the magnetic field, and is used to identify local magnetic flux concentration or leakage phenomena.

[0195] Fourier harmonic analysis was used to analyze the magnetic field sequence B = [B1, B2, ..., B...]. N Perform Discrete Fourier Transform (DFT):

[0196]

[0197] Extract the component with the longest spatial period corresponding to the fundamental amplitude A1=|B(2)|, and the total energy of higher harmonics:

[0198]

[0199] The magnetic field distortion coefficient is defined as:

[0200]

[0201] In the formula, B n denoted as , where represents the magnetic field strength at the nth sampling point, indicating the amplitude of the power frequency magnetic field collected at equal intervals along the axial or circumferential direction of the fitting; N is the total number of sampling points, the number of spatial samples, which determines the frequency resolution; B(k) is the kth order Fourier coefficient, representing the amplitude and phase information of the (k-1)th spatial harmonic in the magnetic field distribution; k is the harmonic index, k=1: DC component, k=2: fundamental wave, k≥3: higher harmonics; E harmonic The total energy of higher harmonics reflects the non-uniformity and distortion of the magnetic field distribution; higher energy indicates more severe local magnetic flux concentration or leakage. B γ is the magnetic field distortion coefficient, under normal conditions. B <0.05; when γ B A value greater than 0.15 is considered a significant distortion, indicating an abnormal hardware structure or connection failure.

[0202] Step 2.4 Image Feature Recognition:

[0203] After acquiring images of the hardware surface using a high-definition imaging module, the following surface defect features are extracted using digital image processing technology:

[0204] Crack pixel ratio: After grayscale conversion, adaptive threshold segmentation, and morphological denoising, the crack region is extracted. Let the total number of pixels in the original image be P. total, the number of detected crack pixels is P crack , and the crack pixel ratio is R c :

[0205]

[0206] Further combined with length measurement calibration, it can be converted into the actual crack area.

[0207] Edge blur: used to evaluate the profile degradation caused by corrosion or wear of the fitting surface. Canny edge detection + gradient amplitude statistical method is adopted:

[0208] Calculate the image gradient amplitude graph The mean value μ g and the standard deviation σ g reflect the edge definition.

[0209] Define the edge blur index:

[0210]

[0211] In the formula, β e is the edge blur index, which is a quantitative index for measuring the profile definition of the fitting surface; σ g is the standard deviation of the image gradient amplitude, which reflects the dispersion degree of the edge intensity change in the image; ε is a smoothing constant, and the value is 10 -6 ; is the image gradient vector, which represents the gray level change rate of the image at each point; μ g is the mean value of the gradient amplitude.

[0212] Rust area: using HSV color space to segment the rust area, hue H ∈ [10°, 30°], saturation S > 0.3:

[0213]

[0214] In the formula, R r is the rust area ratio, which represents the percentage of the rust area on the total area of the fitting surface; P rust is the number of rust pixels; P total is the total number of pixels in the image, which is the total number of all pixels in the effective detection area.

[0215] Step 2.5: Construct the fault feature vector library:

[0216] In order to realize the unified analysis and intelligent discrimination of multi-modal data, the acoustic, electromagnetic and image features are normalized and fused to construct the standardized fault feature vector library.

[0217] Feature parameters: acoustic modal: ultrasonic frequency energy distribution E a , vibration spectrum peak frequency fp , the ultrasonic image spot intensity I s ; image mode: crack pixel ratio R c , the rust area ratio R r , the edge blur degree β e ; electromagnetic mode: magnetic field distortion coefficient γ B , the relative gradient change rate η G .

[0218] Feature normalization: due to the different dimensions of each feature, the minimum-maximum normalization is used to map it to the [0, 1] interval:

[0219]

[0220] In the formula,

[0221] Fault feature vector construction: the normalized features are combined into a dimension feature vector:

[0222] F = [E' a , f' p , I' s , R' c , R' r , β' e , γ' B , η' G ] T ∈R 8 ;

[0223] The vector is stored in the fault feature vector library as the "digital fingerprint" of the current fitting state, which is used for subsequent machine learning classification, threshold comparison and trend analysis.

[0224] In the formula, x is the original feature value, that is, the actual measurement value of a certain feature in the current detection sample; x min is the minimum reference value of the feature, which is the lower limit value obtained by statistical analysis of the feature in the historical normal running state; x max is the maximum reference value of the feature, which is the upper limit value of the feature in the historical normal sample; x' is the normalized feature value, which compresses the original feature to a dimensionless value in the [0, 1] interval; R is the real number set; 8 is an 8-dimensional space; T represents the vector transpose.

[0225] Step 3 Anti-interference processing:

[0226] In addition, the interference of the flight vibration frequency of the unmanned aerial vehicle on the detection result and the influence of the operation of the power transmission line itself on the nondestructive testing result of the power transmission line connecting fitting are realized by using the intelligent shielding technology based on the adaptive filtering algorithm combined with the physical shielding technology in this embodiment. The specific steps are as follows:

[0227] Step 3.1 Adaptive vibration noise cancellation:

[0228] Wherein the adaptive filter passes through real-time estimation of the noise frequency characteristics generated by the unmanned aerial vehicle flight, generates a compensation signal opposite in phase and equal in amplitude to the interference sound wave, and realizes dynamic cancellation of the unmanned aerial vehicle vibration noise. The algorithm is based on the least mean square error LMS criterion, continuously adjusts the filter coefficients to track the noise frequency changes, and accurately matches the interference spectrum characteristics, thereby effectively suppressing the reverse sound wave interference introduced by the unmanned aerial vehicle flight at the sensor receiving end, and ensuring the accuracy of acoustic signal acquisition.

[0229] The overall implementation logic of the adaptive vibration noise cancellation is as shown in Figure 6 The body vibration noise of the unmanned aerial vehicle during stable flight in the previous acquisition is taken as the reference signal x(n);

[0230] An adaptive filter is constructed based on the least mean square error LMS algorithm to generate a reverse compensation signal Y(n) in real time, satisfying Y(n) = -T(n), and realizing dynamic cancellation of the flight vibration noise;

[0231] The filter weight iterative update formula is:

[0232] α(n+1)=α(n)+2μe(n)x(n);

[0233] In the formula, T(n) is the compensation signal output by the adaptive filter; α(n) is the weight vector of the adaptive filter; μ is the step size parameter; e(n) is the error signal, e(n) = z(n) - s(n), z(n) = s(n) + T(n) + Y(n), z(n) is the true output signal, s(n) is the target signal from the acoustic emission or ultrasonic response of the fitting; x(n) is the reference input signal; n is the discrete time sequence number.

[0234] Step 3.2 Physical shielding interference suppression:

[0235] In addition, in order to reduce the influence of the electromagnetic signal of the power transmission line operation on the data acquisition accuracy, as shown in Figure 7 The embodiment also relates to a multi-layer physical shielding technology, a module integrating multiple acquisition means, which is packaged by means of multi-layer physical shielding to resist the interference of the electromagnetic signal of the power transmission line operation. The multi-layer structure of an insulating shell 1-a grid-shaped metal shielding layer 2-an internal insulating shielding layer 3 serves as the basis for module packaging, and a simple structure sample is shown in FIG. 7: it can be customized according to actual application conditions.

[0236] In summary, by combining the adaptive recording method based on the least mean square error LMS criterion with the physical shielding technology, the influence of the unmanned aerial vehicle flight vibration and the electromagnetic signal of the power transmission line operation on the data acquisition accuracy of the data acquisition module is largely solved, and the accuracy and efficiency of data acquisition are improved.

[0237] Step 4 Multimodal data fusion and fault identification:

[0238] On the basis of high-quality features after anti-interference, through multimodal fusion model and dynamic threshold comparison, the collaborative analysis of acoustic, electric and image features is realized, and the leap from single-source judgment to comprehensive diagnosis is completed. The specific steps are as follows:

[0239] Step 4.1 Cross-modal feature fusion analysis:

[0240] In order to realize efficient collaborative analysis of acoustic, electromagnetic and image three types of heterogeneous modal information, and improve the recognition accuracy of composite faults of transmission line connecting fittings such as crack + loosening + corrosion, this embodiment constructs a cross-modal feature fusion model based on deep learning. The model supports nonlinear mapping and adaptive weighting of multi-source features, realizing the leap from “single-modal judgment” to “multimodal joint diagnosis”.

[0241] Input feature vector construction: let the normalized multimodal feature vector be:

[0242] F = [F a ,F e ,F i ] T ∈ R d ;

[0243] In the formula, F a = [E' a ,f' p ,I' s ], that is, acoustic modal: ultrasonic frequency energy distribution E a , vibration spectrum peak frequency f p , echogenicity intensity I s ; F e = [R' c ,R' r ,β' e ], that is, image modal: crack pixel ratio R c , rust area ratio R r , edge blur degree β e ; F i = [γ' B ,η' G ], that is, electromagnetic modal: magnetic field distortion coefficient γ B , relative gradient change rate η G ; total dimension d = 8.

[0244] All features have been normalized to the interval [0, 1] in the preprocessing stage.

[0245] Multilayer Perceptron MLP fusion model: a fully connected feedforward neural network MLP is used for nonlinear fusion, and the structure is as follows:

[0246]

[0247] Finput represents the weight matrix of the l-th layer; is the bias vector; h (l) is the hidden layer output of the l-th layer; σ (l) is the activation function; L is the number of hidden layers, usually 2-3 layers; W (o) is the output layer weight matrix; b o is the output layer bias term; the output y ∈ R C represents the probability distribution of fault categories, C is the number of categories, such as: normal, abnormal, defect, fault.

[0248] Optimize the parameters by back propagation to minimize the cross-entropy loss function:

[0249]

[0250] L CE is the cross-entropy loss function to measure the difference between the model prediction probability distribution y c and the real label distribution t c ; C is the total number of fault categories, indicating the number of categories in the classification task; c is the category index; t c is the encoded value of the real label; y c is the predicted probability output by the model.

[0251] To further enhance the contribution weight of the key modal, a multi-modal attention mechanism is introduced to realize dynamic weighting of feature channels;

[0252] First, group the input features by modal:

[0253]

[0254] Map them to a unified hidden space through linear transformation respectively:

[0255] z m = W m F m + b m , m ∈ {a, e, i};

[0256] F m is the input feature vector of the m-th modal, which represents the normalized feature vector of a certain type of sensor; m is the modal index, indicating the physical modal to which the feature belongs; W m is the modal-specific projection weight matrix, which is a learnable parameter matrix that linearly maps the m-th type of feature from the original space to the unified hidden space, achieving dimension alignment and feature transformation between modalities; b mis the bias vector, the offset of each modality, used to adjust the reference position of feature mapping, and improve the expression ability of the model; z m is the mapped hidden representation vector, the embedding vector of the m-th modality in the unified hidden space after linear transformation, used for subsequent attention fusion or classification processing.

[0257] The importance weight of each modality is calculated:

[0258]

[0259] The weighted fusion output is:

[0260]

[0261] wherein, α m is the attention weight of the m-th modality, indicating the relative importance of the modality in the current sample for fault discrimination; W α is a learnable projection matrix used to map the feature vector z m to a new representation space; v is an attention context vector, which is a learnable global weight vector, used to measure the degree of matching between each modality feature and key information; v T tanh(W α z m ) is an attention scoring function that calculates the relevance score of the m-th modality; z fused is the fused feature vector; z m is the feature embedding vector of the m-th modality, which is the hidden layer representation of the original feature after linear transformation, used to unify the semantic space of different modalities.

[0262] The fusion vector is used as the input of the subsequent classifier to realize "data-driven" feature weight allocation.

[0263] The training strategy adopts an end-to-end joint training method, and the optimization goal is:

[0264]

[0265] wherein, is the minimization operation with respect to the parameters Θ; L is the total loss function, which measures the difference between the model prediction result and the true label; L CE is the classification cross-entropy loss, used for multi-classification tasks, which measures the difference between the predicted probability distribution and the true label; λ is the regularization coefficient, which controls the strength of the L2 regularization term to prevent model overfitting; Θ is the set of all learnable parameters of the model.

[0266] The training data comes from the historical inspection sample library, which contains labeled normal and various fault states.

[0267] The final output is the fault type probability distribution:

[0268] P(class|F)=softmax(·);

[0269] Where P(class|F) is the posterior probability distribution; F is the input normalized multi-modal feature vector; softmax(·) is the normalized exponential function.

[0270] If the maximum probability corresponds to the "fault" class and P>0.9, trigger a level 3 alarm;

[0271] If P∈[0.7,0.9], combined with threshold cross-validation to determine whether to trigger a level 2 alarm;

[0272] Through the attention weight α m Visualization, support output each modal contribution, auxiliary operation and maintenance personnel to understand the criterion.

[0273] Step 4.2 Establish a dynamic threshold benchmark library:

[0274] In order to improve the accuracy and environmental adaptability of the fault identification of the connecting hardware of the power transmission line, and avoid false alarms or missed detection of fixed thresholds under different working conditions, the embodiment proposes a dynamic threshold benchmark library construction method based on historical data and machine learning. Through statistical analysis of normal state data to establish a baseline model, and combined with the operating environment parameters to realize the self-adaptive calibration of the threshold, to ensure the scientificity and robustness of the criterion.

[0275] Baseline feature database construction: Collect a large number of historical inspection data, covering normal operation state samples of hardware under typical working conditions such as 110kV, 220kV, 500kV different voltage levels, seasons, temperature and humidity, wind speed, etc.

[0276] Let the multi-modal feature vector of the ith sample be:

[0277] F i =[f i (1) ,f i (2) ,....,f i (d) T ,i=1,2,...,N;

[0278] For each feature f (j) , calculate its statistical distribution parameters:

[0279] Where the mean is:

[0280]

[0281] The standard deviation is: ​

[0282]

[0283] In the formula, F i is the multi-modal feature vector of the i-th sample, representing the normalized feature set extracted in the i-th inspection, containing acoustic, image, electromagnetic, and other multi-source information; f i (j) is the j-th dimensional feature value of the i-th sample, such as ultrasonic energy, crack proportion, magnetic field distortion coefficient, and other specific feature parameters; i is the sample number; N is the total number of normal samples; d is the feature vector dimension, i.e., the total number of features extracted; μ j is the mean of the j-th feature, describing the average level of the feature in the normal state, serving as the center reference for dynamic threshold calculation; σ j is the standard deviation of the j-th feature, reflecting the fluctuation degree of the feature under normal working conditions.

[0284] A baseline feature database is constructed to store (μ j ,σ j ) and their corresponding operating condition labels C i =[Voltage,T,RH,v ω ], where C i is the operating condition label of the i-th sample; Voltage is the voltage level, which significantly affects the electromagnetic field intensity under different voltages, affecting feature distribution; T is the environmental temperature, which affects the acoustic properties of materials, image contrast, and sensor sensitivity; RH is the relative humidity, which may exacerbate corrosion of fittings in high humidity environments; v ω is the wind speed, which affects the stability of the unmanned aerial vehicle and the propagation of acoustic signals.

[0285] Then, set the initial threshold values of each parameter, and support machine learning adaptive calibration according to operating conditions such as voltage level, temperature and humidity, and wind speed.

[0286] Initial threshold setting: based on the assumption of normal distribution, set the initial upper and lower limit thresholds of each feature, usually using the μ±kσ criterion:

[0287]

[0288] where, is the lower limit threshold of the j-th feature; is the upper limit threshold of the j-th feature; j is the feature index number; k is the safety factor, which is set according to the false alarm tolerance: k=2: about 95.4% confidence interval, suitable for high sensitivity scenarios; k=3: about 99.7% confidence interval, suitable for high reliability requirements scenarios.

[0289] Step 5: multi-level warning and result output:

[0290] Finally, based on the fusion results, multi-level early warning logic is executed to generate visual alarm information and push it to the operation and maintenance system, forming a closed loop of detection-identification-decision-response, and realizing accurate grading and efficient disposal of faults. The specific steps are as follows:

[0291] Step 5.1 executes multi-parameter threshold judgment logic:

[0292] According to the parameter overrun and multi-modal cross-validation results, trigger the early warning:

[0293] Alert level Triggering condition Level 1 alert, i.e. anomaly Single parameter exceeds threshold. Level 2 alert, i.e. defect Single parameter significantly exceeds limit, or multiple parameters slightly exceed limit and sonographic appearance is abnormal. Level 3 alert, i.e. malfunction Multiple parameters consistently exceed limit and electromagnetic distortion coefficient > 15% or intensity of speckle increases dramatically.

[0294] Step 5.2 generates detection report and visual output:

[0295] The early warning results, abnormal position, and feature data are automatically packaged to generate an electronic inspection report. Through the ground station visualization engine, the detection results are mapped to the three-dimensional transmission line model to realize integrated display of "sound source positioning-image labeling-risk level". At the same time, it supports docking with the power operation and management system, and automatically pushes the alarm information to the relevant personnel.

[0296] In addition, the transmission line connecting hardware fault detection method based on acoustics and electromagnetic induction has the advantages of Figure 8 As shown in the figure, the unmanned aerial vehicle is used as the mobile carrier of the multi-modal data acquisition module to build a non-contact inspection hardware system: through the bidirectional communication link between the unmanned aerial vehicle flight control system and the ground control station, remote control of the detection module and data back transmission are realized; the high-frequency ultrasonic sensor array, multi-channel microphone array, electromagnetic induction probe, and high-definition imaging module are integrated in the unmanned aerial vehicle nacelle in a modular ring layout, and the electromagnetic induction probe and the imaging module are coaxially installed to ensure that the detection field overlap rate is ≥80%; the hardware is packaged by a three-layer composite structure of "insulating shell-grid metal shielding layer-internal insulating shielding layer", the inner layer of polytetrafluoroethylene insulating layer isolates the leakage interference, the middle layer of aluminum alloy shielding layer realizes the electromagnetic shielding effectiveness, and the outer layer of silicone shock absorbing layer greatly improves the vibration attenuation rate.

[0297] On the basis of hardware integration, the algorithm is integrated and embedded into the hardware, and the hardware is used as the carrier to realize the overall function. In the algorithm integration dimension, a "data acquisition-intelligent processing-early warning decision" full-link algorithm system is built: as Figure 9As shown, by real-time data stream engine synchronous access high-frequency ultrasound, vibration spectrum, electromagnetic induction and other multi-modal signals, parallel computing architecture completes beamforming, wavelet analysis and other feature extraction; based on deep learning framework, cross-modal feature fusion model is constructed, acoustic imaging spot intensity, electromagnetic distortion coefficient and image crack features are input into multilayer perceptron for joint training, and multi-parameter threshold matrix is dynamically updated; the early warning decision module adopts hierarchical state machine design, generates abnormal mark, defect positioning and fault alarm three-level response instructions according to parameter overrun amplitude and historical trend analysis, and finally maps the detection results to the three-dimensional line model through the visualization engine, realizes the whole process software closed-loop control of "signal processing-feature recognition-early warning push".

[0298] In summary, the transmission line connecting fitting fault detection method based on acoustic and electromagnetic induction proposed in the embodiment realizes the comprehensive breakthrough of detection technology through the collaborative integration of hardware and software. At the hardware level, relying on the unmanned aerial vehicle carrying multi-modal sensor array, combined with the composite shielding structure, the environmental interference is effectively resisted; at the software level, the whole-link intelligent algorithm system is constructed, realizing the deep processing and dynamic early warning of multi-source data. The method breaks through the limitation of traditional single detection technology, through multi-dimensional data acquisition, intelligent algorithm processing and multi-level threshold early warning, significantly improves the accuracy and reliability of the fault detection of fitting surface crack, internal defect and connection loosening, and provides an efficient and accurate technical solution for the intelligent operation and maintenance of the transmission line.

[0299] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for detecting faults in connection hardware of a power transmission line based on acoustics and electromagnetic induction, characterized in that: The power transmission line connecting hardware fault detection method comprises the following steps: Step 1: Multi-modal data collaborative acquisition: through the multi-modal sensor system carried by the unmanned aerial vehicle, the ultrasonic signal, vibration spectrum, high-definition image and power frequency electromagnetic field data of the power transmission line connecting hardware are synchronously collected, and a full-dimensional original data set is constructed, thereby providing a data basis for subsequent intelligent analysis; Step 2: Data preprocessing and feature extraction: based on the full-dimensional original data set, the multi-source signals collected are subjected to noise reduction processing and feature extraction, the original data are converted into quantifiable acoustic, image and electromagnetic characteristic parameters by using wavelet analysis, beam forming and image recognition, and the conversion from data to features is realized; Step 3: Anti-interference processing: based on the full-dimensional original data set, the environmental noise is inhibited by combining the adaptive filtering algorithm with the insulation-metal-damping physical shielding structure, and the accuracy and signal fidelity of feature extraction are ensured; Step 4: Multi-modal data fusion and fault recognition: based on the high-quality features after anti-interference, the multi-modal fusion model and dynamic threshold comparison are used to realize the collaborative analysis of acoustic, electric and image features, and the leap from single-source judgment to comprehensive diagnosis is completed; Step 5: Multi-level early warning and result output: finally, multi-level early warning logic is executed based on the fusion results, visual alarm information is generated and pushed to the operation and maintenance system, a closed loop of detection-recognition-decision-response is formed, and the accurate grading and efficient disposal of faults are realized.

2. The power transmission line fitting failure detection method based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 1, the multi-modal data collaborative acquisition is based on the deployment of a detection platform; the detection platform, which integrates a high-frequency ultrasonic sensor array, a multi-channel microphone array, a high-definition imaging module and an electromagnetic induction probe, is carried on the unmanned aerial vehicle nacelle, flies to the vicinity of the connecting hardware of the power transmission line to be detected, and keeps stable hovering.

3. The power transmission line fitting failure detection method based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 1, four types of original data are synchronously collected; In the same time window, the following multi-modal signals are synchronously acquired: 100kHz-1MHz pulse waves are emitted by the ultrasonic probe to receive the internal reflection echo of the hardware, thereby obtaining internal structure abnormal information, i.e. high-frequency ultrasonic signals; The multi-channel microphone array is used to collect acoustic emission signals and vibration spectrum caused by micro-crack expansion or loosening of the hardware in the running state, i.e. vibration and acoustic emission signals; The high-definition camera is used to take the surface image of the hardware, i.e. high-definition visible light image; The electromagnetic induction probe is used to measure the 50Hz / 60Hz electric field and magnetic field strength and their spatial distribution around the hardware, thereby establishing electromagnetic characteristic benchmarks, i.e. power frequency electromagnetic field data.

4. The power transmission line fitting failure detection method based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 1, the process of constructing the multi-modal original data set is as follows: the ultrasonic signal, vibration spectrum, high-definition image and power frequency electromagnetic field data are aligned according to the time stamp and spatial coordinates, thereby forming a multi-modal original data set containing ultrasonic waveform, vibration spectrum, visible light image and electromagnetic field parameters.

5. The method for detecting a failure of a power line fitting based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 2, the process of using wavelet analysis, beam forming and image recognition is as follows: Wavelet analysis: Band-pass filter is used to remove environmental noise from acoustic signals; Db4 wavelet basis is selected to perform 6-layer wavelet packet decomposition on ultrasonic signals to separate noise and effective signals; Birgé-Massart hard threshold is used for low-frequency approximation coefficients A6 to remove non-stationary noise; improved SURE soft threshold function is used for high-frequency detail coefficients D1-D6 to retain defect features; cross-scale correlation correction algorithm is introduced to repair edge distortion in wavelet reconstruction and improve signal-to-noise ratio to more than 38dB; Beamforming: Beamforming algorithm is used to calculate the phase difference and amplitude difference of signals received by each microphone, construct a spatial filter, and realize sound source direction focusing; line array direction finding algorithm is used to estimate the direction of arrival (DOA) of sound waves to complete the direction positioning; the sound source heat map is superimposed on the high-definition image in the form of rainbow color spots to generate a visual sound image fusion map to intuitively identify abnormal areas; Image recognition: Extract parameters such as magnetic field gradient change rate and magnetic field distortion coefficient, compare with historical normal data, and identify electromagnetic field abnormal distribution; use image processing algorithm to extract crack pixel ratio, edge blur, and rust area as surface defect features.

6. The method for detecting a failure of a power line fitting based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 2, the energy distribution and peak frequency in the acoustic feature parameters, the crack ratio and spot intensity in the image feature parameters, and the distortion coefficient and gradient change in the electromagnetic feature parameters are normalized and fused to form a unified dimension fault feature vector library.

7. The method for detecting a failure of a power line fitting based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 3, the adaptive filtering algorithm process is as follows: The body vibration noise of the unmanned aerial vehicle during stable flight is used as the reference signal x(n); An adaptive filter is constructed based on the least mean square error (LMS) algorithm to generate a reverse compensation signal Y(n) in real time, which satisfies Y(n) = -T(n) and dynamically cancels the flight vibration noise; The filter weight iterative update formula is: α(n+1) = α(n) + 2μe(n)x(n); where T(n) is the compensation signal output by the adaptive filter; α(n) is the weight vector of the adaptive filter; μ is the step size parameter; e(n) is the error signal, e(n) = z(n) - s(n), z(n) = s(n) + T(n) + Y(n), z(n) is the true output signal, s(n) is the target signal from the acoustic emission or ultrasonic response of the fitting; x(n) is the reference input signal; n is the discrete time sequence number.

8. The method for detecting a failure of a power line fitting based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 3, the insulation-metal-damping physical shielding structure is: the sensor module is packaged in an insulation-metal-damping three-layer structure; the inner layer is a polytetrafluoroethylene insulation layer that blocks the leakage current; the middle layer is an aluminum alloy shielding layer that attenuates power frequency electromagnetic interference; the outer layer is a silicone damping layer that reduces external mechanical vibration transmission.

9. The power transmission line fitting failure detection method based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 4, a multi-layer perception (MLP) or attention mechanism network deep learning fusion model is constructed, which inputs acoustic spot intensity, electromagnetic distortion coefficient, and image crack features, and realizes adaptive feature weight allocation through joint training to improve the accuracy of composite fault recognition; A baseline feature database is established based on historical inspection data under normal state of the fitting. Set the initial threshold of each parameter, and perform machine learning adaptive calibration according to voltage level, temperature and humidity, and wind speed.

10. The method for detecting a failure of a power line fitting based on acoustics and electromagnetic induction according to claim 1, characterized by: In step 5, the multi-level early warning logic is triggered by the parameter over-limit condition and the multi-modal cross-validation result: Level 1 warning, i.e. anomaly, triggered condition: single parameter exceeds threshold; Level 2 warning, i.e. defect, triggered condition: single parameter significantly exceeds threshold, or multiple parameters slightly exceed threshold and echocardiogram shows abnormal light spot; Level 3 warning, i.e. failure, triggered condition: multiple parameters continuously exceed threshold, and electromagnetic distortion coefficient > 15% or light spot intensity increases sharply.

Images