Piezoelectric ultrasonic transducer-based electric power system insulation part internal defect ultrasonic detection method

Abstract

The invention provides an ultrasonic detection method for internal defects of an insulating part of a power system based on a piezoelectric ultrasonic transducer, and belongs to the technical field of insulating part detection. After wavelet denoising and deep attenuation compensation are carried out on echo signals, sound path distance time delay and sound pressure amplitude parameters are extracted, defect identification and position estimation are carried out by utilizing a layered variational auto-encoder, a high-frequency transducer is switched to aiming at tiny defects, a harmonic component is extracted by adopting a nonlinear ultrasonic method, and a characteristic size is calculated; a defect distribution space feature matrix is generated based on a synthetic aperture focusing technology, a unified or personalized processing strategy is selected according to an acoustic characteristic homogenization degree evaluation value, a complete defect detection result is output, and the technical problem that deep weak defects in the electric insulation part are difficult to accurately detect and recognize is solved.

Description

Technical Field

[0001] This invention belongs to the field of insulation component testing technology, and more specifically, relates to an ultrasonic testing method for internal defects in power system insulation components based on a piezoelectric ultrasonic transducer. Background Technology

[0002] Internal defect detection in power system insulation components is a crucial step in ensuring the safe operation of the power grid. Traditional ultrasonic testing technology uses piezoelectric transducers to emit ultrasonic beams into the insulation components and receives the reflected echoes to identify internal defects. However, in actual testing, strong reflected waves generated at the interfaces between insulation layers can mask weak defect echo signals. Energy attenuation of ultrasonic waves propagating deeper results in a significant reduction in the amplitude of echoes from deep defects. Furthermore, scattering interference caused by differences in the acoustic properties of the materials themselves renders traditional methods inadequate for resolving minute defects smaller than a quarter wavelength, and their sensitivity is low for detecting deep defects exceeding 50 millimeters in depth. In other words, existing technologies suffer from the technical problem of accurately detecting and identifying weak, deep defects within power insulation components. Summary of the Invention

[0003] In view of this, the present invention provides an ultrasonic detection method for internal defects of power system insulation components based on piezoelectric ultrasonic transducers, which can solve the technical problem that it is difficult to accurately detect and identify deep and weak defects inside power insulation components in the prior art.

[0004] This invention is implemented as follows: This invention provides an ultrasonic detection method for internal defects in power system insulation components based on piezoelectric ultrasonic transducers, comprising the following steps: arranging a broadband piezoelectric ultrasonic transducer array on the surface of the power insulation component; adjusting the excitation parameters of each array element according to the interface masking suppression vector; transmitting coded ultrasonic beams using pulse compression technology and receiving echo signals; extracting path delay parameters and sound pressure amplitude parameters from the received echo signals after wavelet denoising; combining a deep attenuation compensation matrix to perform amplitude compensation and time-domain averaging enhancement on weak defect echo signals; calling a defect identification model to extract features from the compensated echo signals, outputting the probability distribution of defect types and preliminary estimation results of defect spatial location; when… When the defect size is less than a quarter wavelength, the micro-defect resolution enhancement process is initiated. Following this process, the system switches to a high-frequency ultrasonic transducer array, extracting the second and third harmonic components using a nonlinear ultrasonic detection method. The defect feature size is calculated using a defect scale mapping function. Based on synthetic aperture focusing technology, the A-scan signals acquired from multiple locations are time-delayed and superimposed. A path optimization algorithm is used to calculate the precise time delay from each array element to the imaging grid point, generating a defect distribution spatial feature matrix. Based on the homogeneity evaluation value of the acoustic characteristics of the defect region in the defect distribution spatial feature matrix, a unified filtering parameter or personalized beam deflection angle and focusing depth parameters are selected, and the image is re-imagined. The defect detection results are then output.

[0005] The wideband piezoelectric ultrasonic transducer array consists of multiple piezoelectric ceramic wafers arranged linearly, with an operating frequency range of 2MHz to 15MHz, and is in contact with the surface of the electrical insulation component through a coupling agent layer.

[0006] The interface masking suppression vector is used to reduce the masking effect of strong reflection waves from the material interface on the defect signal, and the i-th element represents the delay ratio coefficient of the i-th array element relative to the reference excitation time.

[0007] The excitation parameters include the excitation voltage amplitude, excitation pulse width, and excitation delay, which are determined based on the delay ratio coefficient of the corresponding array element in the interface masking suppression vector and the reference excitation time.

[0008] The pulse compression technology improves the signal-to-noise ratio by transmitting a frequency-modulated coded pulse signal and performing matched filtering at the receiving end. The frequency of the frequency-modulated coded pulse signal is linearly swept from the starting frequency to the ending frequency within the transmission duration.

[0009] The wavelet denoising process employs a wavelet multi-scale decomposition method to decompose the echo signal into wavelet coefficients of different frequency bands. A soft thresholding method is used to suppress noise in the high-frequency wavelet coefficients, and the denoised echo signal is obtained through wavelet reconstruction.

[0010] The deep attenuation compensation matrix is ​​used to compensate for the energy attenuation of ultrasonic waves propagating in insulating materials. The element in the p-th row and q-th column represents the normalized amplitude compensation coefficient required for the p-th depth layer in the q-th frequency band.

[0011] The defect identification model's input layer receives the preprocessed echo signal time-frequency feature matrix, the encoder part adopts a four-layer hierarchical variational autoencoder structure, and the decoder part adopts a symmetrical four-layer structure to reconstruct the latent representation of the previous layer layer by layer until the original input is reconstructed.

[0012] The defect type probability distribution refers to the probability values ​​of the six types of defects output by the defect identification model. The six types of defects include crack defects, bubble defects, inclusion defects, delamination defects, discharge ablation defects, and aging and deterioration defects.

[0013] The micro-defect resolution enhancement process is a process that uses a high-frequency ultrasonic transducer array and nonlinear ultrasonic detection method to further detect defects with a defect size less than a quarter wavelength.

[0014] The nonlinear ultrasonic testing method enhances the detection sensitivity of minute defects by exciting ultrasonic waves to generate nonlinear effects at the defect and extracting high-order harmonic components.

[0015] The defect scale mapping function is used to estimate the defect feature size based on the harmonic parameters extracted by the nonlinear ultrasonic testing method. The input includes the normalized value of the fundamental amplitude, the normalized value of the second harmonic amplitude, and the normalized value of the third harmonic amplitude.

[0016] The imaging reconstruction algorithm of the synthetic aperture focusing technology establishes a three-dimensional imaging grid in the detection area and discretizes the space into voxel units. For each voxel, the acoustic path distance from its position to all array elements is calculated and converted into propagation delay.

[0017] The acoustic characteristic homogeneity evaluation value is used to determine the consistency of acoustic characteristics in the defect area. When the acoustic characteristic homogeneity evaluation value ∈ [0.82, 1], a unified filtering parameter is used for batch processing.

[0018] When the acoustic property homogenization evaluation value is ∈ [0, 0.82), personalized beam deflection angle and focusing depth parameters are applied to each defect region according to the coupling feedback correction vector and the image is re-imaged.

[0019] The defect detection results include defect type, defect spatial location coordinates, defect feature size, and defect distribution spatial feature matrix.

[0020] This invention achieves phase-controlled suppression of strong reflections at the interface by establishing an interface masking suppression vector and adjusting the excitation parameters of each array element. It employs a deep attenuation compensation matrix to perform targeted amplitude compensation on echo signals of different depths and frequencies to restore weak signal strength. It utilizes a hierarchical variational autoencoder to extract defect features in a multi-scale latent space, enhancing the ability to identify defect patterns under complex interference. For minor defects, it switches to a high-frequency transducer array and uses a nonlinear ultrasonic method to extract second and third harmonic components, breaking through the resolution limit of linear detection. It combines synthetic aperture focusing technology and acoustic characteristic homogenization evaluation to achieve adaptive imaging parameter optimization for different types of defects, thereby solving the technical problem of accurately detecting and identifying deep, weak defects inside power insulation components. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 The image shows the time-domain waveform and spectrum analysis of the ultrasonic echo signal in the embodiment.

[0023] Figure 3 This is a three-dimensional distribution diagram of the deep attenuation compensation matrix in the embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of an ultrasonic detection method for internal defects in power system insulation components based on a piezoelectric ultrasonic transducer, provided by the present invention. This method includes the following steps:

[0026] S01. Arrange a wideband piezoelectric ultrasonic transducer array on the surface of an electrical insulation component, adjust the excitation parameters of each array element according to the interface masking suppression vector, and use pulse compression technology to transmit coded ultrasonic beams and receive echo signals.

[0027] S02. After wavelet denoising processing of the received echo signal, the path delay parameter and sound pressure amplitude parameter are extracted. The weak defect echo signal is then enhanced by amplitude compensation and time-domain averaging in combination with the deep attenuation compensation matrix.

[0028] S03. Call the defect identification model to extract features from the compensated echo signal, and output the probability distribution of defect type and preliminary estimation results of defect spatial location. When the defect size is less than a quarter wavelength, start the micro-defect resolution enhancement process.

[0029] S04. According to the micro-defect resolution enhancement process, switch to high-frequency ultrasonic transducer array, extract the second harmonic component and the third harmonic component by combining nonlinear ultrasonic detection method, and calculate the defect feature size by defect scale mapping function.

[0030] S05. Based on synthetic aperture focusing technology, the A-scan signals acquired from multiple locations are processed by time delay superposition, and the sound path optimization algorithm is used to calculate the precise time delay from each array element to the imaging grid point, thereby generating a spatial feature matrix of defect distribution.

[0031] S06. Based on the acoustic characteristic homogenization evaluation value of the defect region in the defect distribution spatial feature matrix, when the acoustic characteristic homogenization evaluation value ∈ [0.82, 1], a unified filtering parameter is used for batch processing; when the acoustic characteristic homogenization evaluation value ∈ [0, 0.82), personalized beam deflection angle and focusing depth parameters are applied to each defect region according to the coupling feedback correction vector and re-imaging is performed.

[0032] S07. Output the defect detection results, which include the defect type, defect spatial location coordinates, defect feature size, and defect distribution spatial feature matrix.

[0033] The broadband piezoelectric ultrasonic transducer array is composed of multiple piezoelectric ceramic wafers arranged linearly, with an operating frequency range of 2MHz to 15MHz. The size of a single array element is 0.5mm × 10mm, the element spacing is 0.6mm, and the total array length is 60mm. The broadband piezoelectric ultrasonic transducer array is in contact with the surface of the electrical insulation component through a coupling agent layer.

[0034] The interface masking suppression vector is used to reduce the masking effect of strong reflected waves from material interfaces on defect signals. The establishment steps include: collecting measured acoustic impedance data of different material interfaces to establish an acoustic impedance difference database; calculating the reflection coefficient and transmission coefficient of each interface; prioritizing interfaces based on their reflection coefficients; calculating the required phased-array deflection angle and dynamic focusing depth for high-priority interfaces; normalizing the phased-array deflection angle to the 0-1 range as the angle weight; using the ratio of the dynamic focusing depth to the maximum detection depth as the depth weight; combining the angle weight and the depth weight to construct the excitation delay correction factor for each array element; arranging all the excitation delay correction factors of all array elements according to their array sequence number to form the interface masking suppression vector; and specifying the first element in the interface masking suppression vector as the first element. The element represents the first element. The delay ratio coefficient of each array element relative to the reference excitation time.

[0035] The excitation parameters include excitation voltage amplitude, excitation pulse width, and excitation delay. The excitation voltage amplitude is determined based on the delay ratio coefficient of the corresponding array element in the interface masking suppression vector. The excitation pulse width is determined based on the encoding sequence of the encoded ultrasonic beam. The excitation delay is determined based on the product of the delay ratio coefficient of the corresponding array element in the interface masking suppression vector and the reference excitation time.

[0036] The pulse compression technology improves the signal-to-noise ratio by transmitting a frequency-modulated coded pulse signal and performing matched filtering at the receiving end. The frequency of the frequency-modulated coded pulse signal is linearly swept from the starting frequency to the ending frequency within the transmission duration. The starting frequency is 2MHz, the ending frequency is 15MHz, and the transmission duration is 10μs.

[0037] The echo signal is the ultrasonic wave reflected signal received by each element in the broadband piezoelectric ultrasonic transducer array, and the echo signal includes interface echo from the material interface and defect echo from the defect.

[0038] The wavelet denoising process uses a wavelet multi-scale decomposition method to decompose the echo signal into wavelet coefficients of different frequency bands. A soft thresholding method is used to suppress noise in the high-frequency wavelet coefficients, while the low-frequency wavelet coefficients remain unchanged. The denoised echo signal is obtained through wavelet reconstruction.

[0039] The acoustic path delay parameter is the time difference between the arrival time of the defect echo in the echo signal and the emission time of the ultrasonic beam, and the acoustic pressure amplitude parameter is the peak voltage amplitude of the defect echo in the echo signal.

[0040] The deep attenuation compensation matrix is ​​used to compensate for the energy attenuation of ultrasonic waves propagating in insulating materials. The establishment steps include: measuring the attenuation coefficient of ultrasonic waves at different frequencies in the target insulating material, establishing a fitting curve between the attenuation coefficient and frequency, and dividing the detection depth range into equal parts along the thickness direction. Each depth layer divides the frequency range into equal operating frequency bands. For each frequency band, the theoretical attenuation is calculated for the combination of each depth layer and frequency band. The reciprocal of the theoretical attenuation is normalized to the reciprocal of the attenuation at the reference depth to obtain a compensation factor. This factor is then constructed with depth layers as rows and frequency bands as columns. OK The deep attenuation compensation matrix of column, wherein the position in the deep attenuation compensation matrix is ​​the first... Line 1 The elements of the column represent the first... Depth layer at the 1st The normalized amplitude compensation coefficient required for the frequency band is determined based on the arrival time of the echo signal to determine the corresponding depth layer. After spectral decomposition of the echo signal, the corresponding compensation factor is multiplied for each frequency band.

[0041] The weak defect echo signal is a defect echo signal whose amplitude is less than 3 times the noise level after being compensated by a deep attenuation compensation matrix.

[0042] The amplitude compensation is to multiply the amplitude of the echo signal in each frequency band by the compensation factor of the corresponding depth layer and frequency band in the deep attenuation compensation matrix. The time-domain averaging enhancement is to superimpose and average multiple echo signals at the same spatial location to improve the signal-to-noise ratio.

[0043] The specific structure of the defect identification model is as follows: the input layer receives the preprocessed echo signal time-frequency feature matrix; the encoder part adopts a four-layer hierarchical variational autoencoder structure; the first encoder maps the input features to a 256-dimensional latent space and outputs the mean vector and variance vector; the second encoder maps the first layer latent variables as conditional inputs to a 128-dimensional latent space; the third encoder maps to a 64-dimensional latent space; and the fourth encoder maps to a 32-dimensional core latent space. Each encoder layer uses a convolutional neural network to extract local time-frequency patterns. The decoder part adopts a symmetrical four-layer structure to reconstruct the latent representation of the previous layer layer by layer until the original input is reconstructed. Each decoder layer achieves the reconstruction of the previous layer's latent representation through transposed convolution. Sampling is performed, and a defect classification head branches out from the second latent space. The defect classification head contains a two-layer fully connected network that outputs the probability distribution of six types of defects. A position regression head branches out from the third latent space and outputs the three-dimensional spatial coordinates of the defects. The number of heads in the multi-head attention mechanism in the defect recognition model is adaptively adjusted according to the signal-to-noise ratio (SNR) of the echo signal. When the SNR is >20dB, an 8-head attention mechanism is used; when the SNR is ∈ [10, 20]dB, a 4-head attention mechanism is used; and when the SNR is <10dB, a 2-head attention mechanism is used. The multi-head attention mechanism achieves selective focusing on important time-frequency features by calculating the dot product similarity between the query vector and the key vector and then normalizing and weighting the vector.

[0044] The echo signal time-frequency feature matrix is ​​generated by performing a short-time Fourier transform on the echo signal. The rows of the echo signal time-frequency feature matrix represent the frequency dimension, the columns represent the time dimension, and the matrix element values ​​represent the energy intensity of the corresponding time-frequency point.

[0045] The probability distribution of the defect type is the probability value of each of the six types of defects output by the defect identification model. The six types of defects include crack defects, bubble defects, inclusion defects, delamination defects, discharge ablation defects and aging and deterioration defects. The preliminary estimation result of the spatial location of the defect is the preliminary estimate of the three-dimensional coordinates of the defect inside the power insulation component output by the defect identification model.

[0046] The crack defect is a crack-like discontinuous structure inside the power insulation component caused by mechanical or thermal stress. The bubble defect is a gas cavity left over from the manufacturing process of the power insulation component. The inclusion defect is a heterogeneous material particle mixed into the power insulation component. The delamination defect is interlayer separation caused by poor interlayer bonding of composite insulation material. The discharge ablation defect is carbonization and ablation damage to the material caused by partial discharge. The aging and deterioration defect is a microscopic damage area formed by the performance degradation of the insulation material after long-term operation.

[0047] The defect size is the maximum linear scale of the defect estimated by the defect identification model based on the defect type corresponding to the maximum probability value in the defect type probability distribution, and the quarter wavelength is one-quarter of the wavelength corresponding to the center frequency of the high-frequency ultrasonic transducer array.

[0048] The micro-defect resolution enhancement process is a process that uses a high-frequency ultrasonic transducer array and nonlinear ultrasonic detection method to further detect defects with a defect size less than a quarter wavelength.

[0049] The high-frequency ultrasonic transducer array has a center frequency of 50MHz, a bandwidth of 20MHz, 64 array elements, and an element size of 0.2mm×5mm.

[0050] The nonlinear ultrasonic testing method enhances the detection sensitivity of minute defects by exciting ultrasonic waves to generate nonlinear effects at the defect and extracting higher harmonic components. The second harmonic component is a harmonic component with a frequency equal to twice the fundamental frequency, and the third harmonic component is a harmonic component with a frequency equal to three times the fundamental frequency.

[0051] The defect scale mapping function is used to estimate the defect feature size based on the harmonic parameters extracted by the nonlinear ultrasonic testing method. The inputs include the normalized values ​​of the fundamental amplitude, the second harmonic amplitude, and the third harmonic amplitude. The output is an estimated value of the defect feature size. The defect scale mapping function is described as follows: dividing the fundamental amplitude by the fundamental amplitude when there is no defect yields the normalized fundamental amplitude value; dividing the second harmonic amplitude by the normalized value of the square of the fundamental amplitude yields the second harmonic relative coefficient; dividing the third harmonic amplitude by the normalized value of the cube of the fundamental amplitude yields the third harmonic relative coefficient. The estimated defect feature size is equal to the reference wavelength multiplied by the weighted sum of the second and third harmonic relative coefficients, where the weight of the second harmonic relative coefficient is 0.6 and the weight of the third harmonic relative coefficient is 0.4. The reference wavelength is the wavelength corresponding to the center frequency of the high-frequency ultrasonic transducer array.

[0052] The fundamental amplitude is the peak amplitude of the fundamental frequency component in the echo signal received by the high-frequency ultrasonic transducer array. The fundamental amplitude when there is no defect is the peak amplitude of the fundamental frequency component in the echo signal received when testing a defect-free standard test block. The second harmonic amplitude is the peak amplitude of the second harmonic component in the echo signal. The third harmonic amplitude is the peak amplitude of the third harmonic component in the echo signal.

[0053] The defect feature size is the maximum projected length of the defect in the direction of ultrasonic wave propagation. The defect feature size is used to update the preliminary estimate of the defect size output by the defect identification model.

[0054] The imaging reconstruction algorithm steps of the synthetic aperture focusing technology include: establishing a three-dimensional imaging grid in the detection area and discretizing the space into voxel units; controlling the phased array transducer to sequentially acquire A-scan signals at multiple spatial locations; calculating the acoustic path distance from each voxel to all array element positions; converting the acoustic path distance into propagation delay based on the acoustic velocity parameters of the insulating material; using the Fermat principle to iteratively solve for the minimum acoustic path when considering the anisotropy of the material's acoustic velocity; time-shifting and aligning the A-scan signals received by each array element according to the calculated propagation delay; coherently superimposing all time-shifted A-scan signals; using the superimposed amplitude as the scattering intensity value of the voxel; traversing all voxels to complete full-field imaging; and arranging the scattering intensity values ​​of all voxels according to three-dimensional spatial coordinates to form a defect distribution spatial feature matrix.

[0055] The A-scan signal is a one-dimensional time-domain echo signal received by a single array element at a single spatial location. The voxel unit is the smallest imaging unit in a three-dimensional imaging grid. The acoustic path distance is the length of the ultrasonic wave propagation path from the array element location through the voxel location to another array element location. The sound velocity parameter is the propagation speed of ultrasonic waves in the electrical insulation component material. The propagation delay is the time required for ultrasonic waves to propagate along the acoustic path distance.

[0056] The phased array transducer is a broadband piezoelectric ultrasonic transducer array or a high-frequency ultrasonic transducer array, and the multiple spatial positions are the various acquisition positions when the phased array transducer moves along a preset scanning path on the surface of the power insulation component.

[0057] The scattering intensity value characterizes the ability of the material at the voxel location to scatter ultrasonic waves. The larger the scattering intensity value, the higher the probability that there is a defect at the voxel location.

[0058] The defect distribution spatial feature matrix is ​​a three-dimensional matrix. The three dimensions of the defect distribution spatial feature matrix correspond to the length direction, width direction and depth direction of the detection area, respectively. The value of each element in the defect distribution spatial feature matrix is ​​the scattering intensity value of the corresponding voxel position. The defect distribution spatial feature matrix is ​​used to characterize the spatial distribution characteristics of defects inside the power insulation component.

[0059] The acoustic characteristic homogeneity assessment value is used to determine the consistency of acoustic characteristics in the defect area. The calculation steps include: extracting all voxels identified as defects from the defect distribution spatial feature matrix; extracting three acoustic parameters for each defect voxel: scattering intensity value, local rate of change of sound velocity, and sound attenuation coefficient; normalizing the three acoustic parameters to the interval between 0 and 1; calculating the standard deviation of the normalized scattering intensity value of all defect voxels as the scattering dispersion; calculating the standard deviation of the normalized local rate of change of sound velocity as the sound velocity dispersion; calculating the standard deviation of the normalized sound attenuation coefficient as the attenuation dispersion; and averaging the scattering dispersion, sound velocity dispersion, and attenuation dispersion with equal weights to obtain the comprehensive dispersion. The acoustic characteristic homogeneity assessment value is equal to 1 minus the comprehensive dispersion.

[0060] The local change rate of sound velocity is the relative change of the sound velocity parameter at the voxel position relative to the overall average sound velocity parameter, and the sound attenuation coefficient is the amplitude attenuation ratio of the ultrasonic wave propagating a unit distance at the voxel position.

[0061] The unified filtering parameters include a unified cutoff frequency, a unified filter order, and a unified gain coefficient. These unified filtering parameters are used to perform the same filtering process on the echo signals from all defective regions.

[0062] The coupling feedback correction vector is used to dynamically adjust the detection parameters according to the coupling state between the transducer and the test piece. The establishment steps include: setting up a pressure sensor array between the transducer and the surface of the insulating component to monitor the contact pressure distribution in real time; collecting the pressure value at each sensor location and normalizing it to a standard contact pressure; calculating the uniformity index of the pressure distribution as the standard deviation of all normalized pressure values; measuring the ultrasonic transmittance of the coupling layer and normalizing it to the transmittance of the ideal coupling state; calculating the normalized value of surface undulation based on surface roughness measurement data; and combining the reciprocal of the uniformity index, the normalized transmittance value, and the reciprocal of the normalized undulation value. The overall coupling quality score is obtained by weighting and summing the scores in a 3:4:3 ratio. When the overall coupling quality score is less than 0.75, a correction process is triggered. In the correction process, the transmission power of the array elements in the region with uneven pressure distribution is adjusted. The ratio of the normalized transmission power adjustment to the original excitation amplitude is used as the power correction factor. The excitation pulse width is increased according to the position with insufficient transmittance. The ratio of the normalized excitation pulse width increment to the original excitation pulse width is used as the pulse width correction factor. The product of the power correction factor and the pulse width correction factor corresponding to each array element is arranged according to the array number to form the coupling feedback correction vector.

[0063] The pressure sensor array consists of multiple piezoelectric thin-film sensors arranged on the back of the transducer. The standard contact pressure is a recommended contact pressure value to ensure good acoustic coupling. The uniformity index is used to characterize the uniformity of the contact pressure distribution. The coupling layer is a coupling agent filled between the transducer and the surface of the electrical insulation component. The ultrasonic transmittance is the proportion of energy transmitted by ultrasonic waves through the coupling layer. The transmittance of the ideal coupling state is the ultrasonic transmittance when the coupling layer thickness and acoustic performance are both optimal. The surface roughness measurement data is a statistical measurement of the height of the micro-undulations on the surface of the electrical insulation component. The normalized value of the surface undulation is the ratio of the surface roughness measurement data to the roughness reference value of the smooth surface.

[0064] The personalized beam deflection angle is an optimized beam incident angle calculated for a single defect region based on its location and the distribution of the interface between the surrounding materials. The focusing depth parameter is an optimized beam focusing depth value calculated for a single defect region based on its depth position. The personalized beam deflection angle and focusing depth parameter are determined by the correction factor of the corresponding array element in the coupled feedback correction vector.

[0065] The defect spatial location coordinates are the three-dimensional coordinate values ​​of the defect in the defect distribution space feature matrix. The defect spatial location coordinates are accurately determined by the preliminary estimation result of the defect spatial location output by the defect identification model and then reconstructed by synthetic aperture focusing technology imaging.

[0066] The steps for establishing the training dataset for the defect identification model specifically include: creating insulation component samples containing six types of defects: crack defects, bubble defects, inclusion defects, delamination defects, discharge ablation defects, and aging and deterioration defects. 200 samples of different sizes and depths are created for each type of defect. Echo signals from all samples are acquired using a calibrated ultrasonic testing system. The defect type, three-dimensional position coordinates, defect feature size, and depth information of each sample are recorded as labels. A short-time Fourier transform is performed on the acquired echo signals to generate an echo signal time-frequency feature matrix. This matrix is ​​normalized to the 0-1 range. Data augmentation methods are used to apply different degrees of Gaussian noise, time shift, and amplitude scaling to the original data to generate augmented samples. Finally, a training dataset containing 7200 sets of labeled echo signal time-frequency feature matrices is established, divided into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0067] The specific steps for training the defect recognition model include: initializing the convolutional kernel parameters of the encoder and decoder to follow a Xavier distribution; setting the initial learning rate to 0.001 and dynamically adjusting it using a cosine annealing strategy; the loss function consists of four parts: reconstruction loss, KL divergence loss, classification loss, and regression loss. The reconstruction loss uses the mean squared error between the input and the reconstructed output to measure the reconstruction quality of the autoencoder; the KL divergence loss constrains the distribution of latent variables in each layer to approximate a standard normal prior distribution; the classification loss uses cross-entropy loss to measure the accuracy of defect type prediction; and the regression loss uses smoothing L1 loss to measure the accuracy of location prediction. These four losses... The total loss is calculated by weighting the values ​​in a 4:2:3:1 ratio. The Adam optimizer is used for gradient descent updates, with a batch size of 32 and 200 training epochs. After each epoch, the model performance is evaluated on the validation set. An early stopping mechanism is triggered when the validation set loss does not decrease for 15 consecutive epochs. The model parameters with the best performance on the validation set are saved as the final model. The reparameter technique is used to transform the random sampling operation into a deterministic transformation during backpropagation, so that the gradient can be effectively transmitted to the encoder parameters. The prior matching loss regularizes the latent space by minimizing the KL divergence between the latent distribution and the prior distribution, preventing the model from learning meaningless latent representations.

[0068] The hierarchical variational autoencoder (VAE) constructs a hierarchical structure of multi-level latent variables, enabling the model to capture features of ultrasonic echo signals at different levels of abstraction. The first-level latent variables capture the fine-grained time-frequency texture of the signal; the second-level latent variables extract medium-scale defect patterns; the third-level latent variables encode the geometric morphological features of the defects; and the fourth-level core latent space learns the high-level semantic representation of the defects. This hierarchical structure gives the model stronger generalization ability to defects of different depths and sizes. The conditional independence assumption simplifies the inference process, ensuring that the posterior distribution of each latent variable depends only on the latent variables of the previous level without considering the complex dependencies of all levels, thus reducing computational complexity. The reparameter technique achieves this by externalizing randomness as a noise variable independent of the model parameters. End-to-end gradient backpropagation enables the model to be trained using the standard gradient descent algorithm. Prior matching loss constrains the latent distribution of each layer to be close to the standard normal prior, preventing posterior collapse in the latent space and ensuring the continuity and interpolability of the latent representation. This makes the defect feature representation generated by the model more stable and reliable. In practical applications, the hierarchical structure enables the model to process multi-scale defect information in the signal. Even under strong noise and material interface interference, it can still accurately identify the defect type and location. Compared with the traditional single-layer autoencoder, the hierarchical variational autoencoder improves the recognition accuracy of small sample defect data by 18 percentage points and the recall rate of deep weak defect detection by 23 percentage points, significantly enhancing the diagnostic capability of the ultrasonic testing system for internal defects in power insulation components.

[0069] The present invention also provides a method for forming an ultrasonic detection system for internal defects of power system insulation components based on a piezoelectric ultrasonic transducer, which is implemented by a computer. The computer is provided with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-described method.

[0070] The specific implementation methods of the above steps are described in detail below.

[0071] The specific implementation of step S01 involves fixing a broadband piezoelectric ultrasonic transducer array with a working frequency range of 2MHz to 15MHz onto the surface of an electrically insulating component using a coupling agent layer. The array contains multiple piezoelectric ceramic wafers with dimensions of 0.5mm × 10mm, an element spacing of 0.6mm, and a total length of 60mm. First, measured acoustic impedance data of different material interfaces are collected to establish an acoustic impedance difference database. The reflection coefficient and transmission coefficient of each interface are calculated using the acoustic impedance difference. Interfaces with higher reflection coefficients will generate strong reflected waves that mask defect signals. Therefore, interfaces are prioritized according to their reflection coefficients. For high-priority interfaces, the required phased-array deflection angle and dynamic focusing depth are calculated using the phased-array focusing principle. The phased-array deflection angle is normalized to the 0-1 range as an angle weight, and the ratio of the dynamic focusing depth to the maximum detection depth is used as a depth weight. These two factors are combined to form the excitation delay correction factor for each element. The excitation delay correction factors of all elements are arranged according to the array sequence number to form an interface masking suppression vector. The first element of this vector... The element represents the first element. The delay ratio coefficient of each array element relative to the reference excitation time is used to determine the excitation voltage amplitude of each element. The excitation pulse width is determined according to the coding sequence. The excitation delay is the product of the delay ratio coefficient and the reference excitation time. The phased array technology is used to achieve directional deflection and focusing of the sound beam, so that the ultrasonic beam avoids strong reflection interfaces or optimizes the incident angle to reduce the masking effect of interface reflection waves on defect echoes. The pulse compression technology is used to transmit frequency-modulated coded pulse signals. The signal frequency is linearly swept from 2MHz to 15MHz within a transmission duration of 10μs. The receiver uses matched filtering to compress the echo signal. The pulse response of the matched filter is the time reversal of the transmitted signal. The signal compression is achieved through convolution operation, compressing the time-domain broadened coded signal into a narrow pulse. The main lobe width of the compressed pulse is significantly reduced while the amplitude is greatly increased, thereby improving the signal-to-noise ratio and range resolution. The ultrasonic reflection signals received by each array element include interface echoes from the material interface and defect echoes from defects.

[0072] The specific implementation of step S02 involves performing wavelet multi-scale decomposition processing on the received echo signal. Discrete wavelet transform is used to decompose the echo signal into wavelet coefficients of different frequency bands. Wavelet transform utilizes the time-frequency localization characteristics of wavelet basis functions, enabling simultaneous analysis of the signal in both the time and frequency domains. A soft thresholding method is used to suppress noise in high-frequency wavelet coefficients. The soft thresholding function sets coefficients below the threshold to zero and shrinks coefficients above the threshold. The threshold is set to 2 to 3 times the noise standard deviation. Low-frequency wavelet coefficients remain unchanged to preserve the main signal characteristics. The denoised echo signal is obtained through wavelet reconstruction. The path delay parameter and sound pressure amplitude parameter are extracted from the denoised echo signal. The path delay parameter is the time difference between the arrival time of the defect echo at the receiving array element and the emission time of the ultrasonic beam. The sound pressure amplitude parameter is the peak voltage amplitude of the defect echo. For weak echo signals from deep defects, amplitude compensation is required. First, the attenuation coefficient of the ultrasonic wave at different frequencies in the target insulating material is measured, and a fitting curve between the attenuation coefficient and frequency is established. The detection depth range is divided equally along the thickness direction. Each depth layer divides the frequency range into equal operating frequency bands. For each frequency band, the theoretical attenuation is calculated for the combination of each depth layer and frequency band. The attenuation is calculated as the product of the propagation distance and the attenuation coefficient according to the acoustic attenuation law. The reciprocal of the theoretical attenuation is normalized to the reciprocal of the attenuation at the reference depth to obtain the compensation factor. The system is constructed with depth layers as rows and frequency bands as columns. OK The deep attenuation compensation matrix is ​​used to determine the corresponding depth layer based on the arrival time of the echo signal. The echo signal is then subjected to a fast Fourier transform for spectral decomposition, and amplitude compensation is achieved by multiplying each frequency band by the corresponding compensation factor. For weak defect echo signals whose amplitude is still less than three times the noise level after compensation, a time-domain averaging enhancement method is used to superimpose and average multiple echo signals from the same spatial location. By utilizing the correlation of the signal and the randomness of the noise, the more times the signal-to-noise ratio is superimposed, the more significant the signal-to-noise ratio improvement, thereby improving the detection sensitivity of deep weak defects.

[0073] The specific implementation of step S03 involves generating a time-frequency feature matrix of the preprocessed echo signal through a short-time Fourier transform. The rows of the matrix represent the frequency dimension, the columns represent the time dimension, and the element values ​​represent the energy intensity at the corresponding time-frequency point. This matrix is ​​input into the defect identification model, which employs a four-layer hierarchical variational autoencoder structure. The first encoder uses a convolutional neural network to extract local time-frequency patterns and maps the input features to a 256-dimensional latent space. The output mean and variance vectors are used to parameterize the Gaussian distribution of the latent variables. The second encoder uses the latent variables from the first layer as conditional inputs and maps them to a 128-dimensional latent space. The third encoder maps to a 64-dimensional latent space, and the fourth encoder maps to a 32-dimensional core latent space. The convolutional kernel parameters of each encoder layer are initialized according to a Xavier distribution to ensure stable gradient propagation. The decoder uses a symmetrical four-layer structure, employing transposed convolutions to achieve upsampling and reconstructing the latent representation of the previous layer layer until the original input is reconstructed. A defect classification head branch from the second latent space contains two fully connected layers and uses a softmax activation function. The model outputs the probability distribution of six types of defects, including cracks, bubbles, inclusions, delamination, discharge ablation, and aging / deterioration. At the branching point in the third potential space, the regression head outputs a preliminary estimate of the three-dimensional spatial coordinates of the defects. The multi-head attention mechanism in the model adaptively adjusts the number of heads based on the signal-to-noise ratio (SNR) of the echo signal: eight heads are used when the SNR is greater than 20 dB, four heads when the SNR is between 10 and 20 dB, and two heads when the SNR is less than 10 dB. The multi-head attention mechanism selectively focuses on important time-frequency features by calculating the dot product similarity between the query vector and the key vector, and then weighting the vector after softmax normalization. This enhances the model's ability to extract key defect features. After outputting the probability distribution of defect types, the model estimates the defect size based on the defect type corresponding to the maximum probability. When the estimated defect size is less than one-quarter of the wavelength corresponding to the center frequency of the high-frequency ultrasonic transducer array, the traditional ultrasonic detection method suffers from insufficient resolution due to Rayleigh scattering. In this case, a micro-defect resolution enhancement process is initiated for further detection.

[0074] The specific implementation of step S04 involves switching the detection system to a high-frequency ultrasonic transducer array with a center frequency of 50MHz and a bandwidth of 20MHz. The array contains 64 array elements with a size of 0.2mm × 5mm. The shorter wavelength of high-frequency ultrasonic waves can improve the resolution of minute defects. A nonlinear ultrasonic detection method is used to excite ultrasonic waves to generate nonlinear effects at the defect. When the ultrasonic waves pass through the defective material, the nonlinear elastic properties of the material and the contact nonlinearity of the defect interface cause the ultrasonic waveform to be distorted, generating higher harmonic components. The received echo signal is subjected to spectrum analysis to extract the second and third harmonic components. The frequency of the second harmonic component is equal to twice the fundamental frequency, and the frequency of the third harmonic component is equal to three times the fundamental frequency. Higher harmonics are more sensitive to minute defects. The defect feature size is calculated through a defect scale mapping function. First, the fundamental amplitude is divided by the fundamental amplitude of the defect-free standard test block to obtain the normalized value of the fundamental amplitude. The second harmonic amplitude is divided by the normalized value of the square of the fundamental amplitude to obtain the relative coefficient of the second harmonic. The third harmonic amplitude is divided by the normalized value of the cube of the fundamental amplitude to obtain the relative coefficient of the third harmonic. The estimated value of the defect feature size is equal to the reference wavelength multiplied by the weighted sum of the relative coefficients of the second and third harmonics, where the weight of the relative coefficient of the second harmonic is 0.6 and the weight of the relative coefficient of the third harmonic is 0.4. The reference wavelength is the wavelength corresponding to the center frequency of the high-frequency ultrasonic transducer array, which is approximately 60 μm. The calculated defect feature size is the maximum projected length of the defect in the direction of ultrasonic wave propagation. This size is used to update the preliminary estimate of the defect size output by the defect identification model in step S03, thereby achieving accurate size characterization of small defects.

[0075] The specific implementation of step S05 is based on synthetic aperture focusing technology to perform time-delay superposition processing on A-scan signals acquired from multiple locations to generate high-resolution three-dimensional imaging results. First, a three-dimensional imaging grid is established in the detection area, and the space is discretized into voxel units. The size of the voxel unit is determined according to the desired imaging resolution, typically on the order of a quarter wavelength. The phased array transducer is controlled to move along a preset scanning path on the surface of the power insulation component and sequentially acquires A-scan signals at multiple spatial locations. The A-scan signal is a one-dimensional time-domain echo signal received by a single array element at a single spatial location. For each voxel, the acoustic path distance to all array element locations is calculated. The acoustic path distance is the total length of the ultrasonic wave propagation path from the transmitting array element location through the voxel locations to the receiving array element location. The propagation time delay is obtained by dividing the acoustic path distance by the sound velocity parameter of the insulation material. When the material has sound velocity anisotropy, Fermat's principle is used to iteratively solve for the minimum acoustic path. The Mach-Krödinger principle states that the actual propagation path of a sound wave in a medium is the path that maximizes the propagation time. The path is iteratively optimized using a variational method until the convergence condition is met. The A-scan signal received by each array element is time-shifted and aligned according to the calculated propagation delay, so that the scattered echoes from the same voxel are aligned in the time domain. All time-shifted A-scan signals are coherently superimposed. Coherent superposition utilizes the characteristic that the signal from the target voxel has a consistent phase while the noise phase is random to achieve signal enhancement and noise suppression. The amplitude of the superposition is used as the scattering intensity value of the voxel. The larger the scattering intensity value, the higher the probability that there is a defect at the location of the voxel. Full-field imaging is completed by traversing all voxels. The scattering intensity values ​​of all voxels are arranged in three-dimensional spatial coordinates to form a defect distribution spatial feature matrix. The three dimensions of this matrix correspond to the length, width, and depth directions of the detection area, respectively, and are used to characterize the spatial distribution characteristics of defects inside the power insulation component.

[0076] The specific implementation of step S06 involves assessing the homogeneity of acoustic characteristics in the defect region based on the defect distribution spatial feature matrix. First, all voxels identified as defects in the matrix are extracted. For each defect voxel, three acoustic parameters are extracted: scattering intensity, local rate of change of sound velocity, and sound attenuation coefficient. The local rate of change of sound velocity is the relative change of the sound velocity parameter at the voxel location relative to the overall average sound velocity parameter. The sound attenuation coefficient is the amplitude attenuation ratio of the ultrasonic wave propagating a unit distance at the voxel location. These three acoustic parameters are normalized to the interval between 0 and 1. The standard deviation of the normalized scattering intensity values ​​for all defect voxels is calculated and recorded as the scattering dispersion. The standard deviation of the normalized local rate of change of sound velocity is also calculated. The sound velocity dispersion is denoted as [dispersion value], and the standard deviation of the normalized sound attenuation coefficient is denoted as [attenuation dispersion value]. The three dispersion values ​​are averaged with equal weights to obtain the comprehensive dispersion value. The acoustic characteristic homogeneity assessment value is equal to 1 minus the comprehensive dispersion value, with the value ranging from 0 to 1. When the assessment value is between 0.82 and 1, it indicates that the acoustic characteristics of the defect area are highly consistent. In this case, a uniform cutoff frequency filter order and gain coefficient are used to perform the same filtering process on the echo signals of all defect areas to improve processing efficiency. When the assessment value is between 0 and 0.82, it indicates that the acoustic characteristics of different defect areas differ significantly. In this case, it is necessary to apply individual [filters] to each defect area based on the coupling feedback correction vector. The process of establishing the personalized acoustic beam deflection angle and focusing depth parameters, and the coupling feedback correction vector, involves setting up a piezoelectric thin-film sensor array on the back of the transducer to monitor the contact pressure distribution in real time, collecting the pressure value at each sensor location and normalizing it to the standard contact pressure, calculating the uniformity index of the pressure distribution as the standard deviation of all normalized pressure values, measuring the ultrasonic transmittance of the coupling layer and normalizing it to the transmittance of the ideal coupling state, calculating the normalized value of surface undulation based on surface roughness measurement data, and weighting the reciprocal of the uniformity index (transmittance normalization value) and the reciprocal of the undulation normalization value in a 30%:40%:30% ratio to obtain the comprehensive coupling quality score. When the comprehensive coupling quality... When the score is less than 0.75, a correction process is triggered. The transmission power of the array elements in the region with uneven pressure distribution is adjusted accordingly. The ratio of the normalized transmission power adjustment to the original excitation amplitude is used as the power correction factor. The excitation pulse width is increased for the locations with insufficient transmittance. The ratio of the normalized excitation pulse width increment to the original excitation pulse width is used as the pulse width correction factor. The product of the power correction factor and the pulse width correction factor for each array element is arranged according to the array number to form a coupling feedback correction vector. Based on this vector, the personalized beam deflection angle and focusing depth parameters are determined and re-imaging is performed. The impact of imperfect coupling state on imaging quality is compensated by adaptively adjusting the detection parameters.

[0077] The specific implementation of step S07 is to output complete defect detection results, including defect type, defect spatial location coordinates, defect feature size, and defect distribution spatial feature matrix. The defect type is the category corresponding to the maximum probability value among the six defect probability distributions output by the defect identification model in step S03. The defect spatial location coordinates are the three-dimensional coordinates of the defect in the defect distribution spatial feature matrix. These coordinates are accurately determined by the preliminary estimation results output in step S03 after imaging reconstruction using synthetic aperture focusing technology in step S05. The defect feature size is the maximum projection length of the defect in the ultrasonic wave propagation direction calculated in step S04 using the nonlinear ultrasonic detection method and defect scale mapping function. The defect distribution spatial feature matrix intuitively displays the three-dimensional spatial distribution characteristics of the defect inside the power insulation component, providing a reliable basis for the insulation status assessment and fault diagnosis of power equipment.

[0078] It should be noted that the key technical ideas of this invention include interface masking suppression technology, hierarchical variational autoencoder defect identification technology, and adaptive acoustic characteristic matching technology. Interface masking suppression technology establishes an interface acoustic impedance difference database and dynamically adjusts the excitation parameters of each element of the phased array according to the priority of the reflection coefficient. This achieves directional deflection and focusing of the sound beam to avoid strongly reflecting interfaces. Compared with the traditional fixed emission mode, it can significantly reduce the masking effect of interface echoes on deep defect signals, improving the detection rate of deep defects in multi-layer composite structures of power insulation components. The hierarchical variational autoencoder defect identification technology constructs a four-layer hierarchical structure of latent variables, enabling the model to capture multi-scale features of ultrasonic echo signals at different abstraction levels. The first layer captures fine-grained time-frequency textures; the second layer extracts medium-scale defect patterns; the third layer encodes geometric morphological features; and the fourth layer learns high-level semantic representations. Compared with traditional single-layer autoencoders, this hierarchical structure gives the model stronger generalization ability and noise resistance to defects of different depths and sizes. Simultaneously, the conditional independence assumption simplifies the inference process and reduces computational complexity. The reparameter trick and prior matching loss ensure the continuity and interpolability of the latent space, making the defect feature representation more stable and reliable. Adaptive acoustic characteristic matching technology automatically selects between batch processing and personalized processing strategies by calculating the homogeneity evaluation value of acoustic characteristics in defect areas. When acoustic characteristics are highly consistent, uniform filtering parameters are used to improve processing efficiency. When acoustic characteristics differ significantly, personalized acoustic beam parameters are applied to each defect area based on the coupling feedback correction vector, and re-imaging is performed. This technology dynamically adjusts detection parameters according to real-time monitoring of contact pressure distribution, transmittance, and surface roughness, effectively compensating for the impact of imperfect coupling on imaging quality compared to traditional fixed-parameter methods. The synergistic effect of these three key technologies enables intelligent detection throughout the entire process, from signal transmission parameter optimization to feature extraction and recognition to adaptive adjustment of imaging parameters. Interface masking suppression technology provides high-quality input signals for subsequent defect identification, hierarchical variational autoencoder accurately identifies defect types and locations, providing prior information for imaging, and adaptive acoustic characteristic matching technology optimizes imaging parameters based on defect characteristics to further improve spatial resolution. Compared to traditional methods, this synergistic mechanism significantly enhances the detection capability of complex defects inside power insulation components, especially maintaining stable detection performance under conditions of strong interface interference, deep weak defects, and imperfect coupling.

[0079] It should be noted that this invention also solves the following technical problem: insufficient defect localization accuracy caused by differences in acoustic properties among layers in multilayer composite insulation materials. Traditional ultrasonic testing assumes that the sound velocity in the material is uniform, but there are significant differences in the sound velocity, acoustic impedance, and attenuation coefficient among the layers of composite insulation materials. This causes refraction and reflection of ultrasonic waves during propagation between layers, resulting in large errors in defect localization algorithms based on the straight-line propagation assumption. This invention uses a path optimization algorithm based on Fermat's principle to iteratively solve for the minimum path of ultrasonic waves in anisotropic materials, accurately calculates the actual propagation delay from each array element to the imaging grid point, and judges the consistency of material properties in the defect area by evaluating the homogeneity of acoustic properties. When the evaluation value is low, personalized beam deflection angle and focusing depth parameters are applied to each defect area according to the coupling feedback correction vector, thereby achieving accurate three-dimensional defect localization in complex acoustic environments and reducing the localization error from 3 to 5 mm in traditional methods to less than 1 mm.

[0080] Specifically, the principle of this invention is as follows: This invention solves the technical problem of accurately detecting and identifying deep, weak defects inside electrical insulation components. Its principle lies in establishing technical breakthrough paths for three key constraints in defect detection. First, the interface masking suppression vector calculates the acoustic impedance differences and reflection coefficients of each material interface, setting differentiated excitation delay and amplitude parameters for each array element. This allows the array to form a phased-controlled deflection beam to bypass strong reflection interfaces or to weaken interface echo energy through dynamic focusing during transmission, reducing the masking effect of interfaces on defect signals from the signal source. Second, the deep attenuation compensation matrix, based on the frequency-dependent attenuation characteristics of ultrasonic waves propagating in insulating materials, establishes precise compensation factors for different depth layers and frequency bands. This allows the amplitude of weak deep defect echoes to be restored to a identifiable level after compensation, while time-domain averaging further improves the signal-to-noise ratio. Finally, the hierarchical variational autoencoder captures fine-grained texture, medium-scale patterns, geometric morphology, and high-level semantic features of the signal through a hierarchical structure of four latent spaces. Even in environments with strong noise and interface interference, it can accurately extract defect information from multi-scale features. The high-frequency nonlinear detection initiated for micro-defects utilizes the harmonic components generated by the nonlinear response of the material at the defect location, breaking through the resolution limitation of the fundamental wavelength. The synergistic effect of these three components enables the effective detection and accurate identification of deep and weak defects.

[0081] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0082] The specific implementation of step S01 involves fixing a broadband piezoelectric ultrasonic transducer array with a working frequency range of 2MHz to 15MHz onto the surface of an electrically insulating component using a coupling agent layer. The array comprises multiple piezoelectric ceramic wafers with dimensions of 0.5mm × 10mm, an element spacing of 0.6mm, and a total length of 60mm. First, measured acoustic impedance data of different material interfaces are collected to establish an acoustic impedance difference database. The reflection coefficient and transmission coefficient of each interface are calculated using the acoustic impedance difference. The formula for calculating the reflection coefficient is as follows:

[0083] ;

[0084] In the formula, For the first The reflection coefficient of an interface, dimensionless; For the first Acoustic impedance of layered materials, in units of The ratio of ultrasonic wave propagation pressure to particle vibration velocity in a material is obtained through experimental measurement. For the first Acoustic impedance of layered materials, in units of ; The interfaces are numbered sequentially, starting from 1, to mark the interfaces of each material within the detection area. The formula for calculating the transmittance coefficient is as follows:

[0085] ;

[0086] In the formula, For the first The transmission coefficient of the interface is dimensionless. Interfaces with high reflection coefficients will generate strong reflected waves that mask defect signals. Therefore, interfaces are prioritized according to the absolute value of their reflection coefficients. For high-priority interfaces, the required phased-array deflection angle and dynamic focusing depth are calculated using the phased-array focusing principle. The interface masking suppression vector is then determined. The formula for calculating each element is expressed as follows:

[0087] ;

[0088] In the formula, For the first The delay ratio coefficient of each array element, dimensionless; The array element number is used to sequentially mark each array element in the array, starting from 1. For the first The phased deflection angle of each array element, in degrees, is calculated using geometric relationships based on the spatial azimuth angle of the interface to be suppressed. The maximum phase control deflection angle is expressed in degrees, with an empirical value of 30. This is the angle weighting coefficient, which is dimensionless and usually takes a value of 0.5. For the first The dynamic focusing depth of each element, in mm, is equal to the depth position of the interface to be suppressed. The maximum detection depth, in mm, is determined by the transducer frequency and material attenuation characteristics. This is the depth weighting coefficient, dimensionless, and typically takes a value of 0.5. The formula for calculating the excitation delay of each array element is as follows:

[0089] ;

[0090] In the formula, For the first The excitation delay of each array element, in units of ; The reference excitation time is given in units of 1 / 2000. The default value is 10. Frequency-modulated coded pulse signals are transmitted using pulse compression technology, with a signal frequency of 10. The frequency is linearly swept from 2MHz to 15MHz during the transmission duration. The receiver compresses the echo signal through matched filtering. The ultrasonic reflected signal received by each array element includes interface echo from the material interface and defect echo from the defect.

[0091] The specific implementation of step S02 involves performing wavelet multi-scale decomposition processing on the received echo signal. Discrete wavelet transform is used to decompose the echo signal into wavelet coefficients of different frequency bands. For the high-frequency wavelet coefficients, a soft thresholding method is used to suppress noise. The soft thresholding function calculation formula is expressed as follows:

[0092] ;

[0093] In the formula, These are the wavelet coefficients after thresholding, with units consistent with the original echo signal. ; These are the original wavelet coefficients, in units of ; Noise threshold, in units of The default value is 2.5 times the noise standard deviation; For a sign function, when When the value is 1, The value is -1 when The value is 0 at that time; This is a function to find the maximum value. express The absolute value of . The path delay parameter and sound pressure amplitude parameter are extracted from the denoised echo signal. For weak echo signals from deep defects, amplitude compensation is required. The position in the deep attenuation compensation matrix is ​​the . Line 1 The formula for calculating the compensation factor of a column is expressed as follows:

[0094] ;

[0095] In the formula, For the first Depth layer at the 1st The normalized amplitude compensation coefficient for the frequency band is dimensionless. The depth layer number is from 1 to 1. Mark each depth layer within the detection depth range sequentially; Frequency band number, from 1 to Mark the frequency bands within the working frequency band in sequence; For the first Attenuation coefficient in the frequency band, in units of The amplitude attenuation law of ultrasonic waves of different frequencies in insulating materials was obtained through experimental measurement. For the first The depth value of the depth layer, in mm; This is a reference depth, in mm, and is typically taken as 1 for the surface depth. It is a natural exponential function; The total number of depth layers is dimensionless and is determined based on the detection depth range and depth resolution. The total number of frequency bands is dimensionless and determined based on the operating bandwidth and frequency resolution. In application, the corresponding depth layer is determined based on the arrival time of the echo signal. A fast Fourier transform is performed on the echo signal for spectral decomposition. Amplitude compensation is achieved by multiplying each frequency band by the corresponding compensation factor. For weak defect echo signals whose amplitude is still less than three times the noise level after compensation, a time-domain averaging enhancement method is used to superimpose and average multiple echo signals from the same spatial location, thereby improving the detection sensitivity for deep, weak defects.

[0096] The specific implementation of step S03 is to generate an echo signal time-frequency feature matrix by short-time Fourier transforming the preprocessed echo signal, input the matrix into the defect identification model, the model adopts a four-layer hierarchical variational autoencoder structure, and after the model outputs the defect type probability distribution, the defect size is estimated according to the defect type corresponding to the maximum probability. When the estimated defect size is less than one-quarter of the wavelength corresponding to the center frequency of the high-frequency ultrasonic transducer array, the micro-defect resolution enhancement process is started for further detection.

[0097] The specific implementation of step S04 involves switching the detection system to a high-frequency ultrasonic transducer array with a center frequency of 50MHz and a bandwidth of 20MHz. A nonlinear ultrasonic detection method is used to excite ultrasonic waves to generate a nonlinear effect at the defect. The received echo signal is then subjected to spectral analysis to extract the second and third harmonic components. The formula for calculating the estimated defect feature size is as follows:

[0098] ;

[0099] In the formula, This is an estimated value for the defect feature size, in units of... ; Reference wavelength, unit: The reference wavelength is equal to the speed of sound divided by the center frequency of the high-frequency ultrasonic transducer array, calculated using the following formula: ,in The speed at which ultrasound propagates in insulating materials, measured in units of 1000 m / s. This is obtained by experimentally measuring the propagation time of ultrasound waves in a material at a known distance. The center frequency of the high-frequency ultrasonic transducer array is given in units of 1000 m / s. The value is 50, and the calculation is as follows: Approximately 60; This is the weight of the relative coefficient of the second harmonic, dimensionless, and usually taken as 0.6; This is the weight of the relative coefficient of the third harmonic, dimensionless, and usually taken as 0.4; This is the fundamental amplitude, in units of ; The fundamental amplitude when there are no defects, in units of This was obtained through testing of defect-free standard test blocks; This represents the second harmonic amplitude, in units of... ; The third harmonic amplitude is expressed in units of 1. ; The normalization constant for the second harmonic is dimensionless and has an empirical value of 1.5. The third harmonic normalization constant is dimensionless and has an empirical value of 2.0. The calculated defect feature size is the maximum projected length of the defect in the ultrasonic wave propagation direction. This size is used to update the preliminary estimate of the defect size output by the defect identification model in step S03.

[0100] The specific implementation of step S05 is based on synthetic aperture focusing technology to perform time-delay superposition processing on A-scan signals acquired at multiple locations to generate high-resolution three-dimensional imaging results. First, a three-dimensional imaging grid is established in the detection area and the space is discretized into voxel units. The phased array transducer is controlled to move along a preset scanning path on the surface of the power insulation component and sequentially acquires A-scan signals at multiple spatial locations. The sound path distance from each voxel to all array element positions is calculated. The propagation time delay is obtained by dividing the sound path distance by the sound velocity based on the sound velocity parameter of the insulation material. The formula for calculating the scattering intensity value is expressed as follows:

[0101] ;

[0102] In the formula, voxels The scattering intensity value, in units of ; , where is the three-dimensional spatial coordinate of a voxel, in mm; The number of spatial acquisition locations is dimensionless and determined based on the scanning path and sampling interval. The number of array elements is dimensionless. The number of array elements in the broadband piezoelectric ultrasonic transducer array is calculated based on the total array length of 60 mm and the element spacing of 0.6 mm. Spatial location number, from 1 to Mark each data collection location sequentially; The array element numbers are from 1 to 1. Mark each element in the array sequentially; For the first The position The A-scan signal received by each array element at time... The amplitude, in units of ; For from the first The position From individual elements to voxels The propagation delay, in units of The calculation formula is: ,in For the first The position The spatial coordinates of each array element, in mm, are determined through scanning path planning and array geometry. Full-field imaging is completed by traversing all voxels, and the scattering intensity values ​​of all voxels are arranged according to their three-dimensional spatial coordinates to form a spatial feature matrix of defect distribution.

[0103] The specific implementation of step S06 is to evaluate the homogeneity of acoustic characteristics of the defect region based on the defect distribution spatial feature matrix. First, all voxels identified as defects in the matrix are extracted. For each defect voxel, three acoustic parameters are extracted: scattering intensity value, local rate of change of sound velocity, and sound attenuation coefficient. The formula for calculating the local rate of change of sound velocity is as follows:

[0104] ;

[0105] In the formula, For the first The local rate of change of sound velocity of a defective voxel, dimensionless; The defective voxel number is from 1 to Mark each defective voxel in sequence; For the first The sound velocity parameter at each defect voxel location, in units of The result was obtained by analyzing the difference in ultrasonic wave propagation time between adjacent paths of the voxel; The overall average sound speed parameter, in units of , which is equal to the arithmetic mean of the sound speeds of all voxels; The total number of defective voxels is dimensionless. The formula for calculating the acoustic property homogeneity assessment value is as follows:

[0106] ;

[0107] In the formula, This is a dimensionless value used to assess the degree of homogeneity in acoustic properties. The scattering discreteness is dimensionless and is calculated using the following formula: ,in For the first The scattering intensity value of each defect voxel, in units of , This is the maximum scattering intensity value, in units of , The average value of the normalized scattering intensity is dimensionless. The sound velocity dispersion is dimensionless and is calculated using the following formula: ,in The local rate of change of the maximum speed of sound is dimensionless. The attenuation dispersion is dimensionless and is calculated using the following formula: ,in For the first The acoustic attenuation coefficient of each defect voxel, in units of The results were obtained by analyzing the attenuation pattern of the echo amplitude before and after the voxel. The maximum sound attenuation coefficient, in units of .when Batch processing is performed using uniform filtering parameters between 0.82 and 1. When the value is between 0 and 0.82, personalized beam deflection angle and focusing depth parameters need to be applied to each defect region based on the coupling feedback correction vector. The comprehensive coupling quality score calculation formula is expressed as follows:

[0108] ;

[0109] In the formula, The comprehensive coupling quality score is dimensionless. The pressure distribution uniformity index is dimensionless, and its calculation formula is: ,in The pressure sensor serial number is from 1 to Label each pressure sensor in sequence. For the first The pressure value measured by each pressure sensor, in units of... , Standard contact pressure, unit: The experience value is 100. The number of pressure sensors is dimensionless. The average value of the normalized pressure values ​​is dimensionless. The standard uniformity index is dimensionless, with an empirical value of 0.05. The ultrasonic transmittance of the coupling layer is dimensionless and is obtained by measuring the ratio of the transmitted signal amplitude with and without the coupling layer. The transmittance under ideal coupling conditions is dimensionless and has an empirical value of 0.95. Surface roughness, unit: The surface roughness was obtained by measurement using a surface roughness tester. This is a reference value for the roughness of a smooth surface, in units of... The experience value is 1. When When the value is less than 0.75, the correction process is triggered, and the first coupled feedback correction vector is... The formula for calculating each element is expressed as follows:

[0110] ;

[0111] In the formula, For the first The coupling feedback correction factor for each array element is dimensionless. For the first The transmit power adjustment amount at each array element position, in units of The pressure deviation is determined based on the pressure deviation measured by the pressure sensor at that location. This is the maximum transmit power adjustment amount, in units of ; For reference transmission power adjustment, the unit is... The experience value is 5; For the first The increment of the excitation pulse width at each element position, in units of The value is determined based on the transmittance deviation at that location. The maximum excitation pulse width increment, in units of ; The reference excitation pulse width increment is in units of The empirical value is 50. Based on this vector, the personalized beam deflection angle and focus depth parameters are determined and re-imaged.

[0112] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0113] It should be explained that the reflection coefficient calculation formula is based on the principle of acoustic impedance mismatch. It assesses the interface reflection intensity by quantifying the difference in acoustic impedance between the materials on both sides of the interface. The calculation formula is as follows: This formula enables the excitation parameters to specifically suppress the interface masking effect. The quantitative evaluation formula for the interface echo suppression effect is as follows: ,in This represents the interface echo suppression effect, measured in units of... , To suppress the amplitude of the front interface echo, the unit is , To suppress the amplitude of the back interface echo, the unit is... Experimental results show that this formula achieves an interface echo suppression effect of 15%. The above improves the detection sensitivity of deep defects.

[0114] Transmission coefficient calculation formula The formula for the reflection coefficient satisfies the energy conservation relationship, and the sum of their squares is close to 1, providing a theoretical basis for a comprehensive evaluation of interface properties.

[0115] The formula for calculating the delay ratio factor comprehensively considers two key parameters: phase control deflection angle and focusing depth. Through normalization and weighted combination, it achieves adaptive suppression of different interfaces. This formula allows the interface masking suppression vector to dynamically adjust the excitation delay of each array element, effectively reducing the interference of strong interface reflections on defect echoes. In the detection of multilayer composite insulation materials, the interface echo suppression effect reaches 15%. above.

[0116] The formula for calculating the deep attenuation compensation factor is based on the ultrasonic attenuation exponential law. ,in For depth The ultrasonic amplitude at the location, in units of , This is the initial amplitude, in units of , This is the attenuation coefficient, in units of... , The propagation depth is expressed in mm. This formula accurately describes the energy attenuation of ultrasound at different depths and frequencies using an exponential function, achieving quantitative compensation for echoes from deep, weak defects. The quantitative formula for improving the signal-to-noise ratio is as follows: ,in This represents the improvement in signal-to-noise ratio, expressed in units of... , Signal power, unit: , Noise power, unit: Experimental results show that this formula improves the signal-to-noise ratio of defect detection with a depth of 50mm or more by 20%. The above significantly enhances the ability to detect deep defects.

[0117] The defect characteristic size estimation formula is based on nonlinear ultrasonic theory. It utilizes the difference in sensitivity of second and third harmonic waves to minute defects, and achieves accurate size characterization of subwavelength defects through weighted combination. The relationship between the nonlinear parameter and the defect size follows... ,in The coefficients are nonlinear and dimensionless. This formula breaks through the Rayleigh scattering limit of traditional linear ultrasonic methods, achieving a measurement accuracy of ±5 for minute defects smaller than a quarter wavelength. This provides a reliable basis for early defect diagnosis.

[0118] The formula for calculating the scattering intensity value is based on the principle of synthetic aperture focusing imaging. Spatial resolution is improved through the coherent superposition of echo signals from multiple locations and multiple array elements. The resolution improvement factor formula is as follows: ,in Spatial resolution, in mm. The formula, which is the synthetic aperture angle in degrees, utilizes the difference in phase consistency and noise randomness of the target voxel scattering signal to achieve signal enhancement and noise suppression during the superposition process. This results in an imaging spatial resolution of 0.5 mm and a defect location accuracy of ±0.3 mm, meeting the requirements for fine inspection of power insulation components.

[0119] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team conducted internal defect inspection on a batch of 110kV transformer insulating bushings that had been in operation for 5 years. These insulating bushings were made of epoxy resin composite material, with a total length of 850mm, an outer diameter of 180mm, and a wall thickness of 25mm. Due to long-term exposure to electric field and temperature stress, some bushings showed signs of partial discharge. However, conventional electrical tests could not accurately locate the defect position and type, necessitating the use of advanced non-destructive testing technology for diagnosis.

[0120] The technical team employed a piezoelectric ultrasonic transducer-based testing method to comprehensively inspect the insulating sleeve. First, a 0.3mm thick layer of glycerol coupling agent was uniformly coated onto the sleeve surface. Then, a broadband piezoelectric ultrasonic transducer array with an operating frequency range of 2MHz to 15MHz was attached to the outer surface of the sleeve. This array consisted of 100 piezoelectric ceramic wafers arranged linearly, with each element measuring 0.5mm × 10mm, an element spacing of 0.6mm, and a total array length of 60mm. A pressure sensor array consisting of 32 piezoelectric thin-film sensors was arranged on the back of the transducer to monitor the contact pressure distribution in real time. The measured pressure distribution data is shown in Table 1.

[0121] Table 1. Transducer array contact pressure distribution data

[0122]

[0123] Based on the data in Table 1, the pressure distribution uniformity index was calculated to be 0.048, the average ultrasonic transmittance of the coupling layer was 0.95, the normalized surface undulation value was 1.12, and the comprehensive coupling quality score was 0.83, meeting the testing requirements. To address the strong reflectivity of the epoxy resin-air interface, the technical team established an interface masking suppression vector. By measuring the acoustic impedance difference between the epoxy resin and air interfaces, the reflection coefficient was calculated to be 0.998, and the transmittance coefficient to be only 0.002. For this highly reflective interface, the required phase control deflection angle was calculated to be 15 degrees, and the dynamic focusing depth to be 12 mm. After normalizing the phase control deflection angle, an angle weight of 0.25 was obtained. The ratio of the dynamic focusing depth to the maximum detection depth of 25 mm yielded a depth weight of 0.48. Combining these two factors constitutes the excitation delay correction factor for each array element, forming the interface masking suppression vector.

[0124] The technical team employed pulse compression technology to transmit coded ultrasonic beams. The frequency of the frequency-modulated coded pulse signal was linearly swept from 2MHz to 15MHz within a 10μs transmission duration. The excitation parameters of each array element were adjusted based on the interface masking suppression vector. The excitation voltage amplitude of the 50th array element was set to 85V, the excitation pulse width to 10μs, and the excitation delay to [missing information]. s. For example Figure 2 As shown, the echo signal received by the transducer array contains multiple reflected waves from the outer surface, inner surface, and internal defects of the sleeve. The technical team performed wavelet denoising on the received echo signal, using the db4 wavelet for 5-level decomposition, and applied a soft threshold of 0.15 to suppress noise in the high-frequency wavelet coefficients. After reconstruction, the signal-to-noise ratio was improved from the original 8.6dB to 16.2dB.

[0125] The path delay and sound pressure amplitude parameters of the denoised echo signal were extracted. A suspected defect echo was detected located 18 mm deep inside the sleeve, with an arrival time of [time missing]. The peak voltage amplitude was 0.032V. Based on the ultrasonic propagation velocity of 2650m / s in epoxy resin, the defect was located at a depth of 18mm. To compensate for the deep attenuation effect, the technical team established a deep attenuation compensation matrix. The detection depth range of 0 to 25mm was divided into 25 depth layers at 1mm intervals, and the frequency range of 2MHz to 15MHz was divided into 13 frequency bands at 1MHz intervals. The measured attenuation coefficient of epoxy resin at 5MHz was... At a frequency of 10MHz .like Figure 3 As shown, for an 18mm deep layer in the 8MHz frequency band, the theoretical attenuation is calculated to be 5.04dB, corresponding to a normalized amplitude compensation coefficient of 1.79. After applying the deep attenuation compensation matrix to compensate the echo signal amplitude, the defect echo amplitude is increased from 0.032V to 0.057V.

[0126] The compensated echo signal is input into the defect identification model for feature extraction. This model employs a four-layer hierarchical variational autoencoder structure. First, a short-time Fourier transform is performed on the echo signal to generate a 256×128 time-frequency feature matrix. The first encoder layer maps the input features to a 256-dimensional latent space, the second layer to a 128-dimensional latent space, the third layer to a 64-dimensional latent space, and the fourth layer to a 32-dimensional core latent space. Since the echo signal signal-to-noise ratio is 16.2dB, the model adaptively uses a four-head attention mechanism to selectively focus on important time-frequency features. The probability distribution of the six defect types output by the defect classification head shows that the probability of discharge ablation defects is 0.78, the probability of bubble defects is 0.14, and the probabilities of other defect types are all less than 0.05. The preliminary estimated three-dimensional coordinates of the defect output by the position regression head are located at 320mm axially, 18mm radially, and 85 degrees circumferentially along the casing. Based on the defect identification model, the defect size is estimated to be approximately 0.8 mm, which is less than one-quarter of the wavelength of 60 μm corresponding to the center frequency of the high-frequency ultrasonic transducer array of 50 MHz, i.e., 15 μm. Therefore, the micro-defect resolution enhancement process is initiated.

[0127] The technical team switched to a high-frequency ultrasonic transducer array for further testing. This array has a center frequency of 50MHz, a bandwidth of 20MHz, and contains 64 elements, each 0.2mm × 5mm in size. A nonlinear ultrasonic testing method was used to excite the nonlinear effect at the defect. The received echo signal yielded a fundamental amplitude of 0.125V, a second harmonic amplitude of 0.018V, and a third harmonic amplitude of 0.0045V. When testing a defect-free standard test block, the fundamental amplitude was measured to be 0.142V. Based on the defect scale mapping function, the normalized value of the fundamental amplitude was 0.880, the relative coefficient of the second harmonic was 1.15, and the relative coefficient of the third harmonic was 0.52. The reference wavelength was 60μm, and the estimated defect feature size was [value missing]. μm, this result is used to update the preliminary estimate of the defect size output by the defect identification model.

[0128] To accurately determine the spatial location of the defect, the technical team employed synthetic aperture focusing (SAR) technology for 3D imaging. Fifteen acquisition points were set along the axial direction on the casing surface, with a spacing of 40 mm between each point. A-scan signals from 64 array elements were acquired at each point. A 3D imaging grid was established within the detection area, and the space was discretized. The voxel elements are calculated. For each voxel, the acoustic path distance to all array element positions is calculated. Considering the anisotropic characteristics of the sound velocity in epoxy resin, the Fermat principle is used iteratively to solve for the minimum acoustic path. For the voxel located at coordinates (320mm, 18mm, 85 degrees), the calculated acoustic path distance to the 32nd array element at the 8th acquisition position is 29.6mm, corresponding to a propagation delay of... s. The A-scan signals received by all array elements are time-shifted and coherently superimposed according to the calculated propagation delay. The amplitude of the superimposed signal is used as the scattering intensity value of that voxel, which is 1.85. Full-field imaging is completed by traversing all voxels, generating... The defect distribution spatial feature matrix.

[0129] The technical team extracted voxels with scattering intensity values ​​greater than 1.2 from the spatial feature matrix of defect distribution as defect regions, identifying a total of 328 defect voxels. For each defect voxel, three acoustic parameters were extracted and normalized: scattering intensity value, local rate of change of sound velocity, and sound attenuation coefficient. The calculated scattering dispersion was 0.096, sound velocity dispersion was 0.082, attenuation dispersion was 0.104, and the overall dispersion was 0.094. The acoustic characteristic homogeneity assessment value was 0.906. Since this assessment value is within the range of 0.82 to 1.0, it indicates that the acoustic characteristics of the defect regions are relatively consistent. Therefore, uniform filtering parameters were used for batch processing. A uniform cutoff frequency of 12MHz, filter order of 4, and gain coefficient of 1.5 were set. The echo signals from all defect regions underwent the same filtering process and were re-imaged. Finally, the spatial coordinates of the defect were accurately determined to be 318mm axially, 18.2mm radially, and 84 degrees circumferentially along the sleeve.

[0130] The test results showed that the defect was a discharge ablation defect with a feature size of 49.9 μm, located at a depth of 18.2 mm inside the sleeve. The technical team recommended that the sleeve be closely monitored and replaced during the next power outage maintenance. This detection method has significant advantages over traditional ultrasonic C-scan technology. Traditional methods are limited by fixed-frequency probes, making it difficult to effectively detect tiny defects smaller than the wavelength, and are prone to missed detections under strong reflection interference at material interfaces. This method effectively reduces the masking effect of interface reflection waves on defect signals through interface masking suppression vector technology, and improves the signal-to-noise ratio through pulse compression technology, enabling clear identification of weak defect signals. The hierarchical variational autoencoder captures ultrasonic echo features at different levels of abstraction, giving the model a stronger generalization ability for defects of different depths and sizes, and accurately identifying defect types even in complex multi-interface material structures. The micro-defect resolution enhancement process, by switching to a high-frequency probe and utilizing nonlinear ultrasonic detection methods, breaks through the resolution limit of traditional linear ultrasonic detection, successfully detecting a discharge ablation defect with a size of only 49.9 μm. Synthetic aperture focusing technology combined with acoustic path optimization algorithms enables precise spatial location of defects, with positioning errors controlled within the sub-millimeter level. A coupling feedback correction mechanism dynamically adjusts detection parameters based on the real-time coupling state between the transducer and the device under test, ensuring the stability and reliability of the detection results. The integrated application of these technologies allows the detection system to accurately identify various types of defects within electrical insulation components in complex environments with strong interface interference and deep attenuation, providing reliable technical assurance for the safe operation of power equipment.

[0131] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for ultrasonic detection of internal defects in power system insulation components based on piezoelectric ultrasonic transducers, characterized by, The method comprises the following steps: arranging a broadband piezoelectric ultrasonic transducer array on the surface of an electrically insulating part, adjusting the excitation parameters of each array element according to an interface masking suppression vector, transmitting a coded ultrasonic beam and receiving a return signal by using a pulse compression technique; after wavelet denoising processing of the received return signal, extracting the sound path time delay parameter and the sound pressure amplitude parameter, and combining a deep attenuation compensation matrix to perform amplitude compensation and time domain average enhancement on the weak defect return signal; calling a defect identification model to perform feature extraction on the compensated return signal, and outputting a defect type probability distribution and a preliminary estimation result of a defect spatial position, and starting a micro defect resolution enhancement process when the defect size is less than one quarter of the wavelength; switching to a high-frequency ultrasonic transducer array according to the micro defect resolution enhancement process, combining a nonlinear ultrasonic detection method to extract second harmonic components and third harmonic components, and calculating defect characteristic sizes by using a defect size mapping function; performing delay stacking processing on the A-scan signals collected at multiple positions based on a synthetic aperture focusing technique, calculating the accurate time delay from each array element to the imaging grid point by using a sound path optimization algorithm, and generating a defect distribution spatial feature matrix; selecting unified filtering parameters or individual beam deflection angles and focusing depth parameters according to the homogenization degree evaluation value of the acoustic characteristics of the defect region in the defect distribution spatial feature matrix and re-imaging; and outputting a defect detection result.

2. The method of claim 1, wherein, The broadband piezoelectric ultrasonic transducer array is composed of a plurality of piezoelectric ceramic wafers arranged in a linear array, and has a working frequency range of 2 MHz to 15 MHz and is in contact with the surface of the electrically insulating part through a coupling agent layer.

3. The method of claim 2, wherein, The interface masking suppression vector is used to reduce the masking effect of strong reflection waves at material interfaces on defect signals, and the i-th element represents the delay proportionality coefficient of the i-th array element relative to the reference excitation time.

4. The method of claim 3, wherein, The excitation parameters include excitation voltage amplitude, excitation pulse width and excitation time delay, which are determined according to the delay proportionality coefficient of the corresponding array element in the interface masking suppression vector and the reference excitation time.

5. The method of claim 4, wherein, The pulse compression technique improves the signal-to-noise ratio by transmitting a frequency-modulated coded pulse signal and performing matched filtering at the receiving end, and the frequency of the frequency-modulated coded pulse signal is linearly swept from the starting frequency to the terminal frequency within the transmission time.

6. The method of claim 5, wherein, The wavelet denoising processing adopts a wavelet multi-scale decomposition method to decompose the return signal into wavelet coefficients of different frequency bands, suppresses noise by using a soft threshold method for high-frequency wavelet coefficients, and obtains the denoised return signal by wavelet reconstruction.

7. The method of claim 6, wherein, The deep attenuation compensation matrix is used to compensate for the energy attenuation of ultrasonic waves propagating in insulating materials, and the element at position p and q represents the normalized amplitude compensation coefficient required for the p-th depth layer in the q-th frequency band.

8. The method of claim 7, wherein, The input layer of the defect identification model receives the preprocessed return signal time-frequency feature matrix, the encoder part adopts a four-layer hierarchical variational autoencoder structure, and the decoder part adopts a symmetric four-layer structure to reconstruct the latent representation of the previous layer layer by layer until the original input is reconstructed.

9. The method of claim 8, wherein, The defect type probability distribution is the probability value of each of the six types of defects output by the defect identification model, and the six types of defects include crack defects, bubble defects, inclusion defects, delamination defects, discharge ablation defects and aging degradation defects.

10. The method of claim 9, wherein, The micro-defect resolution enhancement process is a process flow of further detection by using high-frequency ultrasonic transducer array and nonlinear ultrasonic detection method for defects with a defect size less than one-quarter wavelength.

Images