Method and device for quantifying ultrasonic flaw detection risk of GIS basin insulator and storage medium

Abstract

A GIS basin type insulator ultrasonic flaw detection risk quantification method, device and storage medium belong to the technical field of power equipment detection, solve how to improve the precision of insulator defect ultrasonic flaw detection; the present application optimizes the parameters of multi-element linear array and dynamic focusing rule; the phase compensation superposition of multi-position echo signal is carried out by using synthetic aperture focusing technology, combined with phase cancellation and space-time filtering to suppress interference; the convolutional neural network is constructed to realize hierarchical extraction of defect features, the global features of defects are captured by sine position coding and multi-head self-attention mechanism through the use of transformer, and the feature fusion is completed through the use of feedforward neural network; the defect profile is extracted based on the dynamic threshold segmentation of Otsu method, and the size, depth measurement and risk grade evaluation are realized through pixel quantization and echo time calculation; it is suitable for high-precision detection of hidden defects in insulators, and realizes the intelligentization of the whole process from detection, feature extraction to risk quantification of insulator defects.

Description

Technical Field

[0001] This invention belongs to the field of power equipment testing technology, specifically relating to a method, equipment, and storage medium for quantifying the risk of ultrasonic flaw detection in GIS basin insulators. Background Technology

[0002] With the deepening of smart grid construction, GIS (Gas Insulated Switchgear) has become a core equipment of the power system due to its high reliability. However, its key component, the basin insulator, is prone to hidden defects such as microcracks, air gaps, and metal impurities due to manufacturing process defects and long-term electro-thermal-mechanical stress, which seriously threaten the safe operation of the power grid.

[0003] Traditional ultrasonic testing technology is limited by the number of physical array elements and manual interpretation, resulting in three major bottlenecks: 1) Insufficient resolution: Conventional phased array probes have low sensitivity to detect microcracks smaller than 0.5 mm, and interference from metal inserts can easily lead to echo confusion, making it difficult to identify early defects in complex structures (such as the junction between flanges and the main insulation body); 2) Feature extraction relies on experience: Defect type identification (such as dividing cracks and air gaps) highly depends on the inspector's subjective analysis of the grayscale and texture of ultrasonic images, and the misjudgment rate can reach more than 20% in complex scenarios; 3) Vague risk assessment: There is a lack of quantitative assessment models, which can only qualitatively describe the severity of defects and cannot combine defect size, depth and type to carry out multi-dimensional risk classification, making it difficult to support the formulation of differentiated maintenance strategies.

[0004] Synthetic Aperture Focusing (SAFT) technology can improve resolution by increasing the number of virtual array elements, but single imaging techniques still face the challenge of noise suppression. Convolutional Neural Networks (CNNs), while capable of automatically extracting local defect features, are insufficient at capturing globally correlated features such as long cracks and distributed impurities. Traditional risk assessment methods are mostly based on threshold judgments and do not integrate defect geometric features and type weights, resulting in low reliability of assessment results. How to integrate multi-source data and build a complete technology chain from high-precision imaging to intelligent identification and quantitative assessment has become a key challenge in the current field of GIS basin insulator inspection. Summary of the Invention

[0005] The technical solution of this invention is used to solve the problem of how to improve the accuracy of ultrasonic flaw detection of defects in GIS basin insulators.

[0006] The present invention solves the above-mentioned technical problems through the following technical solutions:

[0007] This invention provides a method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators, comprising:

[0008] S1 employs a multi-element linear ultrasonic array, performs focusing law calculations on the multi-element linear ultrasonic array, optimizes array excitation parameters, and plans the flaw detection and scanning path of the multi-element linear ultrasonic array to achieve accurate coverage and anti-interference acquisition of the ultrasonic beam.

[0009] S2 employs synthetic aperture focusing technology to superimpose phase-compensated echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve the resolution of the detection image.

[0010] S3 constructs a convolutional neural network to achieve hierarchical extraction of defect features and introduces Dropout to enhance generalization ability;

[0011] S4 uses the Transformer model, which captures global defect features through sinusoidal positional encoding and multi-head self-attention mechanism, and completes feature fusion through a feedforward neural network;

[0012] S5 extracts defect contours based on Otsu's method of dynamic threshold segmentation, and achieves size and depth measurement and risk level assessment through pixel quantization and echo time calculation.

[0013] Furthermore, the number of elements in the multi-element linear ultrasonic array mentioned in step S1 is: Array element spacing satisfy ,in, For ultrasonic wavelengths, satisfying conditions, The speed of ultrasonic wave propagation is given by the center frequency. This avoids signal aliasing, and the effective detection width of the probe is... This covers the area of ​​GIS basin insulator flanges.

[0014] Furthermore, the method for calculating the focusing law and optimizing the array excitation parameters of the multi-element linear ultrasonic array described in step S1 is as follows:

[0015] Let the coordinates of the focal point be... , No. The coordinates of each array element are linear array along If the array element is distributed along an axis, then the excitation delay time is:

[0016]

[0017] By delaying the excitation, the sound waves of each array element are superimposed in phase at the focal point to form a high-intensity sound beam.

[0018] Furthermore, the method described in step S2, which uses synthetic aperture focusing technology to superimpose phase compensation on echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve the resolution of the detection image, is as follows:

[0019] The S21 ultrasonic array probe stops at multiple locations point by point, the number of locations being... Trigger all at each position Each element transmits or receives ultrasonic waves, forming an echo matrix. ,in, For scanning the location index, , For array element index, ;

[0020] S22 superimposes the phase compensation signals of the array elements at each position;

[0021] S23 computational array elements in The sound waves emitted by the point reach the measured point and the returned distance value According to the speed of ultrasonic wave propagation The corresponding ultrasonic signal propagation time was obtained. ;

[0022] The effective length of the synthetic aperture is defined as the length corresponding to the farthest position of the array element to which the sound field can radiate, and it is related to the half-power angle of the array element's radiation.

[0023]

[0024] in, The effective length of the synthesized aperture. For the length of the sound wave, Distance to the target point The spacing between array elements;

[0025] Within the effective length, the measured point The relative sound pressure intensity is the superposition amplitude of the radiated sound fields of each array element at this point:

[0026]

[0027] Based on the Rayleigh criterion, the lateral resolution of the synthetic aperture focusing is:

[0028]

[0029] in, The effective length of the synthetic aperture is used to significantly improve the resolution;

[0030] S24 employs a phase cancellation algorithm to suppress echoes from the metal insert;

[0031] The S25 uses spatiotemporal filtering to suppress scattering noise.

[0032] Furthermore, the method for suppressing echoes from metal inserts using a phase cancellation algorithm is as follows:

[0033] Preset metal insert position The echo phase is calculated as follows:

[0034]

[0035] in, The distance from the insert surface to the probe;

[0036] Inverted signals are superimposed during signal synthesis. To cancel out the echo from a fixed interference source and invert the signal. The calculation formula is as follows:

[0037]

[0038] in, This is the echo signal of the metal insert. It is an imaginary factor.

[0039] Furthermore, the method for suppressing scattering noise using spatiotemporal filtering is as follows:

[0040] Spatiotemporal filtering is performed on the signal of each array element. The formula for spatiotemporal filtering is as follows:

[0041]

[0042] in, These are the filter window coefficients. , For the filter scaling factor, the parameter and These are the x and y coordinates of the image, respectively. and All are counting factors. and Values integers, take It can effectively suppress random scattering noise while preserving the details of defect echoes.

[0043] Furthermore, the method described in step S3 for constructing a convolutional neural network to achieve hierarchical extraction of defect features and introducing Dropout to enhance generalization ability is as follows:

[0044] S31 multiplies the weight values ​​of the convolution kernel with the corresponding element values ​​of the feature map within the receptive field point by point and adds them together. The resulting matrix is ​​then combined with a bias matrix to form the output feature map of the convolutional layer.

[0045] S32 adds an activation function between two connected neurons to enhance the network's expressive power;

[0046] S33 samples the image using a pooling layer to reduce the height and width of the input feature map and output the feature map. as follows:

[0047]

[0048] in, The input feature map is a two-dimensional matrix; the pooling window size is [size missing]. Step size is , ; The height of the feature map, The width of the feature map. The height of the feature map after pooling operation. The width of the feature map after pooling.

[0049] S34 uses sub-pixel convolution to sample pixels in one channel, utilizing... The convolution operation reduces the number of channels in the original feature map from... Expand to And maintain and Unchanged, among which, The upsampling rate is used, and then the shape of the feature map is changed from [previous method] to [new method]. Transform into This achieves the goal of increasing the resolution of the feature map;

[0050] S35 transforms intermediate features into the final output through a fully connected layer, classifies information, and uses the Dropout algorithm to randomly discard information from some nodes during information transmission, thereby reducing the correlation between nodes and improving the network's generalization ability.

[0051] Furthermore, in step S4, the Transformer model is used to capture global features of defects through sinusoidal position encoding and multi-head self-attention mechanism, and feature fusion is completed through a feedforward neural network;

[0052] S41 adds a position encoding layer before the Transformer, using sinusoidal fixed position encoding, which assigns a unique value to each position through a combination of sine and cosine functions;

[0053] S42 uses a self-attention mechanism to compute the attention coefficients between each embedding vector and other vectors to model global dependencies;

[0054] In S43, the feedforward neural network in Transformer consists of two fully connected layers and a ReLU activation function to perform mapping and nonlinear transformations, thereby extracting richer defect features and completing feature fusion.

[0055] The present invention also provides an apparatus including a memory and a processor, the memory being used to store a program that supports the processor in executing the above-described ultrasonic flaw detection risk quantification method for GIS basin insulators, and the processor being configured to execute the program stored in the memory.

[0056] The present invention also provides a storage medium storing a computer program, which, when run by a processor, executes the steps of the above-described method for quantifying the risk of ultrasonic flaw detection in GIS basin insulators.

[0057] The beneficial effects of this invention are as follows:

[0058] This invention achieves precise coverage and anti-interference acquisition of ultrasonic beams by optimizing the parameters of a multi-element linear array and dynamic focusing rules; it utilizes Synthetic Aperture Focusing Technique (SAFT) to superimpose phase-compensated echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve detection resolution; it constructs a Convolutional Neural Network (CNN) to achieve hierarchical extraction of defect features through operations such as convolution, activation, and pooling, and introduces Dropout to enhance generalization ability; it employs a Transformer model to capture global defect features through sinusoidal position encoding and a multi-head self-attention mechanism, and completes feature fusion through a feedforward neural network; and it uses Otsu's method for dynamic thresholding. The method extracts the defect contour by segmentation, and achieves size, depth measurement and risk level assessment through pixel quantization and echo time calculation. This invention combines synthetic aperture focusing technology (SAFT), convolutional neural network (CNN) and Transformer model to achieve hierarchical extraction of defect features and global dependency analysis. At the same time, it completes the assessment of defect size, depth and risk level through dynamic threshold segmentation and multi-dimensional quantization algorithm, realizing the intelligent process of defect detection, feature extraction and risk quantification of GIS basin insulators. It is suitable for high-precision detection of hidden defects such as microcracks, air gaps and metal impurities inside basin insulators in gas-insulated switchgear. Attached Figure Description

[0059] Figure 1 This is a flowchart of the ultrasonic flaw detection risk quantification method for GIS basin-type insulators according to an embodiment of the present invention;

[0060] Figure 2 This is an ultrasonic echo signal image processed by synthetic aperture focusing technology in the ultrasonic flaw detection risk quantification method for GIS basin insulators according to an embodiment of the present invention.

[0061] Figure 3This is a CNN defect feature extraction result diagram of the ultrasonic flaw detection risk quantification method for GIS basin insulators according to an embodiment of the present invention;

[0062] Figure 4 This is a Transformer global feature enhancement result diagram of the ultrasonic flaw detection risk quantification method for GIS basin insulators according to an embodiment of the present invention;

[0063] Figure 5 This is the threshold segmentation result of the Otsu method for the risk quantification method of ultrasonic flaw detection for GIS basin insulators according to an embodiment of the present invention;

[0064] Figure 6 This is the defect segmentation and risk level assessment result of the ultrasonic flaw detection risk quantification method for GIS basin insulators according to an embodiment of the present invention. Detailed Implementation

[0065] 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 in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0067] Example 1

[0068] like Figure 1 As shown in the figure, an ultrasonic flaw detection risk quantification method for GIS basin-type insulators according to an embodiment of the present invention includes the following steps:

[0069] 1. A multi-element linear ultrasonic array is adopted. The focusing law of the multi-element linear ultrasonic array is calculated, the array excitation parameters are optimized, and the flaw detection scanning path of the multi-element linear ultrasonic array is planned to achieve accurate coverage and anti-interference acquisition of the ultrasonic beam.

[0070] 1.1 Design of Multi-Element Linear Ultrasonic Array

[0071] A multi-element linear ultrasonic array is used, with the number of elements being: Array element spacing satisfy ,in, For ultrasonic wavelengths, satisfying conditions, The speed of ultrasonic wave propagation is given by the center frequency. This avoids signal aliasing, and the effective detection width of the probe is... This covers the area of ​​GIS basin insulator flanges.

[0072] 1.2. Focusing law calculations are performed on the multi-element linear ultrasonic array to optimize the array excitation parameters.

[0073] Let the coordinates of the focal point be... , No. The coordinates of each array element are linear array along If the array element is distributed along an axis, then the excitation delay time is:

[0074] (1)

[0075] By delaying the excitation, the sound waves of each array element are superimposed in phase at the focal point to form a high-intensity sound beam.

[0076] 1.3 Planning the inspection path for multi-element linear ultrasonic array flaw detection

[0077] 1.3.1 The probe rotates circumferentially along the insulator while simultaneously advancing axially. This forms a spiral scanning trajectory, ensuring there are no blind spots in the detection.

[0078] 1.3.2. For the junction between the flange and the insulation body, which is a high-defect area, the step spacing is reduced to... Improve local resolution.

[0079] By optimizing the array element layout and excitation parameters, flexible focusing and scanning of the ultrasonic beam can be achieved, thereby improving the detection coverage capability of complex structures of basin insulators (such as flanges and inserts).

[0080] Step 2: The Synthetic Aperture Focusing (SAFT) technique is adopted, which involves moving the ultrasonic array probe to collect echo signals from multiple locations to simulate the effect of a large aperture array and overcome the limitation of the number of physical array elements. The synthetic aperture focusing technique performs phase compensation and superposition of echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve the resolution of the detection image.

[0081] 2.1 The ultrasonic array probe stops at multiple locations one by one, the number of locations being... Trigger all at each position Each element transmits or receives ultrasonic waves, forming an echo matrix. ,in, For scanning the location index, , For array element index, .

[0082] 2.2. Superposition of phase-compensated array element signals at each position:

[0083] (2)

[0084] in, For the focus point synthesized signal, The coordinates of the measuring point are... For the first The position The time delay corresponding to the path difference between each array element and the focal point.

[0085] 2.3, Array Elements in The sound waves emitted by the point reach the measured point and the returned distance value The calculation formula is as follows:

[0086] (3)

[0087] According to the speed of ultrasonic wave propagation The corresponding ultrasonic signal propagation time was obtained. :

[0088] (4)

[0089] The effective length of the synthetic aperture is defined as the length corresponding to the farthest position of the array element to which the sound field can radiate, and it is related to the half-power angle of the array element's radiation.

[0090] (5)

[0091] in, The effective length of the synthesized aperture. The wavelength of the sound wave. Distance to the target point The distance between array elements.

[0092] Within the effective length, the measured point The relative sound pressure intensity is the superposition amplitude of the radiated sound fields of each array element at this point:

[0093] (6)

[0094] Based on the Rayleigh criterion, the lateral resolution of the synthetic aperture focusing is:

[0095] (7)

[0096] in, The effective length of the synthetic aperture is used, resulting in a significant improvement in resolution.

[0097] 2.4. Phase cancellation algorithm for suppressing echo in metal inserts

[0098] Preset metal insert position The echo phase is calculated as follows:

[0099] (8)

[0100] in, This is the distance from the insert surface to the probe.

[0101] Inverted signals are superimposed during signal synthesis. As shown in formula (9), it cancels the echo of the fixed interference source.

[0102] (9)

[0103] in, This is the echo signal of the metal insert. It is an imaginary factor.

[0104] 2.5. Use spatiotemporal filtering to suppress scattering noise.

[0105] Spatiotemporal filtering is performed on the signal of each array element, as shown in equation (10):

[0106] (10)

[0107] in, These are the filter window coefficients. , For the filter scaling factor, the parameter and These are the x and y coordinates of the image, respectively. and All are counting factors. and Values integers, take It can effectively suppress random scattering noise while preserving the details of defect echoes.

[0108] Step 3: Construct a convolutional neural network (CNN) to extract defect features hierarchically through operations such as convolution, activation, and pooling, and introduce Dropout to enhance generalization ability.

[0109] 3.1 Multiply the weights of the convolution kernel point by point with the corresponding element values ​​of the feature map within the receptive field and add them together. Add a bias matrix to the resulting matrix to form the output feature map of the convolutional layer. The specific process of one convolution calculation is shown in formula (11):

[0110] (11)

[0111] in, To output the feature map of the th Line 1 The value of the column, , These are the width and height of the convolution kernel, respectively. These are the weights of the convolution kernel. This represents the element value at the corresponding position in the feature map. For the first Line 1 The bias value of the column, and All are counting factors. The value is integers, The value is Integers.

[0112] 3.2. Add activation functions such as Sigmoid (Equation (12)), Tanh (Equation (13)), ReLU (Equation (14)), and Leaky ReLU (Equation (15)) between two connected neurons to enhance the network's expressive power and thus better fit various curves.

[0113] (12)

[0114] (13)

[0115] (14)

[0116] (15)

[0117] in, This represents the gradient of the Leaky ReLU function on negative inputs.

[0118] 3.3. By sampling the image through pooling layers, the process of dimensionality reduction and abstraction of objects by the human visual system is simulated, thereby highly generalizing the input feature map and reducing its height. and width Output feature map :

[0119] (16)

[0120] in, The input feature map is a two-dimensional matrix; the pooling window size is [size missing]. Step size is , The height of the feature map, The width of the feature map. The height of the feature map after pooling operation. The width of the feature map after pooling. .

[0121] 3.4. Subpixel convolution is used to sample pixels in one channel, utilizing... The convolution operation reduces the number of channels in the original feature map from... Expand to And maintain and Unchanged, among which The upsampling rate is used, and then the shape of the feature map is changed from [previous method] to [new method]. Transform into This achieves the goal of increasing the resolution of the feature map.

[0122] 3.5. The intermediate features are transformed into the final output through fully connected layers to classify information. When there are many layers and little original data, overfitting is likely to occur. The Dropout algorithm is used to randomly discard information from some nodes during the information transmission process, thereby reducing the correlation between nodes and improving the network's generalization ability.

[0123] Step 4: Using the Transformer model, the global features of the defect are captured through sinusoidal positional encoding and multi-head self-attention mechanism, and feature fusion is completed through a feedforward neural network.

[0124] 4.1 Adding a position encoding layer before the Transformer can add position information to the vector. Sine fixed position encoding is used, and a unique value is assigned to each position by a combination of sine and cosine functions. The calculation process is shown in formula (17):

[0125] (17)

[0126] in, For the first The first patch Dimensional position encoding, The dimension of the embedding vector for the model.

[0127] 4.2 A self-attention mechanism is employed to calculate the attention coefficients between each model embedding vector and other vectors to model global dependencies, primarily using multi-head self-attention. Multi-head self-attention, building upon scaled dot product self-attention, divides the model into multiple heads, each learning a different attention score matrix in parallel across different subspaces. The number of heads is... Multi-head self-attention processes the input vector. The third linear transformation yields Each ,in , They are the first The query vector, key vector, and value vector of each size. and The shape is , The shape is , They are The dimension of the vector. For each Calculate the scaled dot product attention for each, and finally... The scaling dot product attention outputs of the heads are concatenated along the channels and subjected to a linear transformation. The final multi-head self-attention output is obtained. Despite this... Size, but due to the corresponding reduction in the number of heads... The dimension is not increased, so the computational cost of the network is not increased. The computation process of multi-head self-attention is shown in formulas (18) and (19).

[0128] (18)

[0129] (19)

[0130] in, For the first Attention calculation results based on body size To scale the dot product attention, and It is the size of The matrix, It is the size of The matrix, The overall computational output for multi-head attention. For splicing along the channel dimension, It is the size of The matrix, This represents the dimension of the final output vector.

[0131] 4.3 In Transformer, the feedforward neural network (FFN) generally follows the self-attention mechanism to perform mapping and nonlinear transformations, thereby extracting richer defect features. The FFN consists of two fully connected layers (FC) and a ReLU activation function, and the calculation process is shown in formula (20).

[0132] (20)

[0133] in, For the input vector, It is a bias vector with the same size as the input vector.

[0134] Step 5: Extract the defect contour based on Otsu's method dynamic threshold segmentation, and realize size and depth measurement and risk level assessment through pixel quantization and echo time calculation.

[0135] 5.1. The Otsu method is used to traverse the entire grayscale range, maximizing the variance of grayscale values ​​between parts C1 and C2 under the threshold, thus dividing the image into background and target regions for dynamic thresholding. The Otsu algorithm is calculated as follows:

[0136] (twenty one)

[0137] (twenty two)

[0138] (twenty three)

[0139] (twenty four)

[0140] (25)

[0141] (26)

[0142] Substituting equation (25) into equation (26), we get:

[0143] (27)

[0144] The inter-class variance is obtained by traversing the entire grayscale range. Maximum threshold This threshold is the threshold required for image segmentation; where, This indicates the percentage of the target image pixels in the total number of pixels in the entire image. This represents the average gray level of the pixels in the target image. This indicates the percentage of background image pixels out of the total number of pixels in the entire image. This represents the average gray level of the background image pixels. This represents the calculation of the total average gray level of all pixels in the entire image. Let the image size be denoted as the inter-class variance. , This indicates that the image grayscale value is less than the threshold. The number of pixels, This indicates that the image grayscale value is greater than the threshold. The number of pixels.

[0145] 5.2 Perform defect size quantification, depth measurement, and risk assessment.

[0146] Calculate the length and width of the defect. , ,in , This represents the number of defective pixels. , To achieve resolution, the defect depth is determined using ultrasonic echo signals. , Echo time The speed of ultrasonic wave propagation. (Based on different weights) Calculate the risk level ,in, The weighting of defect length. The weight of the defect width. The weight of defect depth is assigned; a risk assessment is conducted, which is divided into high, medium and low risk levels, and their locations are numbered to increase visibility.

[0147] Experimental Test

[0148] The data for this experiment came from the GIS equipment of a power supply bureau. These devices had varying lifespans, ranging from newly commissioned equipment to equipment that had been in operation for over 10 years. Ultrasonic testing technology was used to inspect the basin-type insulators in the GIS equipment, obtaining a large amount of ultrasonic echo signal data.

[0149] In the experiment, representative samples were selected from these actual detection data. To more comprehensively verify the algorithm's performance, various types and sizes of defects were tested. Simulated defects included microcracks, air gaps, and metallic impurities, with sizes ranging from tiny defects (such as air gaps with a radius of 8 pixels) to larger defects (such as cracks with a major axis of approximately 40 pixels). These simulated defect data, combined with the actual detection data, constituted the dataset for this experiment.

[0150] Professional ultrasonic testing equipment was used to inspect the basin-type insulators in the GIS equipment. During the inspection, multiple test points were arranged on the surface of the basin-type insulator, and a multi-element linear array scanning method was used to acquire ultrasonic echo signals at different locations, such as... Figure 2 As shown, by using synthetic aperture focusing technology, signals that have undergone time delay compensation are superimposed to enhance signals from the same target in phase. By processing signals from different locations, focused imaging of the entire detection area is achieved, thereby improving the resolution.

[0151] Then, CNN defect feature extraction is performed using the Canny edge detection algorithm. The closing operation structuring element is a disk with a radius of 1. The defect contour is located through edge detection, and the closing operation fills small gaps to improve edge continuity, resulting in the CNN defect feature extraction result, such as... Figure 3 As shown.

[0152] Based on the edge information extracted by CNN, pixels with edge detection results greater than 0.5 are enhanced by a factor of 1.5 to improve the contrast of defect areas, resulting in the Transformer global feature enhancement result, such as... Figure 4 As shown.

[0153] Using Otsu's method, the optimal segmentation threshold is determined based on the bimodal characteristics of the image's gray-level histogram, filtering out areas with an area less than 20 pixels, effectively separating defect areas from background noise. Figure 5 The image shows the threshold segmentation result using the Otsu method.

[0154] Extract the geometric features of the defects (area, aspect ratio, etc.), normalize the area and shape to obtain a normalization factor, calculate a comprehensive risk score based on the features, and classify the risk level, such as... Figure 6 As shown, this is the result of defect segmentation and risk level assessment. In the figure, #1 to #12 represent defect numbers, green represents low risk, orange represents medium risk, and red represents high risk, which intuitively reflects the degree of harm of the defects.

[0155] Example 2

[0156] An apparatus includes a memory and a processor, the memory being used to store a program that supports the processor in executing the ultrasonic flaw detection risk quantification method for GIS basin insulators according to Embodiment 1, the processor being configured to execute the program stored in the memory.

[0157] Example 3

[0158] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the ultrasonic flaw detection risk quantification method for GIS basin insulators in Embodiment 1.

[0159] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators, characterized in that, include: S1 employs a multi-element linear ultrasonic array, performs focusing law calculations on the multi-element linear ultrasonic array, optimizes array excitation parameters, and plans the flaw detection and scanning path of the multi-element linear ultrasonic array to achieve accurate coverage and anti-interference acquisition of the ultrasonic beam. S2 employs synthetic aperture focusing technology to superimpose phase-compensated echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve the resolution of the detection image. S3 constructs a convolutional neural network to achieve hierarchical extraction of defect features and introduces Dropout to enhance generalization ability; S4 uses the Transformer model, which captures global defect features through sinusoidal positional encoding and multi-head self-attention mechanism, and completes feature fusion through a feedforward neural network; S5 extracts defect contours based on Otsu's method of dynamic threshold segmentation, and achieves size and depth measurement and risk level assessment through pixel quantization and echo time calculation.

2. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 1, characterized in that, The number of elements in the multi-element linear ultrasonic array mentioned in step S1 is: Array element spacing satisfy ,in, For ultrasonic wavelengths, satisfying conditions, The speed of ultrasonic wave propagation is given by the center frequency. This avoids signal aliasing, and the effective detection width of the probe is... This covers the area of ​​GIS basin insulator flanges.

3. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 2, characterized in that, The method for calculating the focusing law and optimizing the array excitation parameters of the multi-element linear ultrasonic array described in step S1 is as follows: Let the coordinates of the focal point be... , No. The coordinates of each array element are linear array along If the array element is distributed along an axis, then the excitation delay time is: By delaying the excitation, the sound waves of each array element are superimposed in phase at the focal point to form a high-intensity sound beam.

4. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 1, characterized in that, The method described in step S2, which uses synthetic aperture focusing technology to superimpose phase-compensated echo signals from multiple locations, combined with phase cancellation and spatiotemporal filtering to suppress interference and improve the resolution of the detection image, is as follows: The S21 ultrasonic array probe stops at multiple locations point by point, the number of locations being... Trigger all at each position Each element transmits or receives ultrasonic waves, forming an echo matrix. ,in, For scanning location index, , For array element index, ; S22 superimposes the phase compensation signals of the array elements at each position; S23 computational array elements in The sound waves emitted by the point reach the measured point and the returned distance value According to the speed of ultrasonic wave propagation The corresponding ultrasonic signal propagation time was obtained. ; The effective length of the synthetic aperture is defined as the length corresponding to the farthest position of the array element to which the sound field can radiate, and it is related to the half-power angle of the array element's radiation. in, The effective length of the synthesized aperture. For the length of the sound wave, Distance to the target point The spacing between array elements; Within the effective length, the measured point The relative sound pressure intensity is the superposition amplitude of the radiated sound fields of each array element at this point: Based on the Rayleigh criterion, the lateral resolution of the synthetic aperture focusing is: in, The effective length of the synthetic aperture is used to significantly improve the resolution; S24 employs a phase cancellation algorithm to suppress echoes from the metal insert; S25 uses spatiotemporal filtering to suppress scattering noise.

5. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 4, characterized in that, The method for suppressing echoes from metal inserts using a phase cancellation algorithm is as follows: Preset metal insert position The echo phase is calculated as follows: in, The distance from the insert surface to the probe; Inverted signals are superimposed during signal synthesis. To cancel out the echo from a fixed interference source and invert the signal. The calculation formula is as follows: in, This is the echo signal of the metal insert. It is an imaginary factor.

6. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 4, characterized in that, The method for suppressing scattering noise using spatiotemporal filtering is as follows: Spatiotemporal filtering is performed on the signal of each array element. The formula for spatiotemporal filtering is as follows: in, These are the filter window coefficients. , For the filter scaling factor, the parameter and These are the x and y coordinates of the image, respectively. and All are counting factors. and Values integers, take .

7. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 1, characterized in that, The method described in step S3 for constructing a convolutional neural network to achieve hierarchical extraction of defect features and introducing Dropout to enhance generalization ability is as follows: S31 multiplies the weight values ​​of the convolution kernel with the corresponding element values ​​of the feature map within the receptive field point by point and adds them together. The resulting matrix is ​​then combined with a bias matrix to form the output feature map of the convolutional layer. S32 adds an activation function between two connected neurons to enhance the network's expressive power; S33 samples the image using a pooling layer to reduce the height and width of the input feature map and output the feature map. as follows: in, The input feature map is a two-dimensional matrix; the pooling window size is [size missing]. Step size is , ; The height of the feature map, The width of the feature map. The height of the feature map after pooling operation. The width of the feature map after pooling. S34 uses sub-pixel convolution to sample pixels in one channel, utilizing... The convolution operation reduces the number of channels in the original feature map from... Expand to And maintain and Unchanged, among which, The upsampling rate is used, and then the shape of the feature map is changed from [previous method] to [new method]. Transform into This achieves the goal of increasing the resolution of the feature map; S35 transforms intermediate features into the final output through a fully connected layer, classifies information, and uses the Dropout algorithm to randomly discard information from some nodes during information transmission, thereby reducing the correlation between nodes and improving the network's generalization ability.

8. The method for quantifying the risk of ultrasonic flaw detection in GIS basin-type insulators according to claim 1, characterized in that, The Transformer model described in step S4 captures global defect features through sinusoidal position encoding and multi-head self-attention mechanism, and completes feature fusion through a feedforward neural network; S41 adds a position encoding layer before the Transformer, using sinusoidal fixed position encoding, which assigns a unique value to each position through a combination of sine and cosine functions; S42 uses a self-attention mechanism to compute the attention coefficients between each embedding vector and other vectors to model global dependencies; In S43, the feedforward neural network in Transformer consists of two fully connected layers and a ReLU activation function to perform mapping and nonlinear transformations, thereby extracting richer defect features and completing feature fusion.

9. A device comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the ultrasonic flaw detection risk quantification method for GIS basin insulators according to any one of claims 1 to 8, and the processor is configured to execute the program stored in the memory.

10. A storage medium storing a computer program, characterized in that, When a computer program is run by a processor, it executes the steps of the ultrasonic flaw detection risk quantification method for GIS basin insulators as described in any one of claims 1 to 8.

Images