GIS basin-type insulator defect detection method, electronic equipment and storage medium

Abstract

The invention relates to a GIS (Geographic Information System) basin-type insulator defect detection method, electronic equipment and a storage medium, belongs to the technical field of intelligent image identification and processing of power equipment, and solves the problem of how to improve the accuracy of GIS basin-type insulator defect detection. The method effectively improves the resolution and information integrity of defect images, constructs a MobileNetV2 and YOLOv11n lightweight network structure for the technical bottlenecks of large FMC data volume and complex imaging calculation, achieves the intelligent closed-loop process from defect identification, positioning and classification to segmentation evaluation, achieves the automatic identification and geometric quantity extraction of defect areas, and improves the accuracy of the detection result. A data support is provided for subsequent early warning judgment and decision making; according to the method, the ultrasonic detection identification precision and the imaging quality are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent image recognition and processing technology for power equipment, specifically relating to a defect detection method for GIS basin-type insulators, electronic equipment, and storage medium. Background Technology

[0002] Currently, ultrasonic nondestructive testing is widely used in the condition assessment of industrial structural components. GIS basin insulators, as key equipment in high-voltage transmission lines, have highly concealed and complex internal defects (such as microcracks, delamination, and voids), making them difficult to effectively identify using conventional testing methods. These methods suffer from problems such as large blind spots, inaccurate positioning, and reliance on experience for image interpretation. Traditional ultrasonic testing methods often rely on manual image interpretation or single-channel A / B scan images, resulting in low information utilization and limited image resolution, making it difficult to accurately identify internal defects in complex structural components such as basin insulators. With the development of ultrasonic full matrix acquisition (FMC) and total focusing imaging (TFM) technologies, the acquisition method of ultrasonic data has shifted from single-point inspection to multi-channel full waveform acquisition, significantly improving imaging clarity and information dimensionality. FMC systematically acquires the fully interconnected signals between each array element, dividing the structure of the basin insulator into multiple fan-shaped regions during imaging. This achieves finer spatial resolution and defect focusing, effectively solving the positioning and identification problems of traditional methods in the inspection of complex-shaped components, and laying a solid data foundation for subsequent intelligent identification. Summary of the Invention

[0003] The technical solution of this invention is used to solve the problem of how to improve the accuracy of defect detection in GIS basin insulators.

[0004] The present invention solves the above-mentioned technical problems through the following technical solutions:

[0005] This invention provides a method for detecting defects in GIS basin-type insulators, comprising:

[0006] S1 arranges multiple ring probes around the GIS basin insulator, and uses the TFM algorithm to process the signals received by the ring probe array elements to obtain the FMC three-dimensional data matrix; the FMC three-dimensional data matrix is ​​converted into frequency domain form, and the frequency domain form of the FMC three-dimensional data matrix is ​​input into the embedded analysis platform;

[0007] S2 calculates the propagation time of each pixel, finds the value at the propagation time point in the FMC three-dimensional data matrix, converts the frequency domain form of the FMC three-dimensional data matrix into a two-dimensional image by delay summation, and stacks the two-dimensional image slices into a pseudo three-dimensional ultrasound image.

[0008] S3 uses the MobileNetV2 model to extract deep features of images and the YOLOv11n model for defect identification, localization and classification.

[0009] S4 transmits the analysis results to the manual monitoring center for verification and judgment. If there is no fault, it continues to operate; if there is a fault, it immediately terminates the operation and performs precise repairs on the GIS basin insulator based on the fault location identified by the scan.

[0010] Furthermore, the calculation formula for the FMC three-dimensional data matrix obtained by processing the signal received by the ring probe array elements using the TFM algorithm is as follows:

[0011] FMC(r i (t),e j (t),k ij (t))=s ij (t) (1)

[0012] Three parameters r i (t), e j (t) and k ij The following relationship exists between (t):

[0013]

[0014] Where, r i (t) represents the signal received by the i-th array element, e j (t) represents the signal emitted by the j-th element, k ij (t) is the impulse response function between the transmitting element j and the receiving element i of the array. ij (t) represents an element in the FMC three-dimensional data matrix, using mathematical notation. The convolution of the time-domain signal corresponds to the product of the frequency-domain signal, and N represents the number of ring probes.

[0015] Furthermore, the method for calculating the propagation time of a pixel is as follows:

[0016] For a given pixel r, calculate the distance from the i-th emitter r in the array. i From the pixel to the j-th receiver r j The propagation time is as follows:

[0017]

[0018] Where, r i r j is the position of the transmitter and receiver, r is the position of the pixel to be imaged, c is the speed of sound in the propagation medium, and |.| represents the Euclidean distance.

[0019] Furthermore, the method for finding the value at the propagation time point in the FMC three-dimensional data matrix is ​​as follows:

[0020] Linear interpolation is used to estimate the sampled value corresponding to the propagation time. The specific formula is as follows:

[0021]

[0022] in, This is an estimate of the sampled value corresponding to the propagation time;

[0023] To implement the interpolation operation in equation (6), the propagation time T needs to be... ij (r) is mapped to the sampling point location in the discrete time series, since T ij (r) usually falls between two adjacent sampling points and is estimated using linear interpolation. Let t ′ =T ij (r), if t ′ If the value falls between the nth and (n+1)th sampling points, then the interpolation calculation of equation (6) is transformed into the following form:

[0024] s(t ′ )≈(1-α)·s[n]+α·s[n+1](7)

[0025] Wherein s(t) ′ ) represents the signal value to be estimated at a non-integer time point t. ′ The sampled values ​​are s[n], s[n+1], and α, where s[n] is the value of the original signal at the nth sampling point, s[n+1] is the value of the original signal at the (n+1)th sampling point, and α is the interpolation weight.

[0026] Furthermore, the method for converting the frequency domain form of the FMC three-dimensional data matrix into a two-dimensional image through delayed summation is as follows:

[0027] (1) Calculate the gray value of pixel r in the final image. The gray value is the sum of the interpolations corresponding to all channel combinations. The specific calculation formula is as follows:

[0028]

[0029] Where I(r) is the gray value of pixel r;

[0030] (2) After imaging is achieved, time-division windowing, wavefront correction and dynamic aperture control are used to improve imaging quality and suppress noise in TFM;

[0031] Time-division windowing: A windowing function related to the propagation time is introduced for each pixel, as shown in the following expression:

[0032]

[0033] Among them, w t (T ij (r)) is a Gaussian window. T0 represents the propagation time corresponding to the target depth to be enhanced, and σ represents the width of the control window.

[0034] Wavefront correction: Replaces the linear propagation assumption with a non-linear path, calculates a more realistic propagation time, enhances focusing accuracy, and reduces artifacts. The propagation time formula after wavefront correction is:

[0035] T ij (r)=τ i (r)+τ j (r) (10)

[0036] Where, τ i (r) and τ j (r) represents the propagation time from the i-th and j-th probes to pixel r, respectively;

[0037] Dynamic aperture control: At each pixel, a channel selection mask function is designed to select only transmitter-receiver combinations that satisfy geometric conditions for delay summation. The expression for the channel selection mask function is as follows:

[0038]

[0039] So:

[0040]

[0041] Where, χ ij (r) is the channel selection mask function.

[0042] Furthermore, the method for stacking two-dimensional image slices into a pseudo-three-dimensional ultrasound image is as follows:

[0043] Multiple 2D image slices are acquired in one direction, and then the 2D image slices are stacked according to the spacing to form a pseudo 3D data volume. The specific expression is as follows:

[0044] I pseudo-3D (x,y,z k ) = I k (x,y) (13)

[0045] Among them, I pseudo-3D (x,y,z k ) represents the position (x, y, z) in three-dimensional space. k The image gray value at () is I k (x,y) is a two-dimensional image at the k-th layer, z k =z0+k·Δz is the center position of each slice, Δz is the interlayer spacing, and k is the slice number;

[0046] All I k(x, y) are all summed by TFM delay at a fixed z. k The imaging results were obtained for each z layer. k Repeat the TFM imaging process, the specific expression is as follows:

[0047]

[0048] Then all I k (x, y) can be stacked into a three-dimensional array, as shown in the following expression:

[0049] I pseudo-3D (x,y,z)={I0(x,y),I1(x,y),…,I k-1 (x,y)} (15)

[0050] Where N is the number of probe array elements (total number of channels), I k (x,y) represents the imaging layer z. k Above, the pixel intensity value (grayscale value) at the horizontal coordinate (x, y), (x i ,y i ,z i ) represents the position (spatial coordinates) of the i-th transmitting element, I pseudo-3d (x, y, z) represents the position (x, y, z) in three-dimensional space. k The image grayscale value (voxel value) at ().

[0051] Furthermore, the method for extracting deep image features using the MobileNetV2 model is as follows:

[0052] The MobileNetV2 lightweight convolutional neural network is used to extract deep features from images. The feature extraction formula is as follows:

[0053] F l =ReLU6(BN(W l *F l-1 (16)

[0054] Among them, F l F represents the feature map output by the l-th layer. l-1 W represents the feature map output by the (l-1)th layer. l This represents the set of parameters for all convolutional kernels in layer l, ReLU6() represents the activation function, BN represents the normalization of data within the mini-batch in each layer of the network, and * represents the convolution operation.

[0055] Furthermore, the method for defect identification, localization, and classification using the YOLOv11n model is as follows:

[0056] The YOLOv11n model was used to detect the locations of all defect areas from the input ultrasound images and to make a preliminary judgment on whether they were defects. The output of the YOLOv11n model is as follows:

[0057]

[0058] Among them, Box ij Let (i,j) be a bounding box predicted by the object detection model and its related information. This is the predicted value of the regional center point. This is the predicted value for the region width. This is the predicted height for the region. For fault confidence, For class probabilities;

[0059] The formula for calculating the accuracy of defect identification is as follows:

[0060]

[0061] in, Cross-entropy, used to measure classification accuracy, is y k The true label for one-hot encoding C represents the probability of belonging to the k-th class predicted by the YOLOv11n model, where C is the total number of class cases.

[0062] The formula for calculating positioning accuracy is as follows:

[0063]

[0064] in, is the bounding box regression loss, used to measure the positional difference between the predicted box and the ground truth box. (x,y) is the ground truth value of the region center point, w is the ground truth value of the region width, h is the ground truth value of the region height, and b is the normalized center coordinates and width and height of the ground truth box. is the corresponding predicted value of the bounding box, and ||·||1 is the L1 norm.

[0065] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the above-described GIS basin insulator defect detection method, and the processor is configured to execute the program stored in the memory.

[0066] 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 GIS basin insulator defect detection method.

[0067] The beneficial effects of this invention are as follows:

[0068] The method of this invention effectively improves the resolution and information integrity of defect images by introducing FMC three-dimensional data matrix data and TFM imaging algorithm. At the same time, in response to the technical bottlenecks of large FMC data volume and complex imaging calculations, this invention constructs a lightweight network structure of MobileNetV2 and YOLOv11n, realizing an intelligent closed-loop process from defect identification, localization, classification to segmentation evaluation, and automatically identifying and extracting geometric quantities of defect areas, providing data support for subsequent early warning judgment and decision-making. The method of this invention effectively improves the accuracy of ultrasonic detection and identification and imaging quality. Attached Figure Description

[0069] Figure 1 This is a flowchart of a GIS basin-type insulator defect detection method according to an embodiment of the present invention;

[0070] Figure 2 The pseudo-three-dimensional ultrasonic image of the GIS basin insulator reconstructed by the defect detection method for GIS basin insulators in this embodiment of the invention;

[0071] Figure 3 This diagram illustrates the training process of the MobileNetV2 model for the GIS basin insulator defect detection method according to an embodiment of the present invention.

[0072] Figure 4 This is a graph showing the performance test results of the trained MobileNetV2 model for the GIS basin insulator defect detection method according to an embodiment of the present invention.

[0073] Figure 5 This diagram illustrates the training process of the YOLOv11n model for the GIS basin-type insulator defect detection method according to an embodiment of the present invention.

[0074] Figure 6 This is a graph showing the performance test results of the trained YOLOv11n model for the GIS basin insulator defect detection method according to an embodiment of the present invention. Detailed Implementation

[0075] 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.

[0076] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0077] Example 1

[0078] likeFigure 1 As shown, the defect detection method for GIS basin-type insulators based on full-focus imaging and deep learning recognition according to an embodiment of the present invention includes the following steps:

[0079] Step 1: Arrange N ring probes around the GIS basin insulator. Use the TFM algorithm to process the signals received by the ring probe array elements to obtain the FMC three-dimensional data matrix. Convert the FMC three-dimensional data matrix into frequency domain form and input the frequency domain form of the FMC three-dimensional data matrix into the embedded analysis platform.

[0080] The formula for calculating the FMC three-dimensional data matrix by processing the signal received by the ring probe array elements using the TFM algorithm is as follows:

[0081] FMC(r i (t),e j (t),k ij (t))=s ij (t) (1)

[0082] Where, r i (t) represents the signal received by the i-th array element, e j (t) represents the signal emitted by the j-th element, k ij (t) is the impulse response function between the transmitting element j and the receiving element i of the array. ij (t) represents an element in the FMC three-dimensional data matrix (representing the sampling propagation channel formed by transmitting through the j-th channel and receiving through the j-th channel).

[0083] Three parameters r i (t), e j (t) and k ij The following relationship exists between (t):

[0084]

[0085] Among them, mathematical symbols The convolution of the time-domain signal corresponds to the product of the frequency-domain signal, and N represents the number of ring probes.

[0086] The method for converting the FMC three-dimensional data matrix into frequency domain form is as follows:

[0087] According to the Fourier transform, the frequency domain form of equation (2) is as follows:

[0088]

[0089] Equation (3) can be expressed in matrix form as follows:

[0090]

[0091] Among them, K ji (ω) is the frequency domain form of the impulse response function between the transmitting element j and the receiving element i of the array, R i (ω) represents the frequency domain form of the signal received by the i-th array element, E j (ω) represents the frequency domain form of the signal emitted by the j-th array element. Then, in the frequency domain form of the FMC three-dimensional data matrix, the three-dimensional element S... ij (ω)=(R i (ω),E j (ω),K ji (ω)).

[0092] Step 2: Calculate the propagation time of each pixel, find the value at the propagation time point in the FMC 3D data matrix, convert the frequency domain FMC 3D data matrix into a 2D image by delay summation, and stack the 2D image slices to form a pseudo 3D ultrasound image.

[0093] Step 2.1: Calculate the propagation time of each pixel.

[0094] For a given pixel r, calculate the distance from the i-th emitter r in the array. i From the pixel to the j-th receiver r j The propagation time is as follows:

[0095]

[0096] Where, r i r j is the position of the transmitter and receiver, r is the position of the pixel to be imaged, c is the speed of sound in the propagation medium, and |.| represents the Euclidean distance.

[0097] Step 2.2: Find the value at the propagation time point in the FMC three-dimensional data matrix.

[0098] Element s in the FMC three-dimensional data matrix ij (t) (obtained by inverse Fourier transform of the input parameters) is obtained by discrete-time sampling, therefore the propagation time T ij (r) is generally not an integer sampling point, so interpolation is required. In this embodiment of the invention, linear interpolation is used to estimate the sampling value corresponding to the propagation time. The specific formula is as follows:

[0099]

[0100] in, This is an estimate of the sampled value corresponding to the propagation time.

[0101] To implement the interpolation operation in equation (6), the propagation time T needs to be... ij(r) is mapped to the sampling point location in the discrete time series, since T ij (r) usually falls between two adjacent sampling points and can be estimated using linear interpolation. Specifically, let t ′ =T ij (r), if t ′ If the value falls between the nth and (n+1)th sampling points, then the interpolation calculation of equation (6) can be transformed into the following form:

[0102] s(t ′ )≈(1-α)·s[n]+α·s[n+1](7)

[0103] Wherein s(t) ′ ) represents the signal value to be estimated at a non-integer time point t. ′ The sampled values ​​are s[n], s[n+1], and α, where s[n] is the value of the original signal at the nth sampling point, s[n+1] is the value of the original signal at the (n+1)th sampling point, and α is the interpolation weight.

[0104] Step 2.3: Convert the frequency domain form of the FMC three-dimensional data matrix into a two-dimensional image by delaying summation.

[0105] Step 2.3.1: Calculate the grayscale value of pixel r in the final image. The grayscale value is the sum of the interpolations corresponding to all channel combinations. The specific calculation formula is as follows:

[0106]

[0107] Where I(r) is the grayscale value of pixel r.

[0108] Step 2.3.2: After imaging is achieved, time-division windowing, wavefront correction, and dynamic aperture control can be used to improve imaging quality and suppress noise in TFM.

[0109] (1) Time-division windowing: A windowing function related to the propagation time is introduced for each pixel, usually a Gaussian window or a Hamming window, which can be regarded as a weighting of signals at different depths. It can suppress far-field noise, limit interference signals outside the window, and improve the imaging contrast of specific depth areas.

[0110] The specific expression for introducing a windowing function related to the propagation time for each pixel is as follows:

[0111]

[0112] Among them, w t (T ij (r)) is a Gaussian window. T0 represents the propagation time corresponding to the target depth to be enhanced, and σ represents the width of the control window.

[0113] (2) Wavefront Correction: This replaces the straight-line propagation assumption with a non-straight-line path, calculating a more realistic propagation time (based on ray tracing or wave simulation). It can handle non-uniform media (wave velocity variations) or non-flat interfaces, enhancing focusing accuracy and reducing artifacts. The propagation time formula after wavefront correction is:

[0114] T uh (r)=τ u (r)+τ j (r) (10)

[0115] Where, τ i (r) and τ j (r) represents the propagation time from the i-th and j-th probes to pixel r, respectively. This propagation path takes into account the wave velocity variation c(x): Where Γ i→r The true propagation path (ray path) can be obtained through numerical solutions of the eikonal equations or finite difference ray tracing.

[0116] (3) Dynamic aperture control: At each pixel, only transmitter-receiver combinations that meet geometric conditions (such as a certain angle range or distance range) are selected for delay summation. This can eliminate channels with poor signal-to-noise ratio, reduce far-field blur, and improve the local directional resolution of the image. A channel selection mask function needs to be defined:

[0117]

[0118] So:

[0119]

[0120] Where, χ ij (r) is the channel selection mask function.

[0121] Step 2.4: Stack the two-dimensional image slices to form a pseudo-three-dimensional ultrasound image.

[0122] Imaging on a two-dimensional plane (such as the xy plane) yields a two-dimensional image I(x,y) = I(r), while pseudo-three-dimensional ultrasound images are constructed by stacking multiple slices of two-dimensional images to create a "quasi-three-dimensional" visual effect. It is not a true volume reconstruction, but it can show a sense of depth or spatial structure.

[0123] Multiple two-dimensional image slices (e.g., the XY plane) are acquired along one direction (e.g., the Z-axis), and then the two-dimensional image slices are stacked according to the spacing to form a pseudo-three-dimensional data volume. The specific expression is as follows:

[0124] I pseudo-3D (x,y,z k ) = I k (x,y) (13)

[0125] Among them, I pseudo-3D (x,y,z k ) represents the position (x, y, z) in three-dimensional space. k The image grayscale value (voxel value) at point I k (x, y) represents the k-th layer (depth of z). k A two-dimensional image of z k =z0+k·Δz is the center position of each slice, Δz is the interlayer spacing, and k is the slice number.

[0126] All I k (x, y) are all summed by TFM delay at a fixed z. k The imaging results were obtained for each z layer. k Repeat the TFM imaging process, the specific expression is as follows:

[0127]

[0128] Then all I k (x, y) can be stacked into a three-dimensional array, as shown in the following expression:

[0129] I pseudo-3D (x,y,z)={I0(x,y),I1(x,y),…,I K-1 (x,y)} (15)

[0130] Where N is the number of probe array elements (total number of channels), I k (x,y) represents the imaging layer z. k Above, the pixel intensity value (grayscale value) at the horizontal coordinate (x, y), (x i ,y i ,z i ) represents the position (spatial coordinates) of the i-th transmitting element, I pseudo-3D (x, y, z) represents the position (x, y, z) in three-dimensional space. k The image grayscale value (voxel value) at ().

[0131] Step 3: Use the MobileNetV2 model to extract deep features of the image, and use the YOLOv11n model to identify, locate and classify defects, so as to realize an intelligent closed loop from defect identification → location → classification → fine evaluation, and send an alarm immediately when a fault is identified.

[0132] Step 3.1 Feature Extraction

[0133] The MobileNetV2 lightweight convolutional neural network is used to extract deep features from images. The feature extraction formula is as follows:

[0134] Fl =ReLU6(BN(W l *F l-1 (16)

[0135] Among them, F l F represents the feature map output by the l-th layer. l-1 W represents the feature map output by the (l-1)th layer. l It represents the set of parameters for all convolutional kernels in layer l, used to extract higher-level features from the feature map of the previous layer through deep learning. ( A single core used independently for each input channel. (This is channel blending implemented using 1×1 convolution), * represents the convolution operation, BN indicates normalization of the data within the mini-batch in each layer of the network, and ReLU6 is an activation function, ReLU6(x) = min(max(0,x),6).

[0136] Step 3.2: Defect Identification, Location, and Classification

[0137] The YOLOv11n model was used to detect the locations of all defect areas from the input ultrasound images and to make a preliminary judgment on whether they were defects. The output of the YOLOv11n model is as follows:

[0138]

[0139] Among them, Box ij Let (i,j) be a bounding box predicted by an object detection model (such as YOLO in this case) and its related information at position (i,j). This is the predicted value of the regional center point. This is the predicted value for the region width. This is the predicted height for the region. For fault confidence, The probability is the category probability (e.g., whether it is a defect, what kind of defect it is, etc. In the simplest case, C=1, which only distinguishes between defect and non-defect, i.e., defect identification).

[0140] The formula for calculating the accuracy of defect identification is as follows:

[0141]

[0142] in, Cross-entropy, used to measure classification accuracy, is y k The true label for one-hot encoding (1 for class k, 0 for others, representing class k). C represents the probability of belonging to the k-th class predicted by the YOLOv11n model, where C is the total number of class cases.

[0143] The formula for calculating positioning accuracy is as follows:

[0144]

[0145] in, is the bounding box regression loss, used to measure the positional difference between the predicted box and the ground truth box. (x,y) is the ground truth value of the region center point, w is the ground truth value of the region width, h is the ground truth value of the region height, and b is the normalized center coordinates and width and height of the ground truth box. is the corresponding predicted value of the bounding box, and ||·||1 is the L1 norm (absolute difference).

[0146] Step 4: Transmit the analysis results to the manual supervision center for verification and judgment. If there is no fault, continue operation; if there is a fault, stop operation immediately and perform precise repair on the GIS basin insulator based on the fault location obtained from the scan.

[0147] Test

[0148] The dataset used scanned images of basin-type insulators within the GIS equipment of a power company in a certain city. The images were acquired and reconstructed using the method described in this invention to create pseudo-3D images of the basin-type insulators. This dataset covers basin-type insulators used over a six-year period from January 1, 2019 to December 31, 2024, including both in-use and faulty basin-type insulators awaiting repair. 80% of the data was used to train MobileNetV2 and YOLOv11n models, and the remaining 20% ​​was used for evaluation.

[0149] The pseudo-three-dimensional image of the basin-type insulator acquired and reconstructed using the method of this invention is shown below. Figure 2 As shown, the reconstructed image has a good pseudo-3D effect and can clearly show the fault condition of the basin insulator.

[0150] Through cross-validation and evaluation, the optimal hyperparameter configuration of the obtained MobileNetV2 model is: a Dense layer size of 256 and a Dropout value of 0.5.

[0151] The trained model is used to identify faults in basin-type insulators. Based on experimental results, the method of this invention demonstrates high accuracy and good adaptability. Specific evaluation results are as follows:

[0152] Figure 3 The training process of the MobileNetV2 model configured with the above parameters shows that the model's recognition accuracy stabilizes at over 95% and approaches 100% at the end of the training period, while the loss decreases to close to 0. Figure 4The confusion matrix shows the recognition results of the trained MobileNetV2 model on the evaluation set. It can be seen that the model's accuracy in judging lossless / clean and lossy cases has reached 100%.

[0153] Similarly, through cross-validation and evaluation, the optimal hyperparameter configuration of the YOLOv11n model is as follows: 100 total training epochs, 0.937 momentum factor, 3 warmup epochs, 0.1 warmup bias learning rate, 1.5 distribution focus loss weights (dfl), 7.5 weights of the bounding box loss component in the loss function, and 4 segmentation mask downsampling rate.

[0154] The trained model is used to identify and locate faults in basin-type insulators. Based on experimental results, the method of this invention demonstrates high accuracy and good adaptability. Specific evaluation results are as follows:

[0155] Figure 5 The first row lists the evaluation metrics for the training set, from left to right:

[0156] 1) Box loss (train / box_loss): decreased from 2.8 to about 1.7, with a steady decrease, indicating that the model has learned to accurately locate the target boundary;

[0157] 2) Classification loss (train / cls_loss): It decreased from 4 to below 1, indicating a significant improvement in classification ability and rapid convergence of error;

[0158] 3) Distributed focus loss (train / dfl_loss): Decreases steadily, converging to approximately 1.1, reflecting the improvement in bounding box quality and robustness;

[0159] 4) Precision (metrics / precision(B)): It rapidly increased from 0.2 to >0.95, indicating that the model can distinguish defect areas very well;

[0160] 5) Recall rate (metrics / recall(B)): It rises to 0.95+ within dozens of rounds, indicating that almost all defects are detected quickly;

[0161] The second row lists the validation set evaluation metrics, from left to right:

[0162] 1) Box loss (val / box_loss): It fluctuates but decreases overall, and there is no obvious overfitting, indicating good generalization;

[0163] 2) Classification loss (val / cls_loss): It decreased from 4 to around 1, indicating that the performance on the validation set was consistent with that on the training set;

[0164] 3) DFL loss (val / dfl_loss): It decreased significantly and fluctuated little, indicating that the model was also very accurate in localization on the validation set;

[0165] 4) Average detection accuracy (metrics / mAP50(B)) with IoU ≥ 0.5: It remains stable at around 0.95, which is close to that of mainstream state-of-the-art models;

[0166] 5) Average accuracy of IoU from 0.5 to 0.95 (metrics / mAP50-95(B)): reached 0.4 and increased steadily, indicating that the model has strong multi-scale detection capability and good robustness.

[0167] Overall, the model exhibits rapid and stable decreases in all loss functions during training, with consistent performance on both the training and validation sets, showing no signs of overfitting. Furthermore, in terms of detection performance, precision, recall, and mAP all rapidly approach saturation values ​​(>95%), with mAP@0.5 reaching close to 1 and mAP@0.5:0.95 reaching 0.4. This clearly demonstrates the model's extremely high accuracy and robustness in defect detection tasks on simulated images, making it valuable for engineering deployment.

[0168] Figure 6 The performance evaluation results of the trained YOLOv11n model demonstrate its excellent relevance: First, the model exhibits strong recognition and classification capabilities, accurately distinguishing fault types. As shown in the figure, the labels "crack" and "rand" are clearly marked with high positional accuracy and very few false positives or false negatives, fully demonstrating the model's strong generalization ability across multiple fault modes. Second, the model demonstrates high localization accuracy, with most detection boxes closely fitting the defect region boundaries, and the probability confidence levels are generally between 0.6 and 0.9, indicating that the model has a high level of confidence and stability in identifying fault regions. Finally, the model demonstrates excellent multi-target detection capabilities. Even when multiple defect targets appear in the image, the model can accurately label them one by one without overlap or confusion, showcasing good spatial perception and multi-target processing capabilities.

[0169] Based on the above research results and evaluation analysis, this system, combining TFM-FMC pseudo-3D imaging technology with deep learning target detection and recognition algorithms (MobileNetV2 and YOLOv11n), demonstrates extremely high accuracy, robustness, and engineering practical value in the task of identifying and locating defects in basin-type insulators. The system can not only accurately reconstruct pseudo-3D images of the insulator structure but also accurately distinguish multiple fault types and locate their specific positions. It exhibits strong adaptability, a low false detection rate, and extremely high automated intelligent diagnostic capabilities. The model has reached industrial application levels in several key indicators (such as accuracy, recall, and mAP), especially achieving nearly 100% recognition accuracy in identifying complex defects such as non-destructive testing and cracks. This fully verifies the system's promotional value and deployment feasibility in scenarios such as intelligent inspection, condition assessment, and preventive maintenance of power equipment.

[0170] This invention provides a method applicable to real-time inspection and condition monitoring of industrial structural components. By employing FMC data acquisition and combining it with Total Focusing Imaging (TFM) algorithm to construct high-resolution ultrasonic images, and further integrating lightweight deep learning models such as MobileNetV2 and YOLOv11n, it achieves end-to-end closed-loop processing from defect image generation, automatic identification, classification, contour segmentation, and geometric evaluation. This effectively improves the imaging quality, recognition accuracy, and field deployment performance of the ultrasonic inspection system, providing efficient, reliable, and interpretable technical support for industrial safety operation and maintenance. Compared with existing ultrasonic inspection technologies, this invention, by introducing FMC full matrix data and TFM imaging algorithm, effectively improves the resolution and information integrity of defect images, overcoming the problems of blurred imaging and easy missed detection in traditional single-channel acquisition. Simultaneously, addressing the technical bottlenecks of large FMC data volume and complex imaging calculations, this invention constructs an end-to-end data stream adapted to a lightweight network structure, significantly reducing the overall computational burden, enabling both high-quality imaging and rapid processing, meeting the dual demands of high precision and high efficiency in industrial settings. Furthermore, this invention integrates lightweight neural network models such as MobileNetV2 and YOLOv11n to construct an intelligent closed-loop process from defect identification, localization, classification to segmentation and evaluation. This enables automatic identification and geometric extraction of defect areas, providing data support for subsequent early warning judgments and decisions. The system possesses good deployability and interpretability, can be embedded into edge computing platforms such as Zynq, supports real-time detection and remote alarms, and has strong engineering practical value and promising prospects for widespread application.

[0171] Example 2

[0172] An electronic device includes a memory and a processor, the memory being used to store a program that supports the processor in executing the GIS basin insulator defect detection method based on total focusing imaging and deep learning recognition as described in Embodiment 1, the processor being configured to execute the program stored in the memory.

[0173] Example 3

[0174] A storage medium storing a computer program, which, when executed by a processor, performs the steps of the GIS basin insulator defect detection method based on total focusing imaging and deep learning recognition in Embodiment 1.

[0175] 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 detecting defects in GIS basin-type insulators, characterized in that, include: S1 arranges multiple ring probes around the GIS basin insulator, and uses the TFM algorithm to process the signals received by the ring probe array elements to obtain the FMC three-dimensional data matrix; the FMC three-dimensional data matrix is ​​converted into frequency domain form, and the frequency domain form of the FMC three-dimensional data matrix is ​​input into the embedded analysis platform; S2 calculates the propagation time of each pixel, finds the value at the propagation time point in the FMC three-dimensional data matrix, converts the frequency domain form of the FMC three-dimensional data matrix into a two-dimensional image by delay summation, and stacks the two-dimensional image slices into a pseudo three-dimensional ultrasound image. S3 uses the MobileNetV2 model to extract deep features of images and the YOLOv11n model for defect identification, localization and classification. S4 transmits the analysis results to the manual monitoring center for verification and judgment. If there is no fault, it continues to operate; if there is a fault, it immediately terminates the operation and performs precise repairs on the GIS basin insulator based on the fault location identified by the scan.

2. The method for detecting defects in GIS basin-type insulators according to claim 1, characterized in that, The formula for calculating the FMC three-dimensional data matrix by processing the signal received by the ring probe array elements using the TFM algorithm is as follows: FMC(r i (t),e j (t),k ij (t))=s ij (t) (1) Three parameters r i (t), e j (t) and k ij The following relationship exists between (t): Where, r i (t) represents the signal received by the i-th array element, e j (t) represents the signal emitted by the j-th element, k ij (t) is the impulse response function between the transmitting element j and the receiving element i of the array. ij (t) represents an element in the FMC three-dimensional data matrix, using mathematical notation. The convolution of the time-domain signal corresponds to the product of the frequency-domain signal, and N represents the number of ring probes.

3. The defect detection method for GIS basin-type insulators according to claim 2, characterized in that, The method for calculating the propagation time of a pixel is as follows: For a given pixel r, calculate the distance from the i-th emitter r in the array. i From the pixel to the j-th receiver r j The propagation time is as follows: Where, r i r j is the position of the transmitter and receiver, r is the position of the pixel to be imaged, c is the speed of sound in the propagation medium, and |.| represents the Euclidean distance.

4. The defect detection method for GIS basin-type insulators according to claim 3, characterized in that, The method for finding the value at a propagation time point in the FMC three-dimensional data matrix is ​​as follows: Linear interpolation is used to estimate the sampled value corresponding to the propagation time. The specific formula is as follows: in, This is an estimate of the sampled value corresponding to the propagation time; To implement the interpolation operation in equation (6), the propagation time T needs to be... ij (r) is mapped to the sampling point location in the discrete time series, since T ij (r) usually falls between two adjacent sampling points and is estimated using linear interpolation. Let t′=T ij (r), if t′ falls between the nth and (n+1)th sampling points, then the interpolation calculation of equation (6) is transformed into the following form: s(t′)≈(1-α)·s[n]+α·s[n+1](7) Where s(t′) is the sampled value of the signal to be estimated at a non-integer time point t′, s[n] is the value of the original signal at the nth sampling point, s[n+1] is the value of the original signal at the (n+1)th sampling point, and α is the interpolation weight.

5. The method for detecting defects in GIS basin-type insulators according to claim 4, characterized in that, The method for converting a frequency domain FMC three-dimensional data matrix into a two-dimensional image by delay summation is as follows: (1) Calculate the gray value of pixel r in the final image. The gray value is the sum of the interpolations corresponding to all channel combinations. The specific calculation formula is as follows: Where I(r) is the gray value of pixel r; (2) After imaging is achieved, time-division windowing, wavefront correction and dynamic aperture control are used to improve imaging quality and suppress noise in TFM; Time-division windowing: A windowing function related to the propagation time is introduced for each pixel, as shown in the following expression: Among them, w t (T ij (r)) is a Gaussian window. T0 represents the propagation time corresponding to the target depth to be enhanced, and σ represents the width of the control window. Wavefront correction: Replaces the linear propagation assumption with a non-linear path, calculates a more realistic propagation time, enhances focusing accuracy, and reduces artifacts. The propagation time formula after wavefront correction is: T ij (r)=τ i (r)+τ j (r) (10) Where, τ i (r) and τ j (r) represents the propagation time from the i-th and j-th probes to pixel r, respectively; Dynamic aperture control: At each pixel, a channel selection mask function is designed to select only transmitter-receiver combinations that satisfy geometric conditions for delay summation. The expression for the channel selection mask function is as follows: So: Where, χ ij (r) is the channel selection mask function.

6. The method for detecting defects in GIS basin-type insulators according to claim 5, characterized in that, The method for stacking two-dimensional image slices into a pseudo-three-dimensional ultrasound image is as follows: Multiple 2D image slices are acquired in one direction, and then the 2D image slices are stacked according to the spacing to form a pseudo 3D data volume. The specific expression is as follows: I pseudo-3D (x,y,z k )=I k (x,y) (13) Among them, I pseudo-3D (x,y,z k ) represents the position (x, y, z) in three-dimensional space. k The image gray value at () is I k (x,y) is a two-dimensional image at the k-th layer, z k =z0+k·Δz is the center position of each slice, Δz is the interlayer spacing, and k is the slice number; All I k (x, y) are all summed by TFM delay at a fixed z. k The imaging results were obtained for each z layer. k Repeat the TFM imaging process, the specific expression is as follows: Then all I k (x, y) can be stacked into a three-dimensional array, as shown in the following expression: I pseudo-3D (x,y,z)={I0(x,y),I1(x,y),…,I K-1 (x,y)} (15) Where N is the number of probe array elements, I k (x,y) represents the imaging layer z. l Above, the pixel intensity value at the horizontal coordinate (x, y), (x i ,y i ,z i ) represents the position of the i-th transmitting element, I pseudo-3D (x, y, z) represents the position (x, y, z) in three-dimensional space. k The image grayscale value at ().

7. The method for detecting defects in GIS basin-type insulators according to claim 6, characterized in that, The method for extracting deep image features using the MobileNetV2 model is as follows: The MobileNetV2 lightweight convolutional neural network is used to extract deep features from images. The feature extraction formula is as follows: F l =ReLU6(BN(W l *F l-1 )) (16) Among them, F l F represents the feature map output by the l-th layer. l-1 W represents the feature map output by the (l-1)th layer. l This represents the set of parameters for all convolutional kernels in layer l, ReLU6() represents the activation function, BN represents the normalization of data within the mini-batch in each layer of the network, and * represents the convolution operation.

8. The method for detecting defects in GIS basin-type insulators according to claim 7, characterized in that, The method for defect identification, localization, and classification using the YOLOv11n model is as follows: The YOLOv11n model was used to detect the locations of all defect areas from the input ultrasound images and to make a preliminary judgment on whether they were defects. The output of the YOLOv11n model is as follows: Among them, Box ij Let (i,j) be a bounding box predicted by the object detection model and its related information. This is the predicted value of the regional center point. This is the predicted value for the region width. This is the predicted height for the region. For fault confidence, For class probabilities; The formula for calculating the accuracy of defect identification is as follows: in, Cross-entropy, used to measure classification accuracy, is y l The true label for one-hot encoding C represents the probability of belonging to the k-th class predicted by the YOLOv11n model, where C is the total number of class cases. The formula for calculating positioning accuracy is as follows: in, is the bounding box regression loss, used to measure the positional difference between the predicted box and the ground truth box. (x,y) is the ground truth value of the region center point, w is the ground truth value of the region width, h is the ground truth value of the region height, and b is the normalized center coordinates and width and height of the ground truth box. is the corresponding predicted value of the bounding box, and ||·||1 is the L1 norm.

9. An electronic 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 GIS basin insulator defect detection method 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, The computer program is executed by the processor to perform the steps of the GIS basin insulator defect detection method according to any one of claims 1 to 8.

Images