Fault Identification Method for Automated Verification Line of Current Transformers

Abstract

This invention relates to a fault identification method for automated current transformer calibration production lines, addressing the problem of rapidly and accurately identifying faults or anomalies in production lines. It falls under the fields of image processing and fault identification technology. The invention includes: acquiring images of the workstations under test on the production line; preprocessing the acquired images using spectral density analysis to eliminate or reduce white noise effects such as resistance thermal noise from testing equipment during image acquisition; constructing a dataset based on the preprocessed data; augmenting the dataset using an improved Cutmix method based on hybrid cropping; constructing a fault identification model for the secondary side terminal contact of the current transformer based on the YOLOv8 model; training the model using the dataset; and using the trained model for fault identification. This invention provides effective technical support for the development of industrial automated production line monitoring technology.

Description

Technical Field

[0001] This application relates to a fault identification method for an automated current transformer calibration production line, belonging to the fields of image processing and fault identification technology. Background Technology

[0002] Current transformers, as a special type of transformer used for current measurement and control, operate by utilizing the principle of electromagnetic induction to convert a large current on the primary side into a smaller current on the secondary side, supplying power to measurement, protection, and control systems. In power systems, current transformers not only perform current conversion and measurement and provide electrical isolation, but are also a crucial component of various protection and control devices, playing a vital role in ensuring the safe and stable operation of the power system.

[0003] With the rapid development of the power industry and the significant increase in the use of instrument transformers in recent years, the number of instrument transformers to be calibrated by metrology and verification departments has also increased. This development has prompted relevant departments to continuously increase investment and strive to adopt more advanced calibration methods and management tools to make instrument transformer calibration work more efficient and accurate. Currently, automated calibration line systems for current transformers can simultaneously calibrate the errors of multiple transformers, manage them by computer, and process data automatically. However, the high-efficiency continuous operation of the line also brings about high equipment operating intensity, and various faults or deficiencies in the corresponding line stations have begun to emerge. For example, the instrument transformer calibrator may fail to find the corresponding detection point due to the failure of the secondary crimping of the transformer under test, or the voltage output plug-in of the error calibration unit may malfunction, causing the instrument transformer calibrator to malfunction. Currently, in the widely used automated verification production line systems, the electrical control and electrical measurement systems have corresponding monitoring functions for basic fault modes within the system. However, in actual operation, sometimes the electrical control and electrical measurement systems perform their respective functions after a fault occurs, leading to insufficient monitoring of certain faults (such as tilting of the current transformer in the enclosure) or false detections. Compared to traditional manual verification methods, the biggest advantage of the automated verification production line for low-voltage current transformers in the field of electricity metering lies in its fully automated operation, which can improve verification efficiency to a certain extent and eliminate the influence of human factors. Summary of the Invention

[0004] To address the issue of how to more accurately and quickly identify production line faults or abnormalities, this application provides a method for identifying production line faults in automated current transformer calibration.

[0005] The present invention provides a fault identification method for an automated current transformer calibration production line, comprising: Step 1: Collect images of the workstations to be tested on the production line, preprocess the collected images, and construct a dataset based on the preprocessed data; Step 2: Augment the images in the dataset; Step 3: Train the instrument transformer secondary side terminal contact fault identification model using the dataset, and use the trained instrument transformer secondary side terminal contact fault identification model to identify faults.

[0006] Preferably, step 1 includes: Step 11: Select the time for video image extraction. t 0, t The initial value of 0 is t , Step 12, from time t Starting from 0, the extraction time interval is △ t of n Image; Step 13: Collect the data n Each image was converted from RGB color space to HSI color space to obtain the corresponding... H Quantity, S Quantity, I The components, each image has a spatial resolution of 1. M × N ; to each image I Each component is processed by bitmap segmentation to obtain eight bitmap images corresponding to each image, denoted as . , ( x , y ) represents the position coordinates of a pixel within the image. x =0, 1, 2, …, M -1, y =0, 1, 2, …, N -1, i =1,2, …, n ; Step 14: Calculate the coordinates of different pixel positions. exist i Take 1 to n Mean combination matrix ; Step 15: Calculate the position coordinates of each pixel. The difference between the values ​​and the corresponding mean combination matrix forms the difference combination matrix of the coordinates of each pixel. The difference combination matrix is ​​denoted as: and , Proceed to step 12; if The difference combination matrix is ​​denoted as: Proceed to step 16; Step 16, Calculation and The combination matrix of autocorrelation functions at two time points:

[0007] in, This represents a submatrix of the autocorrelation function combination matrix. E [.] represents the expected value of the object within the square brackets; Step 17: Perform a Discrete Fourier Transform on the autocorrelation function combination matrix to obtain the combination matrix of the corresponding spectral density functions:

[0008] in, f To reflect the digital frequency in the frequency domain corresponding to the time domain when the video image changes over time; Step 18: Based on the combination matrix of the spectral density function The sum of the spectral density function values ​​for each plane of the image is obtained. ,like n The corresponding image of the first j The sum of the spectral density function values ​​of each plane image If all are equal to or approximately equal to the same constant, then... I The first component j After filtering out individual plane images, the remaining plane images are recombined to obtain the noise-filtered image. I Component images; Step 19, H Quantity, S Components and noise-filtered I The component images are normalized and then converted into RGB images to complete the preprocessing.

[0009] Preferably, step 2 includes: Step 21: Select image A1 in the dataset as the master image of the augmentation sample. Randomly crop image A1 to obtain multiple rectangular regions. Scale down or up each region rectangle to obtain the specified rectangular region size, denoted as image A2. Step 22: Set up two mutually exclusive structural elements B1 and B2 images of the secondary side wiring bolt of the current transformer. For image A2, use the hit / miss algorithm to identify and locate the structural elements B1 and B2 images of the wiring bolt. Restore the identification and location results to obtain the image of the wiring bolt marked in image A1, denoted as image A3. Extract the rectangular area box where the wiring bolt is located in image A3. Step 23: Set up two mutually exclusive structural elements B1 and B2 images of the terminal block. For image A2, use the hit / miss algorithm to identify and locate the structural elements B1 and B2 images of the terminal block. Restore the identification and location results to obtain the image marked with the terminal block in image A1, denoted as image A4. Extract the rectangular area box where the terminal block is located in image A4. Step 24: Perform binary masking on the rectangular area of ​​the terminal block in image A4 to complete the area with zeros, and obtain the background image of the augmented image, denoted as image A5; Step 25: Blend the rectangular area containing the wiring bolt, the rectangular area containing the wiring terminal, and image A5 to obtain the augmented image.

[0010] As a preferred option, in step 22, based on the external assembly mode of the wiring bolt and the actual imaging form of the monitoring angle of the production line, typical appearances in the monitoring images are selected as structural element B1 images of the wiring bolt, and the complement of the structural element B1 images, which is a limited local rectangular area within the body of the wiring bolt, is set as the structural element B2 image of the wiring bolt. Based on the actual imaging shape of the rectangular area around the thread of the terminal block and the probe it is assembled with, and the monitoring angle of the production line, the typical appearance in the monitoring image is selected as the structural element B1 image of the terminal block. The complement of the structural element B1 image, which is a limited local rectangular area within the body of the terminal block, is set as the structural element B2 image of the terminal block.

[0011] As a preferred approach, the original precise matching method in the pixel value evaluation process is changed to a fuzzy matching method in the hit / miss algorithm. The degree of fuzzy matching is evaluated by measuring the intersection between the target recognition box and the B1 image reference box of the hit / missed pixels in the processed image, and the degree of overlap (IoU) of the structuring element B1 image reference box.

[0012] Preferably, step 25 includes: Based on the size and position of the rectangular area where the wiring bolt is located, mark the corresponding position in image A5 with a rectangular box of the same color. Based on the position and shape of the rectangular area frame where the wiring bolt is located, and combined with the crimping contact position and ratio of the wiring bolt and the wiring terminal, determine the corresponding position area frame to be covered by the wiring terminal in image A5. Cover the corresponding position area frame with the rectangular area frame where the wiring terminal is located extracted by S33, and mark the colored rectangle frame to complete the augmentation.

[0013] As a preferred embodiment, the transformer secondary side terminal contact fault identification model includes a feature extraction network layer, a feature fusion network layer, and a feature detection output layer connected in sequence. The feature extraction network layer consists of the following convolutional layers connected in sequence: Conv1, Conv2, SPD convolution1, C2f_MSDA convolution1, Conv3, SPD convolution2, and C2f_MSDA convolution1. Convolutions 6 and 4 (Conv), 3 (SPD), and 2 (C2f) 6, 5 Convolution Conv, 4 SPD Convolution, 2 C2f_MSDA, SPPF, compression and activation network module SE-Attention; The feature fusion network layer includes Upsampling (number 1), Concat (number 1), and C2f (number 1). 3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) Convolutions 3 and 6 (Conv), concatenation 3 (Concat), and C2f 3. Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4. 3; Upsampling (No. 1), Concat (No. 1), C2f (No. 1) 3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) 3. Top-down connection; 2. C2f The output of 3 is input to convolution 6 (Conv); Convolution 6 (Conv), Concatenation 3 (Concat), C2f 3 Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4. 3. Bottom-up connection; The output of the compression and activation network module SE-Attention is simultaneously input to Upsampling module 1 and Concat module 4. C2f No. 1 The output of 6 is concatenated to input of 2 using Concat; C2f No. 2 The output of 6 is concatenated to input of 1 using Concat; C2f No. 1 The output and input of 3 are concatenated to 3 using Concat; C2f No. 2 Output of 3, C2f of number 3 Output of 3, C2f of number 4 The output of 3 is simultaneously input to the feature detection output layer; the feature detection output layer outputs the recognition result, and the feature detection output layer is implemented using a decoupled head structure.

[0014] Preferably, SPD convolutions 1 through 4 all include SPD layers and non-stretch convolutional layers, and the processing includes: H The length of the input feature map, W The width of the input feature map, C 1 represents the number of channels in the input feature map. SPD layer uses block size S Perform a space-to-depth transformation operation, H × W × C The input feature map of 1 is mapped and segmented into Each feature map is then divided along the channel dimension. The sub-feature maps are spliced ​​and merged to obtain The intermediate feature map is obtained by Conv transformation through a non-stride convolutional layer. The feature map.

[0015] Preferably, both C2f_MSDA No. 1 and C2f_MSDA No. 2 are C2f models that add a multi-scale dilated attention mechanism (MSDA). In MSDA, the channel dimension of the feature map is divided into three heads, and different dilation rates are used within each head for different input image sizes. r The local region patch surrounding the query is sampled, self-attention is calculated, the self-attention calculated from different heads is concatenated, and input into a linear layer for feature aggregation to obtain aggregated semantic information at various scales.

[0016] Preferably, the processing of the compression and activation network module SE-Attention includes a compression step and an activation step: In the compression step, the spatial dimension of each channel of the input feature map is compressed into a scalar to obtain global information; In the activation step, two fully connected layers are used to predict the importance weights of global information for each channel, and the importance weights are nonlinearly transformed and normalized using the activation functions ReLU and Sigmoid to generate channel weights. These channel weights are then multiplied one-to-one with each channel of the input feature map to output the channel weights.

[0017] The beneficial effects of this application are as follows: This application uses images of the workstation under test on the production line as the monitoring object, and utilizes a fault identification model for the secondary side wiring terminals of current transformers based on a deep learning network for fault identification, achieving rapid identification of production line faults or abnormal conditions. Furthermore, this application uses spectral density analysis to achieve noise reduction of the production line video image data, and uses a data augmentation algorithm based on hybrid cropping to perform data augmentation, realizing automatic monitoring and identification of production line faults based on the YOLO deep learning model. This production line fault identification method integrates multiple disciplines such as testing and metrology technology and instruments, reliability fault analysis, computer science and technology, and intelligent science and technology to achieve automated and intelligent identification of production line faults in the automated verification of current transformers. This aligns with the current status, background, development trend, and general direction of the development of massive data and artificial intelligence testing technologies in the digital information age, and is conducive to the routine automatic verification of power metering equipment, etc. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the production line fault image recognition scheme of this application; Figure 2 This is a schematic diagram of the data augmentation algorithm based on hybrid cropping in this application; Figure 3 This is a schematic diagram illustrating the principle of the YOLOv8-MSS model in this application; Figure 4 This is a schematic diagram illustrating the principle of SPD convolution. Figure 5 A schematic diagram illustrating the principle of the Multiscale Hollow Attention (MSDA) mechanism; Figure 6 This is a schematic diagram illustrating the principle of compressing and stimulating the SE-Attention network module. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0021] The present application will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the application.

[0022] Currently, automated verification lines for low-voltage current transformers can achieve efficient and automated verification of metering low-voltage current transformers. These lines, based on the type specifications of metering low-voltage current transformers, have the secondary winding terminals located on the upper part of the transformer, facilitating automatic wiring of the secondary terminals using a crimping device. However, due to dimensional tolerances in low-voltage current transformers, there are inherent deviations in the stopping and stopping of the line and the crimping positioning of the mechanical device. Furthermore, the crimping is not as tight or reliable as manual wiring. Moreover, the secondary wiring impedance may affect the error measurement results of the transformer under test. Therefore, compared to manual verification, automated lines pose a greater risk to verification quality due to issues such as poor secondary terminal contact, necessitating effective quality monitoring.

[0023] This application's assembly line fault image recognition device first collects information from the corresponding workstations and the current transformers under test in the assembly line using a visual imaging device. Then, the collected video image information is transmitted to a host computer and its control device. Finally, the host computer uses control software to call the current transformer secondary side terminal contact fault recognition model to achieve rapid testing and analysis of faults or abnormal conditions. By testing the machine vision-based video monitoring information of the workstation where the current transformer under test is located, and combining this with the code value information corresponding to the faults of the monitoring devices, compartments, or control equipment set up at each workstation by the electrical control system, the control device can achieve the fusion of corresponding video and image information with electrical information such as system electrical control and electrical measurement. This effectively realizes multi-faceted monitoring and alarm of the assembly line's operating status, thereby enabling more comprehensive and accurate rapid identification of assembly line faults or abnormal conditions.

[0024] The fault identification method for the automated current transformer calibration production line of this embodiment includes: Step 1: Collect images of the workstations to be tested on the production line, preprocess the collected images, and construct a dataset based on the preprocessed data; Step 1 involves acquiring video image information from the production line based on machine vision, achieved through a vision imaging device. This device primarily includes a camera and its bracket, a light source and its controller, a host computer, and fault acquisition and detection software. Specifically, the camera used is a Hikvision DS-2CD6425FWD-L20 2MP split-type small network camera, and the light source is an industrial dome light source. During the inspection process, the current transformer under test is fixed on a tooling tray of the production line inspection platform. The dome light source is located above and to the side of the production line, with the camera positioned directly above it. The brightness of the dome light source is adjusted by the light source controller to provide auxiliary illumination for the current transformers on the production line and tooling tray.

[0025] After initial sample acquisition using visual imaging equipment, due to the diverse types of on-site monitoring data, OpenCV software was employed for video format conversion, including reading, writing, and format conversion. During conversion, the spatial resolution and frame rate of the video were adjusted, while simultaneously adjusting video image parameters according to the monitoring environment of the verification scenario. Considering the different storage and compression methods of video image formats such as AVI, MP4, and MPEG, it was necessary to convert them to a consistent format to minimize the computational resource consumption and processing speed reduction caused by directly processing large video surveillance images, thus facilitating the subsequent training and learning of deep learning models. Then, image preprocessing techniques were used to clean, denoise, and enhance the transformer pipeline data acquired by the information acquisition equipment. The automated calibration line system for instrument transformers contains numerous electrical control devices and electrical testing instruments. During operation, these different sensors and control equipment often generate a large amount of complex electromagnetic information, resulting in the inclusion of complex interference or noise in the images and video information collected by video monitoring equipment at various workstations. Extracting effective feature information characterizing line faults from this heterogeneous and overlapping interference information is quite challenging. While commonly used filtering, noise reduction, and signal enhancement techniques and methods, such as smoothing, differentiation, adaptive filtering, and wavelet transform, are employed, they simultaneously enhance or weaken relevant fault information in the image while processing noise signals. Therefore, to improve testing accuracy, step 1 of this implementation method specifically includes: (1) During the video image acquisition process of the workstation under test, the video image extraction time is selected as... t 0, t The initial value of 0 is t ; (2) From t Starting at time 0, extract time intervals Δ from the captured video images. t Image n Amplitude, as obtained from continuous repeated measurements that vary over time. n Images, with a spatial resolution of [number] images. M × N ; (3) Settings i =0; (4) Calculation i The result of adding 1 is assigned to i Then to n The first image i The images are converted from RGB color space to HSI color space to obtain the corresponding... H Quantity, S Quantity, I Quantity;

[0026]

[0027]

[0028]

[0029] (5) Regarding the results obtained in the previous step I The components are processed by bitmap segmentation to obtain 8 bit-plane images, denoted as . , ( x , y ) represents the position coordinates of a pixel within the image. x = 0, 1, 2,…, M -1, y = 0, 1, 2, …, N -1; (6)Judgment i Is it greater than or equal to? n If yes, then jump to (7); otherwise, jump to (4). (7) Calculate position coordinates x The values ​​are taken sequentially as 0, 1, 2, …, M -1 and y The values ​​are taken sequentially as 0, 1, 2, …, N The corresponding pixel points at different positions when -1 ( x , y The combination matrix of the corresponding images exist i Values ​​range from 1 to n Mean combination matrix , where each submatrix in the combination matrix and The matrix dimensions are all M × N ,and j = 0, 1, 2, …, 7; (8) Calculate position coordinates x The values ​​are taken sequentially as 0, 1, 2, …, M -1 and y The values ​​are taken sequentially as 0, 1, 2, …, N The corresponding pixels at different positions (time) x , y The combination matrix of the corresponding images exist i Values ​​range from 1 to n The mean combination matrix described in (7) is used respectively. The difference combination matrix; if ,

[0030] Each submatrix in the difference combination matrix The matrix dimensions are all M × N ,and j =0, 1, 2, …,7, jump to step (9); if The difference combination matrix is ​​denoted as: Each submatrix in the difference combination matrix The matrix dimensions are all M × N ,and j = 0, 1, 2, …,7, jump to step (10); (9) During the video image acquisition process of the workstation under test, the initial time for video image extraction is selected as... ,and Then jump to step (2); (10) Calculate position coordinates x The values ​​are taken sequentially as 0, 1, 2, …, M -1 and y The values ​​are taken sequentially as 0, 1, 2, …, N The corresponding pixels at different positions (time) x , y The autocorrelation function of the corresponding image is in i Values ​​range from 1 to n Combination matrix of time

[0031] in, This represents a submatrix of the autocorrelation function combination matrix. E The brackets [.] represent the expected value of the object within the square brackets. For ease of subsequent description, the autocorrelation function is used. Expressed in the form of a centered autocorrelation function as follows: The above autocorrelation function combination matrix is ​​simplified as follows:

[0032] (11) The autocorrelation function combination matrix Performing a discrete Fourier transform yields the combined matrix of the corresponding spectral density functions, denoted as .

[0033] The transformation formula adopts the Wiener-Khinchin formula, and the resulting combination matrix of the spectral density function is as follows:

[0034] Position coordinates in the formula x The values ​​are taken sequentially as 0, 1, 2, …, M -1, y The values ​​are taken sequentially as 0, 1, 2, …, N , f To reflect the digital frequency in the frequency domain corresponding to the time domain when the video image changes over time; (12) Combination matrix of the spectral density function Further processing yields eight bitmap images obtained from the segmentation of each image bitmap. , recorded as

[0035] In the formula, the combination matrix of the spectral density function submatrix in middle, i =0, 1, 2, …, n , representing the first in the dynamic test video image sequence. i The image number, j =0,1,2,…,7 represents the 7th bit in each of the 8 bitmaps obtained from bitmap cutting. j Individual plane images.

[0036] (13) Based on the characteristic that the spectral density function value of white noise changes with frequency as a constant, the combination matrix of the spectral density function is then processed sequentially. submatrix in Make statistical judgments. ,like n The corresponding image of the first j The sum of the spectral density function values ​​of each plane image If all are equal to or approximately equal to the same constant, then the first... j The image corresponding to each plane represents white noise information, then I The first part obtained by bitmap cutting of the component j The individual plane image is the white noise interference mixed during the image acquisition process, which is filtered out; (14) will I Bitmap obtained by processing components into bitmaps j The remaining planes are recombined to obtain the brightness after noise filtering. I Component images, further processing the above-obtained I Component images, and original H Component images SThe component images are normalized, and then the HSI image is converted to an RGB image model according to the following formula, thus completing the preprocessing.

[0037] when hour,

[0038] when hour,

[0039] when hour,

[0040] This implementation uses video surveillance equipment to collect video images from the production line, which are then transmitted to a host computer. After image enhancement preprocessing as described above, the images are input into a trained deep learning model for fault classification and identification. Finally, the host computer coordinates and controls the operation of each module to achieve rapid monitoring of production line faults. Since deep learning models require a large number of dataset samples to train their learning ability, insufficient dataset samples can lead to overfitting or failure to fit the model. In industrial control scenarios involving automated transformer calibration production lines, the data on contact faults at the secondary side terminals of transformers is limited, making it difficult to obtain a sufficient number of samples during the collection of contact fault image information. To improve the model's generalization ability, data augmentation is needed during dataset creation, in addition to filtering on-site data and prior data.

[0041] Step 2: Augment the images in the dataset: Conventional data augmentation methods often employ mirroring, rotation, cropping, color dithering, and noise addition to enrich data information by augmenting or expanding data samples at the spatial or feature domain level. However, for contact faults at the secondary side terminals of current transformers in automated low-voltage current transformer calibration lines, the background information presented in the video surveillance images of the calibration scene during the pressing process between the tooling tray and the transformer carried on it and the wiring assembly is complex and varied. Furthermore, the sample sizes for different scenarios—partial contact, side contact, and contact between the terminal and the transformer body—are insufficient and inconsistent, exhibiting a lack of diversity and balance. These issues, including insufficient data, variations in terminal contact patterns due to different video image acquisition angles, and uneven sample sizes for different fault images, limit the learning effectiveness of deep neural network models. The complexity of the caliber operation scene and the resulting background interference further increase the difficulty of identification. Therefore, generally applicable basic augmentation methods are insufficient to significantly improve the generalization ability of contact fault identification models. Therefore, this embodiment discloses an improved Cutmix method based on hybrid clipping to augment the dataset of the current transformer secondary side terminal contact fault identification model. The main idea stems from the fact that contact faults occur during the crimping contact process between the terminal of the wiring assembly and the secondary side wiring bolt of the current transformer. The wiring bolts and their embedded nuts are standard mechanical parts, and the specifications of the terminal contacts are also uniformly customized for the assembly line system. Therefore, based on the fact that both are of standard form, the mechanical structural features of the terminal and the current transformer wiring bolt in the video surveillance image of the assembly line scenario are clipped, and these are respectively made into structural element pairs (B1, B2) images based on an improved hit / miss algorithm using fuzzy recognition. By designing and applying a specially designed information probe method using structural elements, we can achieve the research goal of selectively cropping or extracting the feature information of two hosts (terminal ends and secondary side wiring bolts of current transformers) when terminal crimping faults occur in a production line scenario. This not only solves the problem of redundant information filtering caused by random cropping of feature rectangles in the commonly used Cutmix data augmentation method, but also significantly reduces the number of cropped images used for subsequent mixing. This improves the running efficiency of the Cutmix algorithm and enhances the diversity, balance, and generalization of the dataset's crimping faults.

[0042] This implementation method is based on an improved Cutmix dataset augmentation method using hybrid clipping. It utilizes the representational modalities of terminal blocks and bolts to generate corresponding structural element pairs (B1, B2) images in a morphological algorithm. A morphologically improved hit / miss method based on IoU evaluation is used to filter out terminal blocks and bolts that meet the clipping requirements. These clipped and mixed images are then superimposed onto different production line workstation backgrounds. This augmentation derives various sample images that meet the requirements, such as non-contact, partial contact, and lateral contact between the terminal blocks and the secondary side terminals of the transformer, thereby solving the aforementioned data imbalance problem of fault samples. A flowchart illustrating the improved Cutmix data augmentation method based on hybrid clipping in this implementation method is shown below. Figure 2 As shown, the corresponding implementation scheme is achieved through the following steps: (1) Select image A1, label it as... x 1. Use this as the master image for augmentation samples. First, randomly crop a rectangular region from A1, and denot the size of the region as Rectangle ( r x0 , r y0 , r w0 , r h0 ).

[0043] (2) Based on the size of the cut rectangular area frame, Rectangle ( r x0 , r y0 , r w0 , r h0 ), and the specified rectangular frame size (Rectangle) r x1 , r y1 , r w1 , r h1 The corresponding horizontal and vertical scaling ratios are determined by comparing the Rectangle ( r x0 , r y0 , r w0 , r h0 ) Perform several scale reductions or enlargements to the specified rectangular area size (Rectangle). r x1 , r y1 ,r w1 , r h1 (), is called image A2, and the label is denoted as x 2. During scaling, if the rectangular region bounding box (Rectangle) r x0 , r y0 , r w0 , r h0 If the target is medium-sized, the scale remains unchanged; if the target is large, the scale is reduced; if the target is small, the scale is enlarged. Scale enlargement requires subsequent image super-resolution reconstruction. In this embodiment, the image super-resolution reconstruction method uses the IDN (Information Distillation Network) network model algorithm, which has relatively low computational cost and excellent reconstruction performance.

[0044] (3) For the rectangular frame (Rectangle) obtained after scaling, r x1 , r y1 , r w1 , r h1 First, the mathematical morphology hit / miss algorithm is used to identify and locate the secondary side wiring bolt S2 of the current transformer and its surrounding rectangular area. The setting of the structural element pair (B1, B2) in the hit / miss algorithm is shown in [link to relevant documentation]. Figure 2 Images of structural elements B1 and B2 of the secondary side wiring bolt S2 of the current transformer. Then, scale restoration is performed according to the scaling ratio to obtain the image of the wiring bolt S2 marked in image A1, denoted as Rectangle ( r x2 , r y2 , r w2 , r h2 (A), denoted as image A3, extract the rectangular region containing the wiring bolt in image A3, and use it as the label, denoted as... x 3. The hit / miss algorithm in this implementation is an improved hit / miss method, and its operating mechanism and execution method are derived from the following two levels of design and consideration.

[0045] At the first level, the secondary side wiring bolts and the embedded nuts mounted on them of the low-voltage current transformer are standard parts of corresponding specifications. Therefore, based on the external assembly mode and the actual imaging form of the assembly line monitoring angle after their assembly, several typical appearances in the video monitoring image can be selected. The first structural element B1 image in the structural element pair (B1, B2) of the wiring bolt S2 in the hit / miss algorithm is designed. Furthermore, the image B1 is selected to complement a finite local rectangular area within the red body of the wiring bolt, and this is set as the second structural element B2 image in (B1, B2). In actual execution, a batch of representative structural element pairs (B1, B2) can be selected based on the various appearances of the wiring bolt S2 in the assembly line video monitoring image, so that they can be iteratively called sequentially during the execution of the improved hit / miss algorithm.

[0046] On the second level, in video surveillance images in assembly line scenarios, in response to the visual differences caused by inconsistent monitoring angles and the complex background scenes of the transformer under test in the images, the improved hit / miss algorithm disclosed in this application specifically changes the original precise matching method in the pixel value evaluation process to a fuzzy matching method when the structuring element pair (B1,B2) covers the area in the processed image during the attack.

[0047] The color image hit / miss algorithm based on fuzzy matching, and its methods and implementation schemes include: For the RGB image models of the processed image A and the structuring element image pair (B1, B2), perform HSI model conversion respectively; After converting the processed image A and the structuring element image pair (B1, B2) to obtain the corresponding I-component grayscale images, binarize them to obtain their respective binary images. Then, perform an improved hit / miss transform on the corresponding binary images. The steps are as follows: First, perform the improved hit / miss calculation using the following formula:

[0048] In the formula, for The complement of the structuring element pair (B1, B2) in the image. B1 For the target image, B2 for B1 background, ; In the improved hit / miss calculation, The exact matching method in the hit / miss process is changed to fuzzy matching, which creates a rectangular target recognition box containing the localized pixels obtained after the hit / miss in the processed image A. , recorded as The degree of fuzzy matching is measured by The intersection with the image reference box of structuring element B1 and the degree of overlap between the image reference boxes of structuring element B1 are used for evaluation. IoU (Intersection Over Union) is denoted as:

[0049] During fuzzy matching, the ratio of the number of pixels contained in the target bounding box to the number of pixels contained in the reference bounding box is used as an indicator of the degree of IoU overlap. If the IoU index value between the updated bounding box and the reference bounding box is within the set threshold, the fuzzy matching is successful, and the corresponding location and target recognition results of the color image can be derived from the recognition and localization results of the I-component binary image; otherwise, the fuzzy matching fails.

[0050] (4) For the rectangular frame (Rectangle) obtained after scaling, r x1 , r y1 , r w1 , r h1 Furthermore, an improved hit / miss algorithm is used to identify and locate the rectangular area surrounding the wiring terminals and the threads of the probes assembled with them in the assembly line. Then, scale restoration is performed according to the scaling ratio to obtain the image of the marked wiring terminals in image A1, denoted as Rectangle. r x3 , r y3 , r w3 , r h3 (A4), extract the rectangular area containing the terminal in image A4, and use it as the label, denoted as. x 4. Improved settings for the structure element pair (B1, B2) of the terminal block in the hit / miss algorithm. See [link / details]. Figure 2 The structural elements B1 and B2 images of the terminal block are used to select the typical appearance in the monitoring image as the structural element B1 image of the terminal block, based on the actual imaging shape of the rectangular area around the thread of the terminal block and the probe thread assembled with it, and the actual imaging shape of the monitoring angle of the assembly line. The complementary set of the structural element B1 image, which is a limited local rectangular area within the body of the terminal block, is set as the structural element B2 image of the terminal block.

[0051] (5) Based on the labels in image A4 x The size of the terminal of 4 is Rectangle ( r x3 , ry3 , r w3 , r h3 The recognition box is obtained by performing binary masking on the rectangular area of ​​the terminal block in image A4 to complete the setting of the built-in zero in the region, and the background image of the augmented image is denoted as image A5.

[0052] (6) Based on the target labels of images A3 and A4 x 3. x 4. Blend the cropped image label rectangle recognition boxes. The blending method needs to be determined based on the target of the augmentation, such as the terminal block recognition box. x 4. How to identify the wiring bolt recognition box in image A5 compared to that in image A3? x 3. Perform operations such as setting and covering at partial contact, side contact, and other locations: Taking the 50% crimped contact of the terminal as an example, the positioning recognition box Rectangle of the secondary side wiring bolt S2 of the current transformer in image A3 is used. r x2 , r y2 , r w2 , r h2 The size and position of the area are also marked with the same green rectangle in image A5. If the augmentation target is partial crimping contact between the lower 50% of the terminal and the upper edge of the connector bolt, then the size of the rectangular area of ​​the terminal can be calculated as Rectangle (based on the aforementioned position and shape data). r x4 , r y4 , r w4 , r h4 ), recorded as tags x 5.

[0053]

[0054] Based on the Rectangle (the area bounded by the wiring terminals in image A5) that needs to be covered. r x4 , r y4 , r w4 , r h4 (Use the labels on the terminals in image A4) x 4. Rectangle ( r x3 ,r y3 , r w3 , r h3 ) Cover Tag x 5. Rectangular region bounding box (Rectangle) r x4 , r y4 , r w4 , r h4 ), updated new tags x 5. Place the green label box next to the wire terminal to obtain the partial crimping augmentation image A6, which shows 50% of the crimped contact as the augmentation target. Similarly, other crimping augmentation state images, such as the side crimping augmentation image A7, can be obtained.

[0055] Before the overlay operation, first set a binary encoded mask matrix with the same size as image A5, and then perform a mask multiplication operation with image A5 to perform the Rectangle operation on the corresponding region bounding box to be covered in image A5. r x4 , r y4 , r w4 , r h4 The original information is cleared. The mask matrix is ​​Rectangled at its corresponding position. r x4 , r y4 , r w4 , r h4 Within the designated area, set the pixel value to 0; outside the area, set the pixel value to 1. This ensures that the terminal label in image A4 is used correctly. x 4. Rectangle ( r x3 , r y3 , r w3 , r h3 (Covering image A5 labels) x 5. Rectangular area bounding box (Rectangle) r x4 , r y4 , r w4 , r h4 When ), the terminal labelx 4. Rectangle ( r x3 , r y3 , r w3 , r h3 The information will not be interfered with or confused by the information in the original image.

[0056] After completing the selection, augmentation, partitioning and labeling of the dataset samples, the next step is to build and optimize the deep learning model for pipeline fault identification. Step 3: Construct a fault identification model for the secondary side terminals of the current transformer. Train the fault identification model for the secondary side terminals of the current transformer using the dataset. Use the trained fault identification model for the secondary side terminals of the current transformer to identify faults.

[0057] This application presents an improved YOLOv8 model for identifying secondary side terminal contact faults in instrument transformers, referred to as the YOLOv8-MSS model. Based on the YOLOv8 framework, this model employs the MSDA (Multi-Scale Dilated Spatial Attention) mechanism, Squeeze-and-Excitation (SE) compression and excitation network modules, and the SPD (Space-to-Depth Convolution) method to establish an improved YOLOv8-MSS model based on MSDA-SE-SPD. This MSDA-SE-SPD-based improvement not only significantly enhances the model's recognition performance, giving it a greater advantage in feature extraction, but also improves system computational efficiency, thereby achieving the research objective of greatly improving fault identification accuracy and robustness. Figure 3A schematic diagram of the YOLOv8-MSS model is provided. It consists of three parts: a feature extraction network layer (Backbone), a feature fusion network layer (Neck), and a feature detection output layer (Head). The Backbone layer, which is C2f, retains the CSP concept while achieving lightweight design and adapting to different channel counts. The C2f setting allows for skip connections and split operations, optimizes gradient flow branches, and improves model expressiveness and convergence efficiency. The Neck layer has two convolutional layers removed and is further optimized based on YOLOv5's PANet, enhancing the feature pyramid fusion capability and improving feature fusion and context capture capabilities. The Head layer uses a decoupled head structure, making classification and localization more focused and accurate. Anchor-based models are replaced with anchor-free models to reduce anchor box hyperparameter settings and simplify the training process by directly predicting the target's center point. In a specific implementation, based on the different configurations of the CSP module parameters, the model version selected is YOLOv8n, the most lightweight version in the YOLOv8 series, which has fewer parameters, lower computational complexity, and is suitable for real-time automated detection in industrial control production lines. Specifically, The feature extraction network layer consists of the following convolutional layers connected in sequence: Conv1, Conv2, SPD convolution1, C2f_MSDA convolution1, Conv3, SPD convolution2, and C2f_MSDA convolution1. Convolutions 6 and 4 (Conv), 3 (SPD), and 2 (C2f) 6, 5 Convolution Conv, 4 SPD Convolution, 2 C2f_MSDA, SPPF, compression and activation network module SE-Attention; The feature fusion network layer includes Upsampling (number 1), Concat (number 1), and C2f (number 1). 3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) Convolutions 3 and 6 (Conv), concatenation 3 (Concat), and C2f 3. Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4. 3; Upsampling (No. 1), Concat (No. 1), C2f (No. 1) 3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) 3. Top-down connection; 2. C2f The output of 3 is input to convolution 6 (Conv); Convolution 6 (Conv), Concatenation 3 (Concat), C2f 3 Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4. 3. Bottom-up connection; The output of the compression and activation network module SE-Attention is simultaneously input to Upsampling module 1 and Concat module 4. C2f No. 1 The output of 6 is concatenated to input of 2 using Concat; C2f No. 2 The output of 6 is concatenated to input of 1 using Concat; C2f No. 1 The output and input of 3 are concatenated to 3 using Concat; C2f No. 2 Output of 3, C2f of number 3 Output of 3, C2f of number 4 The output of 3 is simultaneously input to the feature detection output layer; the feature detection output layer outputs the recognition result, and the feature detection output layer is implemented using a decoupled head structure.

[0058] Furthermore, considering that in automated verification production lines for low-voltage current transformers, video surveillance images often contain numerous small targets with indistinct features, little information, and are easily confused with the background, making recognition extremely difficult. To address these issues, the YOLOv8-MSS model employs SPD convolution, concatenating a spatial-to-depth (SPD) layer with a non-convolutional stride (Conv) layer. A schematic diagram of SPD convolution is shown below. Figure 4 As shown. Input feature map (set to...) I The size of ) is H (Length) × W (Width) × C 1 (number of channels), using block size for space-to-depth transformation operations. S To divide the space into regions. S × S The data is converted into a new number of channels. Specifically, the SPD layer uses the block size. S Perform a space-to-depth transformation operation, H × W × C The input feature map of 1 is mapped and segmented into Each sub-feature map increases its depth by reducing its spatial resolution, and then it is divided along the channel dimension. The sub-feature maps are spliced ​​and merged to obtain × S 2 CThe intermediate feature map of 1 makes the spatial dimension return to its original value. However, the channel dimension has increased. S 2 Finally, the intermediate feature maps are transformed by a non-stride convolutional layer to obtain the final result. × C Feature map 2. A convolution operation with a stride of 1 is used, allowing the convolution kernel to move pixel-by-pixel across the input feature map. This ensures that the convolution kernel is applied at every location in the feature map, maximizing information preservation and feature processing. This convolutional layer does not reduce the spatial size of the feature map; instead, it extracts important features by learning information from these added channels. Furthermore, considering the varying scales and complex backgrounds of tooling pallets and transformers under inspection in assembly line scenarios, to avoid missed detections, false detections, and other issues, the YOLOv8-MSS model adds a multi-scale dilated attention mechanism (MSDA) to the C2f structure of the YOLOv8n network. A schematic diagram of the MSDA principle is shown below. Figure 5 As shown, in the Multi-Scale Hollow Attention (MSDA) mechanism, the channel dimension of the feature map is divided into three heads. To identify the large, medium, and small sizes of the wiring terminals of the wiring components in different monitoring video images, different dilation rates are used within each head. r The local region patch surrounding the query is sampled, and self-attention computation is performed. These operations are conducted between colored patches within a window surrounding the blue query patch. Finally, the self-attention values ​​calculated from different heads are concatenated and input into a linear layer for feature aggregation. This allows semantic information to be aggregated at various scales within the receptive field of interest without complex operations or additional computational costs, thereby improving the recognition performance of targets at different scales of contact faults. Furthermore, the YOLOv8-MSS model in this implementation uses the SE network attention mechanism to strengthen key features and enhance the model's feature representation efficiency. A schematic diagram of the principle of the compression and activation network module SE-Attention is shown below. Figure 6As shown, the SE-Attention network module, which performs compression and excitation, primarily functions through two key steps: compression and excitation. First, in the compression (Squeeze) step, the spatial dimension of each channel of the input feature map is compressed into a scalar to capture global information. Then, in the excitation step, two fully connected layers (FC) are used to predict the importance weights of the global information for each channel. The ReLU and Sigmoid activation functions are used to perform non-linear transformations and normalization on the importance weights to generate channel weights. This dimensionality reduction is followed by dimensionality increase to quantify the importance of each channel. Finally, the weights are solved and multiplied channel-by-channel with each channel of the original feature map, thereby enhancing the network's response to useful features.

[0059] Specific Applications: Quality monitoring refers to the operational techniques and activities employed by verification or calibration laboratories to ensure the accuracy and reliability of their verification (calibration) data and results. Considering that automated verification lines for low-voltage current transformers are now widely used in the verification production of provincial power metering centers, and compared to traditional manual verification platforms, automated verification lines for metering low-voltage current transformers face greater quality risks due to issues such as poor contact at secondary terminals, effective quality monitoring is necessary. The image recognition method for faults in automated verification lines for current transformers disclosed in this application is applicable to the automated verification of different types of low-voltage current transformers, including busbar type and multi-turn type. This application focuses on faults in the secondary side terminals of the transformer at the insulation withstand voltage testing station and the error testing station, but the method is not limited to these two stations. After adjusting the dataset samples, it is also applicable to the automated identification and analysis of faults in video monitoring images from other stations on the production line.

[0060] This application presents a machine vision-based video image acquisition technology and a deep learning algorithm-based model for identifying contact faults in the secondary side wiring terminals of low-voltage current transformers. This model enables real-time monitoring and analysis of wiring faults during automated production line verification. In addition to contact faults, the method can be further extended to the rapid monitoring and identification of other faults during production line operation, such as whether barcodes and scanners are aligned, whether network cables are loose, and whether gripper components are faulty.

[0061] The pipeline video image preprocessing method based on spectral density analysis has a significant effect on eliminating or reducing the influence of white noise such as resistive thermal noise caused by test equipment during video image acquisition. It can provide a reference method for the extraction of net signals of one-dimensional, two-dimensional or even high-dimensional digital information such as similar images or spectra.

[0062] Because the automated verification line for low-voltage current transformers is a high-level industrial automation system integrating multiple disciplines such as mechanics, electronics, optics, and computing, the system is reliable, versatile, and its verification level is among the leading in China and has reached the advanced level internationally. Therefore, although faults may occur during the operation of a high-quality automated verification line system, its failure rate is relatively low. Thus, in the process of creating deep learning models, the time and economic costs of fault sample collection are both high. Therefore, this application discloses an improved Cutmix dataset sample augmentation method based on hybrid pruning for identifying faults in the secondary side terminal contact of current transformers. This method can specifically solve the problem of insufficient fault sample information. The research on the method and its implementation process have both academic research value and practical significance.

[0063] This application addresses contact faults in the secondary side terminals of current transformers under test in automated calibration lines for low-voltage current transformers. It discloses a YOLOv8-MSS deep learning model based on MSDA-SE-SPD for automatic computer-aided identification and analysis of acquired video images. Considering the imaging characteristics of the wiring terminals at the withstand voltage testing and error testing stations in the automated calibration line system for low-voltage current transformers due to different monitoring angles of the video surveillance equipment, and to improve the extraction capability of different scale target features presented by the wiring terminals and the S1 / S2 wiring bolts of the current transformer secondary side terminals in the video images during crimping, this application, while designing and developing a fault identification model based on the YOLOv8 deep learning model, further improves the YOLO model by adding the MSDA multi-scale hole attention mechanism, introducing SE compression and excitation network modules, and replacing the SPD convolution method, thus establishing an improved YOLOv8-MSS model based on MSDA-SE-SPD. The YOLOv8-MSS method optimizes model feature extraction performance and improves system computational efficiency, thereby enabling accurate identification, tracking, and analysis of corresponding faults in the production line. The method can not only provide an effective monitoring means for fault identification in the automated verification production line of low-voltage current transformers, but also provide a reference and practical means for fault identification or classification monitoring in other similar industrial automated measurement and control production lines.

[0064] In the context of rapid digital information development, automated verification line systems for representative electrical power measurement equipment such as instrument transformers must consider all aspects related to power equipment verification technology in their major technological innovations. This ranges from the structural layout of the testing devices, the connection between inbound and outbound processes, the methods of loading and unloading, the electrical and control environment settings required for equipment operation, the measurement, control, analysis, and management of electrical parameters, and the transportation control of the conveyor lines, all the way to the processing, analysis, and storage of measurement and control data, the connection with relevant management and control platforms, and the evaluation of verification pass / fail. All these aspects are carried out through instrument science and technology, including sensor testing technology, precision instruments and machinery, and testing and measurement technology and instruments. Sensors and other testing instruments and equipment are used to acquire the "raw" data of interest, which must be managed and stored before access. Currently, the challenge is to automatically, quickly, and accurately acquire the largest amount of data and use machine learning to model, analyze, and predict massive amounts of data. This learning is based on the development of artificial neural networks. Therefore, developing and researching deep learning models based on neural networks is an essential tool in the computer information age, applicable to increasingly automated, precise, and accurate automated line measurement and control management. As a machine learning method, YOLO neural networks have a stronger learning ability and better predictive ability with a larger number of typical samples. Modeling, analyzing, and expanding sample datasets for YOLO neural networks has high academic research value and promising application prospects, meeting the future development trend of intelligent information testing and control technology in the digital information age.

[0065] While this application has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of this application. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of this application as defined by the appended claims. It should be understood that different dependent claims and features herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other embodiments.

Claims

1. A fault identification method for an automated current transformer calibration production line, characterized in that, include: Step 1: Collect images of the workstations to be tested on the production line, preprocess the collected images, and construct a dataset based on the preprocessed data; Step 2: Augment the images in the dataset; Step 3: Train the instrument transformer secondary side terminal contact fault identification model using the dataset, and use the trained instrument transformer secondary side terminal contact fault identification model to identify faults. Step 1 includes: Step 11: Select the time for video image extraction. , The initial value is t. Step 12, from time Start by extracting n images at time intervals of Δt; Step 13: Convert the acquired n images from RGB color space to HSI color space to obtain the corresponding H, S, and I components. The spatial resolution of each image is M × N. Perform bitmap segmentation on the I component of each image to obtain 8 bit-plane images corresponding to each image, denoted as . (x, y) are the position coordinates of pixels in the image, x = 0, 1, 2, …, M-1, y = 0, 1, 2, …, N-1, i = 1, 2, …, n; Step 14: Calculate the coordinates of different pixel positions. The mean combination matrix where i takes values ​​from 1 to n ; Step 15: Calculate the position coordinates of each pixel. The difference between the values ​​and the corresponding mean combination matrix forms the difference combination matrix of the coordinates of each pixel. The difference combination matrix is ​​denoted as: ,and , Proceed to step 12; if The difference combination matrix is ​​denoted as: Proceed to step 16; Step 16, Calculation and The combination matrix of autocorrelation functions at two time points: ； in, E[.] represents a submatrix of the autocorrelation function combination matrix, where E[.] is the expected value of the object within the square brackets. Step 17: Perform a Discrete Fourier Transform on the autocorrelation function combination matrix to obtain the combination matrix of the corresponding spectral density functions: ； Where f is the digital frequency in the frequency domain corresponding to the time domain when the video image changes over time; Step 18: Based on the combination matrix of the spectral density function The sum of the spectral density function values ​​for each plane of the image is obtained. If the sum of the spectral density function values ​​of the j-th plane image corresponding to n images If all are equal to or approximately equal to the same constant, then the j-th plane image of the I component is filtered out, and the remaining plane images are recombined to obtain the I component image after noise removal. Step 19: Normalize the H component, S component, and noise-filtered I component images, and then convert them into RGB images to complete the preprocessing.

2. The fault identification method for an automated current transformer calibration production line according to claim 1, characterized in that, Step 2 includes: Step 21: Select image A1 in the dataset as the master image of the augmentation sample. Randomly crop image A1 to obtain multiple rectangular regions. Scale down or up each region rectangle to obtain the specified rectangular region size, denoted as image A2. Step 22: Set up two mutually exclusive structural elements B1 and B2 images of the secondary side wiring bolt of the current transformer. For image A2, use the hit / miss algorithm to identify and locate the structural elements B1 and B2 images of the wiring bolt. Restore the identification and location results to obtain the image of the wiring bolt marked in image A1, denoted as image A3. Extract the rectangular area box where the wiring bolt is located in image A3. Step 23: Set up two mutually exclusive structural elements B1 and B2 images of the terminal block. For image A2, use the hit / miss algorithm to identify and locate the structural elements B1 and B2 images of the terminal block. Restore the identification and location results to obtain the image marked with the terminal block in image A1, denoted as image A4. Extract the rectangular area box where the terminal block is located in image A4. Step 24: Perform binary masking on the rectangular area of ​​the terminal block in image A4 to complete the area with zeros, and obtain the background image of the augmented image, denoted as image A5; Step 25: Blend the rectangular area containing the wiring bolt, the rectangular area containing the wiring terminal, and image A5 to obtain the augmented image.

3. The fault identification method for an automated current transformer calibration production line according to claim 2, characterized in that, In step 22, based on the external assembly mode of the wiring bolt and the actual imaging form of the monitoring angle of the production line, typical appearances in the monitoring images are selected as structural element B1 images of the wiring bolt, and the complement of structural element B1 images, which is a limited local rectangular area within the body of the wiring bolt, is set as structural element B2 images of the wiring bolt. Based on the actual imaging shape of the rectangular area around the thread of the terminal block and the probe it is assembled with, and the monitoring angle of the production line, the typical appearance in the monitoring image is selected as the structural element B1 image of the terminal block. The complement of the structural element B1 image, which is a limited local rectangular area within the body of the terminal block, is set as the structural element B2 image of the terminal block.

4. The fault identification method for an automated current transformer calibration production line according to claim 2, characterized in that, The hit / miss algorithm changes the original precise matching method in the pixel value evaluation process to a fuzzy matching method. The degree of fuzzy matching is evaluated by measuring the intersection between the target recognition box of the hit / missed pixel in the processed image and the B1 image reference box, and the degree of overlap (IoU) of the structuring element B1 image reference box.

5. The method for fault identification in an automated current transformer calibration production line according to claim 2, characterized in that, Step 25 includes: Based on the size and position of the rectangular area where the wiring bolt is located, mark the corresponding position in image A5 with a rectangular box of the same color. Based on the position and shape of the rectangular area frame where the wiring bolt is located, and combined with the crimping contact position and ratio of the wiring bolt and the wiring terminal, determine the corresponding position area frame to be covered by the wiring terminal in image A5. Cover the corresponding position area frame with the rectangular area frame where the wiring terminal is located extracted by S33, and mark the colored rectangle frame to complete the augmentation.

6. The fault identification method for an automated current transformer calibration production line according to claim 1, characterized in that, The model for identifying contact faults in the secondary side terminals of a current transformer includes a feature extraction network layer, a feature fusion network layer, and a feature detection output layer connected in sequence. The feature extraction network layer consists of the following convolutional layers connected in sequence: Conv1, Conv2, SPD convolution1, C2f_MSDA convolution1, Conv3, SPD convolution2, and C2f_MSDA convolution1. Convolutions 6 and 4 (Conv), 3 (SPD), and 2 (C2f) 6, 5 Convolution Conv, 4 SPD Convolution, 2 C2f_MSDA, SPPF, compression and activation network module SE-Attention; The feature fusion network layer includes Upsampling (number 1), Concat (number 1), and C2f (number 1).

3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) Convolutions 3 and 6 (Conv), concatenation 3 (Concat), and C2f 3. Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4. 3; Upsampling (No. 1), Concat (No. 1), C2f (No. 1) 3. Upsampling (No. 2), Concat (No. 2), C2f (No. 2) 3. Top-down connection; 2. C2f The output of 3 is input to convolution 6 (Conv); Convolution 6 (Conv), Concatenation 3 (Concat), C2f 3 Convolutions 3 and 7 (Conv), concatenation 4 (Concat), and C2f 4.

3. Bottom-up connection; The output of the compression and activation network module SE-Attention is simultaneously input to Upsampling module 1 and Concat module 4. C2f No. 1 The output of 6 is concatenated to input of 2 using Concat; C2f No. 2 The output of 6 is concatenated to input of 1 using Concat; C2f No. 1 The output and input of 3 are concatenated to 3 using Concat; C2f No. 2 Output of 3, C2f of number 3 Output of 3, C2f of number 4 The output of 3 is simultaneously input to the feature detection output layer; the feature detection output layer outputs the recognition result, and the feature detection output layer is implemented using a decoupled head structure.

7. The fault identification method for an automated current transformer calibration production line according to claim 6, characterized in that, SPD convolutions 1 through 4 all include SPD layers and non-strut convolutional layers. The processing steps include: H is the length of the input feature map, W is the width of the input feature map, and C1 is the number of channels in the input feature map. The SPD layer uses a block size S to perform a space-to-depth transformation, mapping and segmenting the H×W×C1 input feature map into... Each feature map is then divided along the channel dimension. The sub-feature maps are spliced ​​and merged to obtain The intermediate feature map is obtained by Conv transformation through a non-stride convolutional layer. The feature map.

8. The fault identification method for an automated current transformer calibration production line according to claim 6, characterized in that, Both C2f_MSDA No. 1 and C2f_MSDA No. 2 are C2f models that add a multi-scale dilated attention mechanism (MSDA). In MSDA, the channel dimension of the feature map is divided into three heads. For different input image sizes, different dilation rates (r) are used within each head to sample the local patch around the query and perform self-attention calculation. The self-attention calculated from different heads is concatenated and input into a linear layer for feature aggregation to obtain the aggregated semantic information at each scale.

9. The fault identification method for an automated current transformer calibration production line according to claim 6, characterized in that, The SE-Attention network module's processing steps include compression and activation steps: In the compression step, the spatial dimension of each channel of the input feature map is compressed into a scalar to obtain global information; In the activation step, two fully connected layers are used to predict the importance weights of global information for each channel, and the importance weights are nonlinearly transformed and normalized using the activation functions ReLU and Sigmoid to generate channel weights. These channel weights are then multiplied one-to-one with each channel of the input feature map to output the channel weights.

Images