A transformer core transversely sheared sheet material appearance on-line detection system and method thereof

Abstract

The present application relates to the technical field of transformer core detection, in particular to a kind of transformer core transverse shearing piece material appearance online detection system, the system is installed in the conveying mechanism of transverse shearing machine top, including piece material trigger unit, acquisition imaging unit and shape calculation unit;The piece material trigger unit is used to detect the arrival, departure state of piece material, sends trigger signal, and exports speed signal and frequency signal;The acquisition imaging unit realizes image imaging by line frequency trigger mode;The shape calculation unit is used for the storage, segmentation and contour extraction of image, and the piece material measurement value is structured output.This application also provides a kind of transformer core transverse shearing piece material appearance online detection method.The present application realizes the online detection of core piece material, improves production efficiency.

Description

Technical Field

[0001] This invention relates to the field of transformer core testing technology, and in particular to an online inspection system and method for the appearance of transversely sheared transformer core sheets. Background Technology

[0002] In power systems, transformers play a crucial role, transforming high voltage into low voltage or vice versa. The transformer core is a vital component, responsible for carrying and transmitting electromagnetic energy. The transformer core enhances the transformer's energy efficiency and stability while also preventing electromagnetic interference and losses.

[0003] In traditional transformers, the core material is primarily silicon steel sheets, and the structure and shape of the core vary depending on the application. Generally, a transformer core is composed of multiple stacked silicon steel sheets. The silicon steel sheets undergo two processes during core manufacturing: longitudinal shearing and transverse shearing. Longitudinal shearing cuts the silicon steel coils into strips that meet the width requirements of the core structure. Transverse shearing involves punching holes and cutting corners in the strips, further cutting them into sheets that meet the required length and shape.

[0004] In actual production, the following problems exist in the inspection of silicon steel strips and sheets:

[0005] (1) The inspection of silicon steel strips is generally carried out by manual sampling. During the inspection process, the automatic shearing machine is stopped, which leads to a reduction in production efficiency.

[0006] (2) Inspection of silicon steel sheets: Due to the high requirements of the production process of transformer core sheets, the dimensional accuracy is required to be within 0.2mm, while the typical width value is 600mm and the typical length value is 2000mm. The detection error range is required to be between 0.001% and 0.03%. Therefore, it is extremely challenging to achieve large-size high-precision online inspection of core sheet shapes without stopping the machine and to measure the appearance of the sheets in real time on the production line. Summary of the Invention

[0007] To address the aforementioned shortcomings, the present invention aims to provide an online inspection system and method for the appearance of transversely sheared transformer core sheets.

[0008] To achieve the above objectives, the present invention provides an online inspection system for the appearance of transformer core sheared sheets. The system is installed above the conveying mechanism of the shearing machine and includes a sheet triggering unit, an acquisition and imaging unit, and an appearance calculation unit.

[0009] The sheet material triggering unit is used to detect the arrival and departure status of the sheet material, send a trigger signal, and output a speed signal and a frequency signal;

[0010] The acquisition and imaging unit achieves image imaging through a combination of frame triggering and line frequency triggering.

[0011] The shape calculation unit is used for image storage, segmentation and contour extraction, and outputs the sheet material measurement values ​​in a structured manner.

[0012] Preferably, the sheet material triggering unit includes a triggering device and a speed measuring device; the triggering device includes a color difference sensor, which includes an RGB color sensor and a color mark sensor; the speed measuring device includes a Doppler velocimeter and a detector, which includes a differential speed measuring sensor module and a laser.

[0013] The acquisition and imaging unit includes a line scan camera and a frequency converter;

[0014] The shape calculation unit includes an image storage module, an image segmentation module, and a shape contour extraction module. The image storage module is used to store images, the image segmentation module is used for image segmentation, and the shape contour extraction module is used for extracting the shape contour of the graphic.

[0015] This invention also provides an online inspection method for the appearance of transversely sheared transformer core sheets, including a sheet triggering step, an image acquisition step, and an appearance calculation step;

[0016] The sheet material triggering step detects the beginning and end of the sheet material by sensing the color difference change between the background and the incoming sheet material, while simultaneously measuring the speed and outputting speed and frequency signals to subsequent units.

[0017] In the acquisition and imaging step, the acquisition and imaging unit receives a frequency signal fp from the sheet triggering unit and triggers a start signal s. start and trigger end signal s end The line scan camera received the frame trigger start signal s start Then, the line trigger signal fp is input to the frequency converter and encoder for processing;

[0018] The sheet material shape calculation step involves extracting the image contour through multi-scale image iterative scaling, strip patch image segmentation, and improved adaptive threshold image contour extraction. Finally, the image appearance size is obtained by fitting straight lines and angles.

[0019] Preferably, the sheet material triggering step is as follows:

[0020] Step S101: Detect the color change or grayscale change of the target point, and then proceed to step S102.

[0021] Step S102: Determine whether it is a sheet material head based on whether the color change or grayscale change exceeds the threshold; if yes, proceed to steps S109 and S103; if no, proceed to step S101.

[0022] Step S103: When the detected color change or grayscale change exceeds the threshold, the speed measuring device performs a 2-level differential speed measurement and executes step S104.

[0023] Step S104: After the speed measuring device completes the speed measurement, calculate the frequency, and then execute steps S105 and S110.

[0024] When the lateral velocity of the object being measured is zero, the frequency of the reflected light and the detector light are the same; when the lateral velocity is not zero, the frequency of the reflected light shifts relative to the detector light.

[0025] Step S105: Calculate the instantaneous lateral velocity based on the frequency, and then execute steps S106 and S111.

[0026] The Doppler velocimeter calculates the frequency shift using a fast Fourier transform to obtain the instantaneous lateral velocity v of the measured object. t ;

[0027] Step S106, based on the instantaneous lateral velocity v t Calculate the length of the sheet material and execute steps S107 and S112;

[0028] The formula for calculating the length of the sheet material is:

[0029] l slice ==∑v t ×t interval (3)

[0030] Among them, t interval The speed measurement time interval of the speed measuring device;

[0031] Step S107: Detect whether the color change or grayscale change of the target point exceeds the threshold and determine whether it is a sheet material tail; if yes, proceed to step S108; if no, proceed to step S103.

[0032] Step S108: When the end of the sheet material is detected, the shape calculation unit is triggered, and the shape calculation unit calculates the shape of the sheet material measured this time.

[0033] Step S109, set the sheet length l slice =0, proceed to step S106;

[0034] In step S110, the sheet material triggering unit continuously outputs the sheet material frequency signal fp and sends it to the acquisition and imaging unit;

[0035] Step S111: The sheet material triggering unit continuously outputs the instantaneous speed signal v of the sheet material. t And transmit it to the shape calculation unit;

[0036] Step S112, the sheet material triggering unit continuously outputs the sheet material accumulation length l slice The data is then sent to the shape calculation unit.

[0037] Furthermore, in step S102, when a color change occurs at the detected target point, an RGB color sensor is used. The calculation formula for when the RGB color sensor detects a color change at the target point reaching a threshold is as follows:

[0038]

[0039] Where, r o g o b o Let r be the color component of the target to be measured. b g b b b The color component for the background;

[0040] When a grayscale change occurs at the target point, the color difference sensor calculates the color difference gradient value. The calculation formula for when the color change at the target point detected by the color difference sensor reaches the threshold is as follows:

[0041]

[0042] Among them, g o g represents the grayscale or brightness value of the target to be measured. b The grayscale or brightness value of the background.

[0043] Furthermore, the acquisition and imaging steps are as follows:

[0044] Step S201: The line scan camera receives the trigger start signal s output from the sheet trigger unit. start Execute step S201; the line scan camera receives the trigger end signal s output from the sheet trigger unit. end Execute step S205;

[0045] Step S202: Based on the frequency signal fp output by the sheet trigger unit, the inverter adjusts the actual line frequency of the line scan camera, and then executes step S203.

[0046] Step S203: The line scan camera acquires images line by line, and then proceeds to step S204;

[0047] Execute step S204 to stitch the images acquired line by line in the frequency domain, and then execute steps S205 and S201.

[0048] In step S205, line scanning transmits the image in the buffer to the shape calculation unit.

[0049] Furthermore, the steps for calculating the shape of the sheet material are as follows:

[0050] In step S301, the sheet material shape calculation unit receives the image in the buffer area from the line scan camera and performs multi-scale image iterative scaling.

[0051] Step S302: Perform bar patch image segmentation on the scaled image;

[0052] Step S303: After segmenting the region in the strip patch image, the segmentation result is used as an image mask to perform adaptive image contour extraction on the obtained region of interest.

[0053] Step S304, line detection;

[0054] Step S305: Extract edges, corners, and V-shaped regions;

[0055] Step S306: Output the sheet shape dimensions.

[0056] Furthermore, in step S301, the image is subjected to multi-scale iterative lossless scaling, including image scaling calculation and image grading processing.

[0057] The image scaling calculation is shown in equations (4) and (5):

[0058] hnage resize =Lanczos(Image) (4)

[0059]

[0060] The image grading process is as follows: for each scaling step, the image interpolation is calculated using the eigenvector of the Lanczos matrix in equation (4); equation (4) is called multiple times to scale the image step by step until it reaches the size that the deep learning cache can handle.

[0061] Furthermore, in step S302, a 256x1024 strip patch was used to segment the image using a deep learning network model; the deep learning network model used the LeakyRelu nonlinear activation function to output the segmentation result, and the expression of the LeakyRelu activation function is shown in equation (6):

[0062] LeakyRelu(x)=max(αx,x) (6)

[0063] Where α is a slope coefficient less than 1, usually taken as 0.01;

[0064] When x > 0, LeakyReLU is the same as ReLU;

[0065] When x <= 0, the output of LeakyReLU is αx, meaning there is an output in the negative range.

[0066] 10. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 7, characterized in that, in step S303, adaptive image contour extraction includes the following steps:

[0067] Step S3031: Obtain the region of interest in the original image through the image segmentation results;

[0068] Step S3032: Perform grayscale processing on the acquired image interest region to obtain a grayscale image;

[0069] Step S3033: Obtain the Sobel gradient maps in the horizontal and vertical directions of the image based on the grayscale image;

[0070] Step S3034: Superimpose the Sobel gradient maps in the horizontal and vertical directions of the image to obtain a superimposed image;

[0071] Step S3035: Generate a grayscale histogram based on the overlay image, using the maximum gradient value of the overlay as the upper limit;

[0072] Step S3036: Obtain the gray value Hmax corresponding to the maximum gray frequency, and calculate the pixel extreme value variance Emax;

[0073] Step S3037: Obtain the upper and lower bound parameters for Canny edge extraction by summing Hmax and Emax;

[0074] Step S3038: The image contour is extracted using the Canny edge extraction method, where the upper and lower bound parameters high and low are calculated using equations (7) to (9):

[0075] high = Emax + Hmax (7)

[0076] low = high × ab (8)

[0077] high = low × 2 (9)

[0078] The Canny edge extraction method is shown in equation (10):

[0079] Canny (grayscale, outline, high, low) (10).

[0080] This invention provides an online inspection system and method for the appearance of transversely sheared transformer core sheets, primarily used for real-time monitoring and automatic inspection of the appearance quality of these sheets. The system enables real-time inspection during production, significantly improving inspection efficiency, reducing human error, and ensuring the continuity and stability of the production process. Employing high-precision image recognition technology and sensors, the system accurately judges the appearance quality of the transversely sheared core sheets, ensuring the accuracy of the inspection results. Simultaneously, the system provides timely feedback on the inspection results, allowing production personnel to immediately adjust production processes based on the findings, promptly correcting problems and preventing the generation of batch quality issues. Furthermore, the system utilizes intelligent technologies such as deep learning to reduce reliance on a large number of inspection personnel, significantly reducing labor costs. The system's continuous recording of inspection data, combined with the data analysis functions of the system software, facilitates production process tracking and continuous improvement of quality control. This system fills a gap in the domestic field of online inspection of transversely sheared transformer cores without shutdown, significantly improving production efficiency. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the online inspection system for the appearance of transversely sheared transformer core sheets according to the present invention;

[0082] Figure 2 This is a flowchart of the sheet material triggering process; the sheet material triggering process, the acquisition and imaging process, and the shape calculation process.

[0083] Figure 3 This is a flowchart of the acquisition and imaging steps;

[0084] Figure 4 This is a flowchart of the shape calculation steps;

[0085] Figure 5 This is a schematic diagram illustrating the image acquisition and imaging process to obtain the resolution of the image sheet.

[0086] Figure 6 This is an example of an image segmented by a strip patch that contains interfering factors such as shadows and light / dark boundaries.

[0087] Figure 7 This is a schematic diagram of the UNet network structure used for strip patch image segmentation;

[0088] Figure 8 This is a schematic diagram of the upsampling and downsampling convolutional block structure;

[0089] Figure 9 This is a flowchart of adaptive image contour extraction;

[0090] Figure 10 This is a schematic diagram of the measurement results of the measurement method of the present invention. Detailed Implementation

[0091] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0092] Example 1

[0093] See Figure 1 This invention provides an online inspection system for the appearance of transformer core sheared sheets, which is installed above the conveying mechanism of the shearing machine and includes a sheet triggering unit, an acquisition and imaging unit, and an appearance calculation unit.

[0094] The sheet material triggering unit is used to detect the arrival and departure status of the sheet material, send a trigger signal, and output a speed signal and a frequency signal;

[0095] The acquisition and imaging unit achieves image imaging through a combination of frame triggering and line frequency triggering.

[0096] The shape calculation unit is used for image storage, segmentation and contour extraction, and outputs the sheet material measurement values ​​in a structured manner.

[0097] Specifically, the sheet material triggering unit includes a triggering device and a speed measuring device.

[0098] The triggering device is a color difference sensor; the color difference sensor includes an RGB color sensor and a color mark sensor;

[0099] The speed measuring device includes a Doppler velocimeter and a detector. The Doppler velocimeter includes a differential speed sensing module and a laser.

[0100] The acquisition and imaging unit includes a line scan camera and a frequency converter.

[0101] The shape calculation unit includes an image storage module, an image segmentation module, and a shape contour extraction module. The image storage module is used to store images, the image segmentation module is used for image segmentation, and the shape contour extraction module is used for extracting the shape contour of the graphic.

[0102] During operation, the triggering device detects the incoming sheet material on the cross-cutting machine. When the triggering device detects the presence of sheet material, it transmits a trigger signal to the speed measuring device and sends a frame start signal to the line scan camera.

[0103] The speed measuring device measures the speed and sends a frequency signal to the inverter, while sending length and speed signals to the outline extraction module. The inverter receives the frequency signal and adjusts the frequency. The adjusted frequency controls the line scan camera to receive the frame start signal and begin acquiring images line by line. When the line scan camera receives the frame end signal from the trigger device, the line scan camera stops working.

[0104] The image storage module receives and stores images sent by the line scan camera, and then sends the stored images to the image segmentation module for image segmentation. The shape contour extraction module extracts the segmented images based on the length and speed signals sent by the speed measuring device, and then outputs the results after the image shape detection is formatted.

[0105] Example 2

[0106] This invention also provides an online inspection method for the appearance of transversely sheared transformer core sheets, including a sheet triggering step, an image acquisition step, and an appearance calculation step.

[0107] The sheet material triggering step primarily detects the beginning and end of the sheet material by sensing the color difference change between the background and the incoming sheet material. Simultaneously, a Doppler velocimeter is used to measure the speed, and the speed and frequency signals are output to subsequent units. The sheet material triggering step is as follows: Figure 2 As shown:

[0108] The sheet material triggering step runs when the equipment is powered on.

[0109] Step S101: Detect color or grayscale changes at the target point using a color difference sensor, and then proceed to step S102.

[0110] Step S102: Determine whether it is a sheet material head based on whether the color change or grayscale change exceeds the threshold; if yes, proceed to steps S109 and S103; if no, proceed to step S101.

[0111] When a color change is detected at the target point, an RGB color sensor is used. The calculation formula for when the RGB color sensor detects a color change at the target point that reaches a threshold is as follows:

[0112]

[0113] Where, r o g o b o Let r be the color component of the target to be measured. b g b b b The color component for the background.

[0114] When a grayscale change occurs at the target point, the color difference sensor calculates the color difference gradient value. The formula for calculating the threshold value when the color change at the target point detected by the color difference sensor reaches is:

[0115]

[0116] Among them, g o g represents the grayscale or brightness value of the target to be measured.b The grayscale or brightness value of the background.

[0117] In this step, the color difference sensor used to detect color changes may employ hardware devices including but not limited to RGB color cameras and color mark sensors, none of which affect the implementation principle of this patent.

[0118] Step S109, set the sheet length l slice =0, proceed to step S106.

[0119] Step S103: When the detected color change or grayscale change exceeds the threshold, the speed measuring device performs a 2-level differential speed measurement and executes step S104.

[0120] The speed measuring device is a Doppler velocimeter. The light emitted by the differential speed sensing module inside the Doppler velocimeter is split into two beams and emitted out. The beams converge again on the surface of the object being measured, and the reflected light is then received by the detector.

[0121] Step S104: After the speed measuring device completes the speed measurement, calculate the frequency, and then execute steps S105 and S110.

[0122] When the lateral velocity of the object being measured is zero, the frequency of the reflected light and the detector light are the same; when the lateral velocity is not zero, the frequency of the reflected light shifts relative to the detector light.

[0123] In step S110, the sheet material triggering unit continuously outputs the sheet material frequency signal fp and sends it to the acquisition and imaging unit.

[0124] Step S105: Calculate the instantaneous lateral velocity based on the frequency, and then execute steps S106 and S111.

[0125] The Doppler velocimeter calculates the frequency shift using a fast Fourier transform to obtain the instantaneous lateral velocity v of the measured object. t .

[0126] Step S111: The sheet material triggering unit continuously outputs the instantaneous speed signal v of the sheet material. t The data is then sent to the shape calculation unit.

[0127] Step S106, based on the instantaneous lateral velocity v t Calculate the length of the sheet material and execute steps S107 and S112;

[0128] The formula for calculating the length of the sheet material is:

[0129] l slice =∑v t ×t interbal (3)

[0130] Among them, t intervalThis refers to the speed measurement time interval of the speed measuring device, specifically the speed measurement time interval of the differential speed sensor, with a typical interval of 0.01 milliseconds.

[0131] Step S112, the sheet material triggering unit continuously outputs the sheet material accumulation length l slice The data is then sent to the shape calculation unit.

[0132] Step S107: Detect whether the color change or grayscale change of the target point exceeds the threshold using a color difference sensor to determine whether it is a sheet material tail; if yes, proceed to step S108; if no, proceed to step S103.

[0133] In step S108, after the color difference sensor detects that the end of the sheet material has arrived, it triggers (notifies) the shape calculation unit, and the shape calculation unit calculates the shape of the sheet material measured this time.

[0134] In the acquisition and imaging step, the core of the acquisition and imaging unit is a line scan camera. The line scan camera has external frame trigger signal and line trigger signal functions. The number of lines in a frame is controlled by the frame height register, and the line frequency is controlled by the externally provided line trigger signal, while also being limited by internal settings. Specifically, the line scan camera has 8 lines to receive external triggers, where lines 1 and 2 are line frequency signals, and lines 3 and 4 are frame trigger signals. Figure 1 At the "Arrival Trigger" - "Frame Start Signal" point, frame start triggering is achieved by sending the trigger start signal s. start At the "Leave Trigger" - "Frame End Signal" point, frame end triggering is implemented, and a trigger end signal s is emitted. end The line scan camera receives the frame trigger start signal s. start After that, the horizontal trigger input takes effect, and the horizontal trigger signal fp is input to the frequency converter and encoder for processing. The acquisition and imaging unit receives the frequency signal fp from the sheet trigger unit and triggers the start signal s. start and trigger end signal s end .

[0135] The acquisition and imaging steps are as follows Figure 3 As shown, it includes the following steps:

[0136] Step S201: The line scan camera receives the trigger start signal s output from the sheet trigger unit. start Execute step S201.

[0137] The line scan camera receives the trigger end signal s from the sheet trigger unit. end Execute step S205.

[0138] Step S202: Based on the frequency signal fp output by the sheet trigger unit, the inverter adjusts the actual line frequency of the line scan camera, and then proceeds to step S203.

[0139] Step S203: The line scan camera acquires images line by line, and then proceeds to step S204.

[0140] The line scan camera uses devices such as supplementary lights and CMOS sensors to acquire images of 4K-8K horizontally and 16k-32K vertically, covering the size range of cross-cutting sheet metal from iron cores.

[0141] Execute step S204 to perform frequency domain stitching on the images acquired line by line, and then execute steps S205 and S201.

[0142] In step S205, line scanning transmits the image in the buffer to the shape calculation unit.

[0143] When the line scan camera is activated for line-triggered progressive acquisition, the encoder signal speed controls the camera's actual line frequency after an external encoder trigger signal is connected. If the captured image is slightly compressed or stretched, a frequency converter is required for control. The frequency converter's operation in step S202 is as follows:

[0144] The frequency signal fp output from the sheet trigger unit first enters the pre-divider module. Through integer division, the frequency of the source signal is reduced, and the processed signal is then sent to the multiplier module. The pre-divider module reduces the periodic jitter of the input signal, and signals with frequencies exceeding 100kHz are frequency-reduced by the pre-divider, solving the problem that the multiplier can only accept signals in the 10–100kHz frequency range. Periodic jitter from the signal from the shaft encoder is acceptable.

[0145] After processing by the pre-divider module, the signal is sent to the multiplier. The multiplier multiplies the signal frequency by an integer, increasing the signal frequency. The signal is then sent to the voltage divider module. The adjustment parameter can be set to the rising edge or the falling edge. If set to the rising edge, each rising edge of the signal from the pre-divider module will be locked to match the signal generated by the rising edge, and vice versa. Ensure that you do not use too large a multiplier to increase the signal frequency to avoid triggering the signal frequency to exceed the camera's maximum supported line frequency. Even with a small multiplier, excessively high frequencies may still occur during frequency adjustment, exceeding the camera's maximum line frequency.

[0146] The frequency divider reduces the frequency of the signal by an integer factor and uses the resulting new frequency signal as the camera's trigger signal.

[0147] In the sheet material shape calculation step, the sheet material shape calculation unit receives images from the acquisition and imaging unit. These images are 16K or 32K in size, far exceeding the 2K or 4K images typically processed in image processing. Therefore, this presents significant challenges for image storage and processing. Specifically, the sheet material shape calculation step of this invention involves: extracting the image contour through a multi-scale image iterative scaling step, a strip patch image segmentation step, and an improved adaptive threshold image contour extraction step; finally, obtaining the image appearance dimensions through a line and angle fitting step. The sheet material shape calculation step is as follows: Figure 4 As shown, it includes the following steps:

[0148] In step S301, the sheet material shape calculation unit receives the image in the buffer area from the line scan camera and performs multi-scale image iterative scaling.

[0149] The image is subjected to multi-scale iterative lossless scaling, including image scaling calculation and image grading processing. Specifically, the image scaling calculation is shown in equations (4) and (5):

[0150] Image resize =Lanczos(Image) (4)

[0151]

[0152] The image scaling process is as follows: for each scaling level, the image interpolation is calculated using the eigenvectors of the Lanczos matrix in equation (4). To avoid Lanczos matrix noise, equation (5) incorporates a proportional scaling parameter, i.e., equation (4) is called multiple times to scale the image step by step until it is finally scaled to the size that the deep learning cache can handle. The typical value used here is ∝=0.8, with 8 scaling levels. The resolution is 8192x30000 (8192 multiplied by 30000) pixels, and the scaled size is 1374x5033 (1374 multiplied by 5033) pixels.

[0153] Multi-scale iterative lossless scaling of images is performed to meet the dimensional accuracy requirement of ±0.2mm in the national standard "Electrical Steel Cores for Power Transformers" (GB / T 32288-2020). The image acquisition resolution is 16K or 32K. A single image stitched from the line scan camera occupies more than 180MB of hard disk space and nearly 4GB of memory. For example... Figure 5 The image shown has a resolution of 8129x30000 pixels. Due to the limitations of deep learning technology on video memory capacity, the current 32K resolution image cannot be directly segmented and requires image scaling.

[0154] Step S302: Perform strip patch image segmentation on the scaled image.

[0155] The core material is long and strip-shaped, and the resolution remains high even after scaling. Mainstream graphics cards typically have 8-24GB of video memory, making it impossible to process a single image at a time. Furthermore, there are challenging issues such as shadows and boundaries between light and dark areas at the image edges. Figure 6 As shown. To solve the above problems, a 256x1024 strip patch was used, and a deep learning network model was used for image segmentation training.

[0156] The deep learning network model is an improvement on UNet, and segmentation network structures such as UNet++ and ResUNet can be used without affecting the expression of the method in this invention. The UNet-like network structure proposed in this invention is as follows: Figure 7 As shown, compared to the traditional UNet 4-layer structure, it increases to a 6-layer structure, and uses instance normalization to replace the batch normalization operation. The upsampling and downsampling convolutional block structures are as follows: Figure 8 As shown.

[0157] Both upsampling and downsampling convolutions use 3x3 kernels, while upsampling deconvolutions use 2x2 kernels. Instance normalization is employed, and the LeakyReLU nonlinear activation function is used. LeakyReLU is a variant of ReLU, primarily designed to address the issue of ReLU outputs being zero. When the input is less than 0, although the output value is small, it is not zero, as shown in equation (6).

[0158] LeakyRelu(x)=max(αx,x) (6)

[0159] Here, α is the slope coefficient less than 1, usually taken as 0.01. When x > 0, LeakyReLU is the same as ReLU; when x <= 0, the output of LeakyReLU is αx, that is, there is some output in the negative value range.

[0160] Step S303: After segmenting the region in the strip patch image, the segmentation result is used as an image mask to perform adaptive image contour extraction on the obtained image region of interest.

[0161] Adaptive image contour extraction process as follows Figure 9 As shown, it includes the following steps:

[0162] Step S3031: Obtain the region of interest in the original image through the image segmentation results;

[0163] Step S3032: Perform grayscale processing on the acquired image interest region to obtain a grayscale image;

[0164] Step S3033: Obtain the Sobel gradient maps in the horizontal and vertical directions of the image based on the grayscale image;

[0165] Step S3034: Superimpose the Sobel gradient maps in the horizontal and vertical directions of the image to obtain a superimposed image;

[0166] Step S3035: Generate a grayscale histogram based on the overlay image, using the maximum gradient value of the overlay as the upper limit;

[0167] Step S3036: Obtain the gray value Hmax corresponding to the maximum gray frequency, and calculate the pixel extreme value variance Emax;

[0168] Step S3037: Obtain the upper and lower bound parameters for Canny edge extraction by summing Hmax and Emax;

[0169] Step S3038: Image contour extraction is performed using the Canny edge extraction method. The calculation formulas for the upper and lower bound parameters high and low are shown in equations (7) to (9):

[0170] high = Emax + Hmax (7)

[0171] low = high × ab (8)

[0172] high = low × 2 (9)

[0173] The Canny edge extraction method is shown in equation (10):

[0174] Canny (grayscale, outline, high, low) (10).

[0175] Step S304, line detection.

[0176] Since the outer contour of the sheet material is a straight line, a straight line is fitted using either the Hough transform or the Random Sample Consensus Algorithm (RANSAC) based on adaptive image contour extraction. The resulting straight line information is stored in a straight line list buffer.

[0177] Step S305: Extract edges, corners, and V-shaped regions.

[0178] In the extraction of the outer contour of the sheet material, the final conversion of image pixel information into point, line and surface structured information is completed by extracting the edges, corners and V-shaped regions. The defined structural information includes: length, width, angle (position information, angle value), hole (center coordinates, roundness value, radius), and V-shape (vertex position, angle value, opening direction).

[0179] The image width is calculated as follows:

[0180] (a) Obtain the approximate vertical line of the leftmost image and the approximate vertical line of the rightmost image. Rotate the image according to the offset angle of the approximate vertical line so that the strip is perpendicular to the display screen.

[0181] (b) The width of the piece is as shown in formula (11):

[0182] width = (LineX) right LineX left )×PixelFactor (11)

[0183] Among them, LineX right LineX is the x-coordinate of the rightmost line. left This represents the x-coordinate of the leftmost line. PixelFactor is the actual length of a single pixel in the line scan camera resolution, measured in mm / pixel.

[0184] The method for obtaining the angle is as follows:

[0185] Since the angle will only appear at the intersection of the left and right lines, we can iterate through all lines with an inclination angle of ±40°, and calculate their intersection and angle with the left line, as well as their intersection and angle with the right line. For the angle, we need to further determine the sign of the slope and decide whether to use 180° minus the angle as the result based on the sign of the slope.

[0186] The method for obtaining the V-shape is as follows:

[0187] The V-shape can start in four directions: up, down, left, and right. You can find it by first identifying a slanted straight line at one end, then finding the nearest symmetrical slanted straight line. For example, a V-shape starting to the left. Similarly, for a symbol like the ">", you can first find the top "\" line, then find the nearest " / " line to complete the V-shape.

[0188] The method for calculating image length is as follows:

[0189] Once the angle is determined, find the intersection of the uppermost corner point and the lowermost corner point, calculate the difference between them, and then perform a weighted calculation with the values ​​in the image length segment buffer and the image speed buffer provided by the speed sensor to reduce measurement errors.

[0190] The method for calculating the aperture position in an image is as follows:

[0191] Based on the image segmentation results, within the segmentation mask, a contour search method is used to obtain small contours. The minimum circumcircle of the obtained small contours is calculated, and the area of ​​the minimum circumcircle is compared with the number of pixels within the circumcircle to obtain the image roundness. When the roundness meets a certain threshold, it is considered a hole, and the hole position coordinates, radius, and roundness information are output. The roundness calculation is shown in Equation (12):

[0192]

[0193] Where pixelnum is the number of pixels, and circelouter is the information of the circumcircle of the contour.

[0194] After extracting the edges, corners, and V-shaped areas, the sheet shape dimensions can be output to the MES system, or production anomaly alarm signals can be output based on preset information. Measurement results are as follows: Figure 10 As shown.

[0195] This invention enables online detection of iron core sheets, filling a gap in the domestic field of online detection of iron core cross shears without stopping the machine, and improving production efficiency.

[0196] This invention provides an online inspection system and method for the appearance of transversely sheared transformer core sheets, primarily used for real-time monitoring and automatic inspection of the appearance quality of these sheets. The system enables real-time inspection during production, significantly improving inspection efficiency, reducing human error, and ensuring the continuity and stability of the production process. Employing high-precision image recognition technology and sensors, the system accurately judges the appearance quality of the transversely sheared core sheets, ensuring the accuracy of the inspection results. Simultaneously, the system provides timely feedback on the inspection results, allowing production personnel to immediately adjust production processes, correct problems promptly, and prevent the generation of batch quality issues. Furthermore, the system utilizes intelligent technologies such as deep learning to reduce reliance on a large number of inspection personnel, significantly reducing labor costs. Through continuously recorded inspection data, combined with the data analysis functions of the system software, the system facilitates production process tracking and continuous improvement of quality control, thereby increasing production efficiency.

[0197] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A method for online inspection of the appearance of transversely sheared transformer core sheets, characterized in that: An online inspection system for the appearance of transformer core cross-cutting sheet materials is adopted. The system is installed above the conveying mechanism of the cross-cutting machine and includes a sheet material triggering unit, an acquisition and imaging unit, and an appearance calculation unit. The sheet material triggering unit is used to detect the arrival and departure status of the sheet material, send a trigger signal, and output a speed signal and a frequency signal; The acquisition and imaging unit achieves image imaging through a combination of frame triggering and line frequency triggering. The shape calculation unit is used for image storage, segmentation and contour extraction, and outputs the sheet material measurement values ​​in a structured manner. The sheet material triggering unit includes a triggering device and a speed measuring device; the triggering device includes a color difference sensor, which includes an RGB color sensor and a color mark sensor; the speed measuring device includes a Doppler velocimeter and a detector, which includes a differential speed measuring sensor module and a laser. The acquisition and imaging unit includes a line scan camera and a frequency converter; The shape calculation unit includes an image storage module, an image segmentation module, and a shape contour extraction module. The image storage module is used to store images, the image segmentation module is used for image segmentation, and the shape contour extraction module is used for extracting the shape contour of the graphic. The online detection method includes a sheet material triggering step, an image acquisition step, and a shape calculation step; The sheet material triggering step detects the beginning and end of the sheet material by sensing the color difference change between the background and the incoming sheet material, while simultaneously measuring the speed and outputting speed and frequency signals to subsequent units. In the acquisition and imaging step, the acquisition and imaging unit receives a frequency signal from the sheet triggering unit. Trigger start signal and trigger end signal The line scan camera received a frame trigger start signal. Then, line trigger signal The input is processed by a frequency converter and encoder. The sheet material shape calculation step involves extracting the image contour through multi-scale image iterative scaling, strip patch image segmentation, and adaptive threshold image contour extraction. Finally, the image appearance size is obtained by fitting straight lines and angles. The sheet material triggering steps are as follows: Step S101: Detect the color change or grayscale change of the target point, and then proceed to step S102. Step S102: Determine whether it is a sheet material head based on whether the color change or grayscale change exceeds the threshold; if yes, proceed to steps S109 and S103; if no, proceed to step S101. Step S103: When the detected color change or grayscale change exceeds the threshold, the speed measuring device performs a 2-level differential speed measurement and executes step S104. Step S104: After the speed measuring device completes the speed measurement, calculate the frequency, and then execute steps S105 and S110. When the lateral velocity of the object being measured is zero, the frequency of the reflected light and the probe light are the same; when the lateral velocity is not zero, the frequency of the reflected light shifts relative to the probe light. ; Step S105: Calculate the instantaneous lateral velocity based on the frequency, and then execute steps S106 and S111. The Doppler velocimeter calculates the frequency shift using a fast Fourier transform to determine the instantaneous lateral velocity of the object being measured. ; Step S106, based on instantaneous lateral velocity Calculate the length of the sheet material and execute steps S107 and S112; The formula for calculating the length of the sheet material is: (3) in, The speed measurement time interval of the speed measuring device; Step S107: Detect whether the color change or grayscale change of the target point exceeds the threshold and determine whether it is a sheet material tail; if yes, proceed to step S108; if no, proceed to step S103. Step S108: When the end of the sheet material is detected, the shape calculation unit is triggered, and the shape calculation unit calculates the shape of the sheet material measured this time. Step S109: Set the sheet length Execute step S106; Step S110: The sheet material triggering unit continuously outputs the frequency signal of the sheet material. And transported to the acquisition and imaging unit; Step S111: The sheet material triggering unit continuously outputs the instantaneous speed signal of the sheet material. And transmit it to the shape calculation unit; Step S112: The sheet material triggering unit continuously outputs the sheet material accumulation length. The data is then sent to the shape calculation unit.

2. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 1, characterized in that, In step S102, when a color change occurs at the target point, an RGB color sensor is used. The calculation formula for when the RGB color sensor detects a color change at the target point that reaches a threshold is as follows: (1) in, The color components of the target to be measured. The color component for the background; When a grayscale change occurs at the target point, the color difference sensor calculates the color difference gradient value. The calculation formula for when the color change at the target point detected by the color difference sensor reaches the threshold is as follows: (2) in, The grayscale or brightness value of the target to be measured. The grayscale or brightness value of the background.

3. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 1, characterized in that, The acquisition and imaging steps are as follows: Step S201: The line scan camera receives the trigger start signal output from the sheet trigger unit. Execute step S201; the line scan camera receives the trigger end signal output from the sheet trigger unit. Execute step S205; Step S202: Based on the frequency signal output by the sheet trigger unit. The actual line frequency of the line scan camera is adjusted by the frequency converter, and step S203 is executed. Step S203: The line scan camera acquires images line by line, and then proceeds to step S204; Execute step S204 to stitch the images acquired line by line in the frequency domain, and then execute steps S205 and S201. In step S205, line scanning transmits the image in the buffer to the shape calculation unit.

4. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 1, characterized in that, The steps for calculating the shape of the sheet material are as follows: In step S301, the sheet material shape calculation unit receives the image in the buffer area from the line scan camera and performs multi-scale image iterative scaling. Step S302: Perform bar patch image segmentation on the scaled image; Step S303: After segmenting the region in the strip patch image, the segmentation result is used as an image mask to perform adaptive image contour extraction on the obtained image region of interest. Step S304, line detection; Step S305: Extract edges, corners, and V-shaped regions; Step S306: Output the sheet shape dimensions.

5. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 4, characterized in that, In step S301, the image is subjected to multi-scale iterative lossless scaling, including image scaling calculation and image grading processing. The image scaling calculation is shown in equations (4) and (5): (4) (5) The image grading process is as follows: for each scaling step, the image interpolation is calculated using the eigenvector of the Lanczos matrix in equation (4); equation (4) is called multiple times to scale the image step by step until it reaches the size that the deep learning cache can handle.

6. The online inspection method for the appearance of transversely sheared transformer core sheets according to claim 4, characterized in that, In step S302, a 256x1024 strip patch was used to segment the image using a deep learning network model. The deep learning network model used UNet++ or ResUNet segmentation network structure to output the segmentation results and used the LeakyRelu activation function, the expression of which is shown in equation (6): (6) in, It is a slope coefficient less than 1, taken as 0.01; when At that time, LeakyRelu is the same as Relu; when At that time, the output of LeakyReLU is That is, there is output in the negative value range.

Images