Method and system for accurately positioning and cutting soft package battery tab based on machine vision

Abstract

The invention provides a machine vision-based soft package battery tab accurate positioning and cutting method and system, and relates to the technical field of battery manufacturing, and the method comprises the steps: collecting front and side images of a soft package battery, and constructing a tab three-dimensional grid model; using a deep neural network to extract connection interface morphology features, and analyzing stress distribution; identifying a stress concentration area by adopting spectral imaging, and determining a cutting safety area; and finally planning a tool path and generating machining parameters. The optimal cutting position of the tab can be accurately identified, the cutting precision is improved, the deformation and damage risks of the tab are reduced, and the service life of the battery is prolonged.

Description

Technical Field

[0001] This invention relates to the field of battery manufacturing technology, and in particular to a method and system for precise positioning and cutting of tabs in pouch batteries based on machine vision. Background Technology

[0002] Pouch batteries are widely used in consumer electronics and new energy vehicles due to their advantages such as lightweight design and high energy density. In the production process of pouch batteries, the precise cutting of the tabs is a crucial step in ensuring battery performance and safety. As the electrical connection components of pouch batteries, the cutting quality of the tabs directly affects the battery's electrical performance and lifespan. Traditionally, pouch battery tab cutting relied mainly on manual experience or simple mechanical positioning methods. However, with the development of battery technology and the increasing demands for product quality, higher requirements have been placed on the precision and consistency of tab cutting. Summary of the Invention

[0003] This invention provides a machine vision-based method and system for precise positioning and cutting of tabs in pouch batteries, which can solve the problems in the prior art.

[0004] A first aspect of the present invention provides a machine vision-based method for precise positioning and cutting of tabs in a pouch battery, comprising: Acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, and determine the position coordinates of the pouch battery tabs in the image coordinate system; perform feature point matching on the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The three-dimensional model is input into a deep neural network to extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery. Based on the morphological feature parameters, the stress distribution of the tab during the cutting process is analyzed to generate a bending deformation risk coefficient. A spectral light source is projected onto the surface of the tab of the pouch battery, and spectral images at different wavelengths are acquired. The light intensity response distribution map of the spectral images is calculated. The stress concentration area on the surface of the pouch battery tab is identified based on the light intensity difference. The coordinate information of the stress concentration area is combined with the bending deformation risk coefficient to determine the cutting safety area. Based on the cutting safety area, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed speed and entry angle of the cutting tool at each trajectory point are calculated, and the cutting processing parameters are generated.

[0005] Acquire front and side images of the pouch battery, extract the contour features of the battery tabs, and determine the position coordinates of the battery tabs in the image coordinate system; perform feature point matching on the front and side images, calculate the disparity map, and construct a three-dimensional model of the battery tabs, including: A scale sequence is constructed, and opening and closing operations are performed on the frontal image and the side image at each scale. The opening and closing operation results at each scale are weighted and combined to obtain a morphological gradient map. A pyramid image sequence is constructed based on the morphological gradient map. SIFT feature points are extracted from the top layer image of the pyramid, and a feature point matching correspondence is established between the frontal image and the side image. The feature point matching correspondence obtained in the previous layer is used as the matching constraint for the next layer. Census transformation cost and mutual information cost are calculated in each layer to generate disparity estimates for each layer of the pyramid. The disparity estimates of each layer of the pyramid are fused and optimized to generate a disparity map. Based on the disparity map, the three-dimensional point cloud data of the tabs of the pouch battery is obtained. The moving least squares method is used to fit a local surface to the neighborhood of each point in the three-dimensional point cloud data. The principal curvature and Gaussian curvature of the local surface are calculated, and points whose principal curvature or Gaussian curvature exceeds the curvature threshold are marked as outliers. Calculate the normal vector of each non-abnormal point in the three-dimensional point cloud data, delete the connecting edges between adjacent point pairs whose normal vector angle is greater than the angle threshold, and perform meshing processing on the three-dimensional point cloud data after deleting the connecting edges to generate the three-dimensional mesh model of the tab of the soft pack battery.

[0006] The feature point matching correspondence obtained from the previous layer is used as the matching constraint for the next layer. A constraint penalty term is constructed, and the Census transformation cost and mutual information cost are calculated at each layer to generate disparity estimates for each layer of the pyramid, including: The input frontal and side images are downsampled multiple times to generate a multi-layer pyramid image sequence; In the front and side images of the nth layer of the pyramid, select the pixel to be matched, take a fixed-size neighborhood window centered on the pixel, calculate the size relationship between each pixel in the window and the center pixel to obtain the binary code, and perform an XOR operation on the binary codes of the corresponding windows of the front and side images to obtain the Census transform cost; count the joint occurrence frequency and individual occurrence frequency of the pixel gray values ​​in the corresponding windows of the front and side images, and calculate the logarithmic ratio of the joint frequency to the edge frequency to obtain the mutual information cost. Multiply the disparity value of the nth layer pixel by a scaling factor to obtain the baseline value. Expand the dynamic search range around the baseline value to obtain the search window. The search range is dynamically adjusted exponentially with the number of pyramid layers. This window limits the disparity search range of the corresponding pixel in the (n-1)th layer. For each candidate disparity value in the search window, calculate its deviation from the baseline value. Take the smaller value between the deviation value and the disparity deviation tolerance as the constraint penalty term. The Census transform cost, mutual information cost, and constraint penalty term are weighted and summed to obtain the comprehensive matching cost. Based on the comprehensive matching cost, the disparity value with the minimum cost is selected in the search window as the disparity estimation result for that pixel. This result is used as the constraint condition for the next layer of disparity estimation. All disparity estimations are completed layer by layer from top to bottom, and the dynamic weight coefficients of each layer are adaptively adjusted according to the image texture features.

[0007] Based on the morphological characteristic parameters, the stress distribution of the pouch battery tab during the cutting process is analyzed to generate a bending deformation risk coefficient, including: Based on the three-dimensional mesh model, the three-dimensional topographic coordinate point cloud of the tab connection interface of the soft-pack battery is obtained, the interface contour curve is extracted, and the local curvature distribution and thickness distribution are calculated based on the interface contour curve; the topographic feature parameters are decomposed into basis function components of different feature scales to generate a multi-scale interface topographic feature matrix. The stress distribution is obtained by multiplying the multi-scale interface morphology feature matrix with the elastic stiffness tensor. The acceleration of each point on the interface contour curve is calculated based on the stress distribution. The stress concentration factor is calculated by combining the local curvature distribution and the thickness distribution. The ratio of the stress concentration factor to the preset critical strain threshold is used as the local damage degree. The static risk value is obtained by dividing the number of regions where the local damage exceeds the critical value by the total number of regions. The dynamic risk value is obtained by calculating the proportion of time during which the stress change rate exceeds the critical value. The bending deformation risk coefficient is obtained by weighting the static risk value and the dynamic risk value.

[0008] Based on the difference in light intensity, stress concentration areas on the surface of the pouch battery tabs are identified. The coordinate information of these stress concentration areas is combined with the bending deformation risk coefficient to determine the safe cutting area, including: The square of the difference between the light intensity response distribution map and the preset reference light intensity distribution map is multiplied by the sensitivity coefficient of the corresponding wavelength and summed to generate a stress concentration measurement distribution map. The stress concentration measurement distribution map is input into a pre-trained graph convolutional neural network to extract local features. The feature weights are calculated based on the attention mechanism network. Local features with feature weights higher than the weight threshold are combined into a stress evolution feature set. Based on the stress evolution feature set, the processing efficiency under different process parameters is calculated; the product of stress concentration coefficient and bending deformation risk coefficient is used as the evaluation index. Within the range where the evaluation index is lower than the stability threshold, the process parameter combination with the highest processing efficiency is selected as the optimal parameter combination. The instantaneous position of the cutting tool is calculated through the optimal parameter combination, and the position coordinates of the cutting tool are continuously collected to generate the optimal cutting path sequence. Based on the optimal cutting path, the stress gradient distribution of adjacent regions is calculated, and regions with stress gradient values ​​lower than the critical gradient threshold and local failure probability lower than the safety threshold are selected as cutting safety regions.

[0009] Based on the stated cutting safety zone, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed rate and entry angle of the cutting tool at each trajectory point are calculated, and cutting parameters are generated, including: Obtain the set of boundary points of the cut safety region, calculate the region area and centroid coordinates based on the set of boundary points; divide the stress gradient value of each point in the set of boundary points by the stress failure threshold and take the negative exponent to generate a stress gradient weight distribution map; generate candidate cutting points within the safety region based on the weight distribution map. For each candidate entry point, the local curvature value is calculated on a preset length scale sequence; the variance of the local curvature value is combined with the distance from the candidate entry point to the centroid and its stress gradient weight to obtain the comprehensive evaluation score of the candidate entry point. The comprehensive evaluation scores of all candidate entry points are input into the particle swarm optimizer, and the particle positions are iteratively updated until convergence to obtain the optimal entry point position. Starting from the optimal entry point, the B-spline control point sequence that satisfies the curvature continuity constraint is calculated, and a smooth trajectory point sequence is generated from the control point sequence. The smooth stress gradient weight and smooth curvature value of each trajectory point are calculated along the smooth trajectory point sequence; the sum of the smooth stress gradient weight and the smooth curvature value is superimposed on the process reference angle after being subjected to an adaptive low-pass filter to obtain the real-time cutting angle sequence of each trajectory point; The radius of curvature of each trajectory point is calculated based on the smooth trajectory point sequence. The square root of the product of the radius of curvature of each point and the maximum tangential acceleration of the machining system is used to obtain the local velocity limit value. This limit value is then multiplied by the smooth stress gradient weight to obtain the feed rate curve, thus achieving a smooth speed transition.

[0010] A second aspect of the present invention provides a machine vision-based system for precise positioning and cutting of tabs in a pouch battery, comprising: The first unit is used to acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, determine the position coordinates of the pouch battery tabs in the image coordinate system, match feature points of the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The second unit is used to input the three-dimensional model into a deep neural network, extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery, analyze the stress distribution of the tab during the cutting process based on the morphological feature parameters, and generate a bending deformation risk coefficient. The third unit is used to project a spectral light source onto the surface of the tab of the pouch battery, acquire spectral images at different wavelengths, calculate the light intensity response distribution map of the spectral images, identify the stress concentration area on the surface of the tab of the pouch battery based on the light intensity difference, and combine the coordinate information of the stress concentration area with the bending deformation risk coefficient to determine the cutting safety area; The fourth unit is used to plan the movement trajectory of the cutting tool according to the cutting safety area, determine the entry point position, calculate the feed speed and entry angle of the cutting tool at each trajectory point, and generate cutting processing parameters.

[0011] A third aspect of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0012] Fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0013] The beneficial effects of this application are as follows: This invention achieves precise positioning of the tabs by acquiring images of the front and side of the pouch battery, extracting the tab contour features, and constructing a three-dimensional mesh model. This avoids the errors of traditional two-dimensional positioning methods and improves cutting accuracy and efficiency.

[0014] This invention combines deep neural network analysis and spectral imaging technology to accurately identify the stress distribution and deformation risk at the interface between the tab and the battery body, scientifically determine the safe cutting area, and effectively prevent quality problems such as tab breakage, bending, or internal structural damage caused during the cutting process.

[0015] This invention automatically plans the movement trajectory and processing parameters of the cutting tool based on the cutting safety area, realizing intelligent control of the cutting process, reducing reliance on manual operation, reducing processing errors, and improving production efficiency and product consistency, thus having good industrial application value. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the machine vision-based method for precise positioning and cutting of tabs in a pouch battery according to an embodiment of the present invention. Figure 2 A flowchart illustrating the method for constructing a 3D model of the tabs for a pouch battery. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the machine vision-based method for precise positioning and cutting of tabs in a pouch battery according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, and determine the position coordinates of the pouch battery tabs in the image coordinate system; perform feature point matching on the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The three-dimensional model is input into a deep neural network to extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery. Based on the morphological feature parameters, the stress distribution of the tab during the cutting process is analyzed to generate a bending deformation risk coefficient. A spectral light source is projected onto the surface of the tab of the pouch battery, and spectral images at different wavelengths are acquired. The light intensity response distribution map of the spectral images is calculated. The stress concentration area on the surface of the pouch battery tab is identified based on the light intensity difference. The coordinate information of the stress concentration area is combined with the bending deformation risk coefficient to determine the cutting safety area. Based on the cutting safety area, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed speed and entry angle of the cutting tool at each trajectory point are calculated, and the cutting processing parameters are generated.

[0020] Figure 2 A flowchart of a method for constructing a 3D model of a pouch battery tab. In one optional embodiment, front and side images of the pouch battery are acquired, the contour features of the pouch battery tab are extracted, and the position coordinates of the pouch battery tab in the image coordinate system are determined; feature point matching is performed on the front and side images, a disparity map is calculated, and a 3D model of the pouch battery tab is constructed, including: A scale sequence is constructed, and opening and closing operations are performed on the frontal image and the side image at each scale. The opening and closing operation results at each scale are weighted and combined to obtain a morphological gradient map. A pyramid image sequence is constructed based on the morphological gradient map. SIFT feature points are extracted from the top layer image of the pyramid, and a feature point matching correspondence is established between the frontal image and the side image. The feature point matching correspondence obtained in the previous layer is used as the matching constraint for the next layer. Census transformation cost and mutual information cost are calculated in each layer to generate disparity estimates for each layer of the pyramid. The disparity estimates of each layer of the pyramid are fused and optimized to generate a disparity map. Based on the disparity map, the three-dimensional point cloud data of the tabs of the pouch battery is obtained. The moving least squares method is used to fit a local surface to the neighborhood of each point in the three-dimensional point cloud data. The principal curvature and Gaussian curvature of the local surface are calculated, and points whose principal curvature or Gaussian curvature exceeds the curvature threshold are marked as outliers. Calculate the normal vector of each non-abnormal point in the three-dimensional point cloud data, delete the connecting edges between adjacent point pairs whose normal vector angle is greater than the angle threshold, and perform meshing processing on the three-dimensional point cloud data after deleting the connecting edges to generate the three-dimensional mesh model of the tab of the soft pack battery.

[0021] In one embodiment, the present invention discloses a method for constructing a three-dimensional model of the tabs of a pouch battery. This method requires acquiring front and side images of the pouch battery, extracting the contour features of the tabs using image processing techniques, and determining the position coordinates of the tabs in the image coordinate system.

[0022] When acquiring images, an industrial camera with a resolution of 1920×1080 pixels can be used. The acquisition distance should be maintained at approximately 30 cm, and the lighting conditions should be kept uniform to avoid overexposure due to strong light reflection. After acquisition, the images are preprocessed, including grayscale conversion and histogram equalization, to enhance image contrast. The Canny edge detection algorithm is used to extract the contour features of the electrodes, with a low threshold of 50 and a high threshold of 150, which effectively separates the electrodes from the background. The position coordinates of the electrodes in the image coordinate system are calculated through moment analysis of the contour point set.

[0023] Five different scale values ​​(1, 3, 5, 7, 9 pixels) were selected to form a scale sequence. At each scale, opening and closing operations were performed on the frontal and side images, respectively. For example, for a scale value of 3, an opening operation was performed using a circular structuring element with a radius of 3 pixels to obtain the result after removing bright details smaller than that scale; similarly, a closing operation was performed using a circular structuring element with a radius of 3 pixels to obtain the result after filling in dark details smaller than that scale. The opening and closing operation results at all scales were weighted and combined, with the weight of smaller scales set to 0.5, the weight of medium scales set to 0.3, and the weight of larger scales set to 0.2, finally obtaining a morphological gradient map that highlights the edge features of the tabs.

[0024] A pyramid image sequence was constructed based on morphological gradient maps. A four-layer image pyramid was built, with a resolution ratio of 2:1 between adjacent layers. SIFT feature points were extracted from the top layer (lowest resolution) image of the pyramid. Approximately 200 feature points were extracted from the frontal image and approximately 180 feature points were extracted from the side image. The feature descriptor dimension was 128. Euclidean distance was used for feature point matching, and a distance ratio threshold of 0.7 was set. This established a correspondence between feature points in the frontal and side images, with an initial matching pair of approximately 60 pairs.

[0025] The feature point matching correspondences obtained from the previous layer are used as matching constraints for the next layer. Census transform cost and mutual information cost are calculated at each layer. The Census transform uses a 9×7 window size to generate a 63-bit Census descriptor. The Census transform cost is calculated using Hamming distance, and the mutual information cost is obtained by calculating the joint probability distribution of pixel grayscale values. The two costs are fused at a ratio of 0.6:0.4 to generate a comprehensive matching cost. A semi-global matching algorithm is used to optimize the matching cost, with 8 path directions, a smoothness penalty coefficient P1 of 10, and a discontinuity penalty coefficient P2 of 30. In this way, disparity estimates for each layer of the pyramid are generated.

[0026] The disparity estimates for each layer of the pyramid are fused and optimized. A coarse-to-fine strategy is adopted, upsampling the disparity estimates of the upper layers as the initial values ​​for the disparity estimates of the lower layers, and performing local optimization within a range of ±2 pixels. Median filtering (window size 5×5) and bilateral filtering (spatial domain standard deviation of 3, value domain standard deviation of 10) are used to smooth the disparity map and remove noise. The final generated disparity map has the same resolution as the input image, and the disparity accuracy can reach 0.1 pixels.

[0027] Based on the disparity map and camera calibration parameters (focal length 8mm, baseline length 12cm), two-dimensional pixel coordinates and disparity values ​​were converted into three-dimensional point cloud data using triangulation principles. The acquired point cloud data contains approximately 15,000 points, covering the surface area of ​​the electrode.

[0028] Moving least squares was used to fit a local surface to the neighborhood of each point in the 3D point cloud data. A neighborhood radius of 2 mm was selected, containing approximately 30 neighboring points. A quadratic polynomial was used to fit the neighborhood of each point, calculating the principal curvature and Gaussian curvature of the local surface. A principal curvature threshold of ±0.5 mm was set. -1 The Gaussian curvature threshold is ±0.2mm. -2 Points exceeding a threshold are marked as outliers. In actual processing, 3% of the points are marked as outliers and removed from the point cloud.

[0029] Calculate the normal vector of each non-outlier point in the 3D point cloud data. Estimate the normal vector using principal component analysis with points in the neighborhood, with a neighborhood radius of 1.5 mm. Calculate the angle between the normal vectors of adjacent point pairs, setting an angle threshold of 45 degrees. Delete the connecting edges between adjacent point pairs whose normal vector angle is greater than the threshold. On average, each point has approximately 8 connecting edges; after edge deletion, 85% of the connecting edges are retained.

[0030] The 3D point cloud data after removing connected edges was meshed. A Poisson surface reconstruction algorithm was used, with an octree depth of 8 and a reconstruction weight of 4, to generate a 3D mesh model of the tabs of the pouch battery. The final generated 3D mesh model contains approximately 25,000 faces, with a reconstruction accuracy better than 0.1 mm, fully preserving the geometric features and surface details of the tabs.

[0031] In one optional implementation, the feature point matching correspondence obtained in the previous layer is used as the matching constraint for the next layer, a constraint penalty term is constructed, and the Census transformation cost and mutual information cost are calculated in each layer to generate disparity estimates for each layer of the pyramid, including: The input frontal and side images are downsampled multiple times to generate a multi-layer pyramid image sequence; In the front and side images of the nth layer of the pyramid, select the pixel to be matched, take a fixed-size neighborhood window centered on the pixel, calculate the size relationship between each pixel in the window and the center pixel to obtain the binary code, and perform an XOR operation on the binary codes of the corresponding windows of the front and side images to obtain the Census transform cost; count the joint occurrence frequency and individual occurrence frequency of the pixel gray values ​​in the corresponding windows of the front and side images, and calculate the logarithmic ratio of the joint frequency to the edge frequency to obtain the mutual information cost. Multiply the disparity value of the nth layer pixel by a scaling factor to obtain the baseline value. Expand the dynamic search range around the baseline value to obtain the search window. The search range is dynamically adjusted exponentially with the number of pyramid layers. This window limits the disparity search range of the corresponding pixel in the (n-1)th layer. For each candidate disparity value in the search window, calculate its deviation from the baseline value. Take the smaller value between the deviation value and the disparity deviation tolerance as the constraint penalty term. The Census transform cost, mutual information cost, and constraint penalty term are weighted and summed to obtain the comprehensive matching cost. Based on the comprehensive matching cost, the disparity value with the minimum cost is selected in the search window as the disparity estimation result for that pixel. This result is used as the constraint condition for the next layer of disparity estimation. All disparity estimations are completed layer by layer from top to bottom, and the dynamic weight coefficients of each layer are adaptively adjusted according to the image texture features.

[0032] This embodiment provides a stereo matching disparity estimation method based on pyramid constraints. This method constructs an image pyramid structure, uses the feature point matching correspondence obtained from the previous layer as the matching constraint for the next layer, calculates the Census transform cost and mutual information cost at each layer, and finally completes the full disparity estimation layer by layer from top to bottom.

[0033] Multiple downsampling operations are performed on the input frontal and side images to generate a multi-layered pyramid image sequence. Gaussian blur and quadratic sampling are used to successively reduce the original image (layer 0) to 1 / 2, 1 / 4, 1 / 8 of its original size, thus constructing a four-layer image pyramid structure. For example, for an original image with a resolution of 1920×1080, the first layer image size is 960×540, the second layer is 480×270, and the third layer is 240×135.

[0034] Stereo matching begins at the top of the pyramid (layer 3). In the frontal and side images of layer 3, for each pixel (x, y) to be matched, a 9×7 rectangular window is taken centered on that pixel. When calculating the Census transform, the grayscale value of each pixel within the window is compared with the center pixel: if the grayscale value of surrounding pixels is greater than that of the center pixel, it is encoded as 1; otherwise, it is encoded as 0. Thus, for a 9×7 window (excluding the center pixel), a 62-bit binary code is obtained. The codes of the corresponding windows in the frontal and side images are XORed, and the number of 1s in the XOR result (Hamming distance) is counted as the Census transform cost. For example, if a window in the frontal image is encoded as "1001...0" and the corresponding window in the side image is encoded as "1101...1", the XOR result is "0100...1", which contains 25 "1s", so the Census transform cost is 25.

[0035] Simultaneously, the mutual information cost is calculated by statistically analyzing the joint frequency and individual frequency of pixel grayscale values ​​within the corresponding windows of the frontal and side images. For example, within a 9×7 window, a pixel with a grayscale value of 128 appears 8 times in the frontal image (frequency 8 / 63) and 10 times in the side image (frequency 10 / 63). A pixel with a grayscale value of 128 appears 5 times in both the frontal and side images, with a joint frequency of 5 / 63. Therefore, the corresponding mutual information component is 5 / 63 * log(5 / 63 ÷ (8 / 63 * 10 / 63)). The mutual information cost is obtained by summing the mutual information components for all grayscale value combinations within the window.

[0036] In the third layer of the pyramid, since there are no constraints from the previous layer, a full-range search is performed on each pixel within a fixed disparity range (e.g., 0-15 pixels). The disparity value with the minimum weighted sum of Census transform cost and mutual information cost is selected as the disparity estimation result for this layer. The weighting coefficients can be set to 0.7 and 0.3, and dynamically adjusted according to the richness of image texture.

[0037] When entering the second layer of disparity estimation, for pixel (x', y'), determine the disparity value d3 of its corresponding pixel (x / 2, y / 2) in the third layer. Multiply d3 by the scaling factor 2 (doubling the resolution) to obtain the baseline disparity value d. base = 2d3. In d base Expand the dynamic search range left and right to obtain the search window [d base -△d2, d base +△d2], where △d2=4 (the dynamic search range of the second layer). For example, if the disparity value of a point in the third layer is 5, then the reference disparity of the corresponding point in the second layer is 10, and the search window is [6,14].

[0038] For each candidate disparity value d within the search window cand Calculate the constraint penalty term. Calculate d. cand With d base deviation |d cand- d base |, taking the smaller value between this deviation value and the parallax deviation tolerance t2=3, we obtain the constraint penalty term p=min(|d cand -d base |,t2). For example, for a candidate disparity value of 13, the constraint penalty term p=min(|13-10|,3)=3.

[0039] The Census transformation cost c census Mutual information cost c mi We perform a weighted summation of the constraint penalty term p to obtain the comprehensive matching cost c. total = w1*c census + w2*cmi + w3*p, where w1, w2, and w3 are 0.5, 0.3, and 0.2, respectively. Within the search window, the disparity value with the lowest overall matching cost is selected as the disparity estimate for that pixel.

[0040] The same approach is used to process layers 1 and 0 (original resolution), but the dynamic search ranges are set to Δd1=2 and Δd0=1 respectively, the disparity tolerances to t1=2 and t0=1 respectively, and the weighting coefficients are adjusted accordingly to place greater emphasis on the Census transformation cost and mutual information cost, while reducing the weight of the constraint penalty term. For example, the weights for layer 0 can be set to w1=0.6, w2=0.35, and w3=0.05.

[0041] This top-down pyramid-constrained stereo matching method ensures both the accuracy of global disparity estimation and improves algorithm efficiency. Experimental results show that, on typical stereo image pairs, this method can control the matching error rate to below 5%, and the average processing time is reduced by more than 80% compared to traditional full-range search.

[0042] In one optional implementation, the stress distribution of the pouch cell tab during the cutting process is analyzed based on the morphological characteristic parameters to generate a bending deformation risk coefficient, including: Based on the three-dimensional mesh model, the three-dimensional topographic coordinate point cloud of the tab connection interface of the soft-pack battery is obtained, the interface contour curve is extracted, and the local curvature distribution and thickness distribution are calculated based on the interface contour curve; the topographic feature parameters are decomposed into basis function components of different feature scales to generate a multi-scale interface topographic feature matrix. The stress distribution is obtained by multiplying the multi-scale interface morphology feature matrix with the elastic stiffness tensor. The acceleration of each point on the interface contour curve is calculated based on the stress distribution. The stress concentration factor is calculated by combining the local curvature distribution and the thickness distribution. The ratio of the stress concentration factor to the preset critical strain threshold is used as the local damage degree. The static risk value is obtained by dividing the number of regions where the local damage exceeds the critical value by the total number of regions. The dynamic risk value is obtained by calculating the proportion of time during which the stress change rate exceeds the critical value. The bending deformation risk coefficient is obtained by weighting the static risk value and the dynamic risk value.

[0043] In a specific embodiment, based on the acquired 3D mesh model, a laser confocal microscope is used to scan the interface of the pouch battery tabs to obtain 3D topographic coordinate point cloud data. The point cloud data density is set to no less than 1000 sampling points per square millimeter to ensure accurate capture of topographic features. The acquired point cloud data is processed to extract the interface contour curve. The contour curve is fitted using a B-spline curve with 35 control points and a fitting accuracy controlled within ±2 micrometers. Based on the fitted interface contour curve, the local curvature distribution is calculated using the finite difference method with a sampling interval of 0.05 mm. For the thickness distribution, measurements are taken every 0.1 mm on the contour curve to form a thickness distribution dataset.

[0044] The acquired morphological feature parameters are decomposed into basis function components at different feature scales. Wavelet transform is employed, with the Daubechies wavelet basis function selected. The decomposition level is set to four levels, corresponding to microscale (0.01-0.1 mm), mesoscale (0.1-0.5 mm), macroscale (0.5-2 mm), and global scale (>2 mm). Amplitude, frequency, and phase features are extracted from the components at each scale to form feature vectors, thereby constructing a multi-scale interface morphological feature matrix. For example, for a certain tab connection interface, the surface roughness Ra is 1.8 μm at the microscale, the waviness is 5.3 μm at the mesoscale, and the flatness is 12.5 μm at the macroscale.

[0045] The stress distribution was calculated by multiplying the multi-scale interface morphology feature matrix with the elastic stiffness tensor of the pouch battery tab material. The elastic stiffness tensor was constructed based on the material's Young's modulus, Poisson's ratio, and shear modulus. For aluminum foil tabs, the Young's modulus was taken as 69 GPa and the Poisson's ratio as 0.33; for copper foil tabs, the Young's modulus was taken as 110 GPa and the Poisson's ratio as 0.34. The calculated stress distribution exhibited non-uniform characteristics, with stress concentration often occurring at the points of concavity and convexity of the interface profile curve. Based on the stress distribution, the acceleration at each point on the interface profile curve was calculated. Using the finite element method (FEM), with an element size of 0.02 mm and a time step of 0.01 ms, the cutting process was dynamically simulated.

[0046] The stress concentration factor is calculated by combining the local curvature and thickness distributions. The calculation of the stress concentration factor considers the combined effects of the radius of curvature and local thickness. For areas with a radius of curvature less than 0.1 mm, the stress concentration factor increases significantly. Taking a practical example, in a recessed area at the connection between the electrode tab and the battery cell, the radius of curvature is 0.08 mm and the local thickness is 0.09 mm. The calculated stress concentration factor is 2.7, while the stress concentration factor in a flat area is only 1.1. The ratio of the stress concentration factor to a preset critical strain threshold is used as the local damage degree. The critical strain threshold is set according to the material properties: 1.8% for aluminum foil and 2.5% for copper foil.

[0047] When assessing risk, the contour curve is divided into 200 equal-length regions. The number of regions where the local damage exceeds the critical damage value (set to 0.85) is counted, and this number is divided by the total number of regions to obtain the static risk value. For example, if 35 regions on a tab interface have a local damage value exceeding 0.85, the static risk value is 35 / 200 = 0.175. The dynamic risk value is calculated based on the stress change rate. The time period during the cutting process where the stress change rate exceeds the critical stress change rate value (set to 500 MPa / ms) is recorded, and the proportion of this period to the total cutting time is used as the dynamic risk value. For example, in a 2-millisecond cutting process, if the stress change rate exceeds the critical value for 0.5 milliseconds, the dynamic risk value is 0.5 / 2 = 0.25.

[0048] The final bending deformation risk coefficient is obtained by weighting the static and dynamic risk values. The weighting coefficients are determined based on the characteristics of the cutting process. For high-speed cutting processes, the static risk value has a weight of 0.4, and the dynamic risk value has a weight of 0.6; for low-speed cutting processes, the static risk value has a weight of 0.6, and the dynamic risk value has a weight of 0.4. For example, for a certain high-speed cutting process, the static risk value is 0.175, and the dynamic risk value is 0.25. The calculated bending deformation risk coefficient is 0.175 × 0.4 + 0.25 × 0.6 = 0.22. The grading standard for the bending deformation risk coefficient is: low risk (<0.2), medium risk (0.2-0.5), and high risk (>0.5). Therefore, this tab belongs to the medium risk level during the cutting process, and process parameter optimization or protective measures are required.

[0049] The above methods can accurately assess the stress distribution of the tabs in the cutting process of pouch batteries, predict the risk of bending deformation, provide quantitative basis for tab design and process optimization, and improve the production yield and safety of pouch batteries.

[0050] In one optional implementation, stress concentration areas on the surface of the pouch battery tabs are identified based on light intensity differences. The coordinate information of these stress concentration areas is combined with the bending deformation risk coefficient to determine a safe cutting area, including: The square of the difference between the light intensity response distribution map and the preset reference light intensity distribution map is multiplied by the sensitivity coefficient of the corresponding wavelength and summed to generate a stress concentration measurement distribution map. The stress concentration measurement distribution map is input into a pre-trained graph convolutional neural network to extract local features. The feature weights are calculated based on the attention mechanism network. Local features with feature weights higher than the weight threshold are combined into a stress evolution feature set. Based on the stress evolution feature set, the processing efficiency under different process parameters is calculated; the product of stress concentration coefficient and bending deformation risk coefficient is used as the evaluation index. Within the range where the evaluation index is lower than the stability threshold, the process parameter combination with the highest processing efficiency is selected as the optimal parameter combination. The instantaneous position of the cutting tool is calculated through the optimal parameter combination, and the position coordinates of the cutting tool are continuously collected to generate the optimal cutting path sequence. Based on the optimal cutting path, the stress gradient distribution of adjacent regions is calculated, and regions with stress gradient values ​​lower than the critical gradient threshold and local failure probability lower than the safety threshold are selected as cutting safety regions.

[0051] This embodiment provides a method for identifying stress concentration areas on the surface of a pouch cell battery tab and determining a safe cutting area based on light intensity differences. The method involves obtaining a light intensity response distribution map of the pouch cell battery tab, identifying stress concentration areas through light intensity difference analysis, and determining the optimal safe cutting area by combining this with a bending deformation risk coefficient.

[0052] In this embodiment, the light intensity response distribution map of the pouch battery tab surface can be obtained using a multispectral imaging device. Specifically, a multispectral camera with a wavelength range of 350 nm to 850 nm is used to acquire reflectance spectrum images of the pouch battery tab surface at 10 nm intervals. During the acquisition process, the incident angle of the light source is maintained at 45 degrees, the camera is placed perpendicular to the sample surface, the acquisition distance is 300 mm, and the exposure time is set to 20 ms. After acquisition, the images at each wavelength are standardized to generate a light intensity response distribution map matrix.

[0053] When identifying stress concentration regions, a difference analysis is performed between the acquired light intensity response distribution map and a preset reference light intensity distribution map. The reference light intensity distribution map is obtained from a standard tab sample under stress-free conditions. The square of the light intensity difference at each wavelength for each pixel is calculated and multiplied by the sensitivity coefficient for the corresponding wavelength. For example, the sensitivity coefficient for 550nm is 0.85, for 650nm it is 0.92, and for 750nm it is 0.78. The weighted differences at each wavelength are summed to generate a stress concentration measurement distribution map.

[0054] The generated stress concentration metric distribution map is input into a pre-trained graph convolutional neural network for feature extraction. This network contains four graph convolutional layers, each with 64, 128, 256, and 512 convolutional kernels respectively. The kernel size is 3×3, the stride is 1, and the padding is 1. The local feature matrix extracted by the graph convolutional network is 512×H×W, where H and W are the height and width of the image, respectively. Feature weights are calculated using an attention mechanism network, which combines channel attention and spatial attention, outputting the weight coefficient for each feature. Local features with weights higher than 0.75 are selected to form a stress evolution feature set.

[0055] Based on the stress evolution feature set, the processing efficiency under different process parameters is calculated. Process parameters include cutting speed, cutting pressure, and cutting angle. In this embodiment, the cutting speed is set from 5 mm / s to 20 mm / s in 5 mm / s intervals; the cutting pressure is set from 0.5 MPa to 2 MPa in 0.5 MPa intervals; and the cutting angle is set from 0 degrees to 90 degrees in 15-degree intervals. For each parameter combination, the cutting area per unit time is calculated as the processing efficiency index.

[0056] The product of the stress concentration factor and the bending deformation risk factor is used as the evaluation index. The stress concentration factor is extracted from the stress concentration distribution map, and its value ranges from 0 to 1, with a higher value indicating a higher degree of stress concentration. The bending deformation risk factor is calculated based on the material properties and geometry of the pouch battery tabs, taking into account factors such as the thickness, width, and elastic modulus of the tabs. In practical applications, the bending deformation risk factor for aluminum tabs is approximately 0.65, and for copper tabs it is approximately 0.82.

[0057] Within the range where the evaluation index is below the stability threshold of 0.3, the combination of process parameters with the highest processing efficiency is selected as the optimal parameter combination. Taking a certain type of soft-pack battery as an example, after calculation and comparison, the optimal parameter combination is: cutting speed 15mm / s, cutting pressure 1.5MPa, cutting angle 45 degrees. At this time, the processing efficiency reaches 225mm² / s, and the evaluation index is 0.28.

[0058] The instantaneous position of the cutting tool is calculated using the optimal parameter combination. The initial position of the cutting tool is set at the starting point of the tab edge, with coordinates (0, 0). Based on the cutting speed and direction, the position coordinates of the cutting tool are collected every 10ms to generate the optimal cutting path sequence. For example, in one experiment, some key coordinate points collected were (0, 0), (1.5, 1.5), (3.0, 3.0), and (4.5, 4.5), forming a smooth cutting path from the tab edge to the center region.

[0059] Based on the optimal cutting path, the stress gradient distribution of adjacent regions is calculated. The stress gradient value represents the magnitude of stress change per unit distance, and is calculated by dividing the difference in stress concentration measurement between adjacent pixels by the pixel spacing. Regions with stress gradient values ​​below the critical gradient threshold of 0.05 MPa / mm and local failure probability below the safety threshold of 0.01 are selected as the cutting safety region.

[0060] Using the above method, in the test of 100 pouch battery tabs, the overlap rate between the identified stress concentration area and the actual failure area reached 95.2%. The determined cutting safety area reduced the tab breakage rate of the battery in the cycle charge and discharge test from the original 7.5% to 0.8%, which greatly improved the safety and reliability of pouch batteries.

[0061] In one optional implementation, based on the cutting safety zone, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed rate and entry angle of the cutting tool at each trajectory point are calculated, and cutting processing parameters are generated, including: Obtain the set of boundary points of the cut safety region, calculate the region area and centroid coordinates based on the set of boundary points; divide the stress gradient value of each point in the set of boundary points by the stress failure threshold and take the negative exponent to generate a stress gradient weight distribution map; generate candidate cutting points within the safety region based on the weight distribution map. For each candidate entry point, the local curvature value is calculated on a preset length scale sequence; the variance of the local curvature value is combined with the distance from the candidate entry point to the centroid and its stress gradient weight to obtain the comprehensive evaluation score of the candidate entry point. The comprehensive evaluation scores of all candidate entry points are input into the particle swarm optimizer, and the particle positions are iteratively updated until convergence to obtain the optimal entry point position. Starting from the optimal entry point, the B-spline control point sequence that satisfies the curvature continuity constraint is calculated, and a smooth trajectory point sequence is generated from the control point sequence. The smooth stress gradient weight and smooth curvature value of each trajectory point are calculated along the smooth trajectory point sequence; the sum of the smooth stress gradient weight and the smooth curvature value is superimposed on the process reference angle after being subjected to an adaptive low-pass filter to obtain the real-time cutting angle sequence of each trajectory point; The radius of curvature of each trajectory point is calculated based on the smooth trajectory point sequence. The square root of the product of the radius of curvature of each point and the maximum tangential acceleration of the machining system is used to obtain the local velocity limit value. This limit value is then multiplied by the smooth stress gradient weight to obtain the feed rate curve, thus achieving a smooth speed transition.

[0062] This embodiment provides a method for planning the movement trajectory of a cutting tool based on a cutting safety zone. The method includes determining the entry point position, calculating the feed rate and entry angle of the cutting tool at each trajectory point, and generating cutting processing parameters.

[0063] After obtaining the set of boundary points of the clipping safety region, calculate the area and centroid coordinates of the safety region. Assume the boundary point set is {(x1,y1), (x2,y2),..., (x...} m ,y mThe area of ​​the region is calculated to be 3250 square millimeters using the polygon area calculation formula, with the centroid coordinates at (45.32, 67.89) millimeters. The stress gradient values ​​at each point in the boundary point set are processed; the stress gradient value at the boundary point (23.5, 42.8) is 78.6 MPa / mm. The stress failure threshold is set to 120 MPa / mm, and its negative exponent value e is calculated. (-(78.6 / 120)) The value 0.52 represents the stress gradient weight at that point. Similarly, this process is applied to all boundary points to generate a stress gradient weight distribution map.

[0064] Based on the stress gradient weight distribution map, the system generates candidate entry points within the safe area. Using a grid sampling method, the safe area is divided into a 10mm × 10mm grid, and a candidate entry point is generated at the center of each grid, resulting in a total of 27 candidate points, such as (35.0, 45.0) and (45.0, 45.0). The system then evaluates each candidate entry point.

[0065] For each candidate entry point, the system calculates local curvature values ​​on a preset length scale sequence. The length scale sequence is set to {5mm, 10mm, 15mm, 20mm}. For the candidate point (35.0, 45.0), curvature values ​​are calculated on these four scales, resulting in {0.021, 0.018, 0.015, 0.012}. The variance of this set of curvature values ​​is calculated to be 0.00001353. The distance from the candidate point to the centroid is 25.16mm, and the stress gradient weight is 0.63. The comprehensive evaluation score is calculated using a weighted summation method, with a curvature variance weight of 0.4, a distance weight of 0.3, and a stress gradient weight of 0.3, resulting in a comprehensive score of 0.782 for the candidate point (35.0, 45.0). The same evaluation is performed on all 27 candidate points, generating a complete comprehensive evaluation score set.

[0066] The comprehensive evaluation scores of all candidate entry points are input into the particle swarm optimizer. The population size is set to 30, the maximum number of iterations to 100, the individual learning factor to 2.0, the swarm learning factor to 2.0, and the inertia weight to decrease linearly from 0.9 to 0.4. 30 particles are initialized, each corresponding to an entry point position. During iteration, each particle updates its velocity and position based on its own best position and the global best position. After 87 iterations, the algorithm converges, yielding the optimal entry point position (37.25, 46.83) with a comprehensive evaluation score of 0.915.

[0067] Starting from the optimal incision point, the system calculates a sequence of B-spline control points that satisfy the curvature continuity constraint. Initial control points are determined around the incision point, and additional control points are added based on the region boundary shape and stress gradient distribution. The resulting control point sequence is {(37.25, 46.83), (42.18, 50.62), (48.75, 53.24), (55.36, 52.91),..., (36.82, 46.51)}, totaling 17 control points. Based on these control points, a smooth trajectory point sequence is generated with a sampling interval of 1 mm, resulting in a trajectory point sequence {(37.25, 46.83), (38.02, 47.36),..., (37.15, 46.77)}, totaling 203 trajectory points.

[0068] The smoothed stress gradient weight and smoothed curvature value are calculated for each trajectory point along the smoothed trajectory point sequence. For example, the smoothed stress gradient weight at trajectory point (42.36, 49.28) is 0.73, and the smoothed curvature value is 0.016. The sum of the smoothed stress gradient weight and the smoothed curvature value is then subjected to an adaptive low-pass filter to obtain a filtered value of 0.68. This value is then superimposed on the process reference angle of 45 degrees to obtain the real-time cut-in angle of this trajectory point as 75.4 degrees. The same calculation is performed on all 203 trajectory points to generate a complete real-time cut-in angle sequence.

[0069] The radius of curvature of each trajectory point is calculated based on the smoothed trajectory point sequence. The curvature value of trajectory point (42.36, 49.28) is 0.016, and its radius of curvature is 62.5 mm. The maximum tangential acceleration of the machining system is set to 5 m / s², and the calculated local velocity limit value for this point is √(62.5 × 5000) = 558.3 ​​mm / s. Multiplying this value by the smoothed stress gradient weight of 0.73, the feed rate for this point is 407.6 mm / s. The same calculation is performed on all trajectory points to generate a complete feed rate curve, achieving a smooth speed transition. Finally, the system outputs a complete trajectory plan containing 203 points, each point containing three parameters: position coordinates, cutting angle, and feed rate.

[0070] This method, by comprehensively considering the geometric characteristics, stress distribution characteristics, and dynamic characteristics of the cutting safety zone, achieves precise planning of the cutting tool's motion trajectory, ensuring the safety and processing quality of the cutting process.

[0071] This invention relates to a machine vision-based system for precisely positioning and cutting tabs of pouch batteries, comprising: The first unit is used to acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, determine the position coordinates of the pouch battery tabs in the image coordinate system, match feature points of the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The second unit is used to input the three-dimensional model into a deep neural network, extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery, analyze the stress distribution of the tab during the cutting process based on the morphological feature parameters, and generate a bending deformation risk coefficient. The third unit is used to project a spectral light source onto the surface of the tab of the pouch battery, acquire spectral images at different wavelengths, calculate the light intensity response distribution map of the spectral images, identify the stress concentration area on the surface of the tab of the pouch battery based on the light intensity difference, and combine the coordinate information of the stress concentration area with the bending deformation risk coefficient to determine the cutting safety area; The fourth unit is used to plan the movement trajectory of the cutting tool according to the cutting safety area, determine the entry point position, calculate the feed speed and entry angle of the cutting tool at each trajectory point, and generate cutting processing parameters.

[0072] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0073] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0074] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A machine vision-based method for precise positioning and cutting of tabs in pouch batteries, characterized in that, include: Acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, and determine the position coordinates of the pouch battery tabs in the image coordinate system; perform feature point matching on the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The three-dimensional model is input into a deep neural network to extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery. Based on the morphological feature parameters, the stress distribution of the tab during the cutting process is analyzed to generate a bending deformation risk coefficient. A spectral light source is projected onto the surface of the tab of the pouch battery, and spectral images at different wavelengths are acquired. The light intensity response distribution map of the spectral images is calculated. The stress concentration area on the surface of the pouch battery tab is identified based on the light intensity difference. The coordinate information of the stress concentration area is combined with the bending deformation risk coefficient to determine the cutting safety area. Based on the cutting safety area, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed speed and entry angle of the cutting tool at each trajectory point are calculated, and the cutting processing parameters are generated.

2. The method according to claim 1, characterized in that, Acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, and determine the position coordinates of the pouch battery tabs in the image coordinate system; The process involves matching feature points between the frontal and side images, calculating a disparity map, and constructing a 3D model of the tabs of the pouch battery, including: A scale sequence is constructed, and opening and closing operations are performed on the frontal image and the side image at each scale. The opening and closing operation results at each scale are weighted and combined to obtain a morphological gradient map. A pyramid image sequence is constructed based on the morphological gradient map. SIFT feature points are extracted from the top layer image of the pyramid, and a feature point matching correspondence is established between the frontal image and the side image. The feature point matching correspondence obtained in the previous layer is used as the matching constraint for the next layer. Census transformation cost and mutual information cost are calculated in each layer to generate disparity estimates for each layer of the pyramid. The disparity estimates of each layer of the pyramid are fused and optimized to generate a disparity map. Based on the disparity map, the three-dimensional point cloud data of the tabs of the pouch battery is obtained. The moving least squares method is used to fit a local surface to the neighborhood of each point in the three-dimensional point cloud data. The principal curvature and Gaussian curvature of the local surface are calculated, and points whose principal curvature or Gaussian curvature exceeds the curvature threshold are marked as outliers. Calculate the normal vector of each non-abnormal point in the three-dimensional point cloud data, delete the connecting edges between adjacent point pairs whose normal vector angle is greater than the angle threshold, and perform meshing processing on the three-dimensional point cloud data after deleting the connecting edges to generate the three-dimensional mesh model of the tab of the soft pack battery.

3. The method according to claim 2, characterized in that, The feature point matching correspondence obtained from the previous layer is used as the matching constraint for the next layer. A constraint penalty term is constructed, and the Census transformation cost and mutual information cost are calculated at each layer to generate disparity estimates for each layer of the pyramid, including: The input frontal and side images are downsampled multiple times to generate a multi-layer pyramid image sequence; In the front and side images of the nth layer of the pyramid, select the pixel to be matched, take a fixed-size neighborhood window centered on the pixel, calculate the size relationship between each pixel in the window and the center pixel to obtain the binary code, and perform an XOR operation on the binary codes of the corresponding windows of the front and side images to obtain the Census transform cost; count the joint occurrence frequency and individual occurrence frequency of the pixel gray values ​​in the corresponding windows of the front and side images, and calculate the logarithmic ratio of the joint frequency to the edge frequency to obtain the mutual information cost. Multiply the disparity value of the nth layer pixel by a scaling factor to obtain the baseline value. Expand the dynamic search range around the baseline value to obtain the search window. The search range is dynamically adjusted exponentially with the number of pyramid layers. This window limits the disparity search range of the corresponding pixel in the (n-1)th layer. For each candidate disparity value in the search window, calculate its deviation from the baseline value. Take the smaller value between the deviation value and the disparity deviation tolerance as the constraint penalty term. The Census transform cost, mutual information cost, and constraint penalty term are weighted and summed to obtain the comprehensive matching cost. Based on the comprehensive matching cost, the disparity value with the minimum cost is selected in the search window as the disparity estimation result for that pixel. This result is used as the constraint condition for the next layer of disparity estimation. All disparity estimations are completed layer by layer from top to bottom, and the dynamic weight coefficients of each layer are adaptively adjusted according to the image texture features.

4. The method according to claim 1, characterized in that, Based on the morphological characteristic parameters, the stress distribution of the pouch battery tab during the cutting process is analyzed to generate a bending deformation risk coefficient, including: Based on the three-dimensional mesh model, the three-dimensional topographic coordinate point cloud of the tab connection interface of the soft-pack battery is obtained, the interface contour curve is extracted, and the local curvature distribution and thickness distribution are calculated based on the interface contour curve; the topographic feature parameters are decomposed into basis function components of different feature scales to generate a multi-scale interface topographic feature matrix. The stress distribution is obtained by multiplying the multi-scale interface morphology feature matrix with the elastic stiffness tensor. The acceleration of each point on the interface contour curve is calculated based on the stress distribution. The stress concentration factor is calculated by combining the local curvature distribution and the thickness distribution. The ratio of the stress concentration factor to the preset critical strain threshold is used as the local damage degree. The static risk value is obtained by dividing the number of regions where the local damage exceeds the critical value by the total number of regions. The dynamic risk value is obtained by calculating the proportion of time during which the stress change rate exceeds the critical value. The bending deformation risk coefficient is obtained by weighting the static risk value and the dynamic risk value.

5. The method according to claim 1, characterized in that, Based on the difference in light intensity, stress concentration areas on the surface of the pouch battery tabs are identified. The coordinate information of these stress concentration areas is combined with the bending deformation risk coefficient to determine the safe cutting area, including: The square of the difference between the light intensity response distribution map and the preset reference light intensity distribution map is multiplied by the sensitivity coefficient of the corresponding wavelength and summed to generate a stress concentration measurement distribution map. The stress concentration measurement distribution map is input into a pre-trained graph convolutional neural network to extract local features. The feature weights are calculated based on the attention mechanism network. Local features with feature weights higher than the weight threshold are combined into a stress evolution feature set. Based on the stress evolution feature set, the processing efficiency under different process parameters is calculated; the product of stress concentration coefficient and bending deformation risk coefficient is used as the evaluation index. Within the range where the evaluation index is lower than the stability threshold, the process parameter combination with the highest processing efficiency is selected as the optimal parameter combination. The instantaneous position of the cutting tool is calculated through the optimal parameter combination, and the position coordinates of the cutting tool are continuously collected to generate the optimal cutting path sequence. Based on the optimal cutting path, the stress gradient distribution of adjacent regions is calculated, and regions with stress gradient values ​​lower than the critical gradient threshold and local failure probability lower than the safety threshold are selected as cutting safety regions.

6. The method according to claim 1, characterized in that, Based on the stated cutting safety zone, the movement trajectory of the cutting tool is planned, the entry point position is determined, the feed rate and entry angle of the cutting tool at each trajectory point are calculated, and cutting parameters are generated, including: Obtain the set of boundary points of the cut safety region, calculate the region area and centroid coordinates based on the set of boundary points; divide the stress gradient value of each point in the set of boundary points by the stress failure threshold and take the negative exponent to generate a stress gradient weight distribution map; generate candidate cutting points within the safety region based on the weight distribution map. For each candidate entry point, the local curvature value is calculated on a preset length scale sequence; the variance of the local curvature value is combined with the distance from the candidate entry point to the centroid and its stress gradient weight to obtain the comprehensive evaluation score of the candidate entry point. The comprehensive evaluation scores of all candidate entry points are input into the particle swarm optimizer, and the particle positions are iteratively updated until convergence to obtain the optimal entry point position. Starting from the optimal entry point, the B-spline control point sequence that satisfies the curvature continuity constraint is calculated, and a smooth trajectory point sequence is generated from the control point sequence. The smooth stress gradient weight and smooth curvature value of each trajectory point are calculated along the smooth trajectory point sequence; the sum of the smooth stress gradient weight and the smooth curvature value is superimposed on the process reference angle after being subjected to an adaptive low-pass filter to obtain the real-time cutting angle sequence of each trajectory point; The radius of curvature of each trajectory point is calculated based on the smooth trajectory point sequence. The square root of the product of the radius of curvature of each point and the maximum tangential acceleration of the machining system is used to obtain the local velocity limit value. This limit value is then multiplied by the smooth stress gradient weight to obtain the feed rate curve, thus achieving a smooth speed transition.

7. A machine vision-based system for precise positioning and cutting of tabs for pouch batteries, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire front and side images of the pouch battery, extract the contour features of the pouch battery tabs, determine the position coordinates of the pouch battery tabs in the image coordinate system, match feature points of the front and side images, calculate the disparity map, and construct a three-dimensional mesh model of the pouch battery tabs. The second unit is used to input the three-dimensional model into a deep neural network, extract the morphological feature parameters of the interface between the tab and the battery body of the soft-pack battery, analyze the stress distribution of the tab during the cutting process based on the morphological feature parameters, and generate a bending deformation risk coefficient. The third unit is used to project a spectral light source onto the surface of the tab of the pouch battery, acquire spectral images at different wavelengths, calculate the light intensity response distribution map of the spectral images, identify the stress concentration area on the surface of the tab of the pouch battery based on the light intensity difference, and combine the coordinate information of the stress concentration area with the bending deformation risk coefficient to determine the cutting safety area; The fourth unit is used to plan the movement trajectory of the cutting tool according to the cutting safety area, determine the entry point position, calculate the feed speed and entry angle of the cutting tool at each trajectory point, and generate cutting processing parameters.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images