A laser cutting depth control method for connecting rod fracture forming

Abstract

This invention belongs to the field of intelligent control technology, and particularly relates to a laser cutting depth control method for connecting rod fracture forming. The method first calibrates the fixed rated fracture force of the cylinder in the fracture forming equipment, collects sample data such as structural images, groove depth parameters, and fracture images of connecting rods of different specifications, generates fracture quality qualification score labels, and optimizes and constructs a training dataset; extracts structural visual features such as wall thickness and aperture of the connecting rod to be processed, constructs a deep learning model for laser groove depth prediction, and completes supervised training; inputting the features into the model outputs target groove depth parameters adapted to the fixed fracture force. This invention adopts a constant fracture force design, eliminating the need to change equipment or adjust the cylinder. Through dual-feature image processing, sample screening, and bidirectional iterative decoding, it improves prediction accuracy, effectively solves quality problems such as fracture path deviation and failure to break, improves production line flexibility and changeover efficiency, reduces costs, and achieves intelligent and standardized control of connecting rod fracture.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology, and in particular relates to a laser cutting depth control method for connecting rod fracture forming. Background Technology

[0002] In existing connecting rod fracture processes, the critical fracture force of connecting rods of different specifications (size, material, wall thickness) varies greatly. It is necessary to adjust the cylinder fracture force accordingly, or even replace the fracture equipment. This results in low production line changeover efficiency, poor flexibility, and high cost. The existing laser grooving depth is mostly based on empirical fixed values, without a quantitative theoretical matching model. It is impossible to achieve universal fracture of connecting rods of different sizes under the constraint of constant cylinder fracture force, and it is prone to quality problems such as failure to break, fracture path deviation, and excessive plastic deformation. Summary of the Invention

[0003] In view of the technical problems existing in the background art, the present invention proposes a laser cutting depth control method for connecting rod fracture forming.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Pre-calibrate the fixed rated breaking force of the driving cylinder of the breaking device to maintain constant output throughout the entire process. Under the constraint of the fixed rated breaking force, collect sample data of connecting rods of different specifications to construct a training dataset. The sample data includes a panoramic image of the connecting rod structure, laser grooving depth control parameters adapted to the fixed rated breaking force of the connecting rod, a fracture image of the connecting rod after breaking, and a breaking quality qualification score label generated based on the fracture image.

[0006] S2. For any specification of connecting rod to be processed, acquire its panoramic structural image, and extract the structural visual features of the connecting rod through image processing and feature extraction.

[0007] S3. Construct a deep learning model for predicting laser grooving depth, using the structural visual features extracted from the panoramic image of the linkage structure as the model input, and the laser grooving depth control parameters adapted to the fixed rated breaking force and the breaking quality qualification score as the model output, to complete the supervised training of the model.

[0008] S4. Input the structural visual features of the connecting rod to be processed into the trained deep learning model for predicting the laser grooving depth, and directly output the target laser grooving depth control parameters that are adapted to the rated breaking force.

[0009] Preferably, the specific implementation of the fracture quality qualification rating label generated based on the fracture image includes:

[0010] S11. After the connecting rod completes the fracture, acquire a grayscale image of the fracture surface facing forward and an image of the fracture path perpendicular to the fracture surface, and perform image processing on the two images in sequence to obtain the processed image.

[0011] S12. Extract continuous fracture contour lines from the processed fracture path image and spatially register them with the preset laser grooving reference line. Using the grooving length direction as the x-axis and the workpiece wall thickness direction perpendicular to the grooving as the y-axis, obtain the path deviation sequence along the entire length of the grooving. And extract the path with the largest deviation. Root mean square value of path deviation and path deviation rate The path deviation rate is the proportion of the length whose deviation exceeds a preset warning threshold to the total length. , as well as The input is fed into a preset random forest model to obtain the qualification score of the broken path;

[0012] S13. Perform dual-feature targeted preprocessing on the fracture surface image, extracting and enhancing the gray-level gradient abrupt boundary features of the slag particle region and the shadow features of the semi-detached slag particles to obtain a dual-feature fused image; construct a two-dimensional gray-level histogram based on the gray values ​​of the fracture surface image and the feature response values ​​of the dual-feature fused image, extract peak points to complete the cluster center targeted initialization; perform initial clustering with the initialized cluster centers, calculate the segmentation quality of each cluster using the reconstructed dual feature values, perform iterative clustering optimization with the segmentation threshold as a constraint, and obtain the final clustering segmentation result; calculate the fracture surface qualification score based on the clustering segmentation result;

[0013] S14. Set the minimum threshold for the qualification score of the fracture path and the minimum threshold for the qualification score of the fracture surface. If the score is less than the threshold, it is directly set to 0. The quality qualification score label of the fracture is calculated by weighted summation, and the value range is 0 to 1.

[0014] Preferably, step S13, which obtains the fracture surface qualification score, is specifically implemented as follows:

[0015] S131. Calculate the sum of the gradient magnitude of each pixel in the fracture surface image and the absolute value of the gray level difference in the 5×5 neighborhood. Use the product of the two as the boundary response value of the slag particles. After normalization, obtain the boundary enhancement image. Extract low gray level shadow candidate areas from the fracture surface image. After local contrast enhancement and removal of isolated dark points, obtain the shadow enhancement image. Weightedly fuse the boundary enhancement image and the shadow enhancement image to obtain the dual-feature fusion image.

[0016] S132. Construct a two-dimensional gray-level histogram using the original gray-level value of the fracture surface image as the X-axis and the feature response value of the dual-feature fused image as the Y-axis. Statistical analysis of each The number of pixels corresponding to the coordinates, for Non-maximum suppression is performed, and significant peak points are extracted as initial cluster centers. K-means initial clustering is performed on the dual-feature fusion image to obtain initial clusters. For each cluster, a first feature value representing the dispersion of pixel distribution within the cluster and a second feature value representing the gray-level difference of pixels within the cluster are calculated. The product of the first feature value and the second feature value is used as the initial segmentation quality of the cluster.

[0017] S133. Set a normalized segmentation threshold, mark clusters whose initial segmentation quality is greater than the segmentation threshold, perform secondary clustering on the marked clusters, use the mean of the dual feature response values ​​of the pixels in the current cluster as the new cluster center, and split the original cluster into two sub-clusters.

[0018] S134. For the split sub-clusters, repeat the calculation of feature values ​​and segmentation quality until the segmentation quality of all clusters is less than the preset normalized segmentation threshold, then stop the iteration to obtain the final clustering segmentation result, accurately separating the matrix region and the slag particle region.

[0019] S135. Calculate the proportion of the matrix region to the total fracture surface image to obtain the fracture surface qualification score.

[0020] As a preferred approach, the sample data needs to be filtered before constructing the training dataset. The specific implementation is as follows:

[0021] Based on the specifications of the connecting rod, all the collected original sample data were divided into several sample groups of the same specifications. For each sample group of the same specifications, the sample was sorted in ascending order using the fracture quality qualification score label corresponding to the sample as the only sorting criterion, resulting in an ordered sample sequence in each group arranged from low to high fracture quality.

[0022] For each ordered sample sequence of the same sample group, sort the samples in ascending order and accurately remove the bottom 30% of low-value samples in each group. After removal, the remaining valid sample sequence accounts for 70% of the original sample size in each group.

[0023] For each valid sample sequence retained in the same specification sample group, the sequences are rearranged in descending order according to the fracture quality qualification score label. The sorted sequences are then divided into two functionally distinct sample subsets: the first subset is the core benchmark sample set, consisting of the top 50% of the samples in descending order, which corresponds to the groove depth matching samples with the best fracture quality for this specification under a fixed rated fracture force constraint, and is used to provide the optimal groove depth control benchmark for the model under rated working conditions; the second subset is the generalization boundary sample set, consisting of the bottom 50% of the samples in descending order, which corresponds to the critical matching samples with qualified but not optimal fracture quality for this specification under a fixed rated fracture force constraint, and is used to provide the qualified boundary threshold of groove depth and the adaptation rules for robustness to working condition fluctuations for the model.

[0024] For the two sample subsets that have been divided, differentiated sample augmentation and label weight configuration are performed based on their functional positioning to complete the core construction of the training dataset;

[0025] For all the bi subset samples processed by the same specification sample group, a sample number balancing process is performed to ensure that the difference in the number of samples in the core benchmark sample set and the generalization boundary sample set of different specification groups does not exceed a preset balancing threshold. Finally, the core benchmark sample set and the generalization boundary sample set of all specification groups are merged for supervised training of the deep learning model for laser groove depth prediction.

[0026] Preferably, the image processing operation is as follows:

[0027] S21. For the input grayscale image to be processed, calculate each pixel... For the target pixel, initialize a square filtering neighborhood of fixed size, with a neighborhood range of [value missing]. An 8-neighborhood interval centered at the center, with 1 pixel in each of the top, bottom, left, and right sides;

[0028] S22, Target pixel Calculate the basic feature parameters within its filtering neighborhood, including the gray-level gradient magnitude and gradient direction of each pixel in the neighborhood;

[0029] S23. Using the target pixel as a reference, calculate the pixels in the neighborhood. Relative to target pixel Gray-scale similarity weights: ,in, This is the grayscale similarity adjustment coefficient, which is the standard deviation of the grayscale values ​​of all pixels in the neighborhood. It is used to calculate the similarity of pixels in the neighborhood. Relative to target pixel Gradient consistency weights: ,in, The gradient consistency adjustment coefficient is used, and then the dual-constraint adaptive weights that fuse gray-level similarity and gradient consistency are constructed for each pixel in the neighborhood. ;

[0030] S24. Based on the comprehensive weights of each pixel in the neighborhood, the target pixel is calculated by weighted average. Final grayscale value after denoising: Iterate through all pixels of the input image to obtain the processed image.

[0031] Preferably, the extracted structural visual features of the connecting rod include the thickness and aperture of the connecting rod.

[0032] Preferably, the deep learning model for predicting laser grooving depth includes:

[0033] Input layer: Receives structural visual features, namely the wall thickness and large end aperture of the connecting rod. After performing standardization preprocessing on the two features respectively, the wall thickness feature and aperture feature are output to the split coding layer respectively.

[0034] Encoding layer: Set up encoding branches that correspond one-to-one with the input features. The two branches use structurally independent, parameter-unshared multilayer perceptron encoding units to extract the weight encoding of the influence of the two features on the laser groove depth, and output two independent high-dimensional feature vectors to the cross-feature attention fusion layer.

[0035] Attention Fusion Layer: It has a built-in dual-feature cross-attention interaction unit, which performs feature cross-mapping and adaptive weight allocation on two high-dimensional feature vectors, and fuses them to obtain a comprehensive structural feature vector that simultaneously represents the coupling effect of wall thickness and aperture, and outputs it to the decoding layer;

[0036] Decoding layer: Parallel groove depth regression decoding branch and fracture quality qualification prediction decoding branch are set up. The two branches perform feature cross-transfer and bidirectional iterative optimization in each decoding unit, and output the final depth prediction feature and quality prediction feature to the result output layer.

[0037] Output layer: Outputs the target laser grooving depth control parameters adapted to the fixed rated breaking force, and simultaneously outputs the breaking quality qualification prediction score corresponding to the depth control parameters.

[0038] Preferably, the bidirectional iterative optimization of the decoding layer is achieved through the following steps:

[0039] S31. Input the comprehensive structural feature vector output by the attention fusion layer into the initial fully connected layers of the groove depth regression decoding branch and the fracture quality qualification prediction decoding branch, respectively. After linear transformation and layer normalization, the initial depth decoding features and initial quality decoding features corresponding to the two branches are obtained.

[0040] S32. Set up no less than three layers of cascaded coupled decoding units. Each coupled decoding unit performs iterative updates in the following order: First, depth feature update: concatenate the depth decoding features output by the previous layer with the quality decoding features, perform a linear transformation by a fully connected layer, and then perform a residual connection with the depth decoding features of the previous layer to obtain the updated depth decoding features of the current layer; Second, quality feature update: concatenate the updated depth decoding features of the current layer with the quality decoding features output by the previous layer, perform a linear transformation by a fully connected layer, and then perform a residual connection with the quality decoding features of the previous layer to obtain the updated quality decoding features of the current layer.

[0041] S33. After completing the coupling decoding iteration of all levels, the groove depth regression decoding branch outputs the final depth prediction feature, and the fracture quality qualification prediction decoding branch outputs the final quality score prediction feature.

[0042] Compared with existing technologies, the advantages and positive effects of this invention are as follows: It adopts a fixed rated breaking force design, eliminating the need for equipment replacement or cylinder adjustment, significantly improving production line changeover efficiency, reducing costs, and substantially enhancing flexibility. Through a deep learning model, it accurately predicts the laser grooving depth, eliminating reliance on empirically determined values ​​and adapting to different connecting rod specifications. Combining dual-feature image processing and intelligent quality scoring, it accurately solves quality problems such as fracture path deviation, plastic deformation, and failure to break. Through sample selection optimization and bidirectional iterative decoding, the model prediction accuracy is further improved, achieving intelligent and standardized control of connecting rod fracture, comprehensively enhancing processing quality and production stability. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating a laser cutting depth control method for connecting rod fracture forming. Detailed Implementation

[0045] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0046] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0047] In this embodiment, the present invention provides a laser cutting depth control method for connecting rod fracture forming, such as... Figure 1 As shown, the specific implementation is as follows:

[0048] A fixed rated breaking force is pre-calibrated to ensure a constant output of the cylinder driving the breaking device throughout its entire stroke. Under the constraint of the fixed rated breaking force, sample data of connecting rods of different specifications are collected to construct a training dataset. The sample data includes a panoramic image of the connecting rod structure, laser grooving depth control parameters adapted to the fixed rated breaking force of the connecting rod, a fracture image of the connecting rod after breaking, and a breaking quality qualification score label generated based on the fracture image.

[0049] Furthermore, the specific implementation of the fracture quality qualification rating label generated based on the fracture image includes: after the connecting rod completes fracture, acquiring a grayscale image of the fracture surface viewed from the front and a fracture path image perpendicular to the fracture surface, and performing image processing on the two images sequentially to obtain a processed image; extracting continuous fracture contour lines from the processed fracture path image and spatially registering them with a preset laser grooving reference line, using the grooving length direction as the x-axis and the workpiece wall thickness direction perpendicular to the grooving as the y-axis, obtaining a path deviation sequence along the entire length of the grooving. And extract the path with the largest deviation. Root mean square value of path deviation and path deviation rate The path deviation rate is the proportion of the length whose deviation exceeds a preset warning threshold to the total length. , as well as Inputting the data into a pre-defined random forest model yields a score for the validity of broken paths.

[0050] A dual-feature targeted preprocessing method is performed on the fracture surface image to extract and enhance the gray-level gradient abrupt boundary features of the slag particle region and the shadow features of the semi-detached slag particles, respectively, to obtain a dual-feature fused image. Based on the gray-level values ​​of the fracture surface image and the feature response values ​​of the dual-feature fused image, a two-dimensional gray-level histogram is constructed, and peak points are extracted to complete the targeted initialization of cluster centers. Initial clustering is performed using the initialized cluster centers, and the segmentation quality of each cluster is calculated using the reconstructed dual feature values. Iterative clustering optimization is performed with a segmentation threshold as a constraint to obtain the final clustering segmentation result. The fracture surface qualification score is calculated based on the clustering segmentation result. Specifically, the specific implementation of obtaining the fracture surface qualification score is as follows:

[0051] The gradient magnitude of each pixel in the fracture surface image is calculated as the sum of the absolute values ​​of the gray-level differences within a 5×5 neighborhood. The product of these two values ​​is used as the boundary response value of the slag particles, which is then normalized to obtain a boundary enhancement image. Low-gray-level shadow candidate regions are extracted from the fracture surface image, and after local contrast enhancement and isolated dark spot removal, a shadow enhancement image is obtained. The boundary enhancement image and the shadow enhancement image are then weighted and fused to obtain a dual-feature fusion image. Specifically, the gradient magnitude of each pixel in the image is calculated as follows: Gradient magnitude: ,in, For the image in gradient magnitude at that point These are the horizontal and vertical gradient components, respectively; then the pixel point is calculated. The sum of the absolute values ​​of grayscale differences within a 5×5 neighborhood: ,in For a pixel, the 5×5 neighborhood is for Points within; calculate the boundary response value of slag particles: ,right Normalization from 0 to 1 is performed to obtain the boundary enhancement image. This process strengthens the gradient abrupt boundary of slag particles and suppresses weak gradient interference from the natural grain texture of the fracture surface. Next, low-grayscale shadow candidate regions are extracted from the fracture surface image, and a histogram equalization algorithm is used to amplify the grayscale difference between the shadow region and the surrounding matrix, resulting in a shadow-enhanced image. Enhance the boundary of the image With shadow enhancement image Weighted fusion is performed to obtain a feature-enhanced image.

[0052] A two-dimensional gray-level histogram is constructed using the original gray-level values ​​of the fracture surface image as the X-axis and the feature response values ​​of the dual-feature fused image as the Y-axis. Statistical analysis of each The number of pixels corresponding to the coordinates, for Non-maximum suppression is performed to extract significant peak points as initial cluster centers; the cluster centers are two regions corresponding to the matrix metal with high X and low Y peak points, respectively, which are set as matrix cluster centers, and regions corresponding to the slag particles with high Y and low X peak points, which are set as slag particle cluster centers.

[0053] K-means initial clustering is performed on the dual-feature fused image to obtain initial clusters. For each cluster, a first feature value representing the dispersion of pixel distribution within the cluster and a second feature value representing the gray-level difference of pixels within the cluster are calculated. For the q-th cluster, q=1,2, the first feature value is calculated as follows: ,in, The boundary response value within this cluster The number of pixels exceeding the preset boundary threshold This represents the total number of pixels within a cluster. The standard deviation of the feature response values ​​of all pixels within a cluster in the dual-feature fused image is given. This is the sum of squared deviations of the distances between the nth pixel within a cluster and the dual-feature response values ​​of all other pixels within the cluster. The second feature value is calculated as follows: ,in, This represents the contrast of the clusters in the dual-feature fusion image, i.e., the difference between the maximum and minimum feature response values ​​within the cluster. This represents the difference between the feature response value of the nth pixel within the cluster and the mean feature response value of the pixels in its 5×5 neighborhood. The average of the above differences among all pixels within a cluster is used as the product of the first feature value and the second feature value to determine the initial segmentation quality of the cluster. The higher the segmentation quality, the worse the consistency within the cluster, and the more iterative optimization of the segmentation is needed.

[0054] A preset normalized segmentation threshold is used to mark clusters whose initial segmentation quality exceeds the threshold. Secondary clustering is then performed on these marked clusters, using the average of the dual-feature response values ​​of pixels within the current cluster as the new cluster center. This splits the original cluster into two sub-clusters. For each sub-cluster, feature values ​​and segmentation quality calculations are repeated until the segmentation quality of all clusters is less than the preset normalized segmentation threshold. The iteration then stops, yielding the final clustering segmentation result, accurately separating the matrix region and the slag particle region. Specifically, the normalized segmentation threshold is first preset. Based on the texture complexity of the connecting rod fracture surface, the slag particle size distribution, and the image acquisition resolution, the normalized segmentation threshold is calibrated and preset through process experiments. This threshold is a constant between 0 and 1, used to determine the consistency of pixel features within a cluster. A segmentation quality greater than this threshold indicates high pixel dispersion within the cluster, requiring further segmentation and optimization; a segmentation quality less than this threshold indicates uniform pixel features within the cluster, eliminating the need for further iteration. Clusters with segmentation quality greater than a preset normalized segmentation threshold are marked as clusters to be optimized. For each marked cluster to be optimized, the dual-feature response values ​​of all pixels within the cluster are extracted, and their arithmetic mean is calculated as the new sub-cluster center. Based on this new center, the original cluster to be optimized is split into two sub-clusters with more uniform features. After splitting, the original cluster to be optimized is replaced by two sub-clusters, effectively reducing the pixel feature dispersion within a single cluster. Then, a loop iteration and feature recalculation stage is entered. For all the split sub-clusters, the feature value and segmentation quality calculation process is repeated, recalculating the first feature value, second feature value, and normalized segmentation quality of each sub-cluster. The segmentation quality of each sub-cluster is compared with the preset threshold again. If a sub-cluster segmentation quality is still greater than the threshold, the center update and secondary splitting operation is performed on that sub-cluster. If the sub-cluster segmentation quality is less than the threshold, the sub-cluster is retained and not split further. Finally, when the normalized segmentation quality of all clusters (including the initial unlabeled clusters and stable sub-clusters after multiple splits) is less than the preset normalized segmentation threshold, the iteration process automatically terminates, yielding the final clustering segmentation result. This result accurately segments the effective region of the standardized fracture surface into mutually exclusive clusters, namely the pure matrix region and the slag particle region. The proportion of the matrix region to the total fracture surface image is calculated to obtain the fracture surface qualification score, which ranges from 0 to 1.

[0055] Finally, minimum thresholds for the fracture path qualification score and fracture surface qualification score are set. If the score is less than the threshold, it is directly set to 0. The fracture quality qualification score label is calculated by weighted summation, with a value range of 0 to 1.

[0056] Furthermore, before constructing the training dataset from the sample data, the sample data needs to be filtered. Specifically, this is implemented as follows:

[0057] Based on the specifications of the connecting rods, all the collected raw sample data were divided into several sample groups of the same specifications. For each sample group of the same specifications, the sample was sorted in ascending order using the fracture quality qualification score label corresponding to the sample as the sole sorting criterion, resulting in an ordered sample sequence within each group arranged from low to high fracture quality.

[0058] For each ordered sample sequence of the same sample group, the ranking is sorted in ascending order, and the bottom 30% of low-value samples in each group are accurately removed. After removal, 70% of the original sample volume of each group is retained as valid sample sequences.

[0059] For each valid sample sequence retained in the same specification sample group, the sequences are rearranged in descending order according to the fracture quality qualification score label. The sorted sequences are then divided into two functionally distinct sample subsets based on their ranking: The first subset is the core benchmark sample set, consisting of the top 50% of the samples in descending order. These samples correspond to the groove depth matching samples with the best fracture quality for this specification under a fixed rated fracture force constraint, and are used to provide the optimal groove depth control benchmark for the model under rated working conditions. The second subset is the generalization boundary sample set, consisting of the bottom 50% of the samples in descending order. These samples correspond to the critical matching samples with qualified but not optimal fracture quality for this specification under a fixed rated fracture force constraint, and are used to provide the qualified boundary threshold of groove depth and the adaptation rules for robustness to working condition fluctuations for the model.

[0060] For the two sample subsets that have been divided, differentiated sample augmentation and label weight configuration are performed based on their functional positioning to complete the core construction of the training dataset. For the core benchmark sample set, based on the panoramic image of the linkage structure, micro-scale structural feature perturbation enhancement is performed, including pixel-level micro-deformation of the large-end aperture and wall thickness, and fine-tuning of illumination conditions, generating enhanced samples with labels consistent with the original samples. At the same time, 1.5 to 2.0 times the supervision loss weight is configured for the fracture quality qualification score label of all samples in this set, forcing the model to prioritize learning the core matching relationship between the optimal groove depth and the rated fracture force. For the generalization boundary sample set, based on the laser groove depth control parameters, parameter micro-variation enhancement is performed in the qualified boundary interval. Enhanced samples with parameter gradients are generated within the ±5% critical qualified interval of the groove depth of this sample, and the corresponding gradient score labels are matched. At the same time, 0.8 to 1.2 times the floating supervision loss weight is configured for the label of all samples in this set, guiding the model to learn the qualified boundary and fluctuation tolerance range of the groove depth, and strengthening the model's cross-working condition generalization ability.

[0061] For all bi subset samples processed from the same specification sample group, sample number balancing is performed to ensure that the difference in the number of samples in the core benchmark sample set and the generalization boundary sample set of different specification groups does not exceed the preset balancing threshold, thus avoiding the specification bias problem dominated by large specification samples during model training. Finally, the core benchmark sample set and the generalization boundary sample set of all specification groups are merged to construct a hierarchical supervised training dataset with labeled weighted coefficients, which is directly used for supervised training of the deep learning model for laser groove depth prediction.

[0062] Next, for any specification of connecting rod to be processed, a panoramic image of its structure is acquired. Through image processing and feature extraction, the structural visual features of the connecting rod are extracted. Furthermore, the method for image processing the panoramic image is the same as in the previous step of acquiring a grayscale image of the fracture surface facing forward and an image of the fracture path perpendicular to the fracture surface, and then sequentially performing image processing on both types of images. Specifically, for the input grayscale image to be processed, each pixel... For the target pixel, initialize a square filtering neighborhood of fixed size, with a neighborhood range of [value missing]. An 8-neighborhood interval, 1 pixel above, below, left, and right, centered on the target pixel; Calculate the basic feature parameters within its filtering neighborhood, including the grayscale gradient magnitude and gradient direction of each pixel in the neighborhood; using the target pixel as a reference, calculate the pixel values ​​in the neighborhood. Relative to target pixel Gray-scale similarity weights: ,in, This is the grayscale similarity adjustment coefficient, which is the standard deviation of the grayscale values ​​of all pixels in the neighborhood. It is used to calculate the similarity of pixels in the neighborhood. Relative to target pixel Gradient consistency weights: ,in, The gradient consistency adjustment coefficient is used, and then the dual-constraint adaptive weights that fuse gray-level similarity and gradient consistency are constructed for each pixel in the neighborhood. The target pixel is calculated by weighted average based on the combined weights of each pixel in the neighborhood. Final grayscale value after denoising: Iterate through all pixels of the input image to obtain the processed image.

[0063] Then, a deep learning model for predicting laser grooving depth is constructed. The structural visual features extracted from the panoramic image of the connecting rod structure are used as the model input, and the laser grooving depth control parameters adapted to the fixed rated breaking force and the breaking quality qualification score are used as the model output, completing the supervised training of the model. The extracted structural visual features of the connecting rod include the rod's thickness and aperture.

[0064] Furthermore, the laser grooving depth prediction deep learning model includes an input layer, an encoding layer, an attention fusion layer, a decoding layer, and an output layer.

[0065] The input layer receives structural visual features, namely the wall thickness and large-end aperture of the connecting rod. After performing standardization preprocessing on these two features, the wall thickness and aperture features are output to the split-encoding layer respectively. Specifically, this layer only receives two core structural visual features extracted from the panoramic image of the connecting rod structure: the connecting rod wall thickness and the connecting rod aperture. This layer performs standardization preprocessing on these two features using Min-Max standardization. Based on the training dataset, the global maximum and minimum values ​​of the wall thickness and aperture features are calculated respectively, mapping the two features to the interval between 0 and 1 to eliminate the influence of dimensions.

[0066] The encoding layer sets up encoding branches that correspond one-to-one with the input features. The two branches use structurally independent, non-parameter-sharing multilayer perceptron encoding units to extract the weight encoding of the influence of the two features on the laser groove depth, and output two independent high-dimensional feature vectors to the cross-feature attention fusion layer.

[0067] Specifically, the encoding layer is the core feature extraction unit of the model. Its core responsibility is to perform deep encoding on the structural visual features of two independent inputs, mining the nonlinear mapping relationship between wall thickness features, aperture features, and laser groove depth, and outputting two independent high-dimensional feature vectors. This layer has dedicated encoding branches for wall thickness features and aperture features, each corresponding one-to-one with the input features. The two branches use a multilayer perceptron encoding unit with symmetrical structure but completely non-shared parameters. Both encoding branches employ a three-layer cascaded fully connected encoding structure. The specific implementation of each branch is as follows: The first layer is a feature upscaling layer. The input is a 1D standardized feature, and a 64-dimensional fully connected layer is used with the ReLU activation function to perform non-linear upscaling of the feature, mapping the single-dimensional physical feature to a high-dimensional feature space and uncovering the fundamental non-linear correlation between the feature and the groove depth. The second layer is a depth feature extraction layer, using a 128-dimensional fully connected layer with the GELU activation function and layer normalization. It deeply learns the influence weights of this feature on the stress distribution and fracture critical conditions of the connecting rod under a fixed rated breaking force, extracting depth features strongly correlated with the groove depth. The third layer is a feature normalization layer, using a 64-dimensional fully connected layer with a linear activation function to normalize and denoise the high-dimensional feature, outputting a unified high-dimensional feature vector. After simultaneous encoding by both branches, they output 64-dimensional high-dimensional feature vectors for wall thickness and aperture, respectively, which are then passed to the attention fusion layer.

[0068] The attention fusion layer has a built-in dual-feature cross-attention interaction unit, which performs feature cross-mapping and adaptive weight allocation on two high-dimensional feature vectors, and fuses them to obtain a comprehensive structural feature vector that simultaneously represents the coupling effect of wall thickness and aperture, and outputs it to the decoding layer.

[0069] Specifically, the implementation of the dual-feature cross-attention interaction unit is as follows: First, the input high-dimensional feature vector of wall thickness is processed... High-dimensional feature vector of aperture Each element generates a corresponding query vector (Query), key vector (Key), and value vector (Value) through an independent linear mapping layer, thus generating the wall thickness feature. , , Aperture feature generation , , Then, bidirectional cross-attention calculation is performed, with the first path using the query vector of wall thickness features. Key vectors related to aperture features The attention weights are calculated and normalized using the Softmax function to obtain the value vector of the aperture features. Weighted summation yields thickness-guided aperture interaction features; the second path uses the query vector of aperture features. Key vectors with wall thickness characteristics Calculate the attention weights and normalize the value vector of the wall thickness feature. Weighted summation yields the aperture-guided wall thickness interaction features. Through bidirectional cross-calculation, the coupling influence of the two features is fully captured. After the cross-interaction is completed, this layer dynamically allocates the fusion weights of the two interaction features based on the contribution of the two features learned during training to the fracture quality. The weighted concatenation generates a 128-dimensional comprehensive structural feature vector, which fully preserves the independent influence and coupling effect of the two features. This vector is then output to the decoding layer to complete the final prediction and decoding.

[0070] The decoding layer consists of a parallel groove depth regression decoding branch and a fracture quality qualification prediction decoding branch. The two branches perform feature cross-transfer and bidirectional iterative optimization in each decoding unit, and output the final depth prediction features and quality prediction features to the result output layer. Specifically, the comprehensive structural feature vector output from the attention fusion layer is input into the initial fully connected layers of the groove depth regression decoding branch and the fracture quality qualification prediction decoding branch, respectively. After linear transformation and layer normalization, the initial depth decoding features and initial quality decoding features corresponding to the two branches are obtained. At least three cascaded coupled decoding units are set up, and each coupled decoding unit performs iterative updates in the following order: First, depth feature update: the depth decoding features and quality decoding features output from the previous layer are concatenated, and after linear transformation by the fully connected layer, a residual connection is performed with the depth decoding features of the previous layer to obtain the updated depth decoding features of the current layer. Second, quality feature update: the updated depth decoding features of the current layer are concatenated with the quality decoding features output from the previous layer, and after linear transformation by the fully connected layer, a residual connection is performed with the quality decoding features of the previous layer to obtain the updated quality decoding features of the current layer. After completing the coupled decoding iterations of all layers, the groove depth regression decoding branch outputs the final depth prediction feature, and the fracture quality qualification prediction decoding branch outputs the final quality score prediction feature.

[0071] The output layer outputs the target laser grooving depth control parameters adapted to a fixed rated breaking force, and simultaneously outputs the breaking quality qualification prediction score corresponding to these depth control parameters. Specifically, the grooving depth regression output head receives the depth prediction features output from the decoding layer, performs a linear transformation through a single fully connected layer, and uses a ReLU activation function to ensure the output is non-negative, ultimately outputting a unique scalar value, which is the target laser grooving depth control parameter adapted to a fixed rated breaking force. This parameter directly corresponds to the laser grooving depth that achieves the optimal breaking quality of the connecting rod under a fixed rated breaking force. The breaking quality score prediction output head receives the quality prediction features output from the decoding layer, performs a linear transformation through a single fully connected layer, and uses a Sigmoid activation function to output a scalar value in the range of 0 to 1, which is the breaking quality qualification prediction score corresponding to this grooving depth. The closer the score is to 1, the better the breaking quality.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for controlling the laser cutting depth in connecting rod fracture forming, characterized in that, Includes the following steps: S1. Pre-calibrate the fixed rated breaking force of the driving cylinder of the breaking device to maintain constant output throughout the entire process. Under the constraint of the fixed rated breaking force, collect sample data of connecting rods of different specifications to construct a training dataset. The sample data includes a panoramic image of the connecting rod structure, laser grooving depth control parameters adapted to the fixed rated breaking force of the connecting rod, a fracture image of the connecting rod after breaking, and a breaking quality qualification score label generated based on the fracture image. S2. For any specification of connecting rod to be processed, acquire its panoramic structural image, and extract the structural visual features of the connecting rod through image processing and feature extraction. S3. Construct a deep learning model for predicting laser grooving depth, using the structural visual features extracted from the panoramic image of the linkage structure as the model input, and the laser grooving depth control parameters adapted to the fixed rated breaking force and the breaking quality qualification score as the model output, to complete the supervised training of the model. S4. Input the structural visual features of the connecting rod to be processed into the trained deep learning model for predicting the laser grooving depth, and directly output the target laser grooving depth control parameters that are adapted to the rated breaking force.

2. The laser cutting depth control method for connecting rod fracture forming according to claim 1, characterized in that, The specific implementation of the fracture quality qualification rating label generated based on the fracture image includes: S11. After the connecting rod completes the fracture, acquire a grayscale image of the fracture surface facing forward and an image of the fracture path perpendicular to the fracture surface, and perform image processing on the two images in sequence to obtain the processed image. S12. Extract continuous fracture contour lines from the processed fracture path image and spatially register them with the preset laser grooving reference line. Using the grooving length direction as the x-axis and the workpiece wall thickness direction perpendicular to the grooving as the y-axis, obtain the path deviation sequence along the entire length of the grooving. And extract the path with the largest deviation. Root mean square value of path deviation and path deviation rate The path deviation rate is the proportion of the length whose deviation exceeds a preset warning threshold to the total length. , as well as The input is fed into a preset random forest model to obtain the qualification score of the broken path; S13. Perform dual-feature targeted preprocessing on the fracture surface image, extracting and enhancing the gray-level gradient abrupt boundary features of the slag particle region and the shadow features of the semi-detached slag particles to obtain a dual-feature fused image; construct a two-dimensional gray-level histogram based on the gray values ​​of the fracture surface image and the feature response values ​​of the dual-feature fused image, extract peak points to complete the cluster center targeted initialization; perform initial clustering with the initialized cluster centers, calculate the segmentation quality of each cluster using the reconstructed dual feature values, perform iterative clustering optimization with the segmentation threshold as a constraint, and obtain the final clustering segmentation result; calculate the fracture surface qualification score based on the clustering segmentation result; S14. Set the minimum threshold for the qualification score of the fracture path and the minimum threshold for the qualification score of the fracture surface. If the score is less than the threshold, it is directly set to 0. The quality qualification score label of the fracture is calculated by weighted summation, and the value range is 0 to 1.

3. The laser cutting depth control method for connecting rod fracture forming according to claim 2, characterized in that, The specific implementation of obtaining the fracture surface qualification score in step S13 is as follows: S131. Calculate the sum of the gradient magnitude of each pixel in the fracture surface image and the absolute value of the gray level difference in the 5×5 neighborhood. Use the product of the two as the boundary response value of the slag particles, and obtain the boundary enhancement image after normalization. Low-grayscale shadow candidate regions are extracted from the fracture surface image. After local contrast enhancement and isolated dark spot removal, a shadow-enhanced image is obtained. The boundary enhancement image and the shadow enhancement image are weighted and fused to obtain a dual-feature fused image. S132. Construct a two-dimensional gray-level histogram using the original gray-level value of the fracture surface image as the X-axis and the feature response value of the dual-feature fused image as the Y-axis. Statistical analysis of each The number of pixels corresponding to the coordinates, for Non-maximum suppression is performed, and significant peak points are extracted as initial cluster centers. K-means initial clustering is performed on the dual-feature fusion image to obtain initial clusters. For each cluster, a first feature value representing the dispersion of pixel distribution within the cluster and a second feature value representing the gray-level difference of pixels within the cluster are calculated. The product of the first feature value and the second feature value is used as the initial segmentation quality of the cluster. S133. Set a normalized segmentation threshold, mark clusters whose initial segmentation quality is greater than the segmentation threshold, perform secondary clustering on the marked clusters, use the mean of the dual feature response values ​​of the pixels in the current cluster as the new cluster center, and split the original cluster into two sub-clusters. S134. For the split sub-clusters, repeat the calculation of feature values ​​and segmentation quality until the segmentation quality of all clusters is less than the preset normalized segmentation threshold, then stop the iteration to obtain the final clustering segmentation result, accurately separating the matrix region and the slag particle region. S135. Calculate the proportion of the matrix region to the total fracture surface image to obtain the fracture surface qualification score.

4. The laser cutting depth control method for connecting rod fracture forming according to claim 1, characterized in that, Before constructing the training dataset from the sample data, the sample data needs to be filtered. The specific implementation is as follows: Based on the specifications of the connecting rod, all the collected original sample data were divided into several sample groups of the same specifications. For each sample group of the same specifications, the sample was sorted in ascending order using the fracture quality qualification score label corresponding to the sample as the only sorting criterion, resulting in an ordered sample sequence in each group arranged from low to high fracture quality. For each ordered sample sequence of the same sample group, sort the samples in ascending order and accurately remove the bottom 30% of low-value samples in each group. After removal, the remaining valid sample sequence accounts for 70% of the original sample size in each group. For each valid sample sequence retained in the same specification sample group, the sequences are rearranged in descending order according to the fracture quality qualification score label. The sorted sequences are then divided into two functionally distinct sample subsets based on their ranking: The first subset is the core benchmark sample set, consisting of the top 50% of the samples in descending order. These samples correspond to the groove depth matching samples with the best fracture quality for this specification under a fixed rated fracture force constraint, and are used to provide the optimal groove depth control benchmark for the model under rated working conditions. The second subset is the generalization boundary sample set, consisting of the bottom 50% of the samples in descending order. These samples correspond to the critical matching samples with qualified but not optimal fracture quality for this specification under a fixed rated fracture force constraint, and are used to provide the qualified boundary threshold for groove depth and the adaptation rules for robustness to working condition fluctuations for the model. For the two sample subsets that have been divided, differentiated sample augmentation and label weight configuration are performed based on their functional positioning to complete the core construction of the training dataset; For all the bi subset samples processed by the same specification sample group, a sample number balancing process is performed to ensure that the difference in the number of samples in the core benchmark sample set and the generalization boundary sample set of different specification groups does not exceed a preset balancing threshold. Finally, the core benchmark sample set and the generalization boundary sample set of all specification groups are merged for supervised training of the deep learning model for laser groove depth prediction.

5. A laser cutting depth control method for connecting rod fracture forming according to claim 1 or 2, characterized in that, The image processing operations are as follows: S21. For the input grayscale image to be processed, calculate each pixel... For the target pixel, initialize a square filtering neighborhood of fixed size, with a neighborhood range of [value missing]. An 8-neighborhood interval centered at the center, with 1 pixel in each of the top, bottom, left, and right sides; S22, Target pixel Calculate the basic feature parameters within its filtering neighborhood, including the gray-level gradient magnitude and gradient direction of each pixel in the neighborhood; S23. Using the target pixel as a reference, calculate the pixels in the neighborhood. Relative to target pixel Gray-scale similarity weights: ,in, This is the grayscale similarity adjustment coefficient, which is the standard deviation of the grayscale values ​​of all pixels in the neighborhood. It is used to calculate the similarity of pixels in the neighborhood. Relative to target pixel Gradient consistency weights: ,in, The gradient consistency adjustment coefficient is used, and then the dual-constraint adaptive weights that fuse gray-level similarity and gradient consistency are constructed for each pixel in the neighborhood. ; S24. Based on the comprehensive weights of each pixel in the neighborhood, the target pixel is calculated by weighted average. Final grayscale value after denoising: Iterate through all pixels of the input image to obtain the processed image.

6. The laser cutting depth control method for connecting rod fracture forming according to claim 1, characterized in that, The structural visual features of the extracted connecting rod include the thickness and aperture of the connecting rod.

7. The laser cutting depth control method for connecting rod fracture forming according to claim 1, characterized in that, The deep learning model for predicting laser grooving depth includes: Input layer: Receives structural visual features, namely the wall thickness and large end aperture of the connecting rod. After performing standardization preprocessing on the two features respectively, the wall thickness feature and aperture feature are output to the split coding layer respectively. Encoding layer: Set up encoding branches that correspond one-to-one with the input features. The two branches use structurally independent, parameter-unshared multilayer perceptron encoding units to extract the weight encoding of the influence of the two features on the laser groove depth, and output two independent high-dimensional feature vectors to the cross-feature attention fusion layer. Attention Fusion Layer: It has a built-in dual-feature cross-attention interaction unit, which performs feature cross-mapping and adaptive weight allocation on two high-dimensional feature vectors, and fuses them to obtain a comprehensive structural feature vector that simultaneously represents the coupling effect of wall thickness and aperture, and outputs it to the decoding layer; Decoding layer: Parallel groove depth regression decoding branch and fracture quality qualification prediction decoding branch are set up. The two branches perform feature cross-transfer and bidirectional iterative optimization in each decoding unit, and output the final depth prediction feature and quality prediction feature to the result output layer. Output layer: Outputs the target laser grooving depth control parameters adapted to the fixed rated breaking force, and simultaneously outputs the breaking quality qualification prediction score corresponding to the depth control parameters.

8. The laser cutting depth control method for connecting rod fracture forming according to claim 7, characterized in that, The bidirectional iterative optimization of the decoding layer is achieved through the following steps: S31. Input the comprehensive structural feature vector output by the attention fusion layer into the initial fully connected layers of the groove depth regression decoding branch and the fracture quality qualification prediction decoding branch, respectively. After linear transformation and layer normalization, the initial depth decoding features and initial quality decoding features corresponding to the two branches are obtained. S32. Set up no less than three layers of cascaded coupled decoding units. Each layer of coupled decoding unit performs iterative updates in the following order: First step, depth feature update, the depth decoding features output by the previous layer are concatenated with the quality decoding features, and after linear transformation by the fully connected layer, residual connection is performed with the depth decoding features of the previous layer to obtain the updated depth decoding features of the current layer. The second step, The quality feature update involves concatenating the updated depth decoding features of the current layer with the quality decoding features output from the previous layer, performing a linear transformation through a fully connected layer, and then performing a residual connection with the quality decoding features of the previous layer to obtain the updated quality decoding features of the current layer. S33. After completing the coupling decoding iteration of all levels, the groove depth regression decoding branch outputs the final depth prediction feature, and the fracture quality qualification prediction decoding branch outputs the final quality score prediction feature.

Images