An intelligent detection method for determining forming defects by the offset of the blank contour of a stamping part

Abstract

The present invention discloses an intelligent detection method for determining forming defects by the offset of the blank contour of a stamping part, belonging to the field of sheet metal forming quality detection. In the present invention, the offset amounts of the contour feature points of the blank in the defective state and the defect-free state are obtained through finite element simulation of the stamping process. After manually annotating the defect types, a training data set and a validation data set are obtained, and based on this, a BP neural network model is constructed and trained to obtain a defect detection model. At the stamping site, a laser rangefinder is used to measure the offset amount of the actual stamping blank contour feature points along the material inflow direction, and the offset amount is input into the BP neural network defect detection model to output the defect judgment result. The method of the present invention greatly alleviates the current manual dependence on stamping forming quality detection and has the advantages of high detection efficiency, high accuracy, and low cost.

Description

Technical Field

[0001] The present invention belongs to the field of alloy sheet forming quality inspection, and particularly relates to an intelligent inspection method for determining forming defects based on the offset of the blank contour of a stamping part. Background Art

[0002] In the production process of automobiles, about 60% - 70% of the parts are stamping parts, such as automobile engine covers, vehicle frames, automobile supports, sheet metal parts, etc. These parts have high requirements for the precision of stamping forming technology and strict quality requirements. Changes in various relevant material properties, process parameters, die structures and their related parameters will affect the stamping forming quality. Low manufacturing process level of stamping parts may bring a series of problems, such as defects like cracking, wrinkling, surface scratching, indentation, collapse, dark pits, convex hulls, impact lines, slip lines, springback, warping, etc. If there are defects on the surface of the stamping part, the size of the final product will deviate, and different electromagnetic characteristics will be caused, and the coating effect of the final product will also be affected.

[0003] An efficient and accurate defect detection method is the key to obtaining high-quality stamping products. Only during the production process can the defect information of the stamping part be fed back in real time, so as to avoid generating a large number of defective parts by timely adjusting the stamping parameters. Most domestic manufacturing enterprises rely on professional inspectors to detect product quality. Inspectors detect stamping products in sequence through training and their understanding of different defects. However, manual inspection has limitations such as low efficiency, and the detection ability depends on the inspector's own mood and work attitude.

[0004] In response to this situation, machine vision proposes a new detection method, that is, by processing the image of the stamping part in the computer, using different algorithms to replace the function of the human eye, obtaining the correct and valid information in the image, and then detecting the shape and distribution of the defect area. The advantage of machine vision compared to the human eye is its powerful computing ability and objectivity of the result. However, its detection accuracy depends on the resolution of the input image and the cost is relatively high. Moreover, there are many types of stamping defects. Machine vision is only suitable for detecting obvious surface defects and is powerless for defects such as dark pits, internal cracks, springback, insufficient strain, uneven thickness, excessive thinning, etc. Therefore, the dependence on manual inspection has not been completely solved. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides an intelligent inspection method for determining forming defects based on the offset of the blank contour of a stamping part, which can quickly and accurately determine the presence / absence of stamping defects and the types of defects based on the offset of the blank contour, and effectively improve the inspection efficiency and accuracy of stamping forming quality.

[0006] The present invention is implemented by the following technical solutions: An intelligent detection method for determining forming defects in the outline offset of a stamping blank, comprising the following steps:

[0007] A. Construct a data set of the offset amounts of the feature points on the blank outline;

[0008] 1) Perform several stamping simulation simulations using finite element software; simulate various defective and non-defective situations during the stamping process, and the defects include: cracking, wrinkling, surface scratching, pressing / indentation, collapse, dimple, bulge, warping, and springback; respectively output the blank outline lines corresponding to various situations when the die is fully closed;

[0009] 2) Arrange several feature points along the original blank outline, and measure the offset amounts of the feature points along the material inflow direction; the offset amount is the distance difference of the blank outline along the material inflow direction before and after stamping, and the selection rules of the feature points are as follows: evenly distribute them around the outline, and increase or decrease according to the uniformity of the material inflow amount, that is, increase the setting quantity in the area where the inflow amount changes greatly;

[0010] 3) With the aid of CAD software, calculate the offset amounts of the feature points of the stamping blank outline and the original blank outline in the non-defective and various defective situations simulated by the finite element;

[0011] 4) Manually mark the defect types according to the stamping results after simulation, save the offset amounts of the feature points and the defect types of each data sample in the format of a data set, and randomly divide them into a training data set, a validation data set, and a test data set.

[0012] B. Construct a BP neural network model;

[0013] The BP neural network includes 3 parts: an input layer, a two-layer hidden layer, and an output layer;

[0014] The input layer inputs the offset amounts of the feature points on the blank outline, and the output layer outputs classification labels, and the types include: normal, cracking, wrinkling, surface scratching, pressing / indentation, collapse, dimple, bulge, warping, and springback, etc.;

[0015] The neuron activation function selects the Rectified Linear Unit (ReLU), and its formula is:

[0016] f(x) = max(0, x)

[0017] where x is the input of the neuron.

[0018] The neuron activation function of the output layer is a Softmax classifier, which converts the stamping defect output result into a probability value, and its calculation formula is as follows:

[0019]

[0020] wherein, a i represents the i-th value of the vector input to the Softmax classifier;

[0021] y i represents the i-th value of the output result y, that is, the probability that the input sample is predicted to belong to class i;

[0022] a j represents the j-th value of the vector input to the Softmax classifier, and T represents the number of defect categories.

[0023] The result is calculated through the forward propagation process of the network model, and the network parameters are optimized using the backpropagation and gradient descent algorithms; the optimizer selects the adaptive learning rate optimization algorithm (Adam). Adam is an adaptive parameter update algorithm that can change the value of the learning rate during training and provides an adaptive learning rate based on the first-order moment estimate and second-order moment estimate of the gradient.

[0024] The model is trained using the sample images in the training dataset; then the validation dataset is used to further supervise the training of the model, adjust the parameters in a timely manner, and accelerate the training progress.

[0025] Furthermore, the loss function used in step B is the cross-entropy loss function, and its definition formula is:

[0026]

[0027] wherein, y n is the true value of the n-th sample;

[0028] is the predicted value of the n-th sample;

[0029] N is the total number of test samples;

[0030] C. Measure the offset of the actual stamping blank profile feature points along the material flow-in direction, and use it as the input value of the trained BP neural network model to output the defect classification result.

[0031] In step A1), the specific steps of performing stamping simulation using finite element software are as follows:

[0032] Establish a three-dimensional geometric simulation model of the die and the blank, input the accurately measured material model, optimized process parameters, near-actual boundary conditions and stamping operation processes, simulate the stamping forming of defect-free products, and output the corresponding blank contour simulation results; according to the actual process conditions, set corresponding deviations based on the input parameters of the defect-free product simulation, including setting original blank defects at critical points, changing material parameters, process parameters and boundary conditions, etc., simulate various stamping defects during the forming process, and output the corresponding blank contour simulation results;

[0033] The original blank defects mentioned above include: uneven thickness, uneven surface roughness, dark pits, hidden damages, convex hulls, etc.; the material parameters include: parameters related to yield and hardening models, etc.; the process parameters include: die clearance, blank holder force, draw bead coefficient, etc.; the boundary conditions include: punch and die shapes, friction coefficient, stamping speed, etc.

[0034] In step A4), the method for manually marking the defect type according to the stamping result after simulation is as follows:

[0035] Evaluate the quality of the stamped parts simulated by finite element, and classify them manually, which can be divided into two categories: one is normal stamped parts, marked as "normal"; the other is stamped parts with defects, and the stamping defects are subdivided and marked in turn. The label list is: {0: normal, 1: fracture, 2: wrinkling, 3: surface abrasion, 4: indentation, 4: collapse, 6: dark pit, 7: convex hull, 8: warping, 9: springback,...}. The offset of the characteristic points of the blank contour corresponds one by one to the numbers in the marking list, and a stamping defect database is constructed.

[0036] Furthermore, in step A4), 80% of the samples are randomly selected as the training data set, 10% of the samples are used as the validation data set, and 10% of the samples are used as the test data set.

[0037] Furthermore, in step C, by setting a number of laser distance sensors, adjust the horizontal and vertical heights using the mounting bracket to ensure that the measurement points coincide with the positions of the characteristic points of the blank contour of the workpiece stamped parts, and measure the offset of each characteristic point along the material inflow direction.

[0038] The beneficial effects of the present invention:

[0039] The intelligent modeling for stamping forming quality inspection involved in the present invention is mainly based on accurate finite element simulation. Therefore, the cost of obtaining data samples is low, the modeling efficiency is high, and the detected defects basically cover the vast majority of the current types of stamping drawing defects, far exceeding the existing machine vision inspection methods in this regard. The method of the present invention greatly alleviates the manual dependence of the current stamping forming quality inspection and improves the product quality inspection efficiency. At the same time, the method of the present invention provides a basis for stamping die debugging, changes the current mode mainly based on the experience of die fitters, improves the trial die efficiency, and thus reduces the enterprise cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly introduce the working process of the embodiments of the present invention, the following gives a brief introduction to the drawings required in the description of the embodiments:

[0041] Figure 1 is a flowchart of an intelligent detection method for determining forming defects of the blank contour of a stamping part according to the present invention;

[0042] Figure 2 is a schematic diagram of the offset amount of the characteristic points of the blank contour after calculation and simulation according to the present invention; the short dashed line in the figure is the blank line, and the solid line is the feeding line;

[0043] Figure 3 is a schematic diagram of the neural network model according to the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will describe the present invention in detail with reference to the drawings.

[0045] The present invention provides an intelligent detection method for determining forming defects of the blank contour of a stamping part, and its specific flowchart is as Figure 1 shown, and specifically includes the following steps:

[0046] S1. Stamping defect simulation: Establish a three-dimensional geometric simulation model of the die and the blank, input the accurately measured material model, optimized process parameters, boundary conditions close to the actual situation, and stamping action processes, simulate the stamping forming of a defect-free product, and output the corresponding blank contour simulation results; according to the actual process situation, set corresponding deviations based on the input parameters of the defect-free product simulation, including setting original blank defects at critical points, changing the material model, process parameters, and boundary conditions, etc., simulate various stamping defects during the forming process, and output the corresponding blank contour simulation results;

[0047] The original blank defects include: uneven thickness, uneven surface roughness, dark pits, hidden damages, convex hulls, etc.; the material parameters include: parameters related to yield and hardening models, etc.; the process parameters include: die clearance, blank holding force, draw bead coefficient, etc.; the boundary conditions include: punch and die shapes, friction coefficient, stamping speed, etc.

[0048] S2. Database construction: The finite element output data needs to be processed to obtain a database for the neural network. The specific steps of step S2 are as follows:

[0049] (1) Calculate the offset of the characteristic points of the blank contour after simulation. As Figure 2 shown, arrange several characteristic points along the original blank contour. With the help of CAD software, calculate the offsets of the characteristic points of the stamping blank contour and the original blank contour in the case of no defects and various defects simulated by finite element. The offset is the distance difference of the blank contour along the material inflow direction before and after stamping. Generally, the material inflow direction is the normal direction of the blanking boundary line.

[0050] The selection rules of the characteristic points are as follows: evenly distribute along the perimeter of the contour, and increase or decrease according to the evenness of the material inflow amount, that is, increase the number of its settings in the area where the inflow amount changes greatly;

[0051] (2) Evaluate the quality of the stamping parts simulated by finite element and classify them manually. It can be divided into two categories: one is normal stamping parts, marked as "normal"; the other is stamping parts with defects. Subdivide the existing stamping defects and mark them in sequence. The label list is: {0: normal, 1: fracture, 2: wrinkling, 3: surface abrasion, 4: press / indentation, 4: collapse, 6: dark pit, 7: convex hull, 8: warping, 9: springback,...}. The offset of the blank contour characteristic points corresponds one by one to the numbers in the label list to construct a stamping defect database.

[0052] (3) When establishing the database, randomly select 80% of the samples as the training set, 10% of the samples as the validation set, and 10% of the samples as the test set. After training is completed, the test set does not need to make text labels, and the defect type is predicted under the model of the completed neural network training.

[0053] S3. Stamping prediction and classification: Construct a neural network to realize the prediction of stamping part defects by the blank contour of the stamping part. The specific steps of step S3 are as follows:

[0054] (1) In the present invention, a BP neural network is selected. The BP neural network includes three parts: the input layer, hidden layer and output layer of data. Its specific structure is as Figure 3 shown. The input layer inputs the offset of the blank contour characteristic points, and the number of its neurons is equal to the number of the selected characteristic points. The hidden layer is the connection structure of each neuron in the network. Here, two hidden layers are selected. The output layer outputs the classification labels of the test set. When the input data, model parameters are given and the network structure is determined, the result is calculated through the forward propagation process. It is also necessary to optimize the network parameters by using the backpropagation and gradient descent algorithms. By inputting the data into the network and continuously performing the backpropagation optimization algorithm, the optimal network parameters are finally obtained.

[0055] The activation function used in the network is the Rectified Linear Unit (ReLU), and its formula is:

[0056] f(x) = max(0, x)

[0057] where x is the input of the neuron. Compared with training using activation functions such as Tanh or Sigmoid, the training speed of the ReLU activation function is much faster.

[0058] The optimizer selects the Adaptive Moment Estimation (Adam). Adam is an adaptive parameter update algorithm that can change the value of the learning rate during training and provides an adaptive learning rate based on the first-order moment estimate and second-order moment estimate of the gradient.

[0059] The activation function of the output layer is the Softmax classifier, which converts the stamping defect output result into a probability value. Its calculation formula is as follows:

[0060]

[0061] where a i represents the i-th value of the vector input to the Softmax classifier;

[0062] y i represents the i-th value of the output result y, that is, the probability that the input sample is predicted to belong to class i;

[0063] a j represents the j-th value of the vector input to the Softmax classifier, and T represents the number of defect categories;

[0064] (2) Design the loss function of the defect determination network model, use the training set sample images to train the model, and select the accuracy as the monitoring metric; then use the validation set to further supervise the training of the model, adjust the parameters in a timely manner, and accelerate the training progress. The loss function used here is the binary cross-entropy loss function, and its definition formula is:

[0065]

[0066] where y n is the true value of the n-th sample;

[0067] is the predicted value of the n-th sample;

[0068] N is the total number of test samples;

[0069] (3) Use the trained intelligent judgment model to perform stamping defect judgment and classification, output the types of defect judgments, and count the defect judgment results. During production, the inflow of the sheet material can be quickly and accurately measured by laser ranging. The laser ranging sensor can ensure the coincidence of the measurement point and the position of the contour feature point of the workpiece by adjusting the horizontal and vertical heights of the mounting bracket, and measure the offset of the feature point along the material inflow direction. As the input of the neural network, use the trained neural network to quickly and accurately judge the stamping defects.

Claims

1. An intelligent detection method for determining forming defects based on the offset of the blank contour of a stamping part, comprising the following steps: A. Construct a dataset of the offset amounts of the feature points of the blank contour; 1) Perform several stamping simulation simulations using finite element software; Simulate various defective and non-defective situations during the stamping process. The defects include: fracture, wrinkling, surface abrasion, press / indentation, collapse, dark pit, convexity, warping, and springback; respectively output the blank contour lines corresponding to various situations when the die is fully closed; 2) Arrange several feature points along the original blank contour, and measure the offset amounts of the feature points along the material inflow direction; the offset amount is the distance difference of the blank contour along the material inflow direction before and after stamping. The selection rules for the feature points are as follows: evenly distribute them around the contour, and increase or decrease them according to the uniformity of the material inflow amount, that is, increase the setting quantity in the area where the inflow amount changes greatly; 3) With the aid of CAD software, calculate the offset amounts of the feature points of the blank of the stamping part and the original blank contour in the non-defective and various defective situations simulated by finite element; 4) Manually label normal and defect types according to the stamping results after simulation, save the offset amounts of the feature points and defect types of each data sample in the format of a dataset, and randomly divide them into a training dataset, a validation dataset, and a test dataset; B. Construct a BP neural network model; The BP neural network includes 3 parts: an input layer, a two-layer hidden layer, and an output layer; The input layer inputs the distance deviation values of the feature points of the blank contour, and the output layer outputs classification labels, and the types include: normal, fracture, wrinkling, surface abrasion, press / indentation, collapse, dark pit, convexity, warping, and springback; The neuron activation function selects the rectified linear unit ReLU, and its formula is: f(x) = max(0, x) where x is the input of the neuron; The activation function of the output layer is a Softmax classifier, which converts the output result of the stamping defect into a probability value, and its calculation formula is as follows: where a i represents the i-th value of the vector input to the Softmax classifier; y i represents the i-th value of the output result y, that is, the probability that the input sample is predicted to belong to class i; a j represents the j-th value of the vector input to the Softmax classifier, and T represents the number of defect categories; Calculate the result through the forward propagation process of the network model, and optimize the network parameters using the backpropagation and gradient descent algorithms; the optimizer selects the adaptive learning rate optimization algorithm Adam. Adam is an adaptive parameter update algorithm used to change the value of the learning rate during training, and provides an adaptive learning rate according to the first-order moment estimate and second-order moment estimate of the gradient; Use the samples of the training dataset to train the model; then use the validation dataset to further supervise the training situation of the model, adjust the parameters in a timely manner, and accelerate the training progress; C. Measure the offset amount of the feature points of the blank contour along the material inflow direction after actual stamping, and use it as the input value of the trained BP neural network model to output the defect classification result.

2. The intelligent detection method for determining forming defects based on the offset of the blank contour of the stamping part according to claim 1, characterized in that, The loss function adopted in step B is the cross-entropy loss function, and its definition formula is: where y n is the true value of the nth sample; is the predicted value for the nth sample; N is the total number of test samples.

3. The intelligent detection method for determining forming defects based on the offset of the blank contour of the stamping part according to claim 1, wherein In step A1), the specific steps of performing stamping simulation simulation using finite element software are as follows: A 3D geometric simulation model of the die and the blank is established. The accurately measured material model, the optimized process parameters, the near-actual boundary conditions, and the stamping operation processes are input to simulate the stamping forming of defect-free products, and the corresponding blank contour simulation results are output. According to the actual process conditions, corresponding deviations are set based on the input parameters of the defect-free product simulation, including setting original blank defects at critical points, changing material parameters, process parameters, and boundary conditions, simulating various stamping defects during the forming process, and outputting the corresponding blank contour simulation results. The original blank defects include: uneven thickness, uneven surface roughness, dark pits, hidden damages, and convex bumps. The material parameters include: parameters related to the yield and hardening models. The process parameters include: die clearance, blank holder force, and draw bead coefficient. The boundary conditions include: the shapes of the punch and die, the friction coefficient, and the stamping speed.

4. The intelligent detection method for determining forming defects based on the profile offset of the stamping blank according to claim 1, wherein In step A4), the method for manually marking the defect types according to the stamping results after simulation is as follows: Evaluate the quality of the stamped parts simulated by finite element, and classify them manually. It can be divided into two categories: one is normal stamped parts, marked as "normal"; the other is stamped parts with defects. The existing stamping defects are further divided and marked in sequence. The label list is: {0: normal, 1: fracture, 2: wrinkling, 3: surface abrasion, 4: indentation / dent, 4: collapse, 6: dark pit, 7: convex bump, 8: warping, 9: springback}. The offset of the characteristic points of the blank contour corresponds one by one to the numbers in the marking list, and a stamping defect database is constructed.

5. The intelligent detection method for determining forming defects by stamping blank contour offset according to claim 1, characterized in that, In step A4), 80% of the samples are randomly selected as the training data set, 10% of the samples are used as the validation data set, and 10% of the samples are used as the test data set.

6. The intelligent detection method for determining forming defects by stamping blank contour offset according to claim 1, characterized in that In step C, by setting a number of laser distance sensors, the horizontal and vertical heights are adjusted using the mounting bracket to ensure that the measurement points coincide with the positions of the characteristic points of the blank contour of the workpiece stamped part, and the offset of each characteristic point along the material flow-in direction is measured.

Images