Rebound angle compensation method for bending heat treatment of spring frame

Abstract

The invention discloses a spring frame bending heat treatment rebound angle compensation method. A hybrid convolution feature extraction module, a space attention feature optimization module and a full connection layer output module are included. On the basis of seven technological parameters, confirmed in prior, of spring frame bending and heat treatment, convolution operation is conducted on the seven technological parameters through six one-dimensional convolution kernels of different forms, and feature vectors of different scales are extracted. A space attention mechanism is used for carrying out attention calibration optimization on the feature map, effective features are highlighted, invalid features are inhibited, dimension reduction is carried out on a feature map channel, fitting calculation of multiple factors on single output is achieved by building a full-connection neural network, and a bending angle needing to be set on a numerical control bending machine is output. High-precision, end-to-end and automatic spring frame bending heat treatment springback angle compensation is achieved on the whole, and therefore the purposes of reducing the production cost of enterprises and improving the product quality are achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of spring frames, and in particular relates to a method for compensating a spring frame's springback angle during bending heat treatment. Background Art

[0002] The car seat spring frame is an important component in the automotive industry. The cold-drawn carbon spring steel wire used in it has high strength, hardness and excellent elastic recovery. A simple bending mold can be used to form spring workpieces of various shapes and specifications. The manufacturing process is simple, which makes the spring frame made by it widely used in furniture and car seats, machinery, electronics, aerospace and other fields, with good versatility and applicability. During the bending and forming process of the spring frame, the bending external force causes plastic deformation and elastic deformation of the spring frame at the same time. The plastic deformation part makes it retain the shape after bending, and the elastic deformation part makes the spring frame rebound after the external force is unloaded. At the same time, after bending, the internal grains of high-carbon spring steel are deformed to form a stress concentration area. In terms of process, it is necessary to remove the stress through tempering heat treatment. During the tempering process, the grains of the spring frame steel wire will recrystallize, resulting in a slight increase in grain size and structural relaxation, causing a greater degree of rebound. The springback during the bending heat treatment of the spring frame wire is inevitable in the actual production process, but it will greatly affect the bending angle accuracy of the spring, thereby causing defects in the mechanical properties and service life of the product, and have a negative impact on the company's production quality control and product profits.

[0003] The bending springback phenomenon of spring frame steel wire is affected by many complex factors. In traditional practical operations, in order to reduce or eliminate the springback angle of the spring frame, it is generally necessary to rely on the experience of the technician and repeated angle measurements and manual compensation to determine an approximate angle compensation value, which consumes a lot of manpower and material resources. Moreover, due to the uncertainty of the amount of experience, the compensation result is often not ideal. Therefore, a mathematical analytical method for springback prediction was proposed later, but the high nonlinearity of its mathematical model makes it difficult to express it using a definite mathematical analytical model. In recent years, with the rapid development of artificial intelligence, neural network models have been used for multi-factor model fitting and numerical prediction under more and more complex conditions. It has a huge number of neuron nodes to characterize the mapping relationship between input and output variables, and has high generalization, strong learning ability and excellent nonlinear fitting ability.

[0004] At present, only the methods for predicting and compensating the stamping springback angle of flat plates are discussed, and there is a lack of research on the relevant theoretical and technical methods for compensating the springback angle of the bending heat treatment of cylindrical spring frame steel wire. China Patent Publication No. CN110147602B discloses a method for establishing a bending springback angle prediction model, which is mainly for the stamping forming process of square plates. After obtaining the training samples of the stamping springback angle, the radial basis function approximation model is established using four factors: sheet thickness, elastic modulus, upper die corner radius and lower die opening width, which is used to predict the springback angle of the sheet after stamping. Then, the bending angle is differentiated in the downward pressure calculation formula to obtain the downward pressure compensation value to finally complete the angle compensation. This angle compensation method does not take the heat treatment factor into consideration; and the radial basis function approximation model is usually a local fitting model, and the fitting effect of multi-factor high-dimensional data of global nature is not as good as that of global models such as linear regression networks; in addition, this method performs angle compensation in steps, and its end-to-end capability is insufficient and there is a certain degree of complexity. Summary of the invention

[0005] In order to solve the problems raised in the above background technology, the present invention provides a method for compensating the springback angle of a spring frame during bending heat treatment, which realizes high-precision and automated springback angle compensation of a spring frame during bending heat treatment.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for compensating the springback angle of a spring frame during bending heat treatment, comprising the following steps:

[0007] S101: for each spring frame steel wire to be bent, determine its production process data, and use different one-dimensional convolution kernels to perform mixed convolution calculation on the one-dimensional vector composed of the production process data to obtain feature vectors of different scales; wherein the production process data includes material properties of the spring frame steel wire, diameter of the spring frame steel wire, bending machine related parameters, angle at which the spring frame steel wire needs to be bent and formed, and heat treatment related process parameters;

[0008] S102: Concatenate the feature vectors of different scales calculated in step S101 in the direction of vector length to obtain a feature vector, and perform spatial attention descriptor M S (F) training calculation, and then the spatial attention descriptor M S (F) Multiply the input feature vector to obtain the feature vector after attention calibration;

[0009] S103: performing channel dimension reduction on the vector after spatial attention calibration in step S102, and then using a fully connected neural network to perform multi-factor fitting single-output neural network fitting calculation, fitting calculation of the eigenvalues ​​in the optimized eigenvector, and outputting the final bending angle that needs to be set on the bending machine;

[0010] S104: inputting the angle value required to be set into a CNC wire bending machine for bending to achieve automatic compensation of the bending rebound angle;

[0011] Furthermore, the material properties of the spring frame steel wire are the material elastic modulus and the material yield strength, which are obtained through mechanical testing of the spring frame steel wire materials; the heat treatment related process parameters are the tempering temperature and the tempering time.

[0012] Furthermore, the bending machine includes an outer mold, an inner mold, an inner bending head and an outer bending head, the inner bending head and the outer bending head are respectively located on the inner mold and the outer mold, the inner mold is used to fix the spring frame steel wire, the outer mold is driven by a driving device to rotate around the inner mold, and the outer bending head is used to drive the spring frame steel wire to rotate around the inner bending head for bending; the bending machine related parameters refer to the radius of the inner bending head or the inner mold.

[0013] Furthermore, six different one-dimensional convolution kernels are used to perform mixed convolution calculation on the one-dimensional vector composed of the production process data to obtain six feature vectors of different scales; the specific convolution kernels are:

[0014] The first convolution kernel size is 1, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 7×1×64;

[0015] The second convolution kernel size is 3, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 2, and the size of the output feature vector obtained by convolution is 9×1×64;

[0016] The third convolution kernel size is 4, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 3, and the size of the output feature vector obtained by convolution is 10×1×64;

[0017] The size of the fourth convolution kernel is 2, the number is 64, and the dilation is 2. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 5×1×64;

[0018] The size of the fifth convolution kernel is 3, the number is 64, and the dilation is 2. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 3×1×64;

[0019] The size of the sixth convolution kernel is 7, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 6, and the size of the output feature vector obtained by convolution is 13×1×64;

[0020] After each convolution, the ReLU activation function is used to calculate once and activate the output.

[0021] Furthermore, the six feature vectors obtained in step S101 are concatenated in the length direction of the vector, so as to obtain a feature vector of 47×1×64;

[0022] During the training phase, the spatial attention descriptor M is performed on the 47×1×64 feature vector length × channel number plane. S The training calculation formula of (F) is:

[0023] M S (F) = σ{f (3×3) (F [W×C] )} (1)

[0024] Where F is the input feature vector; f is the convolution step, here a 3×3 two-dimensional convolution is used; σ is the Sigmoid function;

[0025] Then the spatial attention descriptor M S (F) Multiply the input feature vector to obtain the attention-calibrated feature vector.

[0026] Furthermore, the fully connected neural network structure is:

[0027] The number of input layer nodes is 47, corresponding to the 47 data of the input vector, and finally a ReLU function is used for activation.

[0028] The number of nodes in hidden layer 1 is 16, and a ReLU function is used for activation.

[0029] The number of nodes in the hidden layer 2 is 8, and a ReLU function is used for activation.

[0030] The number of nodes in hidden layer 3 is 4, and a ReLU function is used for activation at the end.

[0031] The number of nodes in the output layer is 1, and finally a ReLU function is used for activation, at which point the final bending angle that needs to be set on the CNC bending machine is obtained.

[0032] Compared with the prior art, the beneficial effects of the present invention are: combining multi-modal one-dimensional mixed convolution feature extraction and spatial attention mechanism, and integrating them into the fully connected neural network model of multi-factor fitting single factor, so as to efficiently and accurately calculate the angle value to be set, thereby realizing the compensation of bending springback angle. The input process factor data group is respectively extracted by one-dimensional convolution kernels of 6 different forms, which fully considers the influence correlation between various factors and improves the effectiveness and comprehensiveness of feature extraction; the present invention fully integrates the features of the 6 vector groups extracted in the previous stage by introducing the spatial attention mechanism, and automatically optimizes the features, thereby enhancing useful features and suppressing noise features; the present invention finally directly outputs the bending angle value to be set for angle compensation through a multi-layer fully connected neural network, eliminating other calculation steps and manual intervention, and realizing overall efficient, high-precision, end-to-end spring frame bending springback angle compensation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The overall algorithm flow chart of the spring frame bending heat treatment springback angle compensation of the present invention is shown.

[0034] Figure 2 The form of each convolution kernel and its convolution method diagram in the hybrid convolution feature extraction of the present invention are shown.

[0035] Figure 3 The structural diagram of the spatial attention feature optimization module in the present invention is shown.

[0036] Figure 4 The diagram shows the structure of a multi-layer fully connected network module in the present invention.

[0037] Figure 5 The figure shows the structure of the spring frame wire bending tool used in the present invention. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] The bending machine used in the spring frame bending heat treatment springback angle compensation method provided in this embodiment is as follows: Figure 5As shown, the bending machine includes an outer mold 1, an inner mold 2, a first inner bending head 4, a second inner bending head 5 and an outer bending head 6, wherein the inner bending head and the outer bending head are respectively located on the inner mold and the outer mold, wherein the inner mold 2 is used to fix the spring frame wire 3, the outer mold 1 is driven by a driving device to rotate around the inner mold 2, and the outer bending head is used to drive the spring frame wire 3 to rotate around the inner bending head for bending. Here, the first inner bending head and the second inner bending head are two bending heads with different diameters, which can be selected according to process requirements.

[0040] The bending machine related parameters refer to the diameter or radius of the inner bending head.

[0041] The specific compensation method includes the following steps:

[0042] S101: For each spring frame steel wire to be bent, its production process data is determined, and six different one-dimensional convolution kernels are used to perform mixed convolution calculation on the one-dimensional vector composed of the production process data to obtain six feature vectors of different scales.

[0043] When acquiring production process data, the seven key production process data that need to be acquired are material elastic modulus (GPa), material yield strength (MPa), steel wire diameter (mm), inner mold radius (mm), required bending angle (°), tempering temperature (℃), and tempering time (mins). The inner mold radius can also be replaced by the parameters of the inner bending head. The material elastic modulus and material yield strength belong to the properties of the spring frame steel wire itself, which are obtained based on the mechanical test of each batch of incoming steel wire. The steel wire diameter is the nominal diameter of the current spring frame steel wire to be bent. The required bending angle is the final spring frame steel wire bending forming angle required for this process, which is determined by the workpiece design requirements. The tempering temperature and tempering time are heat treatment parameters determined in the process flow.

[0044] like Figure 2 As shown:

[0045] The convolution kernel size used in the first convolution is 1, the number is 64, the convolution kernel dilation coefficient is 1, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 0. The first position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves 6 times and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 7×1×64.

[0046] The convolution kernel size used in the second convolution is 3, the number is 64, the convolution kernel dilation coefficient is 1, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 2. The third position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves 8 times and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 9×1×64.

[0047] The convolution kernel size used in the third convolution is 4, the number is 64, the convolution kernel dilation coefficient is 1, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 3. The fourth position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves 9 times and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 10×1×64.

[0048] The convolution kernel size used in the fourth convolution is 2, the number is 64, the convolution kernel dilation coefficient is 2, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 0. The first position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves 4 times and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 5×1×64.

[0049] The convolution kernel size used in the fifth convolution is 3, the number is 64, the convolution kernel dilation coefficient is 2, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 0. The first position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves twice and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 3×1×64.

[0050] For the sixth convolution, the convolution kernel size used is 7, the number is 64, the convolution kernel dilation coefficient is 1, and the boundary expansion coefficient (padding) of the data group 7×1 vector is 6. The seventh position of the convolution kernel is aligned with the first position of the data group 7×1 vector to start convolution. The convolution kernel moves 12 times and is finally activated using the ReLU activation function. The output feature vector size W×H×C is 13×1×64.

[0051] It should be noted that in the neural network model training stage of the present invention, it is necessary to obtain multiple sets of the above 7 key production process data and 1 input bending angle of the CNC bending machine as input training samples, and perform network training on them to obtain more robust model parameters. In the inference stage, according to the obtained model parameters, the final bending angle to be set can be directly inferred for each independent set of production process data, that is, only one input data set is required in the inference stage.

[0052] S102: The six feature vectors of different scales calculated in step S101 are concatenated in the direction of the vector length, and the concatenated vectors are calibrated for spatial attention to obtain a calibrated feature vector. Figure 3 As shown, the specific steps include the following:

[0053] The six feature vectors obtained in step S101 are concatenated in the length direction of the vector to obtain a 47×1×64 feature vector.

[0054] During the training phase, the spatial attention descriptor M needs to be performed on the plane of length × number of channels (W×C) of the above 47×1×64 vector, i.e., the 47×64 plane. S The training calculation of (F) can be described by the formula:

[0055] M S (F) = σ{f (3×3) (F [W×C] )} (1)

[0056] Where F is the input feature vector; f is the convolution step, here a 3×3 two-dimensional convolution is used; σ is the Sigmoid function. Then the spatial attention descriptor M S (F) Multiplying the input feature vector to obtain the feature vector after attention calibration, the size of which is still 47×1×64.

[0057] In the inference phase, the spatial attention descriptor M S (F) As part of the trained model parameters, it does not need to be calculated again, and is directly multiplied with the input feature vector to obtain the feature vector after attention calibration. The spatial attention descriptor is used to assign weights to the features of different scales obtained by convolution, automatically focusing on more important features and ignoring unimportant features during the training process, improving the overall effectiveness of the features, and calibrating and calculating to obtain the spatial attention descriptor parameters.

[0058] In the inference stage, the spatial attention descriptor parameters of the known model are directly multiplied by the input feature vector to obtain the attention-calibrated feature vector, whose size is also 47×1×64.

[0059] S103: Perform channel dimension reduction on the vector after spatial attention calibration in step S102, and then use a fully connected neural network to calculate and output the final bending angle that needs to be set on the bending machine.

[0060] S103 specifically includes the following contents:

[0061] Preferably, S103 specifically includes:

[0062] After obtaining the 47×1×64-sized attention-calibrated feature vector, perform a 1×1 convolution on it to reduce the channel dimension, compress the original 64 channels into 1 channel, achieve feature simplification, and obtain a 47×1×1 feature vector. Then build the following Figure 4 The fully connected neural network shown is used to perform multi-factor fitting single-output neural network fitting calculation, and the 47 eigenvalues ​​in the optimized eigenvector are fitted and calculated to obtain the final result.

[0063] Specifically, the fully connected neural network has 1 input layer, 3 hidden layers, and 1 output layer. The specific number of nodes in each layer is [47, 16, 8, 4, 1]. After calculation in each layer, nonlinear activation is performed through the ReLU activation function, and finally 1 output value is obtained in the output layer, which is the angle value that needs to be set.

[0064] Using the attention-calibrated feature vector output in S102, a 1×1 convolution is performed to compress the number of channels to 1, and a 47×1×1 one-dimensional vector is obtained. This vector is then used as input and a fully connected neural network is used to obtain a 1×1×1 output data.

[0065] Preferably, the fully connected neural network structure is:

[0066] The number of input layer nodes is 47, corresponding to the 47 data of the input vector, and finally a ReLU function is used for activation.

[0067] The number of nodes in hidden layer 1 is 16, and a ReLU function is used for activation.

[0068] The number of nodes in the hidden layer 2 is 8, and a ReLU function is used for activation.

[0069] The number of nodes in hidden layer 3 is 4, and a ReLU function is used for activation at the end.

[0070] The number of nodes in the output layer is 1, and finally a ReLU function is used for activation, at which point the final bending angle that needs to be set on the CNC bending machine is obtained.

[0071] S104: Input the angle value required to be set into a CNC wire bending machine for bending to achieve automatic compensation of the bending rebound angle.

[0072] The spring frame bending heat treatment springback angle compensation method based on hybrid convolution and spatial attention neural network of the present invention creatively combines multi-modal one-dimensional hybrid convolution feature extraction and spatial attention mechanism and integrates it into the multi-factor fitting single-factor fully connected neural network model, which is used to efficiently and accurately calculate the angle value to be set, thereby realizing the bending springback angle compensation. The input process factor data group is respectively extracted by one-dimensional convolution kernels of 6 different forms, which fully considers the influence correlation between various factors and improves the effectiveness and comprehensiveness of feature extraction; the present invention fully integrates the features of the 6 vector groups extracted in the previous stage by introducing the spatial attention mechanism, and automatically optimizes the features, thereby enhancing useful features and suppressing noise features; the present invention finally directly outputs the bending angle value to be set for angle compensation through a multi-layer fully connected neural network, eliminating other calculation steps and manual intervention, and realizing overall efficient, high-precision, end-to-end spring frame bending springback angle compensation.

[0073] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for compensating the springback angle of a spring frame during bending heat treatment, characterized in that: The steps include: S101: for each spring frame steel wire to be bent, determine its production process data, and use different one-dimensional convolution kernels to perform mixed convolution calculation on the one-dimensional vector composed of the production process data to obtain feature vectors of different scales; wherein the production process data includes material properties of the spring frame steel wire, diameter of the spring frame steel wire, bending machine related parameters, angle at which the spring frame steel wire needs to be bent and formed, and heat treatment related process parameters; S102: Concatenate the feature vectors of different scales calculated in step S101 in the direction of vector length to obtain a feature vector, and perform spatial attention descriptor M S (F) training calculation, and then the spatial attention descriptor M S (F) Multiply the input feature vector to obtain the feature vector after attention calibration; S103: performing channel dimension reduction on the feature vector after spatial attention calibration in step S102, and then using a fully connected neural network to perform multi-factor fitting single-output neural network fitting calculation, fitting calculation of the eigenvalues ​​in the optimized feature vector, and outputting the final bending angle that needs to be set on the bending machine; S104: inputting the angle value required to be set into a CNC wire bending machine for bending to achieve automatic compensation of the bending rebound angle; 2. The method for compensating the spring frame springback angle during bending heat treatment according to claim 1, characterized in that: The material properties of the spring frame steel wire are the material elastic modulus and the material yield strength, which are obtained through mechanical testing of the spring frame steel wire incoming materials; the heat treatment related process parameters are the tempering temperature and the tempering time.

3. The method for compensating the spring frame bending heat treatment springback angle according to claim 1, characterized in that: The bending machine includes an outer mold, an inner mold, an inner bending head and an outer bending head, the inner bending head and the outer bending head are respectively located on the inner mold and the outer mold, the inner mold is used to fix the spring frame steel wire, the outer mold is driven by a driving device to rotate around the inner mold, and the outer bending head is used to drive the spring frame steel wire to rotate around the inner bending head for bending; the bending machine related parameters refer to the radius of the inner bending head or the inner mold.

4. The method for compensating the spring frame bending heat treatment springback angle according to claim 2, characterized in that: Six different one-dimensional convolution kernels are used to perform mixed convolution calculation on the one-dimensional vector composed of production process data to obtain six feature vectors of different scales; the specific convolution kernels are: The first convolution kernel size is 1, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 7×1×64; The second convolution kernel size is 3, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 2, and the size of the output feature vector obtained by convolution is 9×1×64; The third convolution kernel size is 4, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 3, and the size of the output feature vector obtained by convolution is 10×1×64; The size of the fourth convolution kernel is 2, the number is 64, and the dilation is 2. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 5×1×64; The size of the fifth convolution kernel is 3, the number is 64, and the dilation is 2. At this time, the padding of the 7×1 vector composed of 7 production process data is 0, and the size of the output feature vector obtained by convolution is 3×1×64; The size of the sixth convolution kernel is 7, the number is 64, and the dilation is 1. At this time, the padding of the 7×1 vector composed of 7 production process data is 6, and the size of the output feature vector obtained by convolution is 13×1×64; After each convolution, the ReLU activation function is used to calculate once and activate the output.

5. The method for compensating the spring frame bending heat treatment springback angle according to claim 1, characterized in that: The six feature vectors obtained in step S101 are concatenated in the length direction of the vector to obtain a feature vector of 47×1×64; During the training phase, the spatial attention descriptor M is performed on the 47×1×64 feature vector length × channel number plane. S The training calculation formula of (F) is: M S (F)=σ{f (3×3) (F [W×C] )} (1) Where F is the input feature vector; f is the convolution step, here a 3×3 two-dimensional convolution is used; σ is the Sigmoid function; Then the spatial attention descriptor M S (F) Multiply the input feature vector to obtain the attention-calibrated feature vector.

6. The method for compensating the spring frame springback angle during bending heat treatment according to claim 5, characterized in that: The fully connected neural network structure is: The number of input layer nodes is 47, corresponding to the 47 data of the input vector, and finally a ReLU function is used for activation; The number of nodes in hidden layer 1 is 16, and a ReLU function is used for activation at the end; The number of nodes in hidden layer 2 is 8, and a ReLU function is used for activation at the end; The number of nodes in hidden layer 3 is 4, and a ReLU function is used for activation at the end; The number of nodes in the output layer is 1, and finally a ReLU function is used for activation, at which point the final bending angle that needs to be set on the CNC bending machine is obtained.

Images