Machining surface plastic deformation measurement method and system based on deep learning

Abstract

The invention discloses a deep learning-based machining surface plastic deformation measurement method and system, and belongs to the technical field of metal cutting machining. According to the method, a parallel encoder-decoder structure is adopted, and spatial and temporal features and spatial features of a deformed image pair are fully extracted through multi-layer multi-scale convolution contraction and expansion and are used for predicting a deformation field. A double-attention mechanism module is introduced, position and channel characteristics are combined more effectively, efficient and high-precision measurement of a material deformation field is achieved, and a traditional digital image correlation method can be effectively replaced. According to the method, a Boolean model is used for generating an initial speckle capable of approaching a real environment, a simulated deformation image pair is obtained through image distortion, and the availability of a virtual data set is evaluated by using an autocorrelation full width at half maximum, a system error, a random error and an average gray scale gradient; a large number of high-quality data sets for deep learning training are efficiently provided, so that the model is free from limitation of a real data set, and the generalization of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of metal cutting processing, and specifically relates to a method and system for measuring plastic deformation of a processing surface based on deep learning. Background Art

[0002] The plastic deformation and strain distribution of the machined surface are two key surface integrity parameters used to evaluate the performance and life of the workpiece. During the cutting process, the machined material will be subjected to very high thermal and mechanical loads, which may cause a high degree of plastic deformation on the machined surface. Highly non-uniform plastic deformation will change the residual stress state of the workpiece surface and even cause metallographic changes in the microstructure of the workpiece surface, thereby affecting the surface integrity of the workpiece. Therefore, experimental measurement of plastic deformation and strain distribution of the machined surface is crucial. The patent with publication number CN108426796A proposes a method for calculating the strain rate of the cutting process based on grid flow differential. A grid is machined on the surface of the material to be tested, and the strain rate distribution of the cutting process is determined by measuring the grid deformation.

[0003] The in-situ imaging technology based on the configuration of a high-speed camera and a microscope lens is a non-contact optical measurement method. Due to its extremely high applicability, it is widely used in deformation observation in the field of metal cutting. The patent with publication number CN109333162A discloses an online measurement system and method for high-speed cutting deformation field, which realizes non-contact measurement of transient deformation field and temperature field under high speed and high strain through dual-frame cameras, infrared cameras, etc. The patent with publication number CN106949858B proposes a device and method for measuring the strain performance of high-speed cutting deformation zone using force-sensitive particles, and uses a high-speed camera and a photoelectric conversion system to realize full-field measurement of material deformation under non-contact state. The above method improves the measurement accuracy of material deformation field by innovating the hardware of the measurement system. However, the complex structure leads to high cost and the system reliability needs to be improved.

[0004] To extract deformation field information from cutting process images captured by high-speed cameras, the current mainstream method is to use digital image correlation (DIC) technology to match the position of the same pixel subset in different images to obtain the global deformation field. However, the calculation accuracy of the traditional DIC method for different images and deformation scales is heavily dependent on human experience, and may even fail for highly non-uniform deformations, and the calculation efficiency is low. Therefore, it is of great significance to study a deformation field extraction method with high calculation accuracy and efficiency, and combine it with a high-speed camera to realize the measurement of plastic deformation of the machined surface. Summary of the invention

[0005] In view of the above defects and improvement needs of the prior art, in order to realize the measurement of plastic deformation of the machining surface and further understand the cutting mechanism research, the present invention provides a method and system for measuring plastic deformation of the machining surface based on deep learning. The cutting dynamic image is captured by a high-speed camera, and the deformation field in the dynamic image is extracted by a deep learning model. The hardware equipment is simple, the measurement accuracy is high, and it has good reliability.

[0006] The technical solution adopted by the present invention is:

[0007] A method for measuring plastic deformation of a machined surface based on deep learning, comprising the following steps:

[0008] S1. Use multi-layer convolutional neural network structure to build a deep learning model;

[0009] S2. Generate a reference image and a two-dimensional deformation field, distort the reference image, and generate a virtual data set for model training;

[0010] S3. Use evaluation indicators to evaluate the availability of the dataset generated in S2;

[0011] S4. Based on the models and virtual datasets in S1 and S2, train a deep learning model;

[0012] S5. Build a plastic deformation measurement system for the machined surface to collect images of the cutting process, and use the deep learning model trained in S4 to measure the plastic deformation of the machined surface.

[0013] A deep learning-based plastic deformation measurement system for a machined surface comprises a high-speed camera, a coaxial LED, a workpiece fixture, a displacement adjustment platform and a computer; the high-speed camera is installed on the displacement adjustment platform, the workpiece fixture is used to clamp a workpiece to be measured, the light path is changed by a reflector arranged between the high-speed camera and the workpiece fixture, and the surface of the workpiece to be measured is polished to form random scratches; the coaxial LED is used to illuminate a cutting area to capture continuous dynamic images of the cutting area; the computer stores a computer program for implementing the steps of any one of claims 1 to 8 of a deep learning-based plastic deformation measurement method for a machined surface, and measuring the deformation field of the continuous dynamic images of the cutting area captured by the high-speed camera.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. The measurement method of the present invention adopts a parallel encoder-decoder structure, and fully extracts the spatiotemporal and spatial features of the deformed image pair through multi-layer multi-scale convolution contraction and expansion, which is used to predict the deformation field. The dual attention mechanism module is introduced to more effectively combine the position and channel features to achieve efficient and high-precision measurement of the material deformation field, which can effectively replace the traditional digital image correlation method.

[0016] 2. The measurement method of the present invention uses a Boolean model to generate an initial speckle that can approximate the real environment, obtains a simulated deformed image pair through image distortion, and uses the autocorrelation half-maximum full width, systematic error, random error and average grayscale gradient to evaluate the availability of the virtual data set, efficiently providing a large number of high-quality data sets for deep learning training, freeing the model from the limitations of the real data set and improving the generalization of the model.

[0017] 3. The measuring system of the present invention uses a high-speed camera to capture the dynamic process of cutting. It has simple hardware equipment, high measurement accuracy and good reliability. Combined with the measuring method of the present invention, it can achieve efficient and high-precision measurement of plastic deformation of the machined surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention;

[0019] Figure 2 It is a schematic diagram of the deep learning model structure of the method of the present invention. ;

[0020] Figure 3 is a system schematic diagram of the present invention;

[0021] Figure 4 1 is a diagram of the plastic deformation measurement result of the machined surface of the present invention, wherein (a) is the deformation field u in the x direction, (b) is the deformation field v in the y direction, and (c) is the total deformation field;

[0022] Among them: 1. high-speed camera; 2. coaxial LED; 3. workpiece fixture; 4. displacement adjustment platform; 5. computer; 6. reflector; 7. workpiece to be measured; 8. tool. DETAILED DESCRIPTION

[0023] In order to better understand the purpose, structure and function of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0024] First, a method for measuring plastic deformation of a machined surface based on deep learning is described. The method comprises the following steps: Figure 1 , Figure 2 As shown,

[0025] S1. Build a deep learning model - Use a multi-layer convolutional neural network structure to build a deep learning model:

[0026] The deep learning model uses a multi-layer convolutional neural network to build an encoder-decoder structure and introduces a dual attention module.

[0027] The encoder adopts a parallel structure and consists of two modules: context encoder and feature encoder. The feature encoder processes the deformed image pairs independently, while the context encoder considers both time and space dimensions to form a hybrid feature pyramid of the deformed image pairs. The deformed image pairs are input into the two encoders in different ways to extract image features from different dimensions.

[0028] The feature encoder has two branches, each consisting of 4 convolutional layers, and a single input deformed image pair Extract multi-scale spatial feature pyramid F 1 k 、F 2 k (k=1,2,3,4), each layer of the pyramid contains a convolution layer, a batch normalization layer, an activation function, and a pooling layer. The number of channels is 8, 16, 32, and 64 respectively. Each layer of the pyramid reduces the image size by two times through pooling, and the resolution is reduced to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of the original image respectively.

[0029] The overall structure of the context encoder is similar to that of the feature encoder, which also uses a four-level pyramid structure. The difference is that the deformed image pairs are stacked and spliced ​​as Input context encoder to extract spatiotemporal hybrid feature pyramid M k The number of channels at each level of the pyramid is 16, 32, 64, and 128 respectively, and the step size of the pooling layer is the same as that of the feature pyramid to ensure that the size of the feature map at each level is the same as that of the feature encoder.

[0030] To further enhance the network’s ability to extract relevant and effective information, a dual attention module is added between the encoder and decoder. The dual attention module improves the relevance of local features to the global features by connecting the position attention and channel attention modules in parallel. First, three new feature maps B, C are generated by convolution. Change the shapes of B and C to Where n = h × w is the total number of pixels and c is the number of channels. Then, the feature map C is multiplied by the transpose of B and the softmax layer is used to calculate the spatial attention map. The calculation is as follows:

[0031]

[0032] Among them, s ji It indicates the influence of position i on position j, that is, the more similar the features of two positions are, the greater their correlation is. The feature map D is changed into Multiply S by it and change the result into Finally, multiply it by the scale factor α and sum it element by element with the feature map A to get the final output of the position attention module The calculation is as follows:

[0033]

[0034] Where A is the initial feature map, B, C, and D are the three new feature maps obtained after convolution of the initial feature map, E is the output feature map after position attention weights are assigned, and i and j are the position coordinates of the above feature maps;

[0035] The channel attention module is different from the position attention module. The initial feature map Directly calculate the channel attention map without convolution, that is, change the shape of A to And multiply it with its own transpose, and calculate the channel attention map through the softmax layer The calculation is as follows:

[0036]

[0037] where x ji It is used to measure the influence of channel i on channel j. The channel attention map X is multiplied by the transpose of the original feature map A and the shape is changed to Same as position attention, finally multiplied by a scale factor β to get the final output

[0038]

[0039] Where A is the initial feature map, E is the output feature map after the channel attention weights are assigned, i and j are the channel coordinates of the above feature map; x ji It is used to measure the influence of channel i on channel j, and β is the scale coefficient;

[0040] The deformed image pair is passed through the encoder to extract deep features to form F 1 4. F 2 4. M4, the three deep feature maps are concatenated into (F 1 4, F 2 4, M4) input the dual attention module, aggregate the global information through the two attention modules to generate two new feature maps, and complete the feature fusion by element-by-element summation as the input of the decoder.

[0041] The decoder adopts a five-level pyramid structure in the form of "expansion". In order to avoid the checkerboard artifacts caused by deconvolution, the structure of upsampling-convolution-batch normalization-activation function is adopted. The number of convolution channels is 256, 128, 64, 32, and 16. The upsampling expands the resolution by 2 times in sequence, and finally keeps the same as the input size to obtain the decoded feature map G.k (k=4, 3, 2, 1). It is worth noting that the decoder performs displacement prediction at the current resolution at each level to obtain the predicted displacement field pyramid Y k (k = 4, 3, 2, 1), and the spatial feature pyramid F obtained in the encoder 1 k 、F 2 k , hybrid feature pyramid M k , and the decoding feature pyramid G k Together they form the input of the next level decoder, F 1 k 、F 2 k 、M k The introduction of can avoid the loss of local feature information caused by convolution and pooling, and ensure the original resolution of the output deformation field. The final output layer uses a convolution layer to map the feature map to the displacement in the x and y directions at the original resolution, realizing the displacement extraction at the whole pixel level. The schematic diagram of the deep learning model structure is shown in Figure 2 As shown,

[0042] S2. Virtual dataset generation - Generate reference images and two-dimensional deformation fields, distort the reference images, and generate virtual datasets for model training:

[0043] The quality of image speckle is a key factor that directly affects the quality of the dataset. In order to create high-quality random speckle that conforms to the actual situation, this paper models the speckle image based on the Boolean model. A pattern composed of circular black spots is created on a white background as a reference image. The centers of the spots are evenly distributed on the image, and their radii are exponentially distributed. This infinite spatial resolution image is then integrated into the pixel-level image to obtain the reference image. This drawing scheme strictly follows the image acquisition chain in the digital camera to extract the digital image from the infinite resolution view.

[0044] When the material is deformed without fracture, the deformation field is continuous and smooth, and under the assumption of small deformation, the corresponding strain field can be calculated by the deformation field gradient. The two-dimensional Gaussian function has the characteristics of infinite differentiability and always satisfies the above conditions. Therefore, the local deformation field is composed of several two-dimensional Gaussian functions superimposed. The generalized deformation field Φ is defined as follows:

[0045]

[0046] in The two-dimensional Gaussian function representing the local deformation can be expressed as follows:

[0047]

[0048] Among them, the Gaussian function amplitude and standard deviation Determines the amplitude level and shape of the deformation field, considering random generation within a certain range, and the peak center Determines the location of the local deformation field.

[0049] Under the premise of knowing the reference image and deformation field, the image is distorted and a certain degree of white noise is added. The calculation formula is as follows:

[0050] I def =F(I ref ;Φ)+σ noise

[0051] In the formula, σ noise Represents the added white noise. def is the deformed image, I ref is the reference image, Φ is the deformation field;

[0052] Specifically, we first obtain the grid coordinates of each pixel displacement according to the deformation field, move each pixel of the reference image to the corresponding position, and then form a deformed image by interpolating the uniform grid. The reference and deformed images are randomly cropped to 256×256, white noise is added, and they are divided into validation set, test set, and training set according to 1:2:7.

[0053] Based on the infinitesimal strain assumption, the corresponding strain field can be analytically calculated by the spatial derivative of the deformation field. Therefore, the Green-Lagrange strain calculation on the two-dimensional plane is as follows:

[0054]

[0055] In the formula, ε xx , ε yy , ε xy represents the strain field in three directions, and u and v represent the deformation field in the x and y directions.

[0056] S3. Virtual dataset availability evaluation - Use evaluation indicators to evaluate the availability of the dataset generated in S2:

[0057] In order to characterize the quality of generated speckle, the image information is characterized. The autocorrelation half-maximum full width is selected as the characterization index of the speckle image size, and the autocorrelation function R(a) in the x direction is calculated as follows:

[0058]

[0059] in represents the average grayscale of the entire image, f(x) represents the grayscale value along the x-axis, and m represents the position along the x-axis.

[0060] The system error can be calculated using the interpolated deviation of the entire image, which can be approximated as a sine curve as shown below:

[0061] W≈Csin2πu

[0062] Where u is the unit sub-pixel displacement, C is the interpolation deviation amplitude, and the interpolation deviation kernel E(v x ,v y ) and power spectrum The calculation formula is as follows:

[0063]

[0064]

[0065] In the formula is the two-dimensional interpolation transfer function, The calculation formula is as follows:

[0066]

[0067] The random error is a function of the image noise and the sum of the squares of the grayscale gradients of the image subset, and the calculation formula is as follows:

[0068]

[0069] Where σ represents the standard deviation of image noise, f x (m,n) is the image gradient. In order to reflect the degree of global speckle random error, the average grayscale gradient is used to measure the average level of speckle quality. The calculation formula is as follows:

[0070]

[0071] in is the modulus of the image grayscale gradient vector, f x (x ij ), f y (x ij ) is the first-order grayscale gradient in the x and y directions. The above indicators are used to verify the availability of the virtual dataset.

[0072] S4. Model training - Based on the models and virtual datasets in S1 and S2, train the deep learning model:

[0073] The loss function of the present invention adopts the root mean square error RMSE. In order to compensate for the unclear gradient descent caused by too small RMSE, it is multiplied by an amplification factor, which is set to 5 in this paper. The RMSE can be calculated by the following formula:

[0074]

[0075] Where u pre(i,j) is the deformation prediction value, u gt (i,j) is the true value of the deformation.

[0076] The network is implemented on the tensorflow (version 2.10.0) platform. Adaptive moment estimation Adam is selected as the optimization method, and the two momentum parameters are set to β1 = 0.9 and β2 = 0.999. Adam can adaptively change the learning rate according to the current gradient, so as to have a faster convergence speed. To avoid gradient explosion or gradient disappearance during training, the initial learning rate is set to 0.0001. In order to balance the memory usage and training speed during training, the batch size is set to 8, and the model is trained for 200 epochs.

[0077] S5. Measurement system construction and surface plastic deformation measurement - Build a surface plastic deformation measurement system to collect cutting process images, and use the deep learning model trained in S4 to measure the surface plastic deformation: Figure 3 As shown,

[0078] The whole system is installed on a three-axis CNC milling machine. The high-speed camera 1 is fixed on the displacement adjustment platform 4 by bolts. The displacement adjustment platform 4 is isolated from the three-axis CNC milling machine to avoid the influence of processing vibration. The high-speed camera 1 is equipped with a microscope lens, and the cutting area is illuminated by the coaxial LED 2, and is connected to the computer through an optical fiber; the workpiece 7 to be measured is processed into a sheet and fixed on the workpiece fixture 3, and the surface to be observed is polished with sandpaper to form random scratches; the tool 8 is installed on the tool handle using a spring chuck. The captured image resolution is 960×480, and the spatial resolution is 2.11μm / pix. As the shooting frame rate of the high-speed camera 1 increases, the reduction of exposure time will significantly reduce the amount of light entering the lens, resulting in obvious sensor noise and reduced image quality. Therefore, the selection of frame rate and exposure time needs to be combined with the actual processing situation and the intensity of lighting.

[0079] Pure copper is selected as the workpiece material, with a size of 30×20×2mm. The observation surface is polished flat with #2000 sandpaper, and random scratches are formed to replace the speckle image. The speckle formed in this way does not require the introduction of additional materials on the workpiece surface, so it can change with the plastic deformation of the workpiece. The displacement adjustment platform 4 is adjusted so that the high-speed camera 1 is aimed at the cutting area to be observed, and the shooting parameters of the high-speed camera 1 are adjusted to make the cutting area clearly visible; while the tool 8 rotates forward, the high-speed camera 1 begins to capture the dynamic process of plastic deformation of the cutting surface; the captured image is input into the computer 5, and the deformation field is extracted from the dynamic image by the trained deep learning model to realize the measurement of the plastic deformation field of the processing surface. The image is cropped into 256×256 regions. Each group of deformed image pairs forms an input of 4×256×256×2, and the output is restored to a size of 960×480 in a reversible manner. In addition, the deformation and strain increment fields between frames are obtained by tracking adjacent images, and the final deformation and strain fields are formed by accumulating frame by frame. The measurement results are as follows Figure 4 shown.

[0080] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for measuring plastic deformation of a machined surface based on deep learning, characterized in that: The following steps are involved: S1. Use multi-layer convolutional neural network structure to build a deep learning model; S2. Generate a reference image and a two-dimensional deformation field, distort the reference image, and generate a virtual data set for model training; S3. Use evaluation indicators to evaluate the availability of the dataset generated in S2; S4. Train a deep learning model based on the models and virtual datasets in S1 and S2; S5. Build a plastic deformation measurement system for the machined surface to collect images of the cutting process, and use the deep learning model trained in S4 to measure the plastic deformation of the machined surface.

2. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 1, characterized in that: The deep learning model in S1 uses a multi-layer convolutional neural network to build an encoder-decoder structure, fully extracts the spatiotemporal and spatial features of the deformed image pairs through multi-layer and multi-scale convolution contraction and expansion, which are used to predict the deformation field, and introduces a dual attention module.

3. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 2, characterized in that: In the S1, the encoder adopts a parallel structure, and the encoder includes two modules: a context encoder and a feature encoder. The feature encoder processes the deformed image independently, and the context encoder considers both the time and space dimensions to form a mixed feature pyramid of the deformed image pair. The deformed image pair is input into the two encoders in different ways to extract image features from different dimensions.

4. A method for measuring plastic deformation of a machined surface based on deep learning according to claim 2 or 3, characterized in that: In S1, a dual attention module is added between the encoder and the decoder, and the dual attention module includes: Position attention module, the calculation formula is as follows: Where A is the initial feature map, D is the new feature map obtained after convolution of the initial feature map, E is the output feature map after weighting by position attention, i and j are the position coordinates of the above feature maps, and s ji It indicates the influence of position i on position j, and α is the scale coefficient; Channel attention module, the calculation formula is as follows: Where A is the initial feature map, E is the output feature map after the channel attention weights are assigned, i and j are the channel coordinates of the above feature map, and x ji It is used to measure the influence of channel i on channel j, and β is the scale coefficient.

5. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 1, characterized in that: In step S2, a digital image is extracted from an infinite resolution view using a Boolean model to model the speckle image and generate a reference image.

6. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 5, characterized in that: In the step S2, the two-dimensional strain field is calculated using the spatial derivative of the deformation field based on the infinitesimal strain assumption; The local deformation field is superimposed by several two-dimensional Gaussian functions, which can be expressed as: In the formula, is the Gaussian function amplitude, is the standard deviation, A two-dimensional Gaussian function representing the local deformation, is the peak center coordinate.

7. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 6, characterized in that: In S2, the reference image is distorted and a certain degree of white noise is added to obtain a virtual data set including the reference image, the deformed image, the deformation field and the strain field. The calculation formula is as follows: I def =F(I ref ;Φ)+σ noise In the formula, σ noise represents the added white noise; I def is the deformed image, I ref is the reference image, Φ is the deformation field; The Green-Lagrange strain calculation on a two-dimensional plane is as follows: In the formula, ε xx , ε yy , ε xy represents the strain field in three directions, and u and v represent the deformation field in the x and y directions.

8. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 1, characterized in that: In S3, the autocorrelation full width at half maximum, systematic error, random error and mean gray gradient were used to evaluate the availability of the virtual dataset.

9. The method for measuring plastic deformation of a machined surface based on deep learning according to claim 1, characterized in that: In S4, the loss function uses the root mean square error RMSE. In order to compensate for the unclear gradient descent caused by a small RMSE, it is multiplied by an amplification factor of 5. The RMSE calculation formula is as follows: Where u pre (i,j) is the deformation prediction value, u gt (i,j) is the true value of the deformation.

10. A deep learning-based plastic deformation measurement system for machined surfaces, characterized in that: The invention comprises a high-speed camera (1), a coaxial LED (2), a workpiece fixture (3), a displacement adjustment platform (4) and a computer (5); the high-speed camera (1) is installed on the displacement adjustment platform (4); the workpiece fixture (3) is used to clamp a workpiece to be measured (7); a reflector (6) arranged between the high-speed camera (1) and the workpiece fixture (3) is used to change the light path, and a grinding process is performed on the surface of the workpiece to be measured (7) to form random scratches; the coaxial LED (2) is used to illuminate the cutting area, so as to capture a continuous dynamic image of the cutting area; the computer (5) stores a computer program for implementing the steps of the deep learning-based plastic deformation measurement method for a machined surface according to any one of claims 1 to 8, and performing deformation field measurement on the continuous dynamic image of the cutting area captured by the high-speed camera (1).

Images