A method for estimating the temperature state of a workpiece surface in a machining process

Abstract

The application discloses a kind of methods for estimating workpiece surface temperature state in processing, constructs the infrared visible light image dataset of processing workpiece and adds label, builds and trains the regression model based on neural network;Build the classification model based on neural network, and train using next best viewpoint dataset;From the viewpoint set of the workpiece to be measured, any one is selected as initial viewpoint;Cyclically collect the infrared visible light image of current viewpoint, input regression model, and the output feature is accumulated;The accumulated coverage point cloud of the voxel model of the workpiece to be measured is updated using the current viewpoint, and then input into the classification model to obtain the next best viewpoint;After reaching the cycle termination condition, the accumulated features are averaged and input into the linear layer of the regression model, and the confidence score is output, to judge the temperature state of the workpiece to be measured.The application combines the temperature data collected from multiple angles to realize comprehensive evaluation of the workpiece temperature condition, and to comprehensively evaluate and dynamically feedback the workpiece temperature state during processing.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to a method for comprehensively estimating the surface temperature of a workpiece during processing. Background Technology

[0002] In industrial manufacturing, estimating the surface temperature of workpieces is crucial for ensuring processing quality and production efficiency. This step ensures that the workpiece is at an appropriate temperature at every stage of the processing, thereby improving product performance, reducing manufacturing defects, and effectively reducing the use of materials and energy.

[0003] Establishing a comprehensive vision-based solution is a key technology for monitoring the thermal information of workpiece surfaces. While non-contact sensors can measure localized temperatures, a multi-viewpoint system is essential to obtain the global temperature distribution. The main challenges of this approach include the complexity of data processing and the need for intelligent decision-making. On the one hand, the large amounts of data generated by sensors require rapid processing for real-time feedback, placing high demands on data processing capabilities. On the other hand, in dynamic environments, although it is possible to identify the viewpoints with the greatest information gain, how to utilize the information from these viewpoints to provide a quantitative assessment of temperature status remains an unresolved issue. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method for comprehensively estimating the surface temperature of a workpiece during processing. This method enables rapid and accurate estimation of the workpiece surface temperature during processing, thereby optimizing the processing procedure.

[0005] The specific technical solution is as follows:

[0006] A method for comprehensively estimating the surface temperature state of a workpiece during machining includes the following steps:

[0007] S1: Construct an infrared and visible light image dataset of the processed workpiece and add labels; build a regression model based on a neural network, and train the regression model using the infrared and visible light image dataset to predict the temperature state of the workpiece; the output of the regression model is a linear layer that outputs a single continuous value;

[0008] S2: Convert the 3D model of the workpiece into voxels, obtain a set of viewpoints based on the voxel model of the workpiece, and for each viewpoint, obtain its next best viewpoint to obtain the next best viewpoint dataset of the workpiece; build a classification model based on neural networks, and use the next best viewpoint dataset to train the classification model to predict the next best viewpoint of the workpiece.

[0009] S3: Set the set of viewpoints for the workpiece to be measured, and select any one of the viewpoints in the set as the initial viewpoint;

[0010] S4: Acquire infrared and visible light images of the workpiece at the current viewpoint, input them into the trained regression model, and output features;

[0011] S5: Update the cumulative coverage point cloud of the voxel model of the workpiece under test according to the current viewpoint, and input the voxel model of the workpiece under test after updating the cumulative coverage point cloud into the trained classification model to obtain the next best viewpoint.

[0012] S6: S4 and S5 are executed repeatedly until the loop termination condition is met. In each loop, the features output in S4 are added to the features output in the previous loop to achieve feature accumulation. After the loop termination condition is met, the input terminal of the linear layer is turned on, the accumulated features are averaged and input into the linear layer, and the confidence score is output to determine the temperature state of the workpiece under test.

[0013] Furthermore, in S1, constructing the infrared and visible light image dataset of the processed workpiece and adding labels is achieved through the following sub-steps:

[0014] S1.1: Acquire infrared and visible light images of the workpiece being processed, and perform fusion processing to obtain infrared and visible light images. Multiple infrared and visible light images are combined to form an infrared and visible light image dataset.

[0015] The fusion process is as follows: the infrared image is adjusted to the same size as the visible light image, and the two images are matched pixel by pixel. The pixels are superimposed according to their respective weights. The sum of the weight values ​​of the infrared image pixels and the visible light image pixels is 1.

[0016] S1.2: Manually mark the workpiece area in the visible light image as a target area label in the infrared visible light image; calculate the average gray value of the target area on the infrared image based on the target area label, and use it as a temperature label in the infrared visible light image.

[0017] Furthermore, the infrared image is obtained using a thermal imaging camera, and the visible light image is obtained using a color camera.

[0018] Furthermore, in S1, the regression model is based on a pre-trained ResNet-18, and the output is customized to be a linear layer that outputs a single continuous value.

[0019] The regression model was trained using a stochastic gradient descent optimizer with a momentum of 0.9, a learning rate of 0.001, and a mean squared error loss function.

[0020] Furthermore, in S2, constructing the next optimal viewpoint dataset for the processed workpiece is specifically achieved through the following operations:

[0021] Obtain a 3D model of the workpiece and voxelize it in a 32×32×32 voxel space; create a view sphere around the workpiece voxel model, and uniformly set viewpoints on this view sphere to form a viewpoint set V; for each viewpoint in the viewpoint set, acquire a depth image and convert it into a point cloud z(v); accumulate the coverage point cloud P. acu Updated from the currently explored viewpoint v; traverse the viewpoint set V to find the viewpoint that maximizes the current P. acu viewpoint v * As the next best viewpoint tag, v * With P acu The overlap exceeds a set threshold; the next optimal viewpoint v * The calculation expression is as follows:

[0022]

[0023] In the formula, Cvg(z(v)∪P acu W obj ) is z(v)∪P acu Relative to the workpiece voxel model W obj Coverage rate.

[0024] Furthermore, the voxelization process is as follows: a cubic envelope box of the three-dimensional model of the workpiece is set, and allowances are set in the length, width and height directions; the size of the envelope box is adjusted to a 32×32×32 voxel mesh to obtain the workpiece voxel model.

[0025] Furthermore, the view sphere is an upper hemisphere with the center of the plane where the bottom of the workpiece voxel model is located as the center. Six azimuth viewpoints are evenly arranged on the three polar angle layers of 0°, 30° and 60° respectively, resulting in a viewpoint set with 18 viewpoints.

[0026] Furthermore, in S2, the classification model sequentially includes three consecutive three-dimensional convolutional layers, one flattening layer, three fully connected layers, and one softmax layer; each three-dimensional convolutional layer includes a convolution operation, an activation function, and a max pooling operation; the number of output units in the softmax layer is the same as the number of viewpoints in the viewpoint set.

[0027] Furthermore, in S6, the loop termination condition is: the coverage rate of the workpiece's voxel model reaches 80%, or a total of five viewpoints have been explored; the coverage rate is the cumulative coverage point cloud P. acu The ratio of voxels to total point cloud voxels on the surface.

[0028] Furthermore, in S6, the method for judging the temperature state of the workpiece to be tested is as follows: a confidence threshold range is set; if the confidence score is within the threshold range, the workpiece temperature state is normal; if the confidence score is higher than the threshold range, the workpiece temperature state is overheated; if the confidence score is lower than the threshold range, the workpiece temperature state is undercooled.

[0029] The beneficial effects of this invention are:

[0030] The method of this invention combines temperature data collected from multiple angles to achieve a comprehensive assessment of the workpiece temperature condition, enabling a full evaluation and dynamic feedback of the workpiece temperature condition during the processing. Attached Figure Description

[0031] Figure 1 This is a flowchart of a method for comprehensively estimating the surface temperature of a workpiece during processing, provided by an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the framework of a method for comprehensively estimating the surface temperature of a workpiece during processing, as described in an embodiment of the present invention.

[0033] Figure 3 This is a schematic diagram of the workpiece model voxelization process in an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram of a viewpoint set arranged with the workpiece model as the center in an embodiment of the present invention.

[0035] Figure 5 This is a schematic diagram of the neural network classification model used in the embodiments of the present invention. Detailed Implementation

[0036] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The objectives and effects of the present invention will become clearer as a result. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0037] like Figure 1 and Figure 2 As shown, a method for comprehensively estimating the surface temperature state of a workpiece during processing specifically includes the following steps:

[0038] S1: Construct an infrared and visible light image dataset of the processed workpiece, label the images in the infrared and visible light image dataset, and then build and train a neural network regression model to predict the temperature state of the workpiece.

[0039] In this embodiment, S1 is specifically implemented through the following sub-steps:

[0040] S1.1: Acquire infrared and visible light images of the workpiece being processed, and perform fusion processing to obtain infrared and visible light images. Multiple infrared and visible light images are combined to form an infrared and visible light image dataset.

[0041] In this embodiment, the fusion process is as follows: the infrared image is adjusted to the same size as the visible light image, and pixel-by-pixel correspondence is performed between the infrared and visible light images. Then, the pixels are superimposed, and the pixel value X of the superimposed infrared and visible light image is obtained. c The calculation is as follows:

[0042] X c =pX a +qX b

[0043] In the formula, p represents the weight value of the infrared image pixel, q represents the weight value of the visible light image pixel, and p + q = 1; in this embodiment, p is set to 0.7 and q is set to 0.3. a X represents the pixel value of a specific pixel in an infrared image. b X represents the pixel value of the corresponding pixel in a visible light image. c X represents the pixel value of the corresponding pixel in an infrared-visible image. a X b X c Each pixel corresponds to a specific pixel.

[0044] S1.2: Data annotation is performed on the images in the infrared-visible image dataset. The workpiece area in the visible light image is manually marked as the target area label in the infrared-visible image. Then, the average gray value of the target area is calculated on the corresponding infrared image based on the target area label, and this is used as the temperature label of the infrared-visible image.

[0045] S1.3: Build a regression model for predicting the temperature state of a workpiece. This regression model uses a pre-trained ResNet-18 as its base for feature extraction. The output of the regression model is customized to be a linear layer that outputs a single continuous value so as to directly predict the temperature value of the workpiece. The input of the linear layer will only be turned on when the loop termination condition is met. After the regression model is tuned, it is trained using an infrared and visible light image dataset to optimize the temperature prediction performance.

[0046] In this embodiment, the regression model is trained using a stochastic gradient descent optimizer with a momentum of 0.9 and a learning rate of 0.001. The mean squared error loss function is used. The infrared and visible light image dataset is divided into a training set and a test set, with the ratio of the training set to the test set set set to 3:1.

[0047] S2: Construct a dataset of the next best viewpoint for the current workpiece, and then build and train a neural network classification model to predict the next best viewpoint for the workpiece.

[0048] In this embodiment, S2 is specifically implemented through the following sub-steps:

[0049] S2.1 Constructing a dataset to predict the next optimal viewpoint for the currently being machined workpiece: First, obtain the 3D model of the currently being machined workpiece and voxelize this 3D model in a 32×32×32 voxel space. For example... Figure 3 As shown, in this embodiment, the voxelization process of the workpiece 3D model is as follows: calculate the cubic envelope of the workpiece 3D model, and set a margin m in each of the X-axis, Y-axis, and Z-axis directions of the envelope; adjust the size of the envelope to a 32×32×32 voxel mesh to complete the voxelization of the workpiece 3D model, and obtain the workpiece voxel model W. obj In this embodiment, the margin m is 8 voxel units.

[0050] like Figure 4 As shown, a view sphere is created around the workpiece voxel model, and a set of viewpoints is set on this view sphere. In this embodiment, the process of setting the viewpoint set is as follows: 18 viewpoint positions are evenly placed on the upper hemisphere with the center of the plane where the bottom of the workpiece voxel model is located as the center. That is, six azimuth viewpoints are evenly placed on three polar angle layers (0 degrees, 30 degrees, 60 degrees) to obtain the viewpoint set.

[0051] For each viewpoint in the viewpoint set, acquire a depth image and convert it into a point cloud z(v); iterate through the viewpoints. And obtain the corresponding point cloud z(v), and find the point cloud P that maximizes the current cumulative coverage. acu viewpoint v * As the next best viewpoint label.

[0052] In this embodiment, the viewpoint v is found. * The process is as follows: Construct the cumulative coverage point cloud P acu Update P using the point cloud of the currently explored viewpoint v. acu Traversing viewpoints The next optimal viewpoint v is calculated using the following formula. * :

[0053]

[0054] In the formula, Cvg(z(v)∪P acu W obj ) is z(v)∪P acu Relative to the workpiece voxel model W objCoverage rate.

[0055] At the same time, the next optimal viewpoint v must be satisfied. * With P acu The overlap is higher than a threshold T. In this embodiment, the threshold T is fifty percent.

[0056] S2.2: Build and train a neural network classification model to predict the next optimal viewpoint of the workpiece. This classification model is based on a 3D convolutional neural network, containing three consecutive 3D convolutional layers, followed by a flattening layer and three fully connected layers, and finally a softmax layer with 18 output units for classification; the number of output units is the same as the number of viewpoints set on the workpiece voxel model.

[0057] In this embodiment, the framework of the classification model is as follows: Figure 5 As shown, the framework consists of the following components:

[0058] Conv(64,3×3×3)+Relu+Pool(2×2×2)+Conv(128,3×3×3)+Relu+Pool(2×2×2)+Conv(256,3×3×3)+Relu+Pool(2×2×2)+Flatten+FC(512)+FC(128)+FC(18)+Softmax, where Conv(F1,K1×K1×K1) represents a 3D convolution operation with F1 filters and a kernel size of K1×K1×K1; Pool(K2×K2×K2) represents a 3D max pooling operation with a pooling window size of K2×K2×K2; Flatten represents a flattened layer; FC(F2) represents a fully connected layer with F2 neurons; Relu and Softmax represent two activation functions.

[0059] The classification model was trained using the Adam optimizer with β1 = 0.9, β2 = 0.999, and a learning rate of 0.002. The cross-entropy loss function was used, and the next best viewpoint dataset of the artifact was divided into a training set and a test set with a ratio of 3:1.

[0060] S3: Perform viewpoint initialization on the workpiece to be measured. In this embodiment, step S3 specifically includes the following sub-steps:

[0061] S3.1: Voxelize the 3D model of the workpiece under test in a 32×32×32 voxel space; the calculation process of the voxel model of the workpiece under test is the same as in S2.1. Create a view sphere around the workpiece voxel model, and uniformly set viewpoints on the view sphere to form a viewpoint set V; the method for forming the viewpoint set is the same as in S2.1.

[0062] S3.2: Select any viewpoint from the viewpoint set V as the initial viewpoint.

[0063] S4: Acquire visible light and infrared images of the workpiece at the current viewpoint, perform fusion processing, and input them into the trained regression model for feature accumulation.

[0064] In this embodiment, the infrared image of the workpiece can be obtained by a thermal imaging camera, and the visible light image of the workpiece can be obtained by a color camera; the fusion process is the same as in S1.1.

[0065] S5: Update the cumulative coverage point cloud of the voxel model of the workpiece under test based on the current viewpoint. Input the voxel model of the workpiece under test with the updated cumulative coverage point cloud into the trained classification model to obtain the next optimal viewpoint. Based on the currently explored viewpoints, update the cumulative coverage point cloud P of the workpiece voxel model. acu .

[0066] S6: Repeat S4 and S5. In each iteration, the features output by the regression model in S4 are added to the previously accumulated features until the loop termination condition is met. The accumulated features are then averaged and input into the linear layer of the regression model, outputting a confidence score. The current temperature state of the workpiece is determined based on this confidence score.

[0067] In this embodiment, the loop termination condition in S6 is: the coverage rate of the workpiece voxel model (i.e., the cumulative coverage point cloud P). acu The ratio of voxels to total point cloud voxels on the surface reached 80%, or a total of five viewpoints were explored.

[0068] In this embodiment, the process of determining the current temperature state of the workpiece in S6 is as follows:

[0069] The workpiece temperature state is divided into three states: overheating, normal, and undercooling. A confidence threshold range [100, 150] is defined. If the confidence score output by the regression model is within the threshold range, the temperature state is normal; if it is higher than the threshold range, the temperature state is overheating; if it is lower than the threshold range, the temperature state is undercooling.

[0070] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for comprehensively estimating the surface temperature state of a workpiece during processing, characterized in that, Includes the following steps: S1: Construct an infrared and visible light image dataset of the processed workpiece and add labels. The infrared and visible light images are obtained by fusing infrared and visible light images. Build a regression model based on a neural network and train the regression model using the infrared and visible light image dataset to predict the temperature state of the workpiece. The output of the regression model is a linear layer that outputs a single continuous value. S2: Convert the 3D model of the workpiece into voxels, obtain a set of viewpoints based on the voxel model of the workpiece, and for each viewpoint, obtain its next best viewpoint to obtain the next best viewpoint dataset of the workpiece; build a classification model based on neural networks, and use the next best viewpoint dataset to train the classification model to predict the next best viewpoint of the workpiece. In step S2, the construction of the next optimal viewpoint dataset for the workpiece is specifically achieved through the following operations: Obtain the 3D model of the workpiece and voxelize it in a 32×32×32 voxel space; create a view sphere around the workpiece voxel model, and uniformly set viewpoints on the view sphere to form a viewpoint set V; For each viewpoint in the viewpoint set, acquire a depth image and convert it into a point cloud z(v); accumulate the coverage point cloud P. acu Updated from the currently explored viewpoint v; traverse the viewpoint set V to find the viewpoint that maximizes the current P. acu viewpoint v * As the next best viewpoint tag, v * With P acu The overlap exceeds a set threshold; the next optimal viewpoint v * The calculation expression is as follows: ； In the formula, for Relative to the workpiece voxel model W obj Coverage rate; S3: Set the viewpoint set of the workpiece to be measured, and select any one of the viewpoints in the viewpoint set as the initial viewpoint; S4: Acquire infrared and visible light images of the workpiece at the current viewpoint, input them into the trained regression model, and output features; S5: Update the cumulative coverage point cloud of the voxel model of the workpiece under test according to the current viewpoint, and input the voxel model of the workpiece under test after updating the cumulative coverage point cloud into the trained classification model to obtain the next best viewpoint. S6: S4 and S5 are executed repeatedly until the loop termination condition is met. In each loop, the features output in S4 are added to the features output in the previous loop to achieve feature accumulation. After the loop termination condition is met, the input terminal of the linear layer is turned on, the accumulated features are averaged and input into the linear layer, and the confidence score is output to determine the temperature state of the workpiece under test.

2. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, In step S1, constructing the infrared and visible light image dataset of the processed workpiece and adding labels is achieved through the following sub-steps: S1.1: Acquire infrared and visible light images of the workpiece being processed, and perform fusion processing to obtain infrared and visible light images. Multiple infrared and visible light images are combined to form an infrared and visible light image dataset. The fusion process is as follows: the infrared image is adjusted to the same size as the visible light image, and the two images are matched pixel by pixel. The pixels are superimposed according to their respective weights. The sum of the weight values ​​of the infrared image pixels and the visible light image pixels is 1. S1.2: Manually mark the workpiece area in the visible light image as a target area label in the infrared visible light image; Based on the target region label, the average gray value of the target region is calculated on the infrared image and used as the temperature label of the infrared visible light image.

3. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 2, characterized in that, The infrared image is obtained using a thermal imaging camera, and the visible light image is obtained using a color camera.

4. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, In S1, the regression model is based on a pre-trained ResNet-18, and the output is customized to be a linear layer that outputs a single continuous value. The regression model was trained using a stochastic gradient descent optimizer with a momentum of 0.9, a learning rate of 0.001, and a mean squared error loss function.

5. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, The voxelization process is as follows: a cubic envelope box of the three-dimensional model of the workpiece is set, and allowances are set in the length, width and height directions; the size of the envelope box is adjusted to a 32×32×32 voxel mesh to obtain the workpiece voxel model.

6. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, The view sphere is the upper hemisphere with the center of the plane where the bottom of the workpiece voxel model is located as the center. Six azimuth viewpoints are evenly arranged on the three polar angle layers of 0°, 30° and 60° respectively, resulting in a viewpoint set with 18 viewpoints.

7. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, In S2, the classification model sequentially includes three consecutive 3D convolutional layers, one flattening layer, three fully connected layers, and one softmax layer; each 3D convolutional layer includes a convolution operation, an activation function, and a max pooling operation; the number of output units in the softmax layer is the same as the number of viewpoints in the viewpoint set.

8. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, In S6, the loop termination condition is: the coverage rate of the workpiece's voxel model reaches 80%, or a total of five viewpoints have been explored; the coverage rate is the cumulative coverage point cloud P. acu The ratio of voxels to total point cloud voxels on the surface.

9. The method for comprehensively estimating the surface temperature state of a workpiece during processing according to claim 1, characterized in that, In step S6, the method for judging the temperature state of the workpiece to be tested is as follows: a confidence threshold range is set. If the confidence score is within the threshold range, the workpiece temperature state is normal; if the confidence score is higher than the threshold range, the workpiece temperature state is overheated; if the confidence score is lower than the threshold range, the workpiece temperature state is undercooled.

Images