A prediction method, device, and system based on digital twin technology

Abstract

The present invention discloses a prediction method, device and system based on digital twin technology, belonging to the field of quality prediction of complex thin-walled components. The method includes: acquiring the overall processing data of the manufacturing process of multi-process and multi-technology complex thin-walled components and storing them for preprocessing; converting the collected one-dimensional multi-scale sequence data set into spatial data in a two-dimensional same-scale space and then performing spatial gray mapping to convert it into a gray feature map; constructing a quality prediction model for multi-process and multi-technology parameters based on a multi-scale convolutional neural network with Gaussian pyramid pooling; optimizing the multi-scale convolutional neural network with an improved tuna swarm optimization algorithm; performing feature extraction on the optimized multi-scale convolutional neural network based on the attention mechanism and inputting it into the quality prediction model to output the prediction result of the product quality. The present invention realizes the product quality prediction of the whole process and multiple process parameters by comprehensively considering multiple processes and multiple process parameters, ensuring the accuracy of quality prediction.

Description

Technical Field

[0001] The present invention relates to the field of quality prediction of complex thin-walled components, and particularly to a prediction method, device and system based on digital twin technology. Background Art

[0002] During the manufacturing process of complex thin-walled components, there are often many unpredictable quality problems, such as the shape exceeding the specified accuracy, surface defects, etc. These problems will affect the appearance quality and performance of the product. The quality prediction of complex thin-walled components is a key step to ensure product quality. Producers can adjust process parameters in a timely manner during the production process according to the quality prediction results, optimize the manufacturing and assembly processes, improve the quality and consistency of the product, and thus improve manufacturing efficiency and economic benefits.

[0003] However, the following problems still exist in practical applications: Due to the material removal behavior, internal stress evolution, microstructure evolution, and the highly coupled effects within the mechanical / special / composite energy fields, it is extremely difficult to construct a cross-scale accurate model for the forming and surface forming of complex thin-walled components. Currently, there are many product quality prediction methods based on machine learning algorithms, but these methods often only consider a single process or fewer process parameters and cannot comprehensively reflect the influencing factors of product quality, resulting in inaccurate and unreliable prediction results. There is a need for a new algorithm that can comprehensively consider multiple processes and multiple process parameters to achieve product quality prediction for the entire process and multiple process parameters. Therefore, based on the current research status, the present invention proposes a prediction method, device and system based on digital twin technology to solve the above problems. Summary of the Invention

[0004] To solve the above problems, a prediction method, device and system based on digital twin technology are proposed to solve the problems of difficult comprehensive prediction of the product quality of complex thin-walled components and inaccurate and unreliable prediction results in the prior art. The present invention uses a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for predicting the product quality of complex thin-walled components, realizing the prediction of product quality for the entire process and multiple process parameters, and ensuring the accuracy of model prediction.

[0005] To achieve the above object, the present invention provides the following technical solutions: A prediction method based on digital twin technology proposed by the present invention includes the following steps:

[0006] S1. Obtain the overall processing data of the manufacturing process of multi-process and multi-technology complex thin-walled components, store and preprocess it to obtain a one-dimensional multi-scale sequence data set;

[0007] S2. Convert the collected one-dimensional multi-scale sequence data set into spatial data in a two-dimensional same-scale space, and then perform spatial gray mapping to convert it into a gray feature map;

[0008] S3. Input the obtained grayscale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multiple processes and multiple process parameters;

[0009] S4. Optimize the multi-scale convolutional neural network with an improved tuna school optimization algorithm;

[0010] S5. Perform feature extraction on the optimized multi-scale convolutional neural network based on the attention mechanism, and input the extracted features into the quality prediction model to output the prediction result of the product quality.

[0011] Further, step S1 includes the following steps:

[0012] S101. Use the DNC system, IO box, and MES device to collect real-time data for each process and related process parameters, including data for measuring temperature, pressure, speed, and material properties;

[0013] S102. Record the collected data and organize and label it according to the correspondence between the process and the process parameters; S103. Store the collected full-process processing data in a database and obtain one-dimensional multi-scale sequence data using the data storage method of the database.

[0014] Further, step S2 includes the following steps:

[0015] S201. Arrange the single-process multi-process data into a one-dimensional process sequence vector, and the process sequences of different processes form a one-dimensional process sequence vector of a single process from top to bottom;

[0016] S202. Then arrange the one-dimensional process sequence vectors of multiple processes together left and right to form a two-dimensional matrix data of multiple processes and multiple process parameters, where the horizontal axis represents the process and the vertical axis represents the process parameters of each process;

[0017] S203. Normalize the two-dimensional matrix data to between -1 and 1 through a normalization method;

[0018] S204. Perform spatial grayscale mapping on the normalized two-dimensional matrix data to finally obtain the conversion of the two-dimensional matrix into a grayscale feature map.

[0019] Further, inputting the obtained grayscale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling in step S3 to construct a quality prediction model for multiple processes and multiple process parameters specifically includes the following steps:

[0020] S301. In the single-process multi-process feature diagram, traverse from top to bottom on the data feature axis through one-dimensional multi-scale convolution kernels 3×1, 5×1, 7×1 to complete the convolution operation with a step size of 1. Concatenate and fuse the feature vectors output by different convolution kernels to obtain the relationship between single-process multi-process features;

[0021] S302. On the multi-process axis, use the same method as S301. Traverse from left to right on the data feature axis through one-dimensional multi-scale convolution kernels 3×1, 5×1, 7×1 to complete the convolution operation with a step size of 1. Concatenate and fuse the feature vectors output by different convolution kernels to obtain the relationship between multi-process multi-process features;

[0022] S303. Fuse single-process multi-process features and multi-process multi-process features at different levels by adding cross-layer connections of the Gaussian pyramid module to obtain a multi-dimensional concatenated fusion feature;

[0023] S304. Input the multi-dimensional concatenated fusion feature into a 4-layer DNN fully connected layer network with a sigmoid activation function for non-linear fitting to obtain the product quality prediction value;

[0024] S305. Select the consistency index IA and root mean square error RMSE as performance evaluation indicators to evaluate the difference between the product quality prediction value output by the network and the true label value.

[0025] Furthermore, optimizing the multi-scale convolutional neural network with the improved tuna school optimization algorithm in step S4 specifically includes the following steps:

[0026] S401. Initialize the population with the Circle chaotic map;

[0027] S402. Utilize the search feature of Levy flight for random walk in space to increase the amplitude during spiral foraging of the algorithm;

[0028] S403. Quickly obtain the global optimal position and fitness value as the input of the multi-scale convolutional neural network. S5. Perform feature extraction on the optimized multi-scale convolutional neural network based on the attention mechanism, and input the extracted features into the quality prediction model to output the prediction result of the product quality.

[0029] Furthermore, step S5 also includes: adopting the weighted average smoothing method to perform smoothing and denoising processing on the prediction result of the product quality, removing the influence of local fluctuations on the prediction performance, and the denoising formula is as follows:

[0030] y' s =(y s-m +y s-m+1 +y s-m+2 +…+y s) / (m + 1);

[0031] Where: y' s represents the value after smoothing at time s; y s represents the unsmoothed fluctuating data corresponding to time s.

[0032] This technical solution also provides a device for implementing the above-mentioned prediction method based on digital twin technology, including:

[0033] A dataset construction module, which is used to obtain the overall processing data of the manufacturing process of multi-process and multi-technology complex thin-walled components, store and preprocess it to obtain a one-dimensional multi-scale sequence dataset;

[0034] A grayscale feature map conversion module, which is used to convert the collected one-dimensional multi-scale sequence dataset into spatial data in a two-dimensional same-scale space, and then perform spatial grayscale mapping to convert it into a grayscale feature map;

[0035] A quality prediction model construction module, which is used to input the obtained grayscale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multi-process and multi-process parameters;

[0036] An optimization network module, which is used to optimize the multi-scale convolutional neural network with an improved tuna swarm optimization algorithm;

[0037] A prediction result acquisition module, which is used to perform feature extraction on the optimized multi-scale convolutional neural network based on the attention mechanism, and input the extracted features into the quality prediction model to output the prediction result of the product quality.

[0038] This technical solution also provides a system for implementing the above-mentioned prediction method based on digital twin technology, including:

[0039] A processor;

[0040] A memory;

[0041] And one or more programs, where the one or more programs are stored in the memory and are configured to be executed by the processor, and the programs are used for a computer to execute the prediction method based on digital twin technology.

[0042] With the above technical solution, the present invention provides a prediction method, device and system based on digital twin technology. It has at least the following beneficial effects:

[0043] The present invention constructs a prediction method, device and system based on digital twin technology. This method uses a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for predicting the quality of complex thin-walled component products. By comprehensively considering multiple processes and multiple process parameters, it realizes the prediction of product quality for the entire process and multiple process parameters, ensuring the accuracy of model prediction. The present invention is applicable to the product quality prediction method in the manufacturing process of complex thin-walled components, can achieve accurate prediction of product quality, and can timely adjust key process parameters according to the quality prediction results in subsequent production, optimize the manufacturing and assembly processes, improve the quality and consistency of products, and thus improve manufacturing efficiency and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0045] Figure 1 is a flowchart of the quality prediction method of the present invention;

[0046] Figure 2 is a sample for storing processing data of the present invention;

[0047] Figure 3 is a structural diagram of the grayscale mapping module of the present invention;

[0048] Figure 4 is a network structure diagram of the quality prediction network model of the present invention;

[0049] Figure 5 is a principle block diagram of the quality prediction device of the present invention.

[0050] In the figure: 100 dataset construction module; 200 grayscale feature map conversion module; 300 quality prediction model construction module; 400 optimization network module; 500 prediction result acquisition module. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments. Thereby, a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects can be obtained and implemented accordingly.

[0052] Those of ordinary skill in the art can understand that all or part of the steps in implementing the methods of the above embodiments can be completed by instructing relevant hardware through a program. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0053] Please refer to Figures 1 - 5 , which shows a specific implementation manner of this embodiment. In this embodiment, a quality prediction model is constructed by using a multi-scale convolutional neural network based on Gaussian pyramid pooling to predict the quality of complex thin-walled component products. Considering multiple processes and multiple process parameters comprehensively, the prediction of product quality for the entire process and multiple process parameters is realized, ensuring the accuracy of model prediction, and solving the problems that it is difficult to comprehensively predict the quality of complex thin-walled component products in the prior art and the inaccuracy and low reliability of the prediction results.

[0054] Please refer to Figure 1 , a prediction method based on digital twin technology, includes the following steps:

[0055] S1. Obtain the overall processing data of the manufacturing process of multi-process and multi-technology complex thin-walled components, store and preprocess it to obtain a one-dimensional multi-scale sequence data set;

[0056] Specifically, S1 specifically includes the following steps:

[0057] S101. Use a DNC system, an IO box, and MES equipment to collect real-time data for each process and related process parameters, including data for measuring temperature, pressure, speed, and material properties;

[0058] S102. Record the collected data and organize and label it according to the corresponding relationship between the process and the process parameters;

[0059] S103. Store the collected full-process processing data in a database and obtain one-dimensional multi-scale sequence data using the data storage method of the database. As Figure 2 shown is an example of the storage of processing data in this embodiment of the present invention.

[0060] S2. Convert the collected one-dimensional multi-scale sequence data set into spatial data in a two-dimensional same-scale space, and then perform spatial gray mapping to convert it into a gray feature map;

[0061] Specifically, S2 specifically includes the following steps:

[0062] S201. Arrange the single-process multi-technology data into a one-dimensional process sequence vector. The process sequences of different technologies form a one-dimensional process sequence vector of a single process from top to bottom.

[0063] S202. Then arrange the one-dimensional process sequence vectors of multiple processes side by side to form a two-dimensional matrix data of multiple processes and multiple technologies. The horizontal axis represents the process, and the vertical axis represents the process parameters of each process.

[0064] S203. Normalize the two-dimensional matrix data to the range of -1 to 1 through a normalization method. The calculation formula of the normalization method is as follows:

[0065]

[0066] Where X i represents the process parameter, X i ' represents the normalized process parameter, min(X i ) represents the minimum value, and max(X i ) represents the maximum value.

[0067] S204: Perform spatial gray mapping on the normalized two-dimensional matrix data to finally obtain a two-dimensional matrix converted into a gray-scale feature map. As Figure 3 shown in the structure diagram of the gray mapping module in this embodiment of the present invention, the spatial gray mapping formula is as follows:

[0068] X″ i =|X′ i |·γ

[0069] γ is the spatial gray level, X″ i represents the mapped process parameter. In this study, the gray level γ is set to (0, 255).

[0070] S3. Input the obtained gray-scale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multiple processes and multiple process parameters. As Figure 4 shown in the network structure diagram of the quality prediction network model in this embodiment of the present invention;

[0071] Specifically, S3 specifically includes the following steps:

[0072] S301. In the single-process multi-technology gray-scale feature map, traverse from top to bottom on the data feature axis through one-dimensional multi-scale convolutional kernels 3×1, 5×1, and 7×1 to complete the convolution operation. The feature map contains N one-dimensional process sequence vectors of single processes provided by P processes. Then the feature map formula of a single process and a single technology n is as follows:

[0073]

[0074] Among them, A vector representing P single - process and single - technology n at process t, represents S sets of vectors of Xn in the process range [t, t + S - 1]. T represents matrix transpose, and the convolution operation is performed through the convolution kernel W i on the process axis and multiplied, with a step size of 1. The feature vectors output by different convolution kernels are concatenated and fused to obtain the spatial feature relationship of single - process and multi - technology;

[0075] S302. On the multi - process axis, using the same method as S301, through one - dimensional multi - scale convolution kernels 3×1, 5×1, 7×1, traverse from left to right on the data feature axis to complete the convolution operation. The convolution kernel W i on the process axis and multiplied, with a step size of 1. The feature vectors output by different convolution kernels are concatenated and fused to obtain the spatial feature relationship of multi - process and multi - technology;

[0076] S303. By adding cross - layer connections of the Gaussian pyramid module, fuse the single - process and multi - technology features and multi - process and multi - technology features at different levels to obtain multi - dimensional concatenated fusion features, which helps the network better learn information at different scales; Take the multi - process and multi - technology matrix data as the first group and the first layer of the Gaussian pyramid, and take the image after Gaussian convolution of the first group and the first layer as the second layer of the first group of the pyramid. The expression of the Gaussian convolution function is:

[0077]

[0078] where x is the multi - process sequence, y is the multi - technology sequence, σ is the smoothing factor which is the variance of x and y. Multiply σ by the scale factor k to get the new smoothing factor k*σ, and use it to smooth the image of the first group and the second layer to get the third layer. Repeat this operation until the L - layer image is obtained; In the same group, the size of each layer of the image is the same, but the smoothing coefficients are different, and the smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2) σ;

[0079] Downsample the image of the third - last layer of the first group with a scale factor of 2, and the obtained image is used as the first layer of the second group. Then perform Gaussian smoothing with a smoothing factor of σ on the image of the first layer of the second group to get the second layer of the second group. Repeat this operation to obtain the L - layer image of the second group. Within the same group, their sizes are the same, and the corresponding smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2)σ, but in terms of size, the second group is half of the first group of images, and a total of O groups of images are obtained, where O = [logmin(M,N)] - 2, and M and N are the number of rows and columns of the original image. By repeating the operation, a total of O groups can be obtained, with each group having L layers, resulting in a total of O * L images;

[0080] S304: As Figure 4 shown: Input the fused features of multi-dimensional splicing into a 4-layer DNN fully connected layer network with a sigmoid activation function for non-linear fitting to obtain the product quality prediction values, such as deformation amount, position degree, section profile degree, and roughness. Please refer to Figure 4 , and the formula expression of the activation function is as follows:

[0081]

[0082] Among them, x is the input, and σ(z) is the output after sigmoid transformation.

[0083] S305: Select the consistency index IA and root mean square error RMSE as performance evaluation indicators to evaluate the difference between the product quality prediction value output by the network and the true label value. Train with RMSE as the evaluation indicator until the result converges. The smaller the value of RMSE, the smaller the prediction error. The value range of the IA value is between 0 and 1, and the closer the calculation result is to 1, the better the fitting effect;

[0084] Among them, the formula for the consistency index is as follows:

[0085]

[0086] The calculation formula for the root mean square error is as follows:

[0087]

[0088] In the above formulas, n is the number of data, y i is the true value of the quality of the i-th product, and y' i is the predicted value of the quality of the i-th product. is the average value of the true values.

[0089] In this embodiment, by adding a Gaussian pyramid module to the multi-scale convolutional neural network, the fused features can be better extracted, thereby improving the accuracy and precision of the product quality prediction value.

[0090] S4: Optimize the multi-scale convolutional neural network with an improved tuna school optimization algorithm;

[0091] Specifically, the steps of optimizing the multi-scale convolutional neural network with the improved tuna school optimization algorithm in step S4 specifically include the following steps:

[0092] S401. Initialize the population using the Circle chaotic map to enhance the richness of the population. The mapping formula is as follows:

[0093]

[0094] where mod is the remainder function, and xi + 1 represents the value of the (i + 1)-th mapping;

[0095] S402. The tuna school chases prey by forming a tight spiral. Besides chasing prey, the schooling tuna also exchanges information with each other. Each tuna follows the one in front, so information can be shared between adjacent tunas. Utilize the search characteristics of Levy flight in random walks in space to increase the amplitude of the algorithm during spiral foraging;

[0096]

[0097] where, k 1 and k 2 are weight coefficients that control the moving trends of an individual towards a Levy light individual and the previous individual. a is a constant used to determine the degree to which a tuna follows the best individual and the previous individual in the initial stage. t represents the current iteration number, t max x represents the maximum number of iterations, and D is the dimension of the position vector.

[0098] where Levy(x) is obtained by the following formula:

[0099]

[0100]

[0101] where the value of λ is generally (1, 3). In this study, λ = 1.5, and u and v follow a normal distribution with an expectation of 0 and a standard deviation of σ u 、σ v . The Γ function is defined by the second kind of Euler integral: S403. Quickly obtain the global optimal position and fitness value as the input of the multi-scale convolutional neural network.

[0102] In this embodiment, the multi-scale convolutional neural network is optimized by the tuna school optimization algorithm, which can better use the global optimal position and fitness value as the input of the multi-scale convolutional neural network, thereby improving the accuracy and precision of feature extraction by the multi-scale convolutional neural network.

[0103] S5. Based on the attention mechanism, perform feature extraction on the optimized multi-scale convolutional neural network, and input the extracted features into the quality prediction model to output the prediction result of the product quality.

[0104] In this embodiment, the obtained optimal position and fitness value are reassigned to the multi-scale convolutional neural network to evaluate the difference between the predicted product quality value output by the network and the true label value. Training is performed using RMSE as the evaluation index until the result converges. The smaller the value of RMSE, the smaller the prediction error. The value of the IA ranges from 0 to 1. The closer the calculation result is to 1, the better the fitting effect. If the calculation error is satisfied, the model is successfully established. If the calculation error is not satisfied, training and prediction are repeated until the calculation error meets the design requirements. At this time, the multi-scale convolutional neural network is the optimal multi-scale convolutional neural network optimized by the tuna swarm algorithm, and the predicted result data of the product quality obtained at this time is the accurate predicted value.

[0105] Specifically, step S5 further includes: using the weighted average smoothing method to perform smoothing and denoising processing on the predicted result of the product quality to remove the influence of local fluctuations on the prediction performance. The denoising formula is as follows:

[0106] y' s =(y s-m +y s-m+1 +y s-m+2 +…+y s ) / (m + 1);

[0107] In the formula: y' s represents the value after smoothing at time s; y s represents the unsmoothed fluctuation data corresponding to time s.

[0108] Please refer to Figure 5 , this embodiment also provides a device for implementing the above prediction method based on digital twin technology, including:

[0109] The dataset construction module 100 is used to obtain the overall processing data of the manufacturing process of multi-process and multi-technology complex thin-walled components, store and preprocess it to obtain a one-dimensional multi-scale sequence dataset;

[0110] The grayscale feature map conversion module 200 is used to convert the collected one-dimensional multi-scale sequence dataset into spatial data in a two-dimensional same-scale space and then perform spatial grayscale mapping to convert it into a grayscale feature map;

[0111] The quality prediction model construction module 300 is used to input the obtained grayscale feature map into the multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multi-processes and multi-technology parameters;

[0112] Optimization network module 400, which is used to optimize the multi-scale convolutional neural network with an improved tuna school optimization algorithm;

[0113] Prediction result acquisition module 500, which is used to extract features from the optimized multi-scale convolutional neural network based on the attention mechanism, and input the extracted features into the quality prediction model to output the prediction result of the product quality.

[0114] This embodiment also provides a system for implementing the above prediction method based on digital twin technology, including:

[0115] A processor;

[0116] A memory;

[0117] And one or more programs, where one or more programs are stored in the memory and are configured to be executed by the processor. The programs are used for the computer to execute the prediction method based on digital twin technology.

[0118] The present invention is applicable to the product quality prediction method in the manufacturing process of complex thin-walled structures, can achieve accurate prediction of product quality, and can timely adjust key process parameters according to the quality prediction result in subsequent production, optimize the manufacturing and assembly processes, improve the quality and consistency of products, and thus improve manufacturing efficiency and economic benefits.

[0119] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0120] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A prediction method based on digital twin technology, comprising the following steps: S1. Obtain the overall processing data of the manufacturing process of complex thin-walled components with multiple processes and multiple techniques and store them for preprocessing to obtain a one-dimensional multi-scale sequence data set; S2, converting the collected one-dimensional multi-scale sequence data set into spatial data in a two-dimensional space of the same scale, and then performing spatial grayscale mapping to convert it into a grayscale feature map; It specifically includes the following steps: S201, arranging the single-process multi-process data into a one-dimensional process sequence vector, wherein the process sequences of different processes form a one-dimensional process sequence vector of a single process from top to bottom; S202, arranging the one-dimensional process sequence vectors of the multiple processes together on the left and right to form a two-dimensional matrix data of multiple processes and multiple technologies, where the horizontal axis represents the process and the vertical axis represents the process parameters of each process; S203, normalizing the two-dimensional matrix data to between -1 and 1 by a normalization method; the calculation formula of the normalization method is as follows: Where X i represents the process parameters, X i ' represents the normalized process parameters, min(X i ) represents the minimum value, max(X i ) indicates the maximum value; S204, performing spatial grayscale mapping on the normalized two-dimensional matrix data to finally obtain a two-dimensional matrix converted into a grayscale feature map. The spatial grayscale mapping formula is as follows: X″ i =|X′ i |·c; γ is the spatial grayscale, X″ i Represents the process parameters after mapping, and sets the gray level γ to (0,255); S3, inputting the obtained grayscale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multiple processes and multiple process parameters; which specifically includes the following steps: S301. In the single-process multi-process feature graph, the convolution operation is completed by traversing the data feature axis from top to bottom through one-dimensional multi-scale convolution kernels 3×1, 5×1, and 7×1. The feature graph contains N single-process multi-process one-dimensional process sequence vectors provided by P processes. Then the feature graph formula of a single process n is as follows: in, Represents the vector of P single-process single-process n in t process, Then it means X n In the process range [t, t+S-1], S sets of vectors, T represents the matrix transpose, and the convolution operation is performed through the convolution kernel W i On the process axis Multiply, the step number is 1, the feature vectors output by different convolution kernels are concatenated and fused to obtain the relationship between single-process and multi-process features; S302: Use the same method as s301 on the multi-process axis to complete the convolution operation by traversing the data feature axis from left to right using one-dimensional multi-scale convolution kernels 3×1, 5×1, and 7×1. The convolution kernel W i On the process axis Multiply, the step number is 1, the feature vectors output by different convolution kernels are concatenated and fused to obtain the relationship between multi-process and multi-technique features; S303, by adding the cross-layer connection of the Gaussian pyramid module, the single-process multi-process features and multi-process multi-process features at different levels are fused to obtain the fusion features of multi-dimensional splicing; help the network to better learn information of different scales; use the multi-process multi-process matrix data as the first group and the first layer of the Gaussian pyramid, and use the first group and the first layer image as the second layer of the first group of pyramids after Gaussian convolution, wherein the expression of the Gaussian convolution function is: Where x is a multi-process sequence, y is a multi-process sequence, σ is the smoothing factor, which is the variance of x and y. Multiply σ by the proportional coefficient k to get a new smoothing factor of k*σ, which is used to smooth the second layer of the first group of images to get the third layer. Repeat this operation until the Lth layer of images is obtained. In the same group, the size of each layer of images is the same, but the smoothing coefficients are different. The smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2) σ; The third-to-last layer image of the first group is downsampled by a scale factor of 2, and the obtained image is used as the first layer of the second group. Then, the first layer image of the second group is Gaussian smoothed with a smoothing factor of σ to obtain the second layer of the second group. Repeat this operation to obtain the L-layer image of the second group. The sizes of the images in the same group are the same, and the corresponding smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2) σ, but the size of the second group is half of the first group of images, and a total of O groups of images are obtained, O = [logmin(M,N)]-2, where M and N are the number of rows and columns of the original image. Repeatedly, a total of O groups can be obtained, each with L layers, for a total of O*L images; S304, input the multi-dimensional spliced ​​fusion features into a 4-layer DNN fully connected layer network with a sigmoid activation function for nonlinear fitting to obtain the product quality prediction value; the activation function formula expression is as follows: Among them, x is the input, σ(z) is the output after sigmoid transformation; S305, select consistency index IA and root mean square error RMSE as performance evaluation indicators to evaluate the difference between the network output product quality prediction value and the true label value; use RMSE as the evaluation indicator for training until the result converges, the smaller the RMSE value is, the smaller the prediction error is, and the IA value range is between 0 and 1, the closer the calculation result is to 1, the better the fitting effect is; S4. Optimizing multi-scale convolutional neural networks using an improved tuna school optimization algorithm; S5. Based on the attention mechanism, the optimized multi-scale convolutional neural network is used to extract features, and the extracted features are input into the quality prediction model to output the prediction results of product quality.

2. A prediction method based on digital twin technology according to claim 1, characterized in that: Step S1 includes the following steps: S101. Use DNC system, IO box and MES equipment to collect real-time data of each process and related process parameters, including measuring temperature, pressure, speed and material properties; S102, record the collected data and organize and mark them according to the corresponding relationship between the process and the process parameters; S103, storing the collected processing data of the entire process into a database and using the data storage method of the database to obtain one-dimensional multi-scale sequence data.

3. A prediction method based on digital twin technology according to claim 1, characterized in that: In step S4, the multi-scale convolutional neural network is optimized using an improved tuna school optimization algorithm, which specifically includes the following steps: S401, initializing the population using Circle chaos mapping; S402, using the search characteristics of Levy flight in random walk in space to improve the amplitude of the algorithm in spiral foraging; S403, quickly obtain the global optimal position and fitness value as the input of the multi-scale convolutional neural network.

4. A prediction method based on digital twin technology according to claim 1, characterized in that: Step S5 also includes: using a weighted average smoothing method to smooth and denoise the prediction results of product quality to remove the impact of local fluctuations on the prediction performance. The denoising formula is as follows: and' s (and s-m +y s-m+1 +y s-m+2 +…+and s ) / (m+1); Where: y' s represents the value after smoothing at time s; y s Represents the unsmoothed fluctuation data corresponding to time s.

5. A device for implementing the prediction method based on digital twin technology as described in any one of claims 1 to 4, characterized in that: include: A data set construction module 100, which is used to obtain the overall processing data of the multi-process and multi-technique complex thin-walled component manufacturing process and store and pre-process it to obtain a one-dimensional multi-scale sequence data set; A grayscale feature map conversion module 200 is used to convert the collected one-dimensional multi-scale sequence data set into spatial data in a two-dimensional space of the same scale, and then perform spatial grayscale mapping to convert it into a grayscale feature map; It specifically includes the following steps: S201, arranging the single-process multi-process data into a one-dimensional process sequence vector, wherein the process sequences of different processes form a one-dimensional process sequence vector of a single process from top to bottom; S202, arranging the one-dimensional process sequence vectors of the multiple processes together on the left and right to form a two-dimensional matrix data of multiple processes and multiple technologies, where the horizontal axis represents the process and the vertical axis represents the process parameters of each process; S203, normalizing the two-dimensional matrix data to between -1 and 1 by a normalization method; the calculation formula of the normalization method is as follows: Where X i represents the process parameters, X i ' represents the normalized process parameters, min(X i ) represents the minimum value, max(X i ) indicates the maximum value; S204, performing spatial grayscale mapping on the normalized two-dimensional matrix data to finally obtain a two-dimensional matrix converted into a grayscale feature map. The spatial grayscale mapping formula is as follows: X″ i =|X′ i |·c; γ is the spatial grayscale, X″ i Represents the process parameters after mapping, and sets the gray level γ to (0,255); The quality prediction model construction module 300 is used to input the obtained grayscale feature map into a multi-scale convolutional neural network based on Gaussian pyramid pooling to construct a quality prediction model for multiple processes and multiple process parameters; it specifically includes the following steps: S301. In the single-process multi-process feature graph, the convolution operation is completed by traversing the data feature axis from top to bottom through one-dimensional multi-scale convolution kernels 3×1, 5×1, and 7×1. The feature graph contains N single-process multi-process one-dimensional process sequence vectors provided by P processes. Then the feature graph formula of a single process n is as follows: in, Represents the vector of P single-process single-process n in t process, Then it means X n In the process range [t, t+S-1], S sets of vectors, T represents the matrix transpose, and the convolution operation is performed through the convolution kernel W i On the process axis Multiply, the step number is 1, the feature vectors output by different convolution kernels are concatenated and fused to obtain the relationship between single-process and multi-process features; S302: Use the same method as s301 on the multi-process axis to complete the convolution operation by traversing the data feature axis from left to right using one-dimensional multi-scale convolution kernels 3×1, 5×1, and 7×1. The convolution kernel W i On the process axis Multiply, the step number is 1, the feature vectors output by different convolution kernels are concatenated and fused to obtain the relationship between multi-process and multi-technique features; S303, by adding the cross-layer connection of the Gaussian pyramid module, the single-process multi-process features and multi-process multi-process features at different levels are fused to obtain the fusion features of multi-dimensional splicing; help the network to better learn information of different scales; use the multi-process multi-process matrix data as the first group and the first layer of the Gaussian pyramid, and use the first group and the first layer image as the second layer of the first group of pyramids after Gaussian convolution, wherein the expression of the Gaussian convolution function is: Where x is a multi-process sequence, y is a multi-process sequence, σ is the smoothing factor, which is the variance of x and y. Multiply σ by the proportional coefficient k to get a new smoothing factor of k*σ, which is used to smooth the second layer of the first group of images to get the third layer. Repeat this operation until the Lth layer of images is obtained. In the same group, the size of each layer of images is the same, but the smoothing coefficients are different. The smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2) σ; The third-to-last layer image of the first group is downsampled by a scale factor of 2, and the obtained image is used as the first layer of the second group. Then, the first layer image of the second group is Gaussian smoothed with a smoothing factor of σ to obtain the second layer of the second group. Repeat this operation to obtain the L-layer image of the second group. The sizes of the images in the same group are the same, and the corresponding smoothing coefficients are: 0, σ, kσ, k 2 σ,k 3 σ……k (L-2) σ, but the size of the second group is half of the first group of images, and a total of O groups of images are obtained, O = [logmin(M,N)]-2, where M and N are the number of rows and columns of the original image. Repeatedly, a total of O groups can be obtained, each with L layers, for a total of O*L images; S304, input the multi-dimensional spliced ​​fusion features into a 4-layer DNN fully connected layer network with a sigmoid activation function for nonlinear fitting to obtain the product quality prediction value; the activation function formula expression is as follows: Among them, x is the input, σ(z) is the output after sigmoid transformation; S305, select consistency index IA and root mean square error RMSE as performance evaluation indicators to evaluate the difference between the network output product quality prediction value and the true label value; use RMSE as the evaluation indicator for training until the result converges, the smaller the RMSE value is, the smaller the prediction error is, and the IA value range is between 0 and 1, the closer the calculation result is to 1, the better the fitting effect is; An optimization network module 400, wherein the optimization network module 400 is used to optimize a multi-scale convolutional neural network using an improved tuna school optimization algorithm; The prediction result acquisition module 500 is used to extract features from the optimized multi-scale convolutional neural network based on the attention mechanism, input the extracted features into the quality prediction model, and output the prediction result of the product quality.

6. A system for implementing the prediction method based on digital twin technology as described in any one of claims 1 to 4, characterized in that: include: processor; Memory; and one or more programs, wherein the one or more programs are stored in a memory and configured to be executed by the processor, the programs being used for a computer to execute the method according to any one of claims 1 to 4.

Images