Die casting quality prediction method and device based on key sample depth migration

Abstract

The invention discloses a die-casting quality prediction method and device based on key sample depth migration, and the method gives a higher weight to a quality variable value overlapping region through weighted feature alignment in allusion to the condition that the quality variable value domains of a source domain and a target domain are not completely overlapped, and reduces the limitation on model output. Meanwhile, a feature space scale maintaining mechanism is introduced, the feature space scale is maintained, and the prediction precision of the model under the new working condition is improved. The method provided by the invention has higher prediction precision under the condition of working condition change.

Description

Technical Field

[0001] The present invention relates to the field of industrial process modeling, and in particular to a die-casting quality prediction method and device based on deep transfer of key samples. Background Art

[0002] In die-casting production, changes in working conditions are common phenomena, and there are differences in data distributions under different working conditions. Traditional supervised learning methods assume that the data in the training set and the test set belong to the same distribution, so they cannot provide accurate predictions of quality variables under new working conditions. Existing multi-model methods and immediate learning methods require a large amount of labeled data or sample retrieval time under new working conditions and are difficult to quickly adapt to changes in working conditions. Although deep transfer learning methods can handle the distribution differences between the source domain and the target domain, they do not consider the differences in the value ranges of quality variables, which may lead to limited model output. Summary of the Invention

[0003] Aiming at the deficiencies of the prior art, the present invention proposes a die-casting quality prediction method and device based on deep transfer of key samples for die-casting quality prediction in the process of changing working conditions. Through a weighted feature alignment method, in the case where the value ranges of quality variables in the source domain and the target domain do not completely overlap, higher weights are given to the overlapping regions of the quality variable values, reducing the limitation on the model output. At the same time, a feature space scale maintenance mechanism is introduced to maintain the feature space scale and improve the prediction accuracy of the model under new working conditions. The specific technical solutions are as follows:

[0004] A die-casting quality prediction method based on deep transfer of key samples, the method comprising the following steps:

[0005] S1: Collect labeled data of the main operating conditions during the historical operation of the industrial process as source domain data, collect unlabeled data under the new working conditions as target domain data, and preprocess the data;

[0006] S2: Construct an independent domain discriminator composed of an n-layer fully connected neural network, and train the independent domain discriminator using a training set including the source domain data and the target domain data;

[0007] S3: Use the trained independent domain discriminator to calculate the corresponding weights for the samples in the training set;

[0008] S4: Construct a feature extraction network, a feature domain discriminator, and a regressor, and train them using a training set containing source domain data and target domain data; the feature extraction network is composed of a multi-layer gated recurrent neural network layer, with input samples as input and outputs as embedding layer features; the feature domain discriminator and are both composed of a multi-layer fully connected neural network, with inputs as embedding layer features, the output of the feature domain discriminator is the domain discrimination label, and the output of the regressor is the predicted value of the quality variable;

[0009] S5: Collect unlabeled data of the target domain under the new working condition, input the trained feature extraction network, and output the predicted quality variable by the trained regressor.

[0010] Furthermore, the loss function during training of the independent domain discriminator is:

[0011]

[0012] Where N is the total number of samples in the training set containing source domain data and target domain data, x q represents the input sample at the qth time step, d q Represents x q The corresponding domain label, 0 represents the source domain, and 1 represents the target domain.

[0013] Furthermore, the calculation formula for calculating the corresponding weights for the samples in the training set using the trained independent domain discriminator is as follows:

[0014]

[0015] Where x is the input sample, and w(x) is the calculated weight of the input sample x.

[0016] Furthermore, the loss function for training the feature extraction network, feature domain discriminator, and regressor using a training set containing source domain data and target domain data is as follows:

[0017] L1=L reg (F r ,G)-αL wdisc (F d ,G)+γL scale (G)

[0018] L2=L wdisc (F d ,G)

[0019] Among them, L reg (F r ,G) is the training loss function of the regressor, L wdisc (F d, G) is the weighted domain adversarial training loss function for the adversarial training of the feature extraction network and the feature domain discriminator, L scale (G) is the scale-preserving loss function, and α is L wdisc (F d , G)'s weight coefficient, and γ is L scale 's weight coefficient.

[0020] Furthermore, the calculation formula of the scale-preserving loss function is as follows:

[0021] L scale (G) = ||λ s,i＝1,…,k - λ t,i＝1,…,k ||2

[0022] Among them, G is the feature extraction network, and λ s,i＝1,…,k is the i-th eigenvalue of the feature matrix obtained after the source domain samples are processed by the feature extraction network, and λ t,i＝1,…,k is the i-th eigenvalue of the feature matrix obtained after the target domain samples are processed by the feature extraction network, and k is the number of artificially set main eigenvalues.

[0023] Furthermore, the expression of the training loss function of the regressor is as follows:

[0024]

[0025] Among them, N s represents the number of source domain samples, X j = {x j-l+1 , x j-l+2 , …, x j} represents the input sample matrix corresponding to the j-th time step, l is the input sequence length, x j represents the input sample at the j-th time step, MSE represents the mean squared error loss function for regression, and y j represents the value of the quality variable corresponding to x j .

[0026] Furthermore, the weighted domain adversarial training loss function L wdisc (F d , G) has the following expression:

[0027]

[0028] Among them, N is the total number of samples in the training set containing source domain data and target domain data, X q represents the q-th input sample matrix, d q represents the domain label to which X q corresponds, 0 represents belonging to the source domain, 1 represents belonging to the target domain, xq represents the input sample at the q-th time step, and w(x q ) is the weight of the sample x q .

[0029] A die-casting quality prediction device based on key sample deep transfer, which includes a memory and one or more processors, and is used to implement a die-casting quality prediction method based on key sample deep transfer.

[0030] An electronic device, comprising:

[0031] One or more processors;

[0032] A storage device for storing one or more programs, which when executed by the electronic device, enable the electronic device to implement a die-casting quality prediction method based on key sample deep transfer.

[0033] A computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, it implements a die-casting quality prediction method based on key sample deep transfer.

[0034] The beneficial effects of the present invention are as follows:

[0035] The method and device of the present invention can still maintain a high prediction accuracy for new working conditions, solving the problem that the prediction accuracy of existing methods decreases when the working conditions change; through the weighted feature alignment method and the scale maintenance mechanism, the limitation on the model output is reduced, and the adaptability of the model is improved; it is applicable to the multi-condition quality prediction of the die-casting industrial process and has a wide application prospect. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of the die-casting quality prediction method based on key sample deep transfer in an embodiment of the present invention.

[0037] Figure 2 is a flowchart of the adversarial training in an embodiment of the present invention.

[0038] Figure 3 is a schematic structural diagram of the die-casting quality prediction device based on key sample deep transfer in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become more apparent. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] Die-casting defects usually have the characteristics of diversity, complexity, concealment, and randomness. Therefore, defect detection of die-castings faces difficulties such as scarce data, high diversity and similarity of defects, and difficulty in detecting complex structures. Traditional defect detection methods are difficult to comprehensively detect die-casting defects. In addition, due to the uneven production quantity of products in the factory, the data volume of some models of products is extremely scarce, and it is difficult to migrate and deploy die-casting defect detection methods between product models.

[0041] The die-casting quality prediction method based on deep transfer of key samples of the present invention, aiming at the characteristics that there are coupling and non-linear associations between different process variables of die-casting defects and the die-casting data distributions vary greatly under different working conditions, effectively extracts the association relationships between various process variables through a feature extraction network, and simultaneously obtains feature expressions in a high-dimensional space; calculates the weights of each input sample through an independent domain discriminator to determine key samples, and assists in the domain adversarial training of the quality prediction model and the feature domain discriminator; through the feature domain discriminator, judges the domain to which each input sample belongs during training and conducts adversarial training with the quality prediction model to enhance the processing ability of the quality prediction model for data under different working conditions; uses the feature expression in the high-dimensional space as the input through a regressor to give the predicted value of the quality variable.

[0042] As Figure 1 shown, the die-casting quality prediction method based on deep transfer of key samples of this embodiment includes the following steps:

[0043] Step 1: Collect labeled data of the main operating conditions during the historical operation of the industrial process as source domain data, and divide it into a training set and a verification set for optimizing hyperparameters, denoted as Among them, represents the source domain data of the training set, is the label of represents the source domain data of the verification set, represents the label of

[0044] Collect unlabeled data under the new working condition as the target domain data in the training set, denoted as Collect labeled data under the new working condition as the test set, denoted as

[0045] Preprocess the data, and the specific steps are as follows:

[0046] (1.1) Use the isolation forest method to detect outliers in the screened data. For abnormal data, execute step (1.3); for the remaining data, jump to step (1.3) to execute;

[0047] (1.2) For each piece of abnormal data, use the locally linear regression method to calculate the regression prediction value based on the normal data near the abnormal data, and replace the original abnormal data.

[0048] (1.3) Integrate the normal data and the data obtained in step (1.2) to obtain the training data.

[0049] Step 2: Construct an independent domain discriminator D composed of n layers of fully connected neural networks, and use the training set containing source domain data and target domain data to train the independent domain discriminator D; the specific calculation process of the independent domain discriminator D is as follows:

[0050] D(x) = L n (L n-1 (…L1(x)))

[0051] where x is the input sample, D is the independent domain discriminator, L1 is the first layer of fully connected neural network of the independent domain discriminator, L n-1 is the (n - 1)-th layer of fully connected neural network of the independent domain discriminator, and L n is the n-th layer of fully connected neural network of the independent domain discriminator.

[0052] The loss function during the training of the independent domain discriminator D is:

[0053]

[0054] where N is the total number of samples in the training set containing source domain data and target domain data, x q represents the input sample at the q-th time step, and d q represents the domain label to which x q corresponds, 0 represents belonging to the source domain, and 1 represents belonging to the target domain.

[0055] Step 3: Use the trained independent domain discriminator D to calculate the corresponding weights for the samples in the training set according to the weight calculation formula. The weight calculation formula is:

[0056]

[0057] where x is the input sample, and w(x) is the weight of the input sample x calculated.

[0058] Step 4: Construct a feature extraction network G, a feature domain discriminator F d and a regressor F r , where the feature extraction network G is composed of multiple layers of gated recurrent neural network layers, the input is the input sample, and the output is the embedded layer feature; the feature domain discriminator F dIt consists of multiple fully-connected neural networks. The input is the embedded layer features output by the feature extraction network, and the output is the domain discrimination label. The regressor consists of multiple fully-connected neural networks. The input is the embedded layer features output by the feature extraction network, and the output is the predicted value of the quality variable. And a training set containing source domain data and target domain data is used Train the feature extraction network G and the feature domain discriminator F d and the regressor F r .

[0059] The specific calculation process of the feature extraction network is as follows:

[0060] z = G(x) = GRU m (GRU m-1 (…GRU1(x)))

[0061] where x is the input sample, G is the feature extraction network, m is the number of layers of the gated recurrent neural network, and GRU1, GRU m-1 and GRU m represent the first, the (m - 1)-th, and the m-th gated recurrent neural network layers of the feature extraction network respectively, and z is the embedded layer features.

[0062] The specific calculation process of the feature domain discriminator is as follows:

[0063] F d (z) = L t (L t-1 (…L1(z)))

[0064] where F d is the independent domain discriminator, t represents the number of layers of the fully-connected neural network of the feature domain discriminator, and L1, L t-1 and L t represent the first, the (t - 1)-th, and the t-th fully-connected neural network layers of the feature domain discriminator respectively.

[0065] The specific calculation process of the regressor is as follows:

[0066] F r (z) = L p (L p-1 (…L1(z)))

[0067] where F r is the regressor, p represents the number of layers of the fully-connected neural network. Among them, L1, L p-1 and L p represent the first, the (p - 1)-th, and the p-th fully-connected neural network layers of the feature domain discriminator respectively.

[0068] The feature extraction network G and the feature domain discriminator F dAnd the regressor F r During training, the loss function is as follows:

[0069] L1 = L reg (F r , G) - αL wdisc (F d , G) + γL scale (G)

[0070] L2 = L wdisc (F d , G)

[0071] Among them, L reg (F r , G) is the training loss of the regressor, and L wdisc (F d , G) is the weighted domain adversarial training loss for adversarial training between the feature extraction network and the feature domain discriminator. L scale (G) is the scale-preserving loss function, α is the weight coefficient of L wdisc (F d , G), and γ is the weight coefficient of L scale .

[0072] The representation forms of each loss term are as follows:

[0073] (1) Scale-preserving constraint L scale (G):

[0074] The scale-preserving mechanism aligns the feature scales of the target domain to the source domain explicitly to maintain the feature space scale, and its constraint term is expressed as:

[0075] L scale (G) = ||λ s,i＝1,…,k - λ t,i＝1,…,k ||2

[0076] Among them, G is the feature extraction network. Suppose the feature matrix obtained after the input sample matrix is processed by the feature extraction network is Z ∈ R n×g . In the above formula, λ s,i＝1,…,k is the i-th eigenvalue of the feature matrix corresponding to the source domain samples, and λ t,i＝1,…,k is the i-th eigenvalue of the feature matrix corresponding to the target domain samples. i can take 1,..., k, and k is the number of the main eigenvalues set artificially.

[0077] (2) Training loss L reg (F r , G)

[0078] is defined as follows:

[0079]

[0080] Among them, N s represents the number of source domain samples, and X j ={x j-l+1 , x j-l+2 , …, x j} represents the input sample matrix corresponding to the j-th time step. l is the length of the input sequence, and x j represents the input sample at the j-th time step. MSE represents the mean squared error loss function for regression, and y j represents the value of the quality variable corresponding to x j .

[0081] (3) The weighted domain adversarial training loss L wdisc (F d , G) is defined as follows:

[0082]

[0083] Among them, N is the total number of samples in the training set containing source domain data and target domain data. X q represents the q-th input sample matrix, d q represents the domain label to which X q belongs. 0 represents belonging to the source domain, and 1 represents belonging to the target domain. x q represents the input sample at the q-th time step, and w(x q ) is the weight of the sample x q . The specific adversarial training is as shown in Figure 2 .

[0084] Train the feature extraction network G, the feature domain discriminator F d and the regressor F r using the above training set data and loss function, and adjust the parameters of the independent domain discriminator D, the feature extraction network G, the feature domain discriminator F d and the regressor F r on the validation set.

[0085] Step Five: Collect unlabeled data of the target domain under the new working condition, input it into the trained feature extraction network G to obtain the embedded layer features, and then output the predicted quality variable by the trained regressor F r .

[0086] On the other hand, as shown in Figure 3As shown in the figure, an embodiment of the present invention further provides a die-casting quality prediction device based on deep transfer of key samples. The device includes a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the die-casting quality prediction method based on deep transfer of key samples in the above embodiment.

[0087] The die-casting quality prediction device based on deep transfer of key samples can be applied to any device with data processing capabilities. The any device with data processing capabilities can be a device or apparatus such as a computer. The die-casting quality prediction device based on deep transfer of key samples can be implemented by software, or by hardware, or by a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running them through the processor of any device with data processing capabilities where it is located. From the hardware level, as Figure 3 shown, it is a hardware structure diagram of any device with data processing capabilities where a die-casting quality prediction device based on deep transfer of key samples of the present invention is located. In addition to Figure 3 the processor, memory, network interface, and non-volatile memory shown, generally according to the actual functions of any device with data processing capabilities where the device of the present invention in the embodiment is located, other hardware may also be included, which will not be elaborated here.

[0088] For the specific implementation process of the functions and roles of each unit in the above device, please refer to the implementation process of the corresponding steps in the above method for details, which will not be elaborated here.

[0089] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0090] An embodiment of the present invention further provides a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the die-casting quality prediction method based on deep transfer of key samples in the above embodiment.

[0091] A computer-readable storage medium may be an internal storage unit of any data processing-capable device in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium may also be an external storage device of any data processing-capable device, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc., equipped on the device. Further, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any data processing-capable device. The computer-readable storage medium is used to store computer programs and other programs and data required by any data processing-capable device, and may also be used to temporarily store data that has been output or is to be output.

[0092] The effectiveness of the present invention is verified by combining a specific industrial production case below. The data of this case was obtained from a die-casting factory between January 1 and August 21, 2024, and the time interval of the data is about 30 seconds. Seven features, such as the compression stroke and the highest metal pressure during the boosting stage, are selected as input variables from the database of the die-casting factory, and the remaining casting thickness is selected as the quality variable for prediction. The source domain data are 5,000 production records of side cover products, and the target domain data are 5,000 production records of middle cover products.

[0093] To illustrate the effect of the method of the present invention, other existing quality prediction methods are also used for prediction. The methods for comparison include: Empirical Risk Minimization (hereinafter referred to as ERM), Transfer Component Analysis (hereinafter referred to as TCA), and Domain Adversarial Neural Network (hereinafter referred to as DANN). The model evaluation metrics include Root Mean Square Error (hereinafter referred to as RMSE) and Mean Absolute Error (hereinafter referred to as MAE).

[0094] In this embodiment, the feature extraction network is composed of three layers of gated recurrent neural networks. The length of each layer of gated recurrent neural network is 30. The hidden variable dimension of the first layer of gated recurrent neural network is 16, the hidden variable dimension of the second layer of gated recurrent neural network is 8, and the hidden variable dimension of the third layer of gated recurrent neural network is 8. The regressor is composed of a single layer of fully connected neural network, and the hidden variable dimension is 4. The independent domain discriminator and the feature domain discriminator have the same structure, both of which are composed of three layers of fully connected neural networks. The hidden variable dimension of the first layer of fully connected neural network is 8, the hidden variable dimension of the second layer of fully connected neural network is 4, and the hidden variable dimension of the third layer of fully connected neural network is 1. The weight coefficient α of the weighted domain adversarial training loss function L wdisc takes a value of 0.02, and the weight coefficient γ of the constraint term L scale takes a value of 5.

[0095] The quality prediction effects of each method on the die-casting online samples (test set data) are shown in Table 1.

[0096] Table 1 Quality prediction effects of each method on the die-casting online samples (test set data)

[0097] Method RMSE of the test set MAE of the test set ERM 3.253 2.500 TCA 3.391 2.610 DANN 3.168 2.417 The present invention 3.088 2.377

[0098] It can be seen from the results in Table 1 that the present invention exhibits the highest performance in all evaluation indicators, thus confirming the effectiveness and superiority of the present invention in predicting the quality of die-casting under variable working conditions.

[0099] Those of ordinary skill in the art can understand that the above are only preferred examples of the invention and are not used to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, for those skilled in the art, they can still modify the technical solutions described in the foregoing examples, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, etc. made within the spirit and principle of the invention shall be included within the protection scope of the invention.

Claims

1. A die-casting quality prediction method based on deep transfer of key samples, characterized in that, The method includes the following steps: S1: Collect the labeled data of the main operating conditions during the historical operation of the industrial process as the source domain data, collect the unlabeled data under the new operating conditions as the target domain data, and preprocess the data; S2: Construct an independent domain discriminator composed of n layers of fully connected neural networks, and use the training set containing the source domain data and the target domain data to train the independent domain discriminator; S3: Use the trained independent domain discriminator to calculate the corresponding weights for the samples in the training set; S4: Construct a feature extraction network, a feature domain discriminator, and a regressor, and use the training set containing the source domain data and the target domain data to train the three; the feature extraction network is composed of multiple layers of gated recurrent neural network layers, the input is the input sample, and the output is the feature of the embedding layer; The feature domain discriminator and the regressor are both composed of multiple layers of fully connected neural networks, the input of both is the feature of the embedding layer, the output of the feature domain discriminator is the domain discrimination label, and the output of the regressor is the predicted value of the quality variable; S5: Collect the unlabeled data of the target domain under the new operating conditions, input it into the trained feature extraction network, and the trained regressor outputs the predicted quality variable.

2. The die-casting quality prediction method based on deep transfer of key samples according to claim 1, wherein The loss function during the training of the independent domain discriminator is: where N is the total number of samples in the training set containing source domain data and target domain data, and x q represents the input sample at the q-th time step, and d q represents the domain label to which x q corresponds. 0 represents belonging to the source domain, and 1 represents belonging to the target domain.

3. The die-casting quality prediction method based on deep transfer of key samples according to claim 2, characterized in that, The calculation formula for using the trained independent domain discriminator to calculate the corresponding weights for the samples in the training set is as follows: Where x is the input sample, and w(x) is the weight of the input sample x calculated.

4. The die-casting quality prediction method based on deep transfer of key samples according to claim 1, wherein The loss functions for training the feature extraction network, the feature domain discriminator, and the regressor using the training set containing the source domain data and the target domain data are as follows: L1 = L reg (F r , G) - αL wdisc (F d , G) + γL scale (G) L2 = L wdisc (F d , G) Among them, L reg (F r , G) is the training loss function of the regressor, and L wdisc (F d , G) is the weighted domain adversarial training loss function for the adversarial training of the feature extraction network and the feature domain discriminator. L scale (G) is the scale-preserving loss function, α is the weight coefficient of L wdisc (F d , G), and γ is the weight coefficient of L scale .

5. The die-casting quality prediction method based on deep transfer of key samples according to claim 4, characterized in that The calculation formula for the scale-preserving loss function is as follows: L scale (G) = ||λ s,i＝1,…,k -λ t,i＝1,…,k ||² Among them, G is the feature extraction network, and λ s,i＝1,…,k is the i-th eigenvalue of the feature matrix obtained after the source domain samples are processed by the feature extraction network, and λ t,i＝1,…,k is the i-th eigenvalue of the feature matrix obtained after the target domain samples are processed by the feature extraction network. k is the number of artificially set main eigenvalues.

6. The die-casting quality prediction method based on deep transfer of key samples according to claim 4, wherein The expression of the training loss function of the regressor is as follows: Among them, N s represents the number of source domain samples, X j ={x j-l+1 , x j-l+2 , …, x j} represents the input sample matrix corresponding to the j-th time step, l is the length of the input sequence, x j represents the input sample at the j-th time step, MSE represents the mean squared error loss function for regression, y j represents the value of the quality variable corresponding to x j .

7. The die-casting quality prediction method based on deep transfer of key samples according to claim 4, wherein The weighted domain adversarial training loss function \(L\) for adversarial training of the feature extraction network and the feature domain discriminator wdisc (F d , G) is expressed as follows: where N is the total number of samples in the training set containing source domain data and target domain data, X q represents the q-th input sample matrix, d q represents the domain label to which X q corresponds. 0 represents belonging to the source domain, 1 represents belonging to the target domain, and x q represents the input sample at the q-th time step, w(x q ) is the weight of the sample x q .

8. A die-casting quality prediction device based on deep transfer of key samples, characterized in that, The device includes a memory and one or more processors, and is used to implement the die-casting quality prediction method based on deep transfer of key samples described in any one of claims 1 to 7.

9. An electronic device, characterized in that, Including: One or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the electronic device, enabling the electronic device to implement the die-casting quality prediction method based on deep transfer of key samples described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A program is stored thereon, and when the program is executed by the processor, it implements the die-casting quality prediction method based on deep transfer of key samples described in any one of claims 1 to 7.

Images