Industrial process small sample data generation method and device in combination with domain knowledge

Abstract

The invention provides an industrial process small sample data generation method and device in combination with domain knowledge, and relates to the technical field of data generation. The method comprises the steps that fault attribute description corresponding to industrial process data is collected, and first training data and second training data are obtained; constructing a to-be-trained data generation model combined with domain knowledge based on the network structure of the generative adversarial network; training a to-be-trained data generation model by using the first training data and the second training data to obtain a data generation model; acquiring to-be-generated sample data; and inputting the to-be-generated sample data into the data generation model for data generation to obtain generated sample data. The invention provides an efficient and stable small sample data generation method based on domain knowledge.

Description

Technical Field

[0001] The present invention relates to the technical field of data generation, and in particular to a method and device for generating small sample data of an industrial process in combination with domain knowledge. Background Art

[0002] With the rapid development of information technology and data science, deep learning technology has been widely used in many fields, including but not limited to image recognition, natural language processing, speech recognition, medical diagnosis, autonomous driving, financial risk control, and complex decision support systems. The breakthrough progress in these fields has not only promoted the development of intelligent applications, but also changed the operation mode of traditional industries to a certain extent, and improved the level of automation and intelligent decision-making capabilities. Deep learning relies on large-scale data and powerful computing power, which enables it to achieve performance that surpasses traditional methods in many tasks, especially in processing high-dimensional nonlinear data, automatic feature learning, and end-to-end task optimization.

[0003] Although deep learning has achieved remarkable results, it still faces many challenges in its application. Among them, the model's high dependence on a large amount of labeled data, the black-box nature of the model's decision-making process, unstable generalization ability, and difficult-to-explain learning mechanism are all core issues that restrict its further promotion and implementation. In practical applications, the cost of data acquisition and labeling is often high, especially in the fields of medicine, industrial manufacturing, law, etc. High-quality labeled data is scarce, and the characteristics of the data itself are weakly correlated, resulting in limited generalization capabilities of deep learning models. In addition, the complexity of deep models makes their decision-making process difficult for humans to understand, and the interpretability of model results has become an important bottleneck in industry applications, especially in areas involving security, compliance, and trusted AI. This problem is particularly prominent.

[0004] Especially in industrial scenarios, a large amount of data containing deep domain knowledge has not been fully utilized, making it difficult for the model to effectively learn industry-specific laws and experiences. Complex systems such as industrial manufacturing, energy management, and intelligent scheduling often rely on expert knowledge, physical modeling, and engineering experience, and this knowledge cannot be directly learned automatically through deep learning. Therefore, how to efficiently combine domain knowledge with deep learning methods to improve the generalization ability, interpretability, and stability of the model has become an important issue that needs to be solved urgently. In recent years, academia and industry have carried out many explorations in this direction, such as deep learning methods based on knowledge graphs, neural networks combined with physical models, and the integration of reinforcement learning and expert systems, all of which have provided new ideas for solving this problem.

[0005] In the existing technology, there is a lack of an efficient and stable small sample data generation method based on domain knowledge. Summary of the invention

[0006] To solve the technical problems in the prior art that in industrial scenarios, a large amount of domain knowledge fails to be fully utilized, and in the case of scarce data or unclear relevance of data features, the generalization ability and interpretability of the model are poor, the embodiments of the present invention provide a method and device for generating small-sample data of industrial processes combined with domain knowledge. The technical solutions are as follows:

[0007] On the one hand, a method for generating small-sample data of industrial processes combined with domain knowledge is provided. This method is implemented by a small-sample data generation device for industrial processes, and the method includes:

[0008] Collect the fault attribute descriptions corresponding to the industrial process data to obtain the first training data and the second training data;

[0009] Construct a to-be-trained data generation model combined with domain knowledge based on the network structure of the generative adversarial network;

[0010] Use the first training data and the second training data to train the to-be-trained data generation model to obtain a data generation model;

[0011] Obtain the to-be-generated sample data; input the to-be-generated sample data into the data generation model for data generation to obtain the generated sample data.

[0012] On the other hand, a device for generating small-sample data of industrial processes combined with domain knowledge is provided. This device is applied to the method for generating small-sample data of industrial processes combined with domain knowledge, and the device includes:

[0013] A data acquisition module, configured to collect the fault attribute descriptions corresponding to the industrial process data to obtain the first training data and the second training data;

[0014] A model construction module, configured to construct a to-be-trained data generation model combined with domain knowledge based on the network structure of the generative adversarial network;

[0015] A model training module, configured to use the first training data and the second training data to train the to-be-trained data generation model to obtain a data generation model;

[0016] A data generation module, configured to obtain the to-be-generated sample data; input the to-be-generated sample data into the data generation model for data generation to obtain the generated sample data.

[0017] On the other hand, a small-sample data generation device for industrial processes is provided. The small-sample data generation device for industrial processes includes: a processor; a memory, on which computer-readable instructions are stored. When the computer-readable instructions are executed by the processor, any one of the methods for generating small-sample data of industrial processes combined with domain knowledge as described above is implemented.

[0018] On the other hand, a computer-readable storage medium is provided, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement any one of the above-mentioned industrial process small-sample data generation methods combined with domain knowledge.

[0019] The beneficial effects brought by the technical solutions provided in the embodiments of the present invention at least include:

[0020] The present invention proposes an industrial process small-sample data generation method combined with domain knowledge. Through the fault attribute extractor Robust-Attri-FocusNet, the effective combination of domain knowledge is realized, and the model accuracy and robustness are improved; through the dual-generator GAN architecture and the introduction of a dynamic gradient adjustment technology with soft parameter sharing, it is ensured that the two generators promote each other. The effective integration of these two components ensures the effective generation of small-sample data under the guidance of domain knowledge. The present invention is an efficient and stable small-sample data generation method based on domain knowledge. Description of the Drawings

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a flowchart of an industrial process small-sample data generation method combined with domain knowledge provided by an embodiment of the present invention;

[0023] Figure 2 It is a block diagram of an industrial process small-sample data generation device provided by an embodiment of the present invention;

[0024] Figure 3 It is a schematic structural diagram of an industrial process small-sample data generation device provided by an embodiment of the present invention. Detailed Embodiments

[0025] The following will describe the technical solutions in the present invention with reference to the drawings.

[0026] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.

[0027] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "Of", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.

[0028] In the embodiments of the present invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.

[0029] To make the technical problems, technical solutions, and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0030] The embodiments of the present invention provide a method for generating small-sample industrial process data combined with domain knowledge. This method can be implemented by a small-sample industrial process data generation device, which can be a terminal or a server. As Figure 1 shown in the flowchart of the method for generating small-sample industrial process data combined with domain knowledge, the processing flow of this method can include the following steps:

[0031] S1. Collect the fault attribute descriptions corresponding to the industrial process data to obtain the first training data and the second training data.

[0032] Among them, the data volume of the first training data is greater than or equal to 1000 pieces;

[0033] The data volume of the second training data is 10% of the data volume of the first training data.

[0034] In a feasible implementation manner, the present invention uses the industrial process data with a normal sample volume and the corresponding fault attribute descriptions as the first training data; and uses the industrial process data with a small sample volume and the corresponding fault attribute descriptions as the second training data.

[0035] S2. Construct a training data generation model combined with domain knowledge based on the network structure of the generative adversarial network.

[0036] Among them, the training data generation model includes a first generator, a second generator, a discriminator, and a fault attribute extractor.

[0037] In a feasible implementation manner, a method for generating small-sample industrial process data integrating domain knowledge (Domain-knowledge-guided-Generative Adversarial Network, DKG-GAN) proposed by the present invention aims to integrate domain knowledge into the generation process of a generative adversarial network (Generative Adversarial Network, GAN), guide and expand the distribution of generated data by introducing additional conditional constraints, and increase the diversity of small-sample data. Domain knowledge, as a supplementary learning source of data, guides the generator to explore different data distributions.

[0038] In a traditional GAN, there is a contradiction between the diversity of the generator and the performance of the discriminator, because an efficient discriminator can accurately distinguish real data from generated data, which will inevitably limit the diversity of data generated by the generator. To solve this problem, the present invention proposes a dual-generator network framework based on the generative adversarial network, including a first generator, a second generator, a discriminator, and a fault attribute extractor; the first generator G1 and the second generator G2 share parameters, the first generator is trained using normal data volume data, and the second generator is trained using small-sample data.

[0039] Among them, the fault attribute extractor is composed of multiple classifiers;

[0040] The classifier is constructed based on an attention module and a convolutional neural network;

[0041] The classifier is obtained after being trained using domain knowledge data of the industrial process;

[0042] The number of classifiers corresponds to the number of fault types of the industrial process.

[0043] In a feasible implementation manner, an efficient and stable attribute description extractor is crucial for the knowledge injection process, and can ensure that the generation model generates high-quality results under the guidance of accurate and effective domain knowledge. In practical applications, many generation models often ignore the in-depth mining of data features. By accurately extracting the key attribute features of data, more rich and useful information can be provided for the generation process.

[0044] The present invention proposes a Robust-Attri-FocusNet for fault attribute extraction with high precision and robustness. The fault attribute extractor consists of multiple independent classifiers, each of which is trained for a specific descriptive variable. These classifiers can identify and classify the three states (0-1-2) of fault features, assign corresponding weights to each type of fault feature, and provide a comprehensive knowledge introduction basis for the generation model. Among them, 0 indicates that the device is in normal working state; 1 indicates that the device is in a faulty state; 2 indicates that the device is in a severe faulty state.

[0045] Each independent classifier can analyze fault features from different perspectives, enabling the entire model to capture the details of data at multiple levels and enhancing the model's adaptability to different types of data. In a classification model based on a Convolutional Neural Network (CNN), many features have a strong correlation with the classification results. Therefore, it is particularly important to introduce a feature-based attention mechanism. Through this mechanism, key features can be weighted at the input stage, giving higher weights to important features, so that the model pays more attention to these features during the learning process.

[0046] The attention mechanism can dynamically adjust the weights of features, strengthen the model's perception ability of key data, and improve the learning efficiency and accuracy of the network for these features. Especially when facing complex industrial data, this mechanism can effectively improve the robustness and precision of the model. The fault attribute extractor is designed based on this concept, combining a convolutional neural network and an attention mechanism, and can deeply learn data features at multiple levels and dimensions. The calculation of attention weights uses the softmax function to normalize the weights of each feature, ensuring that the sum of the weights of all features is 1. Through this normalized weight assignment, the model can flexibly adjust the degree of attention to each feature according to the different characteristics of the data and optimize the learning process. Attention weights are calculated according to the following formula (1):

[0047] (1);

[0048] where represents the k-th element of the input feature vector, represents the l-th element of the input feature vector, l is an index parameter, exp represents the exponential function, N is the total number of vectors, and W and b are the weight matrix and bias vector respectively, which determine how each input feature affects its corresponding attention weight.

[0049] Each classifier can identify different subsets of features in the input data set and classify the output results according to the importance of the features. By processing the feature variables separately, these classifiers can focus on the key factors in the data and provide more accurate attribute description information for the generation process. By effectively aggregating the outputs of each classifier, a fault attribute description matrix is ​​formed, which can comprehensively and accurately express the various features and severity of the fault data. Each classifier extracts specific features from the data and analyzes them.

[0050] Since different features have different effects on the manifestation of faults, the classifier can weight them according to their relevance and importance, and reflect the relative importance of these features in the output. This refined classification capability helps the model accurately identify and generate data in complex industrial environments, especially in the case of small sample learning. The output of the classifier will be weighted and summarized according to the importance of the features, and finally form a comprehensive fault attribute description matrix. The process is as follows (2):

[0051] (2);

[0052] in, The nth classifier The output result vector. represents the fault attribute description matrix, where i and j correspond to the rows and columns of the matrix.

[0053] This matrix not only reflects the basic situation of each fault, but also accurately represents the different severity of the fault through three-level coding (0-1-2). In this way, the matrix can provide a structured and efficient input for the subsequent data generation process, so that the generation model can better learn and adjust with the help of domain knowledge.

[0054] In a feasible implementation, the industrial environment itself is highly complex and variable, with factors including temperature fluctuations, mechanical wear, operational errors and other uncertainties, which will have a significant impact on the stability and accuracy of the system. Especially for small sample data environments, the quality and characteristics of each sample may greatly affect the final output of the model. Therefore, how to improve the robustness of the attribute extractor so that it can still accurately identify and classify various fault features under various conditions has become the key to ensuring the stable operation of the system. In order to maintain robustness and accuracy, the attribute extractor must perform well in the face of both natural errors (i.e., normal classification errors) and disturbance errors.

[0055] To maintain its robustness and accuracy, the attribute extractor must perform well in the face of both natural errors (i.e., normal classification errors) and perturbation errors. To this end, an adversarial defense method that combines projected gradient descent and surrogate-loss minimization based on the tradeoff idea (Project Gradient Descent-TRadeoff-inspired Adversarial DEfense via Surrogate-loss minimization, PGD-TRADES) is proposed to enhance the adversarial robustness of the model and maintain the natural accuracy of the model during this process. The core idea of the PGD-TRADES method is to enhance adversarial robustness by pushing the decision boundary away from the data points, while minimizing the impact on the model's classification ability under normal conditions. The PGD-TRADES method achieves this goal by minimizing the difference between the natural samples and the predictions of their corresponding adversarial samples. Adversarial samples are close to the original samples in the input space and are usually generated by adding small perturbations to the original data, but these perturbations are sufficient to cause the model to make incorrect classification judgments.

[0056] Adversarial samples are usually located near the original samples and attempt to cross the model's decision boundary. The smoother the decision boundary, the less likely a small perturbation is to push the input data from one class region to another. Therefore, the robustness of the model under adversarial attacks is enhanced. The introduction of this mechanism enables the model to maintain a high level of stability when dealing with noise and perturbations in the actual environment and ensures that its classification accuracy is not overly lost. In the generation and fault detection of industrial data, this method can help the model effectively resist various uncertainties and external perturbations and improve the practical application ability of the system.

[0057] Through the PGD-TRADES method, the model can not only enhance its robustness in the face of perturbations but also maintain a high natural accuracy in classification tasks. When dealing with small-sample data in complex industrial environments, it can perform fault identification and data generation more precisely and efficiently. The introduction of this method provides strong guarantees for small-sample learning, ensuring that even when the data is scarce, the model can operate stably under different environmental changes and provide accurate results. According to the characteristics of the model, Projected Gradient Descent (PGD) in white-box attacks is adopted to find the perturbation that can maximize the loss by applying small but meaningful perturbations to the input x at each step. The update step of PGD is as follows in Equation (3):

[0058] (3);

[0059] where, is the step size, is the sign function (+1 or -1), is the data after adding perturbations, is the original data, Finally, ensure that the movement does not exceed the feasible region through projection of the limit, (i.e., ), is the loss function of the fault attribute extractor, is the gradient update sign, represents the range of perturbations.

[0060] The fault attribute extractor pre-completes model training with sufficient normal sample data and is integrated into DKG-GAN to ensure that the data generation process can be effectively guided by knowledge. Through the PGD-TRADES method, the fault attribute extractor can not only handle data changes in the natural environment but also effectively defend against adversarial attacks that may occur in practical applications, improving the overall robustness and reliability of the system while ensuring the accuracy of fault detection. The loss function of the fault attribute extractor is as follows in Equation (4):

[0061] (4);

[0062] Among them, the cross-entropy loss (CE) is used to evaluate the difference between the model's prediction of each attribute value and the actual label. is the regularization coefficient, used to balance the original accuracy term and the adversarial robustness accuracy term, is the expected value, represents data sampling. The above settings enable the fault attribute extractor to not only handle natural changes in the data but also effectively defend against adversarial attacks that may occur in practical applications, ensuring the accuracy of fault detection and improving the overall stability and reliability of the system.

[0063] S3. Use the first training data and the second training data to train the data generation model to be trained to obtain the data generation model.

[0064] Optionally, using the first training data and the second training data to train the data generation model to be trained to obtain the data generation model includes:

[0065] Input the first training data into the fault attribute extractor to obtain the first fault attribute description;

[0066] Based on the first fault attribute description, use the first training data to generate data through the first generator to obtain the first generated data;

[0067] Based on the discriminator loss function, calculate the loss function according to the first training data and the first generated data to obtain the discriminator loss; according to the discriminator loss, perform reverse parameter optimization on the discriminator to obtain an optimized discriminator;

[0068] Based on the first generator loss function, calculate the loss function according to the first training data, the first generated data, and the first fault attribute description to obtain the first generator loss; according to the first generator loss, perform reverse parameter optimization on the first generator to obtain an optimized first generator;

[0069] Input the second training data into the fault attribute extractor to obtain the second fault attribute description;

[0070] Assign the model parameters of the first optimized generator to the second generator to obtain a preliminarily optimized second generator;

[0071] Based on the second fault attribute description, use the second training data, and generate data through the preliminarily optimized second generator to obtain the second generated data;

[0072] Based on the second generator loss function, calculate the loss function according to the second training data, the second generated data, and the second fault attribute description to obtain the second generator loss; according to the second generator loss, perform reverse parameter optimization on the preliminarily optimized second generator to obtain an intermediate optimized second generator;

[0073] Based on the soft parameter sharing method with dynamic gradient adjustment, adjust the parameter weights of the intermediate optimized second generator according to the first generator loss and the second generator loss to obtain an optimized second generator;

[0074] Obtain a data generation model according to the optimized discriminator, the optimized first generator, the optimized second generator, and the fault attribute extractor.

[0075] In a feasible implementation manner, in the present invention, the first generator G1 is trained together with the discriminator D and the fault attribute extractor E. The training combination [G1, D, E] enables G1 to generate data from the data categories of the normal data volume, and the discriminator D can effectively distinguish real data from generated data, and the fault attribute extractor E can accurately identify the fault attribute description.

[0076] The second generator G2 is trained together with the fault attribute description extractor E. The combined [G2, E] enables G2 to explore and generate small-sample fault data under the guidance of the fault attribute description of the small-sample data. This design enables G2 to explore in the area outside the data distribution of the normal data volume, improving the model adaptability and generating small-sample fault types.

[0077] Since G1 and G2 share parameters, G1 also indirectly obtains the ability of G2 to generate small-sample data, thereby improving the generality and flexibility of the model. This mechanism ensures that the expansion of the output distribution of the generator does not sacrifice the performance of the discriminator, thus increasing the diversity and coverage of the generated data while ensuring quality. The discriminator loss function, the first generator loss function, the second generator loss function, and the objective function of the model architecture are as shown in the following equations (5), (6), (7), and (8):

[0078] (5);

[0079] where z is noise data sampled from .

[0080] (6);

[0081] where is a hyperparameter, LFD(G1) is the loss of the first generator; LFD(G2) is the loss of the second generator.

[0082] (7);

[0083] (8);

[0084] where A is the fault attribute matrix.

[0085] Optionally, based on the soft parameter sharing method with dynamic gradient adjustment, according to the loss of the first generator and the loss of the second generator, the parameter weights of the intermediate optimized second generator are adjusted to obtain the optimized second generator, including:

[0086] Calculating the shared parameter gradient according to the loss of the first generator and the loss of the second generator;

[0087] Conducting an adaptability analysis on the shared parameter gradient to obtain the adaptability analysis result;

[0088] Based on the adaptability analysis result, dynamically adjusting the weight of the shared parameter gradient to obtain the adjusted shared parameter gradient;

[0089] Based on the adjusted shared parameter gradient, performing reverse parameter optimization on the intermediate optimized second generator to obtain the optimized second generator.

[0090] In a feasible implementation, the setting of parameter sharing may cause interference between the two generators, especially when there are conflicts in the objectives of the generators. If one generator is too biased towards generating a certain type of sample, its bias may affect the training effect of the other generator, resulting in a decrease in the quality of the samples generated by the other generator.

[0091] This conflict weakens the overall performance of the model, especially when dealing with diverse fault types and small sample data, which may prevent the generator from effectively exploring the entire data distribution. To address the problem of interference between models caused by shared parameters, especially when the generator objectives conflict, a soft parameter sharing method with dynamic gradient adjustment is proposed. The shared parameters are defined as Equation (9) below:

[0092] (9);

[0093] where represents a set of fully shared parameters for serving two tasks simultaneously.

[0094] The feature requirements between tasks will be balanced by applying dynamically adjusted weight decay to the shared parameters. For two tasks sharing the same set of parameters, a regularization term is added to the objective function to help the two generators maintain a balance between task similarity and independence under the shared parameters. The new objective function can be expressed as Equations (10) and (11) below:

[0095] (10);

[0096] (11);

[0097] where is a hyperparameter controlling the regularization strength, measures the similarity of the shared parameters of the two tasks using the Euclidean distance.

[0098] This regularization term encourages the values of the shared parameters in the two tasks to remain close, balances the requirements between tasks, and prevents a task from deviating too much from the shared parameters. In addition to introducing the regularization term, the weight decay of the shared parameters is dynamically adjusted through a gradient feedback mechanism to further improve the adaptability of the model under different tasks. The core of this dynamic adjustment mechanism is to adjust the learning rate by feeding back the gradient of the shared parameters, enabling the shared parameters to vary flexibly according to needs under different tasks. The parameter update rules are as follows in Equations (12), (13), and (14):

[0099] (12);

[0100] (13);

[0101] (14);

[0102] where is the learning rate, γ is the adjustment coefficient of gradient feedback, controlling the degree of influence of the gradient on the weight decay, is a hyperparameter Adjust the weight decay according to the gradient change of each task, indicating the gradient of task i with respect to the shared parameter . Here, q, p, and k are all index parameters, and L is the generator loss function as shown in formulas (6) and (7). In this way, the model can dynamically adjust the learning strategy of the shared parameter according to the different requirements of tasks, so as to maintain good adaptability and generalization ability during the training process.

[0103] is the scaling factor for the scaling factor, Adjust the weight decay according to the gradient change of each task. If the gradient change of a certain task is large, increase the weight decay of the corresponding parameter of this task to avoid over-reliance on the shared parameter; indicating the gradient of task i with respect to the shared parameter . The above formula is used to measure the magnitude of the gradient and dynamically adjust the update amplitude of the parameter.

[0104] Through the gradient information of backpropagation, evaluate the adaptability of the shared parameter among tasks, ensure that the shared parameter can be reasonably adjusted among different tasks, and enable the model to coordinate and operate efficiently during multi-task learning. Through the dynamic adjustment mechanism and the proposed soft parameter sharing method, it is possible to minimize the interference between tasks on the basis of ensuring the consistency of the generator target, and promote the collaborative optimization of the generator.

[0105] S4. Obtain the sample data to be generated; input the sample data to be generated into the data generation model for data generation to obtain the generated sample data.

[0106] In a feasible implementation manner, verify the distribution fitting degree of the data generated by the method of the present invention. The data set includes: consider the following types of data as the data types with a normal data volume that can be collected (1000 pieces): normal, top pressure extremely high, and hot hanging charge. Consider the following types of data as having only a small sample data volume (100 pieces, imbalance rate 3.03%): pipeline air flow, wind stoppage, and coal stoppage.

[0107] The attribute descriptions of the working conditions of the above various types of data are collected at the industrial site as shown in Table 1 (Blast Furnace Fault Attribute Description Table) below. According to the fault description, obtain the attribute description matrix of each working condition, as shown in Table 2 (Blast Furnace Fault Attribute Description Matrix Table) below.

[0108] Table 1

[0109]

[0110] Table 2

[0111]

[0112] Select mature generative models in the prior art for horizontal comparison, including GAN, Wasserstein Generative Adversarial Networks (WGAN), Least Squares Generative Adversarial Networks (LSGAN), and Auxiliary Classifier Generative Adversarial Network (ACGAN), to evaluate the advantages of the proposed method over traditional methods.

[0113] Traditional GAN, WGAN, and LSGAN models do not directly utilize domain knowledge. These models usually can only learn implicit features from data and cannot effectively introduce external knowledge. Although ACGAN is improved by using class labels as additional knowledge, in the present invention, ACGAN is further improved by replacing the traditional knowledge introduction method with a fault attribute matrix to better combine domain-specific knowledge.

[0114] Since existing GAN models generally have difficulty effectively combining external knowledge and lack models similar to DKG-GAN that can effectively introduce domain knowledge, it is difficult to find a completely identical comparison model that can effectively combine knowledge. t-Distributed Stochastic Neighbor Embedding (t-SNE) is used to visualize the distributions of generated data and real data, and Jensen-Shannon Divergence (JS) and Maximum Mean Discrepancy (MMD) are applied to quantitatively evaluate the performance of the models in the data generation task. The results are shown in Table 3 (Quantitative Evaluation Results Table for Horizontally Compared Generated Data).

[0115] Table 3

[0116]

[0117] According to Table 3, the method proposed in the present invention is superior to the compared generative models in both evaluation metrics, highlighting the key role of domain knowledge in improving the authenticity and diversity of generated data.

[0118] A longitudinal comparison was made by changing the density of knowledge integration in the model. The density was divided into: low density, medium density, and high density. In the low-density setting, only a single variable of the fault attribute description matrix was used; in the medium-density setting, some variables in the fault attribute description matrix were adopted; in the high-density setting, all variables in the fault attribute description matrix were used. This method can observe the gradual impact of different levels of knowledge input on the model performance. The results are shown in Table 4 below (Quantitative Evaluation Results Table of Data Generated by Longitudinal Comparison). Incorporating more detailed fault knowledge helps to make up for potential data defects or quality problems in the dataset, especially when the sample size is limited. Because the inclusion of detailed knowledge provides additional context information, enabling the model to understand and learn complex patterns that are difficult to extract from sparse data. The success of high-density knowledge guidance proves the ability of the model to capture and utilize the complex interrelationships between features. Therefore, the model can better understand the interactions and dependencies between various fault features, improving the accuracy of fault prediction and the authenticity of the generated data.

[0119] Table 4

[0120]

[0121] From the comparison results, it can be seen that the accuracy of knowledge integration is crucial for the success of the model. Compared with low density, the JS divergence and MMD increase by 67.74% and 160.47% respectively when high-density knowledge is introduced. When the density of knowledge introduction is insufficient, the attribute extractor cannot accurately classify the data categories, leading to misguidance for the data generation of the generative network. It can be clearly seen from the results that the introduction of low-density knowledge performs even worse than the method without any knowledge integration. These findings emphasize that high-quality and well-calibrated knowledge introduction is crucial for improving the performance of the generative model.

[0122] Verify the ability of the method of the present invention to solve practical small-sample problems. GAN, WGAN, LSGAN, ACGAN, and the present method were used to generate data, and each small-sample data type was extended to 1000 samples to obtain a balanced dataset after data augmentation for further analysis.

[0123] To evaluate the effectiveness of the generated data, the augmented dataset and the original dataset were used to train the classifier. Then, the performance of the classifier (Support Vector Machine, SVM) was evaluated on a test set composed of 40 normally operating data samples and 40 samples of each fault type. By comparing the performance of the classifier with the data-augmented dataset generated by different methods as the training set background, the usability of the generated data of the generative model was evaluated. The F1 score and G-means were used as quantitative indicators, and the confusion matrix was used as a visualization tool.

[0124] Table 5

[0125]

[0126] As can be seen from Table 5 (a comparison table of evaluation indicators of the SVM classifier confusion matrix under different data augmentation methods), all augmentation methods have improved in terms of fault classification performance. However, the F1 scores and G-means of fault classification achieved by GAN, WGAN, LSGAN, and ACGAN augmentations are still below 90%, which is not sufficient to support industrial-level fault diagnosis. In contrast, DKG-GAN has significantly improved in terms of the classification F1 scores and G-means of all fault types, and both of these indicators exceed 90% in various fault detections, which demonstrates that the proposed method enhances the accuracy of industrial fault diagnosis under the small sample problem.

[0127] The present invention proposes an industrial process small sample data generation method combining domain knowledge. Through the fault attribute extractor Robust-Attri-FocusNet, the effective combination of domain knowledge is realized, improving the model accuracy and robustness; through the dual-generator GAN architecture and introducing a dynamic gradient adjustment technology with soft parameter sharing, it ensures that the two generators promote each other. The effective integration of these two components ensures the effective generation of small sample data under the guidance of domain knowledge. The present invention is an efficient and stable small sample data generation method based on domain knowledge.

[0128] Figure 2 It is a block diagram of an industrial process small sample data generation device showing according to an exemplary embodiment. This device is used for the industrial process small sample data generation method combining domain knowledge. Refer to Figure 2 , this device includes a data acquisition module 210, a model construction module 220, a model training module 230, and a data generation module 240. Among them:

[0129] The data acquisition module 210 is used to collect the fault attribute descriptions corresponding to the industrial process data to obtain the first training data and the second training data;

[0130] The model construction module 220 is used to construct a to-be-trained data generation model combining domain knowledge based on the network structure of the generative adversarial network;

[0131] The model training module 230 is used to use the first training data and the second training data to train the to-be-trained data generation model to obtain a data generation model;

[0132] The data generation module 240 is used to obtain the to-be-generated sample data; input the to-be-generated sample data into the data generation model for data generation to obtain the generated sample data.

[0133] Among them, the data volume of the first training data is greater than or equal to 1,000 pieces;

[0134] The data volume of the second training data is 10% of the data volume of the first training data.

[0135] Among them, the data generation model to be trained includes a first generator, a second generator, a discriminator, and a fault attribute extractor.

[0136] Among them, the fault attribute extractor is composed of multiple classifiers;

[0137] The classifier is constructed based on the attention module and the convolutional neural network;

[0138] The classifier is obtained after being trained with the domain knowledge data of the industrial process;

[0139] The number of classifiers corresponds to the number of fault types of the industrial process.

[0140] Optionally, the model training module 230 is further configured to:

[0141] Input the first training data into the fault attribute extractor to obtain the first fault attribute description;

[0142] Based on the first fault attribute description, use the first training data and generate data through the first generator to obtain the first generated data;

[0143] Based on the discriminator loss function, calculate the loss function according to the first training data and the first generated data to obtain the discriminator loss; according to the discriminator loss, perform reverse parameter optimization on the discriminator to obtain the optimized discriminator;

[0144] Based on the first generator loss function, calculate the loss function according to the first training data, the first generated data, and the first fault attribute description to obtain the first generator loss; according to the first generator loss, perform reverse parameter optimization on the first generator to obtain the optimized first generator;

[0145] Input the second training data into the fault attribute extractor to obtain the second fault attribute description;

[0146] Assign the model parameters of the first optimized generator to the second generator to obtain the preliminarily optimized second generator;

[0147] Based on the second fault attribute description, use the second training data and generate data through the preliminarily optimized second generator to obtain the second generated data;

[0148] Based on the second generator loss function, calculate the loss function according to the second training data, the second generated data, and the second fault attribute description to obtain the second generator loss; according to the second generator loss, perform reverse parameter optimization on the preliminarily optimized second generator to obtain the intermediate optimized second generator;

[0149] Based on the soft parameter sharing method with dynamic gradient adjustment, adjust the parameter weights of the intermediate optimized second generator according to the first generator loss and the second generator loss to obtain the optimized second generator;

[0150] Obtain the data generation model according to the optimized discriminator, the optimized first generator, the optimized second generator, and the fault attribute extractor.

[0151] Optionally, the model training module 230 is further configured to:

[0152] Calculate the gradient according to the first generator loss and the second generator loss to obtain the shared parameter gradient;

[0153] Conduct an adaptability analysis on the shared parameter gradient to obtain the adaptability analysis result;

[0154] Based on the adaptability analysis result, dynamically adjust the weight of the shared parameter gradient to obtain the adjusted shared parameter gradient;

[0155] Based on the adjusted shared parameter gradient, perform reverse parameter optimization on the intermediate optimized second generator to obtain the optimized second generator.

[0156] The present invention proposes a small sample data generation method for industrial processes combining domain knowledge. Through the fault attribute extractor Robust - Attri - FocusNet, the effective combination of domain knowledge is realized, improving the model accuracy and robustness; through the dual - generator GAN architecture and the introduction of the dynamic gradient adjustment technology with soft parameter sharing, it ensures that the two generators promote each other. The effective integration of these two components ensures the effective generation of small sample data under the guidance of domain knowledge. The present invention is an efficient and stable small sample data generation method based on domain knowledge.

[0157] Figure 3 It is a schematic structural diagram of a small sample data generation device for industrial processes provided by an embodiment of the present invention. As Figure 3 shown, the small sample data generation device for industrial processes may include the above - mentioned Figure 2 small sample data generation device for industrial processes combining domain knowledge shown. Optionally, the small sample data generation device 310 may include a first processor 2001.

[0158] Optionally, the industrial process small sample data generation device 310 may further include a memory 2002 and a transceiver 2003.

[0159] Among them, the first processor 2001, the memory 2002, and the transceiver 2003 may be connected through a communication bus.

[0160] Next, in conjunction with Figure 3 each component of the industrial process small sample data generation device 310 will be specifically introduced:

[0161] Among them, the first processor 2001 is the control center of the industrial process small sample data generation device 310, which may be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or may be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0162] Optionally, the first processor 2001 may execute various functions of the industrial process small sample data generation device 310 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0163] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 3 the CPU0 and CPU1 shown in

[0164] In a specific implementation, as an embodiment, the industrial process small sample data generation device 310 may also include multiple processors, such as Figure 3 the first processor 2001 and the second processor 2004 shown in

[0165] Among them, the memory 2002 is used to store the software program for executing the solution of the present invention, and is controlled by the first processor 2001 for execution. The specific implementation manner can refer to the above method embodiments and will not be elaborated here.

[0166] Optionally, the memory 2002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or may also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through the interface circuit ( Figure 3 not shown) of the industrial process small sample data generation device 310. The embodiments of the present invention do not make specific limitations in this regard.

[0167] The transceiver 2003 is used to communicate with a network device or with a terminal device.

[0168] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 3 not shown separately). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0169] Optionally, the transceiver 2003 may be integrated with the first processor 2001 or may exist independently and be coupled to the first processor 2001 through the interface circuit ( Figure 3 not shown) of the industrial process small sample data generation device 310. The embodiments of the present invention do not make specific limitations in this regard.

[0170] It should be noted that Figure 3 the structure of the industrial process small sample data generation device 310 shown does not constitute a limitation on the router. The actual knowledge structure recognition device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0171] In addition, for the technical effects of the industrial process small-sample data generation device 310, reference may be made to the technical effects of the industrial process small-sample data generation method combining domain knowledge described in the above method embodiments, which will not be elaborated herein.

[0172] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and this processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0173] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM) or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchlink DRAM (SLDRAM) and direct rambus RAM (DR RAM).

[0174] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0175] It should be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. Here, A and B can be singular or plural. Additionally, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood by referring to the context.

[0176] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0177] It should be understood that in various embodiments of the present invention, the magnitudes of the serial numbers of the above processes do not imply the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0178] Those of ordinary skill in the art will realize that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0179] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the devices, apparatuses, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0180] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0181] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0182] In addition, the functional units in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0183] When the above-mentioned function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0184] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for generating small-sample data of industrial processes by integrating domain knowledge, characterized in that, The method includes: Collecting the fault attribute descriptions corresponding to the industrial process data to obtain the first training data and the second training data; Constructing a data generation model to be trained that combines domain knowledge based on the network structure of the generative adversarial network; Using the first training data and the second training data to train the data generation model to be trained to obtain a data generation model; Obtaining the sample data to be generated; inputting the sample data to be generated into the data generation model for data generation to obtain the generated sample data.

2. The method for generating small-sample data of industrial processes by integrating domain knowledge according to claim 1, wherein The data volume of the first training data is greater than or equal to 1000 pieces; The data volume of the second training data is 10% of the data volume of the first training data.

3. The method for generating small-sample data of an industrial process combining domain knowledge according to claim 1, wherein, The data generation model to be trained includes a first generator, a second generator, a discriminator, and a fault attribute extractor.

4. The method for generating small-sample data of industrial processes integrating domain knowledge according to claim 3, wherein, The fault attribute extractor is composed of multiple classifiers; The classifier is constructed according to the attention module and the convolutional neural network; The classifier is obtained after being trained with the domain knowledge data of the industrial process; The number of the classifiers corresponds to the number of fault types of the industrial process.

5. The method for generating small-sample data of industrial processes combining domain knowledge according to claim 3, wherein The step of using the first training data and the second training data to train the data generation model to be trained to obtain a data generation model includes: Inputting the first training data into the fault attribute extractor to obtain the first fault attribute description; Based on the first fault attribute description, using the first training data, and performing data generation through the first generator to obtain the first generated data; Based on the discriminator loss function, calculating the loss function according to the first training data and the first generated data to obtain the discriminator loss; according to the discriminator loss, performing reverse parameter optimization on the discriminator to obtain an optimized discriminator; Based on the first generator loss function, calculating the loss function according to the first training data, the first generated data, and the first fault attribute description to obtain the first generator loss; according to the first generator loss, performing reverse parameter optimization on the first generator to obtain an optimized first generator; Inputting the second training data into the fault attribute extractor to obtain the second fault attribute description; Assigning the model parameters of the first optimized generator to the second generator to obtain a preliminarily optimized second generator; Based on the second fault attribute description, using the second training data, and performing data generation through the preliminarily optimized second generator to obtain the second generated data; Based on the second generator loss function, calculating the loss function according to the second training data, the second generated data, and the second fault attribute description to obtain the second generator loss; according to the second generator loss, performing reverse parameter optimization on the preliminarily optimized second generator to obtain an intermediate optimized second generator; Based on the soft parameter sharing method with dynamic gradient adjustment, adjusting the parameter weights of the intermediate optimized second generator according to the first generator loss and the second generator loss to obtain an optimized second generator; Obtaining a data generation model according to the optimized discriminator, the optimized first generator, the optimized second generator, and the fault attribute extractor.

6. The method for generating small-sample data of industrial processes integrating domain knowledge according to claim 5, characterized in that The step of, based on the soft parameter sharing method with dynamic gradient adjustment, adjusting the parameter weights of the intermediate optimized second generator according to the first generator loss and the second generator loss to obtain an optimized second generator includes: Perform gradient calculation based on the first generator loss and the second generator loss to obtain the shared parameter gradient; Conduct an adaptability analysis on the shared parameter gradient to obtain the adaptability analysis result; Based on the adaptability analysis result, dynamically adjust the weight of the shared parameter gradient to obtain the adjusted shared parameter gradient; Based on the adjusted shared parameter gradient, perform reverse parameter optimization on the intermediate optimized second generator to obtain the optimized second generator.

7. An industrial process small-sample data generation device integrating domain knowledge, the industrial process small-sample data generation device integrating domain knowledge is used to implement the industrial process small-sample data generation method integrating domain knowledge according to any one of claims 1-6, characterized in that The device includes: A data acquisition module, configured to collect fault attribute descriptions corresponding to industrial process data to obtain first training data and second training data; A model construction module, configured to construct a to-be-trained data generation model combining domain knowledge based on the network structure of the generative adversarial network; A model training module, configured to use the first training data and the second training data to train the to-be-trained data generation model to obtain a data generation model; A data generation module, configured to obtain sample data to be generated; input the sample data to be generated into the data generation model for data generation to obtain generated sample data.

8. The small-sample data generation device for industrial processes incorporating domain knowledge according to claim 7, characterized in that The model training module is further configured to: Input the first training data into the fault attribute extractor to obtain the first fault attribute description; Based on the first fault attribute description, use the first training data to perform data generation through the first generator to obtain first generated data; Based on the discriminator loss function, calculate the loss function according to the first training data and the first generated data to obtain the discriminator loss; According to the discriminator loss, perform reverse parameter optimization on the discriminator to obtain the optimized discriminator; Based on the first generator loss function, calculate the loss function according to the first training data, the first generated data, and the first fault attribute description to obtain the first generator loss; According to the first generator loss, perform reverse parameter optimization on the first generator to obtain the optimized first generator; Input the second training data into the fault attribute extractor to obtain the second fault attribute description; Assign the model parameters of the first optimized generator to the second generator to obtain the preliminarily optimized second generator; Based on the second fault attribute description, use the second training data to perform data generation through the preliminarily optimized second generator to obtain second generated data; Based on the second generator loss function, calculate the loss function according to the second training data, the second generated data, and the second fault attribute description to obtain the second generator loss; According to the second generator loss, perform reverse parameter optimization on the preliminarily optimized second generator to obtain the intermediate optimized second generator; Based on the soft parameter sharing method with dynamic gradient adjustment, adjust the parameter weights of the intermediate optimized second generator according to the first generator loss and the second generator loss to obtain the optimized second generator; Obtain the data generation model according to the optimized discriminator, the optimized first generator, the optimized second generator, and the fault attribute extractor.

9. An industrial process small-sample data generation device, characterized in that, The industrial process small sample data generation device includes: A processor; A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code, and the program code can be called by a processor to execute the method according to any one of claims 1 to 6.

Images