Industrial process small sample fault data enhancement method based on meta-adversarial generative model

By using meta-adversarial generative models and conditional generative adversarial networks in industrial processes, combining diversity loss and meta-learning frameworks, the diversity problem of data generation in small sample fault diagnosis is solved, and effective failures in small sample scenarios are achieved Data enhancement and diagnostic effects.

CN119046678AInactive Publication Date: 2024-11-29CHINA JILIANG UNIV
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202411001126.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art is difficult to effectively generate diverse fault data in small sample fault diagnosis in industrial processes, and the traditional transfer learning-based method is not ideal in improving the effectiveness of generative adversarial networks, resulting in pattern collapse and single sample generation.

Method used

Using a method based on meta-adversarial generative model, combining conditional generative adversarial network (CGAN) and model-independent meta-learning framework (MAML), optimize the initial parameters of the generator and discriminator by introducing conditional variables and diversity loss functions to avoid patterns Crash and generate a wealth of varied failure data.

Benefits of technology

Effectively enhance the fault data in small sample scenarios, improve the diversity and generation efficiency of fault data, avoid the problem of pattern collapse, and enhance the model's adaptability and generalization ability on new tasks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046678A_ABST
    Figure CN119046678A_ABST
Patent Text Reader

Abstract

In the practice of equipment fault diagnosis in a large-scale industrial process, the data volume of part of fault samples is insufficient, so that the effect of a fault diagnosis model on small sample fault diagnosis is poorer. Fault samples generated by an existing data enhancement method are single, and the improvement effect is limited. Therefore, the invention provides an industrial process small sample fault data enhancement method based on a meta-adversarial generative model. The method comprises the following steps: firstly, preprocessing a data sample of an industrial process fault; secondly, preliminarily constructing a conditional generative adversarial network (CGAN); secondly, iteratively updating the initial parameters of the CGAN to obtain the optimal initial parameters based on a model independent meta learning (MAML) framework and by taking the maximization of the sensitivity of the initial parameters to a CGAN loss function as a target by utilizing various fault data with a relatively large sample size; further training the CGAN under the optimal initial parameter by using the data sample of the small sample fault; finally, a generator in the trained CGAN model is used for expanding a data set of small sample faults, and data enhancement is achieved. According to the method, the generative adversarial network and meta-learning are combined, the historical data of the small sample faults are effectively enhanced, and the diagnosis capability of the fault diagnosis model on the small sample faults is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an industrial process small sample fault data enhancement method based on a meta-adversarial generation model, and specifically relates to the field of industrial process fault diagnosis in a small sample fault data scenario. Background Art

[0002] In the operation and maintenance of large-scale industrial processes, fault diagnosis is crucial. Existing methods mainly rely on data-driven fault diagnosis models. The effectiveness of such methods depends largely on sufficient historical fault data. However, in practical applications, the amount of historical data for some faults is often insufficient to support effective diagnosis of the model. This is mainly due to the following reasons. First, the probability of occurrence of different types of faults varies, and the probability of occurrence of some faults themselves is very low. Second, in order to prevent these low-probability faults, enterprises have formulated corresponding maintenance plans, which further reduces the frequency of occurrence of these faults, making their samples more scarce. In addition, it is a slow process for a machine to gradually degenerate from a healthy state to a faulty state, which usually requires a long time and high experimental costs to obtain enough fault samples. Therefore, industrial process diagnosis often needs to cope with the challenge of scarce historical fault data for some faults. Taking the industrial production process of a factory producing aramid fiber as an example, there are currently 21 known faults, of which faults 16-20 have an average sample size of only 15.53% of the average sample size of the remaining types of faults due to their small number of occurrences. These faults are classified as small sample faults.

[0003] In order to deal with the above problems, the existing technology has proposed two types of solutions. The first type aims to improve the learning ability of the model itself, including the introduction of more complex model structures and more effective feature extraction technologies. However, this type of method assumes that all target faults are small sample types, while in actual applications, target faults often come from a mixed space of a small number of and sufficient sample fault types. Therefore, they are easily affected by the tendency of large categories and have poor recognition effects on small sample fault types.

[0004] Another method is to enhance the historical data of small sample faults. Among them, the method based on the mechanism model is often too complex and difficult to implement in practice. Although the method based on the generative adversarial network has achieved success in fields such as image generation, it is still essentially a neural network structure. It is prone to mode collapse when the amount of training data is insufficient, and the generated fault samples are single. To this end, attempts have been made to combine the training of the generative adversarial network with the small sample learning method. The current mainstream method is to train the model on a large number of similar samples, then migrate to the target sample and retain the common feature knowledge and update the differentiated knowledge by slightly adjusting the parameters to achieve the training of the generative adversarial network under small samples. However, in the same industrial production system, even similar faults may have large differences in their data patterns. This difference leads to a reduction in the knowledge that the model can learn on similar faults and can be used for small sample learning, which in turn affects the migration effect, so that the generative adversarial network still faces the problem of mode collapse. Therefore, it is urgent to improve the existing fault diagnosis method so that it can effectively generate corresponding fault data for faults with a small sample size, so as to achieve effective diagnosis of industrial process faults in a few sample scenarios. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] Currently, the application of few-shot fault diagnosis technology in industrial process fault diagnosis mainly focuses on data enhancement methods. However, the mechanism models in these methods are often too complex to be applied in practice. Although generative adversarial networks are a viable option, they are prone to mode collapse due to insufficient training data, resulting in single generated fault samples and lack of diversity. At the same time, traditional transfer learning-based methods are not ideal in improving the effect of generative adversarial networks in industrial process fault data enhancement. Therefore, there is an urgent need for an innovative few-shot fault diagnosis method that can effectively achieve data enhancement with limited sample size.

[0007] (II) Technical solution

[0008] In order to solve the above technical problems, the present invention proposes a method for enhancing industrial process small sample fault data based on a meta-adversarial generation model. The method comprises the following steps:

[0009] S1. Collect and organize existing historical fault data sets, preprocess samples, and label them with fault types.

[0010] Furthermore, the standardization processing in S1 includes: data standardization and data noise removal, and the calculation formula is:

[0011]

[0012]

[0013]

[0014] In the formula, N is the number of samples, X is i is the i-th sample feature, μ is the feature mean, σ is the input standard deviation, X std is the standardized feature value.

[0015] S2. Select the conditional generative adversarial network (CGAN) as the generative network infrastructure, which includes two substructures: a generator with auxiliary conditions and a discriminator.

[0016] Preferably, in step S2, the traditional generative adversarial network usually trains an independent generative adversarial network for each fault type dataset. In the MAML pre-training process, fault datasets of different fault types will be used. If the traditional generative adversarial network model structure is used, the training datasets will be mixed and not distinguished during the training process, which will cause the optimal initial parameters obtained by MAML to contain confusing information, which is not conducive to the subsequent training of the CGAN model. In response to the above problems, conditional variables are introduced into the generative adversarial network so that the model selects a specific fault type and generates corresponding fault data in a directionally manner, thereby improving the generation efficiency of small sample fault data.

[0017] Specifically, step S2 is as follows:

[0018] S201. The CGAN is composed of a fault sample generator G and a fault sample discriminator D, both of which are four-layer fully connected neural network architectures. The input layer of the generator G includes a+1 neurons, of which a is used to input random noise and 1 is used to input the fault type of the sample to be generated. The output layer of the generator includes m neurons, which is consistent with the number of features of the fault sample. The input layer of the discriminator D adds a neuron to the input layer of the generator for inputting the fault type, and the output layer is a neuron, which outputs the probability that the input sample is a true sample.

[0019] S202, initialize the neural network parameters θ of the generator G and the discriminator D G and θ D .

[0020] S203: Input the noise data z and the fault label y to be generated into the generator G to generate a fault sample.

[0021] S204: Mix the generated fault samples with the real samples, input them into the discriminator, and output the discrimination result that each sample is a real sample.

[0022] S205, calculate the loss function of the generator G and the discriminator D, and use the back propagation algorithm to adjust the parameters θ G and θ D to update.

[0023] Preferably, the conditional generative adversarial network is characterized in that a sample diversity loss is introduced in the loss function calculation of the generator G, and by measuring the distance of the generated samples in the feature space, the generator is encouraged to generate a sample set with more dispersed features during the training process, thereby covering more potential data distributions. This method effectively avoids the problem of mode collapse and ensures that the generator will not generate only a limited number of sample types, but can generate richer and more diverse samples.

[0024] The loss function of the generator is expressed as:

[0025]

[0026] is the original adversarial loss, expressed as:

[0027]

[0028] is the newly introduced diversity loss, expressed as:

[0029]

[0030] in:

[0031] is the generated sample G(z i |y) feature vector;

[0032] N is the number of samples generated.

[0033] The calculation formula of the discriminator D loss function is:

[0034]

[0035] S206. Repeat S203 to S205 until the loss functions of the two converge. At this time, the CGAN has been initially trained and constructed.

[0036] S3. In the CGAN training process, a model-agnostic meta-learning framework (MAML) is introduced to train the initial optimal parameters of CGAN.

[0037] Preferably, in step S3, it is characterized in that the method integrates the MAML meta-learning framework into the training process of the generative adversarial network (GAN). In traditional GAN ​​training, both the generator and the discriminator are neural network models, and their initial parameters are usually randomly generated. When the amount of training data is small, the generator and the discriminator may converge prematurely after a small number of iterations, causing the generator to repeatedly generate similar samples to deceive the discriminator, the training process to fall into mode collapse, and the samples finally generated lack diversity.

[0038] To address this problem, the MAML meta-learning framework is introduced to find initial parameters for CGAN that are highly sensitive to the loss function, so that CGAN can significantly change the loss function by fine-tuning the parameters of the generator and discriminator in the case of small samples, thereby improving the efficiency of sample utilization and avoiding mode collapse in the training process.

[0039] In the MAML framework, we use historical data of multiple sufficient sample failures in the same industrial process to construct a meta-training task that can simulate the generation of real small sample data to simulate the training scenario of small samples. By maximizing the sensitivity of the initial parameters to the loss function, we repeatedly iteratively adjust the initial parameters of the generator and discriminator to finally obtain the optimal initial parameters.

[0040] Specifically, step S3 is as follows:

[0041] S301, mixing multiple historical data sets with sufficient number of faults in S1 to form a meta-training set.

[0042] S302, build the generator G and discriminator D based on the CGAN framework of S2, and randomly generate the initial network structure parameters θ G ,θ D As a meta-model, its parameter θ G ,θ D are the initial parameters of each base model in the inner loop.

[0043] S303, construct l task data sets with replacement from the data in the meta-training set to form a task pool. Each task data set contains N fault types, each fault type contains K samples, and is divided into a support set and a query set at a ratio of 8:2.

[0044] Preferably, the number of fault types N in the meta-task is less than the total number of fault types contained in the meta-training set, and the number of samples N of each fault type is less than the number of samples of the minimum sample size fault. By limiting the number of categories and samples for each task, the model is forced to learn on a small amount of data, thereby improving the model's ability to quickly adapt and generalize on new tasks, while avoiding overfitting and reducing computing resource consumption.

[0045] S304. Create k base conditional generative adversarial networks CGANT i (referred to as CGANT i ), whose model parameters inherit the initial network structure parameters of the meta-model as θ G ,θ D Then, for each base CGANT, i Assign a task i.

[0046] S305, based on CGANT i Based on the support set T of task i i s Adjust its model parameters and calculate its i q on losses.

[0047] S3051, calculation base CGANT i The support set T of task i i s The loss on . Generator Randomly generate fault samples based on the fault types contained in task i and mix them with the data in the support set and input them into the discriminator In the process, the judgment result is obtained and the losses of the two are calculated based on the judgment result.

[0048] Generator The loss function calculation formula is:

[0049]

[0050] Discriminator The loss function calculation formula is:

[0051]

[0052] S3052, according to the generator loss and the discriminator loss The generator is trained using the back-propagation algorithm and the discriminator The parameters of the two parameters after updating are expressed as

[0053] The update process is expressed as:

[0054]

[0055] where α G and α D is the updated learning rate for the generator and the discriminator.

[0056] S3053, calculation base CGANTi Updated parameters The query set in task i The performance effect on.

[0057] Generator The loss function is:

[0058]

[0059] Discriminator The loss function calculation formula is:

[0060]

[0061] S306. Evaluate the overall performance of the generated model under each task

[0062] Preferably, the CGANT base is calculated for each task i i Generator With the discriminator In the query set The sum of the losses on the above is used as the evaluation indicator.

[0063] The generator loss is:

[0064]

[0065] The discriminator loss is:

[0066]

[0067] Due to the CGANT i Updated parameters Only in task T i The optimal fault generation effect on samples does not mean that they are suitable for the generation of all fault samples. In order to avoid overfitting the model to the data of a specific task, we adopt an additive evaluation strategy to capture the diversity between different tasks. This is to ensure that the initial parameters provided by the meta-model can perform well in the training process of various tasks, while providing more comprehensive gradient information. In this way, the model will take into account the characteristics of various tasks when updating, thereby improving the generalization and adaptability of the meta-model.

[0068] S307: Update the initial parameters θ of the network structure of the meta-model based on the evaluation results. G ,θ D , and its update process is:

[0069]

[0070] Among them, β G and βD is the learning rate.

[0071] Preferably, the model learning rate α of the inner loop G ,α D Much smaller than the model learning rate β of the outer loop G and β D , a smaller learning rate can ensure that the parameter update in the inner loop is smoother and more gradual, thereby avoiding drastic changes and overfitting, maintaining the stability and generalization ability of training, and at the same time laying a good foundation for the large-stride optimization of the outer loop, making the overall parameter optimization more effective.

[0072] S308. Repeat S303-S308 until the maximum number of iterations is reached or the sum of the losses of each task converges.

[0073] S309. Output the optimal initialization parameters of the metamodel generator G and the discriminator D.

[0074] S4: Starting from the initial parameters of CGAN obtained by MAML training, it is characterized by using the historical data of the target fault.

[0075] Preferably, in step S4, the CGAN initial parameters obtained by MAML training do not have the ability to directly generate fault data. Therefore, it is necessary to further train CGAN using the target fault data set. Based on the CGAN training process, the network structure parameters θ of the generator G and the discriminator D are quickly updated. G ,θ D , enabling it to quickly learn the data patterns of faults and effectively generate diverse data patterns.

[0076] Specifically, step S4 is as follows:

[0077] S401, the network structure parameters θ of the optimal CGAN trained by MAML G ,θ D as initial parameters.

[0078] S402, using the data of the generated fault type as the training set, the CGAN of the network structure parameter θ based on the S2 step G ,θ D Further specific optimization.

[0079] S403: After the CGAN training for small sample failures is completed, its generator G and its parameters θ are output. G .

[0080] S5: After the generator training is completed, the small sample fault labels are input into the trained generator to generate these fault samples in a targeted manner and expand the sample space.

[0081] Specifically, step S5 is as follows:

[0082] S501. Randomly generate a set of noise based on Gaussian distribution and combine it with the label of small sample faults to form the initial input of the generator.

[0083] S502: Input the initial input data into the generator G that has completed the MAML-CGAN framework training.

[0084] S503. The data output by the generator G is combined with the training label to form a new fault sample, which is added to the data set.

[0085] S504, repeat S401-S403 until the number of data sets is expanded to meet the requirement. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a flow chart of a method for enhancing industrial process small sample fault data based on a meta-adversarial generation model according to an embodiment of the present application.

[0087] Figure 2 This is a technical example diagram of a conditional generative adversarial network according to one embodiment of the present application.

[0088] Figure 3 This is a training flowchart of a conditional generative adversarial network based on a model-independent meta-learning framework in one embodiment of the present application.

[0089] Figure 4 This is a technical example diagram of a conditional generative adversarial network based on a model-independent meta-learning framework in one embodiment of the present application.

[0090] Figure 5 This is a flowchart of using trained generator data to enhance a small sample fault sample space in one embodiment of the present application. DETAILED DESCRIPTION

[0091] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described examples are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0092] In the specific implementation method, the present invention takes the Tennessee Eastman process as a case. The Tennessee Eastman Process (TE process) is a complex industrial process, which includes five main units: reactor, condenser, compressor, separator and stripping tower, and eight components: A, B, C, D, E, F, G and H. The TE process involves operating variables and measurement variables, totaling 52 observed variables. In the TE process simulation, a total of 21 fault types are set, of which the last 5 are small sample fault types. This paper takes these small sample faults as practical scenarios to describe and verify the process and effect of this method in detail.

[0093] S1. Collect and organize the existing historical fault data sets, and label the samples with the fault type as the label. And standardize the data, including: data standardization and data noise removal, the calculation formula is:

[0094]

[0095]

[0096]

[0097] S2. Preliminary construction of CGAN, which includes two substructures: a generator with auxiliary conditions and a discriminator.

[0098] Specifically, step S2 is as follows:

[0099] S201. The CGAN is composed of a fault sample generator G and a fault sample discriminator D, both of which are four-layer fully connected neural network architectures. The input layer of the generator G includes 101 neurons, 100 of which are used to input random noise and 1 is used to input the fault type of the sample to be generated. The output layer of the generator includes 52 neurons, which is consistent with the number of features of the fault sample. The input layer of the discriminator D includes 53 neurons, 52 of which are observed variables of the fault, and the remaining 1 is a fault label. The output layer is a neuron that outputs the probability that the input sample is a true sample.

[0100] S202, initialize the neural network parameters θ of the generator G and the discriminator G and θ D .

[0101] S203 , randomly generate noise data z based on Gaussian distribution, and input it into the generator G together with the fault label y to be generated to generate a fault sample.

[0102] S204: Mix the generated fault samples with the real samples, input them into the discriminator, and output the discrimination result that each sample is a real sample.

[0103] S205, calculate the loss function of the generator G and the discriminator D, and use the back propagation algorithm to adjust the parameters θ G and θ D to update.

[0104] The calculation formula of the generator G loss function is:

[0105]

[0106]

[0107]

[0108] The calculation formula of the discriminator D loss function is:

[0109]

[0110] S206. Repeat S203 to S205 until the loss functions of the two converge. At this time, the CGAN has been initially trained and constructed.

[0111] S3. Pre-train CGAN under the MAML framework so that it can quickly learn the data characteristics under the fault mode through a small number of fault samples.

[0112] Specifically, step S3 is as follows:

[0113] S301, mixing multiple historical data sets with sufficient number of faults in S1 to form a meta-training set.

[0114] S302, build the generator G and discriminator D based on the CGAN framework of S2, and randomly generate the initial network structure parameters θ G ,θ D , using this as the meta-model.

[0115] S303, construct 30 task data sets with replacement from the data in the meta-training set to form a task pool. Each task data set contains 6 fault types, including a small sample fault type, and each fault type contains 10 samples, which are divided into a support set and a query set at a ratio of 8:2.

[0116] S304, create 10 base conditional generative adversarial networks CGANT i (referred to as CGANT i ), whose model parameters inherit the meta-model, and the network structure parameters are θ G ,θ D Then, for each base CGANT, i Assign a task u.

[0117] S305, based on CGANT i Support set based on task i Adjust its model parameters and calculate its on losses.

[0118] S3051, calculation base CGANT i The support set of task i The loss on . Generator Randomly generate fault samples based on the fault types contained in task i and mix them with the data in the support set and input them into the discriminator In the process, the judgment result is obtained and the losses of the two are calculated based on the judgment result.

[0119] Generator The loss function calculation formula is:

[0120]

[0121] Discriminator The loss function calculation formula is:

[0122]

[0123] S3052, according to the generator loss and the discriminator loss The generator is trained using the back-propagation algorithm and the discriminator The parameters of the two parameters after updating are expressed as

[0124] The update process is expressed as:

[0125]

[0126] where α G and α D is the updated learning rate for the generator and the discriminator.

[0127] S3053, Calculate the base CGANT i Parameters After the update, its query set in task i The performance effect on.

[0128] Generator The loss function is:

[0129]

[0130] Discriminator The loss function calculation formula is:

[0131]

[0132] S306, calculate the base CGANi of each task i in the query set base CGANT i Generator With the discriminator In the query set The sum of the losses is used as an evaluation indicator to evaluate the effect.

[0133] The generator loss is:

[0134]

[0135] The discriminator loss is:

[0136]

[0137] S307: Update the network structure parameter θ of the meta-model based on the evaluation result G ,θ D , and its update process is:

[0138]

[0139] Among them, β G With β D is the learning rate.

[0140] S308. Repeat S303-S308 until the maximum number of iterations is reached or the sum of the losses of each task converges.

[0141] S309. Output the optimal initialization parameters of the metamodel generator G and the discriminator D.

[0142] S4: Taking the initial optimal parameters of CGAN as the starting point, the CGAN is trained using a dataset of small sample failures.

[0143] S401, the network structure parameters θ of the optimal CGAN trained by MAML G ,θ D as initial parameters.

[0144] S402, using the data of the generated fault type as the training set, the CGAN of the network structure parameter θ based on the S2 step G ,θ D Further specific optimization.

[0145] S403: After the CGAN training for small sample failures is completed, its generator G and its parameters θ are output. G .

[0146] S5: After the generator training is completed, the small sample fault labels are input into the trained generator to generate these fault samples in a targeted manner and expand the sample space.

[0147] S501. Randomly generate a set of noise based on Gaussian distribution and combine it with the label of small sample faults to form the initial input of the generator.

[0148] S502: Input the initial input data into the generator G that has completed the MAML-CGAN framework training.

[0149] S503. The data output by the generator G is combined with the training label to form a new fault sample, which is added to the data set.

[0150] S504, repeat S401-S403 until the number of data sets is expanded to meet the requirement.

[0151] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0152] What is disclosed above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or modifications within the technical scope disclosed in the present invention, which should be covered within the protection scope of the present invention.

Claims

1. A method for enhancing small sample fault data of industrial processes based on a meta-adversarial generation model, characterized in that: In order to solve the problem of insufficient historical data of some fault samples in industrial processes, general methods use data augmentation. When training samples are limited, the data pattern generated by traditional generative adversarial networks in the process of data augmentation is too single, resulting in poor enhancement effect. To this end, this method combines the meta-learning framework and uses a variety of fault data with sufficient sample size in the same industrial process to preliminarily train the generative adversarial network to obtain the optimal initial parameter configuration and improve the model's learning ability when faced with a small amount of new data. Subsequently, the optimal initial parameters of the model are optimized in a small number of steps using only the historical data of the target fault, so that it can generate diversified and high-quality target fault data, thereby achieving effective data augmentation and further improving the accuracy of fault diagnosis. 1) Collect and organize existing historical fault data sets, preprocess samples, and label them with fault types. 2) The conditional generative adversarial network (CGAN) is selected as the generative network infrastructure, which includes two substructures: a generator with auxiliary conditions and a discriminator. 3) In order to solve the problem of mode collapse in generative adversarial networks due to insufficient training data, a model-agnostic meta-learning framework (MAML) is introduced to pre-train CGAN and obtain the initial optimal parameters of CGAN. Pre-training with the MAML framework improves the sample learning efficiency of the model and reduces CGAN's dependence on a large amount of labeled data. 4) Taking the initial optimal parameters of CGAN as the starting point, a dataset with small sample failures is used to perform characteristic training on CGAN. 5) After the generator training is completed, the fault labels are input into the trained generator to expand the historical data set.

2. According to claim 1, a method for enhancing industrial process small sample fault data based on a meta-adversarial generation model is characterized in that Step 2), the traditional generative adversarial network usually trains a separate generative adversarial network for each fault type dataset. In the meta-learning process of MAML, fault datasets of different fault types are used. If the traditional generative adversarial network model structure is used, the training datasets are mixed and not distinguished during the training process, which will cause the optimal initial parameters obtained by MAML to contain confusing information, which is not conducive to the training of the subsequent CGAN model. To address the above problems, this method introduces conditional variables into the generative adversarial network, so that the model selects a specific fault type and generates corresponding fault data in a targeted manner, thereby improving the efficiency of generating small sample fault data. The step 2) specifically includes the following steps: 21) The CGAN is composed of a fault sample generator G and a fault sample discriminator D, both of which are four-layer fully connected neural network architectures. The input layer of the generator G includes a+1 neurons, of which a is used to input random noise and 1 is used to input the fault type of the sample to be generated. The output layer of the generator includes m neurons, which is consistent with the number of features of the fault sample. The input layer of the discriminator D adds a neuron to the input layer of the generator for inputting the fault type, and the output layer is a neuron that outputs the probability that the input sample is a true sample. 22) Initialize the neural network parameters θ of the generator G and the discriminator D G and θ D . 23) Input the noise data z and the fault label y to be generated into the generator G to generate a fault sample. 24) The generated fault samples are mixed with the real samples, input into the discriminator, and the discrimination result of each sample as a real sample is output. 25) Calculate the loss function of the generator G and the discriminator D, and adjust their parameters θ according to the back propagation algorithm G and θ D to update. 26) Repeat 23) to 25) until the loss functions of the two converge. At this point, the CGAN has been initially trained and constructed.

3. The method for enhancing industrial process small sample fault data based on a meta-adversarial generative model according to claim 1 is characterized in that Step 3), this method integrates the MAML meta-learning framework into the training process of the generative adversarial network (GAN) to find the initial parameters of the CGAN that are highly sensitive to the loss function, so that in the case of small samples, the loss function can be significantly changed by fine-tuning the parameters of the generator and the discriminator, thereby improving the sample utilization efficiency and avoiding the training process from falling into mode collapse. In the MAML framework, we use historical data of multiple sufficient sample failures in the same industrial process to construct a meta-training task that can simulate the generation of real small sample data to simulate the training scenario of small samples. In this framework, training is divided into two parts: the outer loop and the inner loop. The inner loop uses meta-CGAN to train on a multi-task distribution. The outer loop integrates the training results of the inner loop and updates the meta-CGAN parameters accordingly. With the goal of maximizing the sensitivity of the initial parameters to the loss function, we repeatedly iteratively adjust the initial parameters of the generator and discriminator to finally obtain the optimal initial parameters. The step 3) specifically comprises the following steps: 31) Mix multiple historical data sets with sufficient number of faults to form a meta-training set. 32) Based on the CGAN framework of step 2), the generator G and the discriminator D are constructed, and the initial network structure parameters θ are randomly generated. G ,θ D As a meta-model, its parameter θ G ,θ D are the initial parameters of each base model in the inner loop. 33) The data in the meta-training set are replaced to construct l task data sets to form a task pool. Each task data set contains N fault types, each fault type contains K samples, and is divided into a support set and a query set at a ratio of 8:

2. 34) Create k base conditional generative adversarial networks CGANT i (referred to as CGANT i ), whose model parameters inherit the meta-model, and the network structure parameters are θ G ,θ D Then, for each base CGANT, i Assign a task i. 35) The training process enters the inner loop, based on CGANT i Support set based on task i Adjust its model parameters and calculate its on losses. 36) Evaluate the overall performance of the generated model under each task. 37) Update the network structure parameters θ of the meta-model based on the evaluation results G ,θ D . 38) Repeat 33)-38) until the maximum number of iterations is reached or the sum of the losses of each task converges. 39) Output the network structure parameters θ of the meta-model generator G and the discriminator D G ,θ D is the optimal initial parameter.

4. According to claim 1, it is characterized in that Step 4), based on the optimal initial parameters of CGAN, the historical data of small sample failures is used to quickly update the network structure parameters θ of the generator G and the discriminator D based on the CGAN training architecture G ,θ D , which enables CGAN to generate failure data with small sample failures. 41) The network structure parameters θ of the optimal CGAN trained by MAML G ,θ D as initial parameters. 42) Using the data of generated fault types as the training set, the network structure parameter θ is trained under the CGAN training framework. G ,θ D Further optimization. 43) After the CGAN training for small sample failures is completed, its generator G and its parameters θ are output G .

5. According to step 25) of claim 2, it is characterized in that: Sample diversity loss is introduced into the loss function calculation of the generator G. By measuring the distance of the generated samples in the feature space, the generator is encouraged to generate samples with more dispersed features during the training process, thereby covering more potential data distributions. This method can avoid the problem of mode collapse to a certain extent and ensure that the generator can produce richer and more diverse samples. The loss function of the generator is expressed as: is the original adversarial loss, expressed as: is the newly introduced diversity loss, expressed as: in: is the generated sample G(z i |y); N is the number of samples generated; ||·||2 represents the Euclidean spacing of the feature vector.

6. Step 35) according to claim 3, characterized in that: After entering the inner loop, based on CGANT i Use the support set of the current task i Update the model and in the queryset The purpose of the inner loop is to enable CGAN to quickly adapt to each specific task and lay the foundation for the overall parameter optimization in the outer loop. The step 35) specifically comprises the following steps: 351) Computational base CGANT i The support set of task i The loss on . Generator Randomly generate fault samples based on the fault types contained in task i and mix them with the data in the support set and input them into the discriminator In the process, the judgment result is obtained and the losses of the two are calculated based on the judgment result. Generator The loss function calculation formula is: Discriminator The loss function calculation formula is: 352) According to the generator loss and the discriminator loss The generator is trained using the back-propagation algorithm and the discriminator The parameters of the two parameters after updating are expressed as The update process is expressed as: 353) Calculate the base CGANT i Parameters After the update, its query set in task i The performance effect on.

7. Step 36) according to claim 3, characterized in that: Compared with other models trained with a single neural network and the combination of the MAML framework, CGAN uses two neural networks, the generator and the discriminator, for adversarial training, so the gradient matching mechanism of a single neural network cannot be used to calculate and optimize its loss. To this end, during the evaluation process, the loss of each subtask is summed up, and the performance of the generator and the discriminator on different tasks is comprehensively considered. By optimizing the sum of the losses of multiple tasks, the generator and the discriminator are ensured to perform evenly on diverse data, thereby improving the overall effect of the model. The effect evaluation calculation method is: The generator loss is: The discriminator loss is: 。 8. According to step 37) of claim 3, the metamodel parameter updating process is expressed as: It is characterized in that The model learning rate of the outer loop β G and β D Much larger than the model learning rate α of the inner loop G ,α D In the inner loop, a smaller learning rate makes the parameter update smoother and more gradual, thus avoiding drastic changes and overfitting, and maintaining the stability and generalization ability of training. In the outer loop, a larger learning rate ensures a large-step optimization of the parameters, making the overall parameter optimization more efficient, thereby improving the adaptability and performance of the model on new tasks.

Citation Information

Cited By

  • Application code development method and system based on artificial intelligence model

    CN119473222A

  • An application code development method and system based on artificial intelligence model

    CN119473222B

  • Training method and device of generative adversarial network and storage medium

    CN120455430A

  • Elevator predictive fault early diagnosis method based on data driving

    CN122072699A