Industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network

Abstract

The invention discloses an industrial fault enhanced classification method based on a Gaussian discriminant generative adversarial network, which is used for supplementing relatively scarce fault type data in an industrial production process so as to improve the precision of industrial fault classification. According to the method, a Gaussian discriminant analysis model is used as a discriminator of the generative adversarial network, the authenticity of a generated sample is discriminated by using Gaussian discriminant analysis, the probability that the sample is true is output, the model is called as the Gaussian discriminant generative adversarial network, and the probability that the sample is true is obtained by training the Gaussian discriminant generative adversarial network. According to the method, a synthetic sample similar to a minority class of fault samples can be generated, the synthetic sample is fused with an original scarce fault data sample to generate a new balance fault data set of a data enhancement mode, and the new data set is used for training an industrial fault classifier, so that the precision of industrial fault classification in practical application is improved. According to the method, the problem of unbalanced fault data samples caused by scarcity of partial types of fault samples in industrial fault classification can be effectively solved, the precision of industrial fault classification is improved, and an important basis is provided for real-time fault identification and diagnosis.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault monitoring and diagnosis, and relates to an industrial fault enhanced classification method based on a Gaussian discriminant generative adversarial network. Background Art

[0002] As data in industrial processes increases and becomes easier to obtain, data-driven technologies have gradually become the mainstream technology in the field of process monitoring. Fault classification is an indispensable part of complex industrial process monitoring. It provides an important basis for fault diagnosis and recovery in industrial production processes and is an important technical support for the health assurance and maintenance of industrial systems and equipment. In recent years, more and more statistical machine learning and deep learning technologies have been introduced into the classification application of industrial fault data.

[0003] However, in actual industrial applications, the amount of data on fault types is often much smaller than the amount of data generated by normal process operation. Even for different similar fault modes, the amount of data accumulated in the industrial process is often quite different. This serious imbalance between different types of data will cause the performance of commonly used machine learning classification algorithms to drop significantly. In response to this phenomenon, the present invention proposes an industrial fault enhanced classification method based on a Gaussian discriminant generative adversarial network, which is used to effectively supplement the scarce fault type data and provide a relatively balanced fault data training sample set, thereby improving the accuracy of the industrial fault classification model. Summary of the invention

[0004] In view of the serious imbalance problem of fault type data in the current industrial production process, the amount of fault data of one or several categories is much larger than that of other categories, which makes the classification performance of most of the original fault classification algorithms plummet. The present invention proposes an industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network. The method effectively combines Gaussian discriminant analysis with generative adversarial network. By replacing the discriminator of the original generative adversarial network with Gaussian discriminant analysis, Gaussian discriminant analysis is used as the discriminator to judge the authenticity of samples, thereby realizing data enhancement of minority class fault sample sets in the industrial process, forming a relatively balanced fault training sample set, and thus improving the accuracy of the industrial fault classification model.

[0005] The specific distribution form of Gaussian discriminant analysis is as follows:

[0006] p(y)=φ y (1-φ) 1-y

[0007]

[0008] The estimates of the parameters in the model are obtained by maximizing the log-likelihood function:

[0009]

[0010]

[0011] Input the sample with feature x into the model and use the Bayesian formula The probability that the sample belongs to each category is obtained, and the output formula can be expressed as

[0012]

[0013] The generator of the model is composed of a multi-layer perceptron network. The parameters are updated through back-propagation and the stochastic gradient descent algorithm. The gradient of the generator is calculated based on the output of the samples generated by the generator in the discriminator to guide the model training.

[0014] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network is implemented.

[0015] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the industrial fault enhanced classification method based on a Gaussian discriminant generative adversarial network.

[0016] Beneficial effects of the present invention:

[0017] The present invention can effectively solve the problem of data imbalance in industrial fault classification. By proposing an industrial fault enhancement classification method based on Gaussian discriminant generative adversarial network, Gaussian discriminant analysis and generative adversarial network are effectively combined. By replacing the discriminator in the original generative adversarial network with Gaussian discriminant analysis, Gaussian discriminant analysis is used as a discriminator to discriminate the authenticity of samples, thereby realizing data enhancement of minority class fault sample sets in industrial processes, forming a relatively balanced fault training sample set, thereby improving the performance of industrial fault classification models. This method combines the generalization performance of the generative adversarial network model, and at the same time, according to the characteristics of industrial process fault data, it uses the Gaussian discriminant method to improve it, so that it has better performance in industrial fault data generation. In addition, this method alleviates the problems of gradient vanishing and difficulty in training convergence of the generative adversarial network model to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a Tennessee-Eastman industrial process flow chart. DETAILED DESCRIPTION

[0019] The industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network is further described in detail below in combination with specific implementation methods.

[0020] Embodiment: Industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network, the specific implementation method is as follows:

[0021] Step 1: Collect historical industrial process fault data samples as training data sets. Each type of fault sample can be expressed as Where M is the number of fault samples of this type.

[0022] Step 2: Normalize the fault sample data set D0 collected in step 1 so that the samples in the data set Each variable of is in the range of [0,1]; calculate the mean of the sample, and for the normalized sample, calculate M samples The mean of each variable in , we get a mean vector μ0.

[0023] Step 3: Combine the normalized dataset D0 and the n small batches of noise samples {z (1) ,…,z (n)} are respectively input into the generator and discriminator of the Gaussian discriminant generative adversarial network, and the corresponding discriminator output is obtained for each input sample;

[0024] Step 4: Update the parameters of the discriminator. (1) ,…,z (n)} Input to the network to generate the output Use a sliding window to intercept the synthetic samples generated by the latest N generators and compare them with the real samples in all M training datasets The parameters of the discriminator are updated jointly by the maximum likelihood estimation method.

[0025] Step 5: Update the parameters of the generator. (1) ,…,z (n)} Input to the generator to get the output These synthetic samples are input into the discriminator to obtain the output And use the output of the discriminator to update the parameters of the generator through the stochastic gradient descent method.

[0026] Step 6: Repeat steps 3 to 5 until the model converges, obtain the optimal parameters of the Gaussian discriminant generative adversarial network, establish an industrial fault enhanced classification model based on the Gaussian discriminant generative adversarial network, and save the model parameters.

[0027] Step 7: Generate enhanced samples for the minority class. Sample noise from the same random Gaussian distribution and input it into the generator that has converged in training to obtain synthetic samples similar to the scarce fault samples, so that the number of minority class samples is equal to that of the majority class, and the data set is balanced. Using the enhanced balanced data set for fault classification can improve the accuracy of the classifier to achieve more accurate fault classification.

[0028] The following is a specific example of the Tennessee-Eastman industrial process to illustrate the performance of the industrial fault enhanced classification method based on the Gaussian discriminant generative adversarial network. The Tennessee-Eastman industrial process is a standard experimental platform commonly used in the field of fault monitoring, diagnosis and classification. Its process flow is as follows Figure 1 As shown in Figure 1, the process consists of five operating units, including a gas-liquid separation tower, a continuous stirred reactor, a condenser, a centrifugal compressor, and a reboiler. The entire production process includes a total of 41 routinely measured variables and 21 different failure modes, and the detailed information is shown in Tables 1 and 2.

[0029] In the training set, there are 2000 normal samples and 21 fault samples, and the minority class accounts for 1% of the majority class. In the test set, we selected 200 normal samples and 100 fault samples. The experimental results show that compared with the currently commonly used industrial fault classification method, the industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network proposed in this invention has achieved higher fault classification accuracy. Compared with the benchmark nearest neighbor classification method, the fault classification accuracy of the method of this invention has been improved by more than 20%.

[0030] The following verifies the parallel computing efficiency of the present invention under large data sets. A multi-fault Gaussian discriminant generative adversarial network model is trained and run in parallel on a workstation with 4 Intel(R) Xeon(R) CPU E5-2620 v4@2.10GHz CPUs and 8 GTX 1080TI GPUs. For comparison, only one CPU of the same model is used to train the same model. There are 32 CPU cores in the workstation. The Gaussian discriminant generative adversarial network model corresponding to each fault is trained in a separate core. Combined with the parallel computing capability of the GPU, the model training speed is greatly improved.

[0031] In the speed test, all 21 types of fault samples, 5,000 fault samples, a dimension of 4, and 100,000 batches were used for training. In terms of model training speed, parallel computing training of 21 Gaussian discriminant generative adversarial network models takes a total of 169.8 seconds. In contrast, in non-parallel computing, the average time to train a model is 142.3 seconds, and the time required to train 21 Gaussian discriminant generative adversarial network models is approximately 2,988 seconds. It can be seen that the use of parallel computing with both CPU and GPU has greatly improved training efficiency.

[0032] Table 1: Tennessee-Eastman process commonly measured variables

[0033]

[0034]

[0035] Table 2: Description of Tennessee-Eastman process failure types

[0036]

[0037] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. An industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network, characterized in that: The steps include: Step 1: Collect historical industrial process fault data samples as training data sets. Each type of fault sample is represented as Among them, M is the number of fault samples of this type, Step 2: Normalize the fault sample data set D0 collected in step 1 so that the samples in the data set Each variable is in the range of [0,1], calculate the mean of the sample, and for the normalized sample, calculate M samples The mean of each variable in , we get a mean vector μ0, Step 3: Combine the normalized dataset D0 and the n small batches of noise samples {z (1) ,…,z (n) } are respectively input into the generator and discriminator of the Gaussian discriminant generative adversarial network, and the corresponding discriminator output is obtained for each input sample; Step 4: Update the parameters of the discriminator and convert the sampled {z (1) ,…,z (n) } Input to the network to generate the output Use a sliding window to intercept the synthetic samples generated by the latest N generators and compare them with the real samples in all M training datasets Update the parameters of the discriminator together through the maximum likelihood estimation method, Step 5: Update the parameters of the generator and convert the sampled {z (1) ,…,z (n) } Input to the generator to get the output Input these synthetic samples into the discriminator to get the output And use the output of the discriminator to update the parameters of the generator through the stochastic gradient descent method. Step 6: Repeat steps 3 to 5 until the model converges, obtain the optimal parameters of the Gaussian discriminant generative adversarial network, establish an industrial fault enhanced classification model based on the Gaussian discriminant generative adversarial network, and save the model parameters. Step 7: Generate enhanced samples for the minority class. Sample noise from the same random Gaussian distribution and input it into the generator that has converged in training to obtain synthetic samples similar to the scarce fault samples. This makes the number of minority class samples equal to that of the majority class, and the data set is balanced. Use the enhanced balanced data set for fault classification to improve the accuracy of the classifier and achieve more accurate fault classification.

2. The industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network according to claim 1 is characterized in that: The step three specifically includes the following steps: Constructing the generator: The generator is based on the MLP network structure, which consists of a first hidden layer, a Relu activation function, and a second hidden layer connected in sequence. The weight matrix and bias vector of the first hidden layer and the second hidden layer are W1, b1, W2, b2 respectively. These network parameters are represented as θ = {W1, b1, W2, b2}. Constructing the discriminator: The discriminator is based on Gaussian discriminant analysis, uses the same covariance assumption, and contains two Gaussian distribution assumptions. The parameters of the two Gaussian distributions are μ0, μ1, and Σ, respectively. Assuming that the class probability distribution follows the Bernoulli distribution, the parameter is φ. The parameters of the discriminator are expressed as The discriminator model is described by the following formula, y~Bernouli(φ) x|y=0~N(μ0,Σ) x|y=1~N(μ1,Σ) By Bayesian formula The output model input sample belongs to each category probability, realizing binary classification.

3. The industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network according to claim 1 is characterized in that: The step 4 specifically includes the following steps: The discriminator updates parameters through two different distributions of data One is the real sample, expressed as The other is that the generator generates simulated samples that are close to the real data distribution, expressed as All M real samples are used to update the parameters of the discriminator, expressed as And the N simulation samples newly generated by the sliding window sampling generator are used for parameter update, which is expressed as By assuming that the class prior probabilities are equal, the parameter φ does not need to be estimated; the expectation μ0 of the real data is obtained by estimating the mean of the real data once in the initial stage of training and does not change with training; the expectation μ1 of the simulated sample is obtained by using Updated by the following maximum likelihood estimation formula, The covariance matrix Σ is used and Updated by the following maximum likelihood estimation formula, 4. The industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network according to claim 1 is characterized in that: The step five is as follows: The discriminator outputs the formula Get generated samples Output value in the discriminator By chain conduction rule, we get Update the generator parameters θ using stochastic gradient descent.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network as described in any one of claims 1 to 4 above is implemented.

6. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instruction is executed by the processor, the industrial fault enhanced classification method based on Gaussian discriminant generative adversarial network as described in any one of claims 1 to 4 is implemented.

Images