Dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network

Abstract

The invention discloses a dynamic semi-supervised industrial fault classification method based on a deep convolutional generative adversarial network, and the method is based on the deep convolutional generative adversarial network, expands the deep convolutional generative adversarial network to a semi-supervised form, and employs the information in an unlabeled sample; on the basis, a sliding window technology is further used, data information on a time axis is introduced, a single sample of industrial fault data is expanded to a two-dimensional matrix form, and model training and testing are carried out; according to the method, the capability of capturing the data manifold of the generative adversarial network is utilized, the TangentProp method is adopted to take the captured manifold as a regularization item to limit the complexity of a classifier on a manifold subspace, and the local geometric information of the data on the manifold is utilized to improve the classification precision. In addition, technologies such as Batch Normalization, Dropout, Feature Matching and the like are used for improving the stability and the speed of the training. According to the method, the industrial fault classification precision can be effectively improved in a complex industrial data mode.

Description

Technical Field

[0001] The present invention belongs to the field of industrial process fault diagnosis and classification, and relates to a dynamic fault classification method based on semi-supervised learning and manifold learning. Background Art

[0002] Fault classification is an important part of industrial process monitoring, and machine learning methods have become a mainstream tool in industrial fault classification. However, most classification methods require a large number of labeled samples as support, and it usually takes a lot of time and expert knowledge to obtain labeled data in actual industrial processes. Correspondingly, there are sufficient unlabeled samples in actual industrial data. Moreover, most industrial processes are dynamic and can be regarded as time series, so the time correlation in sequential data is worth using for fault classification. In addition, according to the manifold hypothesis in semi-supervised learning, the basic manifold structure contains additional local geometric information of the training samples, which is also useful for semi-supervised classification. Based on the above three points, the present invention proposes a semi-supervised industrial fault classification method based on deep convolutional generative adversarial network, which makes full use of the information of unlabeled samples, and uses sliding windows to extract information on the time axis for time series data. In addition, in order to further improve the classification performance of the model, a manifold regularization term is introduced into the classifier, so that the classifier remains unchanged on the tangent of the data manifold, further improving the accuracy of industrial fault classification. Summary of the invention

[0003] In actual industrial processes, data labels are usually insufficient, which will greatly reduce the performance of existing classification methods. At the same time, many data in industrial processes have time correlation, and each sample is not independent of each other, but is interrelated on the time axis. Moreover, there is a high degree of coupling between different variables in high-dimensional process data. Furthermore, according to the manifold hypothesis in semi-supervised learning, the basic manifold structure of the data contains some information that is helpful for fault classification in industrial processes. The present invention proposes a dynamic semi-supervised industrial fault classification method based on a deep convolutional generative adversarial network combined with manifold regularization.

[0004] In order to introduce data information on the time axis in fault classification, the present invention uses the sliding window technology to expand a single sample of process data into the form of a two-dimensional matrix, and trains and tests the model with this two-dimensional sample. Correspondingly, a network with CNN as the model is used, that is, the model structure of the deep convolutional generative adversarial network. Similarly, in order to utilize the information in the unlabeled samples, a semi-supervised generative adversarial network is adopted on the basis of the deep convolutional generative adversarial network. To further utilize the information in the unlabeled samples, the invention also introduces a manifold regularization term based on TangentProp, which improves the classification accuracy by limiting the complexity of the classifier on the manifold subspace.

[0005] The semi-supervised generative adversarial network uses both labeled and unlabeled data, and modifies the discriminator into a multi-classification model. Specifically, a K+1-dimensional softmax layer is added to the output layer of the discriminator. When calculating the loss function of the discriminator, it is expected that the K-class labeled data will be divided into corresponding units, and the unlabeled data will be divided into any of the first K units. For the false samples generated by the generator, it is expected that the discriminator will divide them into the last K+1 units. The loss function of the discriminator can be expressed as:

[0006] L discriminator =L supervised +L unsupervised

[0007]

[0008] This allows the discriminator to act as a classifier while distinguishing true from false samples, and a large number of unlabeled samples can help the generator more accurately approximate the probability distribution of the original data. Conversely, the realistic samples generated by the generator can also help improve the accuracy of classification.

[0009] The deep convolutional generative adversarial network applies the CNN network to the model, removes all fully connected hidden layers, and replaces them with deeper convolutional layers and deconvolutional layers. The training process of the deep convolutional generative adversarial network is more stable, and the data representation learned by the model is more accurate.

[0010] Considering the time series of data, the present invention uses a sliding window method to pre-process the data. First, all process data are spliced ​​together along the time T axis in time series to form an N*T data matrix, where N is the number of variables of process data and T is the number of samples. On the basis of this matrix, an N*L sliding window is introduced, where L is the window size of the sliding window. Through this sliding window, the data matrix is ​​cut along the T axis according to a certain step size S to obtain (TL) / S matrix samples of size N*L. By dividing these matrix samples, a test set and a training set can be obtained, and the model can be trained offline. Similarly, in online diagnosis, we should also use a sliding window of the same size to intercept data, and input matrix samples of the same size into the model to obtain online fault diagnosis results.

[0011] In order to model the preprocessed data, the present invention uses a model structure based on a deep convolutional neural network and expands the discriminator to a semi-supervised form so that it can utilize the information in unlabeled samples.

[0012] In terms of model structure, the generator of the present invention uses a multi-layer deconvolution network. Through these deconvolution layers, the generator upsamples a randomly sampled noise vector z to a sample x of the same size as the real sample. When the training reaches convergence, the generator can learn an accurate mapping G:Z→X from the latent variable space to the real sample space. During the training process, the samples generated by the generator will help the discriminator improve the classification accuracy of real data.

[0013] The discriminator in the present invention is composed of a convolutional network. In a semi-supervised generative adversarial network, three types of samples, namely, labeled real samples, unlabeled real samples, and false samples generated by the generator, are input into the discriminator in turn. The discriminator downsamples the input two-dimensional samples layer by layer through multiple convolutional layers to a K+1-dimensional vector corresponding to the number of label categories in the real samples. Each unit in the K-dimensional vector corresponds to a different category, and the last unit represents the false sample generated by the generator. When the training reaches convergence, the discriminator can accurately depict the boundary probability distribution of the real data. At the same time, the discriminator in the semi-supervised form can also be used as a classifier to classify samples of industrial failures.

[0014] In terms of loss function, the semi-supervised discriminator loss function is as described above, and the generator loss function uses the Feature Matching item, which makes the training process more stable and converges faster.

[0015] A manifold regularization term is added to the loss function of the classifier to make the classifier remain unchanged on the tangent space of the manifold. Specifically, from the perspective of the manifold, the generator G:Z→X is a parameterized mapping from a low-dimensional space Z to a manifold M embedded in a high-dimensional space, where the input z∈Z of the generator is the global coordinate of the manifold, and the output G(z) of the generator is a sample on the manifold. Since ordinary generative adversarial networks cannot directly obtain this type of inverse mapping, the structure of a bidirectional generative adversarial network is introduced, and an encoder h:X→Z is added to the original model and trained together with the generator and discriminator. Then the Jacobian matrix of the encoder is obtained to obtain the tangent space T of the real data on the manifold. x =J x h. Finally, the TangentProp method is used to make the output of the classifier invariant in the tangent space of the manifold. The manifold regularization term is Where c is the output of the classifier, J x c is the Jacobian matrix of the classifier, and the loss function of the semi-supervised deep convolutional network with manifold regularization is expressed as

[0016] L discriminator =L supervised +L unsupervised +Ω manifold

[0017]

[0018] Among them, f l Represents the output of a hidden layer of the discriminator.

[0019] An electronic device comprises 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 dynamic semi-supervised industrial fault classification method based on a deep convolutional generative adversarial network is implemented.

[0020] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the dynamic semi-supervised industrial fault classification method based on a deep convolutional generative adversarial network.

[0021] Compared with the prior art, the advantages of the present invention are as follows:

[0022] Based on the historical industrial fault data sample set, the present invention proposes a semi-supervised industrial fault classification method based on deep convolutional generative adversarial network on the basis of data preprocessing and semi-supervised industrial fault data characteristic analysis, which makes full use of the information of unlabeled samples and uses sliding window technology for time series data to extract the time axis information of industrial data sets. On this basis, in order to further improve the performance of the industrial fault classification model, we introduce a manifold regularization term into the classifier, so that the fault classifier maintains the same performance on the tangent of the dynamic data manifold, further improving the accuracy of industrial fault classification and the stability of the model, and providing important technical support for the safe and stable operation and health assessment of industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a Tennessee Eastman (TE) process flow chart. DETAILED DESCRIPTION

[0024] The dynamic semi-supervised fault classification algorithm of the deep convolutional generative adversarial network is further described in detail below in combination with a specific implementation method.

[0025] Embodiment: A dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network, the method comprising the following steps:

[0026] Step 1: Collect historical industrial process data as a training dataset. The training dataset is a labeled dataset containing fault category labels. is a labeled sample, is the corresponding label, N is the number of labeled samples, K is the number of sample categories, and the unlabeled dataset is an unlabeled sample, and M is the number of unlabeled samples.

[0027] Step 2: The dataset D collected in step 1 l ,D u Normalization is to scale each variable of the data so that each variable of the sample in the data set is in the range of [0,1]. The sliding window technique is used to expand the data sample to a two-dimensional matrix form.

[0028] Step 3: Sample n small batches of noise samples D from the random Gaussian noise space z ={z (1) ,…,z (n)} is input into the generator of the convolutional generative adversarial network to obtain the output G(z); the preprocessed dataset D l ,D uand G(z) are input into the discriminator in batches, and for each input sample, the corresponding discriminator output D(x l ),D(x u ),D(G(z)); the data set D l ,D u Input into the encoder in batches and get the encoder output h(x l ),h(x u ).

[0029] Step 4: Update the parameters of the discriminator. Calculate the three inputs D of the discriminator l ,D u , the output corresponding to G(z) is calculated using the loss function, the Jacobian matrix of the encoder is calculated, the tangent direction of the data on the manifold is obtained, the value of the manifold regularization term is calculated using the TangentProp method, and added to the loss value of the discriminator. The parameter value of the discriminator is updated through the back-propagation stochastic gradient descent method.

[0030] Step 5: Update the parameters of the generator and encoder. The loss function of the generator and encoder is the same. z Input to the generator to get output Input these synthetic samples into the discriminator to obtain the output f of a hidden layer l (G(z)), input the real data into the encoder to get the output h(x), use the FeatureMatching item to calculate the loss, and update the parameters of the generator and encoder by stochastic gradient descent.

[0031] Step 6: Repeat steps 3 to 5 until the model converges, obtain the optimal parameters of the industrial failure classification method, and achieve effective classification of historical industrial failure modes.

[0032] Step 7: Extract the discriminator in the model as the classifier for fault classification. For the data that needs online diagnosis, use normalization and sliding window preprocessing to expand the data to be classified into the form of a two-dimensional matrix, input it into the discriminator, output the classification result, and obtain the status of the current process.

[0033] The following is a specific example of the TE process to illustrate the deep convolutional generative adversarial network method for semi-supervised dynamic fault classification. The TE process is a standard data set commonly used in the field of fault diagnosis and fault classification. Figure 1As shown. 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 data set includes 53 process variables, of which 41 variables are conventional measurement variables, as shown in Table 1, which can be used as core variables for fault classification to participate in the construction of the model. In this process, we can easily simulate various fault types. The specific fault category information is shown in Table 2, including step faults, random faults, slow drift faults, sticky faults, and various unknown fault types.

[0034] In this experiment, fault and normal state are used to test the effectiveness of the method of the present invention, and the main 41 conventional process measurement variables are used as the core variables for industrial fault classification. Each type of process state contains 3000 time series samples. For the TE process, a sliding window of size 20*41 is used with a step size of 1, thus forming a two-dimensional matrix sample consisting of 20 consecutive sample points containing 41 variables, the columns of the matrix are the variables of the data, and the rows of the matrix are the time series of the data. For each type of fault, multiple two-dimensional matrix samples are obtained for actual training and testing, of which the training set accounts for 50% of the total number of samples and the test set accounts for 50%.

[0035] The experimental results show that the present invention is compared with other mainstream semi-supervised industrial fault classification methods. In a data set with a label ratio of 1%, the fully supervised convolutional neural network CNN, the long short-term memory network LSTM, and the semi-supervised variational autoencoder model VAE are compared. The semi-supervised deep convolutional generative adversarial network has relative superiority in the accuracy of industrial fault classification, especially after the manifold regularization term is added. The model improves the accuracy by more than 10% based on the original classification performance.

[0036] Table 1: Tennessee-Eastman process commonly measured variables

[0037]

[0038]

[0039] Table 2: Description of Tennessee-Eastman process failure types

[0040]

[0041] 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. A dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network, characterized in that: The steps include: Step 1: Collect historical industrial process data as a training dataset. The training dataset is a labeled dataset containing fault category labels. is a labeled sample, is the corresponding label, N is the number of labeled samples, K is the number of sample categories, and the unlabeled dataset is an unlabeled sample, M is the number of unlabeled samples, Step 2: The dataset D collected in step 1 l ,D u Normalization is to perform equal-scale transformation on each variable of the data so that each variable of the sample in the data set is in the range of [0,1], and use the sliding window technique to expand the data sample to a two-dimensional matrix form. Step 3: Sample n small batches of noise samples D from the random Gaussian noise space z ={z (1) ,…,z (n) } is input into the generator of the convolutional generative adversarial network to obtain the output G(z); the preprocessed dataset D l ,D u and G(z) are input into the discriminator in batches, and for each input sample, the corresponding discriminator output D(x l ),D(x u ),D(G(z)); the data set D l ,D u Input into the encoder in batches and get the encoder output h(x l ),h(x u ), Step 4: Update the parameters of the discriminator and calculate the three inputs D of the discriminator l ,D u , the output corresponding to G(z) is calculated using the loss function, the Jacobian matrix of the encoder is calculated, the tangent direction of the data on the manifold is obtained, the value of the manifold regularization term is calculated using the TangentProp method, and added to the loss value of the discriminator. The parameter value of the discriminator is updated through the back-propagation stochastic gradient descent method. Step 5: Update the parameters of the generator and encoder. The loss function of the generator and encoder is the same. z Input to the generator to get output Input these synthetic samples into the discriminator to obtain the output f of a hidden layer l (G(z)), input the real data into the encoder to get the output h(x), use the FeatureMatching item to calculate the loss, and update the parameters of the generator and encoder by stochastic gradient descent. Step 6: Repeat steps 3 to 5 until the model converges, obtain the optimal parameters of the industrial failure classification method, and achieve effective classification of historical industrial failure modes.

2. According to claim 1, the dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network is characterized in that: The sliding window preprocessing data method of step 2 is specifically as follows: All process data are spliced ​​together along the time T axis in time series to form an N*T data matrix, where N is the number of variables in the process data and T is the number of samples. Based on this matrix, an N*L sliding window is introduced, where L is the window size of the sliding window. Through this sliding window, the data matrix is ​​cut along the T axis according to a certain step size S to obtain (TL) / S matrix samples of size N*L. By dividing these matrix samples, we can obtain the test set and the training set, which can be used to train the model offline.

3. The dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network according to claim 1 is characterized in that: The step three specifically includes the following steps: (3.1) Construct a generator. The generator is based on a CNN network. First, the fully connected layer is transformed into a 3D form, followed by 6 layers of deconvolution layers with different convolution kernel sizes and strides. Each layer uses batchnorm. Except for the last layer, which uses the Tanh activation function, all other CNN layers use the LeakyReLU activation function. The trainable parameters of the generator are represented by θ G , (3.2) Construct an encoder. The encoder is based on a CNN network. First, it consists of six convolutional layers with different convolution kernel sizes and strides connected in sequence. Each layer uses batchnorm. Except for the first layer, which uses the Tanh activation function, all other CNN layers use the LeakyReLU activation function. Then, a fully connected layer is connected and the samples are deformed into a 1-dimensional form. The trainable parameters of the encoder are represented by θ h , (3.3) Construct a discriminator. The discriminator is based on a CNN network structure, which consists of six convolutional layers with different convolution kernel sizes and strides. Dropout is used in each layer, and the LeakyReLU activation function is used each time. Finally, a fully connected layer is connected and flattened to a one-dimensional vector.

4. The dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network according to claim 1 is characterized in that: The step 4 specifically includes the following steps: (4.1) Calculate the Jacobian matrix T of the encoder x =J x h, and calculate the Jacobian matrix J of the classifier x c, where x is the real data input to the encoder and discriminator, h is the output of the discriminator, and c is the output of the classifier. Using the TangentProp method, the manifold regularization term of the classifier is calculated (4.2) Calculate the supervised loss and convert the K-class labeled samples Divide into corresponding K units, according to the formula The supervised loss of the discriminator is obtained. (4.3) Calculate the unsupervised loss. For the unlabeled data, divide it into any of the first K units. For the false samples generated by the generator, divide it into the last K+1 units. According to the formula The unsupervised loss of the discriminator is obtained, (4.3) Adding the three losses together gives the discriminator loss L discriminator =L supervised +L unsupervised +Ω manifold , use the back-propagation stochastic gradient descent method to update the discriminator parameters θ D .

5. The dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network according to claim 1 is characterized in that: The step five is as follows: Get the outputs h(x) and G(z) of the encoder and generator respectively, concatenate G(z) and z, concatenate h(x) and x, and input them into the discriminator to get the output f of a hidden layer. l (z,G(z)) and f l (h(x),x),f l Represents the output of a hidden layer of the discriminator, and uses the featrue matching term to calculate the loss function of the encoder and the generator. Update the encoder parameters θ using back-propagation stochastic gradient descent D and the generator parameters θ G .

6. 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 dynamic semi-supervised industrial fault classification method based on deep convolutional generative adversarial network as described in any one of claims 1 to 5 above is implemented.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, a dynamic semi-supervised industrial fault classification method based on a deep convolutional generative adversarial network as described in any one of claims 1 to 5 is implemented.

Images