A chemical process fault diagnosis method based on a multi-scale convolutional neural network

Abstract

The application discloses a chemical process fault diagnosis method based on a multi-scale convolutional neural network and belongs to the field of chemical safety. First, parameters corresponding to a chemical production process are acquired and preprocessed. Second, the preprocessed data are converted into a picture set, and the picture set is divided into a training picture set and a test picture set. Then, a multi-scale convolutional neural network diagnosis model is designed for the chemical process and trained. Finally, real-time online data are collected from the chemical process, preprocessed and converted into a picture set, and input into the multi-scale convolutional neural network model. If there is a difference between the predicted result and the judgment of an expert, the model is retrained with new data. The application can more accurately diagnose faults of target data, can help diagnose and eliminate faults in time and can avoid some safety accidents.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of chemical safety, relates to matrix color block algorithm and convolutional neural network application technology, and particularly relates to a chemical process fault diagnosis method based on a multi-scale convolutional neural network (MsCNN). BACKGROUND

[0002] With the continuous progress of economic level and the continuous improvement of social science and technology, chemical production equipment is gradually automated and develops towards large-scale, precision and intelligent. At the same time, with the continuous expansion of production scale, it also means that the structure of production equipment and system becomes more and more complex, and the production process often shows characteristics of high nonlinearity, uncertainty, instability and strong correlation.

[0003] The occurrence of chemical safety accidents is generally caused by the failure of chemical device to be found and handled in time. When a fault or anomaly occurs in the chemical production process, if it cannot be correctly identified and quickly handled, it will cause the production equipment to be unable to operate normally and the production to be interrupted in the better case, and it may cause painful loss of life and property and even cause great damage to the environment in the worse case.

[0004] Although the fault detection and diagnosis method based on deep learning has achieved good results, the diagnosis accuracy of some fault types is still low, such as "Fault 09", "Fault 15" and "Fault 16" in the TE process (Teenese Eastman Process). Therefore, how to timely and accurately detect and diagnose all fault types in the chemical production process has always been a hot spot and problem faced by the chemical industry, and it is of great significance to study and improve the fault diagnosis capability in the chemical production process. SUMMARY

[0005] The present application is directed to the deficiencies of the prior art, and proposes a chemical process fault diagnosis method based on a multi-scale convolutional neural network.

[0006] The present application aims to solve the problem of low diagnosis accuracy of some fault types in the fault diagnosis of chemical production process, and proposes a fault diagnosis method based on a multi-scale convolutional neural network. First, the method converts the two-dimensional time series signal data of the chemical process into a matrix color block data set, so as to utilize the advantage of deep learning in picture feature extraction; second, a multi-scale convolutional neural network model is applied to detect and diagnose various faults in the chemical production process.

[0007] The diagnosis process includes offline and online stages, and the specific steps are as follows.

[0008] Step 1: Collect historical data from chemical processes and perform preprocessing.

[0009] Step 1.1: Obtain parameters corresponding to the production process of a chemical enterprise, such as chemical process historical normal data and chemical process historical fault data, etc.

[0010] Step 1.2: Data preprocessing using min-max normalization formula.

[0011] For convenience of calculation, normalization is needed. Data preprocessing is performed using the min-max normalization formula.

[0012] Step 2: Convert the preprocessed historical data into a sample picture set, including normal and fault types.

[0013] In the present application, an effective data preprocessing method is applied, i.e. converting time series original signal data into images. Each set of processed parameters is a time series data, and multiple variables are combined into a data matrix, successfully drawing a data matrix color block diagram, which uses rich colors and shapes to visually display the size of matrix element values. The square shape of each color block in the matrix color block diagram is the same size. The number of color blocks is determined by the number of parameters in the matrix. The color filled in each color block is given in the color standard for generating the color block matrix diagram. Therefore, the color blocks corresponding to each element in the matrix color block diagram control the color properties of the square color blocks through different parameter values.

[0014] Step 3: Divide the sample picture set into two data sets, i.e. training picture set and test picture set.

[0015] Step 4: Design a diagnosis model (multi-scale convolutional neural network) for the chemical process.

[0016] The multi-scale convolutional neural network is composed of three basic parts: local receptive field, shared parameters and time or space downsampling, which guarantee the translation invariance of the network model. The input image of the network is usually preprocessed to a fixed size, and then the input of each subsequent layer is obtained through local connection with the previous layer. In order to prevent network overfitting and increase network convergence speed, batch normalization processing is added after convolution operation. The multi-scale convolutional neural network proposed in the present application includes the following three components: multi-scale feature extraction, feature learning and nonlinear regression.

[0017] Step 5: Train and test the multi-scale convolutional neural network diagnosis model.

[0018] Step 6: Output the fault diagnosis result.

[0019] The multiscale convolutional neural network diagnosis model is applied to process fault diagnosis to evaluate the fault diagnosis performance of the proposed method, and two indexes are used: accuracy and loss function (Loss).

[0020] Step 7: If the fault diagnosis rate in the test meets the requirements, the diagnosis model will be used to perform online fault diagnosis operation. If the fault diagnosis rate is not satisfactory, the multiscale convolutional neural network diagnosis model needs to be redesigned (return to step 4).

[0021] Step 8: Collect real-time online data from the chemical process and perform preprocessing.

[0022] Step 8.1: Obtain parameters corresponding to the production process of the chemical enterprise, such as real-time data of the chemical process.

[0023] Step 8.2: Synchronize step 1.2, use the min-max normalization formula for data preprocessing.

[0024] Step 9: Convert the preprocessed online data into a picture set.

[0025] Synchronize step 2, convert the preprocessed online data into pictures to form an online picture set.

[0026] Step 10: Input the online sample picture set into the multiscale convolutional neural network diagnosis model. The model can provide a predicted diagnosis result for each sample picture set. The diagnosis result is normal or a specific fault type.

[0027] Step 11: If there is a difference between the predicted diagnosis result and the expert's judgment, the multiscale convolutional neural network diagnosis model needs to be retrained using new data.

[0028] The beneficial effects of the present application: The present application combines the difficulties of fault detection and low partial fault diagnosis rate in chemical processes, and explains the effect of the multiscale convolutional neural network model in the chemical process fault diagnosis method. In the present application, the time series signal data of the chemical production process is converted into matrix color block picture data set during data preprocessing, which can fully utilize the advantages of deep learning in picture processing. Then, different scale convolution kernels are applied to extract different feature information from the picture data set, and the network is deepened for processing, so that the network can learn the mapping relationship more comprehensively, thereby achieving more accurate fault diagnosis of target data, and timely diagnosis and elimination of faults can avoid some safety accidents. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 Process flow chart for chemical process fault diagnosis based on multiscale convolutional neural network model;

[0030] Figure 2For two-dimensional time series data to be converted into a matrix color block diagram;

[0031] Figure 3 For a multi-scale convolutional neural network architecture;

[0032] Figure 4 For the TE process process flow chart;

[0033] Figure 5 For each group of 22 training pictures, extract one display picture;

[0034] Figure 6 For each group of 22 test pictures, extract one display picture;

[0035] Figure 7 For the accuracy and Loss curve of the training set and test set of MsCNN. DETAILED DESCRIPTION

[0036] The application is further described below.

[0037] The process of the application is shown in Figure 1 The specific steps are as follows:

[0038] Step 1: Collect historical data from the chemical process and perform preprocessing.

[0039] Step 1.1: Obtain parameters corresponding to the production process of the chemical enterprise, such as chemical process historical normal data and chemical process historical fault data, etc.

[0040] At present, there is too little historical fault data to use deep learning methods for fault detection and diagnosis. Therefore, in general, the data of the Tennessee Eastman Process (TE Process) is used for simulation verification. The TE process is a simulated data, although there are some differences with the actual industrial process, but it can reflect the actual calculation process, at the same time, there is no need to worry that the scheme researched will affect the actual production process. The TE process is a computer simulation program based on actual chemical process, which was proposed by Downs J.J. and Vogel E.F. of Tennessee Eastman Chemical Company in 1993. In 1995, RICKERN.L. et al. improved and optimized the TE process. The TE process is a typical chemical system process, which has strong coupling, time-varying and nonlinearity, and has all the characteristics of actual chemical production process, and its process fault also conforms to the actual situation, which is suitable for research object in the field of fault diagnosis. Process data can be easily obtained, which is conducive to the theoretical research of chemical process. The TE process simulation has 21 preset faults and 53 fault monitoring variables.

[0041] (1) TE process process flow

[0042] The TE process flow diagram of the simulator revision is shown in Figure 1. Figure 4 The overall process has five operating units: reactor, condenser, recycle compressor, gas-liquid separator and desorber.

[0043] (2) TE process chemical reactions

[0044] In the TE process, there are four relevant gaseous reactants, A, C, D and E. There is also a small amount of inert, insoluble B gas component in the various feed gases. The four chemical reactions shown below take place simultaneously in the reactor, with the catalyst, two of which produce the target products G and H, F is a by-product. The first two chemical reactions (Reaction 1 and Reaction 2) are the main reactions of the TE process, the latter two (Reaction 3 and Reaction 4) are side chemical reactions. All four chemical reactions are exothermic and irreversible. The reaction equations are shown below.

[0045] Reaction 1 : A (gas) + C (gas) + D (gas) → G (liquid)

[0046] Reaction 2: A (gas) + C (gas) + E (gas) → H (liquid)

[0047] Reaction 3: A (gas) + E (gas) → F (liquid)

[0048] Reaction 4: 3D (gas) → 2F (liquid)

[0049] (3) TE process variable parameters

[0050] There are 53 process parameter variables in total for the TE process to be monitored for faults, of which 12 are control variables (see Table 1) and 41 are measured variables. Of the 41 process measured variables, 22 are continuous process variable measurements and 19 are component variable measurements. The present application actually acquires 52 parameter variables as data sets, the first 11 process control variables and the 41 process measured variables. The 22 continuous process variable descriptions are XMEAS(1) - XMEAS(22), as shown in Table 2. This data is obtained by sampling every 3 minutes. The 19 component variable descriptions are XMEAS(23) - XMEAS(41), as shown in Table 3, the component contents measured from stream 6, stream 9 and stream 11, stream 6 and stream 9 are sampled every 6 minutes, the same time delay is also set to 6 minutes; stream 11 is sampled every 15 minutes.

[0051] Table 1 12 process control variables

[0052]

[0053]

[0054] Table 2 22 are continuous process variables

[0055]

[0056] Table 3 19 are component variables

[0057]

[0058] (4) TE process fault data

[0059] A normal condition and 21 fault conditions are pre-set in the TE process system data. Among the fault conditions, 16 faults are explicit, and 5 faults are not explicit. Fault 13 is caused by slow drift in the reaction kinetics. Fault 17 is caused by a step change in the process variable, such as a change in the cooling water inlet temperature or a change in the composition of the system feed within a short period of time. Faults 8 and 12 are caused by an increase in the variability of the system process variable, and faults 14, 15, and 21 are related to the viscous valve. The fault description and type of each fault are shown in Table 4.

[0060] Table 4 Fault condition data description parameters

[0061]

[0062] Step 1.2: Data preprocessing using the min-max normalization formula.

[0063] The data selected for fault diagnosis of chemical production processes is obtained from a chemical process data set that contains a large number of various variables such as liquid level, pressure, and temperature. The data range is large, and the dimensions of each variable are different. In order to facilitate calculation, normalization processing is required. Data preprocessing is performed using the min-max normalization formula, as shown in the following formula:

[0064]

[0065] where x' is the normalized data after data processing. x is the original data, and min(x) and max(x) are the minimum and maximum values of the original data.

[0066] Step 2: Convert the preprocessed historical data into a picture set, including normal and fault types.

[0067] In traditional deep learning-based fault diagnosis methods, data preprocessing methods are crucial because most methods cannot directly process raw signals. One of the main functions of data preprocessing methods is to extract features of the original signal from a large amount of historical data. However, extracting appropriate features is a laborious task, and these features have a great impact on the final result.

[0068] In the present application, an effective data preprocessing method is applied, i.e. converting time series original signal data into images. Each set of processed parameters is a time series data, and multiple variables are combined into a data matrix. The data color block matrix diagram is successfully drawn, and the size of the matrix element value is displayed in a rich color and shape. The square shape of each color block in the matrix color block diagram is the same size. The number of color blocks is determined by the number of parameters in the matrix, for example, if the parameter matrix is n1xm1, n1xm1 color block diagram is generated. The color filled in each color block is given in the color standard for generating the color block matrix diagram. Therefore, the color blocks corresponding to each element in the matrix color block diagram control the color properties of the square color blocks through different parameter values. This completes the conversion of two-dimensional time series data into two-dimensional color block matrix pictures. At this time, the advantages of the current deep learning in picture processing can be fully utilized. For example, there is a matrix data x, as shown below, and the color block matrix diagram is as shown in Figure 2 The advantage of this data processing method is that it provides a method for converting two-dimensional time series signals into image features. More importantly, this data preprocessing method can be calculated without any predefined parameters, and can minimize the experience of experts as much as possible.

[0069]

[0070] Step 3: Divide the sample picture set into two data sets, i.e. training picture set and test picture set.

[0071] The normal and fault data sets of the TE process are 22 groups, and after using the original data processing, 22 new data sets are converted. Each data set is a 52-variable time series data matrix, considering that there are 52 variables at each time point. Therefore, 52 time points are extracted to form a 52x52 data matrix, and a 52x52 data matrix color block diagram is drawn, thereby generating a corresponding two-dimensional color block matrix picture set.

[0072] The normal data and fault data of the TE process are also collected from the simulation on MATLAB 2016a. The sampling time is set to 3 minutes (20 samples / hour). The samples under normal working conditions. The training samples are 500 groups of observation data obtained under a 25h running simulation. Repeat 13 times to get 6500 groups of training samples. Take 52 data from each group to generate a 52x52 matrix color block diagram, which can generate a training data set of 125 pictures. While the test samples are 960 groups of observation data obtained under a 48h running simulation, repeated 13 times to get 12480 groups of test samples. Take 52 data from each group to generate a 52x52 matrix color block diagram, which can generate a training data set of 240 pictures. The samples under fault working conditions. The training samples are 480 groups of observation data obtained under a 24h running simulation. Repeat 13 times to get 6240 groups of training samples. Take 52 data from each group to generate a 52x52 matrix color block diagram, which can generate a training data set of 120 pictures. While the test samples are 960 groups of observation data obtained under a 48h running simulation program, and the fault is imported at 8h, the first 160 observation values are normal data, so the test sample is only 800. Repeat 13 times to get 10400 groups of test samples. Take 52 data from each group to generate a 52x52 matrix color block diagram, which can generate a training data set of 200 pictures.

[0073] Finally, a total of 22 groups of training and 22 groups of test picture sets are generated. The 22 groups of picture sets include one group of picture set under normal working conditions and 21 groups of picture set under fault working conditions. As shown in Figure 5 and Figure 6 , one picture is randomly selected from each group of data picture set and arranged together for display. 0 is the picture generated by the normal working condition data set, and 1-21 are the corresponding pictures generated by the 21 fault working condition data sets.

[0074] Step 4: Design a diagnosis model (multi-scale convolutional neural network) for the chemical process.

[0075] The multi-scale convolutional neural network is composed of three basic parts, which are local receptive field, shared parameters and time or space downsampling, which basically guarantee the translation invariance of the network model. The input image of the network is usually preprocessed to a fixed size, and then the units of each subsequent layer are obtained through local connection with the previous layer. In this application, three different scale convolution kernels (7x7, 5x5 and 3x3) are used to extract features from the original image, and a small 3x3 convolution kernel is applied during feature training and learning to increase the depth of the model. In order to prevent network overfitting and increase the convergence speed of the network, batch normalization processing is added after convolution. The multi-scale convolutional neural network proposed in this application includes the following three components: multi-scale feature extraction, feature learning and nonlinear regression. The specific network structure is as follows: Figure 3The relevant main parameters of the network model are shown in Table 5. As can be seen from the multi-scale convolutional neural network architecture, the model has 6 convolutional layers, 5 max pooling layers, 2 batch normalization layers and 4 fully connected layers.

[0076] Table 5 Main parameters of multi-scale convolutional neural network

[0077]

[0078]

[0079] Step 5: Training and testing the multi-scale convolutional neural network diagnostic model.

[0080] Step 6: Output the fault diagnosis result.

[0081] The experimental hardware environment of the present application is: a server, the processor is Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz, the memory is 32GB, the hard disk is 256GB, the graphics processing unit (GPU) is NVIDIA GeForce GTX 1080Ti, and the display memory is 16GB. The software environment is: Windows 10 system. The experimental framework adopts Keras, the development language is Python 3.7.6, the learning rate is set to 5x10 -4 , and the momentum is set to 0.9.

[0082] The multi-scale convolutional neural network model is applied to process fault diagnosis to evaluate the fault diagnosis performance of the proposed method, and two indexes are used here: accuracy and loss function (Loss).

[0083] (1) Accuracy

[0084] Because the number of samples in each category in the data set is not equal, the multi-class evaluation index selected is accuracy (Accuracy), and the calculation process is as shown in the following formula:

[0085]

[0086] (2) Loss function (Loss)

[0087] Cross-entropy is a numerical value for evaluating the difference between the probability distribution obtained by the current training and the true distribution. It describes the difference between the expected and actual output, and reflects the probability distribution of the two output values. Therefore, the smaller the cross-entropy, the more similar the probability distribution. In this paper, the multi-class cross-entropy loss function L is used, as shown in the following formula:

[0088]

[0089] where n is the number of samples and m is the number of classes. y is the expected output of the model, which is 1 if the class is j and 0 otherwise. p is the actual output of the model, which is the probability that the class is j. This output value is calculated using Softmax. j

[0090] The curves of the batch size on the fault detection accuracy and Loss of the training set and the test set are shown in FIG. 2. Figure 7 As can be seen from the figure, the accuracy increases with the increase of the batch size, and when the batch size is greater than 6000, the accuracy of the training set and the test set begins to stabilize, that is, the result of the MsCNN model has stabilized. The time cost of the model for one full coverage training of the sample data, in order to make the performance of the method more optimal, the number of times is selected to be about 6000.

[0091] The training set and the test set data of the TE process are collected in the application, including the data of the normal working condition and 21 fault working conditions. After the time series data is preprocessed, it is converted into 22 new data sets, and then a two-dimensional color block matrix picture set is generated, and finally the MsCNN model is designed and completed for fault diagnosis by using the Tensorflow open source framework. The fault detection accuracy and Loss of the training set and the test set are shown in FIG. 1 and Table 6. As can be seen from the table, the MsCNN model can accurately detect the normal working condition and the 21 fault working conditions, and therefore the comprehensive fault detection rate is obtained. The fault detection accuracy of each type of fault of the training set is as high as 98.44%, and the Loss is 0.09. The fault detection accuracy of each type of fault of the test set is as high as 88.54%, and the Loss is 0.47. The method has good detection results for each type of fault data of the training set and the test set, which shows that the method based on MsCNN can complete the fault diagnosis of the TE process, and confirms the effectiveness thereof. Figure 7 Table 6 Experimental results of MsCNN

[0092]

[0093] Step 7: If the fault diagnosis rate in the test meets the requirements, the model is used to perform online fault diagnosis operation. If the fault diagnosis rate is not satisfactory, the multi-scale convolutional neural network model needs to be redesigned (return to step 4).

[0094] From the result of the model, the fault diagnosis rate meets the requirements, and the three common fault diagnosis types with low accuracy are greatly improved.

[0095] Step 8: Real-time online data are collected from the chemical process and preprocessed.

[0096] Step 8.1: The parameters corresponding to the production process of the chemical enterprise are obtained, such as real-time data of the chemical process.

[0097] ​​

[0098] Step 8.2: Data pre-processing using min-max normalization formula as in step 1.2.

[0099] Step 9: Convert pre-processed online data into picture sets.

[0100] Step 2, Convert pre-processed online data into pictures, compose online picture sets.

[0101] Step 10: Input online sample picture sets into the multi-scale convolutional neural network diagnostic model. The model can provide a predicted diagnostic result for each sample picture set. The diagnostic result is normal or a specific fault type.

[0102] Step 11: If there is a difference between the predicted diagnostic result and the expert's judgment, the multi-scale convolutional neural network diagnostic model needs to be retrained using new data.

Claims

1. A method for fault diagnosis in chemical processes based on multi-scale convolutional neural networks, characterized in that... The specific steps include the following: Step 1: Collect historical data from chemical processes and perform preprocessing on it; Step 1.1: Obtain the parameters of the corresponding chemical enterprise's production process; Step 1.2: Perform data preprocessing using the min-max normalization formula; Step 2: Convert the preprocessed historical data into a sample image set, including normal and fault types; Each set of parameters after preprocessing is a time series data. Multiple time series data are combined into a data matrix, and a data matrix color block diagram is successfully drawn, which uses rich colors and shapes to vividly display the size of the matrix element values. Step 3: Divide the sample image set into two datasets: the training image set and the test image set; Step 4: Design a multi-scale convolutional neural network diagnostic model for chemical processes; The multi-scale convolutional neural network comprises the following three components: multi-scale feature extraction, feature learning, and nonlinear regression. The multi-scale convolutional neural network diagnostic model processes the input image of the network into a fixed size, and then obtains the input of each subsequent layer through local connections with the previous layer; Step 5: Train and test the multi-scale convolutional neural network diagnostic model; Step 6: Output the fault diagnosis results; Accuracy and loss function are used as evaluation metrics to evaluate multi-scale convolutional neural network diagnostic models; Step 7: If the fault diagnosis accuracy in the test meets the requirements, the multi-scale convolutional neural network diagnostic model will be used to perform online fault diagnosis. Otherwise, return to step 4 to redesign the multi-scale convolutional neural network model; Step 8: Collect real-time online data from the chemical process and perform preprocessing; Step 9: Same as Step 2, convert the preprocessed data from Step 8 into images to form an online image set; Step 10: Input the obtained online image set into the multi-scale convolutional neural network diagnostic model obtained in Step 7; The model provides a predicted diagnostic result for each sample in the online image set, which is either normal or a type of fault. Step 11: If there is a discrepancy between the predicted diagnosis and the expert's judgment, the multi-scale convolutional neural network diagnostic model is retrained using new data; In step 2, each color block in the matrix color block diagram has the same square shape size. The number of color blocks is determined by the number of parameters of the matrix. The color filled in each color block is based on the color standard of the generated color block matrix diagram.

2. The chemical process fault diagnosis method based on a multi-scale convolutional neural network according to claim 1, characterized in that: In step 4, the multi-scale convolutional neural network diagnostic model adds batch normalization processing after the convolution operation.

3. The chemical process fault diagnosis method based on a multi-scale convolutional neural network according to claim 2, characterized in that: The multi-scale convolutional neural network consists of three basic parts: local acceptor fields, shared parameters, and temporal or spatial downsampling. These parts ensure the translation invariance of the network model.

4. The chemical process fault diagnosis method based on multi-scale convolutional neural networks according to claim 1, characterized in that: The loss function mentioned in step 6 is the cross-entropy loss function.

5. The chemical process fault diagnosis method based on a multi-scale convolutional neural network according to claim 1, characterized in that: Step 8 is as follows: Step 8.1: Obtain real-time parameter data of the corresponding chemical enterprise's production process; Step 8.2: Perform data preprocessing using the min-max normalization formula.

Images