Robust multivariate fault detection method for blast furnace ironmaking based on time series convolutional network

Abstract

The application discloses a blast furnace ironmaking robust multivariate fault detection method based on a temporal convolution network (RTCDN), which stacks residual modules in a temporal convolution network (TCN) to build a temporal convolution detection network (TCDN). l 1 norm-based robust solution is proposed based on the TCDN, which can separate data with outliers, separate a sparse outlier matrix, and then obtain clean data for model training, so that the model has strong robustness.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of industrial process monitoring, and particularly relates to a robust multivariate fault detection method for blast furnace ironmaking based on a time series convolution network. BACKGROUND

[0002] The steel industry is an important pillar industry of modern industry and national economy in China, and is an important component of the national economy. Blast furnace ironmaking is the core link of the entire steel industry, and is also the key process of energy and mass conversion in the steel manufacturing process, and the energy consumption of blast furnace ironmaking accounts for as high as 70% of the total energy consumption of steel production. Realizing safe operation of the blast furnace ironmaking process is the top priority of deep energy saving and emission reduction and quality improvement and efficiency increase of the steel industry, and an accurate abnormality monitoring model can provide early warning for operators in time, so as to adjust the furnace condition in advance and avoid dangerous situations.

[0003] In the blast furnace ironmaking process, abnormal furnace conditions of the blast furnace often occur due to fluctuations in raw material quality, misoperation of personnel, equipment abnormalities and other reasons. Once the blast furnace furnace condition is abnormal, it will often cause problems such as an increase in fuel ratio, an increase in downtime for maintenance, and off-specification molten iron, which not only causes significant loss of resources and equipment and reduces the service life of the blast furnace, but also may even cause accidents and casualties. Therefore, improving the safety of the blast furnace ironmaking process, especially reducing the frequency of blast furnace accidents and abnormal furnace conditions, has a very important significance for reducing energy consumption, ensuring the safety of equipment and personnel, and improving the economic benefits of the steel production process. However, according to the survey results, the current fault identification mainly relies on the working experience of operators, which has two disadvantages: first, the accumulation of fault experience requires several years, and second, continuous observation of data curves is a tedious task that is prone to fatigue errors. With the continuous improvement of the level of industrial informatization, data-driven process monitoring has become a key technology for improving the safety, quality and operating efficiency of process industries. The data of large blast furnaces also has the characteristics of non-Gaussian distribution and time-varying, and the abnormal monitoring and diagnosis thereof has become a front and difficult point in the research of metallurgical technology in the world. SUMMARY

[0004] The application proposes a robust multivariate fault detection method for blast furnace ironmaking based on a time series convolution network to solve the existing problems in blast furnace abnormality monitoring, which uses a time series convolution network to capture the dynamic characteristics of the process and adapt to the dynamics of blast furnace data, and proposes a robust solution to overcome the influence of outliers in the original data, so that the final model has strong robustness.

[0005] A blast furnace ironmaking robust multivariate fault detection method based on a time series convolution network, a time series convolution detection network (TCDN) is constructed by stacking residual modules in a time series convolution network (TCN), the TCDN extracts dynamic information of a process on a time scale by using a causal time series convolution network, simultaneously transforms a spatial dimension of a variable by using a number of channels of a convolutional neural network, and realizes a fault detection function according to a sum of squares of reconstruction errors of the variable; and a robust solution based on an l1 norm is further proposed for the phenomenon that original data of the blast furnace generally exist outliers, the solution can separate the original data with outliers, separate a sparse outlier matrix, and thus obtain data without outliers for model training, so that the model has the ability to overcome the interference of outliers.

[0006] The time series convolution detection network (TCDN) constructed by stacking residual modules in the time series convolution network (TCN) is:

[0007] The residual modules of the time series convolution network TCN are stacked, and the stacked layer is denoted as , the time step of the model is k, and thus the process variable input form of the model is , , the corresponding latent variable is obtained through each residual block, and the latent variables are , , in turn, the output of the previous residual block will be used as the input of the next residual block; the objective function of the TCDN is to reconstruct the input, that is, to minimize the mean square error (MSE) between the output of the last residual block and the original input, as shown in the following formula:

[0008] .

[0009] Wherein, MSE is a loss function of model training, n is a sample number, k is a set time step, is an input, is the output of the last residual block, that is, a reconstruction result.

[0010] The TCDN extracts dynamic information of a process on a time scale by using a causal time series convolution network, and the TCDN is:

[0011] The convolutional neural network used by the TCDN is a one-dimensional convolutional neural network, and the convolution direction is the time dimension of the variable.

[0012] The TCDN simultaneously transforms a spatial dimension of a variable by using a number of channels of a convolutional neural network, and the TCDN is:

[0013] The residual module is composed of a convolutional neural network, and the channel number of the convolutional neural network structure needs to be designed, wherein the channel number of the original input is the spatial dimension of the variable , the channel number of the convolutional neural network in the middle part of the residual module is designed as , wherein , so that the latent variable space can realize the compression of the spatial dimension of the variable; and the channel number of the convolutional network in the last residual module is designed as , that is, the spatial dimension of the original input is reconstructed.

[0014] The TCDN realizes the fault detection function according to the sum of squares of the reconstruction error of the variable, and the result of the reconstruction is assumed to be , wherein , so that the monitoring statistic can be designed as the sum of squares of the reconstruction error, that is, , and the threshold value of the monitoring statistic can be estimated by the SPE of each sample during training through the kernel density estimation (KDE) method. When online detection is performed, if the SPE value calculated by the new sample is greater than , it is judged that a fault has occurred; if the SPE value calculated by the new sample is less than or equal to , it is considered that the process has not occurred.

[0015] On the basis of the TCDN, a robust solution based on the l1 norm is proposed, which can separate the data with outliers and separate the sparse outlier matrix, and then retain the clean data for model training as follows:

[0016] Since the process variable contains outliers, the original input is decomposed into two parts , wherein is a low-rank matrix, and is a sparse outlier matrix. The following loss function is first constructed:

[0017] ;

[0018] ;

[0019] , wherein represents the TCDN model, represents the reconstruction result of the TCDN, represents the part that can be reconstructed by the TCDN model, is the l1 norm, and is the F norm, is a weight coefficient. The optimization model needs to be solved by using the ADMM method, that is., the inverse gradient propagation optimization , followed by fixing , followed by fixing , followed by fixing , followed by fixing , and so on until convergence. Finally, the clean data , which does not contain outliers, is obtained, and is used to train the TCDN model for fault detection.

[0020] The above proximal method optimizes The proximal method in the above is:

[0021] ;

[0022] where is the iteration number, is a hyperparameter. is the value of at the th iteration. When an element is separated for two times consecutively and the difference is less than , it is directly identified as an outlier and is not optimized any more.

[0023] A procedure of the blast furnace robust multivariate fault detection method based on time series convolutional network is as follows:

[0024] Step 1: offline modeling;

[0025] Step 1.1: collect the process variables of the blast furnace sensors, and the normalized data is denoted as , the sample number is , the variable number is , determine the time step k of the process, and construct the input variable form of the neural network , ;

[0026] Step 1.2: separate according to the loss function and the solving method, and then train the TCDN model with ;

[0027] Step 1.3: calculate the monitoring statistics of all samples , and determine the upper limit of the monitoring statistics according to the kernel density estimation (KDE) method with the confidence level ;

[0028] Step 2: online monitoring;

[0029] Step 2.1: when a new sample of the blast furnace comes, normalize it, and the normalized data is denoted as Then the input is input into the TCDN model with trained parameters in offline modeling to obtain a reconstruction result , and the upper limit of the monitoring statistics is calculated :

[0030] ;

[0031] The contrast statistics and the upper limit of the statistics , if yes , it means that a fault occurs, otherwise it is in normal operation.

[0032] The beneficial effects of the present application are:

[0033] 1. The convolutional neural network is used for fault detection, and the calculation efficiency is higher than that of the traditional time sequence method, and has longer time sequence processing capability.

[0034] 2. The special structure of the network supports parallel time information extraction and spatial dimension transformation.

[0035] 3. A robust solution is proposed, which can effectively avoid the interference caused by outliers.

[0036] 4. The model considers the nonlinearity, dynamics and outlier problems of the process, effectively reduces the false positive rate of the algorithm, and can timely alarm the process fault. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1-1 , Figure 1-2 respectively, are autocorrelation analysis diagrams of hot air temperature and theoretical combustion temperature.

[0038] Figure 2 is the architecture diagram of the detection method.

[0039] Figure 3 is a specific detail diagram for extracting time scale information and spatial dimension transformation of the detection method.

[0040] Figure 4 is the flow chart of the offline training and online detection method.

[0041] Figure 5 is the SPE statistics monitoring diagram of TCDN on test set 1.

[0042] Figure 6 is the SPE statistics monitoring diagram of TCDN on test set 2.

[0043] Figure 7 is the SPE statistics monitoring diagram of TCDN on test set 3.

[0044] Figure 8-1 , Figure 8-2Variable schematic diagram containing outliers and variable schematic diagram after outliers are separated by the method, respectively. DETAILED DESCRIPTION

[0045] The application is further described below in conjunction with the accompanying drawings and examples.

[0046] A blast furnace ironmaking robust multivariate fault detection method based on a time series convolution network, characterized in that residual modules in a time series convolution network (TCN) are stacked to build a time series convolution detection network (TCDN), the TCDN extracts dynamic information of a process on a time scale using a causal time series convolution network, simultaneously transforms a spatial dimension of a variable using a number of channels of a convolutional neural network, and realizes a fault detection function according to a sum of squares of reconstruction errors of the variable; and a robust solution based on an l1 norm is further proposed for the phenomenon that outliers exist in original data of a blast furnace, the solution can separate the original data with outliers, separate a sparse outlier matrix, and thus obtain data without outliers for model training, so that the model has the ability to overcome interference of outliers.

[0047] The residual modules in the time series convolution network (TCN) are stacked to build the time series convolution detection network (TCDN), which is:

[0048] The residual modules of the time series convolution network TCN are stacked, and the stacked layer is denoted as , the time step of the model is k, and thus the input form of the process variable of the model is , , the corresponding latent variable is obtained through each residual block, and the latent variables are , , in turn, the output of the previous residual block will be used as the input of the next residual block; the objective function of the TCDN is to reconstruct the input, that is, to minimize the mean square error (MSE) between the output of the last residual block and the original input, as shown in the following formula:

[0049] ;

[0050] wherein, MSE is a loss function of model training, n is a sample number, k is a set time step, is an input, is the output of the last residual block, that is, a reconstruction result.

[0051] The TCDN extracts dynamic information of a process on a time scale using a causal time series convolution network, which is:

[0052] The convolutional neural network used by the TCDN is a one-dimensional convolutional neural network, and the convolution direction is the variable time dimension.

[0053] The transformation of the channel number of the convolutional neural network to the variable spatial dimension is as follows:

[0054] The residual module is composed of a convolutional neural network, and the channel number of the convolutional neural network structure needs to be designed, wherein the channel number of the original input is the spatial dimension of the variable , the channel number of the convolutional neural network in the middle part of the residual module is designed as , wherein , so that the latent variable space can realize the compression of the variable spatial dimension; and the channel number of the convolutional network in the last layer of the residual module is designed as , that is, the spatial dimension of the original input is reconstructed.

[0055] The TCDN realizes the fault detection function according to the sum of squares of the reconstruction error of the variable, which is assumed to be , wherein , so that the monitoring statistic can be designed as the sum of squares of the reconstruction error, that is, , and the threshold value of the monitoring statistic can be estimated by the SPE of each sample during training through the kernel density estimation (KDE) method. When online detection is performed, if the SPE value calculated by the new sample is greater than , it is judged that a fault has occurred; if the SPE value calculated by the new sample is less than or equal to , it is considered that no fault has occurred in the process.

[0056] On the basis of the TCDN, a robust solution based on the l1 norm is proposed, which can separate the data with outliers and separate the sparse outlier matrix, and then retain clean data for model training as follows:

[0057] Since the process variable contains outliers, the original input is decomposed into two parts , wherein is a low-rank matrix, and is a sparse outlier matrix. First, the following loss function is constructed:

[0058] ;

[0059] ;

[0060] , wherein represents the TCDN model, represents the reconstruction result of the TCDN, This represents the part that can be reconstructed by the TCDN model. It is the l1 norm, and It is the F-norm. These are the weight coefficients. This optimization model requires solving using the ADMM method, i.e., optimization using backpropagation. Then fixed Then optimize using the proximal method. Then fixed Continue to optimize This process is repeated until convergence. Ultimately, the desired result is... It is clean data that does not contain outliers, which is used to train the TCDN model for fault detection.

[0061] The above-mentioned proximal method optimization The proximal method in the text is:

[0062] ;

[0063] in Indicates the number of iterations. It is a hyperparameter. express In the The value at the next iteration. When an element is separated twice consecutively and the difference is less than 1. At that point, it is directly identified as an outlier and is no longer optimized.

[0064] A specific process is as follows:

[0065] Step 1: Offline modeling;

[0066] Step 1.1: Collect the process variables from the blast furnace sensors, and record the normalized data as follows: The sample size is The number of variables is Determine the time step k of the process, and construct the input variable form of the neural network. , .

[0067] Step 1.2: Based on the loss function and solution method, separate and obtain... Then use Train the TCDN model.

[0068] Step 1.3: Calculate the monitoring statistics for all samples. And based on the kernel density estimation (KDE) method, with confidence level Determine the upper limit of monitoring statistics ;

[0069] Step 2: Online monitoring;

[0070] Step 2.1: When a new sample of the blast furnace comes , it is also normalized, and the normalized data is denoted as , which is input into the TCDN model trained by the parameters of offline modeling to obtain the reconstruction result , and the upper limit of the monitoring statistic is calculated:

[0071] ;

[0072] The contrast statistic and the upper limit of the statistic , if , it means that a fault has occurred, otherwise it is in normal operation.

[0073] Embodiment

[0074] 1. Introduction of blast furnace process variables

[0075] This experiment is aimed at the 2650 cubic meter blast furnace of Guangxi Liuzhou Iron and Steel Group. The data of the furnace body is sampled once every 10s on average, containing 16 variables, as shown in Table 1.

[0076] Table 1 Variable list of data set

[0077] 1 Oxygen enrichment rate (%) 9 Furnace top pressure (kPa) 2 BF permeability index 10 Total pressure difference (MPa) 3 Standard wind speed (m / s) 11 Actual wind speed (m / s) [4 Oxygen enriched flow (m 3 / h)] 12 Hot blast temperature (°C) 5 Blowing kinetic energy (KJ) 13 Cold blast pressure (2) (MPa) 6 Furnace top gas quantity (m 3 )]]> 14 Resistance coefficient 7 BF bosh gas index 15 BF top temperature (°C) 8 Theoretical combustion temperature (°C) 16 Hot blast pressure (MPa)

[0078] 2. Analysis of blast furnace process variables

[0079] Dynamic:

[0080] The autocorrelation analysis of 2 representative variables, theoretical combustion temperature and hot blast temperature, is shown in Figure 1-1 , Figure 1-2 . The dynamic nature of data refers to the correlation between variables at different times. As can be seen from the figure, the autocorrelation coefficients of the 2 typical variables are high, and the dynamic nature of the variables is obvious. Therefore, dynamic consideration is necessary for process monitoring.

[0081] 3. Algorithm framework

[0082] The framework of the entire algorithm is shown in Figure 2 , 3 .

[0083] 4. A process of blast furnace abnormality monitoring method of Gaussian mixture model embedded GRU autoencoder

[0084] The specific process is shown in Figure 4 , which can be divided into offline modeling and online detection:

[0085] Step 1: Offline modeling;

[0086] Step 1.1: Collect the process variables of the blast furnace sensor, and the normalized data is denoted as , the number of samples is , the number of variables is , determine the time step k of the process, and construct the input variable form of the neural network , .

[0087] Step 1.2: Calculate the loss function and solution method, separate to get , and then use to train the TCDN model.

[0088] Step 1.3: Calculate the monitoring statistics of all samples , and determine the upper limit of the monitoring statistics according to the kernel density estimation (KDE) method with a confidence level of ;

[0089] Step 2: Online detection;

[0090] Step 2.1: When a new sample of the blast furnace arrives , it is also normalized, and the normalized data is denoted as , then input it into the TCDN model trained with the parameters of offline modeling to get the reconstruction result , calculate its monitoring statistics upper limit :

[0091] ;

[0092] Compare the statistics and the upper limit of the statistics , if , it means a failure has occurred, otherwise it is in normal operation.

[0093] 5, System final output result:

[0094] Test 1: This test is to test the effect of TCDN, and the training set does not contain outliers.

[0095] Take 30000 historical data of blast furnace sensors as the training set, and 3 test sets to verify the system, as shown in Table 2.

[0096] Table 2 Data set introduction

[0097]

[0098] After the algorithm is trained, the monitoring results of the three test sets are as follows Figure 5、 6 , 7. The dots in the figure are the SPE values of each sample, the dotted line is the threshold of the monitoring statistics, and the gray background part is the interval of fault occurrence. As shown in the figure, the method can effectively detect faults and has a low false positive rate and false negative rate.

[0099] Test 2: This test is to test the outlier separation effect of the robust solution.

[0100] The outlier separation effect is shown in FIG. 8. As shown in the figure, the method can effectively separate outliers in the original variables containing outliers to obtain data without outliers. Figure 8-1 、 Figure 8-2

[0101] Test 3: This test is to test the fault detection effect of RTCDN, taking the first temperature anomaly test set as an example.

[0102] The comparison results of different methods under different outlier proportions are shown in Table 3.

[0103] Table 3 Fault detection effect of RTCDN method under different outlier proportions

[0104]

[0105] As shown in Table 3, RTCDN can achieve good detection effect under different outlier proportions, and the false positive rate and false negative rate of detection are also low.

[0106] The embodiments described above can be further combined or replaced, and the embodiments are only used to describe the preferred embodiments of the present application, and do not limit the concept and scope of the present application. Without departing from the design idea of the present application, various changes and improvements to the technical solutions of the present application made by those skilled in the art all belong to the protection scope of the present application. The protection scope of the present application is given by the appended claims and any equivalents thereof.​

Claims

1. A robust multivariate fault detection method for blast furnace ironmaking based on temporal convolutional networks, characterized in that, The residual modules in the Temporal Convolutional Network (TCN) are stacked to construct the Temporal Convolutional Detection Network (TCDN). TCDN uses causal temporal convolutional networks to extract dynamic information of the process on a time scale. At the same time, it uses the number of channels of the convolutional neural network to transform the spatial dimension of variables. TCDN realizes the fault detection function based on the sum of squares of the reconstruction error of variables. For outliers in the original blast furnace data, the original data with outliers is separated to obtain a sparse outlier matrix, thereby obtaining data without outliers for model training. The method of stacking residual modules in a temporal convolutional network to construct a temporal convolutional detection network is as follows: The residual modules of the Temporal Convolutional Network (TCN) are stacked. Layer, assuming the normalized data of the original blast furnace data is denoted as: The model is set to have a time step of k, therefore the input form of the model's process variables is as follows: , Each residual block yields the corresponding latent variables, in the following order: , ,…, The output of the previous residual module will be used as the input of the next residual module; the objective function of TCDN is to reconstruct the input, that is, to minimize the mean square error (MSE) between the output of the last residual module and the original input, as shown in the following formula: ； Where MSE is the loss function for model training, n is the number of samples, and k is the set time step. For input, This is the output of the last residual module, i.e., the reconstruction result; The TCDN described above utilizes causal temporal convolutional networks to extract dynamic information about the process over time: The convolutional neural network used in TCDN is a one-dimensional convolutional neural network, with the convolution direction being the time dimension of the variable; The method of simultaneously transforming the spatial dimension of variables using the number of channels in a convolutional neural network is as follows: The residual module is composed of a convolutional neural network. The structure of a convolutional neural network requires designing the number of channels, where the original input... The number of channels is the spatial dimension of the variable. The number of channels in the convolutional neural network in the intermediate residual module is designed as follows: ,in The latent variable space achieves compression of the variable space dimension; and the number of channels in the convolutional network in the last residual module is designed to be... That is, to reconstruct the spatial dimension of the original input; The TCDN implements fault detection based on the sum of squares of the reconstruction errors of the variables. Assume the reconstruction result is... ,in The monitoring statistic is designed as the sum of squares of the reconstruction error, i.e. The threshold for monitoring statistics The SPE value for each sample during training is estimated using the kernel density estimation (KDE) method. During online detection, the SPE value calculated for new samples is greater than... If the SPE value calculated for the new sample is less than or equal to 0, then a fault is considered to have occurred; If so, it is considered that no malfunction has occurred in the process; The method described above separates the original data containing outliers, extracting a sparse outlier matrix, and then retains the clean data for model training. Because the process variables contain outliers, the original input... Decomposed into two parts ,in It is a low-rank matrix, and Given a sparse outlier matrix, we first construct the following loss function: ； ； in, Represents the TCDN model. This represents the reconstruction result of TCDN. This represents the part that can be reconstructed by the TCDN model. It is the l1 norm, and It is the F-norm. These are the weight coefficients; the loss function is solved using the ADMM method, which optimizes the loss function using backpropagation. Then fixed Then optimize using the proximal method. Then fixed Continue to optimize This process is repeated until convergence; ultimately, the result is... It is clean data that does not contain outliers, which is used to train the TCDN model for fault detection. The above-mentioned proximal method optimization The proximal method in the text is: ； in Indicates the number of iterations. It is a hyperparameter; Represents the j-th variable In the The value at the next iteration; when an element is separated twice consecutively and the difference is less than 1. At that point, it is directly identified as an outlier and is no longer optimized.

2. The method according to claim 1, characterized in that, The process is as follows: Step 1: Offline modeling; Step 1.1: Collect the process variables from the blast furnace sensors, and record the normalized data as follows: The sample size is The number of variables is Determine the time step k of the process, and construct the input variable form of the neural network. , ; Step 1.2: Based on the aforementioned loss function and solution method, separate and obtain... Then use Training the TCDN model; Step 1.3: Calculate the monitoring statistics for all samples. And based on the kernel density estimation (KDE) method, with confidence level Determine the upper limit of monitoring statistics ; Step 2: Online monitoring; Step 2.1: When there is a new sample of the blast furnace Upon arrival, the data is also normalized, and the normalized data is denoted as... Then, it is input into the TCDN model with parameters trained during offline modeling to obtain the reconstruction result. Calculate the upper limit of its monitoring statistics. : ； Comparative statistics and upper limit of statistics ,if If it is normal, it means there is a malfunction; otherwise, it is in normal operation.

Images