Blast furnace ironmaking monitoring method based on depth variation canonical variable analysis of flow model
By integrating the variational autoencoder, flow model and typical variable analysis, the FDVCVA model is designed, which solves the problem of monitoring complex nonlinear and dynamic characteristics in blast furnace ironmaking process, and realizes efficient abnormal detection and early warning, ensuring the production process Safe and stable.
Patent Information
- Application Number
- CN202410938309.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-07-10
- Filing Date
- 2024-07-13
- Publication Date
- 2024-11-05
AI Technical Summary
There are complex nonlinear and dynamic characteristics in the blast furnace iron smelting process, and traditional monitoring methods are difficult to accurately detect abnormal situations. The deep learning model is not robust to noise and outliers, resulting in unsatisfactory monitoring results.
A method of analysis of depth variation typical variables based on flow model was designed, integrating variation autoencoder, flow model and typical variable analysis, extracting deep dynamic uncertainty features through FDVCVA model, and optimizing model parameters using stochastic gradient descent method , in combination with the calculation control limit of nuclear density estimation, the blast furnace iron smelting process is monitored in real time.
It realizes efficient and robust monitoring performance under highly nonlinear and dynamic characteristics, and can promptly detect and warn of abnormal conditions to ensure the safety and stability of the production process.
Smart Images

Figure CN118897946A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of blast furnace ironmaking process monitoring, and in particular to a blast furnace ironmaking monitoring method based on deep variational canonical variable analysis of a flow model. Background Art
[0002] The blast furnace ironmaking process is one of the most important links in the steel industry, and its efficiency and stability directly affect the quality, cost and environmental impact of steel production. The final product of blast furnace ironmaking is molten iron, and its quality directly affects the performance and market competitiveness of subsequent steel products. Through effective process monitoring, key process parameters such as temperature, pressure, raw material composition, etc. can be tracked and controlled in real time to ensure the stability and consistency of molten iron quality. In addition, accurate process monitoring can detect and correct abnormal conditions in a timely manner, reduce fluctuations and downtime in production, thereby improving overall production efficiency and reducing production costs. More importantly, the high temperature, high pressure and complex chemical reactions in the blast furnace ironmaking process bring potential safety risks. Real-time monitoring can detect and warn abnormal conditions in a timely manner, such as excessive temperature, abnormal pressure, equipment failure, etc., to prevent accidents. Effective process monitoring can not only protect the safe operation of equipment and production systems, but also protect the lives of operators and reduce the frequency of industrial accidents.
[0003] Blast furnace ironmaking is a complex thermochemical process involving a variety of physical and chemical reactions. Raw materials (such as iron ore, coke, and limestone) are treated at high temperatures in a blast furnace to produce molten iron and slag. In this process, there are a large number of variables that need to be monitored, including temperature, pressure, gas composition, and the properties of the raw materials. These variables are not only numerous, but also have complex nonlinear relationships and time-varying properties. Abnormal changes in any variable may lead to a decrease in production efficiency or even cause safety accidents. In addition, in actual production, the influence of measurement equipment and environmental factors makes it inevitable that there are noise and outliers in the data. These interference factors increase the difficulty of process monitoring and make it difficult for traditional monitoring methods to accurately detect abnormal conditions. Noise interference may mask real abnormal signals, while outliers may be mistaken for normal fluctuations in the system. Therefore, how to effectively remove noise and identify outliers has become a key issue in building an efficient monitoring model.
[0004] Traditional process monitoring methods, such as principal component analysis (PCA) and canonical variate analysis (CVA), work well when dealing with linear systems and a small amount of noise. These methods can simplify the data structure and capture the main trend of change through dimensionality reduction and feature extraction. However, for the blast furnace ironmaking process with highly nonlinear and dynamic characteristics, the performance of these methods is insufficient. The blast furnace ironmaking process involves complex thermodynamics and chemical reactions, and the changes in the system state are often nonlinear and have significant time dependence. When faced with such a highly complex nonlinear dynamic system, PCA and CVA are difficult to fully capture the deep-level dynamic uncertainty characteristics of the process, resulting in unsatisfactory monitoring results. This deficiency is manifested in reduced detection sensitivity for abnormal conditions and delayed response to changes in system states, which may miss key warning signals or misreport normal fluctuations as abnormalities. In recent years, with the development of deep learning technology, some monitoring methods based on deep neural networks have been proposed in order to better cope with the complexity and nonlinear characteristics of the blast furnace ironmaking process. Deep neural networks can extract complex patterns and relationships from large amounts of high-dimensional data through multi-layer nonlinear transformations and automatic feature learning. However, deep learning models are not robust enough to noise and outliers. In the blast furnace ironmaking process, noise and outliers are common in the data. These interference factors may lead to unstable model output or even wrong anomaly detection. Therefore, it is extremely important to develop efficient blast furnace ironmaking process monitoring technology to address the above problems. Summary of the invention
[0005] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a blast furnace ironmaking monitoring method based on deep variational canonical variable analysis of flow model.
[0006] The technical solution for realizing the technical purpose of the present invention is as follows:
[0007] A blast furnace ironmaking monitoring method based on deep variational canonical variate analysis of flow model is proposed. By integrating variational autoencoder (VAE), flow model and canonical variate analysis (CVA), the overall structure of FDVCVA model and the corresponding loss function are designed to extract the deep dynamic uncertainty characteristics of blast furnace ironmaking process. Subsequently, the model is trained using stochastic gradient descent (SGD) to optimize its parameters and improve the generalization ability and stability of the model. Then, T is established based on the results of FDVCVA model. 2 Statistics, and use the kernel density estimation method (KDE) to calculate the corresponding control limits, and then monitor the blast furnace ironmaking process in real time.
[0008] According to the method described above, the calculation formula of the FDVCVA model loss function is as follows:
[0009]
[0010] Among them, σ1, σ2 and σ3 are the parameters of the neural network model to be trained, E(·) represents the encoder network, D1(·) and D2(·) represent the decoder network, F(·) is the flow model, and P(w) obeys the standard Gaussian distribution.
[0011] The steps of training the model using stochastic gradient descent (SGD) are:
[0012] Step 1: First, randomly initialize the model parameters, including all weights and bias parameters in the encoder, decoder, and reversible function, usually using uniform distribution or Gaussian distribution;
[0013] Step 2: Standardize the input blast furnace data to make its mean 0 and variance 1 to improve the stability and convergence speed of model training;
[0014] Step 3: Define the loss function of the FDVCVA model;
[0015] Step 4: Process the standardized blast furnace data in batches to improve training efficiency and memory utilization. Each batch of data contains training samples.
[0016] Step 5: For each batch of data, first get the hidden variable z through the encoder network E(·) and a series of flow models, and then reconstruct the past Hankel matrix U through the decoder networks D1(·) and D2(·) p and the future blast furnace data Hankel matrix U f ; At the same time, calculate the reversible function and the corresponding posterior distribution;
[0017] Step 6: Use the current batch of data to calculate the value of the loss function through the results obtained by forward propagation of the model;
[0018] Step 7: Based on the calculated loss function value, use the back propagation algorithm to calculate the gradient of the loss function to each model parameter. The gradient calculation involves the derivation of the parameters of the encoder, decoder, and reversible function.
[0019] Step 8: Use stochastic gradient descent (SGD) to update the model parameters according to the calculated gradients;
[0020] Step 9: Repeat steps 5 to 8 until the preset number of training rounds is reached or the model's loss function converges to a satisfactory level.
[0021] The blast furnace ironmaking process monitoring method process is as follows:
[0022] Step 1: Offline modeling;
[0023] Step 1.1: Train the FDVCVA model;
[0024] Step 1.2: Based on the FDVCVA model results, determine T in the canonical variable latent space 2 The monitoring statistic is given by:
[0025]
[0026] Among them, z k (i) and represents the i-th element of z at time k;
[0027] Step 1.3: Use kernel density estimation to estimate T 2 control limits;
[0028] Step 2: Real-time monitoring;
[0029] Step 2.1: Collect real-time data from the blast furnace ironmaking process and perform standardization. Then, use the trained FDVCVA model to generate real-time T 2 Statistics;
[0030] Step 2.2: Compare this statistic to the control limits to assess the performance of the process. 2 If the statistic exceeds the control limit, it indicates that there may be an abnormal situation; on the contrary, if T 2 If the statistic is below the control limit, the process is normal and no maintenance is required.
[0031] The present invention has the beneficial effects:
[0032] 1. The present invention integrates variational autoencoders, flow models and canonical variable analysis, and designs the overall structure of the FDVCVA model and the corresponding loss function. The present invention can effectively extract the deep dynamic uncertainty characteristics in the blast furnace ironmaking process, so that the monitoring model can still maintain efficient and robust monitoring performance in the face of highly nonlinear and dynamic characteristics.
[0033] 2. Based on the results of the FDVCVA model, the present invention determines T in the typical variable latent space 2 Monitor statistics and calculate control limits using kernel density estimation techniques. By collecting and processing blast furnace ironmaking process data in real time, T 2 Statistics and comparison with control limits can detect and warn abnormal conditions in a timely manner to ensure the safety and stability of the production process. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 Shown is an offline modeling and online monitoring flow chart of a blast furnace ironmaking process monitoring method based on a flow model-based deep variational canonical variate analysis (FDVCVA) of the present invention.
[0035] Figure 2The monitoring results of FDVCVA on suspended material failure in blast furnace ironmaking process. DETAILED DESCRIPTION
[0036] In order to more clearly describe the technical content of the present invention, the specific implementation method is further explained below in conjunction with the accompanying drawings.
[0037] like Figure 1 As shown, it is a schematic diagram of the route of the blast furnace ironmaking monitoring method based on deep variational canonical variable analysis of the flow model of the present invention, and the steps are as follows:
[0038] The invention discloses a blast furnace ironmaking monitoring method based on deep variational canonical variable analysis of a flow model, characterized in that the present invention integrates variational autoencoder (VAE), flow model and canonical variable analysis (CVA) to design the overall structure of the FDVCVA model and the corresponding loss function to extract the deep dynamic uncertainty characteristics of the blast furnace ironmaking process. Subsequently, the model is trained using stochastic gradient descent (SGD) to optimize its parameters and improve the generalization ability and stability of the model. Then, T is established based on the results of the FDVCVA model. 2 Statistics, and use the kernel density estimation method (KDE) to calculate the corresponding control limits, and then monitor the blast furnace ironmaking process in real time.
[0039] The overall structure of the FDVCVA model and the corresponding loss function are:
[0040] The inherent nonlinear and dynamic characteristics of the blast furnace ironmaking process, coupled with the presence of data noise, greatly increase the difficulty of process monitoring. Therefore, in this case, extracting effective feature representation is not only crucial but also essential. First, the standardized historical blast furnace data O is divided into the past blast furnace data Hankel matrix U p and the future blast furnace data Hankel matrix U f The probabilistic latent variable model of the canonical variate analysis (CVA) method defines the joint probability distribution of past and future process random variables, which is expressed as the following formula:
[0041] P(U p ,U f ,z)=P(z)P(U p ∣z)P(U f ∣z)
[0042] P(U p ,U f )=∫P(U p ,U f ,z)dz
[0043] However, the linear observation model P(U p |z) and P(U f|z) has limited representation capabilities, especially when faced with highly nonlinear relationships between data variables in the blast furnace ironmaking process. In order to solve the above problem, we consider using a deep neural network (DNN) method to construct a nonlinear observation model, expressed as P(U p ∣z;Ω p ) and P(U f ∣z;Ω f ). However, this approach brings new challenges. The marginal likelihood P Ω (U p ,U f ) cannot be solved analytically, and the potential blast furnace characteristic variable P(z|U p ) also becomes a problem. Therefore, the variational autoencoder (VAE) method is used to use the conditional density Q(z|U p ; Θ) approximates P(z|U p ), where Θ is the parameter of DNN. The lower bound of the log-likelihood of the marginal data can be obtained as follows:
[0044]
[0045] In the above formula, DNN is used to achieve distribution Q Θ (z∣U p ) as the encoder, and implement As a decoder. Therefore, the model can be trained by maximizing the above variational lower bound. In the conventional VAE framework, the Gaussian distribution assumption is conducive to the mathematical processing of the model, but it also limits the model's ability to handle data complexity. In order to resolve this contradiction and enhance the representation ability of the model, a model called unconditional flow is integrated into the existing framework to establish a more powerful posterior distribution. The posterior distribution Q(z|U p ) is not a simple Gaussian distribution, but is constructed as follows:
[0046] Q(z|U p )=∫δ(zF U (w))P(w)dw
[0047] Among them, δ(·) is the Dirac delta function, and P(w) obeys the standard Gaussian distribution. is a reversible function of the variable w, whose parameters are given by U p The new variational lower limit can be obtained by calculating the function of .
[0048]
[0049] Specifically, in order to facilitate the practice in the blast furnace ironmaking process, the reversible function and the posterior distribution P(U p |z) and P(Uf |z) is set as follows:
[0050]
[0051] Among them, σ1, σ2 and σ3 are the parameters of the neural network model to be trained. E(·) represents the encoder network, D1(·) and D2(·) represent the decoder network. The calculation formula of the final FDVCVA model loss function is as follows:
[0052]
[0053] The steps of training the model using stochastic gradient descent (SGD) are:
[0054] Step 1: First, randomly initialize the model parameters, including all weights and bias parameters in the encoder, decoder, and reversible function, usually using a uniform distribution or a Gaussian distribution.
[0055] Step 2: Standardize the input blast furnace data to make its mean 0 and variance 1 to improve the stability and convergence speed of model training.
[0056] Step 3: Define the loss function of the FDVCVA model according to the formula.
[0057] Step 4: Process the standardized blast furnace data in batches (mini-batches) to improve training efficiency and memory utilization. Each batch of data contains several training samples.
[0058] Step 5: For each batch of data, first obtain the hidden variable z through the encoder network E(·) and a series of flow models, and then reconstruct U through the decoder networks D1(·) and D2(·) p and the future blast furnace data Hankel matrix U f At the same time, the reversible function is calculated and the corresponding posterior distribution.
[0059] Step 6: Use the current batch of data and the results obtained through the forward propagation of the model to calculate the value of the loss function.
[0060] Step 7: Based on the calculated loss function value, use the back propagation algorithm to calculate the gradient of the loss function to each model parameter. Gradient calculation involves the derivation of the parameters of the encoder, decoder, and reversible function.
[0061] Step 8: Use stochastic gradient descent (SGD) to update the model parameters according to the calculated gradients.
[0062] Step 9: Repeat steps 5 to 8 until the preset number of training epochs is reached or the loss function of the model converges to a satisfactory level.
[0063] The specific process is as follows:
[0064] Step 1: Offline modeling;
[0065] Step 1.1: Train the FDVCVA model.
[0066] Step 1.2: Based on the FDVCVA model results, determine T in the canonical variable latent space 2 The monitoring statistics can be given by the following formula:
[0067]
[0068] Among them, z k (i) and represents the i-th element of Z at time k.
[0069] Step 1.3: Use Kernel Density Estimation (KDE) to estimate T 2 control limits.
[0070] Step 2: Real-time monitoring;
[0071] Step 2.1: Collect real-time data from the blast furnace ironmaking process and perform normalization. Then, use the trained FDVCVA model to generate real-time T 2 Statistics.
[0072] Step 2.2: Compare this statistic to the control limits to assess the performance of the process. 2 If the statistic exceeds the control limit, it indicates that there may be an abnormal situation; on the contrary, if T 2 If the statistic is below the control limit, the process is normal and no maintenance is required.
[0073] Substituting the present invention into actual industrial data for verification, we take a certain ironmaking plant with a volume of 2650m 3 The data sample of No. 2 blast furnace in 2023 contains 14 parameters and the sampling rate is 10s. 5000 normal samples are used for training and tested on the overhanging material fault. Figure 2 The figure shows the monitoring effect of the method of the present invention on suspended material faults. From the fault monitoring results, it can be seen that the model effect is very good.
[0074] The technical features of the above-described embodiments can be further combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0075] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, all of which belong to the protection scope of the present invention. The protection scope of the present invention is given by the attached claims and any equivalent technical solutions thereof.
Claims
1. A blast furnace ironmaking monitoring method based on deep variational canonical variable analysis of flow model, characterized in that: By integrating variational autoencoder (VAE), flow model and canonical variate analysis (CVA), the overall structure of the FDVCVA model and the corresponding loss function are designed to extract the deep dynamic uncertainty characteristics of the blast furnace ironmaking process. Subsequently, the model is trained using stochastic gradient descent (SGD) to optimize its parameters and improve the generalization ability and stability of the model. Then, based on the results of the FDVCVA model, T 2 Statistics, and use the kernel density estimation method (KDE) to calculate the corresponding control limits, and then monitor the blast furnace ironmaking process in real time.
2. The method according to claim 1, characterized in that The calculation formula of the FDVCVA model loss function is as follows: Among them, σ1, σ2 and σ3 are the parameters of the neural network model to be trained, E(·) represents the encoder network, D1(·) and D2(·) represent the decoder network, F(·) is the flow model, and P(w) obeys the standard Gaussian distribution.
3. The method according to claim 2, characterized in that The steps of training the model using stochastic gradient descent (SGD) are: Step 1: First, randomly initialize the model parameters, including all weights and bias parameters in the encoder, decoder, and reversible function, usually using uniform distribution or Gaussian distribution; Step 2: Standardize the input blast furnace data to make its mean 0 and variance 1 to improve the stability and convergence speed of model training; Step 3: Define the loss function of the FDVCVA model; Step 4: Process the standardized blast furnace data in batches to improve training efficiency and memory utilization. Each batch of data contains training samples. Step 5: For each batch of data, first get the hidden variable z through the encoder network E(·) and a series of flow models, and then reconstruct the past Hankel matrix U through the decoder networks D1(·) and D2(·) p and the future blast furnace data Hankel matrix U f ; At the same time, calculate the reversible function and the corresponding posterior distribution; Step 6: Use the current batch of data to calculate the value of the loss function through the results obtained by forward propagation of the model; Step 7: Based on the calculated loss function value, use the back propagation algorithm to calculate the gradient of the loss function to each model parameter. The gradient calculation involves the derivation of the parameters of the encoder, decoder, and reversible function. Step 8: Use stochastic gradient descent (SGD) to update the model parameters according to the calculated gradients; Step 9: Repeat steps 5 to 8 until the preset number of training rounds is reached or the model's loss function converges to a satisfactory level.
4. The method according to claim 3, characterized in that The blast furnace ironmaking process monitoring method process is as follows: Step 1: Offline modeling; Step 1.1: Train the FDVCVA model; Step 1.2: Based on the FDVCVA model results, determine T in the canonical variable latent space 2 The monitoring statistic is given by: Among them, z k (i) and represents the i-th element of Z at time k; Step 1.3: Use kernel density estimation to estimate T 2 control limits; Step 2: Real-time monitoring; Step 2.1: Collect real-time data from the blast furnace ironmaking process and perform standardization. Then, use the trained FDVCVA model to generate real-time T 2 Statistics; Step 2.2: Compare this statistic to the control limits to assess the performance of the process. 2 If the statistic exceeds the control limit, it indicates that there may be an abnormal situation; on the contrary, if T 2 If the statistic is below the control limit, the process is normal and no maintenance is required.
