Molten iron silicon content real-time intelligent regulation and control method based on edge calculation

Abstract

The invention relates to the technical field of industrial process control, and discloses a molten iron silicon content real-time intelligent regulation and control method based on edge calculation, which comprises the following steps: firstly, in an off-line stage, constructing a thermodynamic constrained variational auto-encoder model, learning a low-dimensional thermodynamic state manifold for describing an internal mechanism of a process from high-dimensional historical data, and calculating the scalar curvature of the manifold to quantify the local stability of each working condition. And then, in an online tracking stage, a curvature-adaptive particle filtering algorithm is operated at an edge node, and process noise is dynamically adjusted according to the curvature of the current state, so that robust tracking of the molten iron state is realized. And finally, in an intelligent regulation and control stage, a model prediction control framework is adopted in a central controller, and geometric risk items are innovatively introduced into a cost function of the model prediction control framework. According to the method, the unstable working condition is avoided by guiding the operation track, and accurate control over key quality indexes and active maintenance of the overall stability of the production process are achieved.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, specifically to a real-time intelligent control method for silicon content in molten iron based on edge computing. Background Technology

[0002] Blast furnace ironmaking is a core component of modern integrated iron and steel enterprises. It is a complex industrial process characterized by large time delays, strong coupling, multiple variables, and high nonlinearity. Achieving safe, stable, long-term, and efficient operation of blast furnaces, and continuously producing high-quality molten iron, is a relentless goal pursued by the metallurgical industry.

[0003] To achieve effective control of blast furnaces, their regulation technology has evolved from relying on manual experience to model-based automated control, and then to data-driven intelligent regulation. In the early days, blast furnace operation mainly relied on experienced operators for manual adjustments. This method was highly subjective, unstable, and difficult to adapt to complex and changing furnace conditions, and it was also difficult to pass on and promote.

[0004] With the development of automation technology, advanced control strategies, such as model predictive control (MPC), have been introduced into blast furnace control. These methods attempt to predict the future dynamic behavior of the system and optimize control by establishing process models that describe the internal mechanisms of the blast furnace. However, the blast furnace interior is a "black box" with extremely complex physicochemical reactions, and its internal state is difficult to measure directly. Constructing a mechanism model that can accurately reflect the dynamic characteristics of the entire process is extremely difficult, which greatly limits the application effectiveness of traditional model predictive control.

[0005] In recent years, artificial intelligence and big data technologies have offered new solutions to this problem. Using methods such as deep learning, complex mapping relationships between blast furnace inputs (e.g., blast volume, coal ratio) and outputs (e.g., molten iron composition, temperature) can be extracted from massive amounts of historical data to construct data-driven predictive models. While these methods have improved the accuracy of predicting key indicators to some extent, existing solutions still have deep-seated shortcomings. On the one hand, many data-driven models merely establish a "shallow" mapping from high-dimensional sensor data to key outputs, failing to delve into the low-dimensional dynamic structures hidden behind the data that truly characterize the system's internal operating state. On the other hand, and more critically, existing control strategies often focus only on tracking the set values ​​of production indicators, representing a "goal-oriented" rather than "risk-averse" control approach. These methods lack an effective mechanism to quantify and predict the stability of current or future furnace conditions, and cannot identify those seemingly stable but abruptly changeable "high-risk" state regions. Therefore, a control decision that appears optimal in the short term may inadvertently push the entire system to the edge of a highly nonlinear and unpredictable unstable operating condition, thereby increasing the risk of anomalies or even accidents. This constitutes a technical bottleneck that urgently needs to be addressed in the field of intelligent blast furnace control. Summary of the Invention

[0006] The technical problem this invention aims to solve is that existing methods for controlling the silicon content in molten iron suffer from several issues: reliance on offline testing leading to control lag; difficulty in accurately describing complex metallurgical processes using mechanistic or statistical models; and insufficient adaptability of conventional artificial intelligence-based prediction models to changes in operating conditions and lack of transparency regarding their internal mechanisms.

[0007] To address the aforementioned technical problems, this invention provides a new technical solution.

[0008] The first aspect of this invention provides a method for real-time intelligent control of silicon content in molten iron based on edge computing, characterized by comprising the following steps:

[0009] a) Offline build phase:

[0010] Collect historical data, including high-frequency process data, low-frequency quality data, and operating parameter data;

[0011] Using the historical data, a thermodynamically constrained variational autoencoder is trained to construct a thermodynamic state manifold that maps the high-frequency process data to a low-dimensional latent space; the points in the thermodynamic state manifold have a mapping relationship with the low-frequency mass data.

[0012] Calculate and store the geometric properties of the thermodynamic state manifold, the geometric properties including at least scalar curvature;

[0013] b) Online tracking phase:

[0014] At the edge computing node, the high-frequency process data is acquired in real time;

[0015] The particle filtering algorithm is used to track the state of molten iron in real time on the thermodynamic manifold to obtain the estimated state of molten iron at the current moment.

[0016] c) Intelligent control stage:

[0017] Based on the estimated state of the molten iron and the scalar curvature, the optimal control command is calculated using a model predictive controller.

[0018] The operating parameters of the blast furnace are adjusted according to the optimal control command.

[0019] In one specific embodiment, in step a), the training of the thermodynamically constrained variational autoencoder is achieved by minimizing a composite loss function, which includes reconstruction loss, KL divergence loss, physical supervision loss, and thermodynamic consistency loss. This design aims to ensure that the thermodynamic state manifold not only accurately reproduces the input process data, but also that its topology conforms to preset physicochemical laws and is aligned with real physical quantities.

[0020] Preferably, the thermodynamic consistency loss is calculated as follows:

[0021]

[0022] in, For thermodynamic consistency loss, This represents the predicted silicon content at time t. C is the predicted temperature of the molten iron at time t. A and C B is a preset thermodynamic constant. This loss term applies the simplified thermodynamic equilibrium relationship of the silicon reduction reaction to the model training process, so that the generated manifold naturally contains this physical constraint.

[0023] Preferably, in step b), the process noise covariance used by the particle filtering algorithm during the prediction step is adaptively adjusted based on the scalar curvature of the particle's location. The scalar curvature characterizes the degree of local bending of the thermodynamic manifold, with high curvature regions corresponding to regions where the system state undergoes significant nonlinear changes.

[0024] In one specific embodiment, the process noise covariance Q t The calculation method is as follows:

[0025] Q t (z)=Q base Q t(z)=Q base ·exp(γ·R(z));

[0026] Where z is the position of the particle on the thermodynamic manifold, R(z) is the scalar curvature at position z, and Q base The base process noise covariance matrix is ​​γ, where γ is a preset curvature sensitivity coefficient. In this way, when particles enter a high curvature region, the process noise increases, increasing the dispersion of the particle cloud and thus improving the tracking robustness of the filtering algorithm when the system state undergoes nonlinear transitions.

[0027] In one specific embodiment, in step b), during the update step, the particle filtering algorithm decodes the particle positions on the thermodynamic state manifold into theoretical high-frequency process data using the decoder in the thermodynamically constrained variational autoencoder, and compares this data with the real-time high-frequency process data to calculate particle weights. This step utilizes the decoder as an observation model, establishing a direct connection between the state space and the observation space.

[0028] Preferably, in step c), the cost function of the model prediction controller includes a geometric risk term, which is related to the scalar curvature of each point on the future prediction path.

[0029] In one specific embodiment, the cost function J(U) of the model predictive controller seq Specifically:

[0030]

[0031] Among them, U seq For control sequences, y is a prediction of the physical quantity at a future time. ref Set the target value, Δu t+k To control the rate of change of the quantity, For the prediction of the scalar curvature at the future state position, λ R Risk avoidance weights are introduced. By introducing the geometric risk term, the optimal control command obtained from the optimization solution tends to choose a path that causes the system state to evolve along the low curvature region of the thermodynamic manifold, thereby improving the stability of the control process.

[0032] Preferably, the low-frequency quality data includes the silicon content and temperature of the molten iron; the high-frequency process data is selected from at least one of flame spectral data, infrared thermal imaging data, and furnace throat acoustic signals.

[0033] A second aspect of the present invention provides a real-time intelligent control system for silicon content in molten iron based on edge computing, characterized in that it comprises:

[0034] Offline training modules, configured in the cloud or on a server, are used for:

[0035] Collect historical data, including high-frequency process data, low-frequency quality data, and operating parameter data;

[0036] A thermodynamically constrained variational autoencoder is trained based on the historical data to construct a thermodynamic state manifold that maps the high-frequency process data to a low-dimensional latent space.

[0037] Calculate and store the geometric properties of the thermodynamic state manifold, the geometric properties including at least scalar curvature;

[0038] Edge computing nodes, configured at the blast furnace production site, are used for:

[0039] Deploy the thermodynamic manifold and the geometric properties;

[0040] Acquire real-time data of the high-frequency process;

[0041] The particle filtering algorithm is run to track the state of molten iron in real time on the thermodynamic manifold and generate an estimate of the state of molten iron at the current moment.

[0042] The central control module is used for:

[0043] Receive the estimated state of the molten iron generated by the edge computing node and the scalar curvature corresponding to the state position;

[0044] The model predictive controller is run to calculate the optimal control command based on the received information.

[0045] The optimal control command is sent to the actuator of the blast furnace to adjust the operating parameters.

[0046] This invention provides a real-time intelligent control method for silicon content in molten iron based on edge computing. It has the following beneficial effects:

[0047] 1. This invention constructs a thermodynamic state manifold and introduces thermodynamic consistency loss to constrain model training. This ensures that the dimensionality-reduced low-dimensional state is not only a compression of high-dimensional data but also a representation of the intrinsic physicochemical state of the blast furnace smelting process. This gives the state estimation results clear physical meaning, significantly improving the reliability of state representation and the effectiveness of subsequent control compared to purely data-driven black-box models.

[0048] 2. This invention introduces the geometric properties (specifically, scalar curvature) of the thermodynamic manifold into the online tracking stage. By adaptively adjusting the process noise covariance in the particle filter algorithm using scalar curvature, the state tracking system can perceive the nonlinear intensity and stability of the current operating condition in real time. When the system enters a region of drastic state change, this method can automatically improve the alertness and robustness of the tracking, thereby more accurately capturing the change in operating condition and avoiding tracking failures caused by model mismatch.

[0049] 3. In the intelligent control stage, this invention incorporates the geometric risk term of the future predicted path (quantized by scalar curvature) into the cost function of the model predictive controller. This design ensures that the controller's optimization objective includes not only tracking error and control cost, but also forward-looking considerations for future process stability. The controller proactively calculates and selects a control path that avoids high-curvature unstable regions, achieving a shift from passive response to proactive risk avoidance, thereby improving the stability of the entire closed-loop control system. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0051] Figure 2 This is a schematic diagram of the thermodynamically constrained variational autoencoder structure of the present invention;

[0052] Figure 3 This is a schematic diagram of the online tracking phase process of the present invention;

[0053] Figure 4 This is a schematic diagram of the intelligent control stage process of the present invention;

[0054] Figure 5 This is a schematic diagram of the thermodynamic state manifold and curvature map of the present invention.

[0055] Among them, 10 is the offline training module; 20 is the edge computing node; 30 is the central control module; 40 is the historical database; and 50 is the blast furnace actuator. Detailed Implementation

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] See attached document Figure 1 , Figure 1This is a schematic diagram of the architecture of a real-time intelligent control system for silicon content in molten iron according to an embodiment of the present invention. The system provided by the present invention may include an offline training module 10, an edge computing node 20, and a central control module 30.

[0058] The offline training module 10 is configured in the cloud or on a server and has the computing resources required to perform large-scale data processing and complex model training. This module is connected to a database 40 that stores historical data, including high-frequency process data, low-frequency quality data, and operating parameter data collected during the blast furnace's historical production cycle.

[0059] The offline training module 10 performs the offline construction phase of the method of this invention. Specifically, this module reads the historical data to train a thermodynamically constrained variational autoencoder. After training, the module uses the trained model to generate a thermodynamic manifold characterizing the blast furnace smelting process and calculates the geometric properties of the manifold, including at least the scalar curvature at each point, forming a curvature map. Finally, the offline training module 10 packages the trained model parameters (including the decoder and physical quantity predictor) with the curvature map and deploys them to the edge computing node 20.

[0060] The edge computing node 20 is configured at the blast furnace production site, physically close to the sensors used to collect high-frequency process data. This node receives and stores model parameters and curvature maps deployed by the offline training module 10. During production, the edge computing node 20 acquires high-frequency process data from the blast furnace sensors in real time, without any delay processing.

[0061] The edge computing node 20 performs the online tracking phase of the method of this invention. This node runs a particle filtering algorithm on its local processor. This algorithm utilizes real-time acquired high-frequency process data, combined with a deployed model and curvature map, to track the current state of the molten iron on the thermodynamic state manifold in real time. The output of this tracking process is an estimate of the molten iron state at the current moment, and the scalar curvature corresponding to the location of this state estimate. Subsequently, the edge computing node 20 reports the comprehensive state information, including the state estimate and the scalar curvature, to the central control module 30.

[0062] The central control module 30 can be integrated into the factory's existing distributed control system (DCS) or programmable logic controller (PLC) system, or it can be a standalone computing unit. This module receives comprehensive status information reported from the edge computing nodes 20.

[0063] The central control module 30 performs the intelligent control phase of the method of this invention. Internally, this module runs a model prediction controller. This controller takes the received estimated state of the molten iron and scalar curvature as input, combined with a preset production target (such as the target silicon content y). ref The optimal control command sequence is calculated by solving a cost function that includes a geometric risk term.

[0064] The central control module 30 sends the first control action of the calculated optimal control command sequence to the blast furnace actuator 50 via the communication interface. The actuator 50 adjusts the blast furnace's operating parameters (such as blast temperature and pulverized coal injection rate) according to the received command. This adjustment causes a change in the blast furnace's state, which is then captured in real-time by the sensors of the edge computing node 20, thus forming a complete closed-loop control process.

[0065] See attached document Figure 2 , Figure 2 This is a schematic diagram of a thermodynamically constrained variational autoencoder structure according to an embodiment of the present invention. The purpose of the offline construction stage described in this invention is to train the encoder to construct a thermodynamic manifold capable of characterizing the physical state of the blast furnace smelting process and quantifying its geometric properties.

[0066] The first step in this stage is data preparation. The collected historical data is divided into three categories: high-frequency process data, low-frequency quality data, and operating parameter data.

[0067] High-frequency process data This refers to the data collected by the blast furnace sensor at time t, with dimension D. x The data vector. In one embodiment, the data may include blast furnace taphole flame spectral data (i.e., a vector composed of light intensity values ​​at multiple preset wavelengths), infrared thermal imaging data (i.e., a vector composed of statistical features such as temperature distribution histogram, mean, and variance extracted from a thermodynamic image), or furnace throat acoustic signal data (i.e., a spectral feature vector obtained after Fourier transform).

[0068] Low-frequency quality data This refers to high-frequency process data x t The time-aligned, delayed acquisition of physicochemical test results for molten iron has a dimension of D. y In one embodiment, this data includes the silicon content [Si] of the molten iron. t and molten iron temperature T t .

[0069] Operating parameter data Includes data synchronized with the above, with dimension D. u Controllable operational variables, such as supply air temperature, supply air humidity, oxygen enrichment rate, and pulverized coal injection rate.

[0070] Before training, the collected data needs to be preprocessed, including: aligning the three types of data using timestamps; handling missing values ​​in the data using linear interpolation or mean filling methods; and normalizing all data, for example, by linearly scaling them to the [0,1] interval to eliminate the influence of different physical units on model training.

[0071] After data preparation, a thermodynamically constrained variational autoencoder is constructed and trained. This model consists of three neural network modules: an encoder q... φ A process decoder p θ and a physical quantity predictor f ψ .

[0072] encoder q φ The function is to input high-frequency process data x t Mapped to a low-dimensional latent variable (where D) z <D x The posterior probability distribution of ). This distribution is set as a diagonal Gaussian distribution with a mean vector μ. z Sum of variance vectors Output from the encoder network. This process can be represented as:

[0073]

[0074] Where φ represents the parameters of the encoder network.

[0075] Process decoder p θ The function is to extract a point z from the low-dimensional latent space. t The original high-frequency process data is reconstructed to obtain the reconstructed data. This process is a deterministic mapping, which can be represented as:

[0076]

[0077] Where d(·) is the decoder network, and its parameter is θ.

[0078] Physical quantity predictor f ψ The function is to extract a point z from the low-dimensional latent space. t The corresponding low-frequency quality data is predicted, and the predicted physical quantity is obtained. This process is also a deterministic mapping, which can be represented as:

[0079]

[0080] Where f(·) is the predictor network with parameter ψ.

[0081] The model is trained by minimizing a composite loss function using optimization algorithms such as stochastic gradient descent on the dataset. To achieve this, the function consists of four weighted terms:

[0082]

[0083] Where β1, β2, and β3 are the weight coefficients of each term, which are preset hyperparameters. Reconstruction Loss Used to measure reconstructed data Compared with the original input data x t The difference between them is calculated using the following formula:

[0084]

[0085] This loss term ensures the latent variable z t It can preserve the main information in the original data. KL divergence loss This is used to regularize the latent space, thereby improving the posterior distribution q of the encoder output. φ (z t |x t Approximates a standard normal distribution prior. The calculation formula is as follows:

[0086]

[0087] This loss term gives the latent space a good topological structure, such as continuity and smoothness. Physically supervised loss. Using real low-frequency quality data y t For physical quantity predictor f ψ The output is supervised, and its calculation formula is:

[0088]

[0089] This loss term ensures that there is an accurate mapping relationship between points in the latent space and the actual physicochemical indicators.

[0090] Thermodynamic consistency loss Prior knowledge from the field of metallurgy is applied to the model as a soft constraint, and its calculation formula is as follows:

[0091]

[0092] in, This represents the predicted silicon content at time t. These are the predicted values ​​of the molten iron temperature at time t; they are predicted physical quantities. Components of. C A and C BThis is a constant pre-defined based on the thermodynamic equilibrium relationship of the silicon reduction reaction. This loss term penalizes predictions that do not conform to this physical law, thus embedding physical constraints into the geometry of the thermodynamic state manifold. After the model training converges, its low-dimensional latent space Z constitutes the thermodynamic state manifold described in this invention. The final step is to calculate the geometric properties of this manifold. First, the Riemannian metric tensor G(z) at any point z on the manifold is calculated. This tensor is obtained from the Jacobian matrix J of the process decoder d(z; θ). d (z) is calculated to yield:

[0093] G(z)=J d (z) T J d (z);

[0094] in,

[0095] Then, based on the calculated Riemannian metric tensor G(z) and its first and second partial derivatives, the scalar curvature R(z) at each point on the manifold is calculated using standard differential geometry formulas. This scalar curvature R(z) is a scalar value that quantifies the intrinsic curvature of the manifold near point z and is used to characterize the local nonlinearity and instability of the blast furnace smelting state corresponding to that point.

[0096] Finally, the trained model parameters θ,ψ (i.e., the process decoder and physical quantity predictor) are combined with the calculated scalar curvature data covering the entire manifold region. Figure 1 It is stored for subsequent online tracking and intelligent control.

[0097] See attached document Figure 3 , Figure 3 This is a flowchart illustrating the online tracking stage according to an embodiment of the present invention. This stage is executed on the edge computing node 20 deployed at the blast furnace production site, and its core is to run a curvature-adaptive particle filter algorithm to achieve real-time tracking of the molten iron state.

[0098] First, before the algorithm starts running, the edge computing node 20 receives and loads the required data and model from the offline training module 10. This includes: the pre-trained process decoder p θ The parameter θ (represented by the function d(z; θ)) and the pre-calculated scalar curvature map R(z) covering the entire thermodynamic manifold.

[0099] The algorithm is executed step-by-step. At the initial time t=0, the algorithm initializes. This step involves initializing from a prior distribution (e.g., the standard normal distribution). Mid-sampling generates a set of particles Where N is the total number of particles, which is a preset parameter. Each particle This represents an assumption about the initial state of the blast furnace and is assigned equal initial weights.

[0100] At any time t>0, the particle filter algorithm executes cyclically according to the prediction step and the update step.

[0101] Propagation step:

[0102] The purpose of this step is to, based on the system model, determine the particle state from the previous moment. Propagate forward to obtain the predicted particle state at the current moment. In this embodiment, the state transition model is set as a random walk model on the manifold:

[0103]

[0104] in, It is process noise sampled from a Gaussian distribution with zero mean, i.e.

[0105] The key here is the covariance matrix Q of the process noise. t It is not a fixed value, but depends on the particle's current position. The scalar curvature is adaptively adjusted. The specific calculation method is as follows:

[0106] Q t (z)=Q base ·exp(γ·R(z));

[0107] Where z represents the particle's position on the manifold; R(z) is the scalar curvature value at position z obtained by querying the deployed curvature map; Q base γ is a preset basic process noise covariance matrix, representing the baseline uncertainty of the system in the flattest region (i.e., the region where R(z) approaches 0); γ ​​is a preset curvature sensitivity positive coefficient, used to adjust the sensitivity of the noise covariance to curvature changes.

[0108] Through this mechanism, when a particle When located in the high curvature region of the manifold (i.e. (The value is relatively large), and its corresponding process noise covariance Q t This also increases accordingly. This allows the particle to perform a larger random walk in the prediction step, increasing the dispersion range of the particle cloud, thereby improving the filter's ability to capture the state when the system state undergoes rapid or highly nonlinear changes.

[0109] Update step:

[0110] The purpose of this step is to utilize the high-frequency process data x acquired in real time at the current time t. t This is used to correct the particle set obtained in the prediction step and update the weight of each particle.

[0111] First, for each particle after the prediction step propagation... Using the deployed process decoder d(z; θ) as the observation model, the theoretical high-frequency process data corresponding to the particle state are calculated.

[0112] Then, by comparing theoretical observations Compared with the actual measured value x t The likelihood of a particle is calculated based on the degree of consistency between the values. Reflects particles Capable of generating real observation data x t The probability of [the likelihood]. In a specific embodiment, this likelihood value can be given by a Gaussian probability density function:

[0113]

[0114] in, It is a preset measurement noise variance hyperparameter.

[0115] Next, the weights of each particle are updated based on the calculated likelihood values:

[0116]

[0117] After updating the weights of all particles, the weights are normalized to obtain a new weight distribution.

[0118]

[0119] At this point, the updated particle set is Its corresponding weight is

[0120] To avoid particle degeneration (i.e., a few particles having much higher weights than others) after multiple iterations, the algorithm also includes a resampling step. This is achieved by calculating the effective number of particles. When N eff When the weights fall below a preset threshold, resampling is initiated. This process is based on the updated weights. For particle sets Perform random sampling with replacement to generate a new set of N particles, and reset all their weights to 1 / N.

[0121] After the update step (or resampling step) is completed, the final estimate of the state of the molten iron at the current time t is... It is calculated by a weighted average of all particles:

[0122]

[0123] Finally, edge computing node 20 will calculate the state estimate. And the scalar curvature corresponding to that location obtained by querying the curvature map. The information is packaged into comprehensive status information and sent to the central control module 30 for subsequent intelligent control decisions.

[0124] See attached document Figure 4 , Figure 4 This is a flowchart illustrating the intelligent control stage according to an embodiment of the present invention. This stage is executed in the central control module 30, which adopts a model predictive control (MPC) framework and generates and issues optimal blast furnace operation instructions through rolling optimization calculations.

[0125] At time t at the beginning of each control cycle, the central control module 30 first receives comprehensive status information reported by the edge computing node 20, which includes the final estimate of the current state of the molten iron. and the scalar curvature corresponding to the location of this state At the same time, the module obtains the preset production target from the system, such as the target silicon content y in molten iron. ref .

[0126] The core of a model predictive controller is to solve an optimization problem at each time step t to find the optimal sequence of control actions over a future period. The controller's operation depends on two preset parameters: the prediction time domain H... p It defines the number of time steps the controller takes to predict the future behavior of the system; and the control time domain H. c It defines the number of control actions that need to be optimized (where H) c ≤H p ).

[0127] At time t, the controller needs to predict the time from t+1 to t+H. p The system state. This prediction requires a state transition model that describes the system state given the current state. and control input u t+k Under these circumstances, the state at the next moment The model g(·) can be trained offline along with the main model, and its form is:

[0128]

[0129] in, By iterating this model repeatedly, the controller can obtain the candidate control sequence. Under its influence, until t+H p Complete state prediction trajectory at time step The goal of this optimization is to minimize a comprehensive cost function J(U t In one embodiment of the invention, the cost function consists of three terms:

[0130]

[0131] The first term is the tracking error term. Among them, It is low-frequency quality data (such as silicon content) predicted from time t to time t+k in the future. It is obtained by predicting the state. Input to the deployed physical quantity predictor f ψ Obtained from, that is This item calculates the predicted output and the reference target y. ref The weighted norm squared between the values ​​is used to penalize deviations from the target. W is a pre-defined positive semi-definite weight matrix used to adjust the tracking priority of different output variables.

[0132] The second term is the control cost term. Where, u t+k This is the control action vector applied at time t+k. This term calculates the weighted norm square of the control action sequence to suppress excessively large or drastic control operations, thereby ensuring the smoothness of the production process. S is a preset positive definite weight matrix.

[0133] The third item is the geometric risk item, which is a key part of this invention. It is obtained by querying a curvature map and predicting the future state. The scalar curvature value at the location. This term is a weighted sum of the scalar curvatures of all points along the future predicted path. Its function is to quantify and penalize control sequences that lead the system to high curvature regions in the thermodynamic manifold. λ R It is a preset risk weight coefficient greater than zero, used to balance the relationship between tracking performance and path stability.

[0134] At each time t, the central control module 30 solves the following problem using a numerical optimization algorithm to obtain the optimal control sequence.

[0135]

[0136] The solution to this optimization problem is constrained by operational constraints, such as the upper and lower limits of the air supply temperature and pulverized coal injection rate. The optimal control sequence is obtained... Then, the controller only processes the first element of the sequence. As the final control command at the current moment, it is sent to the blast furnace actuator 50 through the communication interface.

[0137] At the next time t+1, the system enters a new state, and the central control module 30 will receive the new state estimate. Based on this new information, the entire calculation process from prediction to optimization is repeated to calculate the new optimal control command. This rolling execution method, which re-evaluates and optimizes at each time step, constitutes a closed-loop feedback control system, enabling it to continuously maintain the blast furnace operating status near the target and actively avoid entering unstable operating conditions.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time intelligent control of silicon content in molten iron based on edge computing, characterized in that, Includes the following steps: a) Offline build phase: Collect historical data, including high-frequency process data, low-frequency quality data, and operating parameter data; Using the historical data, a thermodynamically constrained variational autoencoder is trained to construct a thermodynamic state manifold that maps the high-frequency process data to a low-dimensional latent space; the points in the thermodynamic state manifold have a mapping relationship with the low-frequency mass data. Calculate and store the geometric properties of the thermodynamic state manifold, the geometric properties including at least scalar curvature; b) Online tracking phase: At the edge computing node, the high-frequency process data is acquired in real time; The particle filtering algorithm is used to track the state of molten iron in real time on the thermodynamic manifold to obtain the estimated state of molten iron at the current moment. c) Intelligent control stage: Based on the estimated state of the molten iron and the scalar curvature, the optimal control command is calculated using a model predictive controller. The operating parameters of the blast furnace are adjusted according to the optimal control command.

2. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, In step a), the training of the thermodynamically constrained variational autoencoder is achieved by minimizing a composite loss function, which includes reconstruction loss, KL divergence loss, physical supervision loss, and thermodynamic consistency loss.

3. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 2, characterized in that, The thermodynamic consistency loss is calculated as follows: in, This is due to the loss of thermodynamic consistency. This is the predicted value of silicon content at time t; C is the predicted temperature of the molten iron at time t. A and C B These are preset thermodynamic constants.

4. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, In step b), the process noise covariance used by the particle filtering algorithm during the prediction step is adaptively adjusted according to the scalar curvature at the particle's location.

5. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, The process noise covariance Q t The calculation method is as follows: Q t (z)=Q base ·exp(γ·R(z)); Where z is the position of the particle on the thermodynamic manifold; R(z) is the scalar curvature at position z; Q base γ is the covariance matrix of the basic process noise; γ is the preset curvature sensitivity coefficient.

6. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, In step b), when performing the update step, the particle filtering algorithm decodes the position of the particle on the thermodynamic state manifold into theoretical high-frequency process data through the decoder in the thermodynamically constrained variational autoencoder, and compares it with the real-time high-frequency process data to calculate the particle weight.

7. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, In step c), the cost function of the model prediction controller includes a geometric risk term, which is related to the scalar curvature of each point on the future prediction path.

8. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 7, characterized in that, The cost function J(U) of the model predictive controller seq Specifically: Among them, U seq For control sequences, y is a prediction of the physical quantity at a future time. ref Set the target value, Δu t+k To control the rate of change of the quantity, For the prediction of the scalar curvature at the future state position, λ R Weighting for risk aversion.

9. The method for real-time intelligent control of silicon content in molten iron based on edge computing according to claim 1, characterized in that, The low-frequency quality data includes the silicon content and temperature of the molten iron; the high-frequency process data is selected from at least one of flame spectral data, infrared thermal imaging data, and furnace throat acoustic signals.

10. A real-time intelligent control system for silicon content in molten iron based on edge computing, applied to the real-time intelligent control method for silicon content in molten iron based on edge computing as described in any one of claims 1-9, characterized in that, include: Offline training modules, configured in the cloud or on a server, are used for: Collect historical data, including high-frequency process data, low-frequency quality data, and operating parameter data; A thermodynamically constrained variational autoencoder is trained based on the historical data to construct a thermodynamic state manifold that maps the high-frequency process data to a low-dimensional latent space. Calculate and store the geometric properties of the thermodynamic state manifold, the geometric properties including at least scalar curvature; Edge computing nodes, configured at the blast furnace production site, are used for: Deploy the thermodynamic manifold and the geometric properties; Acquire real-time data of the high-frequency process; The particle filtering algorithm is run to track the state of molten iron in real time on the thermodynamic manifold and generate an estimate of the state of molten iron at the current moment. The central control module is used for: Receive the estimated state of the molten iron generated by the edge computing node and the scalar curvature corresponding to the state position; The model predictive controller is run to calculate the optimal control command based on the received information. The optimal control command is sent to the actuator of the blast furnace to adjust the operating parameters.

Images