A real-time carbon temperature identification method and device based on flame image data

Abstract

The application discloses a real-time carbon temperature identification method and device based on flame image data, and relates to the technical field of carbon temperature identification. The method comprises the following steps: inputting spatial feature maps of sample images into a variational space-time coding module in a time reverse order, and performing the following recursion: according to a current latent state, obtaining a predicted state by reverse integration of a solver containing a dynamic function, and constraining the dynamic function by using a mechanism constraint module; inputting the predicted state and the spatial feature map of the previous moment into a convolution gate recurrent unit, and outputting a corrected latent state; repeating the recursive operation to obtain a Gaussian distribution of the initial latent state, inputting the Gaussian distribution into a cross-modal decoding module, mapping the Gaussian distribution into a temperature prediction value through forward integration of the solver, training an initial carbon temperature identification model, and predicting the carbon temperature of a converter mouth flame video based on the trained model. By using the application, the accurate judgment of the change of the carbon content and the temperature with time can be realized, and the identification precision is improved.

Description

Technical Field

[0001] This invention relates to the field of carbon temperature identification technology, and in particular to a real-time carbon temperature identification method and apparatus based on flame image data. Background Technology

[0002] Converter steelmaking is a core process in the modern steel industry, and controlling its endpoint carbon content and temperature (referred to as "carbon temperature") is crucial for determining steel quality, reducing production costs, and improving production efficiency. Traditionally, determining the endpoint carbon temperature relies heavily on experienced steelworkers visually observing the color, shape, brightness, and trends of the furnace flame. This method exposes workers to harsh and even dangerous working environments with high temperatures, high noise levels, and high dust levels, seriously endangering their health. Furthermore, the judgments are driven by subjective experience, lacking stability and consistency. With technological advancements and increasing safety requirements, modern converters generally employ closed designs and monitor the furnace flame using industrial cameras. This renders traditional manual observation methods impractical. Developing precise, real-time, and automated flame image analysis technology to replace manual judgment has become an urgent need for the intelligent upgrading of the steel industry.

[0003] Traditional machine vision-based flame analysis techniques achieve a simple mapping from static image features (such as color histograms and average brightness) to carbon temperature. However, the physicochemical reactions such as decarburization and heating within a converter are an extremely complex, nonlinear, and highly time-varying dynamic process. The furnace flame, as a comprehensive external manifestation of the in-furnace reaction, contains deeper and more crucial information about the current reaction state of the molten pool through its spatiotemporal dynamic characteristics—namely, the flame's morphological evolution, fluctuation rhythm, and texture changes over continuous time. Most existing technical solutions employ traditional image processing combined with shallow machine learning, or use standard convolutional neural networks (CNNs) to analyze single-frame images. Essentially, this discretizes the continuous dynamic process into a series of static time-point "snapshots." Such methods ignore the continuous evolution of the flame over time, failing to effectively capture the dynamic characteristics of the process. Consequently, the models suffer from insufficient accuracy and weak generalization ability, making it difficult to meet the stringent requirements of industrial production for final carbon temperature control. Summary of the Invention

[0004] To address the technical problems of insufficient accuracy and weak generalization ability in existing carbon temperature identification models, this invention provides a real-time carbon temperature identification method and apparatus based on flame image data. The technical solution is as follows: On the one hand, a real-time carbon temperature identification method based on flame image data is provided. This method is implemented by a real-time carbon temperature identification device based on flame image data, and includes: S1. Obtain flame video sample data and an untrained initial carbon temperature recognition model during the converter steelmaking process. The initial carbon temperature recognition model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module. S2. Input the flame video sample image into the image feature encoding module to obtain the spatial feature map; S3. Input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, perform the following recursive operations: Based on the corrected latent state of the next time step, use the neural differential equation solver to perform inverse integration on the dynamic function to obtain the predicted state of the current time step; input the predicted state of the current time step and the spatial feature map of the current time step into the convolutional gated recurrent unit to output the corrected latent state of the current time step; at the same time, use the mechanism constraint module to apply physical mechanism constraints to the dynamic function to obtain the constrained dynamic function; repeat the recursive operation until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state. S4. Input the Gaussian distribution of the initial latent state into the cross-modal decoding module, and obtain the latent features by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network. S5. Based on the loss function between the predicted sample temperature and the true carbon temperature, the initial carbon temperature recognition model is trained to obtain the trained carbon temperature recognition model; the trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing constrained dynamic functions, and a cross-modal decoding module containing constrained dynamic functions. S6. Acquire video data of the converter furnace mouth flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

[0005] On the other hand, a real-time carbon temperature identification device based on flame image data is provided. This device is applied to a real-time carbon temperature identification method based on flame image data, and the device includes: The acquisition unit is used to acquire flame video sample data during the converter steelmaking process and an untrained initial carbon temperature recognition model. The initial carbon temperature recognition model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module. The encoding unit is used to input the flame video sample image into the image feature encoding module to obtain a spatial feature map; The variational spatiotemporal coding and mechanistic constraint unit is used to input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, the following recursive operations are performed: based on the corrected latent state of the next time step, the dynamic function is inversely integrated using a neural differential equation solver to obtain the predicted state of the current time step; the predicted state of the current time step and the spatial feature map of the current time step are input together into a convolutional gated recurrent unit to output the corrected latent state of the current time step; simultaneously, the mechanistic constraint module is used to apply physical mechanistic constraints to the dynamic function to obtain a constrained dynamic function; the recursive operation is repeated until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state. The decoding unit is used to input the Gaussian distribution of the initial latent state into the cross-modal decoding module, and obtain the latent features by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network. The training unit is used to train the initial carbon temperature recognition model based on the loss function between the sample temperature prediction value and the true carbon temperature value, so as to obtain the trained carbon temperature recognition model; the trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing a constrained dynamic function, and a cross-modal decoding module containing a constrained dynamic function. The carbon temperature prediction unit is used to acquire video data of the converter furnace mouth flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

[0006] On the other hand, a real-time carbon temperature identification device based on flame image data is provided. The real-time carbon temperature identification device based on flame image data includes: a processor; a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, any one of the methods described above for real-time carbon temperature identification based on flame image data is implemented.

[0007] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods for real-time carbon temperature identification based on flame image data.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: The core of this invention lies in constructing an end-to-end spatiotemporal model capable of learning continuous-time dynamics. First, leveraging the powerful spatiotemporal feature extraction capabilities of the ConvGRU network, low-dimensional latent feature vectors characterizing the dynamic evolution of the flame are learned and generated from continuous flame image frames. Then, innovatively, the evolution of these latent features is modeled as a continuous dynamic system driven by Neural Ordinary Differential Equations (NeuralODE), and mechanistic constraints are imposed on the system's continuous dynamic trajectory. Unlike traditional recurrent neural networks (RNNs) that update states at discrete time steps, NeuralODE can model the system's state derivatives over continuous time and impose mechanistic constraints, thereby more accurately modeling the continuous physical processes of converter steelmaking. Through this method, the model can learn the continuous dynamics of the underlying reactions from video data, achieving accurate judgment of carbon content and temperature changes over time, greatly improving recognition accuracy and model robustness. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a real-time carbon temperature identification method based on flame image data provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a real-time carbon temperature identification method based on flame image data provided in an embodiment of the present invention; Figure 3 This is a block diagram of a real-time carbon temperature identification device based on flame image data provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a real-time carbon temperature identification device based on flame image data provided in an embodiment of the present invention. Detailed Implementation

[0011] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0012] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0013] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0014] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0015] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0016] This invention provides a real-time carbon temperature identification method based on flame image data. This method can be implemented by a real-time carbon temperature identification device based on flame image data, which can be a terminal or a server.

[0017] The core of the technical solution provided by this invention lies in combining the extraction of high-dimensional spatiotemporal features from video frames with continuous-time dynamic modeling of potential states, and applying physical mechanism constraints in the direction of the system's dynamic evolution. For example... Figure 1 The flowchart shown is for a real-time carbon temperature identification method based on flame image data. Figure 2 The diagram shown illustrates the structure of a real-time carbon temperature identification method based on flame image data. The entire process is trained and optimized in an end-to-end manner, and the processing flow of this method may include the following steps: S1. Obtain flame video sample data and untrained initial carbon temperature recognition model during the converter steelmaking process.

[0018] The initial carbon temperature identification model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module.

[0019] S2. Input the flame video sample image into the image feature encoding module to obtain the spatial feature map.

[0020] In one feasible implementation, the image feature encoding module is responsible for extracting high-quality spatial features from each frame of the input furnace flame video sequence. The input to a single module is a single frame of the flame image from the video stream. ,in Representing the Each sampling time point. Specifically, a deep convolutional neural network is used as the encoder. This network performs multiple convolutions, pooling, and nonlinear transformations on the input image, ultimately outputting a low-dimensional feature map rich in semantic information. This feature map condenses spatial information about the shape, brightness, and color distribution of the flame in a single frame image.

[0021] S3. Input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, perform the following recursive operations: Based on the corrected latent state of the next time step, use the neural differential equation solver to perform inverse integration on the dynamic function to obtain the predicted state of the current time step; input the predicted state of the current time step and the spatial feature map of the current time step into the convolutional gated recurrent unit to output the corrected latent state of the current time step; at the same time, use the mechanism constraint module to apply physical mechanism constraints to the dynamic function to obtain the constrained dynamic function; repeat the recursive operation until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state.

[0022] In one feasible implementation, the mechanism constraint module will be introduced first: The core innovation of this invention lies in the fact that it does not treat the neural network as a purely data fitting tool or a "black box," but rather in its core dynamic function. In the structural design, prior mechanistic constraints of the converter steelmaking process are embedded. This method differs from the existing technology that indirectly applies "soft constraints" using loss functions. By structurally designing the differential equations themselves, it fundamentally ensures the physical rationality and convergence of the evolution trajectory of the potential state. This invention provides two implementation paradigms for physical mechanistic constraints: mechanistic constraints with explicit forms and mechanistic constraints without explicit forms.

[0023] a) Mechanism constraints with explicit form include: decomposing the dynamic function into the sum of a mechanism-guided term with explicit physical meaning and a data-driven neural network residual term.

[0024] Alternatively, the specific operation steps are as follows: Based on the sum of the mechanism-guided term with clear physical meaning and the data-driven neural network residual term, the constrained dynamic function is determined according to the following equation (1): (1) in, This represents the residual dynamics function of a data-driven neural network. For continuous time of Dimensional potential state, It is the instantaneous rate of change of the potential state vector, i.e., the system dynamics. The mechanism guidance function is predefined based on prior knowledge. u(t) represents the external control variable.

[0025] In one feasible implementation, this paradigm is applicable to scenarios where there is a well-defined (or simplified) physicochemical model to describe part of the system's dynamics. Its core idea is to decompose the instantaneous rate of change of the system state into the sum of a mechanistic-guided term and a data-driven residual term. Based on this, a function describing the system's dynamics... The definition is as shown in equation (1) above. The mechanism guidance function is a differentiable function with clear physical meaning, predefined based on prior knowledge (such as mass balance, energy balance, chemical reaction kinetics, etc.). It correlates a portion of the potential state h(t) with actual physical quantities (such as carbon content and temperature), calculating the trends of these physical quantities based on metallurgical mechanisms. External control variables (such as oxygen blowing rate) provide a physically correct baseline direction for system evolution. The residual kinetic function is a trainable function composed of a neural network, whose function is to learn and model the mechanism guidance function. The failure to capture the more complex nonlinear dynamics and model errors in the real process essentially simplifies the task of neural networks from learning the entire complex dynamics to learning the residuals between the physical model and the real process.

[0026] b) Mechanism constraints without explicit form include: constructing an energy functional parameterized by a convolutional neural network, calculating the gradient of the energy functional based on the current potential state, and projecting and correcting the vector field predicted by the dynamic function to force the physical stability constraint of non-increasing energy.

[0027] Optionally, the gradient of the energy functional is calculated based on the potential state at the current time, including: Construct an energy functional parameterized by a convolutional neural network ,in This represents the potential state at the current moment.

[0028] Extracted from convolutional layers The spatial energy distribution is then aggregated into scalar energy values ​​through a global pooling layer.

[0029] Calculate the energy functional with respect to gradient .

[0030] Optionally, the vector field predicted by the dynamic function is projected and corrected, including: The vector field predicted by the dynamic function is projected and corrected according to the following formula (2): (2) in, Represents the dynamic function, This represents the corrected dynamic function. This indicates the correction step size coefficient. The decay factor represents the rate at which energy decay is controlled. That is , representing the energy functional, The Frobenius norm of the feature tensor is used to represent the feature tensor.

[0031] In one feasible implementation, this paradigm primarily targets the control objective of ensuring that the system state stably converges to the desired endpoint (the potential state corresponding to the target carbon temperature). An implicit constraint mechanism based on energy functionals is introduced by constructing a dynamic function. Considering that the steelmaking process is essentially an energy dissipation system tending towards thermodynamic equilibrium, this embodiment constructs a scalar energy function to constrain the evolution direction of the characteristic graph, ensuring that the spatiotemporal evolution of the potential state conforms to physical stability. Given the potential state of this model... To preserve the three-dimensional feature tensor of the spatial structure ( Where C is the number of channels, and H and W are the spatial dimensions), rather than a one-dimensional vector, this paradigm consists of three parts: ① Construct a spatiotemporal energy functional. Define an energy function parameterized by a convolutional neural network. It will be the current feature tensor This is mapped to a non-negative scalar energy value. To adapt to image feature input, the function first extracts the spatial energy distribution through a convolutional layer, and then aggregates the energy distribution through global pooling to obtain the total system energy. ,in This indicates a layer that maintains the change in spatial structure.

[0032] ② Apply dynamic stability constraints. According to the stability theory of physical systems, the evolution of the system state over time should result in non-increasing total energy (i.e., energy dissipation). Therefore, the vector field function describing the system evolution... The following inequality (3) constraint must be satisfied: (3) Here, It is the energy function with respect to the characteristic tensor The gradient, whose dimension is... same( ); This represents the tensor inner product operation; It is the decay factor that controls the rate of energy decay.

[0033] ③ Vector field correction. During model training and inference, if the vector field predicted by the neural network... If the above physical constraints are violated, the vector field is corrected in real time using the projection method to obtain a corrected vector field that conforms to the physical constraints. As shown in equation (2) above. This formula uses mathematical means to force the evolution trajectory of the feature map to "pull back" to the manifold that conforms to the law of physical energy dissipation, thereby ensuring the stability of the spatiotemporal features extracted from the flame image over a long period of time.

[0034] In one feasible implementation, the variational spatiotemporal coding module is described below: This module functions as the system's encoder. It receives the discrete feature map sequence output by the image feature encoding module. By utilizing continuous dynamics under physical constraints and a gated loop mechanism, the spatiotemporal information of the entire video stream is compressed into an initial latent state that follows a specific probability distribution. This state is also preserved in the form of a feature map, retaining spatial structure information.

[0035] The specific implementation process is as follows: a) Reverse temporal coding mechanism. To infer the initial state of the steelmaking process or the global state at the current moment, the model uses reverse temporal processing of video frames. This design is based on the observer concept in control theory, which uses future observations to infer the initial state of the system.

[0036] b) Continuous-discrete hybrid update unit. At two adjacent observation time points... and In between, the model alternately performs two steps: "continuous evolution" and "discrete correction". Step b1: Continuous evolution (prediction). Utilizing the vector field function defined in the mechanism constraint module, modified by physical mechanism constraints. Predicting from neural differential equations using a solver (ODESolver) arrive The system state at time 1. This is equivalent to using the learned physical laws to deduce the state at the previous time, as shown in equation (4): (4) in, It is a function parameterized by a neural network that learns the hidden state. Regarding time The derivative, i.e. It is used to model the dynamics of continuous systems in the converter steelmaking process. ODESolver is a differential solver used to solve problems arising from... The potential state of the defined system dynamic trajectory at the target time , This represents the potential state after ConvGRU correction.

[0037] Step b2: Discrete correction (update). Utilizing Actual observed feature image at time The predicted state is obtained by using a convolutionally gated recurrent unit (ConvGRU). The correction is performed, and the observation information is integrated, as shown in equation (5): (5) During this process, the reset and update gate mechanisms within ConvGRU can adaptively determine how much dynamic prediction information to retain and how much current image observation information to incorporate.

[0038] c) Variational distribution generation. When the entire sequence has been recursively processed and the initial time has been reached... At (or the start point of the sequence), the output of the last ConvGRU unit is no longer a definite state value, but is mapped to the mean feature map of the latent state distribution through two parallel convolutional layers. Characteristic plot of log-variance This step assumes that all the spatiotemporal information contained in the video stream can be encoded as a high-dimensional Gaussian distribution. , The Gaussian distribution represents the initial potential state. This distribution is considered to contain all the spatiotemporal information of the input frame sequence.

[0039] S4. Input the Gaussian distribution of the initial latent state into the cross-modal decoding module. The latent features are obtained by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network.

[0040] In one feasible implementation, the cross-modal decoding module acts as the system's "decoder." This module samples from the Gaussian distribution of the initial latent state, performs forward evolution using the kinetic equations constrained by the physical mechanism, and decodes the carbon content and temperature at any given time. This module achieves a cross-modal mapping from "latest characteristic distribution" to "continuous physical quantities." The specific implementation scheme is as follows: a) Reparameterization sampling. To ensure end-to-end trainability of the model (gradients can be backpropagated), the initial latent state is sampled from the distribution of the encoder output using reparameterization techniques. As shown in equation (6): (6) in, It is a The tensor represents the physical and semantic characteristics of the system at the initial moment. This represents the random noise tensor sampled from the Gaussian distribution of the initial potential state.

[0041] b) Arbitrary-time state reconstruction based on physical constraints. This is achieved using a modified vector field function. ,Will The initial conditions are input into ODESolver. Unlike existing discrete models, this embodiment of the invention can solve for the latent feature map at any time t on the time axis (which can be the time corresponding to a video frame or the time between frames) by changing the upper limit of the integration. As shown in equation (7): (7) in, Indicates time The potential state, It is an integral variable, and its range is from arrive Any time between.

[0042] This step embodies the generative process: based on the learned metallurgical kinetics, the model "generates" the evolutionary trajectory of the entire steelmaking process over continuous time.

[0043] c) Carbon-temperature decoding network. For the latent feature map obtained at any time t... The data is mapped to target scalars (carbon content, temperature) through a decoding network. The decoding network consists of convolutional networks and multilayer perceptrons.

[0044] S5. Based on the loss function between the predicted sample temperature and the true carbon temperature, train the initial carbon temperature identification model to obtain the trained carbon temperature identification model.

[0045] The trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing constrained dynamic functions, and a cross-modal decoding module containing constrained dynamic functions.

[0046] In one feasible implementation, to ensure that the model both conforms to physical laws and can make accurate predictions, this invention uses mean squared error (MSE) as the main reconstruction loss function during the training phase, while combining it with variational inference's KL divergence loss. Total loss function Defined as follows (8): (8) in, To predict the loss, the difference between the predicted carbon temperature value and the actual test value is used. Since test data is typically sparse (not available for every frame), the loss only occurs at times when the true label is available. The calculation is as follows (9): (9) These are actual measured carbon temperature data. This represents the predicted temperature value of the sample. It is a regularization term that makes the posterior distribution close to the standard normal distribution, thus preventing overfitting.

[0047] Through this "encoding-evolution-decoding" architecture, the embodiments of the present invention are not merely fitting data points, but learning a continuous physical system capable of generating observational data. This means that once the model is trained, it can not only predict video frame moments, but also smoothly extrapolate future trends, and it exhibits strong robustness to dropped frames and uneven sampling rates in the input video.

[0048] S6. Acquire video data of the converter furnace mouth flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

[0049] Optionally, S6 inputs video data into the trained carbon-temperature recognition model and outputs the corresponding carbon content and temperature prediction values, including: S61. Input the video data into the image feature encoding module, perform frame segmentation on the video data, input the obtained video frame sequence into the convolutional neural network, and obtain a low-dimensional spatial feature map sequence.

[0050] S62. Input the spatial feature map sequence in reverse order into the variational spatiotemporal coding module containing the constrained dynamic function. Use the neural differential equation solver to integrate the constrained dynamic function to deduce the predicted state of the previous time step. Input the predicted state and the spatial feature map of the corresponding time step into the convolutional gated recurrent unit and output the corrected latent state. Repeat the reverse deduction and correction process until the entire reverse sequence is processed to obtain the Gaussian distribution of the initial latent state.

[0051] Optionally, the repeated backward inference and correction process of S62 continues until the entire reverse sequence has been processed, resulting in a Gaussian distribution of the initial latent state, including: Repeat the recursive operation until the entire sequence is processed, and obtain the mean feature map and log-variance feature map of the latent state distribution mapped by the two parallel convolutional layers, respectively.

[0052] The mean feature map and log-variance feature map of the latent state distribution are encoded into a high-dimensional Gaussian distribution.

[0053] S63. Based on the cross-modal decoding module, the initial potential state is sampled from the Gaussian distribution. The constrained dynamic function is integrated using the neural differential equation solver. The integration is performed from the initial potential state to the current time to obtain the potential characteristics at the current time. The decoding network is then used to map the carbon content and temperature prediction values ​​at the current time.

[0054] In one feasible implementation, when using the trained carbon temperature recognition model to recognize the video data of the converter furnace flame to be identified, the specific processing procedures of the image feature encoding module, the variational spatiotemporal encoding module containing constrained dynamic functions, and the cross-modal decoding module containing constrained dynamic functions in the trained carbon temperature recognition module are consistent with the training process in steps S2-S4 above, and will not be elaborated here.

[0055] The following example, using a 120-ton top-and-bottom blown converter steelmaking process in a steel plant, illustrates the specific implementation of the carbon temperature identification method based on flame image data described in this embodiment of the invention: This system can analyze the video stream of the furnace flame during the converter smelting process, overcome problems such as on-site smoke and dust interference and video frame loss, and output the predicted values ​​of carbon content and temperature of molten steel in real time.

[0056] This implementation method mainly includes the following four steps: data acquisition and preprocessing, model building, model training, and online inference and monitoring.

[0057] 1. Data Acquisition and Preprocessing

[0058] 1.1 Hardware Deployment

[0059] A high-temperature resistant industrial high-speed HD camera is installed on the side and rear of the converter inlet. The camera is equipped with a water-cooled protective cover and an air-cooled dust removal device to ensure stable operation in high-temperature and high-dust environments.

[0060] Sampling frequency: The video stream sampling rate is 30 fps (frames per second).

[0061] Image resolution: The original resolution is 1920×1080 pixels.

[0062] 1.2 Data Preprocessing and Label Matching

[0063] Image processing: The acquired flame video stream is sliced ​​according to timestamps. For computational efficiency, each frame is downsampled to [a specific size / resolution]. The pixels are normalized (pixel values ​​are mapped to [0,1]).

[0064] Tag Acquisition: Actual carbon content and temperature data measured at the end of blowing and during mid-blowout at the secondary lance or rapid carbon and temperature analyzer (TSC / TSO) are used as "true value tags." Since the test data is sparse (e.g., only 1-2 test points per heat of steel), time alignment is used to link the test times... Image frames and corresponding carbon temperature data { Bind it.

[0065] 2. Model Building

[0066] This embodiment constructs an end-to-end deep learning model, whose architecture strictly corresponds to the four modules in the invention: (1) Image feature coding module A lightweight convolutional neural network (CNN) is used as the encoder. The network contains four convolutional blocks, each of which includes a "convolution-BatchNorm-ReLU-max pooling" operation.

[0067] Input: Single-frame flame image .

[0068] Output: Spatial feature map This feature map preserves the spatial texture and brightness distribution information of the flame.

[0069] (2) Mechanism constraint module (adopting an energy constraint paradigm without a clear form)

[0070] To ensure that the evolution of generated features conforms to the laws of physical energy dissipation, the following structure is constructed: Energy function V( Design an energy network consisting of 3 3×3 convolutional layers with 1 output channel, and then output a scalar energy value V(h(t)) through global average pooling.

[0071] Vector Field Correction: Defining the Foundation of Neural Networks (Constituted from 3 layers of ResNet residual blocks) used to fit the dynamics. In computation... When calculating the derivative, the gradient of the energy function with respect to the characteristic tensor is obtained. .

[0072] Projection correction: Apply formula (2). If an increase in energy is detected (i.e., a violation of physical stability), calculate the correction term. Set the attenuation factor. The corrected vector field The forced characteristic manifold evolves towards a lower energy state, simulating the physical characteristics of the steelmaking process tending towards thermodynamic equilibrium.

[0073] (3) Variational Spatiotemporal Coding Module

[0074] Reverse processing: Input a video sequence of length T (e.g., 300 frames from the past 10 seconds) in reverse chronological order.

[0075] Continuous-Discrete Hybrid Update: Continuous prediction: using the Runge-Kutta4 (RK4) solver with the modified vector field. ,from Potential state at time Inverse integral derivation Prior state at time .

[0076] Discrete correction: This involves adjusting the observed features of the current frame. and prior states Input a ConvGRU cell and output the updated state. .

[0077] Distribution generation: when recursively returning to the start of the sequence At that time, the mean tensor is output through two parallel convolutional layers. and log-variance tensor .

[0078] (4) Cross-modal decoding module

[0079] Sampling and Generation: Using Reparameterization Techniques to Generate from Gaussian Distributions Initial state obtained by sampling .

[0080] Forward evolution: As initial values, we again utilize ODESolver and a vector field that conforms to physical constraints. By integrating in the positive direction on the continuous time axis, the target time can be calculated. Potential characteristics .

[0081] (Note: here) It can be any time, and even if the camera drops frames at that time, the model can still calculate the state at that moment through ODE integration, demonstrating robustness.

[0082] Numerical decoding: The input decoding network (consisting of a global pooling layer and three fully connected layers) outputs the final predicted values ​​of carbon content (%) and temperature (°C).

[0083] 3. Model Training Process

[0084] Dataset partitioning: Collect 2000 heats of historical smelting data from the converter over 3 months, and partition them into training, validation and test sets in a ratio of 8:1:1.

[0085] Loss function: The comprehensive loss is defined by formulas (8)-(9).

[0086] Prediction loss: The mean square error (MSE) between the predicted value and the test value is calculated only when test data is available.

[0087] KL divergence loss: constrains the encoding distribution to approximate a standard normal distribution.

[0088] Optimization strategy: Use the Adam optimizer with an initial learning rate of 1e-4.

[0089] Training details: Due to the introduction of ODESolver, the adjoint sensitivity method is used for backpropagation to save GPU memory and accurately calculate gradients. The BatchSize is set to 16, and training is performed for 100 epochs.

[0090] 4. Online Reasoning

[0091] During the model deployment phase, the system's real-time operation process is as follows: Sliding window input: The system maintains a sliding window video stream with a duration of 5 seconds.

[0092] Real-time computation: Whenever a new image frame arrives, the model first extracts features through the encoder, and then uses the variational encoder to inversely deduce the current global initial state distribution.

[0093] Future prediction: Based on the derived initial state, the ODESolver is used to integrate forward to the "current moment" or even the "future 30 seconds" to predict the current carbon temperature value and its future trend.

[0094] Anomaly Handling: Simulation experiments show that when 20% of the input frames are randomly blocked by human intervention (simulating frame loss or strong light interference), since ODE describes a continuous physical process, the model can still fill in the missing information through integration, and the prediction error increases by less than 2%, which is far better than the performance of traditional LSTM / CNN models (the error of traditional models increases by more than 15%).

[0095] The core of this invention lies in constructing an end-to-end spatiotemporal model capable of learning continuous-time dynamics. First, leveraging the powerful spatiotemporal feature extraction capabilities of the ConvGRU network, low-dimensional latent feature vectors characterizing the dynamic evolution of the flame are learned and generated from continuous flame image frames. Then, innovatively, the evolution of these latent features is modeled as a continuous dynamic system driven by Neural Ordinary Differential Equations (NeuralODE), and mechanistic constraints are imposed on the system's continuous dynamic trajectory. Unlike traditional recurrent neural networks (RNNs) that update states at discrete time steps, NeuralODE can model the system's state derivatives over continuous time and impose mechanistic constraints, thereby more accurately modeling the continuous physical processes of converter steelmaking. Through this method, the model can learn the continuous dynamics of the underlying reactions from video data, achieving accurate judgment of carbon content and temperature changes over time, greatly improving recognition accuracy and model robustness.

[0096] Figure 3 This is a block diagram of a real-time carbon temperature identification device based on flame image data, provided in an embodiment of the present invention. This device is used for a real-time carbon temperature identification method based on flame image data. (Refer to...) Figure 3The device includes an acquisition module 310, an encoding unit 320, a mechanism constraint unit 330, a decoding unit 340, a training unit 350, and a carbon temperature prediction unit 360. Wherein: The acquisition unit 310 is used to acquire flame video sample data during the converter steelmaking process and an untrained initial carbon temperature recognition model. The initial carbon temperature recognition model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module. Encoding unit 320 is used to input the flame video sample image into the image feature encoding module to obtain a spatial feature map; The variational spatiotemporal coding and mechanism constraint unit 330 is used to input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, the following recursive operations are performed: based on the corrected latent state of the next time step, the dynamic function is inversely integrated using a neural differential equation solver to obtain the predicted state of the current time step; the predicted state of the current time step and the spatial feature map of the current time step are input together into a convolutional gated recurrent unit to output the corrected latent state of the current time step; simultaneously, the dynamic function is physically constrained using a mechanism constraint module to obtain a constrained dynamic function; the recursive operation is repeated until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state. The decoding unit 340 is used to input the Gaussian distribution of the initial latent state into the cross-modal decoding module, and obtain the latent features by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network. Training unit 350 is used to train the initial carbon temperature recognition model based on the loss function between the sample temperature prediction value and the true carbon temperature value to obtain the trained carbon temperature recognition model; the trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing a constrained dynamic function, and a cross-modal decoding module containing a constrained dynamic function. The carbon temperature prediction unit 360 is used to acquire video data of the converter furnace flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

[0097] Figure 4 This is a schematic diagram of the structure of a real-time carbon temperature identification device based on flame image data provided in an embodiment of the present invention, as shown below. Figure 4 As shown, a real-time carbon temperature identification device based on flame image data may include the above-mentioned... Figure 3 The illustrated real-time carbon temperature identification device based on flame image data. Optionally, the real-time carbon temperature identification device 410 based on flame image data may include a first processor 2001.

[0098] Optionally, the real-time carbon temperature identification device 410 based on flame image data may also include a memory 2002 and a transceiver 2003.

[0099] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0100] The following is combined Figure 4 A detailed description of each component of the real-time carbon temperature identification device 410 based on flame image data is provided below: The first processor 2001 is the control center of the real-time carbon temperature identification device 410 based on flame image data. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0101] Optionally, the first processor 2001 can perform various functions of the real-time carbon temperature identification device 410 based on flame image data by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0102] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0103] In a specific implementation, as one example, the real-time carbon temperature identification device 410 based on flame image data may also include multiple processors, for example... Figure 4 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0104] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0105] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be connected via the interface circuit of the real-time carbon temperature identification device 410 based on flame image data. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0106] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0107] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0108] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected to the interface circuit of the real-time carbon temperature identification device 410 based on flame image data. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0109] It should be noted that, Figure 4 The structure of the real-time carbon temperature identification device 410 based on flame image data shown does not constitute a limitation on the router. Actual real-time carbon temperature identification devices based on flame image data may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0110] Furthermore, the technical effect of the real-time carbon temperature identification device 410 based on flame image data can be referred to the technical effect of the real-time carbon temperature identification method based on flame image data described in the above method embodiments, and will not be repeated here.

[0111] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or it may be any conventional processor, etc.

[0112] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0114] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0115] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0116] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0117] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0119] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0122] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A real-time carbon temperature identification method based on flame image data, characterized in that, The method includes: S1. Obtain flame video sample data and an untrained initial carbon temperature recognition model during the converter steelmaking process. The initial carbon temperature recognition model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module. S2. Input the flame video sample image into the image feature encoding module to obtain the spatial feature map; S3. Input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, perform the following recursive operations: Based on the corrected latent state of the next time step, use the neural differential equation solver to perform inverse integration on the dynamic function to obtain the predicted state of the current time step; input the predicted state of the current time step and the spatial feature map of the current time step into the convolutional gated recurrent unit to output the corrected latent state of the current time step; at the same time, use the mechanism constraint module to apply physical mechanism constraints to the dynamic function to obtain the constrained dynamic function; repeat the recursive operation until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state. S4. Input the Gaussian distribution of the initial latent state into the cross-modal decoding module, and obtain the latent features by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network. S5. Based on the loss function between the predicted sample temperature and the true carbon temperature, the initial carbon temperature recognition model is trained to obtain a trained carbon temperature recognition model. The trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing constrained dynamic functions, and a cross-modal decoding module containing constrained dynamic functions. S6. Acquire video data of the converter furnace mouth flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

2. The real-time carbon temperature identification method based on flame image data according to claim 1, characterized in that, The physical mechanism constraints include mechanistic constraints with explicit forms and mechanistic constraints without explicit forms; The mechanistic constraints with a defined form include: decomposing the dynamic function into the sum of a mechanistic guiding term with a defined physical meaning and a data-driven neural network residual term; The mechanistic constraints without explicit form include: constructing an energy functional parameterized by a convolutional neural network, calculating the gradient of the energy functional based on the potential state at the current moment, and projecting and correcting the vector field predicted by the dynamic function to force the physical stability constraint of non-increasing energy to be satisfied.

3. The real-time carbon temperature identification method based on flame image data according to claim 2, characterized in that, The decomposition of the dynamic function into the sum of a mechanism-guided term with explicit physical meaning and a data-driven neural network residual term includes: Based on the sum of the mechanism-guided term with clear physical meaning and the data-driven neural network residual term, the constrained dynamic function is determined according to the following equation (1): （1） in, This represents the data-driven residual dynamics function of a neural network. For continuous time of Dimensional potential state, It is the instantaneous rate of change of the potential state vector, i.e., the system dynamics. ; The mechanism guidance function is predefined based on prior knowledge; u(t) represents the external control variable.

4. The real-time carbon temperature identification method based on flame image data according to claim 2, characterized in that, The calculation of the gradient of the energy functional based on the current potential state includes: Construct an energy functional parameterized by a convolutional neural network ,in The potential state at the current moment; Extracted from convolutional layers The spatial energy distribution is then aggregated into scalar energy values ​​through a global pooling layer. Calculate the energy functional with respect to gradient .

5. The real-time carbon temperature identification method based on flame image data according to claim 4, characterized in that, The projection correction of the vector field predicted by the dynamic function includes: The vector field predicted by the dynamic function is projected and corrected according to the following formula (2): （2） in, Represents the dynamic function, This represents the corrected dynamic function. This indicates the correction step size coefficient. The decay factor represents the rate at which energy decay is controlled. That is , representing the energy functional, The Frobenius norm of the feature tensor is used to represent the feature tensor.

6. The real-time carbon temperature identification method based on flame image data according to claim 1, characterized in that, The step S6 involves inputting video data into a trained carbon-temperature recognition model and outputting corresponding carbon content and temperature prediction values, including: S61. Input the video data into the image feature encoding module, perform frame segmentation on the video data, input the obtained video frame sequence into the convolutional neural network, and obtain a low-dimensional spatial feature map sequence. S62. Input the spatial feature map sequence in reverse order into the variational spatiotemporal coding module containing the constrained dynamic function. Integrate the constrained dynamic function through the neural differential equation solver to deduce the predicted state of the previous time step. Input the predicted state and the spatial feature map of the corresponding time step into the convolutional gated recurrent unit and output the corrected latent state. Repeat the reverse deduction and correction process until the entire reverse sequence is processed to obtain the Gaussian distribution of the initial latent state. S63. Based on the cross-modal decoding module, the initial potential state is sampled from the Gaussian distribution. The constrained dynamic function is integrated using the neural differential equation solver. The integration is performed from the initial potential state to the current time to obtain the potential characteristics at the current time. The decoding network is then used to map the carbon content and temperature prediction values ​​at the current time.

7. The real-time carbon temperature identification method based on flame image data according to claim 6, characterized in that, The repeated backward inference and correction process of S62 continues until the entire reverse sequence is processed, resulting in a Gaussian distribution of the initial latent state, including: Repeat the recursive operation until the entire sequence is processed to obtain the mean feature map and log-variance feature map of the latent state distribution mapped by the two parallel convolutional layers, respectively. The mean feature map and log-variance feature map of the latent state distribution are encoded into a high-dimensional Gaussian distribution.

8. A real-time carbon temperature identification device based on flame image data, wherein the real-time carbon temperature identification device based on flame image data is used to implement the real-time carbon temperature identification method based on flame image data as described in any one of claims 1-7, characterized in that, The device includes: The acquisition unit is used to acquire flame video sample data during the converter steelmaking process and an untrained initial carbon temperature recognition model. The initial carbon temperature recognition model includes an image feature encoding module, a mechanism constraint module, a variational spatiotemporal encoding module, and a cross-modal decoding module. The encoding unit is used to input the flame video sample image into the image feature encoding module to obtain a spatial feature map; The variational spatiotemporal coding and mechanistic constraint unit is used to input the sequence of spatial feature maps into the variational spatiotemporal coding module in reverse chronological order. For each time step, the following recursive operations are performed: based on the corrected latent state of the next time step, the dynamic function is inversely integrated using a neural differential equation solver to obtain the predicted state of the current time step; the predicted state of the current time step and the spatial feature map of the current time step are input together into a convolutional gated recurrent unit to output the corrected latent state of the current time step; simultaneously, the mechanistic constraint module is used to apply physical mechanistic constraints to the dynamic function to obtain a constrained dynamic function; the recursive operation is repeated until the entire sequence is processed to obtain the Gaussian distribution of the initial latent state; The decoding unit is used to input the Gaussian distribution of the initial latent state into the cross-modal decoding module, and obtain the latent features by positively integrating the constrained dynamic function to the target time through the neural differential equation solver. The latent features are then mapped to the predicted values ​​of carbon content and sample temperature through the decoding network. The training unit is used to train the initial carbon temperature recognition model based on the loss function between the sample temperature prediction value and the true carbon temperature value, so as to obtain the trained carbon temperature recognition model; the trained carbon temperature recognition module includes an image feature encoding module, a variational spatiotemporal encoding module containing a constrained dynamic function, and a cross-modal decoding module containing a constrained dynamic function. The carbon temperature prediction unit is used to acquire video data of the converter furnace mouth flame, input the video data into the trained carbon temperature recognition model, and output the corresponding carbon content and temperature prediction values.

9. A real-time carbon temperature identification device based on flame image data, characterized in that, The real-time carbon temperature identification device based on flame image data includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images