Sintering air bellow temperature forecasting method and system based on local feature Transform network

Abstract

The invention discloses a sintering air bellow temperature forecasting method and system based on a local feature Transform network, and belongs to the technical field of intelligent monitoring in the metallurgical process. The method comprises the steps that waste gas temperature, trolley speed and material layer thickness data of a target air bellow and an adjacent air bellow are collected and preprocessed; performing sliding window slicing on the historical data to construct a sample set; the method comprises the following steps of: constructing a local feature Transform network which is independently designed by adopting a channel, performing Patch division on each variable sequence to extract a time sequence local feature, and capturing an internal dependency relationship of variables by utilizing a self-attention mechanism; and training and optimizing network parameters, and finally realizing accurate prediction of the exhaust gas temperature of the target bellows in a future period of time. The method effectively solves the problems that an existing method depends on a mechanism model, parameters are difficult to obtain, the generalization ability is poor, and a local time sequence mode is insufficient to capture, and the forecasting precision and the working condition adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and forecasting technology for metallurgical industrial processes, and in particular to a method and system for forecasting the temperature of sintering bellows based on a local feature Transformer network. Background Technology

[0002] In the steel smelting process, sintering is a crucial step in ensuring stable blast furnace operation and molten iron quality. The temperature of the blast furnace exhaust gas, as a direct reflection of the sintering process, has significant guiding implications for determining the sintering endpoint, controlling energy consumption, and improving product quality. Currently, blast furnace temperature prediction methods are mainly divided into two categories: mechanistic models and data-driven models.

[0003] Mechanistic models rely on the physicochemical reaction equations of the sintering process, such as sintering zone movement models and computational fluid dynamics models. Although they have clear physical meaning, they are computationally complex, depend on precise physical property parameters, and are difficult to deploy in real-time in actual industrial environments. Data-driven methods, such as support vector machines, random forests, and recurrent neural networks, have improved prediction capabilities to some extent, but still suffer from problems such as sensitivity to input parameters, insufficient capture of temporal dependencies, and weak generalization ability across operating conditions.

[0004] In recent years, Transformer models have demonstrated their advantages in the field of time series forecasting due to their powerful sequence modeling capabilities. However, existing Transformer methods often neglect local temporal patterns between variables in multivariate time series forecasting, resulting in limited ability to model nonlinear and multi-coupling relationships during the sintering process.

[0005] Therefore, there is an urgent need for a time-series prediction method that can integrate local feature extraction and global dependency capture to improve the accuracy and robustness of sintering box temperature prediction. Summary of the Invention

[0006] The purpose of this invention is to provide a sintering box temperature prediction method based on a local feature Transformer network. By introducing channel-independent design and patch partitioning mechanism, the model's ability to extract local temporal features is enhanced. Combined with a self-attention mechanism to capture the internal dependencies of variables, a high-precision and strong-generalization temperature prediction method is achieved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the temperature of a sintering windbox based on a local feature Transformer network, comprising the following steps: Step S1: Collect data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and preprocess the data, including outlier removal and mean value processing. Step S2: Perform sliding window slicing on the historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set; Step S3: Construct a local feature Transformer network. This network processes the input data in a channel-independent manner, inputting the time series of each variable as an independent channel, and performing patching on each channel sequence to extract local temporal features. It also captures the internal dependencies of variables through a self-attention mechanism. Step S4: Determine the network structure parameters, including embedding dimension, patch length, and stride, based on the historical sequence length, predicted sequence length, and number of variables; and set the network's learning rate, number of iterations, and early stopping patience hyperparameters. Step S5: Train the network using the training set, use the validation set to optimize the network parameters, and use the test set to evaluate the model performance. During training, use the loss function to evaluate the error and adjust the network parameters according to the evaluation results. Repeat the training until the expected accuracy is achieved. Step S6: Input the current time and the historical L time-series data sequences of the target wind box and the exhaust gas temperature of the adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

[0008] Optionally, in step S2, the sliding window slicing process is specifically as follows: Multivariate time series are sliced ​​with window length L and step size 1. Each sample input is two-dimensional data of M×L, and the output is a prediction sequence of M×T, where M is the number of variables and T is the prediction length.

[0009] Optionally, in step S3, the channel-independent mode is specifically as follows: The time series of each variable is input into the Transformer network as an independent channel for local feature extraction and encoding.

[0010] Optionally, in step S3, the patch division is specifically as follows: For each variable sequence, a continuous subsequence of length P is extracted with a step size S, forming N = L / S Each local feature block is converted into a fixed-dimensional vector as the encoder input.

[0011] Optionally, the Transformer network encoder consists of multiple stacked encoder units, each of which includes a multi-head self-attention layer, a feedforward neural network layer, and employs residual connections and layer normalization.

[0012] Optionally, in step S5, mean squared error or mean absolute error is used as the loss function to optimize the network weights through backpropagation.

[0013] Optionally, in step S6, the input data includes the target wind box and adjacent wind box temperatures, trolley speed and material layer thickness data at the current time and the historical L times, and the output is the target wind box temperature sequence at the next T times.

[0014] Secondly, the present invention provides a sintering box temperature prediction system based on a local feature Transformer network, used to implement any of the above-described sintering box temperature prediction methods based on a local feature Transformer network, the system comprising: The data acquisition and preprocessing module is configured to acquire data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and to preprocess the data, including outlier removal and mean value processing. The sample construction module is configured to perform sliding window slicing on historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set. The network construction module is configured to build a local feature Transformer network. This network processes input data in a channel-independent manner, inputting the time series of each variable as an independent channel, performing patching on each channel sequence to extract local temporal features, and capturing the internal dependencies of variables through a self-attention mechanism. The parameter setting module is configured to determine network structure parameters, including embedding dimension, patch length and step size, based on the length of historical sequences, the length of predicted sequences and the number of variables, and to set the network's learning rate, number of iterations and early stopping patience hyperparameters. The network training module is configured to train the network using the training set, optimize the network parameters using the validation set, and evaluate the model performance using the test set. During training, a loss function is used to evaluate the error, and the network parameters are adjusted based on the evaluation results. Training is repeated until the expected accuracy is achieved. The temperature prediction module is configured to input the current time and the historical L time-series data sequences of the exhaust gas temperature of the target wind box and adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

[0015] In another aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described methods for predicting the temperature of a sintering box based on a local feature Transformer network.

[0016] On the other hand, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements any of the above-mentioned methods for predicting the temperature of a sintering box based on a local feature Transformer network.

[0017] The above technical solution has the following technical effects: This invention employs a channel-independent design to separate and process multivariate time-series data, and combines local patch partitioning to enhance the extraction of short-term dynamic patterns for each variable. This effectively overcomes the shortcomings of traditional multivariate prediction models, such as mutual interference between variables and the submergence of local time-series features. Furthermore, by combining the self-attention mechanism of the Transformer network to capture global long-range dependencies from the refined local features, a hierarchical feature learning framework from local to global is formed. This technical approach fundamentally reduces the dependence on complex mechanistic models and precise physical property parameters, significantly enhancing the model's ability to characterize the nonlinear and strongly coupled dynamic characteristics of the sintering process. As a result, it has achieved significant benefits in engineering practice, including improved prediction accuracy, enhanced cross-condition generalization ability, and improved adaptability to different operating conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of a sintering windbox temperature prediction method based on a local feature Transformer network according to an embodiment of the present invention.

[0019] Figure 2 This is a structural block diagram of a sintering windbox temperature prediction system based on a local feature Transformer network, according to another embodiment of the present invention.

[0020] Figure 3 This is a structural schematic diagram of another embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0022] The temperature of the exhaust gas from the bellows serves as a real-time barometer and navigator for the "health status" of the sintering process. Accurate prediction of this temperature means we can shift from reactive to proactive intervention, thereby achieving more efficient, higher-quality, and more environmentally friendly production. The predictive value of bellows exhaust gas temperature guidance permeates core aspects of sintering production, including operational optimization, status monitoring, quality control, and energy management.

[0023] The prediction of sintering box exhaust gas temperature currently relies primarily on two technical approaches: establishing mechanistic models such as sintering zone movement models and coupled models of computational fluid dynamics and discrete element method; and data-driven prediction models such as traditional machine learning methods like linear regression, support vector machines, and random forests, or deep learning methods such as fully connected neural networks, recurrent neural networks, convolutional neural networks, and Transformer networks. Among these, the Transformer is a neural network architecture based on a self-attention mechanism. It abandons traditional recurrent and convolutional structures, enabling parallel processing of all elements in a sequence and dynamically calculating dependencies between elements through attention weights. This inherent characteristic makes it excellent at capturing long-range dependencies and achieving high parallel computation efficiency, making it the cornerstone of most models in fields such as natural language processing (e.g., BERT, GPT). However, the sintering process is a complex physical and chemical reaction process. Mechanistic models are costly to calculate and require a large number of physical property parameters that are difficult to obtain accurately. It is difficult to achieve real-time prediction and control in actual industrial settings, and mechanistic models are difficult to simulate accurately. Data models such as SVM, BP network, and RNN network, which rely on raw material parameters such as iron ore powder, fuel, and solvent, as well as state control parameters such as moisture, material layer thickness, and ignition temperature, have poor versatility due to differences in parameter selection, measurement accuracy, and lag.

[0024] Therefore, there is an urgent need for a method to predict the temperature of sintering bellows, aiming to further improve the accuracy and interpretability of existing methods for predicting the temperature of sintering bellows, and to solve the problems of dependence on the measurement of various raw material parameters and state parameters during the sintering process, as well as the low versatility of the calculation methods under different working conditions.

[0025] The following is combined Figure 1 The flowchart below provides a detailed explanation of the method of the present invention.

[0026] A first aspect of the present invention provides a method for predicting the temperature of a sintering windbox based on a local feature Transformer network, comprising the following steps: Step S1: Collect data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and preprocess the data, including outlier removal and mean value processing. Step S2: Perform sliding window slicing on the historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set; Step S3: Construct a local feature Transformer network. This network processes the input data in a channel-independent manner, inputting the time series of each variable as an independent channel, and performing patching on each channel sequence to extract local temporal features. It also captures the internal dependencies of variables through a self-attention mechanism. Step S4: Determine the network structure parameters, including embedding dimension, patch length, and stride, based on the historical sequence length, predicted sequence length, and number of variables; and set the network's learning rate, number of iterations, and early stopping patience hyperparameters. Step S5: Train the network using the training set, use the validation set to optimize the network parameters, and use the test set to evaluate the model performance. During training, use the loss function to evaluate the error and adjust the network parameters according to the evaluation results. Repeat the training until the expected accuracy is achieved. Step S6: Input the current time and the historical L time-series data sequences of the target wind box and the exhaust gas temperature of the adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

[0027] In this embodiment, data acquisition and preprocessing serve as the foundation, ensuring the quality and consistency of the input data and providing reliable input for subsequent analysis. Sliding window slicing transforms continuous time series into discrete samples, constructing a learnable "history-future" mapping relationship for the model. The core local feature Transformer network employs a channel-independent design to separate and process multiple variables, avoiding interference between heterogeneous variables. Patch partitioning then cuts long sequences into local subsequences, enhancing the ability to capture short-term dynamic patterns. Finally, a self-attention mechanism is used to mine global long-term dependencies from these local features. The parameter setting stage configures the network structure and training rules according to the specific task, ensuring that the model capacity matches the task difficulty. During network training, the model parameters are continuously optimized through loss functions and backpropagation, gradually approximating the real physical process. Finally, the temperature prediction module combines the trained model with real-time data to achieve accurate prediction of future temperatures. These interconnected technical features form a complete technical chain from "data cleaning → feature extraction → relationship mining → model optimization → online prediction". Their organic connection helps to overcome the dependence of traditional mechanistic models on precise parameters, and can also solve the shortcomings of existing data-driven methods in capturing local features and long-term dependencies.

[0028] The above embodiment, by employing a combination of "channel independence + patch partitioning," precisely addresses the technical problem of existing Transformer models "ignoring local temporal patterns of variables" in multivariate time series prediction. Specifically, its underlying logic lies in the fact that each variable in the sintering process (such as temperature and machine speed) possesses unique, short-period fluctuation patterns. Traditional methods, by inputting multiple variables together, make it difficult for the model to focus on the local details of each variable. This embodiment first separates the variables and then performs localized slicing on each variable sequence, essentially providing the model with a "magnifying glass," enabling it to precisely extract the local dynamic features of each variable. Then, leveraging the powerful self-attention mechanism of the Transformer, it learns global, complex nonlinear coupling relationships from these refined local features. This constructs a hierarchical feature extraction architecture "from local to global," effectively improving the modeling ability for the complex dynamics of sintering conditions and, to a certain extent, enhancing prediction accuracy and model generalization ability.

[0029] In some embodiments, the sliding window slicing process in step S2 specifically includes: Multivariate time series are sliced ​​with window length L and step size 1. Each sample input is two-dimensional data of M×L, and the output is a prediction sequence of M×T, where M is the number of variables and T is the prediction length.

[0030] In this embodiment, the sliding window slicing is implemented as follows: For a historical dataset containing M=5 variables (e.g., the temperature of the three bellows, the speed of the trolley, and the thickness of the material layer), the window length L=96, the prediction length T=24, and the step size is 1. During processing, starting from the beginning of the time series data, all variable data within the most recent L=96 time steps are extracted each time step to form a 5×96 two-dimensional matrix as the model input; simultaneously, the target bellows temperature data within the next T=24 time steps immediately following the end of the input window are taken to form a 1×24 vector as the expected output of the model. Subsequently, the window slides forward one time step to generate the next sample, until the entire historical dataset has been traversed. It is easy to imagine that the values ​​of L, M, and T in the above embodiment can be set as needed; the preceding example is only to illustrate the sliding window slicing process in detail. For the sample set containing the input and output at each time step obtained after sliding window slicing of the input data, it is randomly sampled proportionally to divide it into a training set, a validation set, and a test set.

[0031] This embodiment effectively solves the fundamental yet crucial technical problem of "how to transform continuous time-series data into discrete samples that can be used for supervised learning model training" by strictly defining the input-output structure of the sliding window. Since the sintering process is a continuous dynamic process, the model needs to predict the future based on a historical context. A fixed-length sliding window ensures that each sample contains the latest and equal amount of historical information required for prediction. Defining the input and output as a two-dimensional data structure perfectly matches the input format requirements of subsequent deep learning models. This method guarantees the causality of time-series data (future data cannot be used to predict the past) and generates rich training samples through numerous overlapping windows, providing a solid foundation for model learning from a data perspective, thereby significantly improving data utilization efficiency and model training stability.

[0032] In some embodiments, the channel-independent mode in step S3 specifically refers to: The time series of each variable is input into the Transformer network as an independent channel for local feature extraction and encoding.

[0033] This embodiment aims to address the problem of "feature extraction interference caused by variable heterogeneity" in multivariate time series. The logical argument is as follows: different variables in the sintering process (such as temperature and velocity) have completely different physical meanings, dimensions, and dynamic characteristics. If they are mixed at the initial input stage, the model needs to make extra effort to distinguish and analyze information from different sources, easily leading to confusion. This embodiment's independent channel design essentially establishes a "dedicated feature extraction channel" for each variable, allowing the model to optimize feature learning for the unique time-series patterns of each variable (such as the inertia of temperature or the step change of velocity). This "separate-then-combine" strategy reduces interference factors in model learning, making the feature extraction process purer and more efficient, thereby enhancing the model's ability to capture the unique dynamic patterns of each variable and ultimately improving overall prediction accuracy.

[0034] In the above embodiments, exemplarily, the channel independence method in step S3 is specifically implemented as follows: when constructing the network input layer, the 5×96 two-dimensional input matrix prepared for each sample is logically separated into 5 independent one-dimensional time series of length 96. That is, the temperature sequence of the first bellows, the temperature sequence of the second bellows, the trolley speed sequence, the material layer thickness sequence, etc., are each treated as an independent "channel" data. During the forward propagation of the network, these 5 sequences are fed into the same feature extraction pipeline (including subsequent patch partitioning, embedding, and encoder) in parallel and independently. Only in the final stage of the encoder are the high-level features extracted from each variable fused for final prediction.

[0035] As a feasible embodiment, the specific steps for constructing the local feature Transformer network in step S3 are as follows: Channel independence: For raw acquired data containing M variables of length L and shape (M, L) By separating each variable as an independent channel, we obtain M input sequences of the form (1, L). Input the Transformer backbone network respectively; Univariate Patch Local Partitioning: For each variable sequence, perform padding and patching operations separately. At every step S, select a sequence of length P as a subsequence. This is particularly important for... If the last value is repeated (PS) times as a subsequence, then the number of subsequences is N = L / S Concatenate N subsequences of length P to obtain a two-dimensional input tensor of the form (P, N). ; Transformer network predicts univariate sequences: It takes a two-dimensional input... As input to the Transformer network, we obtain an output sequence of the form (1,T). As a variable The prediction results are shown. The encoder part of the Transformer network consists of a series of stacked encoder units, each connected by two sub-layers: a multi-head self-attention layer and a feedforward neural network layer. Each layer has a residual connection followed by a normalization step.

[0036] Concatenate prediction results: For each variable's prediction sequence After concatenation, a two-dimensional output of all variables in the form of (M,T) is obtained. Then, the relationship between the forecast results of each variable is captured by mapping to obtain the final single target forecast sequence.

[0037] In some embodiments, in step S3, the patch division specifically involves: For each variable sequence, a continuous subsequence of length P is extracted with a step size S, forming N = L / S Each local feature block is converted into a fixed-dimensional vector as the encoder input.

[0038] In this embodiment, N= L / S In this equation, L is the window length, S is the step size, and N is the number of local feature blocks. This equation represents the result being rounded up. This embodiment focuses on the dynamic control of node behavior in step S3. For example, in this embodiment, the patch partitioning in step S3 is specifically implemented as follows: for each univariate sequence of length L=96 obtained after channel independence, a continuous subsequence of length P=24 is truncated with a fixed step size S=12. For example, starting from the beginning of the sequence, points 1 to 24 are truncated as the first patch, points 13 to 36 as the second patch, and so on. If the last patch is less than length P, it is padded with the values ​​at the end of the sequence. For a sequence of length 96, this operation yields N=96 / 12=8 local feature blocks. Each patch (24-dimensional vector) is then transformed into a fixed-dimensional (e.g., 128-dimensional) vector through a trainable linear projection layer, which serves as the input to the Transformer encoder.

[0039] This embodiment creatively solves the technical challenge of "local temporal patterns being diluted by global attention" in long sequence prediction using traditional RNN / Transformer models by introducing the patch concept from the ViT model in computer vision. The embodiment sets the step size S, window P, and the number of encoder target embedding space features D during the embedding operation based on the sample set size, number of variables, and sequence length. Its core logic lies in the fact that certain key changes in the sintering process (such as a sudden temperature rise) often occur within a short, local time period. Cutting the long sequence into overlapping local patches is equivalent to providing the model with multiple "time slices," enabling it to focus on and reinforce the learning of subtle patterns and short-term dynamics within these local ranges. This is more direct and efficient than having the model directly search for patterns from the entire long sequence. Simultaneously, mapping the local sequences to high-dimensional vectors provides more informative basic computational units for subsequent attention mechanisms. Therefore, this embodiment not only proposes a new sequence processing method but also significantly improves the model's predictive sensitivity to transient conditions and short-term dynamics by strengthening local feature extraction, thereby enhancing the model's prediction accuracy and timeliness to a certain extent.

[0040] In some embodiments, the Transformer network encoder is composed of multiple encoder units stacked together, each unit including a multi-head self-attention layer, a feedforward neural network layer, and employing residual connections and layer normalization.

[0041] In this embodiment, the Transformer network encoder consists of N=3 stacked encoder units with identical structures. Each encoder unit sequentially includes: a multi-head self-attention layer (e.g., 8 heads) for enabling information exchange between patches at different positions in the sequence and calculating global dependencies; and a feedforward neural network layer (typically two fully connected layers and a ReLU activation function) for performing nonlinear transformations and feature refinement on the attention output. A "residual connection" is introduced before each sub-layer (self-attention layer and feedforward layer), which adds the sub-layer's input to its output; and immediately after the residual addition, a "layer normalization" operation is performed.

[0042] This embodiment effectively addresses the common problems of "vanishing / exploding gradients" and "model degradation" in deep neural network training by employing the mature yet crucial structure of "residual connections and layer normalization." The logic is as follows: as the number of network layers increases, information is gradually lost or distorted during transmission. Residual connections introduce a "shortcut," directly transmitting lower-level features to higher layers, ensuring effective gradient backpropagation and greatly alleviating the vanishing gradient problem, making it possible to build deeper, more powerful networks. Layer normalization, by standardizing the output of each layer, stabilizes the data distribution of each layer's input, accelerates the convergence process of model training, and improves robustness to hyperparameter changes. Therefore, this embodiment is not simply a stacking of network layers, but rather, through ingenious substructure design, ensures that the constructed deep Transformer network can be trained stably and efficiently, thereby ensuring that model performance is fully realized and improving the speed and stability of the training process.

[0043] In some embodiments, in step S5, mean squared error or mean absolute error is used as the loss function to optimize the network weights through backpropagation.

[0044] In this embodiment, the network training in step S5 is specifically implemented as follows: During training, the mean squared error (MSE) is selected as the loss function, which calculates the average squared error between the network's predicted temperature sequence for the next T=24 time points and the actual sequence. The optimizer used is Adam, with a learning rate set to 0.0001. In practice, the learning rate, number of iterations (epochs), and early stopping patience, among other hyperparameters, can be flexibly selected based on the sample set size and model complexity. Specifically, during training, the gradient of the loss function with respect to all network weights is calculated using the backpropagation algorithm, and the Adam optimizer updates the network parameters based on these gradients to continuously reduce the loss function value.

[0045] This embodiment directly addresses the core training problem of "how to quantify prediction errors and guide the model to learn in the right direction" by selecting the classic loss function MSE and combining it with the backpropagation mechanism. The underlying logic is that MSE penalizes larger prediction errors more severely, driving the model to prioritize and correct "difficult samples" with larger prediction biases, thus more effectively narrowing the overall distribution between predicted and true values. The backpropagation algorithm is the engine for achieving this goal; it uses a chain rule to precisely distribute the loss signal from the top layer to each parameter in the bottom layer, indicating the direction and magnitude of parameter updates. This combination of "objective (MSE) + method (backpropagation)" forms the cornerstone of deep learning model training, ensuring that the model's learning process is a supervised and systematic optimization process, thereby driving the model parameters to continuously iterate and ultimately approach the optimal solution.

[0046] In some embodiments, in step S6, the input data includes the target wind box and adjacent wind box temperatures, trolley speed and material layer thickness data at the current time and at the historical L times, and the output is the target wind box temperature sequence at the future T times.

[0047] In this embodiment, exemplarily, the prediction step in step S6 can be specifically implemented as follows: After the model is put into online application, the system continuously caches the latest field data at time L=96 (as mentioned above, each parameter can be set as needed). This includes the exhaust gas temperature of the target windbox and its adjacent windboxes, the trolley speed, and the material layer thickness, which together form a 5×96 real-time input matrix. This matrix is ​​input into the deployed and trained local feature Transformer network. After forward computation, the network finally outputs a 1×24 vector, which represents the predicted sequence of the target windbox exhaust gas temperature at time T=24 in the future (e.g., in the next 2 hours, if the sampling interval is 5 minutes).

[0048] This embodiment clarifies the input and output specifications of the model in a real-world industrial scenario, solving the last-mile problem of "how to interface the trained model with real-time production data to generate specific forecasts." The logic is that model training and application must follow a consistent interface. This embodiment strictly defines the source (target and adjacent bellows, machine speed, material layer), time range (L historical time points), and structure (multivariable two-dimensional matrix) of the input data during online prediction, ensuring that the model can perform inference within the same "context" as during training. The clearly defined output (target temperature in the next T time points) directly meets the core requirements of sintering process monitoring and operation optimization.

[0049] refer to Figure 2A second aspect of the present invention provides a sintering box temperature prediction system based on a local feature Transformer network, used to implement any of the above-described sintering box temperature prediction methods based on a local feature Transformer network, the system comprising: The data acquisition and preprocessing module is configured to acquire data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and to preprocess the data, including outlier removal and mean value processing. The sample construction module is configured to perform sliding window slicing on historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set. The network construction module is configured to build a local feature Transformer network. This network processes input data in a channel-independent manner, inputting the time series of each variable as an independent channel, performing patching on each channel sequence to extract local temporal features, and capturing the internal dependencies of variables through a self-attention mechanism. The parameter setting module is configured to determine network structure parameters, including embedding dimension, patch length and step size, based on the length of historical sequences, the length of predicted sequences and the number of variables, and to set the network's learning rate, number of iterations and early stopping patience hyperparameters. The network training module is configured to train the network using the training set, optimize the network parameters using the validation set, and evaluate the model performance using the test set. During training, a loss function is used to evaluate the error, and the network parameters are adjusted based on the evaluation results. Training is repeated until the expected accuracy is achieved. The temperature prediction module is configured to input the current time and the historical L time-series data sequences of the exhaust gas temperature of the target wind box and adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

[0050] exist Figure 2In the system architecture shown, solid and dashed lines represent two different types of key information flows. Solid arrows: represent the system's main data flow. It carries the core processing objects and depicts the complete processing chain from raw data to the final prediction result. The specific path is: Preprocessing data: from raw data to sample construction. Historical sample set and initial network: driving model construction and training. Trained model and real-time data slices: working together in the prediction module to generate the final target bellows temperature forecast sequence. Solid lines are the system's "production line," representing an indivisible, automatically executed, continuous data processing process. Dashed arrows: represent control flow or parameter flow. It does not directly process data but injects instructions and rules into the system on "how to work," issued by the parameter setting module, controlling: Network structure parameters: determining how the model is built (e.g., embedding dimension, patch length). Training hyperparameters: determining how the model learns (e.g., learning rate, number of iterations). Dashed lines are the system's "control switch," representing manual configuration or management instructions; they define and optimize the execution method of the solid line process.

[0051] In some embodiments, the system is specifically implemented as a software module integrated into the sintering plant's MES (Manufacturing Execution System). The data acquisition and preprocessing module communicates with the underlying PLC and sensors via the OPC UA protocol, automatically acquiring raw data and cleaning it using a preprocessing rule base. The sample construction module maintains a historical data buffer in memory and generates samples required for inference in real time according to preset window parameters (L=96, T=24). During system initialization, the network construction module instantiates a Transformer network with a specific structure (number of layers, number of heads, patch parameters) based on a configuration file. The parameter setting module provides a graphical interface, allowing engineers to adjust the network structure and hyperparameters when the model is first deployed or when there are significant changes in operating conditions. The network training module runs on a background server, calling historical datasets to perform training tasks and recording the training process. The temperature prediction module, as a core service, resides in memory, receives real-time samples from the sample construction module, performs high-speed forward inference, and publishes the prediction results to a real-time database for other systems (such as an operation guidance system) to subscribe to and use.

[0052] This embodiment addresses the technical challenge of "how to engineer and systematize advanced algorithms to achieve stable, reliable, and easily maintainable continuous operation in industrial settings" by mapping the aforementioned method entirely into a physical system composed of multiple highly cohesive and loosely coupled modules. The logic is that a single algorithm model cannot exist independently in an industrial environment; it needs to work collaboratively with data sources, human-machine interfaces, scheduling systems, and other components. The system design in this embodiment encapsulates each step of the algorithm as an independent, functionally defined module. This modular architecture ensures clear data and control flows, and upgrades or maintenance of any module (such as improving preprocessing algorithms) will not affect other modules, greatly enhancing the system's maintainability and scalability.

[0053] In some embodiments, such as Figure 3 This application also provides an electronic device, including a processor 301, a memory 302 and a bus 303, and a computer program stored in the memory. When the processor executes the program, it implements any of the above-described methods for predicting the temperature of a sintering box based on a local feature Transformer network.

[0054] Furthermore, as an executable solution, the sintering box temperature prediction system based on a local feature Transformer network can be a computer unit, which can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer unit may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above-described computer unit structure is merely an example and does not constitute a limitation on the computer unit. It may include more or fewer components, or combine certain components, or use different components. For example, the computer unit may also include input / output devices, network access devices, buses, etc., which are not limited in this embodiment of the invention.

[0055] Furthermore, as an executable solution, the processor can be a Central Processing Unit (CPU), or 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 can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer unit, connecting various parts of the entire computer unit via various interfaces and lines.

[0056] In some embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described above in the embodiments of the present invention.

[0057] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the computer unit by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory and non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0058] For example, in some embodiments, the electronic device is specifically implemented as an industrial server deployed in the sintering control room. This server is equipped with a high-performance multi-core processor (such as an Intel Xeon), large-capacity ECC memory (such as 64GB), and a solid-state drive (SSD) as storage. The computer program (i.e., the software system implementing a sintering bellows temperature prediction method based on a local feature Transformer network) is stored on the SSD. When the system starts, the processor loads program instructions and data from the SSD into memory for execution, thereby implementing a complete temperature prediction function from data acquisition, preprocessing, model inference to result output. The computer-readable storage medium is specifically a USB flash drive containing the system image file or an installation package file transmitted over a network, containing all executable code and configuration files of the computer program, used to copy and deploy the prediction system to other identical industrial servers.

[0059] This embodiment, by defining specific physical carriers (servers, processors, memory, storage media), essentially solves the practical technical problem of "how to transform an intangible technical solution into a tangible industrial product." Since any software method must ultimately be executed through hardware, by specifying the hardware environment (industrial server) and program carrier (storage media) for executing the prediction method, the scope of protection is extended from abstract methodological ideas to concrete physical entities and products. Therefore, this embodiment not only affirms the feasibility of the method but also expands its scope of protection to key aspects of industrialization and commercialization.

[0060] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for predicting the temperature of a sintering bellows based on a local feature Transformer network, characterized in that, Includes the following steps: Step S1: Collect data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and preprocess the data, including outlier removal and mean value processing. Step S2: Perform sliding window slicing on the historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set; Step S3: Construct a local feature Transformer network. The network processes the input data in a channel-independent manner, inputting the time series of each variable as an independent channel, performing patching on each channel sequence to extract temporal local features, and capturing the internal dependencies of variables through a self-attention mechanism. Step S4: Determine the network structure parameters, including embedding dimension, patch length, and stride, based on the historical sequence length, predicted sequence length, and number of variables; and set the network's learning rate, number of iterations, and early stopping patience hyperparameters. Step S5: Train the network using the training set, use the validation set to optimize the network parameters, and use the test set to evaluate the model performance. During training, use the loss function to evaluate the error and adjust the network parameters according to the evaluation results. Repeat the training until the expected accuracy is achieved. Step S6: Input the current time and the historical L time-series data sequences of the target wind box and the exhaust gas temperature of the adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

2. The method according to claim 1, characterized in that, In step S2, the sliding window slicing process specifically involves: Multivariate time series are sliced ​​with window length L and step size 1. Each sample input is two-dimensional data of M×L, and the output is a prediction sequence of M×T, where M is the number of variables and T is the prediction length.

3. The method according to claim 1, characterized in that, In step S3, the independent channel mode is specifically as follows: The time series of each variable is input into the Transformer network as an independent channel for local feature extraction and encoding.

4. The method according to claim 1, characterized in that, In step S3, the patch division is specifically as follows: For each variable sequence, a continuous subsequence of length P is extracted with a step size S, forming N = L / S Each local feature block is converted into a fixed-dimensional vector as the encoder input.

5. The method according to claim 1, characterized in that, The Transformer network encoder is composed of multiple encoder units stacked together. Each unit includes a multi-head self-attention layer, a feedforward neural network layer, and uses residual connections and layer normalization.

6. The method according to claim 1, characterized in that, In step S5, mean squared error or mean absolute error is used as the loss function to optimize the network weights through backpropagation.

7. The method according to claim 1, characterized in that, In step S6, the input data includes the target wind box temperature and adjacent wind box temperature, trolley speed and material layer thickness data at the current time and the historical L time points, and the output is the target wind box temperature sequence at the future T time points.

8. A sintering windbox temperature prediction system based on a local feature Transformer network, used to implement the method as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition and preprocessing module is configured to acquire data on exhaust gas temperature, trolley speed and material layer thickness of the target air box and adjacent air boxes, and to preprocess the data, including outlier removal and mean value processing. The sample construction module is configured to perform sliding window slicing on historical data to construct a sample set, and divide the sample set into a training set, a validation set, and a test set. The network construction module is configured to build a local feature Transformer network. The network processes input data in a channel-independent manner, inputting the time series of each variable as an independent channel, performing patch division on each channel sequence to extract temporal local features, and capturing the internal dependencies of variables through a self-attention mechanism. The parameter setting module is configured to determine network structure parameters, including embedding dimension, patch length and step size, based on the length of historical sequences, the length of predicted sequences and the number of variables, and to set the network's learning rate, number of iterations and early stopping patience hyperparameters. The network training module is configured to train the network using the training set, optimize the network parameters using the validation set, and evaluate the model performance using the test set. During training, a loss function is used to evaluate the error, and the network parameters are adjusted based on the evaluation results. Training is repeated until the expected accuracy is achieved. The temperature prediction module is configured to input the current time and the historical L time-series data sequences of the exhaust gas temperature of the target wind box and adjacent wind boxes, the trolley speed and the material layer thickness into the trained network, and output the target wind box exhaust gas temperature prediction sequence for the next T time-series.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images