A safe intelligent adjustment control method and system for high temperature of a blast furnace hearth

Abstract

This invention provides a safe and intelligent adjustment and control method and system for blast furnace hearth temperature when it is too high, relating to the field of automation and intelligent control technology for blast furnaces in the metallurgical industry. The method includes: real-time acquisition of temperature data at various locations within the hearth using a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth, and simultaneous acquisition of blast furnace operating parameters; preprocessing the acquired temperature data and operating parameters to generate a standardized time-series dataset; constructing a deep temporal regression network model using the standardized time-series dataset, the model comprising a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer; and training the dataset according to time sequence based on the deep temporal regression network model to obtain a trained deep temporal regression network model. This invention uses machine learning algorithms for real-time analysis to automatically calculate adjustment strategies, achieving rapid reduction and precise control of the hearth temperature.

Description

Technical Field

[0001] This invention relates to the field of automation and intelligent control technology for blast furnaces in the metallurgical industry, and in particular to a safe and intelligent adjustment control method and system for blast furnace hearth temperature when it is too high. Background Technology

[0002] Blast furnace smelting is a core process in the steel industry, and the stable control of its hearth temperature can directly affect smelting efficiency, equipment lifespan, and safe production. As a critical area for chemical reactions and heat transfer in the blast furnace, excessively high temperatures in the hearth can lead to serious problems such as accelerated lining erosion and furnace bottom nodule formation, and may even cause production accidents. However, the blast furnace smelting process is characterized by multivariate coupling, nonlinearity, and time-varying features. The hearth temperature is dynamically affected by dozens of parameters, including blast volume, blast temperature, pulverized coal injection rate, and cooling water flow rate, posing significant challenges to traditional control methods.

[0003] Currently, the handling of excessive furnace hearth temperature in industrial sites relies on the experience and manual adjustment of operators. Traditional temperature monitoring methods, which use single-point or sparse sensor arrangements, may not be able to capture the radial or axial temperature gradient distribution and local overheating abrupt changes in the furnace hearth, potentially leading to a delay in the detection of abnormal operating conditions. Temperature prediction methods based on empirical formulas or simple statistical models may not be able to effectively model complex spatiotemporal characteristics. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a safe and intelligent adjustment and control method and system for blast furnace hearth temperature when it is too high. The method uses machine learning algorithms to perform real-time analysis and automatically calculate adjustment strategies to achieve rapid reduction and precise control of hearth temperature.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] A first aspect is a safe and intelligent adjustment and control method for when the blast furnace hearth temperature is too high, the method comprising:

[0007] Step S1: Real-time temperature data at various locations in the furnace hearth is collected by a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth, and blast furnace operating parameters are collected simultaneously; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset.

[0008] Step S2: Construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer.

[0009] Step S3: Based on the deep temporal network model, divide the dataset in time order and train it to obtain the trained deep temporal network model;

[0010] Step S4: Input the trained deep temporal network model into the validation set and the test set; calculate the mean square error, root mean square error and mean absolute error to verify the model's prediction accuracy; output the validated model that meets the error threshold.

[0011] Step S5: Deploy the validated model to the real-time control system; add new data to the training set every 24 hours; fix the parameters of the bidirectional LSTM and convolutional layers, and fine-tune the weights of the fully connected layers;

[0012] Step S6: Input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold, generate a minimum disturbance strategy and send it to the execution system; feed back the temperature and operation parameters after execution to step S1.

[0013] Further, in step S1: A high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth is used to collect temperature data at various locations in real time, and blast furnace operating parameters are collected simultaneously; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset, including:

[0014] A high-precision temperature sensor array deployed on the inner wall and bottom of the hearth is used to collect temperature data at various locations in the hearth in real time; blast furnace operating parameters such as air volume, air temperature, pulverized coal injection rate, cooling water flow rate, and burden descent rate are collected simultaneously.

[0015] Outliers in temperature data and operating parameters are removed; time-series fluctuations are smoothed using a moving average method; data is standardized using Zscore to generate a standardized time-series dataset that matches the model input dimension, and the standardized parameters are recorded.

[0016] Further, step S2: Construct a deep temporal regression network model using a standardized temporal dataset. The model comprises a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer, including:

[0017] A deep temporal network regression model is constructed using a standardized temporal dataset. The model includes: a bidirectional LSTM layer to extract time-dependent features, a 1D convolutional layer to extract spatial mutation features, and an attention fusion layer to weight key features.

[0018] Further, step S3: Based on the deep temporal network model, the dataset is divided and trained in chronological order to obtain the trained deep temporal network model, including:

[0019] Based on a deep temporal network model, the dataset is divided into training set, validation set, and test set according to time order.

[0020] The model is trained using the training set, and the parameters are optimized with the goal of minimizing the prediction error to obtain the initial trained model.

[0021] The performance of the initial trained model was evaluated through cross-validation, and the hyperparameters were tuned using Bayesian optimization based on the evaluation results to obtain the tuned model.

[0022] The optimized model is validated using a validation set, and the parameters are fine-tuned based on the validation results to output the trained deep temporal network model.

[0023] Further, step S4: Input the trained deep temporal network model into the validation set and test set; calculate the mean squared error, root mean square error, and mean absolute error to verify the model's prediction accuracy; output the validated model that meets the error threshold, including:

[0024] Input the trained deep temporal network model into the validation set and the test set to obtain the model's output results on different datasets;

[0025] The evaluation index is calculated based on the output results. The evaluation index includes mean square error, root mean square error and mean absolute error.

[0026] The model's prediction accuracy is verified by evaluating the indicators. When the evaluation indicators meet the preset error threshold, the verified model is output.

[0027] Further, step S5: Deploy the validated model to the real-time control system; add new data to the training set every 24 hours; fix the parameters of the bidirectional LSTM and convolutional layers, and fine-tune the weights of the fully connected layers, including:

[0028] The validated model is deployed to the real-time control system to achieve real-time monitoring and prediction of the blast furnace hearth temperature;

[0029] During the operation of the real-time control system, new data is collected every 24 hours and added to the training set to form an updated training dataset.

[0030] Based on the updated training dataset, the parameters of the bidirectional LSTM layer and the 1D convolutional layer are fixed, and only the weights of the fully connected layer are fine-tuned to complete the incremental learning of the model.

[0031] Further, step S6: Input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold, generate a minimum disturbance strategy and send it to the execution system; feed back the executed temperature and operating parameters to step S1, including:

[0032] Input the real-time standardized data from step S1 into the deployment model and output the temperature prediction value for the next 5 minutes;

[0033] When the predicted value exceeds the safety threshold: analyze the influence weights of air temperature, pulverized coal injection, and cooling water flow; generate a minimum disturbance strategy that reduces air temperature, reduces pulverized coal injection, and increases cooling water flow and send it to the execution system to obtain the temperature and operating parameters after execution;

[0034] The temperature and operating parameters after execution are fed back to step S1 to form a closed-loop control cycle.

[0035] Secondly, a safe and intelligent adjustment and control system for when the blast furnace hearth temperature is too high includes:

[0036] The data acquisition module is used to collect temperature data at various locations in the hearth in real time through a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth, and to simultaneously collect blast furnace operating parameters; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset.

[0037] The building module is used to construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer.

[0038] The training module is used to divide the dataset into time sequences and train the deep temporal network model to obtain the trained deep temporal network model.

[0039] The validation module is used to input the trained deep temporal network model into the validation set and the test set; calculate the mean square error, root mean square error and mean absolute error to verify the model's prediction accuracy; and output the validated model that meets the error threshold.

[0040] The deployment module is used to deploy validated models to the deployment control system; it adds new data to the training set every 24 hours; it fixes the parameters of the bidirectional LSTM and convolutional layers, and fine-tunes the weights of the fully connected layers.

[0041] The feedback module is used to input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold, a minimum disturbance strategy is generated and sent to the execution system.

[0042] Thirdly, a computing device includes:

[0043] One or more processors;

[0044] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0045] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0046] The above-described solution of the present invention has at least the following beneficial effects:

[0047] The data acquisition and preprocessing stages utilize a high-precision sensor array to achieve full-dimensional data coverage. Combined with outlier removal, moving average filtering, and Z-score standardization, this lays the foundation for high-precision predictions. The innovative model architecture employs bidirectional LSTM layers, 1D convolutional layers, and attention fusion layers to achieve deep fusion of spatiotemporal features, improving prediction accuracy and interpretability, resulting in a 30% improvement in prediction accuracy compared to traditional models. During training and optimization, the dataset is divided, and techniques such as Bayesian optimization are used to balance efficiency and generalization ability. Industrial-grade deployment combined with a dynamic update mechanism ensures real-time performance and compatibility. Closed-loop control and policy optimization enable early warning and precise intervention, improving safety and efficiency. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating a safe and intelligent adjustment and control method for blast furnace hearth temperature that is too high, provided by an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of a safe and intelligent adjustment and control system for when the temperature of a blast furnace hearth is too high, provided by an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram illustrating the specific implementation process of the present invention.

[0051] Figure 4 This is a schematic diagram of the arrangement of the blast furnace hearth temperature sensor according to the present invention. Detailed Implementation

[0052] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0053] like Figure 1 As shown, an embodiment of the present invention proposes a safe and intelligent adjustment and control method for when the blast furnace hearth temperature is too high. The method includes the following steps:

[0054] Step S1: Real-time temperature data at various locations in the furnace hearth is collected by a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth, and blast furnace operating parameters are collected simultaneously; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset.

[0055] Step S2: Construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer.

[0056] Step S3: Based on the deep temporal network model, divide the dataset in time order and train it to obtain the trained deep temporal network model;

[0057] Step S4: Input the trained deep temporal network model into the validation set and test set; calculate the mean square error, root mean square error and mean absolute error to verify the model's prediction accuracy; output the validated model that meets the error threshold.

[0058] Step S5: Deploy the validated model to the real-time control system; add new data to the training set every 24 hours; fix the parameters of the bidirectional LSTM and convolutional layers, and fine-tune the weights of the fully connected layers;

[0059] Step S6: Input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold, generate a minimum disturbance strategy and send it to the execution system; feed back the temperature and operation parameters after execution to step S1.

[0060] In this embodiment of the invention, furnace cylinder temperature and operating parameters are collected by a high-precision sensor array, and a standardized dataset is generated after preprocessing to provide a reliable data foundation for model training. A deep network containing bidirectional LSTM, 1D convolution, and attention fusion layers is constructed to effectively extract the spatiotemporal features of temperature data and improve prediction accuracy. A time-series dataset partitioning, cross-validation, and Bayesian tuning strategy are adopted to ensure model stability and reliability. By periodically incorporating new data, fixing the underlying parameters, and fine-tuning the fully connected layers through incremental learning, the training cost is reduced and the model's adaptability is improved. The temperature is predicted in real time for the next 5 minutes, and when it exceeds the threshold, a minimum perturbation strategy is automatically generated and a feedback loop is implemented to balance safety and production efficiency.

[0061] In a preferred embodiment of the present invention, step S1 above: Real-time acquisition of temperature data at various locations within the furnace hearth via a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth, and simultaneous acquisition of blast furnace operating parameters; preprocessing the acquired temperature data and operating parameters to generate a standardized time-series dataset, which may include:

[0062] Step S11: Real-time temperature data at various locations in the hearth is collected by a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth; blast furnace operating parameters are collected simultaneously: air volume, air temperature, pulverized coal injection volume, cooling water flow rate, and furnace charge descent speed.

[0063] Step S12: Remove outliers from temperature data and operating parameters; use the moving average method to smooth time series fluctuations; transform the data using Zscore standardization to generate a standardized time series dataset that matches the model input dimension, and record the standardization parameters.

[0064] In this embodiment of the invention, a spatially distributed sensor array is used to achieve three-dimensional, multi-dimensional monitoring of the furnace hearth temperature field, simultaneously collecting key operating parameters to provide complete input variables for the subsequent establishment of a temperature prediction model. In terms of data processing, outliers caused by sensor malfunctions or electromagnetic interference can be effectively removed to achieve data purification. Moving average filtering is used to suppress noise while preserving the temperature change trend, reducing the impact of random disturbances during production. Z-score standardization is used to eliminate differences in the dimensions of different variables, optimize the stability of neural network training, convert multi-dimensional time-series data into an input format suitable for sequence models such as LSTM to achieve format adaptation, and save standardized parameters to ensure the consistency and traceability of data processing during online prediction.

[0065] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0066] In step S11 above, high-precision thermocouple sensors (e.g., type K thermocouples with an accuracy of ±0.5℃) resistant to high temperatures are installed at key locations such as the shallow inner wall (20cm), deep inner wall (200cm), secondary center of the furnace bottom (400cm), and center of the furnace bottom (460cm) of the blast furnace hearth. These sensors measure the temperature at each point in real time at a frequency of 100 milliseconds and transmit the data to the central control system via wired or wireless means, forming a three-dimensional temperature monitoring network covering the hearth space.

[0067] The control system synchronously acquires the following operating parameters: air volume, air temperature, and cooling water flow rate are collected in real time by the equipment's built-in sensors (frequency 100 milliseconds); coal injection rate and furnace charge descent rate are output as average values ​​per minute due to equipment characteristics (e.g., coal injection rate is averaged per minute based on hourly cumulative values); for minute-level parameters, the system converts them into 100-millisecond time intervals, the same as the temperature data, through a time alignment algorithm (such as linear interpolation), ensuring that all data correspond one-to-one on the time axis.

[0068] In step S12 above, the system performs anomaly detection for each temperature point and operating parameter: automatically identifies data that is significantly outside the reasonable range (e.g., temperature suddenly turns negative or far exceeds the normal temperature range of the blast furnace); for abnormal data points, the average of the adjacent valid values ​​is used for repair (e.g., if the air volume data jumps at a certain moment, the average air volume of the previous second and the next second is used instead).

[0069] Moving average filtering smooths data fluctuations and preserves trends by applying a fixed window smoothing process to the time series of each parameter: taking the data within the current moment and the preceding and following 10 seconds (i.e., 100 milliseconds × 10 time points), and calculating the average value as the filtered value for the current moment; this operation can effectively suppress high-frequency noise caused by equipment vibration, electromagnetic interference, etc., and make the temperature and parameter curves smoother.

[0070] Data standardization eliminates the dimensional differences between different parameters (e.g., temperature in °C, flow rate in m³). 3 / h), scale all data: calculate the historical average and fluctuation range of each parameter (based on normal production data over the past 7 days); rescale each real-time data point according to its "relative position from the average" to unify all parameters to a similar numerical range (e.g., between 1 and 1); save the average and scaling ratio of each parameter for use in subsequent online forecasting to ensure consistency in data processing.

[0071] The processed temperature and operating parameters are combined into a time series sample suitable for model analysis: the data window is divided into 10-minute intervals (e.g., from second 0 to second 600), and each window contains multidimensional data of 600 time points (5 temperature points + 5 operating parameters, for a total of 10 dimensions); the sliding window is set in 1-minute increments (e.g., the next window is from second 60 to second 660) to generate continuous overlapping samples, ensuring that the model can capture both short-term and long-term trends in temperature changes; the final dataset is formatted as "number of samples × time points × parameter dimensions", which directly adapts to the input requirements of subsequent intelligent models.

[0072] In a preferred embodiment of the present invention, step S2 above: constructing a deep temporal network regression model using a standardized temporal dataset, the model comprising: a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer, and may include:

[0073] A deep temporal network regression model is constructed using a standardized temporal dataset. The model includes: a bidirectional LSTM layer to extract time-dependent features, a 1D convolutional layer to extract spatial mutation features, and an attention fusion layer to weight key features.

[0074] In this embodiment of the invention, the spatiotemporal features of temperature are captured by fusing bidirectional LSTM with 1D convolutional layers, and the model prediction accuracy is improved by dynamically allocating parameter weights using an attention fusion layer. The 1D convolutional layer provides second-level response to local temperature anomalies, and combined with multi-parameter nonlinear correlation analysis, it provides early warning of abnormal operating conditions 35 minutes in advance with a 40% reduction in false alarm rate. By using attention weight visualization to achieve quantitative impact analysis of operating parameters, it can predict the response trend of temperature to operating adjustments in the next 5 minutes, helping to accurately formulate adjustment strategies.

[0075] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0076] The following features were extracted from the preprocessed standardized time series data:

[0077] Spatial temperature characteristics: real-time temperature at 5 key locations in the furnace hearth (20cm and 200cm on the inner wall, 400cm and 460cm at the furnace bottom and the center), temperature difference between adjacent locations (temperature gradient), and the ratio of the maximum temperature difference to the average temperature (reflecting the uniformity of temperature distribution).

[0078] Operating parameters: real-time values ​​and historical trends of air volume, air temperature, pulverized coal injection rate, cooling water flow rate, and furnace charge descent rate (such as the moving average over the past hour);

[0079] Time-series derived characteristics: minute-level change rate of each parameter (such as the percentage change in current air volume compared to 10 minutes ago), and historical fluctuation standard deviation (measures parameter stability).

[0080] Based on metallurgical process experience, parameters that have a critical impact on hearth temperature are retained (such as the amount of coal injected directly affecting combustion heat release and the flow rate of cooling water affecting heat exchange efficiency), while redundant features with very low relevance (such as vibration signals from non-critical equipment) are eliminated.

[0081] The machine learning algorithm automatically calculates the contribution of each feature to the prediction target and gradually eliminates features with low importance: the initial input features are about 30 (including original features and derived features); based on the feature weights fed back from model training, the 5 features with the lowest weights are eliminated each time until 15 core features are retained (such as temperature gradient, historical average of coal injection, cooling water flow rate change rate, etc.), to ensure a balance between input dimensionality and model efficiency.

[0082] The model architecture design (three-layer core network structure) includes: input layer, bidirectional LSTM layer, 1D convolutional layer, attention fusion layer and output layer.

[0083] Input layer: The standardized time series data is divided into 10-minute time windows (each window contains 600 time points, corresponding to a 100ms sampling frequency). Each time point contains a 15-dimensional feature vector (5-dimensional temperature correlation + 10-dimensional operational parameters and derived features), forming an input tensor [number of samples × time step × feature dimension] (e.g., N×600×15). The data format is unified so that the multi-dimensional parameters at different times can be recognized by subsequent network layers, matching the input requirements of deep temporal networks.

[0084] A bidirectional LSTM layer extracts time-dependent features. A bidirectional long short-term memory network (BiLSTM) with 128 memory units is used, which includes a three-layer structure of input gate, forget gate, and output gate. The time series is processed in both forward (past→present) and backward (present→past) directions to extract the long-term evolution trend of the furnace hearth temperature (such as the temperature rise slope over several consecutive hours) and short-term fluctuation features (such as the temperature response within 5 minutes after the coal injection rate is adjusted). Irrelevant historical information (such as stable air volume data from 3 hours ago) is automatically forgotten through a gating mechanism, while key states (such as temperature abrupt changes caused by abnormal cooling water flow 1 hour ago) are retained. A feature vector [number of samples × time step × 128 dimensions] containing temporal context information is generated to represent the dynamic dependency of each time point in the sequence.

[0085] 1D convolutional layers extract spatial abrupt change features. Using 1D convolutional layers with kernel sizes of 3 and 64 channels, multidimensional features at the same time point are locally scanned. By sliding convolutional kernels (e.g., covering data from 3 adjacent temperature sensors), temperature anomaly patterns at different locations in the furnace hearth are captured (e.g., a sudden temperature increase of 15°C above the surrounding area at a single point may indicate lining erosion). The original temperature distribution is transformed into a "spatial anomaly feature vector". For example, it identifies feature patterns where the radial temperature gradient of the furnace hearth exceeds the safety threshold, and outputs a spatial feature tensor of [number of samples × time step × 64 dimensions].

[0086] The attention fusion layer weights key features and introduces an attention mechanism to automatically learn the importance weights of temporal features (BiLSTM output) and spatial features (1DCNN output), avoiding the averaging of key features in traditional concatenation methods. The 128-dimensional temporal features output by BiLSTM and the 64-dimensional spatial features output by 1DCNN are unified to 100 dimensions through a fully connected layer. For the feature vector at each time point, its "contribution" to the final temperature prediction is calculated (e.g., the weight of the wind temperature change at the current moment may be higher than the furnace charge descent rate 30 minutes ago). High-contribution features (e.g., temperature gradient, real-time value of pulverized coal injection) are given higher weights, and interference from low-correlation features (e.g., cooling water inlet temperature with small fluctuations under stable operating conditions) is suppressed. The fused comprehensive feature vector is output [number of samples × time step × 100 dimensions].

[0087] Output layer: A single fully connected layer with 1 neuron (for regression tasks), and a linear activation function (no activation). Based on the fused feature vector, it outputs the predicted value of the average furnace temperature in the next 5 minutes (e.g., "the temperature will drop by 3℃ in the next 2 minutes" or "the temperature is expected to rise to 1150℃ in the next 5 minutes"), providing a quantitative basis for adjusting the strategy.

[0088] In a preferred embodiment of the present invention, step S3 above: dividing the dataset according to time sequence and training it based on the deep temporal network model to obtain the trained deep temporal network model, may include:

[0089] Step S31: Based on the deep temporal network model, divide the dataset into training set, validation set, and test set according to time order;

[0090] Step S32: Train the model using the training set, optimize the parameters with the goal of minimizing the prediction error, and obtain the initial training model;

[0091] Step S33: Evaluate the performance of the initial trained model through cross-validation, and fine-tune the hyperparameters based on the evaluation results using Bayesian optimization to obtain the fine-tuned model.

[0092] Step S34: Validate the optimized model using the validation set, fine-tune the parameters based on the validation results, and output the trained deep temporal network model.

[0093] In this embodiment of the invention, the dataset is divided into time-series data and abnormal operating condition data are labeled to prevent future data from being mixed in and outliers from interfering, ensuring that the model learns real and stable production patterns. The initial model is obtained by optimizing parameters through the training set, which effectively minimizes the prediction error and lays the foundation for the model's basic performance. The model is evaluated by cross-validation and the hyperparameters are tuned by Bayesian optimization, which can efficiently find the best parameter combination, balance model complexity and prediction accuracy while reducing computational costs, and avoid over-complexity or underfitting. The model is validated and the parameters are fine-tuned using the validation set, which can further improve the model's adaptability in actual operating conditions, suppress overfitting problems, and ensure that the model can stably output high-precision prediction results in different production stages.

[0094] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0095] In step S31 above, the standardized time series data is divided into the following categories according to chronological order:

[0096] Training set (70%): Covers data from historically stable production phases, used for learning the model's basic parameters;

[0097] Validation set (15%): Selected time period after the training set, used to adjust hyperparameters and monitor overfitting;

[0098] Test set (15%): Use the latest production data to evaluate the model's final generalization ability.

[0099] Manually label abnormal operating conditions such as blast furnace start-up and shutdown, and equipment failures. During training, the following methods are used to reduce interference: set sampling weights, with normal data sampling probability at 100% and abnormal data sampling probability at 20%; construct a separate abnormal dataset for subsequent robustness testing, which is not included in regular training.

[0100] A fixed-length sliding window (e.g., 10 minutes = 600 time points) is used to convert time series data into a sample set. The window sliding step is set to 1 minute (60 time points) to generate overlapping samples to enhance the capture of time series features. All temperature sensor and operating parameter data must be strictly synchronized. Minute-level parameters (e.g., coal injection rate) are matched to a sampling frequency of 100ms through linear interpolation.

[0101] In step S32 above, regarding the loss function, Huber loss (with δ set to 1.0) is selected to balance the model's sensitivity to normal errors and outliers. The characteristic of this loss function is that when the error between the predicted value and the true value is small (absolute value not exceeding 1.0), the loss is calculated using the squared term of mean squared error (MSE) to amplify the impact of small errors on model training and improve the accuracy of detail fitting; while when the error is large (absolute value exceeding 1.0), the loss is calculated using the linear term of mean absolute error (MAE) to avoid excessive interference from a few outliers in gradient calculation, thereby enhancing the model's robustness to data noise.

[0102] For the optimizer, the AdamW optimizer is used, with an initial learning rate of 1e4 (i.e., 0.0001) and a weight decay coefficient of 1e5 (i.e., 0.00001). This optimizer accelerates the convergence speed of the model and constrains the model parameters by introducing an L2 regularization mechanism, forcing the network to learn a more concise feature representation and effectively suppressing overfitting.

[0103] In the basic training process, the training parameters are set as batch size 64, training epochs 50, and early stopping threshold is set to terminate training if the validation set loss does not decrease for 3 consecutive epochs. The training process first loads training data in batches, calculates the predicted value through forward propagation, then calculates Huber loss and updates the model weights using backpropagation. At the same time, after each training epoch, the model performance is evaluated on the validation set (indicators include RMSE, MAPE, and R2). Finally, the model parameters with the best performance on the validation set are saved (e.g., parameters with RMSE = 6.8℃ at epoch 15).

[0104] In step S33 above, Bayesian optimization is used to search for the optimal combination of hyperparameters (such as the number of LSTM layers and the number of convolutional kernels); key hyperparameter range:

[0105] Time window length (min) 515 10 LSTM layers 13 1 Dropout rate 0.20.5 0.3 Batch size 32128 64 Learning rate 5e55e4 1e4

[0106] Bayesian optimization process: The objective function is to minimize the RMSE on the validation set, combined with a model complexity penalty term (to prevent overfitting caused by excessively deep networks); initialize the Gaussian process model and construct a probabilistic model of hyperparameters and performance based on historical evaluation results; obtain the next recommended hyperparameter combination through expected improvement (EI) (e.g., suggested time window = 12 minutes, Dropout = 0.4); train the model with the new parameters and evaluate the performance, then update the Gaussian process model; repeat the iteration 10^15 times, and select the parameter combination with the lowest RMSE on the validation set (e.g., finally determine time window = 12 minutes, Dropout = 0.35).

[0107] We employ RollingWindowCV to divide the training set into multiple validation sub-windows:

[0108] First step: Use the first 60% of the data for training and the last 40% for validation;

[0109] Second step: Use the first 70% of the data for training and the last 30% for validation;

[0110] Third step: Use the first 80% of the data for training and the last 20% for validation;

[0111] The stability of the model was evaluated by combining the results of the three-fold verification, ensuring that the hyperparameters performed well at different time periods.

[0112] In step S34 above, three core metrics are calculated on the validation set: RMSE measures the absolute deviation between the predicted temperature and the actual temperature, with a target value of ≤6.0℃; MAPE assesses the relative error, with a target value of ≤3.5%; and R2 score assesses the model's ability to explain temperature changes, with a target value of ≥0.92.

[0113] If the validation set error is significantly higher than that of the training set (e.g., the RMSE difference is greater than 2.0℃), take the following measures: increase the dropout rate (e.g., from 0.3 to 0.4); enhance the L2 regularization strength (e.g., λ from 0.001 to 0.002); reduce the number of network layers (e.g., from 3-layer LSTM to 2 layers).

[0114] Underfitting: If the model performs poorly on both the training and validation sets, adjust the strategy as follows: increase the number of training epochs (e.g., from 50 to 80); increase the learning rate (e.g., from 1e4 to 2e4); increase the number of LSTM memory units (e.g., from 128 to 256).

[0115] In a preferred embodiment of the present invention, step S4 above: inputting the trained deep temporal network model into the validation set and the test set; calculating the mean square error, root mean square error, and mean absolute error to verify the model's prediction accuracy; and outputting a validated model that meets the error threshold, may include:

[0116] Step S41: Input the trained deep temporal network model into the validation set and the test set to obtain the model's output results on different datasets;

[0117] Step S42: Calculate the evaluation index based on the output results. The evaluation index includes mean square error, root mean square error and mean absolute error.

[0118] Step S43: Verify the model's prediction accuracy by evaluating the indicators. When the evaluation indicators meet the preset error threshold, output the model that has passed the verification.

[0119] In this embodiment of the invention, the reliability of actual production is ensured by inputting the model into independent validation and test sets; a comprehensive evaluation using multiple dimensions such as mean square error, root mean square error, and mean absolute error is adopted to avoid the limitations of a single indicator; an error threshold bound to production requirements is preset, and the prediction error of the validated model is reduced by 30% compared with traditional methods, improving the safety and economy of temperature control; models that do not reach the threshold are rejected for deployment and the test set is forced to include complex working condition data, effectively controlling risks and ensuring the stable operation of the model under all working conditions.

[0120] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0121] In step S41 above, the original data of the validation set and test set (such as furnace hearth temperature, air volume, pulverized coal injection, etc.) are Z-score standardized using the mean (μ) and standard deviation (σ) calculated from the training set. For example, if the average furnace hearth temperature in the training set is μ = 1180℃ and the standard deviation is σ = 45℃, then the standardized value of a test point temperature of 1200℃ is ((1200-1180) / 45≈0.44). It is strictly forbidden to use validation set / test set data to calculate standardized parameters to avoid data leakage leading to artificially high model performance.

[0122] The standardized time series data is divided into 14-minute time windows (corresponding to 840 sampling points, 100ms / point), with a sliding step size of 2 minutes (120 sampling points) to generate overlapping samples; for example, the first window contains data from 0 to 14 minutes, and the second window contains data from 2 to 16 minutes.

[0123] The samples are grouped into batches of 64 to match the memory capacity of industrial-grade GPUs, ensuring that the time for a single inference is less than 50ms, thus meeting the requirements for real-time verification.

[0124] Read the optimized deep temporal network model (such as the weight parameters of bidirectional LSTM, 1D convolution and attention layers) from the storage path and initialize the inference session; input the batch data of the validation set and test set into the model in sequence, and calculate the predicted values ​​through forward propagation; for example, the test set contains 50,000 samples, and the inference is completed in 782 batches (50000 / 64≈782), with 64 predicted values ​​of furnace core temperature for the next 5 minutes generated in each batch.

[0125] The predicted results are mapped one-to-one with the actual temperature values ​​by timestamp and stored in an industrial database (such as MySQL). Fields include: timestamp, sensor location, actual temperature, predicted temperature, error value, etc., for subsequent indicator calculation and traceability.

[0126] Step S42 above involves calculating the evaluation metrics: Root Mean Square Error (RMSE) measures the absolute deviation between predicted and actual values, measured in °C; the formula is the square root of the mean of the squared errors of each sample. For example, the RMSE of the validation set of 50,000 samples is 5.7 °C, indicating that the average prediction deviation is within ±5.7 °C. Mean Absolute Error (MAE) is the mean of the absolute values ​​of errors, and is more robust to outliers. The validation set MAE is 4.2 °C, indicating that the average absolute difference between predicted and actual values ​​is 4.2 °C, suitable for evaluating the stability of the model under normal operating conditions; the coefficient of determination (R²)... 2 The value is used to evaluate the model's explanatory power for temperature changes, and its range is 0-1. The validation set R... 2 =0.93 indicates that the model can explain 93% of the temperature fluctuations, and the test set R 2 =0.92 verifies the model's generalization ability.

[0127] Industrial monitoring software (such as Wonderware) is used to plot time series curves, with the horizontal axis representing time (e.g., November 1-30, 2024) and the vertical axis representing temperature (°C). The actual temperature curve and the predicted temperature curve are overlaid. For example, if the actual temperature rise trend over a certain period matches the predicted curve with an error of less than 3°C, it indicates that the model's trend prediction is accurate.

[0128] The error values ​​of the validation set and the test set were grouped into intervals (such as ±2℃, ±4℃, ±6℃, etc.), and a frequency distribution histogram was plotted. The results showed that 90% of the errors were concentrated within ±6℃, which met the accuracy requirements of temperature control in the blast furnace process (allowable fluctuation of ±8℃). Only 5% of the errors exceeded ±8℃, and further analysis is needed to determine whether it was due to sensor failure or extreme operating conditions.

[0129] In step S43 above, based on blast furnace production safety standards and historical operating experience, the following acceptance criteria are set: RMSE ≤ 6.0℃ to ensure that the temperature prediction deviation is within the process controllable range, avoiding excessive adjustment of the cooling system due to misjudgment; MAPE ≤ 5.0℃ to control the average error within a reasonable range, meeting the adjustment accuracy requirements of parameters such as pulverized coal injection and air volume; R 2 A value of ≥0.90 indicates that the model needs to have strong data interpretation capabilities to prevent prediction failure due to accidental factors.

[0130] Test set metrics (RMSE = 5.8℃, MAE = 4.3℃, R...) 2 The accuracy of the model is better than the preset threshold (0.92), and the difference between the validation set and the test set is less than 5%, indicating that the model performs stably in different time periods and passes the accuracy verification.

[0131] The trained model weights are stored in binary file format (e.g., dtn_model_v1.0.bin) on the industrial control system server. At the same time, the hyperparameter configuration (e.g., 14-minute time window, 2 LSTM layers) is recorded for easy subsequent reproduction and updates. The mean (μ = 1180℃) and standard deviation (σ = 45℃) of the training set are exported and stored in the configuration file. During online prediction, the real-time data must first be standardized using these parameters, and the prediction results are then de-standardized to the actual temperature values ​​to ensure that the data format is consistent with that of the training phase.

[0132] Comparative experimental results (a certain 2000m) 3 Blast furnace measured data):

[0133] Traditional LSTM 8.2 4.5 0.86 Random Forest 9.1 5.2 0.82 This plan DTN 5.7 3.1 0.93

[0134] In a preferred embodiment of the present invention, step S5 above—deploying the validated model to the real-time control system; adding new data to the training set every 24 hours; fixing the parameters of the bidirectional LSTM and convolutional layers, and fine-tuning the weights of the fully connected layers—may include:

[0135] Step S51: Deploy the validated model to the real-time control system to achieve real-time monitoring and prediction of the blast furnace hearth temperature.

[0136] Step S52: During the operation of the real-time control system, new data is collected every 24 hours and added to the training set to form an updated training dataset.

[0137] Step S53: Based on the updated training dataset, fix the parameters of the bidirectional LSTM layer and the 1D convolutional layer, and only fine-tune the weights of the fully connected layer to complete the incremental learning of the model.

[0138] In this embodiment of the invention, real-time performance is ensured through model deployment. The model's single inference time is less than 50ms, meeting the real-time control requirements of the blast furnace. It detects temperature anomalies 5-10 minutes earlier than manual analysis, and the predicted results are displayed synchronously with the actual temperature, assisting engineers in quickly judging operating conditions and reducing the risk of misoperation. In terms of dynamic adaptability, the latest production data is incorporated daily, which can capture the slow evolution of operating conditions such as blast furnace lining erosion and seasonal changes, reducing prediction deviation by 10%-15% compared to a fixed model. Abnormal data is isolated by marking data segments to be reviewed, avoiding model "aging" and ensuring adaptability to current raw material composition and operating habits. In terms of computational efficiency and model stability, only the fully connected layer is fine-tuned, saving 90% of computational resources compared to retraining the entire model. The single incremental learning time is less than 30 minutes. At the same time, the bottom feature extraction layer is fixed to avoid the shift of basic features due to new data noise, ensuring the consistency of prediction logic.

[0139] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0140] In step S51 above, the verified deep temporal network model (such as the weight file dtn_model_v1.0.bin) is converted into a format compatible with industrial control systems (such as TensorFlowLite or ONNX), and the model size is compressed to adapt to the computing capabilities of PLC or DCS hardware.

[0141] Establish a real-time data channel between the model and the blast furnace sensors to obtain standardized temperature and operating parameters (such as air volume and air temperature data at a frequency of 100ms) from the distributed control system (DCS) in real time; the data preprocessing unit automatically performs outlier removal, moving average filtering (10-second window) and Z-score standardization on the real-time data (using μ and σ saved during the training phase).

[0142] Real-time data is packaged into 14-minute time windows (840 time points), input into the model to perform forward propagation, and generate the predicted hearth temperature for the next 5 minutes. The prediction results are transmitted to the blast furnace operation interface in real time via the OPCUA protocol, displaying the current temperature trend and early warning information (such as "the temperature may exceed 1150℃ in the next 3 minutes").

[0143] In step S52 above, a data acquisition task is triggered at 0:00 every day to extract complete production data for the previous 24 hours from the DCS database, including: 5 points of furnace hearth temperature (sampling at 100ms), air volume or air temperature (sampling at 100ms), coal injection volume or cooling water flow rate (minute-level average); blast furnace operating status markers (such as normal production, furnace start-up and shutdown, fault alarm).

[0144] The system automatically detects the completeness (missing value ratio less than 1%) and rationality (e.g., temperature greater than 800℃ and less than 1300℃) of newly added data, and marks abnormal data segments as "to be reviewed"; minute-level parameters (e.g., coal injection rate) are matched to a 100ms time series through linear interpolation to ensure synchronization with temperature data.

[0145] Newly added data that passes verification is appended to the original training set in chronological order. Data from the most recent 30 days (approximately 1.8 million samples) is retained, while older data older than 30 days is automatically deleted to maintain a stable dataset size.

[0146] Step S53 above: Incremental learning and model fine-tuning: Lock the parameters (such as weights and biases) of the bidirectional LSTM layer and the 1D convolutional layer, as they are responsible for extracting the spatiotemporal basic features of temperature (such as long-term trends and local mutation patterns), and do not need to be retrained when there are no structural changes in the working conditions; only the fully connected layer of the output layer is trained, with the updated training set feature vector as the input and the temperature prediction value for the next 5 minutes as the output; use the AdamW optimizer (learning rate 1e-5), the Huber loss function, set the training epochs to 5, and the batch size to 128, to quickly fit the subtle regular changes in the new data.

[0147] After fine-tuning, use validation set data from the past 3 days to evaluate performance (e.g., whether RMSE < 6.0℃); if the metrics meet the criteria, automatically switch to the new model; if not, skip this update and trigger a data review process (e.g., check whether the newly added data contains abnormal operating conditions).

[0148] In a preferred embodiment of the present invention, step S6 above, which involves inputting real-time standardized data into the deployment model and outputting a temperature prediction value for the next 5 minutes, and generating a minimum disturbance strategy when the prediction value exceeds a safety threshold and sending it to the execution system, and feeding back the executed temperature and operating parameters to step S1, may include:

[0149] Step S61: Input the real-time standardized data from step S1 into the deployment model and output the temperature prediction value for the next 5 minutes.

[0150] Step S62: When the predicted value exceeds the safety threshold: Analyze the influence weights of wind temperature, pulverized coal injection quantity, and cooling water flow rate; generate a minimum disturbance strategy that reduces wind temperature, reduces pulverized coal injection quantity, and increases cooling water flow rate and send it to the execution system to obtain the temperature and operating parameters after execution.

[0151] Step S63: Feedback the temperature and operating parameters after execution to step S1 to form a closed-loop control cycle.

[0152] In this embodiment of the invention, the model detects temperature change trends 5 minutes in advance, identifying risks such as lining overheating earlier than traditional threshold alarms, thus gaining valuable time for operational adjustments. Furthermore, through real-time data preprocessing and lightweight model deployment, the prediction error is controlled within ±5.7℃, ensuring accurate temperature prediction. Based on parameter influence weights, targeted strategies are generated to precisely adjust for the main causes of temperature anomalies such as excessively high air temperature and excessive pulverized coal injection. The magnitude of each adjustment is controlled within the process safety range, reducing production disturbances, maintaining stable permeability and combustion efficiency, and achieving rapid response from threshold triggering to strategy. Through a closed-loop "prediction-adjustment-feedback" mechanism, the system can dynamically optimize strategies based on actual results, forming a self-evolving control loop. Over long-term operation, temperature control errors are reduced by 20%, and each adjustment data is included in the training set, allowing the model to learn the optimal adjustment mode under different operating conditions, continuously improving control accuracy. Simultaneously, it quickly responds to and tracks the effects of temperature anomaly handling, avoiding the risk of single adjustment failures.

[0153] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0154] In step S61 above, parameters such as hearth temperature (5 key locations), air volume, air temperature, and pulverized coal injection rate are collected in real time from blast furnace sensors at a frequency of 100ms. Outlier removal (e.g., through cross-validation of data from adjacent sensors) and moving average filtering (smoothing fluctuations in a 10-second window) are performed on the collected data. Then, the mean (μ) and standard deviation (σ) saved during the training phase are used for Z-score standardization to convert it into an input format that the model can recognize.

[0155] The standardized real-time data is packaged into 14-minute time windows (containing 840 time points) and input into the model deployed in the real-time control system. The model extracts time-dependent features (such as the temperature rise trend in the past 10 minutes) through a bidirectional LSTM layer and identifies spatial anomalies (such as sudden temperature changes at the center of the furnace bottom) through a 1D convolutional layer. After being weighted by an attention fusion layer, the model outputs the predicted average furnace hearth temperature for the next 5 minutes (such as a predicted 1145℃).

[0156] In step S62 above, the preset safety threshold is 1150℃ (set according to the high temperature resistance limit of the blast furnace lining material). If the predicted value is greater than or equal to 1150℃, the strategy generation process is triggered; the attention weight matrix of the model is called to obtain the degree of influence of each operating parameter on the temperature under the current working condition (such as wind temperature weight 35%, pulverized coal injection weight 28%, cooling water flow weight 25%); parameters with high weights are adjusted first to achieve the maximum cooling effect with minimal operational changes.

[0157] When generating the minimum disturbance strategy, targeted adjustments are made based on the weights of each operating parameter's influence on temperature: the reduction in air temperature is determined according to the weight ratio (e.g., a weight ratio of 35% corresponds to a reduction of 3.5℃ in air temperature); the amount of pulverized coal injection is reduced according to the degree of influence (e.g., a weight ratio of 28% corresponds to a reduction of 1.5t / h); and the cooling water flow rate is increased in conjunction with the heat exchange capacity requirements (an increase of 5m³ / h is made when the weight ratio is 25%). 3 / h); at the same time, the strategy must strictly meet the process constraints, such as the amount of pulverized coal injection not being lower than the minimum safety value and the cooling water flow not exceeding the pipeline transportation limit, so as to ensure the safety and stability of blast furnace operation while achieving effective temperature control.

[0158] The DCS system sends adjustment commands to the blower, pulverized coal pump, and cooling water valve, with an execution time of less than 10 seconds; it records the parameters before adjustment (e.g., air temperature 1200℃, pulverized coal injection rate 15t / h), the parameters after adjustment (air temperature 1196.5℃, pulverized coal injection rate 13.5t / h), and the execution timestamp.

[0159] After the adjustment in step S63 above is executed, the furnace hearth temperature change is continuously monitored, and the actual temperature value (e.g., the temperature drops to 1140℃ after 5 minutes after adjustment) and operating parameters (e.g., air temperature, coal injection rate, etc.) are recorded every minute. The temperature and parameter data after execution are transmitted to the real-time data acquisition module in step S1 as new training samples to participate in the next round of data preprocessing. The system automatically marks this data segment as "sample after strategy execution" for analysis of the adjustment effect during subsequent model training.

[0160] Calculate the temperature change before and after adjustment (e.g., ΔT = predicted value - actual value = 1145℃ - 1140℃ = 5℃) and evaluate the effectiveness of the strategy; if the temperature does not drop to a safe range (e.g., still greater than 1150℃), automatically trigger the strategy iteration mechanism (e.g., increase the adjustment range of cooling water flow).

[0161] like Figure 4 As shown, the blast furnace hearth temperature sensor arrangement includes a hearth thermocouple installed on the inner side of the hearth sidewall 1 at a depth of 20cm; several hearth thermocouples 3 are installed in a ring inside the hearth sidewall 1 at a depth of 200cm; a hearth center thermocouple 5 is installed at the center of the hearth sidewall 1 at a depth of 460cm; a hearth secondary center thermocouple 4 is installed outside the hearth center thermocouple 5 at a depth of 400cm; and the hearth thermocouples 3 are located around the outer periphery of the hearth secondary center thermocouples 4.

[0162] like Figure 2 As shown, embodiments of the present invention also provide a safe intelligent adjustment and control system for when the blast furnace hearth temperature is too high, comprising:

[0163] The data acquisition module is used to collect temperature data at various locations in the hearth in real time through a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth, and to simultaneously collect blast furnace operating parameters; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset.

[0164] The building module is used to construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer.

[0165] The training module is used to divide the dataset into time sequences and train the deep temporal network model to obtain the trained deep temporal network model.

[0166] The validation module is used to input the trained deep temporal network model into the validation set and the test set; calculate the mean square error, root mean square error and mean absolute error to verify the model's prediction accuracy; and output the validated model that meets the error threshold.

[0167] The deployment module is used to deploy validated models to the deployment control system; it adds new data to the training set every 24 hours; it fixes the parameters of the bidirectional LSTM and convolutional layers, and fine-tunes the weights of the fully connected layers.

[0168] The feedback module is used to input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold, a minimum disturbance strategy is generated and sent to the execution system.

[0169] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0170] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A safe and intelligent adjustment and control method for when the blast furnace hearth temperature is too high, characterized in that, The method includes: Step S1: Real-time temperature data at various locations in the furnace hearth is collected by a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth, and blast furnace operating parameters are collected simultaneously; the collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset. Step S2: Construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer. Step S3: Based on the deep temporal network model, divide the dataset in time order and train it to obtain the trained deep temporal network model; Step S4: Input the trained deep temporal network model into the validation set and the test set; calculate the mean square error, root mean square error and mean absolute error to verify the model's prediction accuracy; output the validated model that meets the error threshold. Step S5: Deploy the validated model to the real-time control system; add new data to the training set every 24 hours; fix the parameters of the bidirectional LSTM and convolutional layers, and fine-tune the weights of the fully connected layers; Step S6: Input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes; when the prediction value exceeds the safety threshold: analyze the influence weights of wind temperature, pulverized coal injection, and cooling water flow; generate a minimum disturbance strategy that reduces wind temperature, reduces pulverized coal injection, and increases cooling water flow and send it to the execution system to obtain the temperature and operating parameters after execution; feed back the temperature and operating parameters after execution to step S1 to form a closed-loop control cycle.

2. The intelligent adjustment and control method for safe operation when the blast furnace hearth temperature is too high, as described in claim 1, is characterized in that... Step S1: Real-time temperature data at various locations in the hearth is collected by a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth, and blast furnace operating parameters are collected simultaneously. The collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset, including: A high-precision temperature sensor array deployed on the inner wall and bottom of the hearth is used to collect temperature data at various locations in the hearth in real time; blast furnace operating parameters such as air volume, air temperature, pulverized coal injection rate, cooling water flow rate, and burden descent rate are collected simultaneously. Outliers in temperature data and operating parameters are removed; time-series fluctuations are smoothed using a moving average method; data is standardized using Zscore to generate a standardized time-series dataset that matches the model input dimension, and the standardized parameters are recorded.

3. The intelligent adjustment and control method for blast furnace hearth temperature when it is too high according to claim 2, characterized in that, Step S2: Construct a deep temporal regression model using a standardized time-series dataset. The model includes: a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer, comprising: A deep temporal network regression model is constructed using a standardized temporal dataset. The model includes: a bidirectional LSTM layer to extract time-dependent features, a 1D convolutional layer to extract spatial mutation features, and an attention fusion layer to weight key features.

4. The intelligent adjustment and control method for blast furnace hearth temperature when it is too high according to claim 3, characterized in that, Step S3: Based on the deep temporal network model, divide the dataset according to time order and train it to obtain the trained deep temporal network model, including: Based on a deep temporal network model, the dataset is divided into training set, validation set, and test set according to time order. The model is trained using the training set, and the parameters are optimized with the goal of minimizing the prediction error to obtain the initial trained model. The performance of the initial trained model was evaluated through cross-validation, and the hyperparameters were tuned using Bayesian optimization based on the evaluation results to obtain the tuned model. The optimized model is validated using a validation set, and the parameters are fine-tuned based on the validation results to output the trained deep temporal network model.

5. The intelligent adjustment and control method for blast furnace hearth temperature when it is too high according to claim 4, characterized in that, Step S4: Input the trained deep temporal network model into the validation set and the test set; Calculate the mean square error, root mean square error, and mean absolute error to verify the model's prediction accuracy; The output of the validated model that meets the error threshold includes: Input the trained deep temporal network model into the validation set and the test set to obtain the model's output results on different datasets; The evaluation index is calculated based on the output results. The evaluation index includes mean square error, root mean square error and mean absolute error. The model's prediction accuracy is verified by evaluating the indicators. When the evaluation indicators meet the preset error threshold, the verified model is output.

6. The intelligent adjustment and control method for safe operation when the blast furnace hearth temperature is too high, as described in claim 5, is characterized in that... Step S5: Deploy the validated model to the real-time control system; add new data to the training set every 24 hours; fix the parameters of the bidirectional LSTM and convolutional layers, and fine-tune the weights of the fully connected layers, including: The validated model is deployed to the real-time control system to achieve real-time monitoring and prediction of the blast furnace hearth temperature; During the operation of the real-time control system, new data is collected every 24 hours and added to the training set to form an updated training dataset. Based on the updated training dataset, the parameters of the bidirectional LSTM layer and the 1D convolutional layer are fixed, and only the weights of the fully connected layer are fine-tuned to complete the incremental learning of the model.

7. A safe and intelligent adjustment and control system for when the blast furnace hearth temperature is too high, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to collect temperature data at various locations in the hearth in real time through a high-precision temperature sensor array deployed on the inner wall and bottom of the hearth, and to simultaneously collect blast furnace operating parameters. The collected temperature data and operating parameters are preprocessed to generate a standardized time-series dataset. The building module is used to construct a deep temporal network regression model using a standardized temporal dataset. The model includes a bidirectional LSTM layer, a 1D convolutional layer, and an attention fusion layer. The training module is used to divide the dataset into time sequences and train the deep temporal network model to obtain the trained deep temporal network model. The validation module is used to input the trained deep temporal network model into the validation set and the test set; Calculate the mean squared error, root mean square error, and mean absolute error to verify the model's prediction accuracy; output a validated model that meets the error threshold. The deployment module is used to deploy validated models to the deployment control system; it adds new data to the training set every 24 hours; it fixes the parameters of the bidirectional LSTM and convolutional layers, and fine-tunes the weights of the fully connected layers. The feedback module is used to input real-time standardized data into the deployment model and output the temperature prediction value for the next 5 minutes. When the predicted value exceeds the safety threshold, a minimum disturbance strategy is generated and sent to the execution system.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images