Safe intelligent adjustment control method and system used when temperature of blast furnace hearth is too high

Abstract

The invention provides a safe and intelligent adjustment control method and system when the temperature of a blast furnace hearth is too high, and relates to the technical field of automation and intelligent control of blast furnaces in the metallurgical industry. The method comprises the steps that temperature data of all positions of the hearth are collected in real time through high-precision temperature sensor arrays arranged on the inner wall and the bottom of the hearth, and the temperature data of all the positions of the hearth are obtained; synchronously acquiring blast furnace operation parameters; the collected temperature data and operation parameters are preprocessed, and a standardized time sequence data set is generated; a deep sequential network regression model is constructed through the standardized sequential data set, and the model comprises a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer; and based on the deep sequential network model, dividing a data set according to a time sequence and training the data set to obtain a trained deep sequential network model. According to the invention, real-time analysis is carried out through a machine learning algorithm, and an adjustment strategy is automatically calculated, so that rapid reduction and accurate control of the hearth temperature are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automation and intelligent control of blast furnaces in the metallurgical industry, in particular to a safe and intelligent adjustment control method and system for high furnace hearth temperature. BACKGROUND

[0002] Blast furnace smelting is the core link of the steel industry, and the stable control of the furnace hearth temperature may directly affect the smelting efficiency, equipment life and safety production; as the key area of chemical reaction and heat transfer of the blast furnace, the furnace hearth temperature that is too high may cause serious problems such as intensified lining erosion, furnace bottom nodulation, and even production accidents. However, the blast furnace smelting process has the characteristics of multivariable coupling, nonlinearity and time variation, and the furnace hearth temperature is dynamically affected by dozens of parameters such as air volume, air temperature, coal injection amount and cooling water flow, so the traditional control method may face serious challenges.

[0003] At present, the treatment of high furnace hearth temperature in industrial field depends on the experience judgment and manual adjustment of operators, and the traditional temperature monitoring adopts single-point or sparse sensor arrangement, which may be difficult to capture the radial or axial temperature gradient distribution and local overheating mutation of the furnace hearth, and may cause late discovery of abnormal working conditions; the temperature prediction method based on empirical formula or simple statistical model may not be able to effectively model the complex space-time characteristics. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a safe and intelligent adjustment control method and system for high furnace hearth temperature, which automatically calculates the adjustment strategy through real-time analysis by machine learning algorithm, so as to realize rapid reduction and accurate control of the furnace hearth temperature.

[0005] To solve the above technical problems, the technical scheme of the present application is as follows:

[0006] In a first aspect, a safe and intelligent adjustment control method for high furnace hearth temperature is provided, which comprises:

[0007] Step S1: Real-time collection of temperature data at each position of the furnace hearth and synchronous collection of blast furnace operation parameters through a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace hearth; pre-processing of the collected temperature data and operation parameters to generate a standardized time series data set;

[0008] Step S2: Construction of a deep time series network regression model through the standardized time series data set, wherein the model comprises a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer;

[0009] Step S3: Based on the deep time series network model, the data set is divided in time sequence and trained to obtain a trained deep time series network model;

[0010] Step S4: input the trained deep time series 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 prediction accuracy of the model; output the validation passed 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 bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights;

[0012] Step S6: input real-time standardized data into the deployed model to output the predicted temperature value in the next 5 minutes; when the predicted value exceeds the safety threshold, generate a minimum disturbance strategy and issue it to the execution system; feedback the executed temperature and operation parameters to step S1.

[0013] Further, step S1: real-time collection of temperature data at each position in the furnace through a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace; synchronous collection of blast furnace operation parameters; preprocessing of collected temperature data and operation parameters to generate standardized time series data sets, including:

[0014] Real-time collection of temperature data at each position in the furnace through a high-precision temperature sensor array deployed on the inner wall and bottom of the furnace; synchronous collection of blast furnace operation parameters: air volume, air temperature, coal injection amount, cooling water flow and furnace burden descending speed;

[0015] Remove outliers in temperature data and operation parameters; use the moving average method to smooth the time series fluctuations; convert the data through Zscore standardization to generate a standardized time series data set matching the model input dimension, and record the standardization parameters.

[0016] Further, step S2: construct a deep time series network regression model through the standardized time series data set, which includes a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer, including:

[0017] Construct a deep time series network regression model through the standardized time series data set; the model includes a bidirectional LSTM layer to extract time-dependent features, a 1D convolution layer to extract spatial mutation features, and an attention fusion layer to weight key features.

[0018] Further, step S3: based on the deep time series network model, divide the data set in time sequence and train to obtain a trained deep time series network model, including:

[0019] Based on the deep time series network model, divide the data set into training set, validation set and test set in time sequence;

[0020] Use the training set to train the model to optimize the parameters with the goal of minimizing the prediction error to obtain an initial training model;

[0021] The initial training model is evaluated in performance by cross-validation, and the hyperparameters are optimized based on the evaluation results by Bayesian optimization to obtain an optimized model;

[0022] The optimized model is verified using the verification set, and the parameters are fine-tuned according to the verification results, and the trained deep time series network model is output.

[0023] Further, step S4: input the trained deep time series network model into the verification set and the test set; calculate the mean square error, root mean square error and mean absolute error, and verify the prediction accuracy of the model; output the verification passed model that meets the error threshold, including:

[0024] The trained deep time series network model is input into the verification set and the test set to obtain the output results of the model on different data sets;

[0025] Based on the output results, evaluation indicators are calculated, including mean square error, root mean square error and mean absolute error;

[0026] The prediction accuracy of the model is verified by the evaluation indicators, and when the evaluation indicators meet the preset error threshold, the verification passed model is output.

[0027] Further, step S5: deploy the verification passed model to the real-time control system; add new data to the training set every 24 hours; fix the bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights, including:

[0028] The verification passed model is deployed to the real-time control system to realize 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 data set;

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

[0031] Further, step S6: input the real-time standardized data into the deployed model to output the future 5-minute temperature prediction value; when the prediction value exceeds the safety threshold, generate the minimum disturbance strategy and issue it to the execution system; the executed temperature and operation parameters are fed back to step S1, including:

[0032] The real-time standardized data of step S1 is input into the deployed model to output the future 5-minute temperature prediction value;

[0033] When the predicted value exceeds the safety threshold: analyze the influence weight of the wind temperature, coal injection amount, and cooling water flow; generate a minimum disturbance strategy of reducing the wind temperature, reducing the coal injection amount, and increasing the cooling water flow and issue it to the execution system to obtain the executed temperature and operating parameters;

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

[0035] In a second aspect, a safety intelligent adjustment control system for a high-temperature furnace hearth includes:

[0036] A collection module is configured to collect real-time temperature data of each position of the furnace hearth and synchronous high-temperature furnace operating parameters through a high-precision temperature sensor array arranged on the inner wall and bottom of the furnace hearth, and to preprocess the collected temperature data and operating parameters to generate a standardized time series data set;

[0037] A construction module is configured to construct a deep time series network regression model through the standardized time series data set, wherein the model includes a bidirectional LSTM layer, a 1D convolution layer, and an attention fusion layer;

[0038] A training module is configured to divide the data set in time sequence and train based on the deep time series network model to obtain a trained deep time series network model;

[0039] A verification module is configured to input the trained deep time series network model into a verification set and a test set, calculate mean square error, root mean square error, and mean absolute error, verify the prediction accuracy of the model, and output a verification-passing model that meets the error threshold;

[0040] A deployment module is configured to deploy the verification-passing model to a deployment control system, add new data to the training set every 24 hours, fix the bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights;

[0041] A feedback module is configured to input real-time standardized data into the deployed model to output a 5-minute temperature prediction value, and when the prediction value exceeds the safety threshold, generate a minimum disturbance strategy and issue it to an execution system.

[0042] In a third aspect, a computing device includes:

[0043] One or more processors;

[0044] A storage device is configured to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0045] In a fourth aspect, a computer-readable storage medium stores a program that, when executed by a processor, implements the method.

[0046] The above scheme of the present application at least includes the following beneficial effects:

[0047] The data acquisition and preprocessing link realizes full-dimensional data coverage through a high-precision sensor array, combines outlier rejection, moving average filtering and Z-score standardization, and lays a high-precision prediction foundation for the model; the model architecture innovation adopts a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer, realizes deep fusion of space-time features, improves prediction accuracy and interpretability, and improves the prediction accuracy of the traditional model by 30%; in the training and optimization process, the data set is divided, and technologies such as Bayesian optimization are used, taking into account efficiency and generalization ability; industrial-level deployment combined with a dynamic updating mechanism ensures real-time performance and compatibility; closed-loop control and strategy optimization realize early warning and precise intervention, improving safety and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a flowchart of a safety intelligent adjustment control method for a blast furnace hearth when the temperature of the blast furnace hearth is too high, provided by an embodiment of the present application.

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

[0050] Figure 3 is a specific implementation flowchart of the present application.

[0051] Figure 4 is a blast furnace hearth temperature sensor arrangement diagram of the present application. DETAILED DESCRIPTION

[0052] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0053] As Figure 1 shown, an embodiment of the present application proposes a safety intelligent adjustment control method for a blast furnace hearth when the temperature of the blast furnace hearth is too high, the method comprising the following steps:

[0054] Step S1: Real-time acquisition of temperature data at each position of the hearth by deploying a high-precision temperature sensor array on the inner wall and bottom of the hearth, and synchronous acquisition of blast furnace operating parameters; preprocessing of the acquired temperature data and operating parameters to generate standardized time series data sets;

[0055] Step S2: build a deep time series network regression model by standardizing the time series dataset, which includes a bidirectional LSTM layer, a 1D convolution layer, and an attention fusion layer;

[0056] Step S3: based on the deep time series network model, divide the dataset in time sequence and train to obtain a trained deep time series network model;

[0057] Step S4: input the trained deep time series network model into the validation set and the test set; calculate the mean square error, the root mean square error, and the mean absolute error to verify the prediction accuracy of the model; and output a verification passed model that meets the error threshold;

[0058] Step S5: deploy the verification passed model to a real-time control system; add new data to the training set every 24 hours; fix the bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights;

[0059] Step S6: input real-time standardized data into the deployed model to output a 5-minute temperature prediction value; when the prediction value exceeds the safety threshold, generate a minimum disturbance strategy and issue it to the execution system; and feed back the executed temperature and operation parameters to Step S1.

[0060] In the embodiment of the present application, high-precision sensor arrays are used to collect furnace temperature and operation parameters, and standardized data sets are generated after preprocessing to provide a reliable data basis for model training; a deep network containing bidirectional LSTM, 1D convolution, and attention fusion layer is constructed to effectively extract the spatiotemporal features of temperature data and improve the prediction accuracy; time series division dataset, cross-validation, and Bayesian optimization strategies are used to ensure the stability and reliability of the model; incremental learning by regularly incorporating new data, fixing the bottom layer parameters, and fine-tuning the fully connected layer reduces the training cost and improves the adaptability of the model; real-time prediction of future 5-minute temperature generates a minimum disturbance strategy automatically when the threshold is exceeded and feeds back to the closed loop to balance safety and production efficiency.

[0061] In a preferred embodiment of the present application, the above-mentioned step S1: real-time collection of temperature data at each position in the furnace by deploying high-precision temperature sensor arrays on the inner wall and bottom of the furnace, and synchronous collection of blast furnace operation parameters; preprocessing of the collected temperature data and operation parameters to generate a standardized time series dataset, which can include:

[0062] Step S11, real-time collection of temperature data at each position in the furnace by deploying high-precision temperature sensor arrays on the inner wall and bottom of the furnace; synchronous collection of blast furnace operation parameters: air volume, air temperature, coal injection amount, cooling water flow rate, and furnace burden descent speed;

[0063] Step S12, remove outliers in temperature data and operating parameters; use sliding average method to smooth the time series fluctuations; convert the data by Zscore standardization to generate standardized time series data set matching the input dimension of the model, and record the standardization parameters.

[0064] In the embodiments of the present application, the three-dimensional multi-dimensional monitoring of the hearth temperature field is realized by a spatially distributed sensor array, key operating parameters are synchronously collected, and complete input variables are provided for subsequent establishment of a temperature prediction model; in data processing, outliers caused by sensor failure or electromagnetic interference can be effectively removed to realize data purification, sliding average filtering is used to retain the temperature change trend while suppressing noise, reduce the influence of random disturbances in the production process, eliminate the dimensional differences of different variables by Zscore standardization, 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 realize format adaptation, and save the standardization parameters to ensure the consistency and traceability of data processing during online prediction.

[0065] In the embodiments of the present application, when specifically applied, the following technical solutions can be used to realize it, for example:

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

[0067] The control system synchronously acquires the following operating parameters: air volume, air temperature, and cooling water flow are collected in real time (frequency 100 milliseconds) by the sensors provided with the equipment; the coal injection amount and the burden descent speed are output in the form of average values per minute due to the characteristics of the equipment (for example, the coal injection amount is divided into each minute according to the hourly cumulative value); for minute-level parameters, the system converts them into the same 100-millisecond time interval as the temperature data through a time alignment algorithm (such as linear interpolation), ensuring that all data correspond one by one on the time axis.

[0068] The above step S12, the system performs anomaly detection on each temperature point and operating parameter: automatically identifies data that is obviously outside the reasonable range (such as temperature mutation to negative value or far beyond the normal temperature range of the blast furnace); for abnormal data points, the average value of the previous and subsequent valid values is used for repair (such as the air volume data jumps at a certain time, the average value of the air volume of the previous 1 second and the subsequent 1 second is used instead).

[0069] The moving average filter is used to smooth the data fluctuations and retain the trend. A fixed window smoothing process is applied to the time series of each parameter: take the data within the current time and the previous and next 10 seconds (i.e. 100 milliseconds x 10 time points), calculate the average value as the filtered value at the current time; this operation can effectively suppress high-frequency noise caused by device vibration, electromagnetic interference, etc., making the temperature and parameter curves smoother.

[0070] Data standardization is used to eliminate the dimensional differences of different parameters (such as temperature in ℃, flow in m 3 / h), and to scale all data proportionally: calculate the historical average value and fluctuation range of each parameter (based on normal production data in the past 7 days); rescale each real-time data according to the "relative position to the average value" to make all parameters uniform in a similar numerical range (e.g. between 1 and 1); save the average value and scaling ratio of each parameter for subsequent online prediction, ensuring the consistency of data processing.

[0071] The processed temperature and operating parameters are combined into time series samples suitable for model analysis: divide the data window by 10 minutes (e.g. from 0 seconds to 600 seconds), each window contains 600 time points of multi-dimensional data (5 points of temperature + 5 items of operating parameters, a total of 10 dimensions); slide the window by 1 minute (e.g. the next window from 60 seconds to 660 seconds), generate continuous overlapping samples, ensure that the model can capture short-term and long-term trends of temperature changes; the final data set format is "number of samples x time points x parameter dimensions", which directly meets the input requirements of subsequent intelligent models.

[0072] In a preferred embodiment of the present application, the step S2 of constructing a deep time series network regression model by standardizing the time series data set comprises:

[0073] A deep time series network regression model is constructed by standardizing the time series data set; the model comprises: a bidirectional LSTM layer for extracting time-dependent features, a 1D convolutional layer for extracting spatial mutation features, and an attention fusion layer for weighting key features.

[0074] In the embodiment of the present application, the bidirectional LSTM and 1D convolutional layer are fused to capture temperature spatio-temporal features, and the attention fusion layer is combined to dynamically allocate parameter weights, thereby improving the prediction accuracy of the model; the 1D convolutional layer responds to local temperature anomalies at the second level, combined with multi-parameter nonlinear correlation analysis, to provide 35-minute early warning of abnormal operating conditions with a 40% reduction in false alarm rate; with the help of attention weight visualization, quantitative influence analysis of operating parameters can be realized, the response trend of temperature to operation adjustment in the next 5 minutes can be predicted, and precise adjustment strategies can be formulated.

[0075] In the embodiments of the present application, when specifically applied, the following technical solutions can be used to achieve, for example:

[0076] The following features are extracted from the pretreated normalized time series data:

[0077] Spatial temperature features: real-time temperatures of 5 key positions of the hearth (inner wall 20 cm, 200 cm, hearth bottom 400 cm, 460 cm, and center), temperature difference between adjacent positions (temperature gradient), maximum temperature difference to average temperature ratio (reflecting temperature distribution uniformity);

[0078] Operating parameters: real-time values and historical trends (such as 1-hour moving average) of air volume, air temperature, coal injection amount, cooling water flow, and charge descending speed;

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

[0080] Based on metallurgical process experience, key parameters affecting hearth temperature (such as coal injection amount directly affecting combustion heat release, cooling water flow affecting heat exchange efficiency) are retained, and redundant features with very low correlation (such as vibration signals of non-critical equipment) are excluded.

[0081] The contribution of each feature to the prediction target is automatically calculated by a machine learning algorithm, and features with low importance are gradually removed: about 30 initial input features (including original features and derived features); feature weights are fed back through model training, and the 5 features with the lowest weights are removed each time until 15 core features (such as temperature gradient, coal injection amount historical mean, cooling water flow change rate, etc.) are retained, ensuring the balance between input dimension and model efficiency.

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

[0083] Input layer: normalized 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-related + 10-dimensional operating parameters and derived features), forming an input tensor [sample number x time step x feature dimension] (such as N x 600 x 15); uniform data format allows multi-dimensional parameters at different times to be recognized by subsequent network layers, matching the input requirements of deep time series networks.

[0084] The bidirectional LSTM layer extracts time-dependent features. A bidirectional long short-term memory network (BiLSTM) with 128 memory cells is used, including an input gate, a forgetting gate, and an output gate three-layer structure. At the same time, the time series is processed from the forward direction (past to present) and the reverse direction (present to past), to extract the long-term evolution trend (such as the temperature rising slope for several consecutive hours) and the short-term fluctuation characteristics (such as the temperature response within 5 minutes after adjusting the coal injection amount) of the furnace hearth temperature. Through the gating mechanism, irrelevant historical information (such as stable air volume data 3 hours ago) is automatically forgotten, and key states (such as temperature mutations caused by abnormal cooling water flow 1 hour ago) are retained. A feature vector [sample number x time step x 128 dimension] containing time context information is generated, representing the dynamic dependence of each time point in the sequence.

[0085] The 1D convolution layer extracts spatial mutation features. A 1D convolution layer with a kernel size of 3 and 64 channels is used to locally scan the multi-dimensional features at the same time point. By sliding the convolution kernel (such as covering 3 adjacent temperature sensor data), temperature abnormal patterns at different positions of the furnace hearth (such as a single-point temperature suddenly higher than the surrounding 15°C, which may indicate lining erosion) are captured. The original temperature distribution is converted into a "spatial anomaly feature vector", such as identifying the feature pattern of the radial temperature gradient of the furnace hearth exceeding the safety threshold, and outputting a spatial feature tensor [sample number x time step x 64 dimension].

[0086] The attention fusion layer weights the key features. The attention mechanism is introduced to automatically learn the importance weights of the time sequence features (BiLSTM output) and spatial features (1DCNN output), avoiding the averaging processing of key features in traditional concatenation methods. The 128-dimensional time sequence features output by BiLSTM and the 64-dimensional spatial features output by 1DCNN are unified to 100 dimensions through a fully connected layer. For each time point's feature vector, the "contribution degree" to the final temperature prediction is calculated (such as the current time's wind temperature change weight may be higher than the furnace material descending speed 30 minutes ago). High-contribution-degree features (such as temperature gradient, real-time coal injection amount) are given higher weights, and low-relevance features (such as the small fluctuation of cooling water inlet temperature under stable working conditions) are suppressed. The integrated feature vector [sample number x time step x 100 dimension] after fusion is output.

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

[0088] In a preferred embodiment of the present application, the step S3 of dividing the data set in time sequence based on the deep time sequence network model and training to obtain the trained deep time sequence network model can comprise:

[0089] In step S31, the data set is divided into a training set, a validation set and a test set in time sequence based on the deep time sequence network model.

[0090] In step S32, the model is trained using the training set to optimize the parameters for minimizing the prediction error, and an initial training model is obtained.

[0091] In step S33, the performance of the initial training model is evaluated by cross-validation, and the hyperparameters are optimized based on the evaluation results by Bayesian optimization to obtain an optimized model.

[0092] In step S34, the optimized model is verified using the validation set, and the parameters are fine-tuned according to the verification results, and the trained deep time sequence network model is output.

[0093] In the embodiment of the present application, the data set is divided in time sequence and the abnormal working condition data is labeled to prevent future data mixing and abnormal value interference, and to ensure that the model learns the real and stable production rule; the initial model is obtained by optimizing the parameters of the training set, which effectively minimizes the prediction error and lays a foundation for the performance of the model; the model is evaluated by cross-validation and the hyperparameters are optimized by Bayesian optimization, which can efficiently find the best parameter combination, balance the model complexity and prediction accuracy while reducing the calculation cost, and avoid over-complexity or under-fitting; the parameters are fine-tuned using the validation set, which can further improve the adaptability of the model in actual working conditions, suppress the over-fitting problem, and ensure that the model can stably output high-precision prediction results in different production stages.

[0094] In the embodiment of the present application, when applied specifically, the following technical solutions can be used to achieve the above-mentioned technical effects, for example:

[0095] In step S31, the standardized time sequence data is divided in time sequence as follows:

[0096] Training set (70%): covering the data of stable production stage in history, used for learning basic parameters of the model;

[0097] Validation set (15%): selected from the time period after the training set, used for adjusting hyperparameters and monitoring over-fitting;

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

[0099] The abnormal working condition data of blast furnace start-stop stage and equipment failure are manually labeled, and the interference is reduced during training by the following methods: setting the sampling weight, the sampling probability of normal data is 100%, and the sampling probability of abnormal data is 20%; separately constructing an abnormal data set for subsequent robustness test, which does not participate in regular training.

[0100] The time series data is converted into a sample set by using a fixed length sliding window (such as 10 minutes = 600 time points), and the window sliding step is set to 1 minute (60 time points), to generate overlapping sample enhancement time series feature capture; all temperature sensors and operating parameter data need to be strictly synchronized, and the minute-level parameters (such as coal injection amount) are matched to the 100 ms sampling frequency by linear interpolation.

[0101] In the loss function aspect of the above step S32, Huber loss (set δ as 1.0) is selected to balance the sensitivity of the model to normal error and abnormal value. The characteristic of this loss function is that when the error between the predicted value and the true value is small (the absolute value does not exceed 1.0), the square term form of mean square error (MSE) is used to calculate the loss, so as to reduce the influence of small error on model training and improve the detail fitting precision; when the error is large (the absolute value exceeds 1.0), the linear term form of mean absolute error (MAE) is switched to calculate the loss, so as to avoid excessive interference of a small number of abnormal values on gradient calculation, thereby enhancing the robustness of the model to data noise.

[0102] In the optimizer aspect, AdamW optimizer is used, the initial learning rate is set to 1e4 (i.e. 0.0001), and the weight decay coefficient is set to 1e5 (i.e. 0.00001). This optimizer accelerates the convergence speed of the model while constraining the model parameters through the introduction of L2 regularization mechanism, forcing the network to learn more concise feature representation and effectively suppressing overfitting phenomenon.

[0103] In the basic training process, the training parameters are set as batch size 64, training rounds (Epoch) 50, and early stopping threshold value (if the loss of the validation set does not decrease for 3 consecutive rounds, the training is terminated in advance); the training process first loads the training data by batch, calculates the predicted value by forward propagation, then calculates the Huber loss and updates the model weight by back propagation, simultaneously evaluates the model performance (the indicators include RMSE, MAPE, R2) on the validation set after completing one round of training, and finally saves the model parameters with the best performance on the validation set (for example, the parameters with RMSE = 6.8℃ at the 15th round).

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

[0105] Hyperparameters Search range Initial value Time window length (min) 515 10 LSTM layer number 13 1 Dropout rate 0.20.5 0.3 Batch size 32128 64 Learning rate 5e55e4 1e4

[0106] Bayesian optimization procedure: the objective function is to minimize the RMSE on the validation set, combined with a model complexity penalty term (to prevent overfitting caused by too deep network); initialize the Gaussian process model, build a probabilistic model of hyperparameters and performance based on historical evaluation results; obtain the next recommended combination of hyperparameters (such as recommended time window = 12 minutes, Dropout = 0.4) through expected improvement (EI); train the model with new parameters and evaluate performance, update the Gaussian process model; repeat the iteration 1015 times, select the parameter combination with the lowest validation set RMSE (such as finally determine the time window = 12 minutes, Dropout = 0.35).

[0107] Rolling window cross-validation (RollingWindowCV) is used to divide multiple validation sub-windows on the training set:

[0108] Fold 1: train with the first 60% of the data and validate with the last 40%;

[0109] Fold 2: train with the first 70% of the data and validate with the last 30%;

[0110] Fold 3: train with the first 80% of the data and validate with the last 20%;

[0111] Integrate the results of the three-fold validation to evaluate the stability of the model and ensure that the hyperparameters perform well in different time periods.

[0112] The above step S34, calculate three core indicators on the validation set: RMSE is to measure the absolute deviation of predicted temperature and actual temperature, target value ≤ 6.0℃; MAPE is to evaluate the relative error, target value ≤ 3.5%; R2 score is to evaluate the model's ability to explain temperature changes, target value ≥ 0.92.

[0113] If the validation set error is significantly higher than the training set (such as RMSE difference greater than 2.0℃), take the following measures: increase the Dropout rate (such as from 0.3→0.4); increase the strength of L2 regularization (such as λ from 0.001→0.002); reduce the number of network layers (such as from 3 layers of LSTM→2 layers).

[0114] Underfitting processing: if the model performs poorly on both the training set and the validation set, adjust the strategy: increase the number of training rounds (such as from 50→80); increase the learning rate (such as from 1e4→2e4); increase the number of LSTM memory units (such as from 128→256).

[0115] In a preferred embodiment of the present application, the above step S4: input the trained deep time series 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 prediction accuracy of the model; output the validation passed model that meets the error threshold, which can include:

[0116] Step S41, input the trained deep time series network model into the verification set and the test set to obtain output results of the model on different data sets;

[0117] Step S42, calculate evaluation indexes based on the output results, the evaluation indexes including mean square error, root mean square error and mean absolute error;

[0118] Step S43, verify the prediction accuracy of the model through the evaluation indexes, and output the model passing the verification when the evaluation indexes meet a preset error threshold.

[0119] In the embodiment of the present application, the model is input into independent verification set and test set to ensure the reliability in actual production; the multi-dimensional indexes such as mean square error, root mean square error and mean absolute error are used for comprehensive evaluation to avoid the limitation of single index; the error threshold bound with production demand is preset, the prediction error of the verified model is reduced by 30% compared with the traditional method, and the safety and economy of temperature control are improved; the model not meeting the threshold is rejected for deployment and the test set is forced to contain complex working condition data, which effectively prevents and controls the risk and ensures the stable operation of the model in all working conditions.

[0120] In the embodiment of the present application, when specifically applied, the following technical solutions can be used to realize it, for example:

[0121] The above step S41, the original data (such as furnace temperature, air volume, coal injection volume, etc.) of the verification set and the test set are standardized by using the mean (μ) and the standard deviation (σ) calculated by the training set; for example, if the mean furnace temperature μ of the training set is 1180℃ and the standard deviation σ is 45℃, then the standardized temperature of a test point 1200℃ is ((1200-1180) / 45≈0.44); it is strictly prohibited to use the verification set / test set data to calculate the standardization parameters to avoid data leakage leading to false high model performance.

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

[0123] The samples are grouped according to the batch size of 64 to adapt to the memory capacity of the industrial GPU, and the time consumption of single inference is ensured to be less than 50ms to meet the real-time verification requirement.

[0124] Read the tuned deep time series network model (e.g., containing the weight parameters of bidirectional LSTM, 1D convolution, and attention layers) from the storage path, initialize the inference session; input the batch data of the validation set and the test set into the model in turn, calculate the predicted value by forward propagation; for example, the test set contains 50,000 samples, which are divided into 782 batches (50000 / 64≈782) to complete inference, and 64 future 5-minute furnace temperature prediction values are generated for each batch.

[0125] Map the prediction results to the actual temperature values one by one according to the timestamps, and store them in an industrial database (e.g., MySQL), including fields such as timestamp, sensor location, actual temperature, predicted temperature, error value, etc., for subsequent index calculation and traceability.

[0126] The above step S42, evaluation index calculation: the root mean square error (RMSE) is a measure of the absolute deviation between the predicted value and the actual value, with a unit of ℃; the calculation formula is the square root of the mean of each sample error square, for example, the RMSE of the 50,000-sample validation set is 5.7 ℃, indicating that the average prediction deviation is within ±5.7 ℃; the mean absolute error (MAE) is the mean of the absolute value of the error, which is more robust to outliers. The validation set MAE = 4.2 ℃, indicating that the average absolute difference between the predicted value and the actual value is 4.2 ℃, which is suitable for evaluating the stability of the model under normal working conditions; the coefficient of determination (R 2 ) is an evaluation of the model's ability to explain temperature changes, with a value range of 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 generalization ability of the model.

[0127] Use industrial monitoring software (e.g., Wonderware) to draw a time series curve, with the horizontal axis representing time (e.g., November 1-30, 2024) and the vertical axis representing temperature (℃), and superimpose the actual temperature curve and the predicted temperature curve. For example, the actual temperature trend in a certain period of time coincides with the predicted curve, with an error of less than 3 ℃, indicating that the model is accurate in trend prediction.

[0128] Group the error values of the validation set and the test set by interval (e.g., ±2 ℃, ±4 ℃, ±6 ℃, etc.), and draw a frequency distribution histogram. The results show that 90% of the errors are within ±6 ℃, which meets the precision requirement of the blast furnace process for temperature control (allowable fluctuation ±8 ℃), and only 5% of the errors exceed ±8 ℃, which needs to be further analyzed whether it is due to sensor failure or extreme working conditions.

[0129] The above step S43, combined with the blast furnace production safety standard and historical operation experience, sets the following acceptance criteria: RMSE≤6.0℃ ensures that the temperature prediction deviation is within the controllable range of the process, avoiding excessive adjustment of the cooling system due to misjudgment; MAPE≤5.0℃ controls the average error within a reasonable range, meeting the adjustment accuracy requirements of coal injection rate, air volume and other parameters; R 2 ≥0.90 is the model that needs to have strong data interpretation ability to prevent prediction failure due to accidental factors.

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

[0131] The trained model weight is stored in the industrial control system server in the form of a binary file (such as dtn_model_v1.0.bin), and the hyperparameter configuration (such as time window 14 minutes, LSTM layer 2 layers) is recorded, which is convenient for subsequent reproduction and update; The mean (μ=1180℃) and standard deviation (σ=45℃) of the training set are exported and stored in the configuration file; During online prediction, real-time data needs to be standardized by this parameter first, and the prediction result is then de-standardized to the actual temperature value, ensuring that the data format is consistent with the training phase.

[0132] Comparative experiment results (measured data of a 2000m 3 high furnace):

[0133] Model RMSE (°C) MAPE (%) [R 2 ]] Traditional LSTM 8.2 4.5 0.86 Random forest 9.1 5.2 0.82 The present scheme DTN 5.7 3.1 0.93

[0134] In a preferred embodiment of the present application, the above step S5: deploying the verified model to the real-time control system; adding new data to the training set every 24 hours; fixing the parameters of bidirectional LSTM and convolution layer, and fine-tuning the weight of fully connected layer, which can include:

[0135] Step S51, deploying the verified model to the real-time control system to realize real-time monitoring and prediction of the blast furnace hearth temperature;

[0136] Step S52, during the operation of the real-time control system, collect new data every 24 hours and add it to the training set to form an updated training data set;

[0137] Step S53, based on the updated training data set, fix the parameters of the bidirectional LSTM layer and the 1D convolution layer, and only fine-tune the weight of the fully connected layer to complete the incremental learning of the model.

[0138] In the embodiment of the present application, real-time guarantee is realized through model deployment, the time consumption of single inference of the model is less than 50 ms, the real-time control demand of the blast furnace is met, the temperature anomaly is found 5-10 minutes in advance compared with manual analysis, and the prediction result is displayed synchronously with the actual temperature to assist engineers in quickly judging the working condition and reducing the risk of misoperation; in terms of dynamic adaptability, the latest production data is included every day, slow working condition evolution such as lining erosion and seasonal change in the blast furnace can be captured, the prediction deviation is reduced by 10%-15% compared with a fixed model, abnormal data is isolated by marking the data segment to be reviewed to avoid model aging, and the adaptability to the current raw material composition and operation habit is ensured; in terms of calculation efficiency and model stability, only the fully connected layer is fine-tuned, 90% of the calculation resources are saved compared with retraining the entire model, the time consumption of single incremental learning is less than 30 minutes, the bottom feature extraction layer is fixed to avoid the deviation of the basic features caused by the noise of the new data, and the prediction logic consistency is ensured.

[0139] In the embodiment of the present application, when specifically applied, the following technical solutions can be used to realize it, for example:

[0140] In the above step S51, the verified deep time sequence network model (such as the weight file dtn_model_v1.0.bin) is converted into an industrial control system compatible format (such as TensorFlowLite or ONNX), and the model volume is compressed to adapt to the PLC or DCS hardware calculation capability.

[0141] A real-time data channel of the model and the blast furnace sensor is established, the standardized temperature and operation parameters (such as 100 ms frequency of air volume and air temperature data) are obtained from a distributed control system (DCS) in real time; the data preprocessing unit automatically performs outlier rejection, sliding average filtering (window 10 seconds) and Z-score standardization (using the saved μ and σ in the training stage) on the real-time data.

[0142] The real-time data is packaged according to a 14-minute time window (840 time points), input into the model to perform forward propagation, and the predicted value of the furnace hearth temperature in the next 5 minutes is generated; the prediction result is transmitted to the blast furnace operation interface in real time through the OPCUA protocol, and the current temperature trend and warning information (such as “the temperature in the next 3 minutes may exceed 1150℃”) are displayed.

[0143] In the above step S52, a data collection task is triggered at 0 o'clock every morning, and the complete production data of the previous 24 hours is extracted from the DCS database, including: 5-point temperature of the furnace hearth (100 ms sampling), air volume or air temperature (100 ms sampling), coal injection amount or cooling water flow (minute-level average value); blast furnace operation state markers (such as normal production, start-stop furnace, fault alarm).

[0144] Automatically detect the integrity (missing value ratio less than 1%) and rationality (such as temperature greater than 800℃ and less than 1300℃) of the newly added data, and mark the abnormal data segment as "to be reviewed"; linear interpolation is used to match the minute-level parameters (such as coal injection) to the 100ms time sequence to ensure synchronization with the temperature data.

[0145] The newly added data that passes the verification is appended to the original training set in chronological order, and the data within the last 30 days (about 1.8 million samples) is retained. The old data exceeding 30 days is automatically deleted to maintain the stable size of the data set.

[0146] The above step S53: incremental learning and model fine-tuning: lock the parameters (such as weights and biases) of the bidirectional LSTM layer and the 1D convolution 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 the working condition does not change in structure; only the fully connected layer of the output layer is trained, with the updated training set feature vector as the input and the future 5-minute temperature prediction value as the output; the AdamW optimizer (learning rate 1e-5) is used, the loss function is Huber loss, the training rounds are set to 5 rounds, and the batch size is 128. Quickly fit the subtle regular changes in new data.

[0147] After fine-tuning, the performance (such as whether the RMSE is <6.0℃) is evaluated using the validation set data within the last 3 days; if the indicators meet the standards, automatically switch to the new model; if not, skip this update and trigger the data review process (such as checking whether the newly added data contains abnormal working conditions).

[0148] In a preferred embodiment of the present application, the above step S6: input the real-time standardized data into the deployed model to output the future 5-minute temperature prediction value; when the prediction value exceeds the safety threshold, generate the minimum disturbance strategy and issue it to the execution system; the executed temperature and operation parameters are fed back to step S1, which can include:

[0149] Step S61: input the real-time standardized data of step S1 into the deployed model to output the future 5-minute temperature prediction value;

[0150] Step S62: when the prediction value exceeds the safety threshold: analyze the influence weight of the wind temperature, coal injection, and cooling water flow; generate the minimum disturbance strategy of reducing the wind temperature, reducing the coal injection, and increasing the cooling water flow, and issue it to the execution system to obtain the executed temperature and operation parameters;

[0151] Step S63: feed back the executed temperature and operation parameters to step S1 to form a closed-loop control cycle.

[0152] In the embodiment of the present application, the temperature change trend is perceived 5 minutes in advance by means of the model, the risk of overheating of the lining is discovered earlier than the traditional threshold alarm, valuable time is gained for operation adjustment, and through real-time data preprocessing and lightweight deployment of the model, the prediction error is controlled within ±5.7℃, the temperature prediction accuracy is ensured; the targeted strategy is generated based on the parameter influence weight, the temperature abnormality main cause such as excessively high wind temperature and excessively large coal injection amount is accurately adjusted, the single adjustment amplitude is controlled within the process safety range, the production disturbance is reduced, the air permeability and the combustion efficiency are maintained stable, and the strategy rapid response is realized from the threshold triggering; through the "prediction-adjustment-feedback" closed loop mechanism, the system can dynamically optimize the strategy according to the actual effect, form a self-evolution control loop, the temperature control error is reduced by 20% under long-term operation, and the adjustment data each time is included in the training set, so that the model learns the optimal adjustment mode under different working conditions, the control accuracy is continuously improved, the temperature abnormality processing effect is quickly responded and tracked, and the risk of single adjustment failure is avoided.

[0153] In the embodiment of the present application, when specifically applied, the following technical solutions can be used to achieve, for example:

[0154] The step S61 described above collects the parameters such as the hearth temperature (5 key positions), the air volume, the air temperature, the coal injection amount, etc. from the blast furnace sensor in real time, the frequency is 100 ms; the collected data is subjected to outlier rejection (such as cross verification by adjacent sensor data), sliding average filtering (window 10 seconds smooth fluctuation), and then Z-score standardization is performed using the mean (μ) and standard deviation (σ) saved in the training stage, and is converted into an input format recognizable by the model.

[0155] The real-time data after standardization is packaged according to a 14-minute time window (including 840 time points), and is input into the model deployed in the real-time control system; the model extracts time-dependent features (such as the temperature rising trend in the past 10 minutes) through a bidirectional LSTM layer, identifies spatial anomalies (such as temperature mutation at the center of the hearth) through a 1D convolution layer, and outputs the predicted value of the average hearth temperature in the next 5 minutes (such as 1145℃) after weighted by an attention fusion layer.

[0156] The step S62 described above, the preset safety threshold is 1150℃ (set according to the high temperature resistance limit of the lining material in the blast furnace), 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 influence degree of each operating parameter on the temperature under the current working condition (such as wind temperature weight 35%, coal injection amount weight 28%, and cooling water flow weight 25%); the parameters with high weight are adjusted preferentially to achieve the maximum cooling effect with the minimum operation change.

[0157] When the minimum disturbance strategy is generated, the influence weight of each operating parameter on the temperature is adjusted: the wind temperature reduction amplitude is determined according to the weight proportion (for example, when the weight proportion is 35%, the wind temperature is reduced by 3.5°C), the coal injection amount is reduced according to the influence degree (for example, when the weight is 28%, the coal injection amount is reduced by 1.5t / h), and the cooling water flow is increased according to the heat exchange capacity demand (when the weight is 25%, the cooling water flow is increased by 5m 3 / h); at the same time, the strategy needs to strictly meet the process constraint conditions, such as the coal injection amount being not less than the minimum safety value and the cooling water flow being not more than the pipeline transportation limit, so as to ensure the safety and stability of the blast furnace operation while realizing the effective regulation of the temperature.

[0158] The adjustment instructions are sent to the fan, the coal injection pump and the cooling water valve through the DCS system, and the execution time is less than 10 seconds; the parameters before adjustment (such as the wind temperature of 1200°C and the coal injection amount of 15t / h) and the parameters after adjustment (such as the wind temperature of 1196.5°C and the coal injection amount of 13.5t / h) and the execution time stamp are recorded.

[0159] After the above step S63, the adjustment is executed, the change of the temperature of the furnace hearth is continuously monitored, the actual temperature value (such as the temperature of 1140°C after 5 minutes of adjustment) and the operating parameters (such as the wind temperature, the coal injection amount and the like) are recorded every 1 minute; the temperature and the parameter data after execution are transmitted to the real-time data acquisition module of step S1 as new training samples to participate in the next round of data preprocessing; the system automatically marks the data segment as "sample after strategy execution" for analyzing the adjustment effect in subsequent model training.

[0160] The temperature change amount before and after adjustment (such as ΔT = predicted value - actual value = 1145°C - 1140°C = 5°C) is calculated to evaluate the effectiveness of the strategy; if the temperature has not decreased to the safety range (such as still greater than 1150°C), the strategy iteration mechanism is automatically triggered (such as increasing the adjustment amplitude of the cooling water flow).

[0161] As shown in Figure 4 , the blast furnace hearth temperature sensor is arranged, the hearth thermocouple is arranged inside the furnace hearth side wall 1, the installation depth is 20cm, a plurality of hearth thermocouples 3 are arranged in a ring shape inside the furnace hearth side wall 1, the installation depth is 200cm; the hearth center thermocouple 5 is installed at the inner center position of the furnace hearth side wall 1, the installation depth is 460cm; the hearth sub-center thermocouple 4 is installed outside the hearth center thermocouple 5, the installation depth is 400cm; the hearth thermocouple 3 is arranged outside the periphery of the hearth sub-center thermocouple 4.

[0162] As shown in Figure 2 , the embodiment of the present application also provides a safe and intelligent adjustment control system for the high-temperature blast furnace hearth, which comprises:

[0163] The collection module is used for collecting temperature data of each position of the furnace hearth and synchronously collecting blast furnace operation parameters in real time through a high-precision temperature sensor array arranged on the inner wall and the bottom of the furnace hearth; and the collected temperature data and operation parameters are preprocessed to generate a standardized time series data set;

[0164] The construction module is used for constructing a deep time series network regression model through the standardized time series data set, and the model comprises a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer;

[0165] The training module is used for dividing the data set in time sequence and training based on the deep time series network model to obtain a trained deep time series network model;

[0166] The verification module is used for inputting the trained deep time series network model into a verification set and a test set; calculating mean square error, root mean square error and mean absolute error to verify the prediction accuracy of the model; and outputting a verification passed model meeting an error threshold;

[0167] The deployment module is used for deploying the verification passed model to a deployment control system; adding new data to the training set every 24 hours; fixing the parameters of the bidirectional LSTM and the convolution layer and fine-tuning the weight of the full connection layer;

[0168] The feedback module is used for inputting real-time standardized data into the deployed model to output a 5-minute temperature prediction value; and generating a minimum disturbance strategy and issuing it to an execution system when the prediction value exceeds a safety threshold.

[0169] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0170] The above is the preferred embodiment of the present application, and it should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A safe intelligent adjustment control method for high temperature of a blast furnace hearth, characterized in that, The method comprises: Step S1: collecting temperature data of each position of the furnace by a high-precision temperature sensor array arranged on the inner wall and the bottom of the furnace, and synchronously collecting blast furnace operation parameters; and pre-processing the collected temperature data and operation parameters to generate a standardized time series data set; Step S2: constructing a deep time series network regression model through the standardized time series data set, wherein the model comprises a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer; Step S3: based on the deep time series network model, dividing the data set in time sequence and training to obtain a trained deep time series network model; Step S4: inputting the trained deep time series network model into a validation set and a test set; calculating mean square error, root mean square error and mean absolute error to verify the prediction accuracy of the model; and outputting a verified model meeting the error threshold; Step S5: deploying the verified model to a real-time control system; adding new data to the training set every 24 hours; fixing the parameters of the bidirectional LSTM and the convolution layer, and fine-tuning the weights of the fully connected layer; Step S6: inputting real-time standardized data into the deployed model to output temperature prediction values in the next 5 minutes; generating a minimum disturbance strategy and issuing it to an execution system when the prediction value exceeds a safety threshold; and feeding back the executed temperature and operation parameters to step S1.

2. The safety intelligent adjustment control method for the high temperature of the blast furnace hearth according to claim 1, characterized in that, Step S1: collecting temperature data of each position of the furnace by a high-precision temperature sensor array arranged on the inner wall and the bottom of the furnace, and synchronously collecting blast furnace operation parameters; The collected temperature data and operation parameters are pre-processed to generate a standardized time series data set, including: Collecting temperature data of each position of the furnace by a high-precision temperature sensor array arranged on the inner wall and the bottom of the furnace; and synchronously collecting blast furnace operation parameters: air volume, air temperature, coal injection amount, cooling water flow and charge descending speed; Removing outliers in the temperature data and operation parameters; smoothing the time series fluctuations by using the moving average method; converting the data by Zscore standardization to generate a standardized time series data set matching the model input dimension, and recording the standardization parameters.

3. The safety intelligent adjustment control method for high temperature of the blast furnace hearth according to claim 2, characterized in that, Step S2: constructing a deep time series network regression model through the standardized time series data set, wherein the model comprises a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer, including: Constructing a deep time series network regression model through the standardized time series data set; the model comprises a bidirectional LSTM layer extracting time-dependent features, a 1D convolution layer extracting spatial mutation features, and an attention fusion layer weighting key features.

4. The safety intelligent adjustment control method for the high temperature of the blast furnace hearth according to claim 3, characterized in that, Step S3: based on the deep time series network model, dividing the data set in time sequence and training to obtain a trained deep time series network model, including: Based on the deep time series network model, dividing the data set into a training set, a validation set and a test set in time sequence; Training the model using the training set to optimize the parameters with the objective of minimizing the prediction error to obtain an initial training model; Evaluating the performance of the initial training model through cross-validation, and based on the evaluation result, adjusting the hyperparameters by using Bayesian optimization to obtain an adjusted model; Using the validation set to verify the adjusted model, and according to the verification result, fine-tuning the parameters to output a trained deep time series network model.

5. The safety intelligent adjustment control method for high temperature of the blast furnace hearth according to claim 4, characterized in that, Step S4: input the trained deep time series 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 prediction accuracy of the model; Output the validation passed model that meets the error threshold, including: Input the trained deep time series network model into the validation set and the test set to obtain the output results of the model on different data sets; Based on the output results, calculate evaluation indicators, including mean square error, root mean square error and mean absolute error; Verify the prediction accuracy of the model through the evaluation indicators. When the evaluation indicators meet the preset error threshold, output the validated model.

6. The safety intelligent adjustment control method for the high temperature of the blast furnace hearth according to claim 5, 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 bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights, including: Deploy the validated model to the real-time control system to realize real-time monitoring and prediction of the blast furnace hearth temperature; During the operation of the real-time control system, collect new data every 24 hours and add it to the training set to form an updated training data set; Based on the updated training data set, fix the parameters of the bidirectional LSTM layer and the 1D convolution layer, and only fine-tune the weights of the fully connected layer to complete the incremental learning of the model.

7. The safety intelligent adjustment control method for high temperature of the blast furnace hearth according to claim 6, characterized in that, Step S6: input real-time standardized data into the deployed model to output 5-minute temperature prediction value; When the prediction value exceeds the safety threshold, generate the minimum disturbance strategy and issue it to the execution system; The executed temperature and operation parameters are fed back to step S1, including: Input real-time standardized data of step S1 into the deployed model to output 5-minute temperature prediction value; When the prediction value exceeds the safety threshold: analyze the influence weight of the wind temperature, coal injection amount and cooling water flow; generate the minimum disturbance strategy of reducing the wind temperature, reducing the coal injection amount and increasing the cooling water flow, and issue it to the execution system to obtain the executed temperature and operation parameters; The executed temperature and operation parameters are fed back to step S1 to form a closed-loop control cycle.

8. A safe intelligent adjustment control system for high temperature of a blast furnace hearth, which realizes the method as claimed in any one of claims 1 to 7, characterized in that, Including: The acquisition module is used to collect temperature data at each position of the hearth and synchronous acquisition of blast furnace operation parameters through high-precision temperature sensor arrays deployed on the inner wall and bottom of the hearth; The collected temperature data and operation parameters are preprocessed to generate standardized time series data sets; The construction module is used to construct a deep time series network regression model through the standardized time series data set, which includes a bidirectional LSTM layer, a 1D convolution layer and an attention fusion layer; The training module is used to divide the data set in time sequence based on the deep time series network model and train it to obtain a trained deep time series network model; The verification module is used to input the trained deep time series 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 prediction accuracy of the model; output the validation passed model that meets the error threshold; The deployment module is used to deploy the validated model to the deployment control system; add new data to the training set every 24 hours; fix the bidirectional LSTM and convolution layer parameters, and fine-tune the fully connected layer weights; The feedback module is used to input real-time standardized data into the deployed model to output 5-minute temperature prediction value; When the prediction value exceeds the safety threshold, a minimum disturbance strategy is generated and issued to an execution system.

9. A computing device, comprising: Comprising: one or more processors; a memory device for storing one or more programs, which when executed by the one or more processors, cause the one or more processors to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program, which when executed by a processor, implements the method of any one of claims 1 to 7.

Images