Blast furnace covering agent performance prediction method and equipment based on data driving

Abstract

The invention relates to the technical field of blast furnace covering agents, and discloses a blast furnace covering agent performance prediction method and equipment based on data driving. The method comprises the following steps: acquiring raw material characteristic data, process control parameter data and service environment parameter data of the blast furnace covering agent; performing preprocessing and feature fusion on the three types of data to obtain a covering agent performance influence factor data set; the data set is input into a preset multi-dimensional prediction model, and a melting performance prediction value, a heat preservation performance prediction value and a crystallization performance prediction value of the blast furnace covering agent are obtained through model calculation; and comparing each prediction value with a preset performance threshold range, and generating a blast furnace covering agent performance evaluation report according to a comparison result. According to the method, comprehensive prediction and evaluation of multiple key performances of the covering agent are realized by integrating multi-source data and utilizing a multi-dimensional model, and reference is provided for application optimization of the covering agent.

Description

Technical Field

[0001] This invention relates to the field of blast furnace covering agent technology, specifically to a data-driven method and equipment for predicting the performance of blast furnace covering agents. Background Technology

[0002] In the steelmaking process, after blast furnace tapping, the molten iron needs to rely on covering agents to maintain its temperature stability and prevent secondary oxidation during transportation and handling. The performance of the blast furnace covering agent directly affects the quality of the molten iron, production efficiency, and energy consumption. Its melting performance determines whether the covering agent can quickly spread on the surface of the molten iron to form a uniform covering layer, its heat-insulating performance affects the rate of temperature drop of the molten iron, and its crystallization performance is related to the structural stability and duration of action of the covering layer.

[0003] The performance control of blast furnace covering agents largely relies on the accumulated experience of operators or small-scale experimental exploration. In terms of performance prediction, early methods often established empirical formulas based on a single factor, such as inferring melting performance solely based on the content of a certain mineral in the raw material, or judging the insulation effect based on the holding time. This approach ignores the interaction between raw material characteristics, process control, and the service environment, leading to significant deviations between predicted results and actual operating conditions.

[0004] With the development of smelting technology, some studies have begun to introduce experimental testing methods, which involve preparing samples and conducting performance tests in a laboratory environment that simulates a blast furnace. However, such methods have significant limitations: the experimental process consumes a large amount of raw materials and energy, and the cycle is long, often taking several days from sample preparation to performance testing, making it difficult to meet the needs of rapid adjustment of process parameters in steel production; the laboratory simulation environment differs from the actual blast furnace service environment, and factors such as ambient humidity and actual temperature fluctuations of molten iron are difficult to completely replicate, thus limiting the reference value of the test results.

[0005] Existing data processing methods also have shortcomings. Raw material characteristic data, process control parameter data, and service environment parameter data often come from different acquisition systems, resulting in differences in data format and accuracy, and containing a large amount of noise and redundant information. Traditional processing methods mostly involve simple screening or single-dimensional analysis, which cannot effectively integrate multi-source data and struggle to capture the complex relationships between various factors. For example, the synergistic effect of raw material particle size distribution and cooling rate on crystallization performance is often overlooked in single-factor analysis, thus affecting the accuracy of predictions.

[0006] Existing predictive models are mostly designed for a single property of the covering agent, and cannot simultaneously output prediction results for multiple key properties such as melting, insulation, and crystallization. In actual production, the various properties of the covering agent are interrelated and mutually influential. Meeting the standard of a single property does not mean that the overall performance meets the requirements. This makes it difficult for traditional models to meet the needs of comprehensively evaluating the suitability of the covering agent, often leading to problems such as uneven spreading of the covering layer, insufficient insulation effect, or premature cracking in production. Summary of the Invention

[0007] The purpose of this invention is to provide a data-driven method for predicting the performance of blast furnace covering agents, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a data-driven method for predicting the performance of blast furnace covering agents, the method comprising:

[0009] Acquire raw material characteristic data, process control parameter data, and service environment parameter data for blast furnace covering agents. Raw material characteristic data includes raw material particle size distribution parameters, raw material mineral composition ratio, and raw material physical property parameters. Process control parameter data includes smelting temperature parameters, holding time parameters, and cooling rate parameters. Service environment parameter data includes molten iron temperature parameters, slag layer thickness parameters, and ambient humidity parameters.

[0010] The raw material characteristic data, process control parameter data, and service environment parameter data of blast furnace covering agent were preprocessed and feature fused to obtain a dataset of covering agent performance influencing factors.

[0011] The dataset of factors affecting the performance of the covering agent is input into the preset multi-dimensional prediction model, and the predicted values ​​of the melting performance, heat preservation performance and crystallization performance of the blast furnace covering agent are obtained through model calculation.

[0012] The predicted values ​​of melting performance, heat preservation performance, and crystallization performance are compared with the preset performance threshold ranges, and a blast furnace covering agent performance evaluation report is generated based on the comparison results.

[0013] Preferably, the raw material particle size distribution parameters are specifically the uniformity of raw material particle size distribution, the maximum particle diameter of raw material, and the minimum particle diameter of raw material; the raw material mineral composition ratio is specifically the molar ratio of calcium oxide to silicon dioxide, the mass fraction of magnesium oxide, and the volume fraction of aluminum oxide; the raw material physical property parameters are specifically the raw material bulk density, the raw material porosity, and the raw material specific heat capacity; the smelting temperature parameters are specifically the peak melting temperature, the heating rate, and the isothermal holding temperature; the holding time parameters are specifically the high-temperature holding time, the medium-temperature holding time, and the low-temperature holding time; the cooling rate parameters are specifically the rapid cooling stage rate, the slow cooling stage rate, and the isothermal transformation stage duration; the molten iron temperature parameters are specifically the initial molten iron temperature, the molten iron cooling rate, and the target molten iron temperature; the slag layer thickness parameters are specifically the initial slag layer thickness, the slag layer solidification thickness, and the slag layer uniformity coefficient; and the environmental humidity parameters are specifically the environmental relative humidity, the water vapor partial pressure, and the humidity change rate.

[0014] Preferably, the specific steps for preprocessing and feature fusion of the raw material characteristic data of blast furnace covering agent are as follows:

[0015] Outliers in the particle size distribution parameters of raw materials are judged and corrected using the Laida criterion; missing values ​​in the mineral composition ratio of raw materials are subjected to multiple interpolation; and duplicate data in the physical property parameters of raw materials are deduplicated based on timestamps.

[0016] The particle size distribution parameters, mineral composition ratios, and physical property parameters of the pretreated raw materials were standardized and converted to obtain dimensionless basic data of raw material characteristics.

[0017] Based on dimensionless raw material characteristic data, the influence weight of each parameter on the performance of the covering agent is calculated by grey relational analysis. The parameters with influence weights greater than a preset threshold are combined to obtain a raw material characteristic fusion feature set.

[0018] Preferably, the specific steps for preprocessing and feature fusion of process control parameter data are as follows:

[0019] Time series alignment was performed on melting temperature parameters, holding time parameters, and cooling rate parameters to unify parameters with different acquisition frequencies into time series data with the same time interval;

[0020] Sliding window sampling is performed on the time-series process control parameter data, and the maximum, minimum, average and standard deviation within each window are extracted as statistical features;

[0021] By concatenating the statistical features with the original values ​​of the process control parameters, and reducing the data dimensionality through principal component analysis, a fusion feature set of process control is obtained.

[0022] Preferably, the specific steps for preprocessing and feature fusion of service environment parameter data are as follows:

[0023] The dimensional consistency of parameters such as molten iron temperature, slag layer thickness, and ambient humidity was verified, and parameters in different units were converted into standard data under the International System of Units (SI).

[0024] Trend terms are extracted from standard data, and the linear trend in the data is eliminated by the difference method to obtain stable service environment data.

[0025] Based on stationary service environment data, wavelet transform is used to extract fluctuation features in different frequency bands. The fluctuation features are then fused with the stationary data to obtain a fused feature set of the service environment.

[0026] Preferably, the specific steps for obtaining the dataset of factors affecting the performance of covering agents are as follows:

[0027] Align the sample dimensions of the raw material characteristics fusion feature set, process control fusion feature set, and service environment fusion feature set to ensure that the number of samples in each feature set corresponds one-to-one with the timestamp.

[0028] The three aligned feature sets are horizontally concatenated to form an initial feature matrix containing multi-dimensional information on raw materials, processes, and environment.

[0029] Feature selection was performed on the initial feature matrix, and redundant features were removed using a recursive feature elimination method. Feature variables that are strongly correlated with the performance of the covering agent were retained, resulting in a dataset of covering agent performance influencing factors.

[0030] The preferred, pre-defined steps for constructing a multi-dimensional prediction model are as follows:

[0031] A deep learning network architecture consisting of an input layer, a hidden layer, and an output layer was constructed. The number of neurons in the input layer was consistent with the feature dimensions of the coating agent performance influencing factor dataset. The output layer contained three prediction branches: melting performance, heat preservation performance, and crystallization performance.

[0032] The hidden layer adopts a residual network structure. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function layer. Feature reuse is achieved between adjacent residual blocks through skip connections.

[0033] The output layer uses a combination of a fully connected layer and a softmax activation function to output the probability distributions of the predicted values ​​of melting performance, heat preservation performance, and crystallization performance, respectively.

[0034] Preferably, the specific steps for obtaining the predicted values ​​of melting performance, heat preservation performance, and crystallization performance of the blast furnace covering agent through model calculation are as follows:

[0035] The dataset of influencing factors on the performance of covering agents was divided into training sample set, validation sample set and test sample set;

[0036] The multidimensional prediction model is trained using a training sample set, and the model parameters are optimized using an adaptive moment estimation algorithm. The prediction error of the model is monitored in real time using a validation sample set.

[0037] When the prediction error of the validation sample set no longer decreases for 5 consecutive iterations, stop model training, use the test sample set to evaluate the model's performance, and if the evaluation metric meets the preset standard, output the trained multi-dimensional prediction model.

[0038] The data of the influencing factors of the coating agent performance to be predicted are input into the trained multi-dimensional prediction model, and the predicted values ​​of melting performance, heat preservation performance and crystallization performance are obtained by forward propagation calculation.

[0039] Preferably, the specific steps for generating a blast furnace covering agent performance evaluation report are as follows:

[0040] The predicted melt performance value is compared with the preset melt performance threshold range. If the predicted value is within the threshold range, the melt performance is deemed qualified; otherwise, it is deemed unqualified and marked as out of tolerance.

[0041] The predicted value of thermal insulation performance is compared with the preset thermal insulation performance threshold range. If the predicted value is within the threshold range, the thermal insulation performance is deemed to be qualified; otherwise, it is deemed to be unqualified and marked as out of tolerance.

[0042] The predicted crystallization performance value is compared with the preset crystallization performance threshold range. If the predicted value is within the threshold range, the crystallization performance is deemed qualified; otherwise, it is deemed unqualified and marked as out of tolerance.

[0043] Summarize the judgment results of melting performance, heat preservation performance, and crystallization performance, and generate a blast furnace covering agent performance evaluation report that includes qualified items, unqualified items, and analysis of the reasons for deviations.

[0044] Preferably, the present invention further includes an apparatus comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement the above-described data-driven blast furnace covering agent performance prediction method.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This data-driven method for predicting the performance of blast furnace covering agents comprehensively captures the factors influencing their performance by systematically integrating three types of data: raw material characteristics, process control, and service environment. Parameters in raw material characteristics, such as particle size distribution and mineral composition, are directly related to the inherent properties of the covering agent; parameters in process control, such as smelting temperature and holding time, reflect the key conditions in the covering agent formation process; and parameters in the service environment, such as molten iron temperature and slag layer thickness, reflect the scenario characteristics of the covering agent's actual function. The comprehensive inclusion of these three types of data breaks through the limitations of traditional methods that focus only on a single aspect, and can more completely present the complex network of factors affecting covering agent performance.

[0047] The preprocessing and feature fusion of the data effectively solved problems such as inconsistent formats and noise interference from multiple data sources. By cleaning outliers, unifying data dimensions, and extracting key features, the originally scattered and messy data was transformed into a dataset of influencing factors with clear physical meaning. This processing method reduces the interference of data redundancy on prediction results, while highlighting the intrinsic correlation between different factors, such as the law that the composition of raw material minerals and smelting temperature jointly affect melting performance, making the data input into the model closer to the causal relationships in actual working conditions.

[0048] The pre-defined multi-dimensional prediction model overcomes the limitations of traditional single-performance prediction, simultaneously outputting predicted values ​​for melting performance, heat retention performance, and crystallization performance. This model can identify the differences in the influencing mechanisms behind different performance characteristics. For example, melting performance depends more on the matching degree between the raw material's melting characteristics and the smelting temperature; heat retention performance is related to the thermal conductivity coefficient of the covering agent and the temperature gradient of the service environment; and crystallization performance is jointly regulated by the cooling rate and mineral composition. The multi-dimensional prediction results can more comprehensively reflect the overall performance status of the covering agent, avoiding the one-sidedness of single-performance evaluation.

[0049] By comparing predicted values ​​with preset performance threshold ranges and generating evaluation reports, the performance evaluation process becomes more objective and operational. Clear threshold comparisons clearly demonstrate the compliance status of various performance parameters, eliminating reliance on operator subjective judgment. When a predicted performance value exceeds the threshold range, the specific cause can be traced using influencing factor data. For example, substandard melting performance may stem from unreasonable raw material particle size distribution or deviations in smelting temperature settings, providing a clear direction for targeted adjustments. This complete process, from data acquisition to evaluation report generation, shifts the performance prediction of covering agents from experience-driven to data-driven, better adapting to the requirements of refined processes and precise parameters in modern steel production, and reducing production fluctuations caused by covering agent performance issues. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the data-driven blast furnace covering agent performance prediction method described in this invention.

[0051] Figure 2 A flowchart for process control parameter data preprocessing and feature fusion;

[0052] Figure 3 A flowchart for preprocessing and feature fusion of service environment parameter data;

[0053] Figure 4 Design diagram for building a multi-dimensional prediction model. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides a data-driven method and equipment for predicting the performance of blast furnace covering agents. The method includes the following steps:

[0056] Acquire raw material characteristic data, process control parameter data, and service environment parameter data for blast furnace covering agents. Raw material characteristic data includes raw material particle size distribution parameters, raw material mineral composition ratio, and raw material physical property parameters. Process control parameter data includes smelting temperature parameters, holding time parameters, and cooling rate parameters. Service environment parameter data includes molten iron temperature parameters, slag layer thickness parameters, and ambient humidity parameters.

[0057] The raw material characteristic data, process control parameter data, and service environment parameter data of blast furnace covering agent were preprocessed and feature fusion was performed to obtain a dataset of covering agent performance influencing factors.

[0058] The dataset of factors affecting the performance of the covering agent is input into a preset multi-dimensional prediction model. The predicted values ​​of the melting performance, heat preservation performance, and crystallization performance of the blast furnace covering agent are obtained through model calculation.

[0059] The predicted values ​​of melting performance, heat preservation performance, and crystallization performance are compared with the preset performance threshold ranges, and a blast furnace covering agent performance evaluation report is generated based on the comparison results.

[0060] Example 1: See Figure 2The raw material particle size distribution parameters involve the uniformity of particle size distribution, the maximum particle diameter, and the minimum particle diameter. The uniformity of particle size distribution reflects the distribution of particle size, while the maximum and minimum particle diameters represent the upper and lower limits of particle size, respectively. The raw material mineral composition ratio includes the molar ratio of calcium oxide to silicon dioxide, the mass fraction of magnesium oxide, and the volume fraction of aluminum oxide. These parameters directly reflect the composition of various minerals in the raw material. The raw material physical properties parameters cover bulk density, porosity, and specific heat capacity. Bulk density refers to the mass per unit volume of the raw material in its natural packed state; porosity reflects the number of pores within the raw material; and specific heat capacity represents the amount of heat required to raise the temperature of a unit mass of raw material by 1°C.

[0061] The melting temperature parameters in the process control data specifically include peak melting temperature, heating rate, and isothermal holding temperature. Peak melting temperature is the highest temperature reached by the raw material during melting; heating rate is the rate of temperature increase per unit time; and isothermal holding temperature refers to the constant temperature value maintained during melting. Holding time parameters include high-temperature holding time, medium-temperature holding time, and low-temperature holding time, corresponding to the duration of holding at different temperature ranges. Cooling rate parameters consist of rapid cooling stage rate, slow cooling stage rate, and isothermal transition stage duration. Rapid cooling stage rate refers to the temperature change per unit time during rapid cooling; slow cooling stage rate is the temperature change rate during slow cooling; and isothermal transition stage duration is the time required for the transition at a constant temperature.

[0062] The service environment parameters include molten iron temperature parameters such as initial molten iron temperature, cooling rate, and target temperature. Initial molten iron temperature is the temperature of the molten iron when it enters the service environment; cooling rate describes how quickly the temperature drops; and target temperature is the temperature the molten iron is expected to reach during service. Slag layer thickness parameters include initial slag layer thickness, solidified slag layer thickness, and slag layer uniformity coefficient. Initial slag layer thickness is the thickness of the slag layer at the start of service; solidified slag layer thickness is the thickness of the solidified portion of the slag layer; and the uniformity coefficient reflects the uniformity of the slag layer thickness distribution. Environmental humidity parameters include relative humidity, partial pressure of water vapor, and rate of change of humidity. Relative humidity is the percentage of actual water vapor pressure in the air to saturated water vapor pressure at the same temperature; partial pressure of water vapor is the pressure generated by water vapor in a gas mixture; and rate of change of humidity is the change in environmental humidity per unit time.

[0063] When preprocessing and feature fusion of raw material characteristic data for blast furnace covering agents, outliers in the raw material particle size distribution parameters are addressed first. Using the Laida criterion, the mean and standard deviation of the raw material particle size distribution parameter dataset are calculated first. Data deviating from the mean by more than three times the standard deviation are then identified as outliers. These outliers are corrected by averaging the adjacent normal data in the sample containing that parameter, thus ensuring the outlier data is appropriately adjusted.

[0064] For missing values ​​in the raw material mineral composition ratio, multiple imputation is employed. Specifically, multiple complete datasets are constructed based on existing complete data. In each dataset, missing values ​​are imputed differently according to the data distribution characteristics and correlations. Subsequently, the analysis results of multiple datasets are synthesized, taking into account the imputation results of each dataset, to obtain the final imputation result, thereby compensating for the impact of missing data.

[0065] For duplicate data in the physical property parameters of raw materials, a timestamp-based deduplication process is performed. The timestamp of each data record is checked, the record with the most recent timestamp is retained, and other duplicate records with the same parameter information are deleted, ensuring the uniqueness and accuracy of the raw material physical property parameter data.

[0066] After preprocessing, the raw material particle size distribution parameters, raw material mineral composition ratios, and raw material physical property parameters are standardized. By subtracting the average value of each parameter from the average value of the entire dataset and then dividing by the standard deviation of the parameter, the transformed parameter values ​​are made to be of the same order of magnitude, eliminating the influence of dimensions and obtaining dimensionless basic data of raw material characteristics.

[0067] Based on dimensionless raw material characteristic data, grey relational analysis was used to calculate the influence weight of each parameter on the performance of the covering agent. Grey relational analysis determines the weight by analyzing the degree of correlation between each parameter and the performance of the covering agent. A preset threshold was set, and parameters with influence weights greater than this threshold were selected for feature combination. These parameters were considered key parameters with a significant impact on the performance of the covering agent. The resulting combination formed a raw material characteristic fusion feature set, which integrated relevant information on raw material characteristics that significantly affected the performance of the covering agent.

[0068] Example 2: See Figure 3When preprocessing and feature fusion of process control parameter data, the first step is to align the time series of melting temperature, holding time, and cooling rate parameters. Because the acquisition frequencies of different parameters vary—for example, melting temperature might be acquired every 30 seconds, holding time every 2 minutes, and cooling rate every minute—the time series lengths and intervals for these parameters are inconsistent. Therefore, interpolation or resampling is needed to convert these parameters into time series data with the same time interval. Specifically, a suitable time interval, such as 1 minute, can be chosen as the standard. For parameters with acquisition frequencies higher than this standard, data at fixed time points is extracted to reduce the frequency; for parameters with acquisition frequencies lower than this standard, linear interpolation is used to supplement data at intermediate moments, ensuring that all process control parameters form a continuous time series with a 1-minute interval, guaranteeing that the parameters are matched in the time dimension.

[0069] After time series alignment, sliding window sampling is performed on the time-series process control parameter data. The size and step size of the sliding window can be set according to the data variation characteristics. For example, a window size of 5 minutes and a step size of 1 minute mean that each window contains data from 5 consecutive time points, with a one-time-point difference between adjacent windows. Within each window, the maximum, minimum, average, and standard deviation of the melting temperature, holding time, and cooling rate parameters are calculated. The maximum value reflects the highest parameter value within the window, the minimum value reflects the lowest parameter value, the average value reflects the overall level of the parameter within the window, and the standard deviation reflects the degree of fluctuation of the parameter within the window. By extracting these statistical features, the local variation characteristics in the time series data can be transformed into quantifiable indicators, enriching the information dimensions contained in the data.

[0070] The extracted statistical features are concatenated with the original values ​​of the process control parameters to form a combined dataset containing both the original parameter values ​​and the statistical features. For example, for the melting temperature parameter at a certain moment, its original value, along with the maximum, minimum, average, and standard deviation of the window at that moment, constitutes the extended features for that moment. After concatenation, the combined dataset is processed using principal component analysis (PCA) to reduce data dimensionality. PCA constructs new feature variables (principal components) to transform multiple related original features into a few unrelated composite features, which retain most of the information in the original data. During processing, the covariance matrix of the combined dataset is calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The number of principal components is determined based on the magnitude of the eigenvalues. Typically, the top few principal components with a cumulative contribution rate reaching a certain proportion (e.g., 85%) are selected as the final features, thereby eliminating redundant information and obtaining a process control fusion feature set with lower dimensionality and more concentrated information.

[0071] When preprocessing and feature fusion of service environment parameter data, the first step is to verify dimensional consistency. Hot metal temperature parameters may be recorded in degrees Celsius, slag layer thickness parameters in centimeters, and water vapor partial pressure in ambient humidity parameters in kilopascals. These different units can lead to significant differences in the numerical scale of the data. All parameters need to be converted to standard data in the International System of Units (SI). For example, hot metal temperature parameters should be converted from degrees Celsius to Kelvin (i.e., the degree Celsius value plus 273.15), slag layer thickness parameters from centimeters to meters (i.e., the centimeter value divided by 100), and water vapor partial pressure from kilopascals to pascals (i.e., the kilopascal value multiplied by 1000). This ensures that all service environment parameters maintain unit consistency and eliminates the impact of dimensional differences on subsequent processing.

[0072] After unit conversion, trend terms are extracted from the standard data, and the finite difference method is used to eliminate linear trends. A linear trend refers to a continuous upward or downward trend in the data over time, which may mask the inherent volatility of the data. The finite difference method calculates the difference between data points at two adjacent time points—that is, the data at the later time point minus the data at the previous time point—to obtain a difference sequence. For example, for time-series data of molten iron temperature, the difference between the temperature value at time t+1 and the temperature value at time t is calculated to form a new difference sequence. This sequence effectively removes the linear trend from the original data, making the data exhibit a more stable characteristic. For data that still shows a slight trend after finite difference processing, multiple finite difference operations can be performed until stable service environment data is obtained.

[0073] Based on stationary service environment data, wavelet transform is used to extract fluctuation features in different frequency bands. Wavelet transform can decompose time series data into different frequency channels, thereby identifying fluctuation patterns at different time scales. For example, by selecting appropriate wavelet basis functions and decomposition levels, stationary data can be decomposed into low-frequency and high-frequency components. The low-frequency component reflects long-term, slow changes in the data, while the high-frequency component reflects short-term, rapid fluctuations. The coefficients of each frequency band obtained from the decomposition are reconstructed to obtain fluctuation sequences for different frequency bands, which correspond to fluctuation features within different frequency ranges. These fluctuation features are then fused with the stationary data, that is, the fluctuation sequences of each frequency band are added as new features to the stationary data, forming a service environment fusion feature set that includes the original stationary features and multi-frequency fluctuation features. This feature set can comprehensively reflect the variation characteristics of service environment parameters in both time and frequency dimensions.

[0074] Example 3: See Figure 4In obtaining the dataset of influencing factors for coating agent performance, the first step is to align the sample dimensions of the raw material characteristic fusion feature set, the process control fusion feature set, and the service environment fusion feature set. Each feature set contains multiple samples, each corresponding to a specific timestamp that marks the exact moment of data collection or generation. The number of samples in each of the three feature sets needs to be checked one by one. If there are discrepancies, it is necessary to check for missing or duplicate samples, delete redundant samples, or supplement missing samples through interpolation to ensure that the number of samples in the three feature sets is consistent. Simultaneously, for each sample, it is essential to ensure that its corresponding timestamp is exactly the same in all three feature sets. For example, if the timestamp of a sample in the raw material characteristic fusion feature set is "2023-10-01 08:00:00", then the corresponding samples in the process control fusion feature set and the service environment fusion feature set must also have the same timestamp to avoid data association errors between different feature sets due to time mismatches.

[0075] After aligning the sample dimensions, the three feature sets are horizontally concatenated. Horizontal concatenation refers to combining the raw material characteristic fusion feature, process control fusion feature, and service environment fusion feature at the same time point to form a comprehensive feature vector containing multi-dimensional information. For example, at a certain time point, the raw material characteristic fusion feature set contains 5 feature values, the process control fusion feature set contains 8 feature values, and the service environment fusion feature set contains 6 feature values. After concatenation, the feature vector corresponding to that time point will contain 5 + 8 + 6 = 19 feature values. The feature vectors corresponding to all time points together constitute the initial feature matrix. The number of rows in the matrix equals the number of samples, and the number of columns equals the sum of the number of features in the three feature sets. Each element in the matrix represents the specific value of a certain feature at a certain time point.

[0076] After the initial feature matrix is ​​formed, feature selection is required, using a recursive feature elimination method to remove redundant features. The core of the recursive feature elimination method is to repeatedly build the model and select features, removing the least important features each time, until a predetermined number of features remain. Specifically, a base model is first trained using the initial feature matrix. The least important features are determined by the model's output feature importance score and removed, resulting in a new feature subset. Then, the model is retrained based on this new feature subset, and the least important features are removed again. This process is repeated until the number of features in the feature subset meets the predetermined requirement. A stopping condition needs to be set during this process, for example, stopping when the number of features decreases to 50% of the initial number. The remaining feature variables have a strong correlation with the coverage agent performance, and these feature variables together constitute the coverage agent performance influencing factor dataset.

[0077] The construction of the pre-defined multi-dimensional prediction model begins with network architecture design, which includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer strictly corresponds to the feature dimensions of the cover agent performance impact factor dataset. For example, if the cover agent performance impact factor dataset contains 20 feature dimensions, then the input layer will have 20 neurons, each receiving a value from one feature dimension as input, ensuring that the dataset can be completely fed into the model for processing. The output layer is divided into three prediction branches, corresponding to the prediction of melting performance, heat preservation performance, and crystallization performance, respectively. Each branch contains a number of neurons, the specific number of which is determined according to the required prediction accuracy. For example, each branch may have one neuron, directly outputting the corresponding performance prediction value.

[0078] The hidden layers employ a residual network structure, which consists of multiple residual blocks connected in series. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function layer. The convolutional layers perform convolution operations on the input data using a sliding window, extracting local features. The window size and stride can be set according to the data characteristics, for example, a 3×3 window with a stride of 1. The batch normalization layer standardizes the output data from the convolutional layers, adjusting it to a distribution with a mean of 0 and a standard deviation of 1, reducing the impact of data distribution variations on model training. The activation function layer uses a nonlinear function to transform the output of the batch normalization layer, introducing nonlinear factors to enable the model to fit complex nonlinear relationships. Commonly used activation functions include ReLU.

[0079] Feature reuse is achieved between adjacent residual blocks through skip connections. A skip connection directly adds the input of the previous residual block to the output of the next residual block to form a new output. For example, if the input of the previous residual block is X, and the output after convolution, batch normalization, and activation function processing is F(X), then the output after the skip connection is X+F(X). This connection method can effectively alleviate the gradient vanishing problem that may occur in deep networks, allowing the model to maintain good training performance while increasing its depth.

[0080] The output layer uses a combination of a fully connected layer and a softmax activation function. The fully connected layer transforms the feature vector output from the hidden layer into a vector that matches the dimension of the predicted target. For example, if the feature vector output from the hidden layer has a dimension of 128, the fully connected layer can have 3 neurons to convert the 128-dimensional feature vector into a 3-dimensional vector, corresponding to the original predicted values ​​of melting performance, heat preservation performance, and crystallization performance, respectively. The softmax activation function then transforms these original predicted values ​​into a probability distribution, the calculation formula of which is:

[0081]

[0082] Where P(y=k) represents the probability that the predicted result is the k-th performance category, z kThis represents the k-th original prediction output of the fully connected layer, where e is the natural constant. This represents the exponential sum of the three original predicted values ​​output by the fully connected layer. After processing with the softmax activation function, the sum of the probabilities of the three performance prediction values ​​is 1. The magnitude of each probability value reflects the model's confidence in the performance prediction result. The final output prediction values ​​of melting performance, heat preservation performance, and crystallization performance are determined based on these probability distributions.

[0083] Example 4: In the process of obtaining the predicted values ​​of melting performance, heat preservation performance, and crystallization performance of blast furnace covering agent through model calculation, it is first necessary to divide the dataset of influencing factors of covering agent performance into training sample set, validation sample set, and test sample set. The division can be carried out according to a certain proportion, for example, randomly selecting 70% of the samples from the dataset as the training sample set, 20% as the validation sample set, and the remaining 10% as the test sample set. During the division process, it is necessary to ensure that the data distribution of the three sample sets is consistent, that is, the distribution of characteristics such as raw material properties, process control, and service environment contained in the training sample set, validation sample set, and test sample set is similar, to avoid deviations in model training or evaluation due to differences in sample distribution. For example, if the dataset contains influencing factor data corresponding to different batches of blast furnace covering agent produced, it is necessary to ensure that the data of each batch is distributed in a certain proportion in the three sample sets during the division, rather than assigning all the data of one batch to the training sample set and all the data of another batch to the validation or test sample set.

[0084] When training a multi-dimensional prediction model using a training sample set, an adaptive moment estimation algorithm (IME) is employed to optimize the model parameters. Model parameters include the weights of convolutional layers in the hidden layers, the mean and variance of batch normalized layers, and the weights of fully connected layers. During optimization, the IME maintains a momentum term and an adaptive learning rate term for each parameter. The momentum term accumulates the gradient direction of the parameter in previous update steps, similar to the concept of inertia in physics, accelerating parameter updates when the gradient direction is consistent and decelerating them when the gradient direction changes frequently. The adaptive learning rate term adjusts the learning rate based on the accumulation of the squares of the parameter gradients, using a smaller learning rate for parameters with large gradient changes and a larger learning rate for parameters with small gradient changes, thus achieving adaptive updates for different parameters. During training, feature data from the training sample set is input into the model, and predicted values ​​are calculated through forward propagation. The loss function value is then calculated based on the difference between the predicted and actual performance values. The IME is then used to backpropagate and update the model parameters based on the loss function value. This process is iterated repeatedly until a preset number of training epochs or other stopping conditions are reached.

[0085] During model training, the model's prediction error is monitored in real time using a validation set. After a certain number of training epochs (e.g., every 10 epochs), the validation set is input into the current model, and the error between the model's predicted values ​​and actual performance values ​​on the validation set is calculated. Common error calculation methods include mean absolute error (MAE) or mean squared error (MSE). These error values ​​are recorded according to the training epochs to form an error change curve. By observing the trend of the curve, the training status of the model can be judged. For example, if the error gradually decreases with the increase of training epochs, it indicates that the model is continuously optimizing; if the error no longer decreases after a certain epoch, or even increases, it may mean that the model has overfitting or other problems.

[0086] Training is stopped when the prediction error on the validation sample set no longer decreases for five consecutive iterations. An iteration cycle here refers to the process of training the entire training sample set once; for example, each iteration cycle includes one forward and backpropagation update of all samples in the training sample set. If the error no longer decreases for five consecutive iterations (i.e., the error in iterations n+1 to n+5 is greater than or equal to the error in iteration n), the model is considered to have reached a relatively stable state. Continuing training at this point may not significantly improve model performance and could even lead to overfitting. After training stops, the model's performance is evaluated using a test sample set. The test sample set is input into the trained model, and evaluation metrics such as prediction error and prediction accuracy on the test sample set are calculated. If these metrics meet preset standards, such as the prediction error being below a certain threshold, the trained multi-dimensional prediction model is output as the final model used for actual prediction.

[0087] The data on the performance influencing factors of the coating agent to be predicted are input into the trained multi-dimensional prediction model, and the predicted values ​​of each performance are obtained through forward propagation. The data to be predicted needs to undergo the same preprocessing and feature fusion steps as the training data to ensure that its feature dimensions and data distribution are consistent with the training sample set. The data first enters the input layer of the model, where each neuron receives the corresponding feature value and passes it to the hidden layer. In the hidden layer, the data is processed through multiple residual blocks. Each residual block extracts local features through a convolutional layer, adjusts the data distribution through a batch normalization layer, and then introduces a nonlinear transformation through an activation function. Skip connections between adjacent residual blocks fuse the previous features with the current processing result, enhancing the feature reuse capability. The feature vector after being processed by the hidden layer is passed to the output layer. The fully connected layer of the output layer converts the feature vector into a vector related to performance prediction, and then converts it into a probability distribution through a softmax activation function. Finally, the predicted values ​​of melting performance, heat preservation performance, and crystallization performance are determined based on the probability distribution, completing the entire prediction process.

[0088] Example 5: When generating a blast furnace covering agent performance evaluation report, the process begins with comparing the predicted values ​​of melting performance. Melting performance is a crucial indicator of the melting characteristics of the blast furnace covering agent under high-temperature conditions. The preset melting performance threshold range is determined based on the actual needs of blast furnace production and historical data, covering reasonable ranges for key indicators such as melting start temperature, melting completion temperature, and melting rate. The predicted melting performance values ​​obtained through a multi-dimensional prediction model are compared item by item with this threshold range. For example, if the predicted melting start temperature is 1300℃, and the lower limit of the threshold range is 1250℃ and the upper limit is 1350℃, then the predicted value is within the threshold range, and the indicator is deemed qualified in terms of melting performance. If the predicted melting rate is 5℃ / min, and the upper limit of the threshold range is 4℃ / min, then the predicted value exceeds the threshold range and is deemed unqualified. This out-of-tolerance item is marked as "melting rate too fast" in the report. All predicted values ​​related to melting performance are compared in this manner to determine the qualification or non-qualification of each indicator.

[0089] Next, the predicted values ​​of heat preservation performance are compared. Heat preservation performance directly affects the temperature maintenance effect of molten iron in the blast furnace. The preset threshold range for heat preservation performance includes standard ranges for indicators such as the fluctuation range of heat preservation temperature, the ability to maintain heat preservation duration, and the rate of heat loss. The predicted values ​​of heat preservation performance are compared with these thresholds to check whether each indicator is within a reasonable range. For example, if the predicted temperature fluctuation range is ±3℃, while the threshold range is ±5℃, then the indicator is qualified; if the predicted heat loss rate is 20kJ / h, while the upper limit of the threshold is 15kJ / h, then the indicator is unqualified, and the out-of-tolerance item is marked as "excessive heat loss rate". During the comparison process, the deviation between the predicted value and the threshold for each indicator should be recorded in detail to provide a basis for subsequent cause analysis.

[0090] Then, the predicted crystallization performance values ​​are compared. Crystallization performance is related to the structural stability of the covering agent after solidification. The preset crystallization performance threshold range involves the allowable range of indicators such as crystallization initiation temperature, crystallization rate, and crystal distribution uniformity. The predicted crystallization performance values ​​are checked against the threshold range one by one to determine the pass / fail status of each indicator. For example, if the predicted crystallization initiation temperature is 1100℃ and the threshold range is 1080℃-1120℃, then the indicator is qualified; if the predicted crystal distribution uniformity is 80%, but the lower limit of the threshold is 85%, then the indicator is unqualified, and the out-of-tolerance item is marked as "insufficient crystal distribution uniformity". For each crystallization performance indicator, it is necessary to determine whether it meets the standard to ensure the comprehensiveness of the evaluation.

[0091] After comparing various performance parameters, summarize the results of melting performance, heat preservation performance, and crystallization performance. Compile a list of qualified indicators, noting the predicted value and corresponding threshold range for each qualified indicator; compile a list of unqualified indicators and their deviations, recording the specific values ​​and directions of deviations (e.g., above the upper limit or below the lower limit). Subsequently, analyze the reasons for the deviations, considering influencing factors such as raw material characteristics, process control, and service environment, to explore possible factors leading to performance deviations. For example, for the deviation of "excessively fast melting rate," analyze whether the molar ratio of calcium oxide to silicon dioxide in the raw material is too high, or whether the heating rate in the melting temperature is too fast; for "excessively high heat loss rate," check whether the high-temperature holding time in the heat preservation time parameter is insufficient, or whether the specific heat capacity in the raw material's physical properties is too low; for "insufficient crystal distribution uniformity," check whether the slow cooling stage rate in the cooling rate parameter is unreasonable, or whether the particle size distribution uniformity in the raw material particle size distribution is poor. This analysis clarifies the relationship between each deviation item and influencing factor, providing direction for subsequent adjustments to the covering agent formulation or optimization of the process.

[0092] Finally, the qualified and unqualified items, along with the analysis of the reasons for exceeding tolerances, are integrated to form a complete performance evaluation report for the blast furnace covering agent. The report should be clearly structured, first outlining the overall assessment of each performance metric, then detailing the qualified and unqualified items, and finally explaining the analysis process and conclusions regarding the reasons for exceeding tolerances. The language of the report should be concise and clear, avoiding excessive use of technical jargon that could hinder comprehension, ensuring that relevant technical personnel can quickly grasp the performance status and existing problems of the covering agent, providing a reference for its application in blast furnace production.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-driven method for predicting the performance of blast furnace covering agents, characterized in that, Includes the following steps: Acquire raw material characteristic data, process control parameter data, and service environment parameter data for blast furnace covering agents. Raw material characteristic data includes raw material particle size distribution parameters, raw material mineral composition ratio, and raw material physical property parameters. Process control parameter data includes smelting temperature parameters, holding time parameters, and cooling rate parameters. Service environment parameter data includes molten iron temperature parameters, slag layer thickness parameters, and ambient humidity parameters. The raw material characteristic data, process control parameter data, and service environment parameter data of blast furnace covering agent were preprocessed and feature fused to obtain a dataset of covering agent performance influencing factors. The dataset of factors affecting the performance of the covering agent is input into the preset multi-dimensional prediction model, and the predicted values ​​of the melting performance, heat preservation performance and crystallization performance of the blast furnace covering agent are obtained through model calculation. The predicted values ​​of melting performance, heat preservation performance, and crystallization performance are compared with the preset performance threshold ranges, and a blast furnace covering agent performance evaluation report is generated based on the comparison results.

2. The data-driven blast furnace covering agent performance prediction method according to claim 1, characterized in that, The raw material particle size distribution parameters specifically include the uniformity of raw material particle size distribution, the maximum particle diameter, and the minimum particle diameter. The raw material mineral composition ratio specifically includes the molar ratio of calcium oxide to silicon dioxide, the mass fraction of magnesium oxide, and the volume fraction of aluminum oxide. The raw material physical property parameters specifically include the raw material bulk density, the raw material porosity, and the raw material specific heat capacity. The smelting temperature parameters specifically include the peak melting temperature, the heating rate, and the isothermal holding temperature. The holding time parameters specifically include the high-temperature holding time, the medium-temperature holding time, and the low-temperature holding time. The cooling rate parameters specifically include the rapid cooling stage rate, the slow cooling stage rate, and the isothermal transformation stage duration. The molten iron temperature parameters specifically include the initial molten iron temperature, the molten iron cooling rate, and the target molten iron temperature. The slag layer thickness parameters specifically include the initial slag layer thickness, the slag layer solidification thickness, and the slag layer uniformity coefficient. The environmental humidity parameters specifically include the environmental relative humidity, the water vapor partial pressure, and the humidity change rate.

3. The data-driven blast furnace covering agent performance prediction method according to claim 2, characterized in that, The specific steps for preprocessing and feature fusion of raw material characteristic data for blast furnace covering agents are as follows: Outliers in the particle size distribution parameters of raw materials are judged and corrected using the Laida criterion; missing values ​​in the mineral composition ratio of raw materials are subjected to multiple interpolation; and duplicate data in the physical property parameters of raw materials are deduplicated based on timestamps. The particle size distribution parameters, mineral composition ratios, and physical property parameters of the pretreated raw materials were standardized and converted to obtain dimensionless basic data of raw material characteristics. Based on dimensionless raw material characteristic data, the influence weight of each parameter on the performance of the covering agent is calculated by grey relational analysis. The parameters with influence weights greater than a preset threshold are combined to obtain a raw material characteristic fusion feature set.

4. The data-driven blast furnace covering agent performance prediction method according to claim 3, characterized in that, The specific steps for preprocessing and feature fusion of process control parameter data are as follows: Time series alignment was performed on melting temperature parameters, holding time parameters, and cooling rate parameters to unify parameters with different acquisition frequencies into time series data with the same time interval; Sliding window sampling is performed on the time-series process control parameter data, and the maximum, minimum, average and standard deviation within each window are extracted as statistical features; By concatenating the statistical features with the original values ​​of the process control parameters, and reducing the data dimensionality through principal component analysis, a fusion feature set of process control is obtained.

5. The data-driven blast furnace covering agent performance prediction method according to claim 4, characterized in that, The specific steps for preprocessing and feature fusion of service environment parameter data are as follows: The dimensional consistency of parameters such as molten iron temperature, slag layer thickness, and ambient humidity was verified, and parameters in different units were converted into standard data under the International System of Units (SI). Trend terms are extracted from standard data, and the linear trend in the data is eliminated by the difference method to obtain stable service environment data. Based on stationary service environment data, wavelet transform is used to extract fluctuation features in different frequency bands. The fluctuation features are then fused with the stationary data to obtain a fused feature set of the service environment.

6. The data-driven blast furnace covering agent performance prediction method according to claim 5, characterized in that, The specific steps to obtain the dataset of factors affecting the performance of covering agents are as follows: Align the sample dimensions of the raw material characteristics fusion feature set, process control fusion feature set, and service environment fusion feature set to ensure that the number of samples in each feature set corresponds one-to-one with the timestamp. The three aligned feature sets are horizontally concatenated to form an initial feature matrix containing multi-dimensional information on raw materials, processes, and environment. Feature selection was performed on the initial feature matrix, and redundant features were removed using a recursive feature elimination method. Feature variables that are strongly correlated with the performance of the covering agent were retained, resulting in a dataset of covering agent performance influencing factors.

7. The data-driven blast furnace covering agent performance prediction method according to claim 6, characterized in that, The steps for constructing the pre-defined multi-dimensional prediction model are as follows: A deep learning network architecture consisting of an input layer, a hidden layer, and an output layer was constructed. The number of neurons in the input layer was consistent with the feature dimensions of the coating agent performance influencing factor dataset. The output layer contained three prediction branches: melting performance, heat preservation performance, and crystallization performance. The hidden layer adopts a residual network structure. Each residual block contains a convolutional layer, a batch normalization layer, and an activation function layer. Feature reuse is achieved between adjacent residual blocks through skip connections. The output layer uses a combination of a fully connected layer and a softmax activation function to output the probability distributions of the predicted values ​​of melting performance, heat preservation performance, and crystallization performance, respectively.

8. The data-driven blast furnace covering agent performance prediction method according to claim 7, characterized in that, The specific steps for obtaining the predicted values ​​of melting performance, heat preservation performance, and crystallization performance of the blast furnace covering agent through model calculation are as follows: The dataset of influencing factors on the performance of covering agents was divided into training sample set, validation sample set and test sample set; The multidimensional prediction model is trained using a training sample set, and the model parameters are optimized using an adaptive moment estimation algorithm. The prediction error of the model is monitored in real time using a validation sample set. When the prediction error of the validation sample set no longer decreases for 5 consecutive iterations, stop model training, use the test sample set to evaluate the model's performance, and if the evaluation metric meets the preset standard, output the trained multi-dimensional prediction model. The data of the influencing factors of the coating agent performance to be predicted are input into the trained multi-dimensional prediction model, and the predicted values ​​of melting performance, heat preservation performance and crystallization performance are obtained by forward propagation calculation.

9. The data-driven blast furnace covering agent performance prediction method according to claim 8, characterized in that, The specific steps for generating a blast furnace covering agent performance evaluation report are as follows: The predicted melt performance value is compared with the preset melt performance threshold range. If the predicted value is within the threshold range, the melt performance is deemed qualified; otherwise, it is deemed unqualified and marked as out of tolerance. The predicted value of thermal insulation performance is compared with the preset thermal insulation performance threshold range. If the predicted value is within the threshold range, the thermal insulation performance is deemed to be qualified; otherwise, it is deemed to be unqualified and marked as out of tolerance. The predicted crystallization performance value is compared with the preset crystallization performance threshold range. If the predicted value is within the threshold range, the crystallization performance is deemed qualified; otherwise, it is deemed unqualified and marked as out of tolerance. Summarize the judgment results of melting performance, heat preservation performance, and crystallization performance, and generate a blast furnace covering agent performance evaluation report that includes qualified items, unqualified items, and analysis of the reasons for deviations.

10. A device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement the data-driven blast furnace covering agent performance prediction method as described in any one of claims 1 to 9.

Images