Multi-source-domain asynchronous data-oriented converter steelmaking process prediction method and system

Abstract

The invention relates to the technical field of converter steelmaking, particularly discloses a converter steelmaking process prediction method and system for multi-source-domain asynchronous data, and aims to solve the problems that converter steelmaking multi-source-domain data is lagged in information, asynchronous data feature extraction and fusion are difficult, and converter steelmaking working conditions are diversified and complicated. According to the method and system, a converter steelmaking working condition process prediction model is provided, the model uses a feature pyramid network to construct a multi-scale feature extraction network, and the multi-source domain asynchronous data feature learning ability of the model is improved through fusion; according to the model, a Transform module is adopted, a self-attention mechanism, a multi-head attention mechanism and residual connection are utilized to capture a dependency relationship in multi-scale fusion features, probability values of different working conditions are output, and an abnormal category corresponding to the maximum probability value is output as a working condition prediction result. Experiments prove that the method has relatively high accuracy on identification of five converter steelmaking working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of converter steelmaking, and in particular to a converter steelmaking process prediction method and system oriented to multi-source domain asynchronous data. Background Art

[0002] Converter steelmaking is a steelmaking process with high temperature, high pressure, multiple influencing factors and strong physical and chemical reactions. The main task of the modern converter steelmaking process is to smelt blast furnace molten iron according to a specific proportion of raw materials (scrap steel, molten iron, slag, etc.) into molten steel with temperature and composition that meet the requirements of various steel types. The common converter steelmaking production flow chart is as follows: Figure 1 As shown, the process mainly goes through the following stages: charging, oxygen blowing, endpoint control, deoxidation and steelmaking, slag splashing and furnace protection, and slag pouring.

[0003] As an important reaction vessel in the entire molten steel smelting process, the converter is very likely to cause serious splashing and slag overflow accidents if the smelting process is not properly operated, causing unnecessary casualties. At the same time, during the converter steelmaking process, due to the violent reaction between oxygen and molten iron, a large amount of gas will be produced, accompanied by a strong chemical reaction, resulting in the splashing of molten metal and slag. This phenomenon not only affects the normal production process, but also brings a series of negative effects to the entire steelmaking process, such as endangering the safety of workers, reducing production efficiency, reducing product quality, and forcing production costs to increase.

[0004] In modern process flow, converter endpoint control is the main method, supplemented by converter process control. The converter endpoint control technologies include manual experience control, static control and dynamic control. Manual experience control refers to the guidance and control of steelmaking on-site production by steel workers based on their own production operation experience and steelmaking knowledge summarized over the years. However, manual experience control is subject to the different operators and the uneven control levels of each operator. In addition, the actual production process is accompanied by uncertain factors such as complex steelmaking conditions, high-quality steel composition, various steelmaking raw materials, and large fluctuations in furnace composition. It is difficult to ensure the stability of control and a high hit rate of the converter endpoint. The static control model requires manual experience to assist in adjustment, and lacks learning of the entire converter smelting process, and only considers the initial and final states of smelting. The dynamic control model is mainly based on detection methods, mainly dynamic control technologies based on auxiliary gun detection, flue gas analysis, and furnace mouth flame.

[0005] Problems faced at this stage: The actual generation time of flue gas data (CO, CO2) is inconsistent with the time of detection through the pipeline, resulting in information lag. When analyzing the data, it was found that CO and CO2 had different mutation trends a few seconds before splashing, and the two were out of place in time and space relative to the detection of splashing. Existing prediction methods rarely solve the problems of information lag in multi-source domain data, the difficulty of asynchronous data feature extraction and fusion, and the diversification and complexity of converter steelmaking conditions.

[0006] The on-site smelting process of converter steelmaking will be affected by various factors. The thresholds of CO and CO2 or the trend of dynamic changes are difficult to determine, that is, it is difficult to determine the precise control amount of CO and CO2. From the data analysis, it can be seen that there is indeed such a law in the CO and CO2 flue gas change curves, and this is used as the in-depth point of subsequent work. At the same time, the experiment conducts such a prediction, that is, to predict the future situation based on multiple factors. If only the threshold of CO or CO2 cannot be judged, it is necessary to combine the information detected by other on-site detectors for comprehensive analysis and prediction. At the same time, the control of the oxygen gun position (i.e., the height of the oxygen gun) affects the splashing and overflowing slag: Since oxygen blowing will generate high pressure, if the oxygen gun position is too high, it will cause serious foaming of the slag in the converter, and the generated slag will overflow from the furnace mouth; if the oxygen gun position is too low, there will be too little foam in the converter, resulting in back drying.

[0007] In general, the existing prediction methods cannot solve the problems of information lag in multi-source domain data, difficulty in extracting and fusing asynchronous data features, and the diversification and complexity of converter steelmaking conditions. Summary of the invention

[0008] The present invention provides a converter steelmaking process prediction method and system for multi-source domain asynchronous data, and solves the technical problem that in the converter steelmaking process, multi-source domain data information lags, asynchronous data feature extraction and fusion are difficult, and the diverse steelmaking conditions of the converter cannot be accurately predicted.

[0009] In order to solve the above technical problems, the present invention provides a converter steelmaking process prediction method for multi-source domain asynchronous data, comprising the steps of:

[0010] S1. constructing an original data set consisting of asynchronous data from multiple source domains in a converter steelmaking process, and performing equalization preprocessing on the original data set so that the amount of data of each identification category in the original data set remains consistent; normalizing the original data set subjected to equalization preprocessing to obtain a normalized data set; and dividing the normalized data set into a training set and a test set;

[0011] S2. Construct a converter steelmaking process prediction model, which includes a feature pyramid and a Transformer module. When the data samples of the training set or the test set are input into the feature pyramid, the feature pyramid combines sample features from different levels through bottom-up feature extraction, top-down feature fusion path and lateral connection to obtain multi-scale fusion features; the Transformer module uses self-attention mechanism, multi-head attention mechanism and residual connection to capture the dependency relationship in the multi-scale fusion features, obtains an output feature map, and calculates and outputs the probability values ​​of different working conditions through activation function, and outputs the working condition prediction result with the maximum probability value;

[0012] S3, using the training set and the test set to train and test the converter steelmaking process prediction model;

[0013] S4. Use the trained and tested converter steelmaking process prediction model to predict any input data sample to obtain a process prediction result.

[0014] Furthermore, the calculation steps of the feature pyramid are as follows:

[0015] (1) Time-series multi-source domain asynchronous data with time windows as the unit quantity is used as input;

[0016] (2) Extract features at different levels from asynchronous data in time series multi-source domains;

[0017] (3) Starting from the deepest layer of feature extraction, use point-by-point convolution to generate the number of channels;

[0018] (4) Use transposed convolution to increase the spatial resolution of features and double the spatial dimension so that the spatial scale of macro features matches that of micro features;

[0019] (5) Merge features at different levels, integrate the detail information of low-level features with the semantic information of high-level features, and use point-by-point convolution to compress and integrate the number of merged channels to obtain a multi-scale feature pyramid;

[0020] (6) Repeat (4) and (5) until the top-level features are extracted;

[0021] (7) Output the fusion results of feature vectors of different scales.

[0022] Furthermore, by setting time windows of different sizes for comparative experiments, the time window with the best model evaluation index is used as the time window of the converter steelmaking condition process prediction model participating in the prediction of the actual converter steelmaking process in step S4.

[0023] Furthermore, the Transformer module includes an encoder and a decoder;

[0024] The processing flow of the encoder is:

[0025] The input sequence first passes through the embedding layer and position encoding to obtain an input sequence embedding representation with position information;

[0026] The input sequence embedding means inputting N stacked encoder layers for hierarchical feature extraction. In each encoder layer, the relationship between different positions is first extracted through a multi-head attention mechanism, and then residual connections and layer normalization are performed. Then, the information is passed to the position-aware feedforward neural network, and residual connections and layer normalization are performed again.

[0027] Finally, through the stacking of N encoder layers, the high-level feature representation of the input sequence is obtained and input to the decoder;

[0028] The processing flow of the decoder is:

[0029] The target sequence first passes through the embedding layer and position encoding to obtain an embedded representation of the target sequence with position information;

[0030] The target sequence embedding means inputting N stacked decoder layers for hierarchical feature extraction. In each decoder layer, the input data first passes through the masked multi-head attention mechanism to extract the relationship between different positions before the current prediction, then performs residual connection and normalization, and then passes it together with the encoder output to the multi-head attention mechanism for feature extraction, and then performs residual connection and normalization;

[0031] Through the stacked N decoder layers, the final output features are obtained;

[0032] The final output features are passed through a linear layer and Softmax to map the features to the probability distribution of the vocabulary, and the most likely prediction result is selected as the output of the decoder.

[0033] Furthermore, the encoder is provided with four encoder layers, and the decoder is provided with four decoder layers.

[0034] Furthermore, the cross entropy loss function is used as the loss function for training the converter steelmaking process prediction model.

[0035] Furthermore, during the training process, an Adam optimizer is used to update the parameters of the converter steelmaking process prediction model.

[0036] Furthermore, based on the grid search method, the optimal hyperparameter set of the converter steelmaking process prediction model is obtained.

[0037] Furthermore, the data collected in the original data set are flue gas data, oxygen flow, oxygen pressure, oxygen lance height, rope tension, sonar intensity and converter mouth flame recognition results at the corresponding time in the converter smelting process; the operating condition categories include splash warning, slag overflow, slight splash, severe splash and dry back, splash warning means that the system detects signs of impending molten steel splashing, slag overflow refers to the phenomenon that slag overflows from the converter mouth, slight splash refers to splashing reaching a first preset degree, severe splashing refers to splashing reaching a second preset degree deeper than the first preset degree; dry back refers to the phenomenon that high-melting-point substances are precipitated in the slag, the slag becomes thicker, and cannot effectively cover the metal liquid surface.

[0038] The present invention also provides a converter steelmaking process prediction system for multi-source domain asynchronous data, the key of which is: an intelligent agent is provided, and the intelligent agent is used to implement the converter steelmaking process prediction method for multi-source domain asynchronous data.

[0039] The present invention provides a converter steelmaking process prediction method and system for multi-source domain asynchronous data. Aiming at the problems of information lag, asynchronous data feature extraction and fusion difficulties, and diversified and complicated converter steelmaking conditions in multi-source domain data of converter steelmaking, the present invention proposes a converter steelmaking process prediction model based on a self-attention mechanism for multi-scale feature extraction. The model uses a feature pyramid network (FPN) to construct a multi-scale feature extraction network, and improves the model's learning ability for multi-source domain asynchronous data features through fusion; the model uses a Transformer module to capture the dependencies in multi-scale fusion features using a self-attention mechanism, a multi-head attention mechanism, and a residual connection, outputs probability values ​​of different conditions, and outputs the abnormal category corresponding to the maximum probability value as the condition prediction result. Experiments have confirmed that the present invention has a high accuracy rate in identifying five converter steelmaking conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a converter steelmaking production flow chart provided by the background technology of the present invention;

[0041] Figure 2 It is a flow chart of a converter steelmaking process prediction method and system for multi-source domain asynchronous data provided by an embodiment of the present invention;

[0042] Figure 3 It is a graph showing the changing rules of abnormal furnace CO, CO2 and sonar curves provided by an embodiment of the present invention;

[0043] Figure 4 is a graph showing the change trend of oxygen lance height and oxygen flow rate over time for abnormal furnaces provided by an embodiment of the present invention;

[0044] Figure 5 is a diagram of an FPN network structure provided by an embodiment of the present invention;

[0045] Figure 6 It is a structural diagram of the internal encoder and decoder of the Transformer model provided by an embodiment of the present invention;

[0046] Figure 7 is a calculation flow chart of the FPN-Transformer model provided by an embodiment of the present invention;

[0047] Figure 8 is a comparison diagram of the ablation experiment model loss and F1 score provided in an embodiment of the present invention;

[0048] Fig. 9This is a comparison diagram of confusion matrices of two models in the ablation experiment provided by an embodiment of the present invention;

[0049] Fig.10 is a graph showing the change trend of model loss and F1 score corresponding to different time windows provided by an embodiment of the present invention;

[0050] Fig.11 It is a confusion matrix diagram of a 15-second time window provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0052] The embodiment of the present invention provides a converter steelmaking process prediction method for multi-source domain asynchronous data, such as Figure 2 As shown, it includes the steps of:

[0053] S1. constructing an original data set consisting of asynchronous data from multiple source domains in a converter steelmaking process, and performing equalization preprocessing on the original data set so that the amount of data of each identification category in the original data set remains consistent; normalizing the original data set subjected to equalization preprocessing to obtain a normalized data set; and dividing the normalized data set into a training set and a test set;

[0054] S2. Construct a converter steelmaking process prediction model, which includes a feature pyramid and a Transformer module. When the data samples of the training set or the test set are input into the feature pyramid, the feature pyramid combines sample features from different levels through bottom-up feature extraction, top-down feature fusion path and lateral connection to obtain multi-scale fusion features; the Transformer module uses self-attention mechanism, multi-head attention mechanism and residual connection to capture the dependency relationship in the multi-scale fusion features, obtains an output feature map, and calculates and outputs the probability values ​​of different working conditions through activation function, and outputs the working condition prediction result with the maximum probability value;

[0055] S3, using the training set and the test set to train and test the converter steelmaking process prediction model;

[0056] S4. Use the trained and tested converter steelmaking process prediction model to predict any input data sample to obtain a process prediction result.

[0057] The data used in this example are flue gas data (CO, CO2 volume fraction, CO refers to carbon monoxide, CO2 refers to carbon dioxide), oxygen flow (gas_flow), oxygen pressure (gas_press), oxygen lance height (lance_height), rope tension (lance_force a1, lance_forcea2, lance_force b1, lance_force b2, lance_force a1, lance_force a2 respectively represent the tension of the two ropes pulling the oxygen lance A, lance_force b1, lance_force b2 respectively represent the tension of the two ropes pulling the oxygen lance B) and sonar intensity (sound_intensity) and the flame recognition results of the converter mouth at the corresponding time provided by a steel plant. The recognition results are divided into 5 categories: splash warning, slag overflow, slight splash, severe splash and back drying. The size of the data set is 878 heats, which are divided into training set and test set in a ratio of 8:2. The data of 702 heats are used for training and the data of 176 heats are used for testing. In order to improve the prediction accuracy of the model and obtain the best prediction model, data preprocessing was performed on all the heat data.

[0058] There is a serious imbalance in the data of all furnaces. After statistics, there are 617,299 normal data in the data of 878 furnaces, and the data of splash warning, slag overflow, slight splash, severe splash and return drying are 3905, 4861, 1565, 6178 and 6773 respectively. It can be seen that there is a serious imbalance in the data of all furnaces. Because the experiment needs to classify and predict the abnormal data, this example uses the synthetic minority over-sampling technique (SMOTE) to enhance the data of the abnormal category. By synthesizing new abnormal class samples, the data volume of the five abnormal categories is kept consistent to balance the data set, thereby improving the classification performance of the model. It should be noted here that since the converter steelmaking and its detection process are continuous, the prediction of the normal category should also be considered. It is just that when the model performs probability output, the prediction of the normal category is set not to be output in the final output, and only the abnormal category is output.

[0059] Through analysis, it is found that the dimensional fluctuations between all variables are large and the consistency of the data is poor, so it is necessary to eliminate the differences between the data before conducting the experiment to improve the reliability of the data. In this example, Z-Score standardization is used to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1 for standardization.

[0060] After data standardization, the model performance and training speed are significantly improved, which avoids certain features from having a greater impact on the model training process, reduces model bias, avoids the problem of gradient disappearance or explosion, and enhances model stability.

[0061] An important task in predicting converter operating conditions is to accurately predict five common operating conditions, namely splash warning, slag overflow, slight splash, severe splash and back drying, in advance, and to issue early warnings at the same time so that real-time adjustments can be made at the smelting site. Figure 3 It is a trend chart of CO, CO2 and sonar over time for five abnormal conditions of splash warning, slag overflow, slight splash and severe splash during the converter process. Figure 4 It is a trend chart of the oxygen lance height and oxygen flow rate over time when five abnormal conditions occur during the converter process: splash warning, slag overflow, slight splash, and severe splash. Figure 3 and Figure 4 It can be found that abnormal operating conditions mainly occur in the middle stage of blowing.

[0062] In the converter steelmaking process, splash warning means that the system detects signs of impending splashing of molten steel and issues a warning signal so that operators can take measures to prevent splashing. Before the splash warning occurs, the CO content increases rapidly, and the CO2 content decreases slowly in a step-like manner, with the CO change range being higher than that of CO2; the sonar signal decreases, and the decrease range is slightly larger than before and after; the oxygen flow and oxygen pressure increase rapidly to balance the changes in the furnace; the oxygen gun rope tension changes in a large range, and the overall large drop reaches a certain value and then increases slightly.

[0063] In the converter steelmaking process, slag overflow refers to the phenomenon that slag overflows from the converter mouth, which is usually caused by excessive slag volume or improper operation in the furnace. Slag overflow will affect the normal operation of the converter and may even cause damage to equipment and personnel. When slag overflow occurs, the CO concentration will drop suddenly, and the CO2 concentration will increase in a step-like manner, depending on the reaction situation in the furnace and the chemical composition of the slag; the sonar signal will change significantly, with violent fluctuations or abnormal signals, mainly showing a stable downward curve, because the overflow of slag will change the path and intensity of sonar propagation; the oxygen flow rate will be stable first and then drop sharply to avoid further overflow; the change in oxygen pressure is similar to the change in oxygen flow rate, and the pressure value is reduced to reduce the reaction intensity; the oxygen lance height is raised to reduce the impact force of oxygen injection on the surface of the molten pool, reduce the turbulence in the molten pool, and reduce slag overflow; the oxygen lance rope tension fluctuates to adapt to the rapid adjustment of the oxygen lance height. After the slag overflow ends, the CO and CO2 concentrations gradually return to normal levels, and the curve tends to be stable. The sonar signal will gradually return to stability, and the oxygen flow and pressure will return to normal levels. The oxygen lance height returns to the normal working position.

[0064] The splashing condition of converter steelmaking is divided into two types according to the splashing degree and the intensity of the reaction, namely, slight splashing and severe splashing. Slight splashing refers to splashing with a splashing degree reaching a first preset degree, and severe splashing refers to splashing with a splashing degree reaching a second preset degree deeper than the first preset degree. Under normal operating conditions, the CO and CO2 concentration curves in the converter are relatively stable. When the oxidation reaction is carried out in the molten steel, the concentration of CO will fluctuate slightly. When the temperature inside the converter rises and the molten steel and slag react violently, the CO concentration rises sharply because the carbon oxidation reaction intensifies at high temperature. As the concentration of CO rises sharply, the concentration of CO2 will also change accordingly, usually rising first and then falling, which is manifested as a sharp fluctuation of the curve. After the splashing condition disappears, the concentrations of CO and CO2 gradually return to a stable state, and the curve tends to be flat.

[0065] Based on the occurrence of severe splashing, the CO and CO2 contents change in opposite trends; the sonar signal will fluctuate abnormally, reflecting the bubble bursting and splashing phenomenon in the furnace. This fluctuation is usually manifested as a sudden increase in the signal or a change in frequency. Sharp and harsh sonar signals can be detected. The sonar curve will show an upward trend, but the corresponding actual plotted curve is a downward trend. It can be seen that the measured sonar information has a time lag problem; the oxygen flow rate suddenly increases, and the oxygen pressure also fluctuates. The oxygen gun height is appropriately lowered to cope with the intense reaction process in the furnace; the oxygen gun rope tension shows an unstable increase and decrease. After the splashing ends, the oxygen flow rate and pressure will return to normal levels, and the sonar signal will gradually return to a stable state.

[0066] By analyzing a large amount of abnormal heat data, it is found that if severe splashing occurs, it must be accompanied by splashing warnings, slag overflow and minor splashing in order. Therefore, it is necessary to accurately control the converter steelmaking process to ensure smooth smelting and effectively suppress severe splashing conditions. It is imperative to make high-precision converter condition predictions when splashing warnings occur to ensure subsequent adaptive dynamic adjustments and the production of high-quality molten steel in the converter.

[0067] In the converter steelmaking process, dryout refers to the phenomenon that some high-melting-point substances in the slag are precipitated due to factors such as oxygen pressure, too low gun position and low TFe (total iron) content, causing the slag to thicken and fail to effectively cover the metal liquid surface. When dryout occurs in the furnace, the CO concentration rises, the carbon oxidation reaction weakens, the CO2 concentration decreases, and the overall gas generation decreases, resulting in a curve decline; the sonar signal will change significantly, with violent fluctuations or abnormal signals, mainly showing a "U"-shaped curve that first drops and then rises, because the solidified slag and molten steel affect the sonar propagation; the oxygen flow rate fluctuates and rises in an attempt to restore the furnace temperature; the oxygen pressure rises in a step-by-step manner to cope with the temperature changes in the furnace; the oxygen gun rope tension fluctuates in a pulse shape, and the tension has dropped significantly 2 seconds before the dryout occurs, indicating the occurrence of the dryout condition. After the dryout disappears, the temperature rises, the CO and CO2 concentrations will gradually return to normal levels, and the curve tends to be stable. The sonar signal will gradually return to stability, reflecting the normal state of the liquid steel and slag in the furnace. Oxygen flow, pressure and lance rope tension return to normal levels.

[0068] After data analysis, it is found that the dependencies between multi-source domain data in converter steelmaking are very complex. It is impossible to achieve efficient prediction of converter conditions based on a single data feature, and the actual situation in the furnace cannot be fully reflected due to incomplete feature extraction. It is necessary to explore the time correlation and multi-scale features between different data attributes collected from multiple detection devices, so that the model has the ability to extract multi-scale features, so as to simultaneously process and utilize information at different time scales, thereby improving the accuracy and robustness of the prediction.

[0069] The importance of characteristics at different time scales for accurate prediction and classification of operating conditions is reflected in the following two points:

[0070] (1) Short time scale: It can capture rapidly changing furnace conditions, such as instantaneous splashing or sharp changes in gas content, which is particularly important for short-term warning and response.

[0071] (2) Long time scale: It can reveal slower trends and cyclical changes, and help identify long-term anomalies or potential problems.

[0072] In order to better predict the working conditions of converter steelmaking, this example proposes to use the feature pyramid network to extract and fuse the features of asynchronous data from multiple source domains in combination with the above analysis. Feature Pyramid Network (FPN) is a network structure used for computer vision tasks such as target detection and semantic segmentation. The core idea of ​​FPN is to build a feature pyramid from low resolution to high resolution so that the network can capture rich semantic information at different scales. Through top-down paths and lateral connections, multi-scale feature fusion is achieved by combining features from different levels.

[0073] FPN can capture feature information of different scales and has important implications for asynchronous data processing in time series and multi-source domains. Figure 5 As shown. The specific calculation steps of the FPN structure are as follows:

[0074] (1) Input: Time-series multi-source domain asynchronous data with time windows as the unit quantity is used as input;

[0075] (2) Multi-scale feature extraction: Extract features from different levels of asynchronous data in time series multi-source domains. When using CNN for feature extraction, a deeper level is usually selected as the starting point for feature extraction;

[0076] (3) Top-level features: Starting from the deepest layer of feature extraction, point-by-point convolution is used to generate fewer channels and reduce the amount of model calculation;

[0077] (4) Upsample: Use transposed convolution (nn.ConvTranspose2d) to increase the spatial resolution of features, such as upsampling the input features from 14×14 to 28×28, while doubling the spatial dimension so that the spatial scale of macro features matches that of micro features;

[0078] (5) Fusion: Merge features at different levels, fuse the detail information of low-level features with the semantic information of high-level features, and use point-by-point convolution to compress and integrate the number of merged channels to obtain a multi-scale feature pyramid;

[0079] (6) Repeat: Repeat (4) and (5) until the top-level features are extracted;

[0080] (7) Output: The final feature pyramid outputs the fusion result of feature vectors of different scales (output).

[0081] When the FPN network implements multi-scale feature fusion, FPN extracts time series data into feature maps of multiple scales through a series of convolutional layers and pooling operations. Each feature map represents information details at different time scales. In the top-down path of FPN, higher-level feature maps are gradually upsampled and laterally connected with low-level feature maps. This multi-scale feature fusion method not only retains the detailed information of low-level features, but also combines the abstract information of high-level features to form a rich multi-scale feature representation. Finally, these fused multi-scale features are input into the Transformer model to further capture the complex dependencies in the time series data. In this way, the multi-scale feature fusion mechanism provided by FPN effectively improves the feature representation capability of time series data, and the powerful sequence modeling capability of Transformer further enhances the prediction performance of the model.

[0082] For model training, the input parameters of the model (based on the start time of the furnace) are: the amount of data (data features are: CO, CO2, lance_height, gas_flow, gas_press, lance_force a1, lance_force a2, lance_force b1, lance_force b2, sound_intensity, and data label is: type) of different time window sizes (taking the 15-second window size as an example). Set the window sliding parameter Δt of the same size as the window, that is, the output parameter of the model is the type fitting result of the 30th second. Subsequently, the model input is performed in the same way for different time windows. When processing the parameter settings of different models, RNN, LSTM and Transformer are tested with different time window sizes, and the 15-second window is selected as a comparative experiment with the same parameter input and output method. For model prediction, the data is divided into proportions before the experiment, and the trained model file is called. The input of the model is the same as that during training, and the data features and label values ​​of the first 15 seconds of the furnace must be provided to drive the model.

[0083] The Transformer model is a deep learning architecture specifically designed to process sequence data. Figure 6 As shown, it includes an encoder and a decoder. The encoder and decoder combine technologies such as self-attention mechanism, multi-head attention mechanism and residual connection.

[0084] The self-attention mechanism is a technique that allows the model to pay attention to the interdependencies between different positions in the sequence when processing sequence data. It is also called the internal attention mechanism. By calculating the correlation between each element in the sequence and all other elements, a weighted vector is generated based on this, which reflects the representation of the element in the current context. In this way, the model can automatically learn the dependencies between different positions and capture richer contextual information.

[0085] The main implementation steps of the self-attention mechanism are:

[0086] (1) Query-Key-Value decomposition: For each element in the sequence, it is converted into a query vector (Query), a key vector (Key), and a value vector (Value) through three different linear transformations (usually weight matrix multiplication). Query represents the degree to which the current position is paying attention to other positions; Key represents the information of other positions, which is used to match the Query; Value contains the actual content information of other positions, which is used to aggregate information according to the attention weight.

[0087] (2) Attention score calculation: For each position i in the sequence, calculate the attention score between it and all other positions j. This is usually done by taking the dot product of the query and the key and then scaling it. The scaling factor is used to stabilize the gradient of the softmax function.

[0088] (3) Attention weight allocation: The calculated attention score is converted into a probability distribution through the softmax function as the attention weight of each position. This means that for position i, each position j has a weight indicating how much i should pay attention to j.

[0089] (4) Weighted summation to obtain output: Finally, the Value vectors of all positions are weighted summed according to the calculated attention weights to obtain the context representation of the current position.

[0090] The multi-head attention mechanism is based on the self-attention mechanism, and both the encoder and decoder contain multiple multi-head attention mechanisms. Unlike the self-attention mechanism, it uses multiple independent attention heads, each of which can be calculated independently in different subspaces. Multiple independent attention mechanisms are run in parallel to obtain the attention distribution of different subspaces of the input sequence, thereby more comprehensively capturing the potential multiple semantic associations in the sequence.

[0091] The converter steelmaking process prediction model proposed in the present invention combines the Transformer model idea and model idea, also known as the FPN-Transformer model, and the Transformer part is mainly divided into two parts: two symmetrical parts, the encoder (Encoder) and the decoder (Decoder), each of which contains multiple layers (Layers). Generally, the deeper the network layer, the larger the field of view of the feature map that can be obtained. The shallow convolution focuses on surface features, and the deep network focuses on essential features, but both are meaningful. The fusion of all feature information is achieved by splicing features to ensure the best prediction effect.

[0092] like Figure 6 As shown, the processing flow of the Transformer encoder is:

[0093] 1. The input sequence first passes through the embedding layer and position encoding to obtain an input sequence embedding representation with position information;

[0094] 2. Input sequence embedding means inputting N stacked encoder layers for hierarchical feature extraction. In each encoder layer, the relationship between different positions is first extracted through a multi-head attention mechanism, and then residual connection and layer normalization are performed, and then passed to the position-aware feedforward neural network, and residual connection and layer normalization are performed again;

[0095] 3. Finally, through the stacking of N encoder layers, the high-level feature representation of the input sequence is obtained and input to the decoder.

[0096] The main task of the encoder is to convert the input sequence into a series of hidden representations. It extracts multi-source domain asynchronous data features by capturing context information, and realizes the abstraction and extraction of low-level to high-level features by setting parameters such as convolution kernel size and convolution step size.

[0097] like Figure 6 As shown, the processing flow of the Transformer decoder is:

[0098] 1. The target sequence first passes through the embedding layer and position encoding to obtain an embedded representation of the target sequence with position information;

[0099] 2. The target sequence embedding means inputting N stacked decoder layers for hierarchical feature extraction. In each decoder layer, the input data first passes through the masked multi-head attention mechanism to extract the relationship between different positions before the current prediction, then performs residual connection and normalization, and then passes it together with the encoder output to the multi-head attention mechanism (by using the encoder output to focus on the relevant information of the source sequence) for feature extraction, and then performs residual connection and normalization;

[0100] 3. The final output features are obtained by stacking N decoder layers;

[0101] 4. The final output features are passed through a linear layer and Softmax to map the features to the probability distribution of the vocabulary, and the most likely prediction result is selected as the output of the decoder.

[0102] The target sequence of the decoder is Figure 6 The output of the decoder in the above example actually refers to the feature representation of the encoder's input sequence after being processed by the "input embedding layer" and "position encoding layer". This representation contains the content information and position information of the input sequence, and is the initial feature map of the input sequence. These feature maps are passed to the encoder's multi-layer structure in turn, and are further processed by multi-head attention and feedforward neural networks, and finally the encoder's final feature map is output at the top layer. This output is passed to the decoder for subsequent decoding and output generation.

[0103] The main task of the decoder is to restore the original data features and position information to solve the problem of multi-source domain data position restoration and high-precision positioning. The decoder combines the output of the encoding part after FPN network feature extraction with the upsampling result so that it can match the position information of the decoder, and introduces the retained original position information to improve the accuracy of data position positioning.

[0104] In this example, N is preferably 4.

[0105] Finally, the probability values ​​of different working conditions are calculated and output through the ReLU activation function, and the working condition prediction result with the maximum probability value is output.

[0106] This example combines the feature pyramid network (FPN) with the Transformer to process asynchronous data in time series multi-source domains in order to combine the advantages of both: FPN's multi-scale feature extraction and Transformer's powerful sequence modeling capabilities. The fused multi-scale features are input into the Transformer model, and its powerful sequence modeling capabilities are used to further capture the complex dependencies in time series data.

[0107] Take one epoch as an example to simulate the calculation process of the FPN-Transformer model. The calculation flow chart is as follows Figure 7As shown. The input tensor is a 15×10 shape information to track the changes in each layer. It is necessary to add a channel dimension in the first dimension and adjust the input to a four-dimensional tensor of (batch_size, 1, 15, 10). The batch_size value is 32. Through the first layer of convolution (3×3conv, p=1, p is padding, indicating the padding size) and the maximum pooling layer (maxpool, 2×2) and batch normalization to accelerate convergence, the activation function ReLU is used to add nonlinearity. The output shape after convolution is (32, 16, 15, 10), and the output shape of this layer is (32, 16, 7, 5); through the second layer of convolution (3×3conv, p=1) and the maximum pooling layer (maxpool, 2×2) and batch normalization, the activation function ReLU is used. The output shape after convolution is (32, 32, 7, 5), and the output shape of this layer is (32, 32, 3, 2); FPN is used to fuse features of different resolutions to increase the detail information of the features and the output of the previous layer is normalized. Upsampling (Upconv) and using transposed convolution (stride(s) = 2) to restore the spatial dimension to the same shape as the output of the first layer of convolution, the output shape is (32,16,6,4), this shape and the output of the first layer of convolution cropped to the same spatial dimension are concatenated using the channel dimension to fuse feature maps of different resolutions; the number of channels of the concatenated feature map is compressed to 16 again through 1×1 convolution, and the output shape is (32,16,6,4); flatten (32,16,6,4) to automatically calculate the size of other dimensions, so that the output dimension is (32,16*6*4) and increase the time step dimension (Adding The output of the upper layer is used as the input of the encoder and decoder. The Transformer will globally model the features (1, 32, 384) and output the same dimension of the sequence transformed[-1]. The output features (32, 384) of the last time step (lasttime step) are extracted, and two linear transformations (Linear 1 and Linear 2) are performed to reduce the dimension (32, 32) → (32, 5). The output is sent to the final classification layer (classifier), and its dimension is adjusted to the number of output categories (5 categories). The final output of the model is (32, 5), which represents the 5 class probability scores of each sample in the batch (each sample).

[0108] The following is the experiment and result analysis.

[0109] The computer processor used in the experiment is the 12th Gen Intel(R)Core(TM)i5-12490F 3.00GHz, the graphics card is NVIDIA Geforce RTX 3070, the programming environment is Python3.7, and the corresponding deep learning framework is Pytorch.

[0110] The parameter Δt is used to solve the time asynchrony problem of multi-source domain data and adjust the asynchronous data to synchronous data, which is convenient for the early training of the prediction model. In this case, multiple comparative experiments were conducted to verify the effectiveness and authenticity of different time windows and key parameters Δt.

[0111] Due to the asynchronous phenomenon of data, that is, the timestamp of the detection data is not synchronized with the time when the actual data is generated, there is a deviation. The experiment used a time sliding window, such as using a 15-second window size prediction (data features and labels from 1-15 seconds) at the 1st second to predict the output value of the 30th second to solve this asynchronous problem. The window size is set to be consistent with the interval time, which is set as a parameter Δt. Different window sizes correspond to different interval times. The setting of Δt reflects the time feature extraction strategy in data processing. By selecting a suitable Δt value, the time span of feature extraction is determined according to actual needs, so as to capture as complete context information as possible. It can be understood that Δt as a window size controls the coverage of data features, enabling the model to integrate information from different time periods and reduce the prediction deviation caused by a single time point.

[0112] Δt enables the conversion of multi-source asynchronous data to synchronous data by ensuring consistent data sampling and allowing flexibility in dealing with data asynchrony and latency.

[0113] (1) Δt ensures the consistency of data sampling. By keeping the sliding window size consistent with the sampling interval (i.e., the same Δt value), the feature data used for each prediction covers the same amount of historical information. This consistency is particularly important when facing asynchronous data, because it can help the model average the impact of historical information, allowing the model to better capture the overall trend of the time series instead of being overly affected by abnormal data at a certain point in time.

[0114] (2) Adjustment of Δt allows for flexible response to asynchrony and delay. Setting Δt not only allows the model to adapt to different delays, but also can mitigate the time deviation caused by asynchrony by adjusting the window size. Even if the observed data and the real data are not completely synchronized, a reasonable setting of Δt can bridge this gap to a certain extent and provide a more accurate correspondence between time points. In this way, the data has been processed into features that can reflect the real time dynamics before entering the model, and the impact of asynchrony is greatly weakened.

[0115] When dealing with the problem of time asynchrony, the setting of the Δt parameter makes the model more robust by controlling the coverage time range of data features, and can effectively extract key timing information, thereby improving the accuracy of predictions.

[0116] The effectiveness of FPN is explored by comparing the performance of the Transformer model with FPN integrated and the Transformer model without FPN under the same experimental conditions. In the experiment, both model configurations were trained for 100 cycles to ensure sufficient learning. All experiments used the same dataset, loss function, optimizer, and time window size settings to exclude the influence of other variables and ensure the comparability of the results. In terms of learning rate, in order to ensure the optimality of different models, the experiment was debugged many times to meet the optimal learning rate values ​​of different models themselves: the optimal learning rate value of RNN and LSTM is 1e-4, while the learning rate value of Transformer and its FPN-Transformer series models is 1e-5.

[0117] In order to comprehensively evaluate the performance of the model, this example uses multiple evaluation indicators, including precision, recall, F1 score and confusion matrix. Precision and recall provide information about the model's ability to predict the positive class, and the F1 score is the harmonic average of precision and recall, providing a comprehensive evaluation of the model's performance under a single indicator. The confusion matrix provides detailed information about the correct and incorrect predictions of the model in each category, which helps to further analyze the performance differences of the model in different categories.

[0118] Through the above training process and evaluation method, we can understand in detail the learning and prediction performance of the model when processing data in different time windows, thus providing a scientific basis for the monitoring and optimization of the converter steelmaking process.

[0119] In this example, the model and the ablation model have the following Precision, Recall, and F1 scores after model training convergence:

[0120] Table 1 Ablation experiment evaluation indicators

[0121]

[0122] As can be seen from Table 1, the FPN-Transformer model outperforms the traditional Transformer model in all major performance indicators. This improvement is mainly due to the introduction of the FPN structure, which effectively integrates feature information at different levels to enhance the model's ability to capture complex data patterns. The multi-scale feature integration strategy of FPN significantly improves the model's ability to distinguish samples of various categories when facing different features at different stages of change. The loss comparison chart and F1 comparison chart are shown in Figure 1. Figure 8 shown.

[0123] like Figure 8 As shown in the figure, FPN-Transformer is superior to traditional Transformer in terms of loss reduction speed and stability. Especially in the early stage of training, the loss of FPN transformer decreases more rapidly and the learning efficiency is higher. At the same time, the loss curve of FPN-Transformer is smoother, indicating that it is more stable during the learning process. FPN-Transformer also shows high performance in F1 score. From the beginning to the end of training, the F1 score of FPN-Transformer remains at a high level and fluctuates less, indicating that the accuracy and robustness of the model in classification tasks have been enhanced. In contrast, although Transformer can also achieve a high F1 score in the middle and late stages, its volatility is larger and the overall score is lower than that of FPN-Transformer.

[0124] The confusion matrix can reflect the model's prediction performance for different categories. Fig. 9 To compare the prediction effects of the two models for different categories, the performance of the Transformer model is as follows: Category 0 (splash warning): The Transformer model correctly predicted 52 samples, incorrectly predicted 5 samples as category 1, 5 as category 2, 2 as category 3, and 21 as category 4. Category 1 (slag overflow), category 2 (mild splash), and category 3 (severe splash) were all correctly predicted. Category 4 (return dry): 71 samples were correctly predicted, 7 samples were incorrectly predicted as category 0, 3 as category 1, and 4 as category 2. Performance of the FPN-Transformer model: Category 0 (splash warning): The FPN-Transformer model performed slightly better in this category, correctly predicting 57, incorrectly predicting 2 as category 1, incorrectly predicting 3 as category 2, incorrectly predicting 2 as category 3, and 21 as category 4. Category 1 and category 3 were all correctly predicted. Category 2: 85 were correctly predicted, and 2 were incorrectly predicted as category 1. Class 4: 72 were predicted correctly, 8 were mistakenly predicted as class 0, and 1 each were mistakenly predicted as class 1, 2, and 3.

[0125] From the comparison of the two models, it can be seen that FPN-Transformer performs better in reducing certain types of misclassification. In particular, when dealing with categories with unclear boundaries (such as category 0 and category 4), FPN-Transformer can better distinguish different categories. This is because the FPN structure can more effectively capture local and global information when processing features, thereby improving the accuracy and robustness of classification. The FPN structure extracts features at different scales and uses these multi-scale features for prediction, helping the model to better understand the different levels of information in the image or data, thereby improving the classification effect of complex or similar samples. In addition, the introduction of the FPN structure can also help the model to perform feature fusion at a lower level, enhancing the model's ability to capture details, which is of great help in distinguishing subtle differences in classification tasks.

[0126] This example aims to explore the impact of different time window settings on the performance of deep learning-based models, especially in predicting and classifying converter conditions (splash warning, slag overflow, minor splash, severe splash, and back drying). The experiment will compare how the model processes and predicts multi-source domain asynchronous data collected from the converter (including key indicators such as CO gas content, CO2 gas content, oxygen lance height, oxygen flow rate, and oxygen pressure) under time windows of 5, 10, 15, 20, 25, 30, and 60 seconds. Considering the sensitive requirements for warning and response time in actual operation, special attention is paid to the prediction accuracy and responsiveness of the model under different window settings.

[0127] The training of different models all adopts the cross-entropy loss function, which is a commonly used loss function in classification problems and is suitable for multi-classification scenarios. The cross-entropy loss function can quantify the difference between the probability distribution predicted by the model and the probability distribution of the true label, thereby guiding model optimization.

[0128] This example uses the Adam optimizer to update the model's parameters. The Adam optimizer is an adaptive learning rate optimization algorithm that combines the advantages of the AdaGrad and RMSProp optimizers. It can automatically adjust the learning rate according to different parameters and is suitable for processing deep learning models with large amounts of data and parameters. A smaller learning rate (1e-5) is set during training to ensure the stability and convergence of the model during the learning process.

[0129] In order to verify the effectiveness of the basic model Transformer, this example uses RNN, LSTM and Transformer for comparison. By setting the same experimental parameters, the experimental results are shown in Table 2:

[0130] Table 2 Evaluation indicators of different basic models

[0131] Model Precision Recall F1 Score Model inference time (s) RNN 0.7232 0.7327 0.726 0.00000032 LSTM 0.8374 0.8425 0.8374 0.00003123 Transformer 0.8881 0.8878 0.8819 0.00012746 FPN-Transformer 0.913 0.9152 0.9111 0.00013467

[0132] Among the above models, FPN-Transformer showed the best performance in precision (0.901), recall (0.9002) and F1 score (0.897), which shows that by combining the multi-scale feature extraction capability of feature pyramid network FPN and the sequence modeling capability of Transformer, the classification accuracy of the model in complex scenarios has been significantly improved. The model inference time (0.00013467 seconds) is slightly higher than that of the pure Transformer model, but the difference is small and will not have a significant impact on practical applications.

[0133] The Precision, Recall, and F1 scores after model training convergence in different time windows (5 seconds, 10 seconds, 15 seconds, 20 seconds, 25 seconds, 30 seconds, and 60 seconds) are shown in Table 3.

[0134] Table 3 Model evaluation indicators corresponding to different time windows

[0135]

[0136] The changing trends of model loss and F1 score during training under different time windows are as follows: Fig.10 shown.

[0137] Combining the above indicators and Fig.10 It can be seen that the 15-second time window provides the best performance for the model, with excellent performance in terms of precision, recall, and F1 score. This is because the 15-second window is long enough to capture enough context information without introducing noise or irrelevant information due to the window being too long. In contrast, a window shorter than 15 seconds may be too short to effectively capture enough feature information, resulting in lower performance; a window longer than 15 seconds can capture more context, but may also introduce some unnecessary information, resulting in a decrease in model performance (combined with Fig.10 It can be clearly observed that overfitting occurs in the 20, 25, 30, and 60 second time windows.) This comparative experiment can significantly improve the model's precision, recall, and F1 score by adjusting the time window size, thereby improving the model's overall performance.

[0138] From the analysis of the loss function value of the test set, within the first 20 epochs, the loss of all time windows dropped rapidly, indicating that no matter how large the time window is, the model can quickly learn useful features from the data in the early stages of training. The loss of windows below 15 seconds dropped most significantly, because shorter time windows limit the complexity of the input data, making the model easier to optimize. As training progresses, the loss of all models tends to stabilize. The loss of models with windows above 15 seconds fluctuates more frequently, indicating that longer time windows can help the model capture time dependencies and long-term patterns. The loss of these windows increases slightly in the later stages, because the model begins to capture too much irrelevant information or noise in the long time window, causing the model to overfit slightly. The loss function curve of the model with a 15-second window not only drops steadily, but also eventually reaches the lowest loss function value among all models, with the best effect.

[0139] From the F1 value analysis of the test set, the F1 scores of all time windows increased rapidly in the early stage, indicating the general effectiveness of FPN-Transformer in capturing key features. The 15-second time window quickly reached a high and stable level in F1 score, indicating that these window sizes provide the model with enough information to make effective predictions while avoiding the information that may be missed by short time windows and the noise interference of long time windows. After about 40 epochs, the F1 scores of 15-second and 20-second windows showed high stability, while the scores of 5-second and 10-second windows fluctuated slightly and had low F1 scores. This shows that when dealing with data with significant time dependence, too short a window is not enough to capture all relevant information. The window loss fluctuates greatly for windows above 15 seconds, and obvious overfitting occurs at about 30 epochs. This is because longer windows introduce unnecessary context information to a certain extent, thus affecting the prediction accuracy of the model.

[0140] As can be seen from the comparative experiment graph, the 15-second window provides a good balance point in the entire training process, with lower loss and higher F1 score while maintaining good stability. Therefore, this example chooses 15 seconds as the best time window for processing such data.

[0141] In addition, by analyzing the confusion matrix under the 15-second time window setting, we can have a deeper understanding of the classification performance of the model.

[0142] like Fig.11As shown in the figure, the main diagonal in the confusion matrix of the 15-second time window is the predicted category corresponding to the true category, showing the number of correct classifications for each category. There are 57 correct predictions for splash warning, 68 correct predictions for slag spill, 85 correct predictions for slight splash, 88 correct predictions for severe splash, and 72 correct predictions for dry return. The off-diagonal elements show the misclassification cases, that is, the cases where the category predicted by the model does not match the true category: there are 2 cases where the splash warning is misclassified as slag spill, 3 cases as slight splash, 2 cases as severe splash, and 21 cases as dry return.

[0143] The model performed best in predicting slag spills and severe splashing conditions without any misclassification, showing that the model has strong sensitivity and accuracy for these two events. The prediction of slight splashing has a high number of correct predictions, indicating that the model can identify this condition well. The model shows a certain degree of accuracy in distinguishing different splashing and operating events, especially for slag spills and severe splashing. The distinction between splash warnings and backwash needs to be further optimized to reduce misclassification and improve the overall performance of the model. Future work can further improve the model's ability to recognize subtle changes by increasing the diversity of training data, adjusting model parameters, or introducing more complex feature extraction mechanisms.

[0144] According to the grid search method, the optimal set of network hyperparameters is obtained. To verify this method, this example studies the impact of different hyperparameters on different models, mainly comparing RNN, LSTM, Transformer and FPN-Transformer to show the advantages of the model. Since the F1 value can more comprehensively reflect the performance of the model than the precision and recall rate, because F1 combines the two and provides a more balanced evaluation, the F1 value and Test Loss (training loss) are selected as evaluation indicators. The smaller the training loss, the better the model fits the test data and the more accurate the prediction results. The larger the F1, the better the model is in the classification task, which can accurately predict the positive class and ensure that most of the positive classes are captured. The following shows the evaluation scores of different models under different hyperparameter settings.

[0145] While keeping other conditions the same, the performance of the models under different optimizers is compared. All optimizers are suitable for multi-classification tasks. The performance results of different models are shown in Table 4:

[0146] Table 4 Comparison of model performance under different optimizers

[0147]

[0148] Based on the experimental results in the above table, the Adam optimizer has better performance in both training Loss and F1 Score in these models.

[0149] At the same time, the effects of Cross Entropy Loss and Focal Loss on the performance of different models are compared, as shown in Table 5.

[0150] Table 5 Comparison of model performance under different loss functions

[0151]

[0152] As shown in Table 5, from the perspective of Test Loss value, Focal Loss performs better than the former, indicating that Focal Loss makes the model fit the data better; but from the perspective of F1 Score value, Cross Entropy Loss has a better effect, indicating that Cross Entropy Loss makes the model more capable of classification. At the same time, it is verified that the optimal network hyperparameter set obtained by the grid search method is effective.

[0153] In actual industrial applications, there may be failures in the detection equipment, resulting in errors or missing detection data, so that the model cannot make accurate predictions. In order to explore the effect of different input conditions on the prediction accuracy of the model, this example also conducted 5 exploratory experiments. The input conditions and experimental results are shown in Table 6:

[0154] Table 6 Different input conditions and experimental results of 5 exploratory experiments

[0155] Experiment\Input conditions <![CDATA[CO CO2]]> Oxygen flow Rope tension Sonar strength F1 1 √ √ √ √ 0.9111 2 √ √ √ 0.8625 3 √ √ √ 0.9105 4 √ √ √ 0.8549 5 √ √ √ 0.909

[0156] As shown in Table 6, when the above four devices are working normally, the F1 score of the model is 91.11%; when the smoke data is missing, the F1 score of the model is 86.25%; when the oxygen flow data is missing, the F1 score of the model is 91.05%; when the rope tension data is missing, the F1 score of the model is 85.49%; when the sonar intensity data is missing, the F1 score of the model reaches 90.9%. It can be seen that the missing rope tension data has the greatest impact on the model prediction, followed by the missing smoke concentration data, and the missing oxygen flow has the least impact on the model prediction.

[0157] An embodiment of the present invention also provides a converter steelmaking process prediction system for asynchronous data in multiple source domains, which is provided with an intelligent agent, and the intelligent agent is used to implement a converter steelmaking process prediction method for asynchronous data in multiple source domains.

[0158] In view of the problems of information lag, difficulty in extracting and fusing asynchronous data features, and diversification and complexity of converter steelmaking conditions in multi-source domain data of converter steelmaking, the embodiment of the present invention carries out a study on the converter steelmaking condition process prediction model for multi-source domain asynchronous data under complex steelmaking environment, and proposes a converter steelmaking condition process prediction model based on self-attention mechanism of multi-scale feature extraction. A multi-scale feature extraction network is constructed by using feature pyramid network (FPN), and the model's learning ability of multi-source domain asynchronous data features is improved by fusion. Different time windows and window sliding size parameters Δt are set in the FPN-Transformer model to obtain the best model performance. The probability values ​​of different working conditions are calculated and output by ReLU activation function, and the abnormal category corresponding to the maximum probability value is output as the working condition prediction result. By evaluating the prediction effect of different models and different time windows under the same data set and conducting ablation experiments with traditional Transformer, the effectiveness of the proposed model is verified. The best time window is 15s, the accuracy rate reaches 90.1%, the recall rate reaches 90.02%, and the F1 score reaches 89.7%. At the same time, the influence of different hyperparameters on the performance of different models is briefly discussed.

[0159] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for predicting converter steelmaking process based on asynchronous data from multiple source domains, characterized in that: Includes steps: S1. constructing an original data set consisting of asynchronous data from multiple source domains in a converter steelmaking process, and performing equalization preprocessing on the original data set so that the data volume of each identification category in the original data set remains consistent; Normalize the original data set after equalization preprocessing to obtain a normalized data set; Divide the normalized dataset into training and testing sets; S2. Construct a converter steelmaking process prediction model, which includes a feature pyramid and a Transformer module. When the data samples of the training set or the test set are input into the feature pyramid, the feature pyramid combines sample features from different levels through bottom-up feature extraction, top-down feature fusion path and lateral connection to obtain multi-scale fusion features; the Transformer module uses self-attention mechanism, multi-head attention mechanism and residual connection to capture the dependency relationship in the multi-scale fusion features, obtains an output feature map, and calculates and outputs the probability values ​​of different working conditions through activation function, and outputs the working condition prediction result with the maximum probability value; S3, using the training set and the test set to train and test the converter steelmaking process prediction model; S4. Use the trained and tested converter steelmaking process prediction model to predict any input data sample to obtain a process prediction result.

2. The method for predicting converter steelmaking process based on multi-source domain asynchronous data according to claim 1, characterized in that: The calculation steps of the feature pyramid are as follows: (1) Time-series multi-source domain asynchronous data with time windows as the unit quantity is used as input; (2) Extract features at different levels from asynchronous data in time series multi-source domains; (3) Starting from the deepest layer of feature extraction, use point-by-point convolution to generate the number of channels; (4) Use transposed convolution to increase the spatial resolution of features and double the spatial dimension so that the spatial scale of macro features matches that of micro features; (5) Merge features at different levels, integrate the detail information of low-level features with the semantic information of high-level features, and use point-by-point convolution to compress and integrate the number of merged channels to obtain a multi-scale feature pyramid; (6) Repeat (4) and (5) until the top-level features are extracted; (7) Output the fusion results of feature vectors of different scales.

3. The method for predicting converter steelmaking process based on multi-source domain asynchronous data according to claim 2, characterized in that: By setting time windows of different sizes for comparative experiments, the time window with the best model evaluation index is used as the time window of the converter steelmaking condition process prediction model participating in the actual converter steelmaking process prediction in step S4.

4. The method for predicting converter steelmaking process based on asynchronous data in multiple source domains according to claim 3, characterized in that: The Transformer module includes an encoder and a decoder; The processing flow of the encoder is: The input sequence first passes through the embedding layer and position encoding to obtain an input sequence embedding representation with position information; The input sequence embedding means inputting N stacked encoder layers for hierarchical feature extraction. In each encoder layer, the relationship between different positions is first extracted through a multi-head attention mechanism, and then residual connections and layer normalization are performed. Then, the information is passed to the position-aware feedforward neural network, and residual connections and layer normalization are performed again. Finally, through the stacking of N encoder layers, the high-level feature representation of the input sequence is obtained and input to the decoder; The processing flow of the decoder is: The target sequence first passes through the embedding layer and position encoding to obtain an embedded representation of the target sequence with position information; The target sequence embedding means inputting N stacked decoder layers for hierarchical feature extraction. In each decoder layer, the input data first passes through the masked multi-head attention mechanism to extract the relationship between different positions before the current prediction, then performs residual connection and normalization, and then passes it together with the output of the encoder to the multi-head attention mechanism for feature extraction, and then performs residual connection and normalization; Through the stacked N decoder layers, the final output features are obtained; The final output features are passed through a linear layer and Softmax to map the features to the probability distribution of the vocabulary, and the most likely prediction result is selected as the output of the decoder.

5. The method for predicting converter steelmaking process based on multi-source domain asynchronous data according to claim 4, characterized in that: The encoder is provided with four encoder layers, and the decoder is provided with four decoder layers.

6. The method for predicting converter steelmaking process based on asynchronous data from multiple source domains according to claim 5, characterized in that: The cross entropy loss function is used as the loss function for training the converter steelmaking process prediction model.

7. The method for predicting converter steelmaking process based on asynchronous data in multiple source domains according to claim 6, characterized in that: During the training process, the Adam optimizer is used to update the parameters of the converter steelmaking process prediction model.

8. The method for predicting converter steelmaking process based on asynchronous data in multiple source domains according to claim 7, characterized in that: According to the grid search method, the optimal hyperparameter set of the converter steelmaking process prediction model is obtained.

9. The method for predicting converter steelmaking process based on asynchronous data in multiple source domains according to any one of claims 1 to 8, characterized in that: The data collected in the original data set are flue gas data, oxygen flow, oxygen pressure, oxygen lance height, rope tension, sonar intensity and converter mouth flame recognition results at the corresponding time in the converter smelting process; the operating condition categories include splash warning, slag overflow, slight splash, severe splash and dry back. Splash warning means that the system detects signs of impending molten steel splashing, slag overflow refers to the phenomenon that slag overflows from the converter mouth, slight splash refers to splashing reaching a first preset degree, and severe splash refers to splashing reaching a second preset degree deeper than the first preset degree; dry back refers to the phenomenon that high-melting-point substances are precipitated in the slag, the slag becomes thicker, and cannot effectively cover the metal liquid surface.

10. A converter steelmaking process prediction system for asynchronous data in multiple source domains, characterized by: An intelligent agent is provided, and the intelligent agent is used to implement the converter steelmaking process prediction method for multi-source domain asynchronous data as described in any one of claims 1 to 9.

Images