A sintering endpoint prediction method, device, equipment and readable storage medium

Abstract

The application discloses a sintering endpoint prediction method, device, equipment and readable storage medium, obtains to-be-processed data, divides the to-be-processed data into t moments, obtains a feature vector based on the to-be-processed data of the current moment, obtains a factor attention result based on the feature vector and a hidden vector of the previous moment, determines a time attention result, calculates a new feature vector through matrix addition of the feature vector and the time attention result, obtains the hidden vector of the current moment and a prediction result of the current moment based on the new feature vector, judges whether the current moment is the t moment, if not, the next moment is a new current moment, and the step of constructing an adjacency matrix based on the to-be-processed data is returned to be executed, and if yes, a final prediction result is determined based on the t prediction results. In the application, the factor time attention is focused, the contribution of each variable and each time to the prediction is captured, and the accuracy of predicting the sintering endpoint is improved to a certain extent.

Description

Technical Field

[0001] This application relates to the field of iron and steel smelting, and in particular to a method, apparatus, equipment and readable storage medium for predicting sintering endpoint. Background Technology

[0002] In iron and steel smelting technology, sintering is a crucial step, and the sintered ore is the main raw material for blast furnace ironmaking. The sintering process primarily aims to improve the physicochemical properties and metallurgical performance of iron ore, making it more suitable for subsequent smelting processes, thereby improving smelting efficiency and product quality. Because the sintering process consumes a significant amount of energy, and the quality of the sintered ore has a substantial impact on the efficiency and quality of the entire smelting process, iron ore sintering has attracted increasing attention from researchers. The sintering process typically includes batching, mixing and granulation, feeding, ignition and sintering, cooling, and finished product screening. The main sintering steps are as follows: Figure 1 As shown.

[0003] The burning through point (BTP) is one of the most important parameters in the sintering process, indicating the point at which sintering is complete. If the BTP occurs too early, the sintered material layer is burned through prematurely, resulting in overburning. The effective area of ​​the sintering machine is not fully utilized, wasting energy and reducing quality. If the BTP occurs too late, the sintered material layer is not completely burned, resulting in underburning. The sintered ore quality is substandard and cannot meet the requirements for subsequent smelting. Therefore, predicting the BTP in advance is extremely important.

[0004] However, due to the long lag time and strong nonlinearity and polymorphism of the sintering process, along with the strong coupling between variables, there is currently no instrument that can directly measure BTP. Pre-adjustment of the sintering endpoint heavily relies on experienced workers. If workers rely solely on experience without reference data, the deviation in the sintering endpoint prediction can be significant. Therefore, a sintering endpoint prediction method is urgently needed to provide workers with fundamental parameters for prediction, thereby improving the accuracy of sintering endpoint prediction. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a sintering endpoint prediction method, apparatus, device and readable storage medium. It uses historical data to predict the sintering endpoint at multiple future times and provides workers with parameter data of the sintering endpoint based on the prediction results, thereby improving the accuracy of sintering endpoint prediction to a certain extent.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a sintering endpoint prediction method, comprising the following steps:

[0007] S1. During the sintering process, select the initial time predicted for the sintering endpoint;

[0008] S2. Obtain historical data from the sintering process prior to the initial moment as data to be processed, and use this data to predict the sintering endpoint for multiple consecutive future prediction moments. Based on the data to be processed and the prediction model, output the sintering prediction results, including:

[0009] (1) Construct an adjacency matrix based on the data to be processed;

[0010] (2) Based on the data to be processed at the current time and the adjacency matrix, obtain the feature vector;

[0011] (3) Based on the feature vector and the hidden vector of the previous time step, the factor attention result is obtained;

[0012] (4) Based on the factor attention results and the hidden vector of the previous time step, the temporal attention results are obtained;

[0013] (5) Calculate a new feature vector by matrix addition using the feature vector and the time attention result;

[0014] (6) Based on the new feature vector, the hidden vector at the current prediction time and the sintering endpoint prediction result at the current prediction time are obtained;

[0015] (7) Determine whether the current prediction time is the last of multiple consecutive prediction times. If yes, the multiple sintering endpoint prediction results are the sintering endpoint parameters. If no, return to step (3) to predict the sintering endpoint of the next time.

[0016] The method for constructing the adjacency matrix in step (1) includes:

[0017] Calculate the Euclidean distance between each variable in the data to be processed at the current time, and obtain the distance matrix;

[0018] The distance matrix is ​​then subjected to min-max normalization to obtain the standard matrix;

[0019] The distance matrix is ​​sparsified using the standard matrix to determine the graph nodes;

[0020] Based on the graph nodes, the adjacency matrix at the current time is obtained;

[0021] Based on the data to be processed, construct a global graph matrix;

[0022] The adjacency matrix at the current time step is added to the global graph matrix to obtain the final adjacency matrix.

[0023] The methods for obtaining the feature vector in step (2) include:

[0024] Calculate the Laplace matrix using the adjacency matrix;

[0025] Based on the Laplacian matrix and the data to be processed at the current time, the feature vector is obtained.

[0026] The methods for obtaining the factor attention results in step (3) include:

[0027] The feature vectors are processed to obtain the key vectors of each variable at the current time.

[0028] Using the hidden vector from the previous time step, determine the importance score of each variable to the output;

[0029] Factor attention results are obtained based on the key vector of each variable at the current time and the importance score of each variable to the output.

[0030] The methods for obtaining the time attention results in step (4) include:

[0031] The factor attention results are processed to obtain the key vector for each sub-time point at the current time.

[0032] Using the hidden vector from the previous time step, determine the importance score for each sub-time point;

[0033] The temporal attention result is obtained based on the key vector of each sub-time point at the current time and the importance score of each sub-time point.

[0034] The process of obtaining the hidden vector at the current moment and the sintering endpoint prediction result at the current moment based on the new feature vector includes:

[0035] The new feature vector is input into the recurrent neural network model to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time.

[0036] The recurrent neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to determine the information that needs to be forgotten in the previous time step, the input gate is used to determine the importance of the information in the current time step, and the output gate is used to determine the information that needs to be output as a hidden state.

[0037] The recurrent neural network model calculates its loss in the following ways:

[0038] When the absolute value of the difference between the predicted value and the target value is less than or equal to the preset threshold, the loss is calculated using the mean square error.

[0039] When the absolute value of the difference between the predicted value and the target value is greater than a preset threshold, the loss is calculated using the mean absolute error in conjunction with the preset threshold.

[0040] A sintering endpoint prediction device, comprising:

[0041] The data acquisition module is used to acquire the data to be processed.

[0042] The data preprocessing module is used to divide the data to be processed according to time sequence and execute it starting from the first moment in time order;

[0043] A matrix construction module is used to construct an adjacency matrix based on the data to be processed;

[0044] The first feature vector determination module is used to obtain feature vectors based on the data to be processed at the current time and the adjacency matrix;

[0045] The factor attention calculation module is used to obtain the factor attention result based on the feature vector and the hidden vector of the previous time step;

[0046] The temporal attention calculation module is used to obtain the temporal attention result based on the factor attention result and the hidden vector of the previous time step;

[0047] The second feature vector determination module is used to calculate a new feature vector by matrix addition using the feature vector and the time attention result.

[0048] The prediction module is used to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time based on the new feature vector.

[0049] The judgment module is used to determine whether the current time is time t. If not, the next time is the new current time, and the matrix construction module is returned to execute the step of constructing the adjacency matrix based on the data to be processed. If it is, the final prediction result is determined based on t prediction results.

[0050] An electronic device includes: a memory and a processor;

[0051] The memory is used to store programs;

[0052] The processor is used to execute the program, which is used to implement the sintering endpoint prediction method.

[0053] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the sintering endpoint prediction method.

[0054] The advantages of this invention are: it can predict the sintering endpoint with relatively high accuracy, improve sintering quality, and provide a foundation for procedural production. By focusing on the time factors and capturing the contribution of each variable and each time point to the prediction, the accuracy of predicting the sintering endpoint is improved to a certain extent. Attached Figure Description

[0055] The following is a brief explanation of the contents of each of the accompanying drawings and the markings in the drawings:

[0056] Figure 1 A sintering process flow diagram provided for an embodiment of this application;

[0057] Figure 2 A flowchart of a sintering endpoint prediction method provided in this application embodiment;

[0058] Figure 3 This application provides a result diagram of adjusting the number of layers in a graph convolutional network according to an embodiment of the present application.

[0059] Figure 4 A schematic diagram of an optional architecture for a sintering endpoint prediction model provided in this application embodiment;

[0060] Figure 5 This application provides an example of a result diagram showing the adjustment of a scaling factor.

[0061] Figure 6 An optional architecture diagram of an attention network provided in an embodiment of this application;

[0062] Figure 7 A structural diagram of a sintering endpoint prediction device provided in an embodiment of this application;

[0063] Figure 8 This is a hardware structure block diagram of a sintering endpoint prediction device disclosed in an embodiment of this application. Detailed Implementation

[0064] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the description of the preferred embodiments.

[0065] Current technologies often use risk temperature to determine the sintering endpoint. Due to the complexity of the sintering process, there are currently no instruments that can directly measure BTP. In actual production, the location of the sintering endpoint is mostly determined by the bellows temperature. Among them, the classic soft measurement method uses a quadratic polynomial or a higher-order polynomial to fit the bellows temperature and takes the highest point of the fitted curve as the sintering endpoint.

[0066] First, calculate the bellows temperature:

[0067] T(p i ,ω)=ω0+ω1p i 1 +ω2p i 2 +...+ω k p i k

[0068] Where i is the bellows number, p i Let ω be the temperature of the i-th bellows, and ω be the optimization parameter. k is the polynomial order of the fitted function; if k = 4, then the highest-order term of the function is 4. The optimization parameter ω is a vector, w0, w1, ..., w k It is a vector element.

[0069] Then, the parameter ω is optimized using the loss function:

[0070]

[0071] The parameters in the formula are as follows: i is the number of the bellows, T(p i T(w) is the calculated temperature of the i-th bellows. i ω is the actual value of the bellows temperature, and L(w) is the loss function.

[0072] Finally, the temperature curve is obtained, and the highest temperature is taken as the sintering endpoint.

[0073] However, predicting the sintering endpoint solely through the above calculations faces numerous challenges. First, due to the harsh environment of the sintering site and the dynamic nature of the process, the data acquired by on-site sensors contains significant noise. Processing this noise is a daunting task. Second, the sintering process involves numerous parameters, and the intrinsic relationships between these parameters are difficult to quantify accurately. Finally, the parameters in the sintering process exhibit a certain temporal sequence; better capturing the temporal information of the characteristic sequences is also a crucial component of sintering endpoint prediction. Therefore, existing technologies, whether relying on manual experience or calculating the highest temperature, all suffer from significant errors and cannot meet the requirements for predicting the sintering endpoint.

[0074] This application provides a sintering endpoint prediction method, comprising: determining an initial prediction time and multiple future consecutive prediction times after the initial time; using historical sintering process data corresponding to the initial time as data to be processed; predicting the sintering endpoint for the multiple future consecutive prediction times; sorting the multiple consecutive prediction times according to time sequence; and performing the following steps starting from the first prediction time in chronological order: constructing an adjacency matrix based on the data to be processed; obtaining a feature vector based on the data to be processed at the current time and the adjacency matrix; obtaining a factor attention result based on the feature vector and the hidden vector at the previous time; and based on the factor attention result... The hidden vector from the previous time step is used to obtain the temporal attention result. The feature vector and the temporal attention result are then combined using matrix addition to obtain a new feature vector. Based on the new feature vector, the hidden vector and the sintering endpoint prediction result for the current time step are obtained. It is then determined whether the current prediction time step is the last prediction time step. If not, it means that the time point has not been predicted completely. Therefore, the next time step is set as the new current time step, and the process returns to execute the step of constructing an adjacency matrix based on the data to be processed. If so, it means that the initial time step and several nearby time points have been predicted completely, resulting in multiple prediction results. The final prediction result is determined based on the multiple prediction results.

[0075] This application improves the accuracy of predicting the sintering endpoint to some extent by focusing on factor time attention and capturing the contribution of each variable and time to the prediction. The method uses time t during the sintering process as the initial time, and uses historical data from time t to predict the sintering endpoint position (BTP) 40 minutes later. Then, t is incremented by 1, and time t+1 is used as the prediction time point. Historical data before time t+1 is obtained, and similarly, historical data before time t+2 is obtained. The sintering endpoint positions at three time points—40 minutes, 41 minutes, and 42 minutes—are then predicted. The final sintering endpoint can be obtained from these three time points, allowing for the adjustment and control of the production process. Therefore, in the decoder section, this method incorporates the hidden vector h from the previous prediction result to improve the performance of the current prediction (see appendix). Figure 4 (This shows the source and destination of h). For example, when predicting the BTP at 41 minutes, the hidden vector of the 40-minute prediction result is added. This hidden vector h is used as the query vector in the attention calculation. Based on this, the hidden vector of the 41-minute prediction result is added when predicting at 42 minutes. However, when predicting at 40 minutes, there is no previous prediction, so the hidden vector h is initialized to 0. That is, when the time point is the first time point sorted according to time order, the hidden vector h of the previous time point is 0. The specific scheme is as follows:

[0076] This application provides a method for predicting the sintering endpoint, such as... Figure 2 As shown, Figure 2A flowchart of a sintering endpoint prediction method provided in this application embodiment is included. The method may include the following steps:

[0077] Step S100: Acquire the data to be processed. The data to be processed refers to the various variable data during the sintering process, which are collected using different sensors. The sintering process includes stages such as batching, mixing, material distribution, ignition, and exhaust sintering. Multiple sensors are placed at key locations in each stage to detect the sintering status. The data collected by these sensors is saved, and then outliers are removed and missing values ​​are filled in, resulting in 12,000 data points, with each data point spaced one minute apart. These include parameters such as raw material ratio, moisture content in the raw materials after mixing with water, furnace temperature, and main flue temperature.

[0078] This method uses 12,000 data points acquired during the sintering process, each containing 22 parameters, i.e., data shape (12000, 22). A sliding window strategy is used to create the dataset, with a window size of 40, resulting in 11,960 samples. Each sample has a shape of (40, 22), representing 40 values ​​of the 22 parameters over the past 40 minutes. This sample with shape (40, 22) is the data to be processed (the data with shape (40, 22) and its corresponding sintering endpoint position constitute one sample. Subsequent graph convolution, factor temporal attention, and LSTM are all part of the prediction model). The following definitions of T=40 and n=22 originate from this. The data used for prediction at each time point contains historical data of length 40 for each of the 22 parameters. In this model, T=40.

[0079] Specifically, in steel smelting, data on various variables during the sintering process is acquired. This acquisition can be achieved through data collection using different sensors. Sintering is a continuous process with significant time delays. Assuming the steel plant's sintering trolley is 90 meters long and travels at a speed of 1.5–2.0 m / h, it takes approximately 40 minutes to travel from the initial material placement to the sintering endpoint. In this application, the problem can be defined as a sequence-to-sequence regression prediction task with a time delay of 40 minutes. Given a feature sequence at time t... Where T is the size of the time window. That is, the different feature values ​​of a feature at time t in historical data; then all input feature sequences at time t are identified as follows:

[0080] Predicted output. The aforementioned definition can be expressed by the formula:

[0081]

[0082] Where f is the mapping function and θ are all learnable parameters.

[0083] Step S101: Determine the initial prediction time t, and then determine the prediction time. The prediction times are set sequentially as t+40, t+41, t+42...t+m. Generally, it is only necessary to predict the sintering endpoint at the 40th, 41st, and 42nd minutes after the initial time to provide parameters for the workers and obtain historical data during the sintering process as data to be processed. The value of t can be adjusted according to actual needs.

[0084] Step S102: Construct an adjacency matrix based on the data to be processed.

[0085] Specifically, execution proceeds sequentially, starting from the first prediction time t. Constructing the adjacency matrix requires combining global information with local information at the current time step.

[0086] Step S103: Based on the data to be processed at the current time and the adjacency matrix, obtain the feature vector.

[0087] Specifically, in this step, Graph Convolutional Networks (GCNs) can be used. To effectively filter noise in the data, the GCN loops N times. For example... Figure 3 As shown, Figure 3 The image provided in this application illustrates the results of adjusting the number of layers in a graph convolutional network (GCN). As the number of GCN layers increases from 1 to 6, the mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE) all show a decreasing trend. This is because the designed filters perform deep denoising on the data. The model reaches its optimal performance when the number of GCN layers is equal to 5. Further increasing the number of layers leads to a deterioration in model performance, possibly due to excessive network layers causing model degradation and difficulty in continuing learning.

[0088] The formulas for calculating MAE, RMSE, and MSE are as follows:

[0089]

[0090] Among them, y i For the target value, These are predicted values.

[0091] Step S104: Based on the feature vector and the hidden vector from the previous time step, obtain the factor attention result.

[0092] Specifically, the sintering process is dynamic. As various parameters change during sintering, the importance of different parameters also changes. For example, when the proportion of input quicklime changes, the importance of the raw material's moisture content increases. However, ordinary attention cannot accurately capture the importance relationships between variables along the feature dimension. Therefore, factor attention can be designed along the feature dimension to learn the importance between variables. At the current time step (time step 1), the hidden vector from the previous time step is 0.

[0093] Step S105: Based on the factor attention results and the hidden vector from the previous time step, obtain the temporal attention results.

[0094] Specifically, the sintering process is continuous. Each sintering parameter at the current moment changes with the sintering environment based on the previous moment, exhibiting temporal sequence. To focus on the temporal characteristics of each variable, temporal attention can be applied to the time dimension of each feature. When the current moment is the first moment, the hidden vector of the previous moment is 0.

[0095] Step S106: Calculate a new feature vector by combining the feature vector and the temporal attention result using matrix addition.

[0096] Step S107: Based on the new feature vector, obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time.

[0097] Specifically, in this step, the new feature vector can be input into the recurrent neural network model to obtain the hidden vector at the current time step and the sintering endpoint prediction result at the current time step. The recurrent neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to determine the information that needs to be forgotten in the previous time step, the input gate is used to determine the importance of the information at the current time step, and the output gate is used to determine the information that needs to be output as the hidden state.

[0098] The recurrent neural network (RNN) model can utilize a Long Short-Term Memory (LSTM) network to further extract temporal relationships within the sequence and obtain the final prediction output. While RNNs can capture dependencies in sequences, they may face vanishing or exploding gradient problems when processing long sequences. LSTM introduces three gating structures: a forget gate, an input gate, and an output gate. The forget gate controls which information from the cell state in the previous time step is forgotten, the input gate determines the importance of new information in the current time step, and the output gate controls which part of the cell state is output to the hidden state. Through these three gating structures, LSTM effectively mitigates the vanishing and exploding gradient problems. The specific processing can be as follows:

[0099] i t =σ(Wi a t +U i h t-i +b i )

[0100] f t =σ(W f a t +U f h t-i +b f )

[0101] o t =σ(W o a t +U o h t-i +b o )

[0102] c t =f t ⊙c t-1 +i t ⊙tanh(W c a t +U c h t-i +b c )

[0103] h t =c t ⊙tanh(C t )

[0104] Among them, i t f t o t These are the input gate, forget gate, and output gate, respectively, where a t The output of the factorial temporal attention is currently used as the input to the LSTM; i t f t o t These are the input gate, forget gate, and output gate, respectively. t In cellular state, h t For the output of the LSTM network; W f W i W o W c U f U i U i U c b f b f b f b cThese are the parameters learned by the model; ⊙ represents the element-wise multiplication operation; σ represents the ReLU activation function, and tanh represents the tanh activation function.

[0105] Step S108: Determine whether the current time is the last prediction time. Specifically, determine whether the current time is the last prediction time. If not, the next time after the current time is the new current time, and return to step S104 to execute the data to be processed using the new current time. If yes, execute step S109.

[0106] Here, time t refers to a specific point in time, which serves as the starting point for the prediction. Because the raw material at time t will undergo ignition and be fully burned in approximately 40 minutes, the model predicts the sintering endpoint 40 minutes later at this time. Therefore, when making the prediction, the ignition time is selected as the first time t, and then the sintering endpoint is predicted at 40, 41, and 42 minutes after time t.

[0107] To improve prediction performance, we prepared historical data (X) for the 40 minutes prior to time t. t This allows the model to learn the influence of various sintering parameters on the sintering endpoint from the data collected in the first 40 minutes, and to predict the sintering endpoint position 40 minutes later.

[0108] Step S109: If yes, then based on the prediction results corresponding to multiple prediction times, basic parameters are provided for the worker to determine the final sintering endpoint. After steps S101 to S107, the data corresponding to each prediction time will yield a prediction result. For example, for prediction times t to t+m, there are m+1 prediction results. Then, based on the m+1 prediction results, the final sintering endpoint prediction result is obtained. This application improves the accuracy of predicting the sintering endpoint to a certain extent by focusing on factor time attention and capturing the contribution of each variable and each time to the prediction. Specifically, as follows... Figure 4 As shown, Figure 4 This application provides an optional architecture diagram for a sintering endpoint prediction model, which includes an encoder and a decoder. Data processing methods include soft-sensing target vectors (using a multinomial fit to the bellows temperature to calculate BTP) and creating a sliding window of data (i.e., using 40 historical data points to create samples for each input to the model). Graph convolution and attention are part of the model; they have the function of automatically learning complex relationships in the data and filtering out some noise, but they are not part of the data preprocessing.

[0109] In this application, m is set to 2, resulting in three predictions: one for the sintering endpoint 40 minutes later at times t, t+1, and t+2. These three predictions are then processed to obtain the final result. Predicting the sintering endpoint at three different times is primarily to verify the model's stability. By making predictions at multiple time steps, the model's generalization ability to unknown data (i.e., the future sintering endpoint) can be tested. Since this work does not yet cover the automatic control of sintering parameters, these three predictions are mainly intended to provide a reference for operators.

[0110] In some embodiments of this application, the variables in the process are intertwined, making it difficult to derive a clear quantitative relationship. In order to establish a connection between the variables, the process of constructing the adjacency matrix based on the data to be processed in step S102 may include:

[0111] S10. Calculate the Euclidean distance between each variable in the data to be processed at the current time, and obtain the distance matrix. The model input is sequence data, where each sequence represents a sintering parameter. Calculate the Euclidean distance d for each pair of sequences i and j. ij There are n sequences in total. After calculating the distance between all sequence pairs, we get a distance matrix of size n*n. The distance matrix can be represented as:

[0112]

[0113] Specifically, to obtain a graph matrix that better reflects the current time period, one can choose to calculate the Euclidean distance of the sequence data for each batch:

[0114]

[0115] Where, d ij Let be the Euclidean distance between variables i and j at time t. Let variables i and j at time t be expressed as... T represents the size of the time window.

[0116] S11. Perform min-max normalization on the distance matrix to obtain the standard matrix.

[0117] Specifically, we can first calculate the maximum value d of the distance matrix obtained in the previous step. max and minimum value d min Then, perform max-min normalization using the following formula:

[0118]

[0119] S12. Using the standard matrix, sparsify the distance matrix to determine the graph nodes.

[0120] Specifically, the distance matrix can be sparsified to determine the graph nodes. The graph nodes can be represented as follows:

[0121]

[0122] Where ε is the Euclidean distance threshold. Each node after normalization is compared with the Euclidean distance threshold according to the above formula, and then the graph node is established based on the comparison result, and the graph node is assigned a value of 1 or 0.

[0123] S13. Based on the graph nodes, obtain the adjacency matrix at the current time.

[0124] Specifically, the adjacency matrix at time t can be represented as A′∈R n×n .

[0125] S14. Construct a global graph matrix based on the data to be processed.

[0126] Specifically, a global graph A is constructed using the K-Nearest Neighbor (KNN) algorithm based on the training set. * ∈R n×n .

[0127] S15. Add the adjacency matrix at the current time step to the global graph matrix to obtain the final adjacency matrix.

[0128] Specifically, the adjacency matrix A′ at the current moment and the global graph matrix A′ are used to... * Adding them together yields the final adjacency matrix A = A * +A′.

[0129] In some embodiments of this application, a smoothing filter can be designed to suppress the impact of noise on prediction. This filter can filter noise to a certain extent. Step S103, the process of obtaining the feature vector based on the data to be processed at the current time and the adjacency matrix, may include:

[0130] S20. Calculate the Laplace matrix using the adjacency matrix.

[0131] Specifically, the Laplacian matrix L is calculated using the adjacency matrix:

[0132] L=DA

[0133] Where D is the degree matrix of the adjacency matrix A.

[0134] Design a filter F with a scaling factor α using the Laplace matrix L:

[0135] F = I + α(LI)

[0136] Where I is the identity matrix and α is a hyperparameter.

[0137] For the filter scaling factor α, it can be set to 0.1, 0.2, 0.3, 0.4, 0.5, such as... Figure 5 As shown, Figure 5 This diagram illustrates the results of adjusting the scaling factor according to an embodiment of this application. The MEA, RMSE, and MSE parameters all show the same trend: the model achieves optimal performance when α = 0.3, and the error gradually increases after reaching values ​​greater than 0.3. Furthermore, it can be observed that model performance decreases when the scaling factor is removed. Based on the experimental results, we found that α = 0.3 yields the best results; therefore, the model hyperparameter α is set to 0.3.

[0138] S21. Based on the Laplace matrix and the data to be processed at the current time, obtain the eigenvector.

[0139] Specifically, graph convolution operations can be performed using GCN:

[0140]

[0141] in, W represents matrix multiplication. g For model parameters, c t W is the result of graph convolution, i.e., the feature vector output by the encoder. g These are the parameters learned by the model, and are represented in matrix form. X t This refers to the data to be processed.

[0142] In the above embodiments, a smoothing filter F is designed based on the Laplacian matrix and integrated into the graph convolution module to filter noise signals in the sequence, thereby reducing the impact of noise in the data on model performance. The filter F is used to filter the input data Xt at the current time step, filtering out some noise interference.

[0143] In some embodiments of this application, in order to further capture the complex relationships between input variables, a factor attention mechanism can be designed in the decoder. Step S104, the process of obtaining the factor attention result based on the feature vector and the hidden vector from the previous time step, may include:

[0144] S30. Process the feature vectors to obtain the key vectors of each variable at the current time.

[0145] Specifically, the feature vector c t The input factor attention network, after convolution, max pooling, average pooling, and activation operations, yields two n×1×1 feature maps. Adding these two feature maps together produces the key vector. Among them, the key vector The calculation process can be as follows:

[0146]

[0147] Here, nn includes convolution, ReLU activation, and convolution operations; avg_pool represents average pooling; and max_pool represents max pooling. t This refers to the output of the model encoder. It is the k-th feature sequence in the encoder output at time t, where k = 1, 2, ..., n. It is the key vector of the k-th feature sequence computed using the attention mechanism in the feature dimension.

[0148] S31. Using the hidden vector from the previous time step, determine the importance score of each variable to the output.

[0149] Specifically, after performing convolution, batch normalization, and ReLU activation on the hidden vector h from the previous time step, the importance score of each variable to the output is calculated. The importance score can be calculated in the following ways:

[0150]

[0151] This refers to the importance score of the nth variable at time t. `softmax()` is the softmax function. It is the key vector of factor attention. It is the kth sequence of the previously predicted hidden vector.

[0152] S32. Based on the key vector of each variable at the current time and the importance score of each variable to the output, the factor attention results are obtained.

[0153] Specifically, attention values ​​are assigned to the input of each variable to obtain the final factor attention result a. f (The factor attention result is a matrix.)

[0154]

[0155] a f This is the result of factor attention calculation. It is the query score, namely the importance score mentioned above. It is the kth feature sequence in the encoder output at time t.

[0156] In some embodiments of this application, temporal attention can be added after factor attention to determine the importance of each time point in the time dimension of the sequence. Step S105, the process of obtaining the temporal attention result based on the factor attention result and the hidden vector of the previous time step, may include:

[0157] S40. Process the factor attention results to obtain the key vector for each sub-time point at the current time.

[0158] Specifically, the factor attention results are input into the temporal attention network. After convolution, max pooling, average pooling, and activation operations, two n×1×1 feature maps are obtained. The two feature maps are then added together to obtain the key vector e. j :

[0159] e j =nn(avg_pool(a f,j )+max_pool(a f,j ))

[0160] a f It is the result of factor attention, a f，j It is the vector of the j-th historical time in the factor attention output at time t, where j = 1, 2, ..., T. j It is the key vector of the j-th historical time, computed using an attention mechanism in the time dimension.

[0161] Here, nn includes convolution, ReLU activation, and convolution operations, avg_pool represents average pooling, and max_pool represents max pooling.

[0162] S41. Using the hidden vector from the previous time step, determine the importance score for each sub-time point.

[0163] Specifically, similar to factor attention, temporal attention also uses the hidden vector h of the LSTM in the decoder at the previous time step as input. The difference is that temporal attention performs attention calculations in the time dimension, rather than the feature dimension.

[0164] After performing convolution, batch normalization, and ReLU activation on the hidden vector h from the previous time step, calculate the importance score s of the current variable at each time step within a time window of length T. f ={s1,s2,...,s j ,...,s T The importance score can be calculated as follows: s t =softmax(e1h1,e2h2,...,e j h j ,...,e T h T Among them, e j It is the key vector of temporal attention, h j It is the vector of the j-th historical time of the previously predicted hidden vector.

[0165] S42. Based on the key vector of each sub-time point at the current time and the importance score of each sub-time point, obtain the time attention result.

[0166] Specifically, the input at each time step is assigned attention values ​​to obtain the final time-based attention calculation result a. t :

[0167] a t =(s1c1,s2c2,...,s j c j ,...,s T c T )

[0168] The result of the attention function is a matrix of size (batch_size, channel_size, window_size, feature_number), which represents (batch size, number of channels, window length for historical data, number of parameters). In this model, the size is (64, 32, 40, 22). t This is the result of time attention calculation, s j It is the query score, i.e., the importance score mentioned above, c j It is the data at the j-th time point in the encoder output at time t.

[0169] In the above embodiments, a factor-temporal attention network was designed. Factor attention and temporal attention focus on important components in the variable and time dimensions, respectively, enhancing their role in model prediction and thus improving prediction accuracy to some extent. The architectures of the factor attention network and temporal attention network can be found in [reference needed]. Figure 6 As shown, Figure 6 This is a diagram of an optional architecture for an attention network provided in an embodiment of this application.

[0170] The feature vectors and time attention results from step S21 are used to calculate a new feature vector through matrix addition.

[0171] The new feature vector is input into the recurrent neural network model to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time.

[0172] The recurrent neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to determine the information that needs to be forgotten in the previous time step, the input gate is used to determine the importance of the information in the current time step, and the output gate is used to determine the information that needs to be output as a hidden state.

[0173] When training a recurrent neural network (RNN) model, its loss function is designed as follows: The methods for calculating the loss of an RNN model include:

[0174] When the absolute value of the difference between the predicted value and the target value is less than or equal to the preset threshold, the loss is calculated using the mean square error.

[0175] When the absolute value of the difference between the predicted value and the target value is greater than a preset threshold, the loss is calculated using the mean absolute error in conjunction with the preset threshold.

[0176] In some embodiments of this application, SmoothL1Loss can be used as the objective function, and a threshold can be used to determine whether to use mean squared error loss or mean absolute error loss with a threshold.

[0177] Specifically, the recurrent neural network model in this application calculates the loss by using mean squared error when the absolute value of the difference between the predicted value and the target value is less than or equal to a preset threshold; and by using mean absolute error in conjunction with the preset threshold when the absolute value of the difference between the predicted value and the target value is greater than the preset threshold. The calculation formula can be as follows:

[0178]

[0179] Where y is the target value. The predicted value is δ = 1.

[0180] In the above embodiments, by using SmoothL1Loss as the objective function, the model can be made more robust.

[0181] At time t, using T historical data points (containing 22 parameters), predict the location of the sintering endpoint at three time points: t+40, t+41, and t+42 minutes.

[0182] The model outputs predictions for the sintering endpoint at 40, 41, and 42 minutes after time t, resulting in three final predictions. The three evaluation metrics are also calculated using these three results.

[0183] The recurrent neural network LSTM transmits information from the previous LSTM unit to the current LSTM unit through a cell state, and then forgets and updates the cell state information within the LSTM unit before inputting it into the next LSTM unit. In this way, LSTM can preserve and transmit historical information. The prediction model for predicting the sintering endpoint in this application consists of an encoder and decoder, and includes not only LSTM (recurrent neural network), but also graph convolution (graph construction and Laplacian-based filters also serve the graph convolutional neural network (GCN) in the model) and factor temporal attention. Figure 4This describes the structure of the encoder and decoder. Data processing methods include soft-sensing of the target vector (using a multinomial fit to the bellows temperature to calculate BTP) and creating a sliding window of data (i.e., using 40 historical data points to create samples for each input to the model). Graph convolution and attention are both part of the model; they have the function of automatically learning complex relationships in the data and filtering out some noise, but they are not part of the data preprocessing.

[0184] The following describes a sintering endpoint prediction device provided in the embodiments of this application. The sintering endpoint prediction device described below and the sintering endpoint prediction method described above can be referred to and correspond to each other.

[0185] like Figure 7 As shown, Figure 7 A structural diagram of a sintering endpoint prediction device provided in this application embodiment is shown. The device may include:

[0186] Data acquisition module 10 is used to acquire data to be processed;

[0187] Data preprocessing module 20 is used to divide the data to be processed into t time periods;

[0188] Execution will proceed in chronological order, starting from the first moment.

[0189] Matrix construction module 30 is used to construct an adjacency matrix based on the data to be processed;

[0190] The first feature vector determination module 40 is used to obtain feature vectors based on the data to be processed and the adjacency matrix at the current time.

[0191] The factor attention calculation module 50 is used to obtain the factor attention result based on the feature vector and the hidden vector of the previous time step;

[0192] The temporal attention calculation module 60 is used to obtain the temporal attention result based on the factor attention result and the hidden vector of the previous time step;

[0193] The second feature vector determination module 70 is used to calculate a new feature vector by matrix addition of the feature vector and the time attention result.

[0194] The prediction module 80 is used to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time based on the new feature vector;

[0195] The judgment module 90 determines whether the current time step is the last time step. If not, the next time step is the new current time step, and the result is returned to the factor attention calculation module. If it is, reference data is provided based on multiple prediction results to determine the final prediction result. If the process is not completed at this point, it returns to the decoder section to execute factor temporal attention and LSTM. At this point, the hidden vector h from the previous time step's prediction output is added to the factor temporal attention to improve the prediction performance at the current time step.

[0196] As can be seen from the above technical solution, the sintering endpoint prediction device provided in this application includes: a data acquisition module 10 acquiring data to be processed; a data preprocessing module 20 dividing the data to be processed into t time moments, executing from the first time moment in chronological order; a matrix construction module 30 constructing an adjacency matrix based on the data to be processed; a first feature vector determination module 40 obtaining a feature vector based on the data to be processed at the current time moment and the adjacency matrix; a factor attention calculation module 50 obtaining a factor attention result based on the feature vector and the hidden vector at the previous time moment; a time attention calculation module 60 obtaining a time attention result based on the factor attention result and the hidden vector at the previous time moment; a second feature vector determination module 70 calculating a new feature vector by matrix addition using the feature vector and the time attention result; a prediction module 80 obtaining the hidden vector at the current time moment and the sintering endpoint prediction result at the current time moment based on the new feature vector; and a judgment module 90 judging whether the current time moment is time moment t. If not, the next time moment is the new current time moment, and the step of constructing an adjacency matrix based on the data to be processed moment moment is returned to execution. If so, the final prediction result is determined based on the t prediction results. This application improves the accuracy of predicting the sintering endpoint to some extent by focusing on the time attention of factors and capturing the contribution of each variable and each time point to the prediction.

[0197] Optionally, the matrix construction module 30 of the sintering endpoint prediction device, which performs the step of constructing an adjacency matrix based on the data to be processed, may include:

[0198] Calculate the Euclidean distance between each variable in the data to be processed at the current time, and obtain the distance matrix;

[0199] The distance matrix is ​​then subjected to min-max normalization to obtain the standard matrix;

[0200] The distance matrix is ​​sparsified using the standard matrix to determine the graph nodes;

[0201] Based on the graph nodes, the adjacency matrix at the current time is obtained;

[0202] Based on the data to be processed, construct a global graph matrix;

[0203] The adjacency matrix at the current time step is added to the global graph matrix to obtain the final adjacency matrix.

[0204] Optionally, the first feature vector determination module 40 of the sintering endpoint prediction device, which performs the step of obtaining the feature vector based on the data to be processed and the adjacency matrix at the current time, may include:

[0205] Calculate the Laplace matrix using the adjacency matrix;

[0206] Based on the Laplacian matrix and the data to be processed at the current time, the feature vector is obtained.

[0207] Optionally, the factor attention calculation module 50 of the sintering endpoint prediction device, when performing the process of obtaining the factor attention result based on the feature vector and the hidden vector of the previous time step, may include:

[0208] The feature vectors are processed to obtain the key vectors of each variable at the current time.

[0209] Using the hidden vector from the previous time step, determine the importance score of each variable to the output;

[0210] Factor attention results are obtained based on the key vector of each variable at the current time and the importance score of each variable to the output.

[0211] Optionally, the time attention calculation module 60 of the sintering endpoint prediction device, performing the step of obtaining the time attention result based on the factor attention result and the hidden vector of the previous time step, may include:

[0212] The factor attention results are processed to obtain the key vector for each sub-time point at the current time.

[0213] Using the hidden vector from the previous time step, determine the importance score for each sub-time point;

[0214] The temporal attention result is obtained based on the key vector of each sub-time point at the current time and the importance score of each sub-time point.

[0215] Optionally, the prediction module of the sintering endpoint prediction device, performing the step of obtaining the hidden vector at the current time and the sintering endpoint prediction result at the current time based on the new feature vector 90, may include:

[0216] The new feature vector is input into the recurrent neural network model to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time.

[0217] The recurrent neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to determine the information that needs to be forgotten in the previous time step, the input gate is used to determine the importance of the information in the current time step, and the output gate is used to determine the information that needs to be output as a hidden state.

[0218] This application also provides an electronic device. Figure 8 A hardware block diagram of the electronic device is shown, with reference to... Figure 8 The hardware structure of an electronic device may include: at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4;

[0219] In this embodiment, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4.

[0220] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0221] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0222] The memory stores a program, which the processor can call. The program is used to implement each step of the aforementioned sintering endpoint prediction method.

[0223] This application embodiment also provides a storage medium that can store a program suitable for execution by a processor, the program being used to implement the various steps of the aforementioned sintering endpoint prediction method.

[0224] Finally, 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0225] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined with each other, and the same or similar parts can be referred to each other.

[0226] Obviously, the specific implementation of this invention is not limited to the above-described methods. Any non-substantial improvements made using the inventive concept and technical solution of this invention are within the protection scope of this invention.

Claims

1. A method for predicting the sintering endpoint, characterized in that, Includes the following steps: S1. During the sintering process, select the initial time predicted for the sintering endpoint; S2. Obtain historical data from the sintering process prior to the initial moment as data to be processed, and use this data to predict the sintering endpoint for multiple consecutive future prediction moments. Based on the data to be processed and the prediction model, output the sintering prediction results, including: (1) Construct an adjacency matrix based on the data to be processed; (2) Based on the data to be processed at the current time and the adjacency matrix, obtain the feature vector; (3) Based on the feature vector and the hidden vector of the previous time step, the factor attention result is obtained; (4) Based on the factor attention results and the hidden vector of the previous time step, the temporal attention results are obtained; (5) Calculate a new feature vector by matrix addition using the feature vector and the time attention result; (6) Based on the new feature vector, the hidden vector at the current prediction time and the sintering endpoint prediction result at the current prediction time are obtained; (7) Determine whether the current prediction time is the last of multiple consecutive prediction times. If yes, the multiple sintering endpoint prediction results are the sintering endpoint parameters. If no, return to step (3) to predict the sintering endpoint of the next time.

2. The sintering endpoint prediction method according to claim 1, characterized in that, The methods for constructing the adjacency matrix in step (1) include: Calculate the Euclidean distance between each variable in the data to be processed at the current time, and obtain the distance matrix; The distance matrix is ​​then subjected to min-max normalization to obtain the standard matrix; The distance matrix is ​​sparsified using the standard matrix to determine the graph nodes; Based on the graph nodes, the adjacency matrix at the current time is obtained; Based on the data to be processed, construct a global graph matrix; The adjacency matrix at the current time step is added to the global graph matrix to obtain the final adjacency matrix.

3. The sintering endpoint prediction method according to claim 1, characterized in that, The methods for obtaining the feature vector in step (2) include: Calculate the Laplace matrix using the adjacency matrix; Based on the Laplacian matrix and the data to be processed at the current moment, the feature vector is obtained.

4. The sintering endpoint prediction method according to claim 1, characterized in that, The methods for obtaining the factor attention results in step (3) include: The feature vectors are processed to obtain the key vectors of each variable at the current time. Using the hidden vector from the previous time step, determine the importance score of each variable to the output; Factor attention results are obtained based on the key vector of each variable at the current time and the importance score of each variable to the output.

5. The sintering endpoint prediction method according to claim 1, characterized in that, The methods for obtaining the time attention results in step (4) include: The factor attention results are processed to obtain the key vector for each sub-time point at the current time. Using the hidden vector from the previous time step, determine the importance score for each sub-time point; The temporal attention result is obtained based on the key vector of each sub-time point at the current time and the importance score of each sub-time point.

6. The sintering endpoint prediction method according to claim 1, characterized in that, The process of obtaining the hidden vector at the current moment and the sintering endpoint prediction result at the current moment based on the new feature vector includes: The new feature vector is input into the recurrent neural network model to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time. The recurrent neural network model includes a forget gate, an input gate, and an output gate. The forget gate is used to determine the information that needs to be forgotten in the previous time step, the input gate is used to determine the importance of the information in the current time step, and the output gate is used to determine the information that needs to be output as a hidden state.

7. The sintering endpoint prediction method according to claim 6, characterized in that, The recurrent neural network model calculates its loss in the following ways: When the absolute value of the difference between the predicted value and the target value is less than or equal to the preset threshold, the loss is calculated using the mean square error. When the absolute value of the difference between the predicted value and the target value is greater than a preset threshold, the loss is calculated using the mean absolute error in conjunction with the preset threshold.

8. A sintering endpoint prediction device, characterized in that, include: The data acquisition module is used to acquire the data to be processed. The data preprocessing module is used to sort the data to be processed by time. Execution will proceed in chronological order, starting from the first predicted moment. A matrix construction module is used to construct an adjacency matrix based on the data to be processed; The first feature vector determination module is used to obtain feature vectors based on the data to be processed at the current time and the adjacency matrix; The factor attention calculation module is used to obtain the factor attention result based on the feature vector and the hidden vector of the previous time step; The temporal attention calculation module is used to obtain the temporal attention result based on the factor attention result and the hidden vector of the previous time step; The second feature vector determination module is used to calculate a new feature vector by matrix addition using the feature vector and the time attention result. The prediction module is used to obtain the hidden vector at the current time and the sintering endpoint prediction result at the current time based on the new feature vector. The judgment module is used to determine whether the current prediction time is the last of multiple prediction times. If not, the next time is the new current prediction time, and the matrix construction module is returned to execute the step of constructing an adjacency matrix based on the data to be processed. If so, the final prediction result is determined based on multiple prediction results.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is configured to execute the program, which is configured to implement the sintering endpoint prediction method as described in any one of claims 1-7.

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

Images