BTP prediction method based on multi-cycle dynamic spatio-temporal feature extraction

Abstract

The invention provides a BTP prediction method based on multi-cycle dynamic spatial-temporal feature extraction. The method comprises the following steps: S1, obtaining sintering time sequence data detected by each sensor in a sintering process, and screening to obtain original data; s2, introducing FFT (Fast Fourier Transform) operation of a one-dimensional time sequence, and performing two-dimensional convolution operation on original data to extract a feature variable; s3, obtaining a static global graph based on a Pearson correlation result, introducing a self-attention learning mechanism to obtain a dynamic local graph, and then performing para-position addition to obtain a graph adjacency matrix; s4, inputting each feature variable into a spatial-temporal feature extraction module to obtain an output result; and S5, converting the channel dimension of each output result into an expected dimension, and then adopting a full connection layer for connection to obtain a BTP prediction result. The method has the beneficial effects that time features at different moments can be captured, and potential spatial information among feature variables is integrated in spatial feature extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of BTP prediction of sintering process, and in particular to a BTP prediction method based on multi-period dynamic spatiotemporal feature extraction. Background Art

[0002] In the sintering process, predicting the position of the sintering end point (BTP) is crucial to optimizing production and improving the quality of sintered ore. Traditional sintering BTP prediction methods mainly include curve fitting and statistical models. Curve fitting is a commonly used mathematical technique used to generate a smooth curve through a given set of data points, but the fitting results may be biased. For complex data sets, multiple fitting curves may be required to better describe the data, which increases the complexity of the model and the computational cost.

[0003] Statistical data driven models can be divided into two categories, namely linear statistical models and nonlinear statistical models. The method based on linear statistical models establishes the relationship between relevant features and BTP through linear models. Since linear statistical models cannot fit the nonlinear relationship between relevant features, nonlinear statistical models are more widely used in BTP prediction. In comparison, nonlinear statistical models are more suitable for BTP prediction. However, there is still room for improvement in their prediction accuracy.

[0004] Deep learning has a powerful ability in extracting big data features. In the field of BTP prediction in sintering, deep learning methods are used to explore time correlation. Previous methods mainly only considered the time relationship of the sintering process. Many studies lacked the analysis and processing of the position relationship of detection points in the data. By integrating spatial features into the field of sintering BTP prediction, the position information between different sensor detection points can form a graph structure, which can be integrated into the process of time series data feature extraction, thereby improving the prediction accuracy.

[0005] The currently used BTP prediction methods lack the processing of capturing time features at different moments and do not consider the potential spatial information between feature variables in spatial feature extraction, which makes the final BTP prediction results not very accurate. Summary of the invention

[0006] The technical problem to be solved by the present invention is how to capture the time characteristics at different moments and integrate the potential spatial information between feature variables in the spatial feature extraction to improve the BTP prediction accuracy. In order to overcome the defects of the above-mentioned prior art (or related technology), the present invention provides a BTP prediction method based on multi-period dynamic spatiotemporal feature extraction.

[0007] The present invention provides a BTP prediction method based on multi-period dynamic spatiotemporal feature extraction. A plurality of sensors are pre-configured in a sintering device. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction comprises the following steps:

[0008] Step S1, obtaining sintering time series data detected by each sensor during the sintering process and screening out features related to target features as raw data;

[0009] Step S2, for a time window of a period of time, introducing an FFT operation of a one-dimensional time series to obtain multiple sub-periods, and performing a two-dimensional convolution operation on the original data in the time window to extract characteristic variables corresponding to each sub-period;

[0010] Step S3, processing the Pearson correlation results between the feature variables and constructing a static global graph based on the Pearson correlation results, and introducing a self-attention learning mechanism to obtain a dynamic local graph according to the feature variables, and then performing positional addition and normalization processing on the static global graph and the dynamic local graph to obtain a graph adjacency matrix;

[0011] Step S4, inputting each of the feature variables into a plurality of pre-configured spatiotemporal feature extraction modules to obtain corresponding output results based on the graph adjacency matrix and the one-dimensional time series;

[0012] Step S5, converting the channel dimension of the output results of each of the spatiotemporal feature extraction modules to a desired dimension through a convolutional layer, and then using a fully connected layer to connect the output results to obtain a BTP prediction result.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] In the present invention, the sintering time series data and the original data are processed by step S1, the sub-period division of the time window and the feature variable extraction are performed by step S2, the graph adjacency matrix is ​​calculated by step S3, the spatiotemporal feature extraction module is used by step S4, and the output result conversion and the BTP prediction result calculation are performed by step S5. In the present invention, considering that the previous prediction methods of BTP have not studied the time characteristics at different moments, a time series alignment method based on multi-period segmentation is introduced, and the sub-periods in the time window are divided by combining FFT operation. Considering that the previous prediction methods of BTP have not considered the integration of potential spatial information between feature variables in spatial feature extraction, multiple sensors at different positions are used to detect and obtain the sintering time series data, and the time features are extracted by two-dimensional convolution on the time dimension, in order to capture the time dependencies of different scales in the time window and realize the capture of time features at different moments. At the same time, the potential spatial information between feature variables is integrated in the spatial feature extraction, and the graph structure construction of static global graph and dynamic global graph is carried out, which fully considers the spatial features and can improve the BTP prediction accuracy.

[0015] In a possible implementation, the step S1 includes:

[0016] Step S11, taking the roller speed, the bellows negative pressure, the bellows temperature and the raw material ratio as data variables, obtaining the roller speed, the bellows negative pressure, the bellows temperature and the raw material ratio at different times during the sintering process as the sintering time series data;

[0017] Step S12, filling missing values ​​in the collected sintering time series data by using the mean value interpolation method;

[0018] Step S13, using the maximum and minimum normalization algorithm to normalize the time series D of each data variable i ={d1,d2,d3,...,d t-1 ,d t} Normalization is performed in the time dimension, and the calculation formula of the maximum and minimum normalization algorithm is as follows:

[0019]

[0020] Among them, max(D i ) represents the time series D i The maximum value, min(D i ) represents the time series D i The minimum value in d t represents the data before normalization, D′ i represents the normalized data;

[0021] Step S14, performing Pearson correlation analysis on the target feature and other features in the sintering time series data to obtain corresponding correlation coefficients, and selecting features with absolute values ​​of the correlation coefficients greater than 0.5 as the original data.

[0022] In a possible implementation, in step S14, the correlation coefficient is obtained by the following calculation formula:

[0023]

[0024] in,

[0025] r represents the correlation coefficient;

[0026] T represents the number of time series;

[0027] Y represents the time series of the target feature;

[0028] Z represents the time series of other features;

[0029] Y t The value of the time series representing the target feature at time t;

[0030] Z t The value of the time series representing other features at time t;

[0031] an average value of the time series representing the target feature;

[0032] Represents the mean of the time series of other features.

[0033] In a possible implementation, step S2 includes:

[0034] Step S21, introduce the FFT operation of the one-dimensional time series, and the FFT operation formula of the time series with a time window length of T is as follows:

[0035] Alpha i =Avg(Amp(FFT(X i )))

[0036]

[0037] Among them, FFT(·) represents FFT operation, Amp(·) represents the size of the calculated amplitude value, Avg(·) represents the average operation, and A i Indicates the calculated amplitude value corresponding to each cycle, X i represents the data of the i-th characteristic variable, p j represents the length of the jth selected cycle;

[0038] Step S22, according to the selected frequency set {f1, ..., f k} and the corresponding period lengths {ρ1,...,ρ k}, extract features from the original data through two-dimensional convolution to obtain feature variables and unify their dimensions, and decompose the one-dimensional time series with a time window length of T into a two-dimensional tensor composed of multiple sub-periods.

[0039] In a possible implementation manner, after executing step S22, the method further includes:

[0040] A fuzzy curve analysis algorithm is used to estimate the delay in importance between each of the characteristic variables and the target characteristic variable to obtain an optimal time delay, and time delay alignment is performed on each of the corresponding sub-cycles based on the optimal time delay.

[0041] In a possible implementation, step S3 includes:

[0042] Step S31, processing the Pearson correlation results between the feature variables and constructing the static global graph based on the Pearson correlation results;

[0043] Step S32, introducing the self-attention mechanism to extract the dependencies in the learning window, the formula is as follows:

[0044] Q=X i W q , K = X i W k , V = X i W v

[0045] in, represents the trainable weight matrix, represents the query vector matrix, represents the key vector matrix, represents a matrix of value vectors;

[0046] Each of the feature variables is then embedded into a learnable positional encoding, using the following self-attention formula:

[0047]

[0048] in, represents the value after scaling the inner product of the matrix, represents a learnable positional encoding, Represents the result of self-attention calculation;

[0049] Then, for each of the feature variables captured in different sub-periods, k features of N feature variables are combined using learnable parameters to obtain the dynamic local graph;

[0050] Step S33, performing positional addition and normalization processing on the static global graph and the dynamic local graph to obtain the graph adjacency matrix, and the calculation formula used is as follows:

[0051] A=normal(A S +A D )

[0052] in, represents the graph adjacency matrix, A S represents the static global graph, represents the dynamic local graph.

[0053] In a possible implementation, in step S31, the static global graph is obtained by the following calculation formula:

[0054]

[0055]

[0056] in,

[0057] Represents the data of the i-th feature variable;

[0058] Represents the data of the jth feature variable;

[0059] n represents the total number of characteristic variables;

[0060] n-1 indicates an unbiased estimate;

[0061] Represents the expected value of the data of the i-th feature variable;

[0062] Represents the expected value of the data of the jth feature variable;

[0063] σ xi represents the standard deviation of the i-th characteristic variable;

[0064] represents the standard deviation of the jth feature variable;

[0065] The static global graph represents the ith feature variable and the jth feature variable.

[0066] In a possible implementation, in step S32, the dynamic local map is obtained by the following calculation formula:

[0067]

[0068] in,

[0069] The dynamic local graph representing the ith characteristic variable and the jth characteristic variable;

[0070] w l represents the learnable parameter;

[0071] Α' i Represents the last data in the feature variable extracted in each sub-period.

[0072] In a possible implementation, step S4 includes:

[0073] Step S41: give a given one-dimensional time series Respectively with four convolution kernels f of different dimensions 1×2 ,f 1×3 ,f 1×5 ,f 1×7 Perform convolution operation, the calculation formula is as follows:

[0074]

[0075] in, Represents the features extracted by the corresponding convolution kernel, and s represents the expansion coefficient;

[0076] Then, the features extracted by the convolution kernel are connected across channels through the connection layer and nonlinearly transformed. The calculation formula is as follows:

[0077] z=Tanh(concat(X f1 ,X f2 ,X f3 ,X f4 ))

[0078] in, Represents the result of nonlinear transformation, Tanh() indicates that the activation function provides nonlinear transformation to suppress the gradient disappearance problem;

[0079] Subsequently, the time series is passed through a fully connected layer containing one hidden layer to extract features, and a sigmoid activation function is added. The calculation formula is as follows:

[0080] z T =sigmoid(FC(X))

[0081] in, Represents the feature extraction result, FC() represents the fully connected layer, and sigmoid() represents the activation function;

[0082] Then, z is extended by channel expansion T Multiply it with z to get the characteristic data result;

[0083] Step S42, input the feature data results into the pre-configured multiple spatiotemporal feature extraction modules, based on the graph adjacency matrix A and the corresponding one-dimensional time series X, the one-dimensional time series X is used as a hidden state, and a dynamic local graph is added to control the selection of features to obtain a comprehensive feature that integrates the relevant dependencies between feature variables.

[0084] In a possible implementation, in step S42, the comprehensive feature is obtained by the following calculation formula:

[0085] H (i) =α i AH in +(1-α i )H (i-1)

[0086]

[0087] in,

[0088] H in Representing the characteristic data result;

[0089] α i represents trainable parameters;

[0090] A represents the graph adjacency matrix;

[0091] Linear() indicates that the linear layer is used to vertically filter the information in each of the spatiotemporal feature extraction modules and add their corresponding positions;

[0092] Represents the comprehensive characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 is a flow chart of the steps of the present invention;

[0094] Figure 2 is a specific flow chart of step S1 of the present invention;

[0095] Figure 3 is a specific flow chart of step S2 of the present invention;

[0096] Figure 4 is a specific flow chart of step S3 of the present invention;

[0097] Figure 5 It is a specific flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0098] First, those skilled in the art should understand that these implementations are only used to explain the technical principles of the embodiments of the present invention, and are not intended to limit the protection scope of the embodiments of the present invention. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0099] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0100] See also Figure 1 The embodiment of the present invention discloses a BTP prediction method based on multi-period dynamic spatiotemporal feature extraction, comprising:

[0101] Step S1: Obtain the sintering time series data monitored by each sensor, and use the Pearson correlation coefficient to perform feature screening, and select features with a strong correlation with the target variable as the original data;

[0102] Step S2: Dynamic feature extraction is performed on the sintering time series data. The FFT operation is used to identify different signal patterns in the sintering time series data, thereby obtaining various sub-periods. Subsequently, a two-dimensional convolution operation is introduced to extract shallow spatiotemporal features. After that, different sub-periods are time-delay aligned to align shallow features and integrate them into shallow fusion features that consider multiple signal patterns. For the extracted shallow features, graph learning and spatiotemporal feature extraction are used for learning.

[0103] Step S3, using graph learning methods to process shallow spatiotemporal features, the mutual relationships between different feature variables can be obtained within a time window of length T. As the time series evolves, these relationships are combined with the static global graph to generate dynamic relationships between different feature variables;

[0104] Step S4: The extracted shallow fusion features are used as the input of the spatiotemporal feature extraction module. The spatiotemporal feature extraction module and the spatial feature extraction module work together to further extract potential spatiotemporal features, following a strategy similar to that of a recursive neural network, thereby realizing a bottom-up to top-down extraction strategy, and combining residual connections to alleviate the problem of gradient disappearance;

[0105] Step S5: The connection layer connects the outputs of each spatiotemporal feature extraction module and projects them onto the output dimension, i.e., the obtained BTP prediction result.

[0106] See also Figure 2 The specific implementation process of step S1 is as follows: step S11 to step S14:

[0107] Step S11: Sintering process related data collection: Use different sensors to monitor and obtain a large amount of data such as roller speed, bellows negative pressure, bellows temperature, various raw material ratios, etc. The expected position of BTP is generally at the position of the second to last bellows. Achieving the best BTP at this position can improve the efficiency of the sintering machine and maximize its potential;

[0108] Step S12: Data preprocessing: using the mean value interpolation method to fill in the missing values ​​in the collected sintering time series data;

[0109] Step S13: Data normalization: Use the maximum and minimum normalization algorithm to normalize the sintering time series data of each data variable in the time dimension. Since the data scales of the sintering time series data are different, it is necessary to normalize the data of different characteristic variables in the sintering time series data separately. For the time series D corresponding to each data variable i ={d1,d2,d3,...,d t-1 ,d t}, the maximum and minimum normalization algorithm formula is as follows:

[0110]

[0111] Among them, max(D i ) represents the time series D i The maximum value, min(D i ) represents the time series D i The minimum value in d t represents the data before normalization, D′ i represents the normalized data;

[0112] Step S14: Feature selection: Perform Pearson correlation analysis on the target feature in the sintering time series data and other features, and select features with an absolute value of correlation coefficient greater than 0.5. The Pearson correlation analysis calculation formula is as follows:

[0113]

[0114] in,

[0115] r represents the correlation coefficient;

[0116] T represents the number of time series;

[0117] Y represents the time series of the target feature;

[0118] Z represents the time series of other features;

[0119] Y t Represents the value of the time series of the target feature at time t;

[0120] Z t Represents the value of the time series of other features at time t;

[0121] represents the average value of the time series of the target feature;

[0122] Represents the average of the time series of other features.

[0123] Continue to see Figure 2 The method can construct a BTP prediction model. After executing step S14, the collected raw data is preprocessed and normalized, and then divided into a training set, a validation set, and a test set. Then, the method also includes:

[0124] Step S15: Dataset division: Divide the original data into a training set, a validation set, and a test set in a ratio of 7:2:1. The specific execution steps include:

[0125] Step S151: For supervised learning, a corresponding label needs to be set for each data sample. In the time series prediction problem, the data sample is the historical time series data of a fixed window size, and the label is the future time series of a certain time step that needs to be predicted. The historical time series data of the window size is selected as the original data, and the variable Windows is defined as the length of the historical time series to be used for prediction. Horizon represents the future time step that needs to be predicted. At this time, the dimension of a data sample becomes N represents the time dimension, T represents the size of the historical data window, and its corresponding labels are T′ represents the time step of the prediction;

[0126] Step S152: Divide the data set: Divide the original data into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set is used to train the model parameters during the model training process, the validation set is used to verify the training parameters to prevent overfitting during the training process and improve the training efficiency, and the test set is used to evaluate the training results.

[0127] See also Figure 3 The specific implementation process of step S2 is as follows: step S21 to step S22:

[0128] Step S21: In order to capture local characteristic trends in a longer time window, the model introduces the FFT operation of the one-dimensional time series to find the sub-period. The FFT operation formula of the time series with a time window length of T is as follows:

[0129] Alpha i =Avg(Amp(FFT(X i )))

[0130]

[0131] Wherein, FFT(·) represents the FFT operation, Amp(·) represents the magnitude of the calculated amplitude value, and V represents the calculated amplitude corresponding to each potential cycle;

[0132] Step S22: According to the selected frequency set {f1, ..., f k} and the corresponding period lengths {ρ1,...,ρ k}, extract features from the original data and unify its dimensions through two-dimensional convolution, decomposing the one-dimensional time series with a time window of T into a two-dimensional tensor consisting of multiple sub-periods:

[0133]

[0134] Since the selected cycle length may not be divisible by the window length T, the Padding operation is extended along the time dimension to make its cycle length divisible. The Reshape operation converts the extended one-dimensional time series into rows and columns representing the intra-cycle changes and inter-cycle changes under cycle k respectively. Then, for the two-dimensional tensor, a two-dimensional convolution operation can be introduced to extract features from the original data and unify its dimensions. Finally, each feature variable obtains the corresponding k feature sequences as subsequent inputs. For convenience, it is used as {X1,X2,...,X k},in Where h is the general dimension of the model after feature extraction.

[0135] Continue to see Figure 3 , after executing step S22, it also includes:

[0136] Step S23: Fuzzy curve analysis (FCA) is used to estimate the delay of importance between each feature variable and the target feature variable. The fuzzy curve analysis (FCA) method estimates the importance of each feature variable to the target feature variable by delay. The specific method of FCA processing delay is as follows:

[0137] (1) For the input time series Expert experience is introduced to estimate the maximum time delay between the input feature variable and the target feature variable and recorded as T max , so that we can get the reconstructed time series X corresponding to a certain feature variable i (t) = {x i (t),x i (t-1),...,x i (tT max )},i∈{1,2,...N};

[0138] (2) Obtain the fuzzy vector corresponding to each reconstructed time series through centroid defuzzification The specific calculation formula is as follows:

[0139]

[0140]

[0141] Among them, ρ is used to limit the result size, which is conducive to quickly finding the optimal delay, x BTP is the time series of the corresponding target characteristic variable BTP;

[0142] (3) The maximum and minimum values ​​of the fuzzy vector obtained from each characteristic variable are screened and the optimal delay is obtained. The specific calculation is shown in the following formula:

[0143]

[0144] Among them, argmax() means taking the maximum value. After N operations, the aligned sequence for the input time series is obtained. For convenience, the aligned data is still recorded as When the subsequent window moves backward, the above steps are repeated to find the aligned data corresponding to the latest window time series;

[0145] For the time series obtained from multiple sub-periods, there will be k different time delays for each characteristic variable. The present invention proposes a weighted cost to measure the time delay. When the variable changes slowly, even if the time delay is not aligned, the impact on model training will be much smaller than the data features with drastic changes. Therefore, the variance is extracted as the weight cost indicator to measure the corresponding time delays in different sub-periods to obtain more accurate alignment data. The specific calculation process is shown in the following formula:

[0146]

[0147]

[0148] in, is the mean value corresponding to different sub-periods, normal() is the normalization operation, and finally the alignment data X corresponding to each feature variable is obtained. 1 ,x 2 ,...x N},in

[0149] See also Figure 4 The specific implementation process of step S3 is as follows: step S31 to step S33:

[0150] Step S31: construct a static global graph using the Pearson correlation results of different feature variables; selecting the Pearson coefficient can better avoid outliers to explore the trend between feature variables, and its formula is as follows:

[0151]

[0152]

[0153] Among them, x i It is defined as the i-th characteristic variable for all the data used in the experiment, and n is the total number of all the data used in the experiment. Since the sintering process is always going on, from a macroscopic point of view, the experimental data taken is only a part of the sample, so n-1 represents an unbiased estimate, which allows the sample to better approximate the population. represents the expected value of the overall data of the i-th characteristic variable, σ xi represents the standard deviation of the i-th characteristic variable;

[0154] Step S32: Due to the existence of operators or specific disturbances, the relationship between the characteristic variables in the sintering process is not fixed, and there is a dynamic local relationship between the characteristic variables, that is, this relationship changes with the sliding of the learning window. In order to capture the dynamic local correlation, the signal fluctuation in the learning window is used to extract the features, and the self-attention mechanism is introduced to extract the dependency in the learning window. The formula is as follows:

[0155]

[0156] in, is the trainable weight matrix, is the query vector matrix, is the key vector matrix, is a value vector matrix, similar to the relative position in natural language processing. Although feature variables have no positional relationship, experiments have shown that adding trainable positional encoding is effective in improving performance. The self-attention formula after each feature variable is embedded in the learnable positional encoding is as follows:

[0157]

[0158] in, It is used to scale the values ​​after the inner product of the matrix to ensure that the range of the attention weight is between [0,1]. The softmax activation function is used to convert the output result into a probability distribution output so that the sum of the N values ​​in each row of N rows is 1. Output the learnable positional encoding In order to comprehensively understand the input data, the model uses learnable parameters to comprehensively consider the k features of N feature variables for different dynamic features captured in different sub-periods, and obtains the dynamic graph of the features. The formula to follow is as follows:

[0159]

[0160] Among them, w l is a learnable parameter, Α' i The last comprehensive data of the features extracted for each sub-period is A D is a N-column matrix with N rows repeated;

[0161] Step S33: perform positional addition and normalization processing on the obtained dynamic local map and the static global map to obtain the feature map corresponding to the input data It dynamically adjusts the weights of connected nodes according to different input data. The specific formula is shown in the figure below:

[0162] A=normal(A S +A D ).

[0163] See also Figure 5 The specific implementation process of step S4 is as follows: step S41 to step S42:

[0164] Step S41: Use the time domain convolutional network TCN to extract high-level time features, and then apply an activation function to ensure nonlinear transformation of the data; Respectively with four convolution kernels f of different dimensions 1×2 ,f 1×3 ,f 1×5 ,f 1×7 Perform convolution operation, the specific formula is as follows:

[0165]

[0166] in, Represents the features extracted by the corresponding convolution kernel, s represents the expansion coefficient, and then the convolution results are connected across channels through the connection layer and nonlinearly transformed. The specific operations are as follows:

[0167]

[0168] Among them, the Tanh() activation function provides a nonlinear transformation to suppress the gradient disappearance problem, and the result is The fully connected layer on the other side extracts the features of the time series through a fully connected layer containing one hidden layer, and then adds a sigmoid activation function. The specific formula is as follows:

[0169] z T =sigmoid(FC(X))

[0170] Among them, FC() is the fully connected layer, sigmoid() is the activation function, By channel expansion, z T Multiply it with z to get the output of the time module;

[0171] Step S42: Each module in the horizontal module of the spatial feature learning module will synthesize the information that the historical module considers important and pass this information backward. The vertical information selection fuses the dynamic graph information obtained by graph learning through trainable parameters, and outputs the key information measured after the spatial feature evaluation of the module. Given the graph adjacency matrix A and the corresponding time series X, the time series X is used as a hidden state. When it is in the transmission process, a dynamic local graph is added to control the selection of time series features. The depth is I. The horizontal selection module selects information on the time feature, and finally obtains a comprehensive feature that integrates the correlation and dependency between feature variables. The horizontal propagation formula is shown in the following formula:

[0172] H (i) =α i AH in +(1-α i )H (i-1)

[0173] Among them, H in is the initial input feature data, α i The trainable parameters corresponding to each horizontal feature screening module are used to achieve the ratio of the original state preservation of the i-th node, and H is taken (0) is the feature data initially input, and the calculation formula for selecting I feature information in the information selection operation is as follows:

[0174]

[0175] Among them, Linear() is a linear layer used to vertically filter the information in each module and add their corresponding positions to get the output

[0176] Continue to see Figure 1 , step S5 is specifically as follows:

[0177] The channel dimension of each spatiotemporal feature extraction module output is converted to the desired dimension through the convolution layer, and finally the feature dimension is converted using the fully connected layer. The specific calculation formula is:

[0178]

[0179] in, It represents the output of each layer of the spatiotemporal feature extraction module, P represents the number of layers of the spatiotemporal feature extraction module, FC() is the fully connected layer, and ReLU() is the activation function for nonlinear transformation of the result.

[0180] Embodiment 1

[0181] This example will combine the actual data of daily measurements in the sintering plant to explain in detail the specific parameter meanings, data calculation process and model training process of the BTP prediction model;

[0182] Dataset introduction: The original data was collected from a steel plant for three consecutive months. The original sampling interval was 5s, but considering the amount of data suitable for deep network training, the time interval of the time series was set to 7min. Due to sensor failure or other factors, the dataset contains missing and outlier data. The Lloyd's criterion is used to detect outliers and remove them, and linear interpolation is used for filling;

[0183] Model parameter settings: The time window of the entire input data is set to 128, the number of layers of spatiotemporal extraction is set to 5 layers, the model is trained by the Adam optimizer, and the learning rate is set to 0.0001. In order to prevent overfitting problems, L1 regularization is used, the dropout of the BTP prediction model and each module is set to 0.3, the entire training is repeated for 100 epochs, the connection layer and the hidden dimension are set to 256, the number of layers of the spatiotemporal extraction module is set to 5, the graph node embedding dimension is set to 40, and the batch size is set to 4. This experiment uses two evaluation indicators: root mean square error (RMSE) and mean absolute error (MAE). The smaller the value of the two indicators, the higher the accuracy of the prediction model. Their respective equations are as follows:

[0184]

[0185]

[0186] Among them, y i is the observed value, is the predicted value, Ω is the indicator set of the observed sample;

[0187] The data calculation process of the model: the input of the model is a third-order tensor Where b represents the batch size; T represents the window size, N represents the data feature dimension, and the output data dimension is Where T' represents the time step of the prediction, and the training model assigns weights to the trainable parameters based on the prediction results and the target results;

[0188] The training process of the model: First, input the training set into the model for model training: input the training data into the model, calculate the loss value between the model output and the data label according to the loss function calculation formula, and update the model parameters according to the gradient update strategy of the Adam optimizer; use the validation set to verify the model loss: after each training, input the validation set data to verify the model parameters to prevent the model from overfitting; test set: after the model training is completed, input the test set data to evaluate the effect of the model;

[0189] Verify the effectiveness of the method proposed in the present invention: the method proposed in the present invention is tested with other standard baseline algorithms under the same experimental settings and using the same data set, and the algorithm output results are compared using the same evaluation indicators, and the obtained evaluation indicators: mean absolute error (MAE) and root mean square error (RMSE) values ​​are listed in Table 1; the standard baseline algorithms include: statistical method for time series prediction using autoregression and moving average (ARIMA), long short-term memory network algorithm (LSTM), gated recurrent unit (GRU), and time domain convolutional network (TCN). Table 1 is as follows:

[0190] Table 1 Comparison of mean absolute error (MAE) and root mean square error (RMSE)

[0191]

[0192] As can be seen from Table 1, the method proposed in this paper has a significant decrease in prediction error compared with other standard baseline algorithms in the prediction task, and the effect is optimal;

[0193] To verify the effectiveness of each component of the proposed method, ablation experiments were performed using model variants, and the obtained evaluation indicators: mean absolute error (MAE) and root mean square error (RMSE) values ​​are listed in Table 2. Specifically, these variants include: w / o DCE, the dynamic period extraction module is removed and replaced with a simple fully connected layer; w / o GRL, we simply remove the graph learning module; w / o TFE, the spatiotemporal feature extraction module of each layer is replaced by a fully connected layer; w / o SFE, the spatial feature extraction module of each layer is replaced by a fully connected layer. Table 2 is shown below:

[0194] Table 2 Comparison of mean absolute error (MAE) and root mean square error (RMSE)

[0195]

[0196] It can be seen from Table 2 that the prediction errors of each module in the method proposed in this paper are significantly reduced in the prediction task.

[0197] In the description of the present invention, the description with reference to the terms "one embodiment", "some embodiments", "in the present embodiment", "specific example", or "some examples" etc. means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0198] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A BTP prediction method based on multi-period dynamic spatiotemporal feature extraction, characterized in that: A plurality of sensors are pre-configured in the sintering device. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction comprises the following steps: Step S1, obtaining sintering time series data detected by each sensor during the sintering process and screening out features related to target features as raw data; Step S2, for a time window of a period of time, introducing an FFT operation of a one-dimensional time series to obtain multiple sub-periods, and performing a two-dimensional convolution operation on the original data in the time window to extract characteristic variables corresponding to each sub-period; Step S3, processing the Pearson correlation results between the feature variables and constructing a static global graph based on the Pearson correlation results, and introducing a self-attention learning mechanism to obtain a dynamic local graph according to the feature variables, and then performing positional addition and normalization processing on the static global graph and the dynamic local graph to obtain a graph adjacency matrix; Step S4, inputting each of the feature variables into a plurality of pre-configured spatiotemporal feature extraction modules to obtain corresponding output results based on the graph adjacency matrix and the one-dimensional time series; Step S5, converting the channel dimension of the output results of each of the spatiotemporal feature extraction modules to a desired dimension through a convolutional layer, and then using a fully connected layer to connect the output results to obtain a BTP prediction result.

2. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 1 is characterized in that: The step S1 comprises: Step S11, taking the roller speed, the bellows negative pressure, the bellows temperature and the raw material ratio as data variables, obtaining the roller speed, the bellows negative pressure, the bellows temperature and the raw material ratio at different times during the sintering process as the sintering time series data; Step S12, filling missing values ​​in the collected sintering time series data by using the mean value interpolation method; Step S13, using the maximum and minimum normalization algorithm to normalize the time series D of each data variable i ={d1,d2,d3,...,d t-1 ,d t } Normalization is performed in the time dimension, and the calculation formula of the maximum and minimum normalization algorithm is as follows: Among them, max(D i ) represents the time series D i The maximum value, min(D i ) represents the minimum value in the time series Di, d t represents the data before normalization, D' i represents the normalized data; Step S14, performing Pearson correlation analysis on the target feature and other features in the sintering time series data to obtain corresponding correlation coefficients, and selecting features with absolute values ​​of the correlation coefficients greater than 0.5 as the original data.

3. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 2 is characterized in that: In step S14, the correlation coefficient is obtained by the following calculation formula: in, r represents the correlation coefficient; T represents the number of time series; Y represents the time series of the target feature; Z represents the time series of other features; Y t The value of the time series representing the target feature at time t; Z t The value of the time series representing other features at time t; an average value of the time series representing the target feature; Represents the mean of the time series of other features.

4. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 1 is characterized in that: The step S2 comprises: Step S21, introduce the FFT operation of the one-dimensional time series, and the FFT operation formula of the time series with a time window length of T is as follows: Α i =Avg(Amp(FFT(X i ))) Among them, FFT(·) represents FFT operation, Amp(·) represents the size of the calculated amplitude value, Avg(·) represents the average operation, and A i Indicates the calculated amplitude value corresponding to each cycle, X i represents the data of the i-th characteristic variable, p j represents the length of the jth selected cycle; Step S22, according to the selected frequency set {f1, ..., f k } and the corresponding period lengths {ρ1,...,ρ k }, extract features from the original data through two-dimensional convolution to obtain feature variables and unify their dimensions, and decompose the one-dimensional time series with a time window length of T into a two-dimensional tensor composed of multiple sub-periods.

5. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 4 is characterized in that: After executing step S22, the method further includes: A fuzzy curve analysis algorithm is used to estimate the delay in importance between each of the characteristic variables and the target characteristic variable to obtain an optimal time delay, and time delay alignment is performed on each of the corresponding sub-cycles based on the optimal time delay.

6. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 1 is characterized in that: The step S3 comprises: Step S31, processing the Pearson correlation results between the feature variables and constructing the static global graph based on the Pearson correlation results; Step S32, introducing the self-attention mechanism to extract the dependencies in the learning window, the formula is as follows: Q=X i W q ,K=X i W k ,V=X i W v in, represents the trainable weight matrix, represents the query vector matrix, represents the key vector matrix, represents a matrix of value vectors; Each of the feature variables is then embedded into a learnable positional encoding, using the following self-attention formula: in, represents the value after scaling the inner product of the matrix, represents a learnable positional encoding, Represents the result of self-attention calculation; Then, for each of the feature variables captured in different sub-periods, k features of N feature variables are combined using learnable parameters to obtain the dynamic local graph; Step S33, performing positional addition and normalization processing on the static global graph and the dynamic local graph to obtain the graph adjacency matrix, and the calculation formula used is as follows: A=normal(A S +A D ) in, represents the graph adjacency matrix, A S represents the static global graph, represents the dynamic local graph.

7. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 6 is characterized in that: In step S31, the static global graph is obtained by the following calculation formula: in, Represents the data of the i-th feature variable; Represents the data of the jth feature variable; n represents the total number of characteristic variables; n-1 indicates an unbiased estimate; Represents the expected value of the data of the i-th feature variable; Represents the expected value of the data of the jth feature variable; represents the standard deviation of the i-th characteristic variable; represents the standard deviation of the j-th characteristic variable; The static global graph represents the ith feature variable and the jth feature variable.

8. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 6, characterized in that: In step S32, the dynamic local map is obtained by the following calculation formula: in, The dynamic local graph representing the ith characteristic variable and the jth characteristic variable; w l represents the learnable parameter; A' i Represents the last data in the feature variable extracted in each sub-period.

9. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 1, characterized in that: The step S4 comprises: Step S41: give a given one-dimensional time series Respectively with four convolution kernels f of different dimensions 1×2 ,f 1×3 ,f 1×5 ,f 1×7 Perform convolution operation, the calculation formula is as follows: in, Represents the features extracted by the corresponding convolution kernel, and s represents the expansion coefficient; Then, the features extracted by the convolution kernel are connected across channels through the connection layer and nonlinearly transformed. The calculation formula is as follows: in, Represents the result of nonlinear transformation, Tanh() indicates that the activation function provides nonlinear transformation to suppress the gradient disappearance problem; Subsequently, the time series is passed through a fully connected layer containing one hidden layer to extract features, and a sigmoid activation function is added. The calculation formula is as follows: z T =sigmoid(FC(X)) in, Represents the feature extraction result, FC() represents the fully connected layer, and sigmoid() represents the activation function; Then, z is extended by channel expansion T Multiply it with z to get the characteristic data result; Step S42, input the feature data results into the pre-configured multiple spatiotemporal feature extraction modules, based on the graph adjacency matrix A and the corresponding one-dimensional time series X, the one-dimensional time series X is used as a hidden state, and a dynamic local graph is added to control the selection of features to obtain a comprehensive feature that integrates the relevant dependencies between feature variables.

10. The BTP prediction method based on multi-period dynamic spatiotemporal feature extraction according to claim 9, characterized in that: In step S42, the comprehensive feature is obtained by the following calculation formula: H (i) =a i AH in +(1-a i )H (i-1) in, H in Representing the characteristic data result; α i represents trainable parameters; A represents the graph adjacency matrix; Linear() indicates that the linear layer is used to vertically filter the information in each of the spatiotemporal feature extraction modules and add their corresponding positions; Represents the comprehensive characteristics.

Images