Sintering mixture moisture intelligent prediction method based on deep learning

Abstract

The invention discloses a sintering mixture moisture intelligent prediction method based on deep learning, and belongs to the technical field of metallurgy sintering. According to the method, the local receptive field and parameter sharing characteristics of a CNN module are utilized, multi-scale features are extracted, local correlation of key variables in moisture prediction is captured, self-adaptive modeling of global features is realized through a Transform module, a long-distance dependency relationship of moisture change is excavated, an attention mechanism is introduced to endow the features with weights, and the moisture prediction accuracy is improved. And key features highly related to moisture prediction are highlighted, so that the prediction precision is improved. Raw material moisture values collected in real time and key parameters in the mixing process are input into the model in an end-to-end mode, high-precision prediction of future moisture changes in the mixture moisture control process is achieved through deep learning, and accurate and reliable data support is provided for the mixture mixing process.

Description

Technical Field

[0001] The present invention belongs to the technical field of metallurgical sintering, and more specifically, relates to an intelligent prediction method for the moisture content of sintering mixtures based on deep learning. Background Art

[0002] Sintering is currently the primary method for agglomerating iron ore fines in China, with sintered ore accounting for over 70% of blast furnace charge. Sintered ore with a stable composition is crucial for ensuring blast furnace smelting efficiency and quality. In sintering production, adding water, mixing, and pelletizing is a key step that directly impacts the yield and quality of the sintered ore. Appropriate moisture content in the mix ensures good particle size distribution and a well-ventilated bed.

[0003] However, existing sintering processes face challenges in controlling the moisture content of the mixture, including hysteresis, nonlinearity, low detection accuracy, multivariate interference, and data fluctuation. Furthermore, due to operational complexity and safety costs, most sintering plants in China still rely on remote manual control of water addition. This approach, due to differences in operator habits and experience, can lead to inconsistent quality of the final sintered ore during actual production. Therefore, accurate mixture moisture prediction is crucial. It not only helps detect potential anomalies promptly but also allows for optimized water addition strategies to further enhance the stability of the mixing process.

[0004] After searching, the Chinese patent publication number is CN 112986491 A, the publication date is June 18, 2021, and the name of the invention is: A method for correcting the moisture detection value of a mixture based on a feedback adaptive prediction model. The method of this patent includes: step 1: collecting raw material moisture content information, step 2: calculating the moisture content of the mixture through the ingredient calculation system, step 3: using the MIV algorithm to screen the model input, using the collected samples to train the BP neural network, the trained adaptive neural network system predicts the moisture value of the mixture and corrects it, regularly comparing the predicted value with the target value, and if the deviation exceeds a certain value, feedback correction is performed. The technical solution of this patent can predict and correct the moisture value of the mixture after training based on known sintering mixed raw materials, which improves the detection accuracy to a certain extent. However, this solution relies on the BP neural network in conjunction with the MIV algorithm for prediction. It has insufficient generalization ability when facing non-steady-state working conditions or new ingredient combinations. In addition, there is an obvious time delay and dynamic relationship between the moisture content of the raw materials, water addition and the final moisture content of the mixture during the sintering process. This patent uses a static regression model to deal with dynamic problems, which makes it difficult to capture the dynamic changes in moisture, and the prediction ability is limited. The accuracy of the prediction results needs to be improved. Summary of the Invention

[0005] In response to the problems of hysteresis and low detection accuracy in the existing sintering mixture moisture prediction, the present invention provides an intelligent prediction method for sintering mixture moisture based on deep learning to meet the demand for precise control of sintering mixture moisture in complex production environments.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides a method for intelligently predicting moisture content of sintering mixtures based on deep learning, comprising the following steps:

[0008] S1. Construct a dataset based on historical data of the mixture water addition control process and divide the dataset into a training set, a validation set, and a test set;

[0009] S2. Construct an MST-Attention network model, which includes a multi-scale feature extraction module, a feature connection module, and a learning and prediction module.

[0010] The multi-scale feature extraction module has a four-layer structure. The output of the previous layer is saved and used as the input of the next layer for convolution processing, and the output is saved;

[0011] The feature connection module is used to perform cross-layer skip connection and channel attention weighted fusion on the output results of each layer of the multi-scale feature extraction module;

[0012] The learning and prediction module embeds the Transformer encoder to extract global temporal dependencies and introduces an attention mechanism to assign weights;

[0013] S3. Set hyperparameters and initialize the network model. Use the training set to train the model until the maximum epoch value is reached, end the training, and save the model.

[0014] S4. Use the validation set to evaluate the model performance. When the coefficient of determination R 2 If ≥0.9, the requirement is met, the training is ended and saved as the optimal model;

[0015] S5. Use the optimal model to predict the test set data, evaluate the performance, and finally output the mixture moisture prediction value.

[0016] Furthermore, before processing the MST-Attention network model, the input features are first processed by a standard convolution to output feature A1.

[0017] Furthermore, when the multi-scale feature extraction module performs processing, the specific processing method is as follows:

[0018] 1) Input feature A1 into the first-layer multi-scale feature extraction module, which uses different-scale convolutional networks to extract features of different scales. Then, the Add layer is used to add the features extracted by the different-scale convolutional networks element by element to achieve feature fusion, save and output feature A2;

[0019] 2) Input feature A2 into the second-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A3;

[0020] 3) Input feature A3 into the third-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A4;

[0021] 4) Input feature A4 into the fourth-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A5.

[0022] Furthermore, the convolution kernel size is [1×1] during standard convolution processing, and the convolution kernel sizes of different scale convolution networks used in each layer of multi-scale feature extraction module are [1×1, 4×1, 7×1].

[0023] Furthermore, before performing the Add layer fusion, a batch normalization layer is added after the convolution layer. The calculation formula is as follows:

[0024]

[0025] Where, is the standardized eigenvalue, x i is a certain dimension eigenvalue of the input data, μ B represents the mean of all samples in the current batch, Represents the variance of all samples in the current batch, and ε is a very small positive value to prevent the denominator from being zero.

[0026] Furthermore, the fusion calculation of the Add layer is performed using the following formula:

[0027]

[0028] Where, They represent the eigenvalues of the i-th channel from the three convolutional networks in MFE, c is the number of channels, and Y add is the output of the Add layer.

[0029] Furthermore, a learnable weight vector is added to the feature connection module, and through softmax normalization processing, the weight of each feature is randomly assigned before connection and integration.

[0030] Furthermore, in step S1, the preprocessing of historical data includes:

[0031] S1.1. Clean historical data;

[0032] S1.2. Perform Min-Max normalization on the cleaned data;

[0033] S1.3. Screen the normalized features to identify those that have a greater impact on the moisture content of the mixture;

[0034] S1.4. Introduce a sliding window mechanism to divide the sliding time window.

[0035] Furthermore, the sliding time window size is 40.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The present invention provides an intelligent prediction method for moisture content of sintered mixtures based on deep learning. The model utilizes the local receptive field and parameter sharing characteristics of the CNN module to extract multi-scale features, capture the local correlation of key variables in moisture prediction, and realizes adaptive modeling of global features through the Transformer module. It mines the long-distance dependency of moisture changes, rather than introducing an attention mechanism to assign weights to features, and highlights key features that are highly relevant to moisture prediction, thereby improving prediction accuracy.

[0038] (2) The present invention's deep learning-based intelligent prediction method for sintering mix moisture content achieves a balance between global trends and local details by optimizing the convolution kernel size of the multi-scale feature extraction module (MFE). This balances the extraction capabilities of global features and local details, thereby significantly improving model performance. Furthermore, the number of layers in the multi-scale feature extraction module (MFE) is designed, preferably four layers, to achieve an optimal balance in the model's feature extraction capabilities, enabling sufficient data fit while avoiding overfitting.

[0039] (3) The present invention provides an intelligent prediction method for moisture content of sintered mixtures based on deep learning. The sliding window mechanism is introduced and the sliding window size is determined, preferably 40. This can well adapt to the local characteristics of moisture-related data of sintered mixtures and find a suitable balance point, so that the model can simultaneously capture the global trend and local changes of moisture-related data, while also taking into account the amount of calculation. This is beneficial to greatly reduce the amount of calculation while ensuring the accuracy of the prediction results, and significantly improve the efficiency of the algorithm.

[0040] (4) The present invention provides an intelligent prediction method for the moisture content of sintering mixtures based on deep learning. By optimizing the model structure and operation process, the method can effectively predict the moisture content of the mixture with high prediction accuracy, which is beneficial to improving the stability of the mixture mixing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the operation of the sintering mixture moisture prediction model of the present invention;

[0042] Figure 2 This is the MST-Attention network framework diagram of the present invention;

[0043] Figure 3 When the MFE module of the present invention is set with different MFE layers, the model R 2 Comparison chart of value, RMSE value and MAE value;

[0044] Figure 4 When the MFE module of the present invention sets different convolution kernel sizes, the model R 2 Comparison chart of value, RMSE value and MAE value;

[0045] Figure 5 When different sliding window sizes are set for the present invention, the R 2 Comparison chart of value, RMSE value and MAE value;

[0046] Figure 6 This is a comparison chart of the processing results when the model of the present invention, the SVR model, and the random forest model are used to process the same sample data set. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0048] The mixture moisture prediction model used in the prior art factories usually only makes a rough estimate based on the raw material moisture, ingredient amount and water addition amount, or relies on single-scale feature extraction, which is difficult to meet the demand for precise control in complex production environments. At the same time, it is inefficient when processing time series data with long-term dependencies. To solve these problems, the present invention provides an intelligent prediction method for sintering mixture moisture based on deep learning, combined with Figure 1 , the prediction method includes the following steps:

[0049] Step S1: Collect historical data of the mixture mixing process and perform preprocessing. After preprocessing, the data is randomly divided into three parts according to a certain ratio, namely, a training set, a validation set and a test set.

[0050] The dataset collected by this method includes actual operating data from the long-term mixing process of the mixture, which can comprehensively reflect the operating status of the mixer under different operating conditions. The actual operating data collected includes raw material composition, moisture content, and equipment operating parameters, all of which are obtained through real-time sensors, automated control systems, or manual sampling.

[0051] When preprocessing the collected data set, considering the possible impact of the mixer operating under various working conditions on the original data sequence, the front-end method is used to fill the blank data to complete the preliminary data cleaning.

[0052] Subsequently, the data was processed using the Min-Max normalization method. The normalization method is shown in the following formula:

[0053]

[0054] Where, represents the normalized ith feature scaling value, which is between 0 and 1; X ij represents the jth original data of the i-th feature, X imin Represents the minimum value in the original data of the i-th feature, X imax Represents the maximum value in the original data of the i-th feature.

[0055] Because hybrid machine datasets suffer from high-dimensional feature redundancy and noise interference, this paper performs feature screening on the normalized data. This paper uses the Person correlation analysis method to analyze the correlation between input and output parameters. When the correlation between the input and output parameters is extremely low, features that contribute little to the model or are irrelevant are directly eliminated, reducing feature dimensionality, thereby reducing model complexity and improving training efficiency.

[0056] When there is a strong correlation between multiple input parameters, it indicates that these three input parameters are redundant in the overall prediction. To simplify the model and avoid multicollinearity problems, only one input feature with the highest correlation with the output parameter is retained. The above method is used to screen the control parameters related to the moisture control process of the sintering mixture. The parameters that enter the model for training after screening are shown in the following table:

[0057]

[0058] Finally, the sliding window mechanism is introduced to divide the sliding time window.

[0059] The primary purpose of introducing sliding time windows is to capture local dynamic features and temporal dependencies in time series. Sliding windows decompose long time series into multiple short time segments, allowing the model to focus on learning local features and improving its ability to capture complex temporal patterns. Sliding window processing relies on two key parameters: the window length W and the sliding step size S. For a time series of length N, with a window size of W and a step size of S, the sliding window can be split into NW / S+1 time windows.

[0060] It is worth noting that the sliding window size in the present invention is preferably 40, which can effectively capture the characteristics of time series data. The applicant established prediction models with sliding window sizes of 10, 20, 30, 40, 50, and 60 (only sliding window size), and compared and analyzed their prediction performance. Figure 5 As shown in the figure, the sliding window size is preferably 40, and the R 2 It reaches a peak of 0.9009, and RMSE is at its lowest. At this time, the performance of the model reaches the optimal state.

[0061] S2. Process the training set data by establishing an MST-Attention network model and initialize the weights. Specifically, the MST-Attention network framework constructed by the present invention is as follows: Figure 2 As shown, it includes a multi-scale feature extraction module MFE (for feature extraction and fusion), a feature connection module FCB (connecting features at different levels), and a learning and prediction module LPB (for learning long-term dependencies of time series and performing mixture moisture prediction).

[0062] Specifically, the multi-scale feature extraction module (MFE) in the present invention is provided with four layers. The core function of the multi-scale feature extraction module MFE is to extract multi-scale features. When the number of layers is small, the feature extraction capability of the model is insufficient and it is difficult to fully capture the complex patterns in the data. As the number of MFE layers increases, the model can extract richer features and capture multi-scale information in the data. At 4 layers, the feature extraction capability of the model reaches the best balance, which can fully fit the data while avoiding overfitting. However, when the number of MFE layers increases further, the model may extract too many redundant features and noise. Redundant features will increase the complexity of the model and lead to overfitting of the training data, while noise features will interfere with the learning of the model and reduce its generalization ability. The applicant found through a large number of experimental studies that the MST-Attention model with a 4-layer structure achieved the optimal RMSE value (0.0705), and the R 2 Reached 0.9009, as shown in the following example: Figure 3 As shown in Figure 3, it shows that the MST-Attention model with 4-layer MFE modules performs best in capturing the global trends and local details of moisture content features.

[0063] Before processing the MST-Attention network model, the input features are first subjected to a standard convolution process with a convolution kernel size of [1×1], and the output feature A1. When using the multi-scale feature extraction module for processing, follow the following process:

[0064] 1) Input feature A1 into the first-layer multi-scale feature extraction module, which uses different-scale convolutional networks to extract features of different scales. Then, the Add layer is used to add the features extracted by the different-scale convolutional networks element by element to achieve feature fusion, save and output feature A2;

[0065] 2) Input feature A2 into the second-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A3;

[0066] 3) Input feature A3 into the third-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A4;

[0067] 4) Input feature A4 into the fourth-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A5.

[0068] Furthermore, in step 1), the calculation formula of the convolutional neural network used is as follows, and padding is used to keep the feature dimension unchanged after convolution.

[0069] y=ReLU(x l-1 ·ω l +b l )

[0070] Where y is the output of the current layer and the input of the next layer, Relu is the activation function of the convolutional layer, which is defined as ReLU(z)=max(0,z), x l-1 is the output of the previous layer, ω l represents the weight matrix of the current layer, b l is the bias vector of the current layer.

[0071] In addition, the applicant found through research that when performing convolution in each layer of the multi-scale feature extraction module MFE, the convolution kernel size [1*4*7] is the best combination, balancing the global trend and local details. In the experiment, when setting different convolution kernel sizes, the prediction model R 2 The comparison chart of the value and RMSE value is as follows Figure 4 As shown in Figure 2, the convolution kernel combination [1*4*7] achieved the best RMSE value (0.0738), while R 2The value reached 0.8724, indicating that this combination performs best in capturing the global trend and local details of moisture content characteristics. The second is the combination [1*2*6], whose RMSE value is 0.0739, and its performance is close to the best. The RMSE value of the combination [1*6*7] is 0.0769, R 2 The value is 0.8616, which is the worst performance. This may be due to the fact that the larger convolution kernel size is not sensitive enough to capture local details, resulting in a decrease in model performance. As can be seen from the figure, R 2 The trend of change is basically consistent with that of RMSE, indicating that the improvement in model performance is closely related to the choice of convolution kernel combination. Larger convolution kernels can extract large-scale features and reflect the overall trend of moisture content, but may ignore details. Smaller convolution kernels focus more on local details but may not fully capture global features. The optimal combination [1*4*7] shows that a moderate convolution kernel size can achieve a balance between global trends and local details. The larger or smaller the convolution kernel size, the better the model performance. This may be because different convolution kernels have different sensitivities to features. In summary, the experimental results show that the convolution kernel combination [1*4*7] performs best in the moisture content feature extraction task, effectively balancing the extraction capabilities of global features and local details, thereby significantly improving model performance.

[0072] More optimally, in order to reduce initialization sensitivity, speed up network convergence, improve operation efficiency and improve training quality, a batch normalization layer is added after the convolution layer. Its calculation formula is as follows:

[0073]

[0074] Where, is the standardized eigenvalue, x i is a certain dimension eigenvalue of the input data, μ B represents the mean of all samples in the current batch, Represents the variance of all samples in the current batch, and ε is a very small positive value to prevent the denominator from being zero.

[0075] Then, the Add layer is used to add the features extracted by different convolutional networks element by element to achieve feature fusion. The fusion calculation formula of the Add layer is as follows:

[0076]

[0077] Where, They represent the eigenvalues of the i-th channel from the three convolutional networks in MFE, c is the number of channels, and Y add is the output of the Add layer. The above root mean square error RMSE and determination coefficient R 2 The mean absolute error (MAE) is used to evaluate the model performance, and the characterization formula is as follows:

[0078]

[0079]

[0080] Where, Indicates the predicted value of mixture moisture, y i Indicates the actual value of the mixture moisture.

[0081] Then, the features A2, A3, A4, and A5 extracted by the 4-layer multi-scale feature extraction module MFE are connected and integrated using the feature connection module FCB to integrate multi-scale and multi-dimensional features. The formula is as follows:

[0082] F output =Concat(F1, F2, F3, F4)

[0083] Where, F outputt Concat represents the feature concatenation operation, which combines the output features of the concatenation layer. F1, F2, F3, and F4 represent the features extracted by the four parallel convolutional networks, namely Feature A2, Feature A3, Feature A4, and Feature A5 mentioned above.

[0084] Combine Figure 2 The method of the present invention makes up for the deficiency of the traditional multi-scale feature extraction module that can only extract features from a single dimension by introducing a feature connection module, and fully integrates the feature information of different scales and dimensions to capture the multi-dimensional feature representation of the moisture content of the mixture.

[0085] Furthermore, since some moisture-related feature information may be weakened after processing by each MFE module layer, this operation can further filter and supplement features, thereby strengthening the expression of moisture-related features and achieving multi-scale and multi-dimensional feature fusion. This design not only retains the refined features extracted by each MFE module, but also enhances the feature expression capability through additional convolution, providing richer and more comprehensive feature support for subsequent modules.

[0086] At the same time, more importantly, the feature connection module FCB also introduces a learnable weight vector and normalizes it through softmax, so that the model can dynamically adjust the weight of each branch. The code and comments for performing the above operations are as follows:

[0087] Perform conv1 convolution and ReLU activation on the input x

[0088] x = F.relu(self.conv1(x))

[0089] The first set of convolution operations uses convolution layers with different kernel sizes to process x

[0090] x1_1 = self.conv1_1(x) # Apply conv1_1 convolution directly to x

[0091] x_padded3_1=F.pad(x,(1,0))#Pad the left side of x by 1 unit

[0092] x1_4 = self.conv1_4(x_padded4_1) # Apply conv1_4 convolution to the padded x

[0093] x1_7 = self.conv1_7(x) # Apply conv1_7 convolution directly to x

[0094] Add the results of the three convolutions to get x1

[0095] x1=x1_1+x1_4+x1_7

[0096] The second set of convolution operations has the same structure as the first set, but the input is x1

[0097] x2_1 = self.conv2_1(x1) # Apply conv2_1 convolution to x1

[0098] x_padded3_1=F.pad(x1,(1,0))#Pad the left side of x1 by 1 unit

[0099] x2_4 = self.conv2_4(x_padded3_1) # Apply conv2_4 convolution to padded x1 x2_7 = self.conv2_7(x1) # Apply conv2_7 convolution to x1

[0100] The three convolution results are added together to get x2

[0101] x2=x2_1+x2_4+x2_7

[0102] The third set of convolution operations, the input is x2

[0103] x3_1 = self.conv3_1(x2) #conv3_1 convolution

[0104] x_padded3_1=F.pad(x2,(1,0))#Pad 1 unit on the left

[0105] x3_4 = self.conv3_4(x_padded3_1) #conv3_3 convolution

[0106] x3_7 = self.conv3_7(x2) #conv3_7 convolution

[0107] Add the three convolution results and get x3

[0108] x3=x3_1+x3_4+x3_7

[0109] The fourth set of convolution operations, the input is x3

[0110] x4_1 = self.conv4_1(x3) #conv4_1 convolution

[0111] x_padded3_1=F.pad(x3,(1,0))#Pad 1 unit on the left

[0112] x4_4 = self.conv4_4(x_padded3_1) #conv4_4 convolution

[0113] x4_7 = self.conv4_7(x3) #conv4_7 convolution

[0114] The three convolution results are added together to get x4

[0115] x4=x4_1+x4_3+x4_5

[0116] Process x1, x2, x3, and x4 through the same conv1_2_3 convolutional layer x123_1 = self.conv1_2_3(x1)

[0117] x123_2=self.conv1_2_3(x2)

[0118] x123_3 = self.conv1_2_3(x3)

[0119] x123_4 = self.conv1_2_3(x4)

[0120] Add the four processed results to get the final feature xx = x123_1 + x123_2 + x123_3 + x123_4

[0121] Get the batch size, number of channels, and sequence length of x

[0122] batch_size,_,seq_len=x.shape

[0123] Adjust the dimension order to (batch_size, seq_len, channels) for easy subsequent processing x = x.permute(0,2,1)

[0124] Flatten the tensor into two dimensions, i.e. (batch_size, seq_len*channels)

[0125] x = x.reshape(batch_size, -1)

[0126] Feature transformation is performed sequentially through three fully connected layers

[0127] x = self.dense1(x)

[0128] x = self.dense2(x)

[0129] x = self.dense3(x)

[0130] Return the final result

[0131] return x

[0132] The above processing can better capture the information brought by different convolution operations and further improve the overall performance of the model. This feature fusion strategy can effectively integrate multi-dimensional feature information, significantly enhance the prediction ability of the model, and establish a connection between the multi-scale feature extraction module MFE module and the learning and prediction module LPB, effectively improving the accuracy of moisture prediction.

[0133] Finally, the connected and fused features are input to the encoder in the learning and prediction module for processing. The encoder layer of the present invention includes a Transformer module and an Attention module. The attention layer of the Transformer module mainly has two structures: a self-attention module and a multi-head attention module, which are as follows:

[0134] Self-attention module: The input matrix X is respectively combined with the weight matrix W Q 、W K and W V Multiply them together to get the query vector (Query, Q), the key vector K (Key, K) and the content vector (Value, V). The calculation formula for the scaled dot product is as follows:

[0135]

[0136] In the formula, softmax normalizes the score of each row to obtain the attention weight matrix; The scaling factor is used to alleviate the gradient vanishing problem caused by excessive dot product values. T Represents the calculation of the similarity between the query and the key, with an attention matrix of shape (N,N).

[0137] Multi-head attention module: It is composed of multiple single-head attention heads, each of which represents a scaled dot product. Due to their different emphasis on correlation, the weights obtained during training are also different. The multi-head self-attention mechanism focuses more on the inherent connections of the data itself. The input sequence with position information is fed into the multi-head self-attention mechanism as Q, K, and V. It efficiently processes the long-term dependencies of time series data through parallel computing. Its calculation formula is as follows:

[0138] MultiHead(Q,K,V)=Concat(head1,…,head h )W O

[0139] head i =Attention(Q i ,K i ,V i )

[0140] Where h is the number of heads in the multi-head self-attention mechanism, head i represents the output of the i-th attention head, Qi, Ki, and Vi are the query, key, and value vectors of the i-th head, respectively, and their dimensions are controlled by the model hyperparameters; Wo is the weight matrix, which is used to linearly transform the concatenated output back to the original feature dimension to ensure that the merged information is still suitable for subsequent processing by the model.

[0141] By adding the Attention mechanism to the CNN-Transformer module, the addition of the Attention mechanism can not only make up for the local receptive field limitation of the convolutional neural network (CNN), but also give play to the unique advantages of the Transformer in global feature modeling, thereby effectively improving the model performance.

[0142] The core of the Attention module lies in capturing global dependencies and enhancing the representation of key features through adaptive weight assignment. Specifically, the self-attention mechanism of the Transformer module can effectively model complex global features, but its efficiency in capturing local details (such as short-term patterns or sudden changes) is relatively low. Therefore, the introduction of the Attention mechanism into the Transformer module significantly improves its ability to express small-scale features, compensating for the limitations of the self-attention mechanism. By combining the strengths of CNN and Transformer, the model achieves comprehensive integration of multi-scale features. The CNN's efficiency in extracting local patterns and the Transformer's ability to accurately model global dependencies complement each other, enabling the model to simultaneously focus on local details and global features. This design not only improves the model's adaptability and robustness in complex data analysis tasks, but also significantly enhances overall performance. In particular, in the task of mixture moisture prediction, the Attention mechanism can spontaneously assign greater weight to key features, highlighting feature information closely related to moisture prediction.

[0143] Finally, the knowledge learned by the encoder layer is mapped to the predicted value of the mixture moisture through a linear regression layer containing one neuron. The calculation formula of the linear layer is as follows:

[0144] Y=σ(x op w+b)

[0145] Where x op is the output of the encoder layer, w is the weight matrix, b is the bias vector, σ is the sigmoid activation function, and Y is the predicted value.

[0146] S3. Set hyperparameters and initialize the network. Then use the constructed MST-Attention network model for training until the maximum epoch value is reached, end the training, and save the model.

[0147] Specifically, the present invention uses grid search to determine the optimal hyperparameters, as shown in the following table:

[0148] parameter describe Maximum number of cycles 100 Batch size 512 Window length T 40 L2 regularization 0.001 <![CDATA[d model ]]> 512 num_heads 128 CNN-num_layers 4 Transformer-num_layers 5 Transformer-layers 1 Learning rate 0.0085

[0149] S4. Use the validation set to evaluate the model performance. When the coefficient of determination R 2 If ≥0.9, the requirement is met, the training is ended and saved as the optimal model;

[0150] S5. Use the optimal model to predict the test set data, evaluate the performance, and finally output the mixture moisture prediction value.

[0151] Example 1

[0152] Historical data from the mixture mixing process, totaling 100,001 data sets, was collected. 5,000 data sets covering various operating conditions were selected as the experimental set, divided into training, validation, and test sets in an 8:1:1 ratio. This dataset, comprised of actual operating data from long-term continuous production, comprehensively reflects the mixer's operating status under various operating conditions. The collected data, including raw material composition, moisture content, and equipment operating parameters, was obtained through real-time sensors, automated control systems, and manual sampling.

[0153] In order to verify the accuracy of the model of the present invention, the optimal model saved by the present invention is compared with the SVR and Radom Forest models. The processing results are as follows: Figure 6 shown.

[0154] SVR model parameters: kernel function is radial basis function (rbf), penalty coefficient C = 1.0, kernel function parameter γ = scale, tolerance error ∈ = 0.1, convergence tolerance = 0.001, and maximum number of iterations is -1.

[0155] Random forest model parameters: the number of trees is 100, the maximum depth is unlimited, the maximum number of features is 1.0, the minimum number of sample splits is 2, the minimum number of leaf node samples is 1, the random seed is 42, and the squared error is used as the splitting criterion.

[0156] Compared with other moisture prediction methods, the MST-Attention model's significant improvement in error hit rate demonstrates its ability to better adapt to the nonlinear characteristics and temporal dependencies of complex operating conditions. Traditional methods typically struggle to achieve a high hit rate within an error of 1%, while the proposed model achieves hit rates of 72.3%, 91.6%, and 99.4% within error ranges of 1%, 2%, and 3%, respectively, significantly improving moisture prediction accuracy.

[0157] Random Forest performs best in MAE, but its R 2 The poor RMSE index indicates that it is inferior to the deep learning method in terms of overall prediction accuracy and is more suitable for tasks that require high local errors but low overall fitting accuracy. SVR has the worst performance, R 2 The lowest and the highest RMSE, indicating that it is not suitable for handling the complex nonlinear characteristics of this task.

[0158] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently predicting moisture content in sintering mixtures based on deep learning, characterized in that: The steps include: S1. Construct a dataset based on historical data of the mixture water addition control process and divide the dataset into a training set, a validation set, and a test set; S2. Construct an MST-Attention network model, which includes a multi-scale feature extraction module, a feature connection module, and a learning and prediction module. The multi-scale feature extraction module has a four-layer structure. The output of the previous layer is saved and used as the input of the next layer for convolution processing, and the output is saved; The feature connection module is used to perform cross-layer skip connection and channel attention weighted fusion on the output results of each layer of the multi-scale feature extraction module; The learning and prediction module embeds the Transformer encoder to extract global temporal dependencies and introduces an attention mechanism to assign weights; S3. Set hyperparameters and initialize the network model. Use the training set to train the model until the maximum epoch value is reached, end the training, and save the model. S4. Use the validation set to evaluate the model performance. When the coefficient of determination R 2 If ≥0.9, the requirement is met, the training is ended and saved as the optimal model; S5. Use the optimal model to predict the test set data, evaluate the performance, and finally output the mixture moisture prediction value.

2. The intelligent prediction method for sintering mixture moisture according to claim 1, characterized in that: Before processing the MST-Attention network model, the input features are first processed by standard convolution to output feature A1.

3. The intelligent prediction method for sintering mixture moisture according to claim 2, characterized in that: When the multi-scale feature extraction module performs processing, the specific processing method is as follows: 1) Input feature A1 into the first-layer multi-scale feature extraction module, which uses different-scale convolutional networks to extract features of different scales. Then, the Add layer is used to add the features extracted by the different-scale convolutional networks element by element to achieve feature fusion, save and output feature A2; 2) Input feature A2 into the second-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A3; 3) Input feature A3 into the third-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A4; 4) Input feature A4 into the fourth-layer multi-scale feature extraction module and perform the same operation as in step 1), save and output feature A5.

4. The intelligent prediction method for sintering mixture moisture according to claim 3, characterized in that: The convolution kernel size is [1×1] during standard convolution processing, and the convolution kernel sizes of different scale convolution networks used in each layer of multi-scale feature extraction module are [1×1, 4×1, 7×1].

5. The intelligent prediction method for sintering mixture moisture according to claim 3, characterized in that: Before performing Add layer fusion, add a batch normalization layer after the convolution layer. The calculation formula is as follows: Where, is the standardized eigenvalue, x i is a certain dimension eigenvalue of the input data, μ B represents the mean of all samples in the current batch, Represents the variance of all samples in the current batch, and ε is a very small positive value to prevent the denominator from being zero.

6. The intelligent prediction method for sintering mixture moisture according to claim 3, characterized in that: The fusion calculation of the Add layer is performed using the following formula: Where, They represent the eigenvalues of the i-th channel from the three convolutional networks in MFE, c is the number of channels, and Y add is the output of the Add layer.

7. The method for intelligently predicting moisture content of sintering mixture according to any one of claims 1 to 6, characterized in that: A learnable weight vector is added to the feature connection module, and after softmax normalization, the weight of each feature is randomly assigned before connection and integration.

8. The method for intelligently predicting moisture content of sintering mixture according to any one of claims 1 to 6, characterized in that: In step S1, the preprocessing of historical data includes: S1.

1. Clean historical data; S1.

2. Perform Min-Max normalization on the cleaned data; S1.

3. Screen the normalized features to identify those that have a greater impact on the moisture content of the mixture; S1.

4. Introduce a sliding window mechanism to divide the sliding time window.

9. The intelligent prediction method for sintering mixture moisture content according to claim 8, characterized in that: The sliding time window size is 40.

Images