Rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement

Abstract

The invention provides a rotary kiln sintering temperature prediction method based on improved multi-scale WGAN industrial rotary kiln data enhancement, and the method comprises the steps: constructing an MSC-WGAN network model which comprises a generator and a discriminator; the generator is used for processing the input noise data to generate rotary kiln sintering data; in the generator, feature extraction is carried out based on a multi-scale depth separable convolution structure, and a channel attention mechanism is introduced to enhance the expression ability of different feature channels; and discriminating the generated industrial data and the original rotary kiln sintering data by using the discriminator in combination with a loss function formed by a Wasserstein distance, a gradient penalty term and a root-mean-square error term so as to train the MSC-WGAN network model. And generating final rotary kiln sintering data by using the trained MSC-WGAN network model, combining the final rotary kiln sintering data with original rotary kiln sintering data, and inputting the combined rotary kiln sintering data into a long short-term memory network, so as to predict the sintering temperature of the rotary kiln. According to the method, the calculation cost and parameter quantity of the model are reduced, and meanwhile, the data generation quality is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial data processing, and in particular to a rotary kiln sintering temperature prediction method based on data enhancement. Background Art

[0002] In the operation and optimization of industrial rotary kilns, data generation and analysis technologies are of great significance for fault detection, process optimization, and prediction of key parameters (such as sintering temperature). Industrial processes involve a large amount of high-dimensional, multi-scale, nonlinear, and time-series related data. These data are often affected by noise interference and data loss, and suffer from sample imbalance problems (such as scarce fault data). With the large-scale development of process industries, industrial production systems are becoming increasingly complex. The use of soft measurement technology to monitor the working conditions and predict quality parameters of production processes is crucial to improving the stability of complex industrial processes. In recent years, driver-based soft measurement technology has been widely used in process industries. However, the performance of data-driven models depends on a large amount of data. Only sufficient data can guarantee the accuracy of predictions. However, due to objective conditions, it is difficult to obtain sufficient data in industrial production.

[0003] The invention patent with publication number CN117891289A discloses a method for predicting the probability interval of sintering temperature of a rotary kiln. By combining a parallel multi-head self-attention DA-LSTM-GPR model with Gaussian process regression, it solves the problem that the sintering temperature prediction model in the existing technology is difficult to explore deep structure and quantify uncertainty. However, it uses raw data for prediction and does not introduce data enhancement technology. When faced with common problems such as sample imbalance and noise interference in industrial data, the model is prone to poor prediction performance due to insufficient data or poor quality. It relies on mRMR (minimum redundancy maximum relevance) for manual feature screening, and requires manual intervention to select input variables (such as kiln head temperature, blast flow rate, etc.). This process increases the complexity of model deployment and may lead to feature selection bias due to insufficient domain knowledge. In addition, a parallel multi-head self-attention mechanism is introduced, which requires maintaining multiple attention heads (Q, K, V matrices) and using Gaussian process regression (GPR) to output probability intervals. When processing large-scale data, the training and inference speed of GPR will drop sharply, resulting in a significant increase in the number of parameters and computational complexity. Summary of the Invention

[0004] In response to the technical problems that existing methods have poor ability to extract spatiotemporal features from complex industrial process data, resulting in limited diversity and quality of generated data and complex calculations, the present invention proposes a rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement, which significantly reduces the computational cost and parameter amount of the model while ensuring the quality of data generation.

[0005] In order to achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] A rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement is characterized by comprising the steps of:

[0007] S1: Acquire real rotary kiln sintering data, and pre-process the real rotary kiln sintering data to obtain original rotary kiln sintering data;

[0008] S2: Build the MSC-WGAN network model, including the generator and discriminator;

[0009] S3: The generator is used to process the input noise data to generate rotary kiln sintering data; the generator extracts features from the noise data based on a multi-scale depth-separable convolutional structure, and introduces a channel attention mechanism to enhance the expression capabilities of different feature channels;

[0010] S4: using the discriminator in combination with a loss function consisting of Wasserstein distance, gradient penalty term, and root mean square error term to discriminate the generated rotary kiln sintering data and the original rotary kiln sintering data to train the MSC-WGAN network model;

[0011] S5: The trained MSC-WGAN network model is used to generate the final rotary kiln sintering data and combined with the original rotary kiln sintering data to input the long short-term memory network to predict the rotary kiln sintering temperature.

[0012] The method of using the generator to process the input noise data to generate rotary kiln sintering data is as follows: taking a random noise tensor as input, using the multi-scale depth-separable convolution structure to perform multi-scale feature extraction to obtain the output feature map of the multi-scale separable convolution, using the channel attention mechanism to dynamically adjust the importance of each channel in the output feature map of the multi-scale depth-separable convolution, and obtaining an enhanced feature map; the enhanced feature map is output after passing through the flattening layer and the dense layer to generate the generated rotary kiln sintering data.

[0013] The multi-scale depth-wise separable convolution structure includes multiple depth-wise separable convolution layers of different scales. Convolution kernels of different sizes are used in the depth-wise separable convolution layer of each scale, and features of the current scale are extracted based on depth-wise convolution operations and point-wise convolution operations. Multiple scales are spliced based on feature map stacking operations to obtain features of multiple time scales in random noise.

[0014] The depth convolution operation is:

[0015] The input random noise tensor is Where T is the sequence length and M is the number of features. The depthwise separable convolutional layers at each scale use different convolution kernels to independently convolve the random noise tensor, thereby extracting features of multiple time scales in the depthwise separable convolutional layers at different scales:

[0016]

[0017] in, Represents the output feature map after depth convolution, c represents the channel number, m represents the feature number, t' represents the time position after depth convolution, is the i-th weight value in the depth convolution kernel, Indicates the time position as t i The eigenvalue of K c is the size of the convolution kernel, and * is the convolution operation.

[0018] The point convolution operation is as follows: combining all the feature maps output at each scale during depthwise convolution, performing 1×1 convolution using a point convolution kernel, changing the number of features from M to N, and calculating the output feature map at each scale:

[0019]

[0020] Among them, F c represents the output feature map of each scale, Represents the point convolution kernel, n is the channel number after point convolution;

[0021] The feature map stacking operation is:

[0022] F MSC =Concat(F 1 ,F 2 ,.....F c );

[0023] in, represents the output feature map of multi-scale depth-wise separable convolution, T 2 , N and C are the sequence length, number of features and number of channels respectively, and Concat(·) is the concatenation operation.

[0024] The channel attention mechanism obtains global semantic information of the channel dimension by performing global average pooling and global maximum pooling on each channel of the output feature map of multi-scale depth-separable convolution; the global semantic information is input into a fully connected network for nonlinear transformation to obtain channel attention weights; the channel attention weights are used to perform weighted summation on the output feature map of multi-scale depth-separable convolution to obtain an enhanced feature map.

[0025] The loss function described in step S4 includes a generator loss based on Wasserstein loss and RMSE loss and a discriminator loss based on Wasserstein distance and gradient penalty term;

[0026] The discriminator includes a multi-scale convolutional structure and a fully connected layer connected in sequence, wherein the scale of the multi-scale convolutional structure corresponds to the scale of the multi-scale depth-separable convolutional structure in the generator; each scale of the multi-scale convolutional structure includes a convolutional layer, a LeakyReLU activation function and a batch normalization layer connected in sequence, and the discriminator adopts a multi-scale convolutional structure symmetrical to the generator.

[0027] The generator loss is:

[0028]

[0029] Among them, k is a weight parameter used to balance the impact of Wasserstein loss and RMSE loss. M is the root mean square error, Loss G is the generator loss, Express expectations, To generate data.

[0030] The discriminator loss is:

[0031]

[0032] in, To generate data expectations, is the original data expectation item, is the gradient penalty term, D(·) represents the discriminator, and Expresses expectation, P g is the generated rotary kiln sintering data distribution, P r This is the original rotary kiln sintering data distribution.

[0033] When training the MSC-WGAN network model, KL divergence, JS divergence, maximum average difference and average Pearson correlation coefficient are used to comprehensively evaluate the quality of the rotary kiln sintering data finally generated by the MSC-WGAN network model.

[0034] Beneficial effects of the present invention:

[0035] Through multi-scale deep separable convolution, the model can simultaneously capture multi-scale features such as short-term fluctuations and long-term trends, fully adapting to the complexity and diversity of industrial data. Furthermore, the introduction of deep separable convolution significantly reduces the model's computational cost and parameter count while ensuring the quality of data generated, making it suitable for real-time industrial applications.

[0036] Based on the traditional WGAN-GP, the RMSE (root mean square error) loss term is introduced to ensure the alignment of the generated data at the distribution level and instance level, thereby improving the numerical accuracy of the generated data and its consistency with the real data.

[0037] Aiming at the complexity of industrial data generation, an auxiliary evaluation strategy is designed. Through comprehensive evaluation of indicators such as KL divergence, JS divergence, MMD and average Pearson correlation, it can comprehensively measure the distribution matching and correlation between variables of the generated data, and ensure the reliability and applicability of the generated data.

[0038] By using generated data to expand the training set, the common sample imbalance problem in industrial data is solved, and the accuracy and generalization ability of the LSTM model for sintering temperature prediction are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 It is a schematic diagram of the overall structure of the present invention.

[0041] Figure 2 Schematic diagram of the generator structure of the present invention.

[0042] Figure 3 Schematic diagram of the structure of the discriminator of the present invention.

[0043] Figure 4 It is a comparison diagram of the normal distribution curves of the original data and the generated data of the present invention.

[0044] Figure 5 This is a PCA visualization diagram of the original data and generated data of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0046] A rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement, such as Figure 1As shown, the steps include:

[0047] S1: Acquire real rotary kiln sintering data, and pre-process the real rotary kiln sintering data to obtain original rotary kiln sintering data.

[0048] In this chapter, real-world individual data were selected to validate the proposed method. Industrial data was collected from the No. 6 rotary kiln of the Aluminum Corporation of China. The rotary kiln is a cylindrical structure approximately 70 meters long and 5 meters in diameter, capable of producing 2,000 tons of clinker per day. The kiln is installed in a slightly inclined position and rotates continuously driven by a motorized system. During operation, a mixed slurry of raw materials (lime, alumina, carbon, etc.) is fed into the kiln from the rear (the kiln tail). Due to the rotation and inclination, the mixed slurry gradually moves toward the front (the kiln head). Simultaneously, pulverized coal is injected into the kiln head and flows toward the kiln tail, providing the necessary heat for the sintering process. The material passes through four key zones in sequence: preheating, decomposition, sintering, and cooling. The final product, clinker, exits the kiln and enters the cooling system. The sintering temperature is the most critical parameter in this process, as it directly affects the quality of the clinker and the efficiency of energy and material consumption.

[0049] Industrial rotary kiln data has significant temporal characteristics, so a sliding window mechanism is used to construct the input of the Generative Adversarial Network (GAN). A sliding window is a fixed-size window that moves stepwise along the time series dataset to capture sequence information for model training. The sliding window is defined by two key parameters: window size (W) and step size (S). Formally, for a given time series spanning 1 to T, the input matrix The construction is as follows:

[0050]

[0051] Specifically, the first seven rows of the matrix represent the time series values of the seven variables in the W-step window, while the eighth row corresponds to the sintering temperature value at each time step. This matrix serves as the input of the model and includes the time dynamics of the process variables and the target variable (sintering temperature).

[0052] In order to verify the effectiveness of the data generation method of the present invention, the sintering temperature prediction of an industrial rotary kiln is selected as an example, where X includes 7 variables, y is the sintering temperature, and the original data consists of 7000 samples divided into 80% and 20% ratios.

[0053] After training the data augmentation model, the generator is used to generate N synthetic samples, where N is equal to the size of the training set (5,600 samples). This ensures that the augmented dataset remains large enough while maintaining its diversity. The final training set consists of 5,600 real samples and 5,600 synthetic samples, for a total of 11,200 samples.

[0054] S2: Construct the MSC-WGAN network model, including the generator and discriminator, such as Figure 1 shown.

[0055] S3: The generator is used to process the input noise data to generate rotary kiln sintering data; the generator extracts features from the noise data based on a multi-scale depth-separable convolution structure, and introduces a channel attention mechanism to enhance the expression capabilities of different feature channels.

[0056] The method of using the generator to process the input noise data to generate rotary kiln sintering data is as follows:

[0057] like Figure 2 As shown, with a random noise tensor as input, the multi-scale depthwise separable convolution (MSDSCNN) structure is used to perform multi-scale feature extraction to obtain the output feature map of the multi-scale separable convolution, and the channel attention mechanism (ChannelAttention) is used to dynamically adjust the importance of each channel in the output feature map of the multi-scale depthwise separable convolution to obtain an enhanced feature map; the enhanced feature map is output after the flattening layer (Flatten) and the dense layer (Denes) to generate the rotary kiln sintering data.

[0058] The multi-scale depth-wise separable convolution structure includes multiple depth-wise separable convolution layers of different scales. Convolution kernels of different sizes are used in the depth-wise separable convolution layer of each scale, and features of the current scale are extracted based on depth-wise convolution operations and point-wise convolution operations. Multiple scales are spliced based on feature map stacking operations to obtain features of multiple time scales in random noise.

[0059] Traditional GANs typically use standard convolutional layers to extract features, which only support fixed scales and are difficult to capture multi-scale features simultaneously. This paper introduces multi-scale depthwise separable convolutions into the generator, combining convolution kernels of different scales to achieve multi-scale feature extraction. In this embodiment, the convolution kernels are 3×3, 5×5, and 7×7.

[0060] The specific calculation process of the multi-scale depth-separable convolution structure is as follows:

[0061] Depthwise convolution: The input random noise tensor is Where T is the sequence length (the total time step is T), and M is the number of features. The depthwise separable convolutional layers of each scale use different convolution kernels to independently convolve the random noise tensor, thereby extracting features of multiple time scales in the depthwise separable convolutional layers of different scales:

[0062]

[0063] in, Represents the output feature map after depth convolution, c represents the channel number, which is also the scale number, m represents the feature number, and t' represents the time position after depth convolution. is the i-th weight value in the depth convolution kernel, Indicates the time position as t i The eigenvalue of K c is the size of the convolution kernel, and * is the convolution operation.

[0064] Point convolution: Combine all the feature maps output at each scale during depth convolution, use the point convolution kernel to perform 1x1 convolution, the number of features changes from M to N, and calculate the output feature map of each scale:

[0065]

[0066] Among them, F c represents the output feature map of each scale, Represents the point convolution kernel, and n is the channel number after point convolution.

[0067] Feature map stacking:

[0068] F MSC =Concat(F 1 ,F 2 ,.....F c )

[0069] in, Represents the output feature map of multi-scale depth-wise separable convolution, T 2 , N, and C are the sequence length, number of features, and number of channels, respectively, and Concat(·) is a concatenation operation. Depthwise separable convolution can significantly reduce the number of parameters while retaining the ability to express features for high-dimensional data. To integrate multi-scale depthwise separable convolution into the generator, the present invention uses a concatenation method to merge the outputs of multi-scale modules. By adopting different kernel sizes, the model can capture both fine-grained and global features, which is crucial for applications involving industrial process data exhibiting multiple temporal or spatial scales.

[0070] The channel attention mechanism obtains global semantic information in the channel dimension by performing global average pooling and global maximum pooling on each channel of the output feature map of the multi-scale depth-separable convolution:

[0071] Specifically, the average pooling descriptor z is generated by aggregating spatial information through the average pooling AvgPool and maximum pooling MaxPool operations. avg and the maximum pooling descriptor z max , represents the global semantic information of the channel dimension:

[0072] z avg=Avgpool(F MSC );

[0073] z max =Maxpool(F MSC );

[0074] The present invention dynamically adjusts the importance of each channel through average pooling and maximum pooling operations, enhances the expression ability of key information, and suppresses noise and redundant features.

[0075] The global semantic information is input into the fully connected network for nonlinear transformation to obtain the channel attention weight:

[0076] L c =σ(MLP(z avg )+MLP(z max ));

[0077] Here, MLP(·) represents a fully connected network and σ(·) represents the Sigmoid activation function used to normalize the weights.

[0078] The output feature map F of the multi-scale depth-separable convolution is processed using the channel attention weights MSC Perform weighted summation to obtain the enhanced feature map.

[0079] F MSC =F MSC =FMSC⊙LcL c

[0080] where ⊙ represents element-wise multiplication.

[0081] Industrial data typically has high-dimensional features, where some channels (such as specific sensor data) may be more critical. The multi-topological scale local spatiotemporal feature maps generated by the multi-scale convolutional layer contain local behavioral information at different scales, which has different effects on the prediction output of the final soft sensor model. In order to emphasize useful features and suppress irrelevant noise, the present invention introduces a channel attention mechanism (CAM) in the generator. CAM adaptively assigns importance weights to each feature channel, enabling the model to focus on key temporal features extracted at multiple scales. This mechanism enhances feature representation and improves the fidelity of generated data by dynamically recalibrating the channel importance of feature maps.

[0082] S4: Using the discriminator in combination with a loss function consisting of Wasserstein distance, gradient penalty term, and root mean square error term, the generated rotary kiln sintering data and the original rotary kiln sintering data are discriminated to train the MSC-WGAN network model. The loss function includes a generator loss based on Wasserstein loss and RMSE loss, and a discriminator loss based on Wasserstein distance and gradient penalty term.

[0083] The discriminator loss is:

[0084]

[0085] in, To generate data expectations, is the original data expectation item, is the gradient penalty term, is the generated data, that is, the generated rotary kiln sintering data, D(·) represents the discriminator, and Expresses expectation, P g is the generated rotary kiln sintering data distribution, P r This is the original rotary kiln sintering data distribution.

[0086] The root mean square error term is introduced into the generator loss, and the generator loss is:

[0087]

[0088] Among them, k is a weight parameter used to balance the impact of Wasserstein loss and root mean square error (RMSE) loss. M is the root mean square error, Loss G is the generator loss.

[0089] In this paper, the Wasserstein distance is used as the loss function to solve the problems of training instability and mode collapse in traditional GAN. The gradient penalty (GP) term is introduced to ensure that the discriminator satisfies the 1-Lipschitz continuity constraint, further enhancing training stability. The Wasserstein loss of traditional WGAN-GP focuses on the distribution matching of generated data and real data, but cannot guarantee the numerical accuracy of generated data. To this end, the present invention adds a mean square error (MSE) term to the loss function, which further improves the numerical accuracy of generated data by directly comparing the values of generated data with real data.

[0090] By minimizing the RMSE between generated data and real data, the quality of the generated data can be improved, making the generated data closer to the real data in terms of statistical characteristics. The addition of MSE helps stabilize the training process because it provides a smoother optimization target, which helps avoid the problem of vanishing or exploding gradients. Because the generator maximizes the discriminator's score while minimizing MSE, this can encourage the discriminator to learn richer feature representations, thereby improving its ability to distinguish between real data and generated data. By adjusting the weight parameter k, the influence of MSE in the loss function can be flexibly controlled to adapt to different training stages or dataset characteristics.

[0091] like Figure 3 As shown in Figure 1, the discriminator extracts and learns the features of the input data mainly through a multi-scale convolution structure (CONV 1D), a batch normalization layer (BatchNormalization), and a LeakyReLU activation function.

[0092] The discriminator includes a multi-scale convolutional structure and a fully connected layer connected in sequence, where the scale of the multi-scale convolutional structure corresponds to the scale of the multi-scale depth-wise separable convolutional structure in the generator. Each scale of the multi-scale convolutional structure includes a convolutional layer, a LeakyReLU activation function, and a batch normalization layer connected in sequence. After the convolutional layer, the LeakyReLU activation function is applied to the output, and then enters the batch normalization layer. The batch normalization layer is used to standardize the output of each layer to avoid gradient explosion or vanishing and improve training stability. The discriminator adopts a multi-scale convolutional structure symmetrical to the generator, ensuring that the discriminator can effectively recognize multi-scale features.

[0093] In this embodiment, the specific processing process in the discriminator is: taking the generated rotary kiln sintering data and the original rotary kiln sintering data as input respectively, using the three convolutional structures to gradually extract the spatial features of the input data, and after the extracted spatial features are evaluated in the fully connected layer, outputting a scalar representing the authenticity of the input data through the dense layer.

[0094] The KL divergence, JS divergence, maximum mean difference, and average Pearson correlation coefficient were used to comprehensively evaluate the quality of the rotary kiln sintering data generated by the MSC-WGAN network model. These were used after training to directly assess the distribution differences between the generated data and the real data.

[0095] KL divergence D KL : It is used to measure the difference between the generated data and the real data distribution. The formula is as follows:

[0096]

[0097] Among them, P r(x) is the real data distribution, P g (x) is the generated data distribution.

[0098] JS Divergence D JS : As a symmetric form of KL divergence, it more stably reflects the distribution overlap. The formula is as follows:

[0099]

[0100] Where M is the average distribution.

[0101] Maximum average difference D MMD : Calculate the distribution distance between generated data and real data in high-dimensional space through the kernel function. The formula is as follows:

[0102]

[0103] Among them, X and Y are the real data sample set and the generated data sample set respectively, Φ(x) is the mapping function, which maps the original data x to the high-dimensional feature space H, ||·|| H Represents the norm of the high-dimensional feature space H.

[0104] Average Pearson Correlation Coefficient: Evaluates the variable correlation between generated data and real data to ensure that the generated data retains the dependencies between key features. The average Pearson correlation is an extension of the point-by-point Pearson correlation. It calculates the average correlation between generated data and original data between multiple features.

[0105]

[0106] Among them, x i and y i are arbitrary points of generated data and original data respectively, and Divided into the mean of generated data and original data.

[0107] By using these metrics to conduct a multi-dimensional analysis, we ensure that the generated data meets high-quality requirements in terms of distribution matching, feature relevance, and other aspects. Compared to traditional GANs that rely on a single evaluation metric (such as FID), this method comprehensively measures the reliability and applicability of generated data from multiple perspectives, making it particularly suitable for the complex nature of industrial data.

[0108] S5: The trained MSC-WGAN network model is used to generate final rotary kiln sintering data. This data is then combined with the original rotary kiln sintering data and input into a long short-term memory network to implement a method for predicting rotary kiln sintering temperature. Industrial data often suffers from sample imbalance (e.g., scarcity of fault data). This method addresses this data imbalance by generating high-quality synthetic data and expanding the training set for the LSTM model.

[0109] The generated final rotary kiln sintering data was combined with the first 80% of the real-world original rotary kiln sintering data to form the training set. The remaining 20% of the real-world data was used as the test set to train the LSTM network. The trained LSTM model was then tested using the test set to evaluate its predictive performance. The model trained using the generated data was compared with the model trained using only real-world data to see if prediction metrics (root mean square error and mean absolute error) improved. If the generated data significantly improved the model's predictive performance, it indicates that the generated data is highly practical and reliable.

[0110] Finally, the trained LSTM network is used to predict the sintering temperature of the rotary kiln and capture the time dependence and dynamic change rules of the rotary kiln sintering data.

[0111] This invention significantly improves the computational efficiency and feature extraction capabilities of industrial data generation, overcoming the limitations of traditional generative adversarial networks in industrial data processing. This method reduces computational complexity while ensuring data quality, making it suitable for real-time applications in industrial environments and providing effective data support for industrial process optimization and fault diagnosis.

[0112] In this example, the present invention is compared with the existing method.

[0113] The invention uses distribution KL and JS divergence, average Pearson correlation and numerical precision (MMD) to evaluate model performance, and the structure is shown in Table 1. The effect of the baseline GAN is limited, with high KL (0.077) and JS divergence (0.016) and poor feature correlation (0.971). CGAN greatly improves the distribution alignment rate by 55.8% (KL: 0.034) and feature retention rate (correlation: 0.995). WGAN-GP further improves stability through Wasserstein distance and gradient penalty, achieves better distribution indicators (KL: 0.033, JS: 0.008), and establishes a baseline numerical similarity (MMD: 0.00053). The proposed model MSC-WGAN achieved excellent performance in all indicators, with a 6.9% reduction in KL value (0.031), a 5.1% increase in MMD value (0.00051), and a stronger feature correlation (0.997). These improvements stem from multi-scale feature extraction, attention-based feature emphasis, and RMSE loss optimization, making MSC-WGAN an effective solution for industrial data generation.

[0114] Table 1. Model performance evaluation results using distribution KL and JS divergences, average Pearson correlation and numerical accuracy

[0115]

[0116]

[0117] Table 2. Results of systematic evaluation experiments on LSTM models using data augmentation techniques.

[0118]

[0119] To analyze the impact of different data generation methods on sintering temperature prediction results, this paper systematically evaluated the LSTM model using data augmentation techniques. To evaluate the effectiveness of the proposed method, an LSTM model was used as a baseline model. The experiment involved the following steps: First, the LSTM model was trained using only the original training data (without data augmentation) to establish a performance baseline. Then, the same LSTM model was trained using the augmented dataset. By comparing the performance of the two models, the effectiveness of various data augmentation methods was analyzed. The baseline LSTM model used the RMSprop optimizer with a learning rate set to 0.001. It consisted of two LSTM layers, with the first and second layers consisting of 64 and 32 units, respectively. All layers used ReLU as the activation function to capture nonlinear temporal patterns. To reduce overfitting, the model incorporated L2 regularization (coefficient 0.001) and 60% dropout. The input data used a sliding window approach with a window size of 8 and a step size of 1, meaning that each input sequence consisted of 8 consecutive time steps. The experimental results are shown in Table 2. The baseline is an LSTM without data augmentation. As can be seen, the baseline model without data augmentation performs the weakest, with the highest RMSE and MAE values (RMSE: 35.223, MAE: 28.715, R2: 0.869). In comparison, the LSTM model augmented with GAN data shows slight improvement (RMSE: 34.065, MAE: 27.059, R2: 0.877), but its adversarial training mechanism suffers from mode collapse, limiting the diversity of generated data. CGAN alleviates this mode collapse by introducing conditional inputs, significantly improving distribution alignment (RMSE: 31.269, MAE: 25.932, R2: 0.897), but its reliance on conditional guidance limits the flexibility of data generation. By employing the Wasserstein distance and gradient penalty mechanism, WGAN-GP further improves stability and distribution alignment capabilities, achieving even better performance (RMSE: 30.354, MAE: 25.083, R²: 0.920), generating higher-quality data and effectively capturing complex patterns. Ultimately, the proposed model performed best among all methods (RMSE: 27.556, MAE: 22.382, R²: 0.919). These results validate the advantages of this approach for industrial data augmentation and significantly improve the performance of LSTM prediction.

[0120] The normal distribution curve is used to qualitatively evaluate the diversity and authenticity of the generated data. Figure 4 Since the dimensionality of industrial rotary kiln data is high, the present invention adopts principal component analysis (PCA) dimensionality reduction visualization technology to map the original high-dimensional data into a low-dimensional space and extract the two-dimensional data with the largest amount of information. The distribution characteristics of the original data and the generated data can be compared intuitively, such as Figure 5 As shown in the figure, the distribution of the generated data represented by the red dots and the original data represented by the blue dots in two-dimensional space has a high degree of overlap. This shows that the generated data performs well in capturing the main characteristics of the original data, especially in the central area of the data, where the distributions of the two sets of data are almost identical. However, the generated data shows some discreteness in the edge areas, which may be related to the scarcity of boundary samples or the insufficient learning ability of the generative model for extreme samples. Overall, this distribution comparison verifies the high similarity between the generated data and the original data, further demonstrating the effectiveness of the model in generating high-quality industrial data.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement, characterized in that: Including steps: S1: Acquire real rotary kiln sintering data, and pre-process the real rotary kiln sintering data to obtain original rotary kiln sintering data; S2: Build the MSC-WGAN network model, including the generator and discriminator; S3: The generator is used to process the input noise data to generate rotary kiln sintering data; the generator extracts features from the noise data based on a multi-scale depth-separable convolutional structure, and introduces a channel attention mechanism to enhance the expression capabilities of different feature channels; S4: using the discriminator in combination with a loss function consisting of Wasserstein distance, gradient penalty term, and root mean square error term to discriminate the generated rotary kiln sintering data and the original rotary kiln sintering data to train the MSC-WGAN network model; S5: The trained MSC-WGAN network model is used to generate the final rotary kiln sintering data and combined with the original rotary kiln sintering data to input the long short-term memory network to predict the rotary kiln sintering temperature.

2. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 1 is characterized in that: The method of using the generator to process the input noise data to generate rotary kiln sintering data is as follows: taking a random noise tensor as input, using the multi-scale depth-separable convolution structure to perform multi-scale feature extraction to obtain the output feature map of the multi-scale separable convolution, using the channel attention mechanism to dynamically adjust the importance of each channel in the output feature map of the multi-scale depth-separable convolution, and obtaining an enhanced feature map; the enhanced feature map is output after passing through the flattening layer and the dense layer to generate the generated rotary kiln sintering data.

3. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 1 or 2, characterized in that: The multi-scale depth-wise separable convolution structure includes multiple depth-wise separable convolution layers of different scales. Convolution kernels of different sizes are used in the depth-wise separable convolution layer of each scale, and features of the current scale are extracted based on depth-wise convolution operations and point-wise convolution operations. Multiple scales are spliced based on feature map stacking operations to obtain features of multiple time scales in random noise.

4. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 3 is characterized in that: The depth convolution operation is: The input random noise tensor is Where T is the sequence length and M is the number of features. The depthwise separable convolutional layers at each scale use different convolution kernels to independently convolve the random noise tensor, thereby extracting features of multiple time scales in the depthwise separable convolutional layers at different scales: in, Represents the output feature map after depth convolution, c represents the channel number, m represents the feature number, t' represents the time position after depth convolution, is the i-th weight value in the depth convolution kernel, Indicates the time position as t i The eigenvalue of K c is the size of the convolution kernel, and * is the convolution operation.

5. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 4 is characterized in that: The point convolution operation is as follows: combining all the feature maps output at each scale during depthwise convolution, performing 1×1 convolution using a point convolution kernel, changing the number of features from M to N, and calculating the output feature map at each scale: Among them, F c represents the output feature map of each scale, Represents the point convolution kernel, n is the channel number after point convolution; The feature map stacking operation is: F MSC =Concat(F 1 ,F 2 ,.....F c ); in, Represents the output feature map of multi-scale depth-wise separable convolution, T 2 , N and C are the sequence length, number of features and number of channels respectively, and Concat(·) is the concatenation operation.

6. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 1 or 2, characterized in that: The channel attention mechanism obtains global semantic information of the channel dimension by performing global average pooling and global maximum pooling on each channel of the output feature map of multi-scale depth-separable convolution; the global semantic information is input into a fully connected network for nonlinear transformation to obtain channel attention weights; the channel attention weights are used to perform weighted summation on the output feature map of multi-scale depth-separable convolution to obtain an enhanced feature map.

7. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to any one of claims 1, 2, 4 or 5, characterized in that: The loss function described in step S4 includes a generator loss based on Wasserstein loss and RMSE loss and a discriminator loss based on Wasserstein distance and gradient penalty term; The discriminator includes a multi-scale convolutional structure and a fully connected layer connected in sequence, wherein the scale of the multi-scale convolutional structure corresponds to the scale of the multi-scale depth-separable convolutional structure in the generator; each scale of the multi-scale convolutional structure includes a convolutional layer, a LeakyReLU activation function and a batch normalization layer connected in sequence, and the discriminator adopts a multi-scale convolutional structure symmetrical to the generator.

8. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 7 is characterized in that: The generator loss is: Among them, k is a weight parameter used to balance the impact of Wasserstein loss and RMSE loss. M is the root mean square error, Loss G is the generator loss, Express expectations, To generate data.

9. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 8, characterized in that: The discriminator loss is: in, To generate data expectations, is the original data expectation item, is the gradient penalty term, D(·) represents the discriminator, and Expresses expectation, P g is the generated rotary kiln sintering data distribution, P r This is the original rotary kiln sintering data distribution.

10. The rotary kiln sintering temperature prediction method based on improved multi-scale WGAN data enhancement according to claim 1 or 2, characterized in that: When training the MSC-WGAN network model, KL divergence, JS divergence, maximum average difference and average Pearson correlation coefficient are used to comprehensively evaluate the quality of the rotary kiln sintering data finally generated by the MSC-WGAN network model.

Images