Rotary kiln sintering temperature prediction method and system based on multi-scale three-dimensional convolution model, medium and equipment

Abstract

The invention discloses a rotary kiln sintering temperature prediction method and system based on a multi-scale three-dimensional convolution model, a medium and equipment, dynamic prediction of key quality indexes of a rotary kiln is realized, three-dimensional features of multiple scales are generated by utilizing one-dimensional convolution of different scales, then the three-dimensional features are fused into multi-scale three-dimensional features, and the multi-scale three-dimensional features are obtained by utilizing a three-dimensional convolutional neural network (3D Convolutional Neural Network). According to the method, spatial-temporal characteristics in data are extracted through a three-dimensional neural network (3DCNN), the prediction performance and generalization ability of a soft measurement system are improved, meanwhile, three different evaluation indexes are used for evaluating the performance of a model, and more powerful support is provided for intelligent monitoring and optimization of the industrial production process of the rotary kiln.

Description

Technical Field

[0001] The present invention belongs to the technical field of complex rotary kiln industry, and in particular relates to a method, system, medium and equipment for predicting the sintering temperature of a rotary kiln using a multi-scale three-dimensional convolution model. Background Art

[0002] In the complex rotary kiln industrial process, time series problems exhibit distinct temporal and spatial characteristics. By comprehensively considering these two characteristics, we can more comprehensively understand the characteristics and patterns of data in the rotary kiln industrial process. Then, using appropriate data processing and modeling techniques, we can achieve precise control and optimization of the rotary kiln industrial process. The effective application of modeling technology requires the full extraction of spatiotemporal characteristics, because only by fully exploring the spatiotemporal characteristics of the data can we achieve a more comprehensive understanding of the rotary kiln industrial process, thereby achieving more precise control and optimization.

[0003] In the field of time series analysis, CNNs have become crucial due to their superior spatial feature extraction capabilities. CNNs capture local spatiotemporal features in data by applying convolution kernels of varying sizes. The kernel size determines the size of its receptive field, which in turn influences the scale of the extracted features. Leveraging local connectivity and a specific receptive field size, CNNs are able to accurately identify local patterns in time series data, such as edges and textures, which often indicate key signals. As network depth increases, CNNs are able to extract multi-level features, ranging from basic to complex. This is crucial for parsing complex spatiotemporal patterns in time series data using multi-scale three-dimensional convolutional models. Their parallel processing mechanism is particularly well-suited for processing large datasets, which is particularly important for real-time industrial data analysis. Compared to traditional two-dimensional convolutional neural networks, 3D CNNs offer superior modeling capabilities for spatiotemporal data, making them suitable for applications such as weather forecasting, industrial process optimization, and human-computer interaction. For example, convolution has been used to learn local spatial features for industrial modeling, and its effectiveness has been demonstrated in industrial hydrocracking processes. However, two-dimensional convolutions cannot capture temporal correlations between different samples. In order to fill the research gap, a prediction model based on three-dimensional convolution was subsequently proposed to jointly capture the spatiotemporal features of low-level and high-level layers.

[0004] In order to deeply explore the intrinsic connections between data in industrial processes and adapt the input dimensions of three-dimensional convolutional neural networks, this study used one-dimensional convolution and time delay technology to enhance the data and construct three-dimensional sample data. This method clearly improves the efficiency of complex industrial data processing and the overall performance of the model. However, the convolution kernel of a single scale is limited to capturing feature information within a limited range and cannot extract the feature relationships between multiple coupled production equipment and units at different scales and levels. To this end, this paper proposes a new multi-scale three-dimensional convolutional network model. This model forms multi-scale three-dimensional features through parallel convolution of different scales. Multi-scale three-dimensional features capture spatiotemporal features at different scales, thereby more comprehensively understanding feature information at different scales and levels in the production system. This design further enhances the model's insight into complex production processes. In addition, we effectively integrate these feature maps extracted at different scales to construct a comprehensive feature representation. These features not only enrich the information content of the model, but also improve the model's adaptability and prediction accuracy to dynamic changes in the production process. By comprehensively using multi-scale convolution, the model's accuracy in predicting production system performance has been significantly improved, while its robustness to variable production conditions has been enhanced. This provides strong decision-making support for the optimization of the production process and opens up new ideas for the intelligent management of complex rotary kiln industrial systems. Summary of the Invention

[0005] The present invention focuses on the problems existing in the above-mentioned prior art and provides a method, system, medium and equipment for predicting the sintering temperature of a rotary kiln using a multi-scale three-dimensional convolution model.

[0006] A first aspect of the present invention provides a method for predicting the sintering temperature of a rotary kiln using a multi-scale three-dimensional convolution model, comprising:

[0007] Step S100: using one-dimensional convolution to perform data augmentation and reconstruction to generate three-dimensional data, thereby increasing the size and diversity of the data;

[0008] Step S200: Generate three-dimensional features of multiple scales using one-dimensional convolutions of different scales, and then fuse them into multi-scale three-dimensional features;

[0009] Step S300: using the deep network architecture of 3D CNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system;

[0010] Step S400: Use three different evaluation indicators to evaluate the performance of the model, where the evaluation indicators include mean absolute error, root mean square error, and correlation coefficient.

[0011] In step S100, the data augmentation and reconstruction using one-dimensional convolution to generate three-dimensional data specifically includes:

[0012] Step S110: Collect multiple one-dimensional sample vectors from the industrial process at fixed intervals; each one-dimensional sample vector refers to the measured values ​​of m process variables at a specific sampling time t. Specifically, the original variable data at multiple time points are collected from the industrial system at a fixed sampling period, and each sample is an m-dimensional vector. The sample at the tth time is expressed as:

[0013]

[0014] Where: m represents the number of process variables; t represents the time sampling point index; x i (t) represents the measured value of the i-th variable at time t, and the value of i is between 1 and m.

[0015] Step S120: For each one-dimensional sample Apply m 1D convolution kernels to form a two-dimensional dynamic matrix , the i-th row of the matrix is ​​the response of the i-th convolution kernel to the input vector:

[0016]

[0017] in:

[0018] is the variable in the i-th row and j-th column, which represents the perceptual response of the i-th convolution kernel to the j-th variable. The values ​​of i and j are between 1 and m.

[0019] X(t): The two-dimensional dynamic feature map at time t, representing the embedded structure between variables;

[0020] Among them, 1D convolution can be understood as a local embedding operation, mapping the variable relationship to a two-dimensional space. The essence of the convolution operation is to extract the interaction information between variables from multiple angles using different receptive fields or weight kernels;

[0021] Step S130: The two-dimensional feature matrix constructed at each time t As the t-th slice of the 3D tensor x, stacked step by step along the "time channel" dimension:

[0022]

[0023] in:

[0024] T is the time window length (i.e., the number of lag frames), which means that the continuous information of the latest T time points is used;

[0025] The constructed 3D tensor contains T two-dimensional embedding graphs.

[0026] Step S140: Generating a 3D feature map by sliding stacking the time window specifically includes:

[0027] The time series is sliced ​​using a sliding window (stride = 1) and the two-dimensional matrices of consecutive T frames are stacked to obtain a three-dimensional feature tensor:

[0028]

[0029] in:

[0030] Each tensor sample contains the time dimension T, the variable dimension m, and the inter-variable embedding dimension m;

[0031] Preferably, the step S200 further includes:

[0032] The input feature map of the multi-scale convolutional layer is defined as X∈R C×H′×W′ Description, where C, H′, and W′ are the number of channels, height, and width of the feature map, respectively. The height H and width W are calculated as follows:

[0033] H=[(H′-f m +2P H ) / stride]+1

[0034] W=[(H′-f m +2P W ) / stride]+1

[0035] Among them, [] means rounding up, stride is the step size of the convolution kernel, and the step size of all convolution kernels is consistent, P H and P W are the number of vertical and horizontal zero padding in the input feature map during convolution operation, setting stride = 1, H = H′, W = W′, then P H =P W =(f m -1) / 2;f m Indicates the size of the convolution kernel.

[0036] The convolution kernel sizes are C1, C2, ..., C p The one-dimensional convolution performs three-dimensional data processing on the sample data to generate p T×m×m tensors, whose three-dimensional features are F1, F2, ..., F p , and transform the features F1, F2, ..., F p Data fusion is performed in three ways, which are as follows:

[0037] 1) MAX function

[0038] F 1-p=MAX(F1,F2,...,F p )

[0039] Among them, MAX() means taking the maximum value of the corresponding position elements of all input features; this method takes the maximum value of the feature value of each position at different scales, retains the most significant response area, and improves the model's sensitivity to mutation information;

[0040] 2) ADD function

[0041] F 1-p =ADD(F1,F2,...,F p )

[0042] Among them, ADD() represents the addition of elements at corresponding positions of all input features; ADD fusion retains the superposition information of features at each scale by adding feature maps of different scales element by element at each position, which helps to enrich the expressive power of features and enhance the robustness and generalization ability of the model.

[0043] 3) STACK function

[0044] F 1-p =STACK(F1,F2,...,F p )

[0045] Among them, STACK() means that all input features are connected along a new dimension, and STACK fusion splices all features along the new channel dimension. This method retains the feature differences of each scale to the greatest extent without losing the original information.

[0046] The step S300 specifically includes:

[0047] Define input feature map Where C is the channel dimension, W and H represent the width and height of the l-th layer feature map respectively, and T represents the time dimension. Its output can be formally expressed as:

[0048]

[0049] In the formula, * represents the three-dimensional convolution operation, σ is the activation function, and is a learnable parameter. To extract spatiotemporal features, a three-dimensional convolutional network with L layers is stacked, where L represents the total number of network layers. By stacking multiple layers of three-dimensional convolutional networks, the model can gradually extract high-level global temporal features from the underlying local features.

[0050] After being processed by the multi-scale 3DCNN module, the fused three-dimensional feature map is input into the fully connected layer network to achieve sintering temperature prediction.

[0051] Where W is the weight and b is the bias;

[0052] Finally, the optimization objective is to minimize the mean square error (MSE) between the predicted output and the true label, and the loss function is defined as:

[0053]

[0054] Among them, N is the number of training samples, y i is the actual value, is the predicted value.

[0055] Preferably, the step S400 further includes:

[0056] The mean absolute error reflects the actual situation of the error between the model output and the true value. The root mean square error more comprehensively reflects the deviation between the model output and the true value. The correlation coefficient is used to evaluate the degree of fit of the input-output relationship in the multiple regression equation. The calculation method is as follows:

[0057]

[0058]

[0059] Among them, y i is the actual value of sintering temperature, is the predicted value, N is the number of samples in the validation set, is the mean of the observations, MAE is the mean absolute error, RMSE is the root mean square error, and R2 is the correlation coefficient.

[0060] A second aspect of the present invention provides a system for predicting the sintering temperature of a rotary kiln based on a multi-scale three-dimensional convolution model according to the method, the system comprising:

[0061] The data reconstruction unit uses one-dimensional convolution to perform data amplification and reconstruction to generate three-dimensional data, increasing the size and diversity of the data;

[0062] The feature fusion unit uses one-dimensional convolutions of different scales to generate three-dimensional features of multiple scales, and then fuses them into multi-scale three-dimensional features;

[0063] The feature extraction unit uses the deep network architecture of 3D CNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system;

[0064] The model evaluation unit uses three different evaluation indicators to evaluate the performance of the model, including mean absolute error, root mean square error and correlation coefficient.

[0065] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the aforementioned methods are implemented.

[0066] According to a fourth aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of any of the aforementioned methods when executed by the processor.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] This study employed one-dimensional convolution and time delay techniques to enhance data and construct three-dimensional sample data. This approach significantly improved the efficiency of complex industrial data processing and the overall performance of the model. A new multi-scale three-dimensional convolutional network model was proposed. This model generates multi-scale three-dimensional features through parallel convolutions at different scales. Multi-scale three-dimensional features capture spatiotemporal characteristics at different scales, enabling a more comprehensive understanding of feature information at different scales and levels within the production system. This design further enhances the model's insight into complex production processes.

[0069] Furthermore, these feature maps extracted at different scales are effectively integrated to construct a comprehensive feature representation. These features not only enrich the information content of the model but also enhance its adaptability to dynamic changes in the production process and its prediction accuracy. By comprehensively utilizing multi-scale convolution, the model significantly improves its accuracy in predicting production system performance while enhancing its robustness to variable production conditions. This provides strong decision-making support for optimizing the production process and opens up new avenues for the intelligent management of complex rotary kiln industrial systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Flowchart of the rotary kiln sintering temperature prediction method of the present invention;

[0071] Figure 2 The three-dimensional sample data of the present invention at time t;

[0072] Figure 3 The structure diagram of the MS-3D CNN model of the present invention;

[0073] Figure 4 Comparison of prediction results of six algorithms for rotary kiln sintering temperature. DETAILED DESCRIPTION

[0074] The techniques described below are susceptible to various modifications and embodiments, and are described in detail herein with reference to specific embodiments in conjunction with the accompanying drawings. However, this is not intended to limit the techniques described below to the specific embodiments. It should be understood that the present invention includes all similar modifications, equivalents, and alternatives that do not depart from the spirit and technical scope of the techniques described below.

[0075] Example 1

[0076] like Figure 1 As shown, the present invention provides a method for predicting the sintering temperature of a rotary kiln using a multi-scale three-dimensional convolution model, comprising:

[0077] Step S100: using one-dimensional convolution to perform data augmentation and reconstruction to generate three-dimensional data, thereby increasing the size and diversity of the data;

[0078] Step S200: Generate three-dimensional features of multiple scales using one-dimensional convolutions of different scales, and then fuse them into multi-scale three-dimensional features;

[0079] Step S300: using the deep network architecture of 3D CNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system;

[0080] Step S400: Use three different evaluation indicators to evaluate the performance of the model, where the evaluation indicators include mean absolute error, root mean square error, and correlation coefficient.

[0081] Preferably, the step S100 of performing data augmentation and reconstruction to generate three-dimensional data by using one-dimensional convolution specifically includes:

[0082] Assume that in an industrial process, there are m real-valued process variables measured by hardware sensors. Each variable corresponds to a time series. According to the industrial topology, the m process variables at time t can be expressed as x 1,(t) ,x 2,(t) ,...,x m,(t) , that is, x t =[x 1,(t) ,x 2,(t) ,...,x m,(t) ], Furthermore, assuming the number of time lags is T, the size of the 3D tensor is T×m×m.

[0083] Step S110: Collect multiple one-dimensional sample vectors from the industrial process at fixed intervals; each one-dimensional sample vector refers to the measured values ​​of m process variables at a specific sampling time t. Specifically, the original variable data at multiple time points are collected from the industrial system at a fixed sampling period, and each sample is an m-dimensional vector. The sample at the tth time is expressed as:

[0084]

[0085] Where: m represents the number of process variables; t represents the time sampling point index; x i (t) represents the measured value of the i-th variable at time t, and the value of i is between 1 and m.

[0086] Step S120: For each one-dimensional sample Apply m 1D convolution kernels to form a two-dimensional dynamic matrix The i-th row of the matrix is ​​the response of the i-th convolution kernel to the input vector:

[0087]

[0088] in:

[0089] is the variable in the i-th row and j-th column, which represents the perceptual response of the i-th convolution kernel to the j-th variable. The values ​​of i and j are between 1 and m.

[0090] X(t): The two-dimensional dynamic feature map at time t, representing the embedded structure between variables;

[0091] Among them, 1D convolution can be understood as a local embedding operation, mapping the variable relationship to a two-dimensional space. The essence of the convolution operation is to extract the interaction information between variables from multiple angles using different receptive fields or weight kernels;

[0092] Step S130: The two-dimensional feature matrix constructed at each time t As the t-th slice of the 3D tensor x, stacked step by step along the "time channel" dimension:

[0093]

[0094] in:

[0095] T is the time window length (i.e., the number of lag frames), which means that the continuous information of the latest T time points is used;

[0096] The constructed 3D tensor contains T two-dimensional embedding graphs.

[0097] Step S140: Generating a 3D feature map by sliding stacking the time window specifically includes:

[0098] The time series is sliced ​​using a sliding window (s tride = 1) and the two-dimensional matrices of consecutive T frames are stacked to obtain a three-dimensional feature tensor:

[0099]

[0100] in:

[0101] Each tensor sample contains the time dimension T, the variable dimension m, and the inter-variable embedding dimension m;

[0102] Preferably, the step S200 further includes:

[0103] The core of MS_3DCNN is the multi-scale convolution layer, which sets multiple convolution kernels of different sizes in parallel, with receptive fields of different sizes. It can extract local behaviors of different scales in the input feature map and generate feature maps containing multi-topological scale local spatiotemporal features. These multi-topological scale feature maps obtained in parallel will be re-merged and stacked into a multi-topological scale local spatiotemporal feature map set for subsequent network layers to perform high-order feature conversion and combination. However, the size of the output feature map after the convolution operation is related to the convolution kernel size, step size and number of zero padding. If the sizes of the output feature maps of convolution kernels of different sizes are inconsistent, these feature maps cannot be merged. For this reason, the multi-scale convolution layer can be designed in parallel as described below:

[0104] The input feature map of the multi-scale convolutional layer is defined as X∈R C×H′×W′ Description, where C, H′, and W′ are the number of channels, height, and width of the feature map, respectively. The height H and width W are calculated as follows:

[0105]

[0106] Among them, [] means rounding up, stride is the step size of the convolution kernel, and the step size of all convolution kernels is consistent, P H and P W are the number of vertical and horizontal zero padding in the input feature map during convolution operation, setting stride = 1, H = H′, W = W′, then P H =P W =(f m -1) / 2;f m Indicates the size of the convolution kernel.

[0107] The convolution kernel sizes are C1, C2, ..., C p The one-dimensional convolution performs three-dimensional data processing on the sample data to generate p T×m×m tensors, whose three-dimensional features are F1, F2, ..., F p , and transform the features F1, F2, ..., F p Data fusion is performed in three ways, which are as follows:

[0108] 1) MAX function

[0109] F 1-p =MAX(F1,F2,...,F p )

[0110] Among them, MAX() means taking the maximum value of the corresponding position elements of all input features; this method takes the maximum value of the feature value of each position at different scales, retains the most significant response area, and improves the model's sensitivity to mutation information;

[0111] 2) ADD function

[0112] F 1-p =ADD(F1,F2,...,F p )

[0113] Among them, ADD() represents the addition of elements at corresponding positions of all input features; ADD fusion retains the superposition information of features at each scale by adding feature maps of different scales element by element at each position, which helps to enrich the expressive power of features and enhance the robustness and generalization ability of the model.

[0114] 3) STACK function

[0115] F 1-p =STACK(F1,F2,...,F p )

[0116] Among them, STACK() means that all input features are connected along a new dimension, and STACK fusion splices all features along the new channel dimension. This method retains the feature differences of each scale to the greatest extent without losing the original information.

[0117] Step S300 specifically includes:

[0118] Define input feature map Where C is the channel dimension, W and H represent the width and height of the l-th layer feature map respectively, and T represents the time dimension. Its output can be formally expressed as:

[0119]

[0120] In the formula, * represents the three-dimensional convolution operation, σ is the activation function, and is a learnable parameter. To extract spatiotemporal features, a three-dimensional convolutional network with L layers is stacked, where L represents the total number of network layers. By stacking multiple layers of three-dimensional convolutional networks, the model can gradually extract high-level global temporal features from the underlying local features.

[0121] After being processed by the multi-scale 3DCNN module, the fused three-dimensional feature map is input into the fully connected layer network to achieve sintering temperature prediction.

[0122] Where W is the weight and b is the bias;

[0123] Finally, the optimization objective is to minimize the mean square error (MSE) between the predicted output and the true label, and the loss function is defined as:

[0124]

[0125] Among them, N is the number of training samples, y i is the actual value, is the predicted value.

[0126] Preferably, the step S400 further includes:

[0127] During the model performance evaluation process, three different evaluation indicators are used to assess the performance of the model. These indicators are the mean absolute error (MAE), the root mean square error (RMSE), and the correlation coefficient (R2). MAE can better reflect the actual situation of the error between the model's output and the true value. RMSE can more comprehensively reflect the deviation between the model's output and the true value. The smaller the MAE and RMSE values, the better the performance of the model. R2 is a statistic used in a multiple regression equation to evaluate the degree of fit of the regression equation to the input-output relationship. This indicator ranges from 0 to 1, and the closer the value is to 1, the better the model fits the input-output relationship.

[0128]

[0129] Among them, y i is the actual value of sintering temperature, is the predicted value, N is the number of samples in the validation set, is the mean of the observations.

[0130] This paper compares the performance of six algorithms, namely LSTM, CNN_LSTM, 1DCNN, 2DCNN, 3DCNN, and MS-3DCNN, in predicting the sintering temperature of a rotary kiln. The evaluation results of the mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R2) are shown in the following table:

[0131]

[0132] It can be seen that the MS-3D CNN algorithm proposed in this invention performs better in the evaluation of MAE, RMSE and R2.

[0133] Furthermore, the results of the six algorithms in predicting the sintering temperature of the rotary kiln are compared. Figure 4 shown.

[0134] Example 2

[0135] The present invention also provides a rotary kiln sintering temperature prediction system based on a multi-scale three-dimensional convolution model, comprising:

[0136] The data reconstruction unit uses one-dimensional convolution to perform data amplification and reconstruction to generate three-dimensional data, increasing the size and diversity of the data;

[0137] The feature fusion unit uses one-dimensional convolutions of different scales to generate three-dimensional features of multiple scales, and then fuses them into multi-scale three-dimensional features;

[0138] The feature extraction unit uses the deep network architecture of 3D CNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system;

[0139] The model evaluation unit uses three different evaluation indicators to evaluate the performance of the model, including mean absolute error, root mean square error and correlation coefficient.

[0140] Example 3

[0141] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the aforementioned methods when executed by a processor.

[0142] Example 4

[0143] The present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of any of the aforementioned methods are implemented.

[0144] The memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), and may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory), hard disk drive (English: hard disk drive, abbreviated: HDD) or solid-state drive (English: solid-state drive, abbreviated: SSD). The memory may also include a combination of the above types of memory;

[0145] The processor may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and a NP.

[0146] The processor may further include a hardware chip, which may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof, and the PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.

[0147] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0148] Although the present invention has been described in detail above using general descriptions and specific embodiments, modifications and improvements may be made based on the present invention. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Other changes and modifications made by those skilled in the art without departing from the spirit and scope of protection of the present invention are also included within the scope of protection of the present invention.

Claims

1. A method for predicting sintering temperature of a rotary kiln based on a multi-scale three-dimensional convolution model, characterized in that: The method comprises the following steps: Step S100: using one-dimensional convolution to perform data augmentation and reconstruction to generate three-dimensional data, thereby increasing the size and diversity of the data; Step S200: Generate three-dimensional features of multiple scales using one-dimensional convolutions of different scales, and then fuse them into multi-scale three-dimensional features; Step S300: using the deep network architecture of 3DCNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system; Step S400: Use three different evaluation indicators to evaluate the performance of the model, where the evaluation indicators include mean absolute error, root mean square error, and correlation coefficient.

2. The method according to claim 1, wherein The step S100 of using one-dimensional convolution to perform data augmentation and reconstruction to generate three-dimensional data specifically includes: Step S110: Collect multiple one-dimensional sample vectors from the industrial process at fixed intervals. Each one-dimensional sample vector refers to the measured values ​​of m process variables at a specific sampling time t. Specifically, the original variable data at multiple time points are collected from the industrial system at a fixed sampling period. Each sample is an m-dimensional vector. The sample at the tth time is expressed as: in: m represents the number of process variables; t represents the time sampling point index; x i (t) represents the measured value of the i-th variable at time t, and the value of i is between 1 and m; Step S120: For each one-dimensional sample Apply m 1D convolution kernels to form a two-dimensional dynamic matrix The i-th row of the matrix is ​​the response of the i-th convolution kernel to the input vector: in: is the variable in the i-th row and j-th column, which represents the perceptual response of the i-th convolution kernel to the j-th variable. The values ​​of i and j are between 1 and m. X(t): The two-dimensional dynamic feature map at time t, representing the embedded structure between variables; Step S130: The two-dimensional feature matrix constructed at each time t As the t-th slice of the 3D tensor x, stacked stepwise along the time dimension: in: T is the length of the time window, which means using the continuous information of the latest T time points; The constructed 3D tensor contains T two-dimensional embedding graphs; Step S140: Generating a 3D feature map by sliding stacking the time window specifically includes: The time series is sliced ​​using a sliding window (stride = 1) and the two-dimensional matrices of consecutive T frames are stacked to obtain a three-dimensional feature tensor: in: Each tensor sample contains the time dimension T, the variable dimension m, and the embedding dimension m between variables.

3. The method according to claim 1, wherein The step S200 specifically includes: The input feature map of the multi-scale convolutional layer is denoted by X∈R C×H′×W′ Description, where C, H′, and W′ are the number of channels, height, and width of the feature map, respectively. The height H and width W are calculated as follows: H=[(H′-f m +2P H ) / stride]+1 W=[(H′-f m +2P W ) / stride]+1 Among them, [] means rounding up, stride is the step size of the convolution kernel, and the step size of all convolution kernels is consistent, P H and P W are the number of vertical and horizontal zero padding in the input feature map during convolution operation, setting stride = 1, H = H′, W = W′, then P H =P W =(f m -1) / 2,f m Indicates the size of the convolution kernel, and the convolution kernel sizes used are C1, C2, ..., C p The one-dimensional convolution performs three-dimensional data processing on the sample data to generate p T×m×m tensors, whose three-dimensional features are F1, F2, ..., F p , the three-dimensional features F1, F2, ..., F p Data fusion is performed in three ways, which are as follows: 1) MAX function F 1-p =MAX(F1,F2,...,F p ) Among them, MAX() means taking the maximum value of the corresponding position elements of all input features; the MAX function takes the maximum value of the feature value of each position at different scales, retains the most significant response area, and improves the model's sensitivity to mutation information; 2) ADD function F 1-p =ADD(F1,F2,...,F p ) Among them, ADD() represents the addition of elements at corresponding positions of all input features; ADD fusion retains the superposition information of features at each scale by adding feature maps of different scales element by element at each position, which helps to enrich the expressive power of features and enhance the robustness and generalization ability of the model. 3) STACK function F 1-p =STACK(F1,F2,...,F p ) Among them, STACK() means that all input features are connected along a new dimension, and STACK fusion splices all features along the new channel dimension. The STACK function retains the feature differences of each scale to the greatest extent without losing the original information.

4. The method according to claim 1, wherein The step S300 specifically includes: Define input feature map Where C is the channel dimension, W and H represent the width and height of the l-th layer feature map respectively, and T represents the time dimension. Its output can be formally expressed as: In the formula, * represents the three-dimensional convolution operation, σ is the activation function, and is a learnable parameter. To extract spatiotemporal features, a three-dimensional convolutional network with L layers is stacked, where L represents the total number of network layers. By stacking multiple layers of three-dimensional convolutional networks, the model can gradually extract high-level global temporal features from the underlying local features. After being processed by the multi-scale 3DCNN module, the fused three-dimensional feature map is input into the fully connected layer network to achieve sintering temperature prediction. Where W is the weight and b is the bias; Finally, the optimization objective is to minimize the mean square error (MSE) between the predicted output and the true label, and the loss function is defined as: Among them, N is the number of training samples, y i is the actual value, is the predicted value.

5. The method according to claim 1, wherein The step S400 specifically includes: The mean absolute error reflects the actual situation of the error between the model output and the true value. The root mean square error more comprehensively reflects the deviation between the model output and the true value. The correlation coefficient is used to evaluate the degree of fit of the input-output relationship in the multiple regression equation. The calculation method is as follows: Among them, y i is the actual value of sintering temperature, is the predicted value, N is the number of samples in the validation set, is the mean of the observations, MAE is the mean absolute error, RMSE is the root mean square error, and R2 is the correlation coefficient.

6. A rotary kiln sintering temperature prediction system based on a multi-scale three-dimensional convolution model for implementing the method according to any one of claims 1 to 5, characterized in that: The system comprises: The data reconstruction unit uses one-dimensional convolution to perform data amplification and reconstruction to generate three-dimensional data, increasing the size and diversity of the data; The feature fusion unit uses one-dimensional convolutions of different scales to generate three-dimensional features of multiple scales, and then fuses them into multi-scale three-dimensional features; The feature extraction unit uses the deep network architecture of 3D CNN to extract feature information of multi-scale three-dimensional features at different scales, thereby improving the prediction performance and generalization ability of the soft measurement system; The model evaluation unit uses three different evaluation indicators to evaluate the performance of the model, including mean absolute error, root mean square error and correlation coefficient.

7. The system according to claim 6, characterized in that Also includes: A display module is used to receive and display the results of the rotary kiln sintering temperature prediction; The storage module is used to store the prediction result of the rotary kiln sintering temperature.

8. The system according to claim 7, characterized in that The display module includes: The display panel includes a pixel array and a driving backplane, wherein the pixel array is composed of a plurality of sub-pixels arranged in a preset manner; a driving circuit, electrically connected to the display panel, comprising a timing controller, a source driver chip, and a gate driver chip, and configured to provide a driving signal to the pixel array; A protective layer covering the light-emitting surface of the display panel, comprising tempered glass or a transparent polymer cover; a storage medium unit comprising at least one volatile memory and at least one non-volatile memory; a control unit connected to the storage medium unit, comprising a main controller and an error correction engine, wherein the main controller is configured to perform data read and write operations and bad block management, and the error correction engine is configured to perform real-time error correction on stored data; An interface unit, supporting high-speed serial bus protocols and low-speed control protocols, for communicating with the storage module; The power management unit provides multi-voltage domain power supply for the storage medium unit, the control unit and the interface unit.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images