A forward process-guided steel thick plate shape quality prediction method and device

Abstract

This application proposes a forward process-guided steel thick plate shape quality prediction method and device, which belongs to the technical field of intelligent modeling of complex industrial processes. The prediction method includes: obtaining multi-source heterogeneous data of the entire thick plate production process; preprocessing the multi-source heterogeneous data of the entire thick plate production process; extracting features from the preprocessed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain features corresponding to different processes; inputting the features corresponding to different processes into a pre-trained collaborative attention mechanism model to obtain fused features; inputting the fused features into a pre-trained prediction model to obtain steel thick plate shape quality prediction results. The collaborative attention mechanism is used to fuse the three extracted process features, fully considering the connection between each process and improving the accuracy of thick plate shape quality prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent modeling of complex industrial processes, and in particular relates to a method and device for predicting the shape quality of thick steel plates guided by a forward process. Background Art

[0002] As a key product in the steel industry, heavy plate boasts diverse product types, stable quality, diverse performance, and a wide range of applications. Consequently, countries around the world attach great importance to its quality. Improving heavy plate quality not only ensures safety during subsequent use but also reduces losses for heavy plate manufacturers, further driving the steel manufacturing industry towards intelligent and digital development.

[0003] The production process of thick plate is highly complex and typical of process industries. Its production goal is to produce products with superior performance, qualified quality, and conforming to standard plate shapes. To achieve this goal, the thick plate production process must be monitored and effectively controlled in real time. Thick plate production consists of a series of steps, with the output of each step serving as the input for the next. These steps primarily include heating, rolling, and cooling. The quality of the final thick plate product is currently assessed based on five key aspects: dimensions, shape, performance, surface finish, and internal quality. Thick plate shape quality is closely linked to the production process, requiring the strictest control throughout the entire production process. Furthermore, as a key quality indicator, plate shape directly impacts the final yield rate. Therefore, thick plate manufacturers must optimize product quality while minimizing production costs and energy losses. Therefore, identifying the relationship between the production process and final thick plate shape quality and establishing an accurate prediction model for thick plate shape quality are critical challenges that need to be addressed. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, this application proposes a forward process-guided steel plate shape quality prediction method and device.

[0005] In a first aspect, the present application proposes a forward process-guided steel plate shape quality prediction method, comprising:

[0006] Acquire multi-source heterogeneous data for the entire thick plate production process;

[0007] Preprocessing multi-source heterogeneous data of the entire thick plate production process;

[0008] Feature extraction is performed on the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain the features corresponding to different processes;

[0009] Input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fused features;

[0010] The fusion features are input into a pre-trained prediction model to obtain a prediction result of the shape quality of the steel thick plate.

[0011] The preprocessing of multi-source heterogeneous data of the entire thick plate production process includes:

[0012] Performing data cleaning on multi-source heterogeneous data of the entire thick plate production process;

[0013] Interpolate the multi-source heterogeneous data of the entire thick plate production process after data cleaning to obtain multi-source heterogeneous data of equal length;

[0014] Dimensionality reduction is performed on the high-dimensional data in the multi-source heterogeneous data of the entire thick plate production process to obtain the reduced-dimensional source heterogeneous data.

[0015] The feature extraction is performed on the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain features corresponding to different processes, including:

[0016] According to different processes, multi-source heterogeneous data of equal length and source heterogeneous data after dimensionality reduction are divided into variables;

[0017] According to the time sequence of different processes, the multi-source heterogeneous data of equal length after the variables are divided and the source heterogeneous data after dimensionality reduction are spliced;

[0018] Normalize the spliced multi-source heterogeneous data;

[0019] The normalized multi-source heterogeneous data are input into different convolutional neural networks corresponding to different processes for feature extraction to obtain the features corresponding to different processes.

[0020] The different convolutional neural networks include: a convolution layer, an attenuation layer, a flattening layer and a fully connected layer.

[0021] The different processes include: a heating process, a rolling process and a cooling process.

[0022] The pre-trained collaborative attention mechanism model includes:

[0023] A first collaborative attention layer and a second collaborative attention layer; the first collaborative attention layer includes: a first query vector, a first key vector and a first value vector; the second collaborative attention layer includes: a second query vector, a second key vector and a second value vector; the characteristics of the heating process are superimposed on the first query vector as the input of the first collaborative attention layer, and the characteristics of the rolling process are superimposed on the first key vector and the first value vector respectively as the attention weight of the first collaborative attention layer, the output of the first collaborative attention layer is superimposed on the second query vector as the input of the second collaborative attention layer, the cooling process is superimposed on the second key vector and the second value vector respectively as the attention weight of the second collaborative attention layer, and the output of the second collaborative attention layer is used as the fusion feature.

[0024] The pre-trained prediction model is a single-layer fully connected network.

[0025] In a second aspect, the present application proposes a forward process-guided steel plate shape quality prediction device, comprising:

[0026] The data acquisition module is used to obtain multi-source heterogeneous data of the entire thick plate production process;

[0027] The data preprocessing module is used to preprocess the multi-source heterogeneous data of the entire thick plate production process;

[0028] The process feature extraction module is used to extract features from the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes, and obtain the features corresponding to different processes;

[0029] The fusion feature extraction module is used to input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fusion features;

[0030] The quality prediction module is used to input the fusion features into the pre-trained prediction model to obtain the steel plate shape quality prediction result.

[0031] In a third aspect, the present application proposes an electronic device comprising: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the forward process-guided steel thick plate shape quality prediction method.

[0032] In a fourth aspect, the present application proposes a computer-readable storage medium storing executable instructions, which, when executed, enable a processor to execute the forward process-guided steel thick plate shape quality prediction method.

[0033] In a fifth aspect, the present application proposes a computer program product, comprising a computer program or instructions, which, when executed by a processor, implements the forward process-guided steel thick plate shape quality prediction method.

[0034] Beneficial effects:

[0035] This application proposes a forward process-guided steel thick plate shape quality prediction method and device. Different convolutional neural networks are designed for feature extraction based on the characteristics of different process data, and a collaborative attention mechanism is used to fuse the three extracted process features. The connection between each process is fully considered, which can better predict the thick plate shape quality, thereby assisting on-site production decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flowchart of the method for predicting the shape quality of thick steel plates guided by the forward process according to an embodiment of the present application;

[0037] Figure 2 Schematic diagram of the steel plate shape quality prediction training model and prediction method according to an embodiment of the present application;

[0038] Figure 3 Flow chart of the actual production process of thick steel plates according to the embodiment of the present application;

[0039] Figure 4 Schematic diagram of the steel plate shape quality prediction training model and prediction process according to the embodiment of the present application;

[0040] Figure 5 Schematic diagram of the feature extraction process based on the collaborative attention mechanism in an embodiment of the present application;

[0041] Figure 6 ROC curve experimental results of the embodiment of the present application;

[0042] Figure 7 Confusion matrix experimental results of the embodiment of the present application;

[0043] Figure 8 Comparison chart of the box shape results of the thick plate shape quality prediction in the embodiment of the present application;

[0044] Figure 9 Principle block diagram of the forward process-guided steel thick plate shape quality prediction device in an embodiment of the present application. DETAILED DESCRIPTION

[0045] Controlling the shape quality of thick plate products is a challenging task in evaluating their quality. The production process involves complex physical processing and chemical reactions, and changes in any variable can affect other variables, leading to shape issues. Furthermore, since thick plate production involves small batches, multiple varieties, and multiple processes, identifying the specific process variables that influence shape is particularly difficult. This makes it difficult to select appropriate process variables for accurate shape prediction when developing shape quality prediction models. Furthermore, it is impossible to deeply analyze the significance of individual process variables in affecting thick plate shape.

[0046] Current research on thick plate shape prediction primarily focuses on the production mechanism, deeply analyzing each process in the thick plate production process. By understanding and applying the physical and chemical changes that thick plate undergoes during production and incorporating expert experience, a systematic analysis of the various potential impacts of production processes on thick plate shape is conducted. Ultimately, a prediction model is established that uses production process indicators and equipment operating parameters as inputs and outputs the final shape of the thick plate. However, most traditional methods focus on specific processes or issues, directly using various parameters as model inputs, and fail to address shape quality from a comprehensive perspective. Thick plate production is a continuous, full-process process in which each process interacts and influences each other. Mechanism-based modeling and data-driven approaches that focus solely on a single process may overlook the inter-process dependencies, leading to modeling errors. Furthermore, data-driven prediction methods have also become a research hotspot. For example, a paper developed a random forest prediction model using historical data from the thick plate rolling stage. It then applied an association rule mining algorithm to analyze key control variables in the production process. These results were then applied to actual thick plate control processes, improving shape quality control. However, most of these methods target specific processes or issues, failing to address plate quality from a comprehensive perspective. They often directly utilize acquired data without classifying and extracting features based on the process, nor do they consider the fusion of data across processes. However, thick plate production is a continuous, full-process process, with each process interacting and influencing each other. Mechanism-based modeling and data-driven approaches that focus solely on a single process can overlook the inter-process connections, leading to modeling errors.

[0047] In response to this problem, the present application proposes a forward process-guided steel thick plate shape quality prediction method and device. This method uses the full process data (including 100 one-dimensional process variables and 36 two-dimensional variables) in the thick plate production process as the initial input of the model based on expert knowledge and production technology, and then divides these data into processes and performs data preprocessing according to the actual production process of the thick plate; uses a convolutional neural network to extract features of the three divided process features to further obtain a deep representation of the input data; and on this basis, uses a collaborative attention mechanism to fuse the extracted process features, and finally performs subsequent plate shape quality prediction. The method of the present application fully considers the impact of the interaction between the various processes of thick plate production on the final plate shape quality, thereby more accurately predicting the thick plate shape quality. Practice has shown that the method of the present application is an effective method for predicting the shape quality of steel thick plates, and has good guiding significance for actual thick plate production operations.

[0048] Example 1:

[0049] This embodiment proposes a forward process-guided steel plate shape quality prediction method, such as Figure 1 、 Figure 2 As shown, including:

[0050] Step S1: Acquire multi-source heterogeneous data of the entire thick plate production process;

[0051] Step S2: preprocessing the multi-source heterogeneous data of the entire thick plate production process;

[0052] Step S3: extracting features from the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain features corresponding to different processes;

[0053] Step S4: Input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fused features;

[0054] Step S5: inputting the fusion features into the pre-trained prediction model to obtain the steel plate shape quality prediction result.

[0055] In this embodiment, the first step is to obtain multi-source heterogeneous data of the entire thick plate production process. The multi-source heterogeneous data of the entire thick plate production process include: one-dimensional data, one-dimensional time series data and two-dimensional data. Under normal circumstances, the result of pre-processing the multi-source heterogeneous data of the entire thick plate production process is directly used as the input of the prediction model. In practice, it is found that even if the collaborative attention mechanism model is subsequently used for feature fusion, the result of pre-processing the multi-source heterogeneous data of the entire thick plate production process is used as the input of the collaborative attention mechanism model. No matter how the adjustment is made, the final prediction result cannot be further improved. However, the present application extracts features from the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain features corresponding to different processes; then the features corresponding to different processes are input into the pre-trained collaborative attention mechanism model to obtain fused features, and the fused features are input into the pre-trained prediction model to obtain the prediction results of the plate shape quality of the steel thick plate, and its prediction accuracy is significantly improved, breaking the inherent thinking in this field that the use of the collaborative attention mechanism model cannot further improve the prediction accuracy.

[0056] In step S1, in the specific prediction process, multi-source heterogeneous data of the entire thick plate production process is first obtained. The data can be real-time data of the thick plate production process or training data of the thick plate production process. The obtained multi-source heterogeneous data of the entire thick plate production process is used as input data for prediction. Before making the prediction, the collaborative attention mechanism model and the prediction model need to be pre-trained. Before pre-training, historical multi-source heterogeneous data of the entire thick plate production process needs to be collected. The construction process of historical multi-source heterogeneous data of the entire thick plate production process will be described in detail in the pre-trained collaborative attention mechanism model.

[0057] In step S2, the preprocessing of multi-source heterogeneous data of the entire thick plate production process includes:

[0058] Step S2.1: performing data cleaning on the multi-source heterogeneous data of the entire thick plate production process;

[0059] Step S2.2: Interpolate the cleaned multi-source heterogeneous data of the thick plate production process to obtain multi-source heterogeneous data of equal length;

[0060] Step S2.3: Perform dimensionality reduction processing on the high-dimensional data in the multi-source heterogeneous data of the entire thick plate production process to obtain the reduced-dimensional source heterogeneous data.

[0061] The different processes include: heating process, rolling process and cooling process, such as Figure 3As shown in the figure. In the actual thick plate production process, due to the different lengths of the thick plates produced, the lengths of the one-dimensional time series data in the heating process, rolling process, and cooling process vary greatly based on the original sampling rates. Therefore, in order to fully utilize the correlation between the data in the three processes, it is necessary to perform linear interpolation processing on the unequal length sample data in the extracted high-dimensional data to unify the dimensions of all sample data.

[0062] Data cleaning is performed on the multi-source heterogeneous data of the entire thick plate production process. At the same time, the linear interpolation method is used to preprocess the input data to unify the dimensions of all sample data due to the differences in data dimensions caused by the different data sample sizes. To address the problem of inconsistent lengths of heterogeneous data, the traditional data padding method selects the value with the largest variable length and performs subsequent zero padding operations on related variables. This method will lead to a large number of null values in the future and does not take into account the problem of excessive differences in data length between different processes in the process industry, which is not conducive to modeling feature extraction. Another method is to perform up and down sampling and interpolation processing on the data. Among them, the linear interpolation algorithm is easy to calculate and is the most commonly used method. Because the changing trends of actual thick plate variable data do not have an obvious functional relationship, it is impossible to use mathematical expressions to accurately describe the functional relationship. Therefore, linear interpolation is a very effective method in this case. Finally, in order to reduce the redundancy between features and facilitate subsequent fusion with one-dimensional process data, high-dimensional data needs to be reduced in dimensionality.

[0063] In step S3, feature extraction is performed on the pre-processed multi-source heterogeneous data of the thick plate production process according to different processes to obtain features corresponding to different processes, including:

[0064] Step S3.1: Divide the variables of the multi-source heterogeneous data of equal length and the source heterogeneous data after dimensionality reduction according to different processes;

[0065] Step S3.2: Splice the multi-source heterogeneous data of equal length after variable division and the source heterogeneous data after dimensionality reduction according to the time sequence of different processes;

[0066] Step S3.3: normalize the spliced multi-source heterogeneous data;

[0067] Step S3.4: Input the normalized multi-source heterogeneous data into different convolutional neural networks corresponding to different processes to perform feature extraction and obtain features corresponding to different processes.

[0068] In this embodiment, the pre-processed multi-source heterogeneous data of the thick plate production process are divided into variables according to the three actual production processes of the thick plate: heating process, rolling process, and cooling process, and the divided one-dimensional data and the high-dimensional data after dimensionality reduction are spliced according to the three processes. In the thick plate production process, there are large differences in the distribution and range of each variable data. Therefore, before predicting and analyzing the thick plate shape quality, it is necessary to normalize each variable data in the thick plate sample. Finally, three different convolutional neural networks are used to further extract features from the normalized multi-source heterogeneous data to obtain a deeper expression of the input data of the three processes. Each convolutional neural network contains: a convolution layer (Conv layer), a dropout layer (Dropout layer), a flattening layer (Flatten layer) and a fully connected layer (Dense layer).

[0069] The pre-trained collaborative attention mechanism model includes:

[0070] A first collaborative attention layer and a second collaborative attention layer; the first collaborative attention layer includes: a first query vector, a first key vector and a first value vector; the second collaborative attention layer includes: a second query vector, a second key vector and a second value vector; the characteristics of the heating process are superimposed on the first query vector as the input of the first collaborative attention layer, and the characteristics of the rolling process are superimposed on the first key vector and the first value vector respectively as the attention weight of the first collaborative attention layer, the output of the first collaborative attention layer is superimposed on the second query vector as the input of the second collaborative attention layer, the cooling process is superimposed on the second key vector and the second value vector respectively as the attention weight of the second collaborative attention layer, and the output of the second collaborative attention layer is used as the fusion feature.

[0071] The pre-trained prediction model is a single-layer fully connected network.

[0072] In this embodiment, after extracting features from each process, in order to fully consider the effects and connections between variables in each process, a collaborative attention mechanism is used to fuse the features of the three processes, and then the fused features are used to perform subsequent steel plate shape quality prediction. In this embodiment, step S3.4 obtains features corresponding to three different processes, including: features corresponding to the heating process, features corresponding to the rolling process, and features corresponding to the cooling process. The features corresponding to the heating process, the rolling process, and the cooling process are used as three input sources X, respectively. heat 、X roll and X cool ,in and n1, n2, n3 are the number of variables in the heating, rolling, and cooling processes, respectively, and d is the variable dimension. Each input source corresponds to the feature representation of the heating process, rolling process, and cooling process. To calculate the fused feature representation, the specific implementation steps are as follows:

[0073] First, the input source X of the heating process heat Convert it into a feature vector and add this feature vector to the input source X of the rolling process roll On top, generate data X heat Based on data X roll The attention weight A' (i.e., Q1 of the first collaborative attention layer comes from the heating process, K1 and V1 come from the rolling process) is used to achieve the purpose of the forward process (heating process) guiding the backward process (rolling process).

[0074] Similarly, convert X' into a feature vector and add this feature vector to the input source X of the cooling process cool On top, generate data X cool Attention weight A based on data X' (here Q2 comes from the fused feature X' of the first two processes, and K2 and V2 come from the cooling process). In this embodiment, through the design of the query vector Q (Query), key vector K (Key) and value vector V (Value) in the collaborative attention mechanism model, it not only realizes the fusion of the three process features, but also reflects the guiding role of the forward process on the backward process, and fully considers the connection between the various processes in the actual production process.

[0075] Finally, the fused features are input into a single-layer fully connected network to extract the thick plate product quality label information implicit in the production process data from the data features. Furthermore, the extracted quality label information is spatially normalized using the Softmax function to obtain the final thick plate shape quality label prediction value.

[0076] In this embodiment, before predicting the shape quality of thick steel plates, it is necessary to collect historical multi-source heterogeneous data from the entire thick plate production process as sample data and pre-train the collaborative attention mechanism model. The training process is as follows:

[0077] Step S100: Collect historical multi-source heterogeneous data of the entire thick plate production process, including:

[0078] Step S100.1: Collect historical full-process multi-source heterogeneous production data from actual industrial sites as sample data and store it in a relational database;

[0079] Step S100.2: Select the required variables from all historical full-process multi-source heterogeneous production data as input, and use the thick plate shape quality label as the prediction output;

[0080] In this example, based on the actual process and the analysis of the thick plate production process mechanism, 100 one-dimensional process variables and 36 two-dimensional variables were selected as the model input, and the thick plate shape quality label was used as the model's predicted output. The specific one-dimensional process variables and high-dimensional variables selected are shown in Tables 1 and 2 below:

[0081] Use SQL statements in the relational database to extract data and divide the specifications of thick plate samples, and select thick plates of the same specifications and store their data in .csv format for the next step of data preprocessing;

[0082] Specifically, in this embodiment, a total of 5,000 production data of the same specifications from January 2019 to June 2021 are extracted from the historical database of the thick plate production process of a steel plant as input data, and the quality label data indicating the quality of the plate shape is used as output data. These data are used to establish a quality prediction model for the final plate shape quality prediction.

[0083] In this example, because historical data contains numerous types of products that can be filtered based on specifications, we extracted a specific specification of thick plate from a historical database of thick plate production processes at a domestic steel mill for experimental analysis. After data cleaning, screening, and preprocessing, we obtained 4,504 thick plate shape quality sample data, including 2,947 normal thick plate sample data and 1,557 abnormal thick plate sample data. Using a sample partitioning method, we divided the training and test sets into a ratio of 8:2, with 3,603 training data sets and 901 test data sets. The training data was used to construct the model, while the test data was used to validate the model and find the optimal model parameters.

[0084] Table 1 Selected statistical table of one-dimensional process variable information (100 variables)

[0085]

[0086]

[0087] Table 2 Selected statistics of high-dimensional variable information (36 variables)

[0088]

[0089] Step S101: Preprocessing the historical multi-source heterogeneous data of the entire thick plate production process, including:

[0090] Step S101.1: Cleaning the historical multi-source heterogeneous data of the entire thick plate production process;

[0091] Step S101.2: Interpolate the historical multi-source heterogeneous data of the entire thick plate production process after data cleaning to obtain historical multi-source heterogeneous data of equal length;

[0092] Step S101.3: Dimensionality reduction processing is performed on the high-dimensional data in the historical multi-source heterogeneous data of the entire thick plate production process to obtain the historical source heterogeneous data after dimensionality reduction.

[0093] The preprocessing process for historical multi-source heterogeneous data from the entire thick plate production process is identical to the preprocessing process in step S2, differing only in the input data: historical data is input in step S101, while test data or real-time data is input in step S2. In the actual thick plate production process, due to the varying lengths of the plates produced, the lengths of the one-dimensional time series data in the heating, rolling, and cooling processes vary significantly, based on the original sampling rates. Therefore, to fully leverage the correlation between the data in these three processes, linear interpolation is performed on the unequal-length samples in the extracted high-dimensional data to unify the dimensions of all the sample data.

[0094] Specifically, this embodiment performs data cleaning and preprocessing on the selected 5000 thick plate data of the same specification, removes abnormal data samples in the process data, and obtains the remaining 4504 thick plate data. The linear interpolation method is implemented as follows:

[0095] According to the Lagrange interpolation method, the intermediate values of the adjacent data points on the left and right of the one-dimensional time series data are estimated;

[0096] If the coordinates (x0, y0) and (x1, y1) are known, calculate the value of x at a certain position in the interval [x0, y0], and calculate the polynomial y = L1(x) = a0 + a1x, satisfying L1(x0) = y0, L1(x1) = y1. From analytic geometry, we know that:

[0097]

[0098] The mathematical expression of the linear interpolation method is:

[0099]

[0100] Wherein, L(x) is the estimated value of the intermediate value, a0 and a1 are the coefficients of the difference polynomial, specifically, a0 is the intercept of the linear function, and a1 is the slope of the linear function.

[0101] In order to reduce the redundancy between features and facilitate subsequent fusion with one-dimensional process data, high-dimensional data needs to be reduced in dimensionality. This embodiment uses PCA (Principal Component Analysis) for dimensionality reduction.

[0102] Step S102: extracting features from the pre-processed historical multi-source heterogeneous data according to different processes to obtain features corresponding to different processes;

[0103] Step S102.1: Divide the historical multi-source heterogeneous data of equal length and the historical source heterogeneous data after dimensionality reduction into variables according to different processes;

[0104] Step S102.2: splicing the historical multi-source heterogeneous data of equal length after variable division and the historical source heterogeneous data after dimensionality reduction according to the chronological order of different processes;

[0105] Step S102.3: normalize the spliced historical multi-source heterogeneous data;

[0106] Step S102.4: Input the normalized historical multi-source heterogeneous data into different convolutional neural networks corresponding to different processes to perform feature extraction and obtain features corresponding to different processes.

[0107] In this embodiment, the process of extracting the corresponding features of different processes is the same as step S3, and the only difference is the input data. Step S102 inputs historical data, while step S3 inputs test data or data acquired in real time. Specifically, first, according to different processes (including: heating process, rolling process, cooling process), the historical multi-source heterogeneous data of equal length and the historical source heterogeneous data after dimensionality reduction are divided into variables; secondly, the historical multi-source heterogeneous data of equal length after variable division and the historical source heterogeneous data after dimensionality reduction are spliced according to the process. Furthermore, the spliced historical multi-source heterogeneous data are normalized; in the thick plate production process, there are large differences in the distribution and range of each variable data. Therefore, before predicting and analyzing the thick plate shape quality, it is necessary to normalize each variable data in the thick plate sample. The normalization calculation formula is as follows:

[0108]

[0109] Among them, x np represents the pth variable data in the nth data sample, and x min Indicates the minimum value of the variable in all data samples, x max Indicates the maximum value of the variable in all data samples.

[0110] In a specific implementation, the normalized historical multi-source heterogeneous data can be divided into a training set and a test set in a ratio of 8:2, wherein the training set sample data is used to train the parameters of the model, while the test set sample data does not participate in the training and is used for the final evaluation of the model performance after the model training is completed. Finally, the normalized historical multi-source heterogeneous data is input into different convolutional neural networks corresponding to different processes for feature extraction to obtain features corresponding to different processes. This embodiment uses three convolutional neural networks to further extract features from the normalized historical multi-source heterogeneous data to obtain a deeper expression of the input data of the three processes. Each feature extraction network contains a convolutional layer (Conv layer), a dropout layer (Dropout layer), a flattening layer (Flatten layer) and a fully connected layer (Dense layer).

[0111] In step S102.4, the normalized historical multi-source heterogeneous data is input into different convolutional neural networks corresponding to different processes to perform feature extraction to obtain features corresponding to different processes, including:

[0112] Step S102.4.1: Input the normalized historical multi-source heterogeneous data into a convolutional layer with different convolution kernels and the same filter;

[0113] In this embodiment, the convolution layer is used to extract features from the input data. The convolution layer in each network is set with different sizes of convolution kernels (the convolution kernel of the convolution layer of the heating process is 5×5, the convolution kernel of the convolution layer of the rolling process is 9×9, and the convolution kernel of the convolution layer of the cooling process is 7×7) and the same number of filters (32) according to the different input data. The convolution layer operation process is as follows:

[0114]

[0115] Among them, x(i,j) represents the pixel value of the i-th row and j-th column of the input X; K represents the size of the convolution kernel, W(m,n) represents the weight of the m-th row and n-th column, and W b represents the bias term of the convolution kernel, X is the normalized historical multi-source heterogeneous data, u CNN,i is the output of the i-th convolutional layer.

[0116] Step S102.4.2: Use the output of each convolutional layer as the input of the attenuation layer;

[0117] In this example, a Dropout layer is applied after each convolutional layer, with the dropout rate parameter set to 0.2 to prevent overfitting during training. The output after Dropout is calculated as follows:

[0118]

[0119] Among them, ⊙ represents the Hadamard product, that is, element-wise multiplication, and M is the same as the input u CNN,i The binary mask matrix of the same size, Y is the output of the attenuation layer, and p is the dropout rate.

[0120] Step S102.4.3: Use the output of the attenuation layer as the input of the flattening layer;

[0121] Step S102.4.4: Convert the output of the flattening layer into a one-dimensional vector, and input the one-dimensional vector into the fully connected layer to obtain features corresponding to different processes;

[0122] In this example, after the Dropout layer, the output is flattened, converting the multidimensional output into a one-dimensional vector that can be input into the fully connected layer. After each Flatten layer, a fully connected layer with 32 neurons is applied. These layers help learn complex patterns in the data. The calculation formula for the fully connected layer is as follows:

[0123] y flatten =W T x+b

[0124] Among them, x is the input data of the fully connected layer, W is the weight, b is the bias, and y is the flatten is the output of the fully connected layer.

[0125] Among them, both the convolutional layer and the fully connected layer use Relu as the activation function:

[0126] f Relu (x r )=max(0,x r )

[0127] Among them, x r Represents the activation function input data, f Relu (x r ) is the activation function.

[0128] Compared with other activation functions such as Sigmoid and Tanh, Relu is easier to optimize and can effectively avoid the problem of gradient disappearance during gradient descent, thereby reducing the risk of information loss.

[0129] Step S103: using features corresponding to different processes as input to train the collaborative attention mechanism model;

[0130] Step S104: Adaptive Moment Estimation (ADAM) is used to optimize the loss function of the collaborative attention mechanism model to obtain a pre-trained collaborative attention mechanism model.

[0131] In this embodiment, after extracting features from each process, in order to fully consider the effects and connections between variables in each process, a collaborative attention mechanism is used to fuse the features of the three processes, and then the fused features are used for subsequent flatness quality prediction. In this embodiment, step S102 obtains features corresponding to three different processes, including: features corresponding to the heating process, features corresponding to the rolling process, and features corresponding to the cooling process. The features corresponding to the heating process, the rolling process, and the cooling process are used as three input sources X heat 、X roll and X cool ,in and n1, n2, n3 are the number of variables in the heating, rolling, and cooling processes, respectively, and d is the variable dimension. Each input source corresponds to the feature representation of the heating process, rolling process, and cooling process. To calculate the fused feature representation, the specific implementation steps are as follows:

[0132] First, the input source X of the heating process heat Convert it into a feature vector and add this feature vector to the input source X of the rolling process roll On top, generate data X heat Based on data X roll The attention weight A' (i.e., Q1 of the first collaborative attention layer comes from the heating process, K1 and V1 come from the rolling process) is used to achieve the purpose of the forward process (heating process) guiding the backward process (rolling process).

[0133] Specifically, let and The collaborative attention weight is calculated as follows:

[0134]

[0135] in, is the learned projection matrix, d k is the dimension of the key-value set (also the dimension of the query set), f o is a linear layer X' that projects the output weights to the target channel dimension, represents the dot product calculation result of the query vector Q1 and the key vector, represents the value calculated by query vector Q1, represents the result of weighted summation of the value vector V1 using the attention weight A', The final attention output result of all the above steps is represented by Attn(·), which represents the operation function of the attention mechanism. Here, it specifically refers to the entire process of calculating the query, key, and value vectors through the attention mechanism.

[0136] Similarly, convert X' into a feature vector and add this feature vector to the input source X of the cooling process cool On top, generate data X cool Based on the attention weight A of data X' (where Q2 comes from the fusion feature X' of the first two processes, K2 and V2 to cool the process), the specific calculation formula is the same as above. The collaborative attention fusion module structure is as follows Figure 5 shown.

[0137] The above process not only realizes the fusion of the three process features, but also reflects the guiding role of the forward process on the backward process through the design of the attention mechanism query vector Q (Query), key vector K (Key) and value vector V (Value), and fully considers the connection between the various processes in the actual production process. Compared with the current research method, this embodiment pays more attention to how to integrate multi-source heterogeneous data in the thick plate production process, that is, to perform feature fusion of thick plate data of different dimensions (including one-dimensional data and two-dimensional data in the thick plate production process) to achieve plate shape quality prediction. This article, based on the consideration of the characteristics of multi-source heterogeneous data of thick plates, further considers the characteristics of multiple processes and strong coupling in the actual production process of thick plates (that is, there are interactions and influences between the front and back processes of thick plate production). Through the collaborative attention mechanism, the Q, K, and V of the three processes are cross-designed to explore the influence of the interaction between the variables of each process on the final thick plate shape quality.

[0138] Finally, during the training of the entire network, Adam is used to optimize the loss function to adapt to the update speed of different parameters. The network weight and bias update calculation formula is as follows:

[0139] (1) Update the estimation of the momentum and bias of the loss function. The specific calculation formula is as follows:

[0140]

[0141] Among them, S t and R t They represent the first-order and second-order moment estimates of the gradient of the loss function; β1 and β2 are the decay rates, g t Represents the gradient value at time step t.

[0142] (2) Correct the deviation. The specific calculation formula is as follows:

[0143]

[0144] in, and represents the corrected first-order and second-order moment estimates.

[0145] (3) Update the parameters of the entire network. The specific update calculation formula is as follows:

[0146]

[0147] Among them, θ t is the parameter at the tth iteration, θ t+1 is the parameter updated after the tth iteration, η represents the learning rate, and ε represents the smoothing term, which is set to 10-8 to prevent division by zero.

[0148] In order to further illustrate the effect of this embodiment, the experimental process is described in detail as follows:

[0149] 1. Experimental environment configuration and parameter setting

[0150] This experiment uses Python to extract and normalize the data, merge and resample the multi-source heterogeneous data, and use the PyTorch framework to establish a steel plate shape quality prediction model based on the forward process guided attention mechanism feature fusion. The activation function uses the Relu function and Softmax is used for thick plate shape classification. Figure 4 、 Figure 5 The algorithm was designed based on the feature fusion model structure shown in the figure, and the network was trained using the cross-entropy loss function. Backpropagation used the Adam optimization algorithm to learn network parameters. After multiple experiments, the optimal learning rate was selected as 0.001, the random seed number random_state was set to 42, the batch_size was set to 512, and the number of epochs for the entire model was set to 50. Network training was performed on a computer with an Intel(R) Xeon(R) W-2123, 3.60GHz, 8-core CPU, 32GB of RAM, and a Quadro P2000 graphics card.

[0151] 2. Evaluation Metrics

[0152] In this example, the accuracy, F1 score, and area under the receiver operating characteristic curve (ROC) (AUC), referred to as AUC-ROC, are used to evaluate the prediction performance of the method of the present invention. The larger the AUC-ROC value, the better the prediction performance of the model. The calculation formulas for each evaluation index are as follows:

[0153] (1) Accuracy

[0154]

[0155] (2) F1-Score

[0156]

[0157] in, represents the accuracy, Recall represents the recall rate. TP represents the number of samples correctly identified as positive by the model, FP represents the number of samples that are incorrectly labeled as positive but actually belong to the negative class, TN represents the number of samples correctly identified as negative, and FN represents the number of samples that are incorrectly classified as negative but actually belong to the positive class.

[0158] 3. Experimental results

[0159] The experiment was conducted on the above sample data and the parameters were set as above. The ROC curve and confusion matrix of the obtained model are as follows: Figure 6 and Figure 7 shown.

[0160] In addition, in order to more fully illustrate the effectiveness of the forward process-guided steel thick plate shape quality prediction method proposed in this embodiment, the results of the prediction method of this embodiment are compared with other common algorithms. The comparative experimental results are shown in Table 3 below.

[0161] Table 3 Comparison of prediction results between the prediction method of the present invention and common algorithms

[0162]

[0163]

[0164] It can be seen from Table 3 that the forward process-guided steel thick plate shape quality prediction method (i.e., Co-Attention Fusion) of this embodiment has the highest accuracy, and the AUC value of the area under the ROC curve is also closer to 1, indicating that the forward process-guided steel thick plate shape quality prediction method proposed in this embodiment has better classification performance.

[0165] In the table of experimental results, the first five are the comparison results of plate shape quality prediction with other common algorithms. The sixth and seventh (DNN-CNN Fusion and SE-CNN-Attention Fusion) are the comparison results of plate shape quality prediction by directly fusing data between different dimensions. The last one is the forward process-guided steel thick plate shape quality prediction method proposed in this embodiment. It can be seen that the accuracy of thick plate shape quality prediction is higher by considering the connection between the front and back processes. The prediction method of the embodiment is compared with the prediction results of the direct feature fusion method DNN-CNN Fusion and SE-CNN-AttentionFusion, and the box shape is obtained. Figure 8 As shown. Figure 8 It can also be seen that the steel thick plate shape quality prediction method guided by the forward process of this embodiment has the highest accuracy and the best effect.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present application.

[0167] Example 2:

[0168] In the second aspect, the present application proposes a forward process guided steel plate shape quality prediction device, such as Figure 9 As shown, it includes: data acquisition module, data preprocessing module, process feature extraction module, fusion feature extraction module and quality prediction module;

[0169] The data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the process feature extraction module, the process feature extraction module is connected to the fusion feature extraction module, and the fusion feature extraction module is connected to the quality prediction module;

[0170] The data acquisition module is used to obtain multi-source heterogeneous data of the entire thick plate production process;

[0171] The data preprocessing module is used to preprocess the multi-source heterogeneous data of the entire thick plate production process;

[0172] The process feature extraction module is used to extract features from the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes, and obtain the features corresponding to different processes;

[0173] The fusion feature extraction module is used to input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fusion features;

[0174] The quality prediction module is used to input the fusion features into the pre-trained prediction model to obtain the steel plate shape quality prediction result.

[0175] Example 3:

[0176] This embodiment proposes an electronic device, comprising: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the forward process-guided steel thick plate shape quality prediction method.

[0177] The electronic device can be a mobile phone, computer, or tablet computer, and includes a memory and a processor. The memory stores a computer program that, when executed by the processor, implements the forward-process-guided steel plate shape quality prediction method described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and a communication component.

[0178] The processor is configured to execute all or part of the steps of the forward-process-guided steel plate shape quality prediction method described in the above-described embodiment. The memory is configured to store various types of data, such as instructions for any application or method in the electronic device, as well as data related to the application.

[0179] The processor can be an application specific integrated circuit (ASIC), a digital signal processor (DSP), a programmable logic device (PLD), a field programmable gate array (FPGA), a controller, a microcontroller, a microprocessor or other electronic components, and is used to execute the forward process guided steel thick plate shape quality prediction method described in the above embodiment.

[0180] Example 4:

[0181] This embodiment provides a computer-readable storage medium storing executable instructions. When the instructions are executed, if they are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0182] The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the forward process-guided steel thick plate shape quality prediction method described in each embodiment of the present application.

[0183] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (for example, SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR abbreviation, memory data register) memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, CD, server, APP (Application, abbreviation of application software) application store and other media that can store program verification codes, on which a computer program is stored. When the computer program is executed by the processor, the various steps of the above-mentioned forward process-guided steel thick plate plate shape quality prediction method can be implemented.

[0184] Example 5:

[0185] This embodiment provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the forward process-guided steel plate shape quality prediction method.

[0186] Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution can be embodied in the form of a computer program product.

[0187] The various embodiments in this application are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0188] The scope of protection of this application is not limited to the above-described embodiments. Obviously, those skilled in the art may make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, the disclosure is intended to include such modifications and variations.

Claims

1. A forward process-guided steel plate shape quality prediction method, characterized in that: include: Acquire multi-source heterogeneous data for the entire thick plate production process; Preprocessing multi-source heterogeneous data of the entire thick plate production process; Feature extraction is performed on the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain the features corresponding to different processes; Input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fused features; Inputting the fusion features into a pre-trained prediction model to obtain a prediction result of the shape quality of the steel thick plate; The different processes include: a heating process, a rolling process and a cooling process; The pre-trained collaborative attention mechanism model includes: A first collaborative attention layer and a second collaborative attention layer; the first collaborative attention layer includes: a first query vector, a first key vector and a first value vector; the second collaborative attention layer includes: a second query vector, a second key vector and a second value vector; the characteristics of the heating process are superimposed on the first query vector as the input of the first collaborative attention layer, and the characteristics of the rolling process are superimposed on the first key vector and the first value vector respectively as the attention weight of the first collaborative attention layer, the output of the first collaborative attention layer is superimposed on the second query vector as the input of the second collaborative attention layer, the cooling process is superimposed on the second key vector and the second value vector respectively as the attention weight of the second collaborative attention layer, and the output of the second collaborative attention layer is used as the fusion feature.

2. The forward process-guided steel plate shape quality prediction method according to claim 1, characterized in that: The preprocessing of multi-source heterogeneous data of the entire thick plate production process includes: Performing data cleaning on multi-source heterogeneous data of the entire thick plate production process; The multi-source heterogeneous data of the entire thick plate production process after data cleaning are interpolated to obtain multi-source heterogeneous data of equal length; the high-dimensional data in the multi-source heterogeneous data of the entire thick plate production process are subjected to dimensionality reduction processing to obtain the source heterogeneous data after dimensionality reduction.

3. The forward process-guided steel plate shape quality prediction method according to claim 2, characterized in that: The feature extraction is performed on the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes to obtain features corresponding to different processes, including: According to different processes, multi-source heterogeneous data of equal length and source heterogeneous data after dimensionality reduction are divided into variables; According to the time sequence of different processes, the multi-source heterogeneous data of equal length after the variables are divided and the source heterogeneous data after dimensionality reduction are spliced; Normalize the spliced multi-source heterogeneous data; The normalized multi-source heterogeneous data are input into different convolutional neural networks corresponding to different processes for feature extraction to obtain the features corresponding to different processes.

4. The forward process-guided steel plate shape quality prediction method according to claim 3, characterized in that: The different convolutional neural networks include: a convolution layer, an attenuation layer, a flattening layer and a fully connected layer.

5. A forward process guided steel plate shape quality prediction device, characterized in that: include: The data acquisition module is used to obtain multi-source heterogeneous data of the entire thick plate production process; The data preprocessing module is used to preprocess the multi-source heterogeneous data of the entire thick plate production process; The process feature extraction module is used to extract features from the pre-processed multi-source heterogeneous data of the entire thick plate production process according to different processes, and obtain the features corresponding to different processes; The fusion feature extraction module is used to input the features corresponding to different processes into the pre-trained collaborative attention mechanism model to obtain fusion features; The quality prediction module is used to input the fusion features into the pre-trained prediction model to obtain the steel plate shape quality prediction result; The different processes include: a heating process, a rolling process and a cooling process; The pre-trained collaborative attention mechanism model includes: A first collaborative attention layer and a second collaborative attention layer; the first collaborative attention layer includes: a first query vector, a first key vector and a first value vector; the second collaborative attention layer includes: a second query vector, a second key vector and a second value vector; the characteristics of the heating process are superimposed on the first query vector as the input of the first collaborative attention layer, and the characteristics of the rolling process are superimposed on the first key vector and the first value vector respectively as the attention weight of the first collaborative attention layer, the output of the first collaborative attention layer is superimposed on the second query vector as the input of the second collaborative attention layer, the cooling process is superimposed on the second key vector and the second value vector respectively as the attention weight of the second collaborative attention layer, and the output of the second collaborative attention layer is used as the fusion feature.

6. An electronic device, characterized in that: include: One or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the forward process-guided steel thick plate shape quality prediction method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that The device stores executable instructions, which, when executed, enable a processor to execute the forward process-guided steel thick plate shape quality prediction method according to any one of claims 1 to 4.

8. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the forward process-guided steel plate shape quality prediction method according to any one of claims 1 to 4 is implemented.

Images