Coiling temperature prediction method based on multi-level feature fusion under cloud edge-end cooperation

Abstract

The invention discloses a coiling temperature prediction method based on multi-level feature fusion under cloud edge-end cooperation, and belongs to the technical field of industrial control, and the method comprises the steps: deploying a manufacturing execution layer, a process control layer and a real-time control layer at a cloud side, an edge side and an end side respectively; the method comprises the following steps: collecting current strip steel process data of a process control layer and steel coil setting data of a manufacturing execution layer, and preprocessing the collected data on an end side; feature extraction is conducted on the preprocessed current strip steel process data on the side, and bottom layer working condition information of the current plate blank is obtained; and feature extraction is conducted on the preprocessed steel coil setting data on the cloud side, upper-layer production scheduling setting information of the next slab is obtained, bottom-layer working condition information of the current slab and the upper-layer production scheduling setting information of the next slab are fused, multi-level fusion features are obtained, and coiling temperature prediction is achieved based on the multi-level fusion features. According to the scheme, the coiling temperature prediction precision can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of industrial control technology, and in particular to a coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration. Background Art

[0002] In complex industrial processes, key process parameters have an irreplaceable impact on product quality. For example, in the hot strip rolling mill (HSRM) process, the coiling temperature (CT) has a significant impact on the mechanical properties of the steel coil. Most existing CT prediction methods only consider a single level of information, and do not fully consider the impact of multiple levels of information in the actual process, resulting in poor prediction results.

[0003] The working conditions of hot rolling process are complex and changeable, and the continuous change of temperature is affected by many factors, such as strip quality, speed, thickness, cooling water volume, water pressure, final rolling temperature, heat conduction, convection, radiation conditions, etc. The interaction and influence between these factors make the prediction of coiling temperature extremely complicated. In order to meet the requirements for the performance of finished strip steel in the production process, the coiling temperature needs to have a high prediction accuracy. Therefore, the accuracy of the prediction model has become an urgent problem to be solved. Coiling temperature control is one of the core links in the production of hot-rolled strip steel. It not only directly affects the final performance of the strip steel, but also is the decisive factor for whether the strip steel can be coiled smoothly and safely produced. High-precision control of coiling temperature can optimize the microstructure of the strip steel and improve the mechanical properties and dimensional accuracy of the product. Therefore, precise control of coiling temperature is a key link to ensure the production quality of hot-rolled strip steel.

[0004] In the modern steel industry, laminar cooling is a process that improves the structural properties of steel strips, improves the quality and output of steel strips by forced water cooling after rolling, which has a direct and huge impact on the accuracy of coiling temperature control. At the same time, the laminar cooling process has complex industrial characteristics such as drastic changes in working conditions, strong nonlinearity, time-varying parameters, and mathematical models that are difficult to accurately describe. Therefore, the method based on the mechanism model is difficult to accurately describe, which brings certain difficulties to process modeling. The data-driven method directly learns and extracts features from data and is suitable for complex and changeable industrial fields. It has been successfully applied in the fields of suspended magnetization roasting process, blast furnace ironmaking process, hydrocracking process, petroleum refining, wastewater treatment, etc., and has achieved good results.

[0005] With the continuous development of machine learning algorithms and the continuous accumulation of process history data, machine learning-based methods have been widely developed. Support Vector Machine (SVM), Random Forest (RF) and other methods are difficult to fully extract the time series characteristics of the complex hot rolling process, and the prediction accuracy is limited. In addition to traditional machine learning-based methods, many improved neural network methods based on intelligent optimization algorithms have also been developed for temperature prediction. These methods do not fully consider the characteristics of multi-level information in the process and cannot guarantee high accuracy of temperature prediction modeling.

[0006] In order to extract the features of continuous time series data, Long Short Term Memory (LSTM) networks and Gated Recurrent Unit (GRU) networks developed for time series data modeling are usually used. The GRU network is a variant of the LSTM network, which has the advantages of simpler architecture, fewer parameters and higher computational efficiency. At the same time, the attention mechanism has been incorporated into various GRU models to extract local features. However, the above methods do not fully consider the influencing variables of the target variable and do not fully extract the multi-scale features of the variable. In addition, in modern industrial processes, traditional prediction methods only use information of a single scale of the variable during data preprocessing, which may lose some useful information. Summary of the invention

[0007] The present invention provides a coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration to solve the technical problem of low prediction accuracy of traditional coiling temperature prediction methods.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] On the one hand, the present invention provides a coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration, which is used for coiling temperature prediction in a hot rolling industrial process, wherein the hot rolling industrial process is divided into a manufacturing execution layer, a process control layer, and a real-time control layer from top to bottom; the method comprises:

[0010] The manufacturing execution layer is deployed on the cloud side, the process control layer is deployed on the edge side, and the real-time control layer is deployed on the device side.

[0011] Collect the current strip process data of the process control layer and the steel coil setting data of the manufacturing execution layer, and pre-process the collected current strip process data and steel coil setting data on the terminal side;

[0012] On the side, feature extraction is performed on the pre-processed current strip process data to obtain the underlying working condition information of the current slab;

[0013] On the cloud side, feature extraction is performed on the pre-processed coil setting data to obtain the upper-level production setting information for the next slab;

[0014] On the cloud side, the underlying working condition information of the current slab is fused with the upper-level production scheduling setting information of the next slab to obtain multi-level fusion features, and the coiling temperature prediction is achieved based on the multi-level fusion features.

[0015] Furthermore, the current strip process data includes: the actual value of the finishing exit thickness and the set value of the finishing exit thickness, the actual value of the finishing exit width and the finishing exit width deviation value, the finishing exit speed, the first intermediate speed, the second intermediate speed, the finishing exit temperature, the laminar cooling intermediate temperature, the water tank water temperature, the rapid cooling section actuator temperature, the rapid cooling first end, the rapid cooling second section, and the water temperature of each actuator above and below the rapid cooling 18th section, the valve position, the rapid cooling section actuator pressure, and the flow increment.

[0016] Furthermore, the steel coil setting data includes: a finishing temperature setting value, a thickness setting value, a width setting value, a coiling temperature setting value, and an actual coiling temperature of a previous steel coil.

[0017] Furthermore, the collected current strip process data and coil setting data are preprocessed, including:

[0018] Perform data interception and alignment operations on the current strip process data and coil setting data;

[0019] The discrete variables in the current strip process data are subjected to time window truncation and downsampling processing to obtain multi-scale discrete data; wherein the discrete variables include the switch status of the manifold at the top and bottom of each zone.

[0020] Furthermore, the data interception and alignment operation includes:

[0021] The collected data is used for slab tracking, that is, the data is intercepted and processed according to different slab numbers, and the processed data is stored separately by slab numbers. Specifically, the time interval of the same slab is intercepted according to the finishing exit and coiling mark positions, and then the position and time of the manifold passing through different sections are calculated according to the speed and time of the slab. Finally, the data in the length direction with an interval of 1 meter is aligned based on the total length of the slab.

[0022] Furthermore, the time window interception and downsampling processing includes:

[0023] Perform time window interception on discrete variables, where the step size of the time window is set to 1, the width of the time window is w, the number of time windows is n, and w and n are both preset values;

[0024] After completing the time window interception, the time window is downsampled to obtain multi-scale discrete data.

[0025] Furthermore, feature extraction is performed on the preprocessed current strip process data, including:

[0026] The GRU layer with fusion attention mechanism is used to extract features of the current strip process data that has completed the alignment operation, so as to extract the continuous feature vector of the steel coil on the process control layer; and SlowFast Networks is used to extract features of multi-scale discrete data, so as to extract the discrete feature vector of the steel coil on the process control layer.

[0027] Furthermore, when SlowFast Networks is used to extract features from multi-scale discrete data, the lateral connection result of the second residual network in SlowFast Networks is used as the extracted feature.

[0028] Furthermore, feature extraction is performed on the preprocessed coil setting data, including:

[0029] SENet is used to extract features from the setting data of the steel coil that has completed the alignment operation, and the feature vector of the next steel coil in the manufacturing execution layer is obtained.

[0030] Furthermore, the bottom-level working condition information of the current slab and the upper-level production scheduling setting information of the next slab are feature-fused to obtain multi-level fusion features, and the coiling temperature prediction is realized based on the multi-level fusion features, including:

[0031] The discrete feature vector of a steel coil on the process control layer is averaged and flattened, and then spliced ​​with the continuous feature vector of a steel coil on the process control layer to obtain the feature vector of the steel coil on the process control layer;

[0032] The feature vector of a steel coil on the process control layer is averaged over time steps and then concatenated with the feature vector of a steel coil on the manufacturing execution layer to obtain a multi-level fusion feature.

[0033] The multi-level fusion features are passed through two fully connected layers to obtain the predicted value of the coiling temperature.

[0034] On the other hand, the present invention further provides an electronic device, comprising a processor and a memory; wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the above method.

[0035] In yet another aspect, the present invention further provides a computer-readable storage medium, wherein at least one instruction is stored in the storage medium, and the instruction is loaded and executed by a processor to implement the above method.

[0036] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0037] 1. The present invention proposes an industrial process coiling temperature method that integrates multi-level features. The method uses the GRU layer and SlowFast Networks that integrate the attention mechanism to extract the features of the hot rolling process control layer, extract the multi-scale features of the process variables, and obtain the status information such as the bottom operating conditions of the slab. Then, a higher-level feature extraction is performed on the manufacturing execution layer, and the features of the setting data are extracted through SENet to obtain the top-level production scheduling information of the slab. After that, the multi-level extracted from the manufacturing execution layer and the process control layer are fused through the fully connected layer to finally obtain the predicted value of the coiling temperature.

[0038] 2. The present invention deploys the proposed prediction model in a cloud-edge-end collaborative framework for application verification. The manufacturing execution layer, process control layer, and real-time control layer are deployed on the cloud side, edge side, and end side, respectively. The end-side real-time control layer implements preprocessing tasks such as process data alignment and sliding time windows; the edge-side process control layer performs feature extraction on the strip process data, extracts the underlying working condition information, and obtains the process control layer features; the cloud-side manufacturing execution layer performs feature extraction on the steel coil setting data, extracts the top-level scheduling and production scheduling, and also receives the features extracted by the edge-side process control layer, and then integrates the multi-system-level features of the manufacturing execution layer and the process control layer, and finally realizes the real-time prediction of the coiling temperature. The effectiveness and real-time performance of this method were verified using actual industrial data. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0040] Figure 1 It is a schematic diagram of hot rolling process;

[0041] Figure 2 It is a schematic diagram of the laminar cooling process;

[0042] Figure 3 It is a framework diagram of a coiling temperature prediction method based on multi-level feature fusion provided by an embodiment of the present invention;

[0043] Figure 4 is a schematic diagram of the end-side preprocessing provided by an embodiment of the present invention;

[0044] Figure 5 is a schematic diagram of a data alignment process provided by an embodiment of the present invention;

[0045] Figure 6is a schematic diagram of coiling temperature prediction results provided by an embodiment of the present invention;

[0046] Figure 7 It is a deployment diagram of a model under a cloud-edge-device collaboration framework provided by an embodiment of the present invention;

[0047] Figure 8 It is a system block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0049] First of all, it should be noted that in the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "exemplary" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "exemplarily" is intended to present the concept in a concrete way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0050] First embodiment

[0051] This embodiment provides a coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration, which is used for coiling temperature prediction in hot rolling industrial process. First, complex industrial processes (such as hot rolling industrial processes) can be divided into manufacturing execution layer, process control layer and real-time control layer from top to bottom. The process control layer receives the job plan of the manufacturing execution layer and the operation data of the process control layer, and feeds back the actual operation and operation instructions to them respectively. Secondly, considering the high efficiency of cloud-side computing and the flexibility and real-time performance of edge-side computing, the manufacturing execution layer, process control layer and real-time control layer are deployed on the cloud side, edge side and end side respectively. The end-side real-time control layer implements preprocessing such as process data alignment and sliding time window; the edge-side process control layer extracts features of process data, extracts the underlying working condition information, and obtains the process control layer features; the cloud-side manufacturing execution layer extracts features of steel coil setting data, extracts top-level scheduling and production scheduling information, and also receives the process control layer features extracted by the edge-side process control layer, and then fuses the multi-system-level features of the manufacturing execution layer and the process control layer, and finally realizes coiling temperature prediction. The application of this method has been verified in the prediction of coiling temperature in an actual hot rolling production line: the average absolute error of the method in this paper is reduced by 53.01% on average compared with other traditional methods; the effectiveness and real-time performance of the method are verified using actual industrial data.

[0052] The method can be implemented by an electronic device. Specifically, the method includes the following steps:

[0053] S1, deploy the manufacturing execution layer, process control layer, and real-time control layer on the cloud side, edge side, and device side respectively;

[0054] S2, collecting the current strip process data of the process control layer and the steel coil setting data of the manufacturing execution layer, and preprocessing the collected current strip process data and steel coil setting data on the terminal side;

[0055] S3, extracting features of the pre-processed current strip process data on the side to obtain the underlying working condition information of the current slab;

[0056] S4, extracting features from the pre-processed coil setting data on the cloud side to obtain the upper-level production setting information for the next slab;

[0057] S5, on the cloud side, the bottom-level working condition information of the current slab is integrated with the upper-level production scheduling setting information of the next slab to obtain a multi-level integrated feature, and the coiling temperature is predicted based on the multi-level integrated feature.

[0058] Among them, it should be noted that the hot rolling process is as follows Figure 1 As shown. The process is generally composed of the following units: heating furnace, rough rolling reversible rolling mill, insulation cover, flying shear, descaling machine, finishing rolling mill, laminar cooling, and coiling unit. During the laminar cooling process, the strip undergoes complex water cooling, air cooling heat exchange and internal heat conduction processes. It has complex industrial characteristics such as drastic changes in working conditions, strong nonlinearity, time-varying parameters, and mathematical models that are difficult to accurately describe. Therefore, a data-driven deep learning method is used to fully extract the characteristics of variables at multiple time scales. Figure 2 As shown in the figure, laminar cooling mainly controls the cooling speed and coiling temperature of the steel plate by changing the roller speed and the switch state and flow rate of the water curtain header. Therefore, the switch state of the header in the laminar cooling process has an irreplaceable influence on the coiling temperature. These process data of the strip reflect the working condition information of the bottom layer of the equipment end, and feature extraction of these data can reflect the working status of the equipment.

[0059] The multi-level framework of industrial processes such as Figure 3 As shown in the figure, it is mainly divided into real-time control layer, process control layer and manufacturing execution layer. The bottom real-time control layer uploads the operation data to the process control layer, and the operation performance is further uploaded to the manufacturing execution layer in this layer. The manufacturing execution layer formulates and sends the operation plan to the process control layer according to the production control requirements, and then provides operation instructions for the bottom real-time control layer.

[0060] The framework of the coiling temperature prediction method based on multi-level feature fusion is as follows Figure 3As shown in the figure, this method makes full use of multi-level information such as the bottom-level working condition information and the upper-level production scheduling relationship. First, the method obtains data at different levels, including the current strip process data obtained from the process control layer and the next steel coil setting data obtained from the manufacturing execution layer. Then, the current strip is feature extracted at the process control layer to obtain the bottom-level working condition information. Then, the manufacturing execution layer feature extraction is performed on the setting data of the top-level manufacturing execution layer to obtain the top-level production scheduling information. Then, the features extracted from the manufacturing execution layer and the features extracted from the process control layer are multi-level feature fused to finally realize the coiling temperature prediction of the next steel coil.

[0061] Combining the above modules and techniques, we get a coiling temperature prediction model that integrates multi-level information. The multi-level coiling temperature is conducive to timely discovering potential problems and providing production scheduling suggestions for the manufacturing execution layer. On the other hand, it is conducive to optimizing the process parameters of the process control layer and improving the safety and reliability of the process.

[0062] The data preprocessing and feature extraction operations are described in detail below.

[0063] 1. Data preprocessing on the client side

[0064] The real-time control layer is deployed on the client side, mainly for preprocessing operations such as data alignment and sliding time windows. The implementation principle is as follows: Figure 4 As shown. The data of the same slab during hot rolling is continuous, while the data of different slabs is discontinuous. The coiling temperature prediction model established in this paper is mainly based on effective rolling data. Therefore, it is necessary to intercept the effective rolling data of different slabs and perform corresponding preprocessing to obtain partial slab data. After collecting the field data, it is necessary to track the data, that is, intercept and process it according to different slab numbers, and store the preprocessed data according to the slab number. The number of slab tracking is m. Because the effective data of different cooling sections of the same slab laminar cooling are not aligned in the time dimension, it is necessary to align the data according to the length percentage. Data alignment first intercepts the time interval of the same slab according to the finishing exit and coiling mark, and then calculates the position and time of the manifolds passing through different sections according to the speed and time of the slab. Finally, the data alignment in the length direction with an interval of 1 meter is achieved based on the total length L of the slab, and the data of each part of the slab passing through each section of the manifold is obtained.

[0065] In order to fully extract the characteristics of discrete data at multiple time scales, it is also necessary to perform time window operations on the data. The discrete variables with a width of w and a number of variables of d are intercepted, and n time windows are intercepted as a group, corresponding to the continuous variables with a number of variables of c at the last moment of the time window, to obtain the final input form of the model.

[0066] 2. Side process data feature extraction

[0067] GRU effectively alleviates the gradient vanishing and gradient exploding problems that traditional RNNs are prone to encounter when dealing with long-term dependency problems. At the same time, in dealing with complex tasks of extracting features from discrete data, GRU has a simpler structure and faster training speed while maintaining performance compared to LSTM. The GRU layer input is obtained by tracking multi-slab information. The GRU unit can be expressed by the following mathematical expression:

[0068] V t =σ(W v ·[H t-1 ,X t ]+b v ) (1)

[0069] R t =σ(W r ·[H t-1 ,X t ]+b r ) (2)

[0070]

[0071] Among them, R t 、V t and H t represents the reset gate, update gate, and output gate, and H t-1 Represents the current memory information and the previous hidden state, W v represents the weight matrix, b v represents the bias vector of the update gate, W r represents the weight matrix, b r Represents the bias vector of the reset gate, W h represents the corresponding weight matrix, b h represents the bias vector, I represents a matrix whose elements are all 1, represents the Hadamard product of two vectors, σ represents the sigmoid activation function, and tanh represents the hyperbolic tangent activation function.

[0072] The attention mechanism is then used to assign weights to different parts of the GRU layer output, highlighting more relevant information and helping the model focus on the parts of the input sequence that are considered more important for the current task. K, Q, and V represent the key, query, and value in the attention mechanism, respectively. The input to the attention mechanism consists of key-value pairs <K, V>. The calculation of the attention mechanism is divided into three steps: first, the correlation between K and Q is calculated, then the weight coefficient of each K corresponding to V is obtained, and finally the weighted sum of the obtained weight coefficient and the V value is obtained to obtain the output of the attention. The calculation method of the attention mechanism weight coefficient is as follows:

[0073] θ=Q·KT (5)

[0074]

[0075] Where, θ is the correlation coefficient between Q and K; θ j is the attention weight coefficient of the jth sample; n represents the number of key-value pairs in the input. The output of the GRU layer is weighted and summed under the attention mechanism to obtain the feature vector z as part of the input of the feature fusion layer. C , z C The calculation is as follows:

[0076]

[0077] Furthermore, the discrete data of the hot rolling laminar cooling process has the characteristics of multiple time scales, and it is necessary to capture the characteristics of fast and slow changes. In this regard, this embodiment uses the SlowFast network to extract features from discrete variables, obtains discrete data of different sampling frequencies through a sliding time window, and uses the SlowFast network for multi-scale feature extraction, and uses lateral connections for fusion. Specifically: the discrete data is intercepted and downsampled through a time window as the input of the fast path and the slow path. The SlowFast Networks used in this article are composed of two paths with different sampling frequencies for feature extraction and fusion. First, two sequences of high frequency and low frequency are obtained through different sampling intervals and are input into the network respectively. Both paths use 3D Resnet convolutional neural networks for spatiotemporal feature extraction. The slow path processes the input data at a low sampling rate to capture spatial features that change slowly over time, and the fast path processes the input data at a higher sampling rate to capture features that change faster. The feature information extracted from the two paths is fused through lateral connections.

[0078] Considering the complexity of the network and the characteristics of the data, this paper uses the lateral connection result of the second residual network in the SlowFast network structure as the feature extracted from the final discrete quantity. The discrete data is first input and sampled through the data layer. The fast path data intercepted by the time window is downsampled in the slow path to obtain the fast path and slow path inputs with a sampling frequency ratio of α, that is, the ratio of the sampling density of the two is α. This generates an n×w×d sequence framework. The generated sequence is then input into the 3D convolution layer for feature extraction.

[0079] The multi-scale feature fusion after each stage is completed through the lateral connection operation. The information extracted by the fast path is fused into the slow path through the lateral connection, which enables the slow path to perceive the information extracted by the fast path and realize the extraction and fusion of multi-scale features. Since the two paths in this paper have different time dimensions, the information obtained from the fast path needs to undergo certain transformation operations before it can be fused into the slow path. The slow path output feature map is represented as {n,w×d,c}, and the fast path output feature map is represented as {α×n,w×d,β×c}. In this paper, the time stride-based convolution is used for lateral connection: the stride is set to α and the size of the output feature map is controlled. The information extracted from the fast path is 3D convolved, and then spliced ​​with the information from the slow path.

[0080] First, in both paths, the variable x of the fast path F and the variable x of the slow path S F is obtained by convolution and pooling operations respectively. 1 and S 1 ,in:

[0081] S 1 =maxpool(relu(conv(x S ))) (8)

[0082] F 1 =maxpool(relu(conv(x F ))) (9)

[0083] Then, the F obtained by the fast path is transformed into 1 Fusion to S 1 Get S 1 ',satisfy:

[0084] S 1 '=[S 1 ;relu(batchnorm(conv(F 1 )))] (10)

[0085] In the second stage, S 1 ' and F1 are respectively subjected to the residual network composed of three groups of convolution operations to obtain S 2 and F 2 ,in:

[0086] S 2 =res 2 (S 1 ') (11)

[0087] F 2 =res2 (F 1 ) (12)

[0088] Then similarly, F 2 Fusion to S via lateral connections 2 Get S 2 ',in:

[0089] S' 2 =[S 2 ;relu(batchnorm(conv(F 2 )))] (13)

[0090] Finally, S' 2 Channel averaging and flattening are performed to obtain the features finally extracted by the SlowFast network.

[0091] 3. Cloud-side setting data feature extraction

[0092] There are abundant multi-level data in the industrial process, including setting data of the manufacturing execution layer, process parameters of the process control layer, and process data of the real-time control layer. These data are interrelated and coupled in a multi-level manner, so it is necessary to fully integrate multi-level information to improve prediction accuracy. Squeeze-and-excitation networks (SENet) enhance the ability of convolutional neural networks to perceive different features by adaptively adjusting the weights of each channel. This mechanism enables the network to pay more attention to information-rich feature channels, which is conducive to extracting and improving the accuracy and efficiency of feature extraction at the manufacturing execution layer.

[0093] Based on this, this embodiment uses a SENet-based model to extract features from the data of the top manufacturing execution layer. The model can automatically learn to enhance the model's attention to important channels and weaken the attention to unimportant channels by learning a weight coefficient on each channel and weighting them separately, thereby making more effective use of input features. The slab setting information is obtained through convolution and batch normalization to obtain E, which satisfies:

[0094] E = batchnorm(conv(x E )) (14)

[0095] SENet mainly includes three processes: squeezing, excitation, and recalibration. In the squeezing process, the feature map in the input channel is compressed into a feature vector through global average pooling. In the excitation process, the weight of each channel is learned by a fully connected layer and a nonlinear activation function to capture the relationship between channels. In recalibration, the weight of each channel is multiplied by the feature on the original channel to calculate the weighted output feature. E is obtained by the squeezing excitation block to satisfy:

[0096] E'=E·sigmoid(W 1 ·relu(W·GAP(E))) (15)

[0097] Where GAP represents global average pooling, W and W 1 represents two layers of full connection, relu and sigmoid represent activation functions. Then, after two sets of convolution, batch normalization, relu activation function and maximum pooling, feature z is obtained. E :

[0098] z E =maxpool(relu(batchnorm(conv(maxpool(relu(batchnorm(conv(E')))))))(16)

[0099] 4. Multi-level feature fusion and parameter learning on the cloud side

[0100] The multi-level feature fusion process uses the continuous feature vector of the steel coil on the process control layer Discrete eigenvector and the next coil feature vector z at the manufacturing execution layer E In this paper, the discrete features are averaged and flattened, and then spliced ​​with the continuous features to obtain Z L , then Z L Perform time step averaging with z E Splice and get the input of the fully connected layer, denoted as Z S ,satisfy:

[0101]

[0102] Z S =[Z L ,z E ] (18)

[0103] Enter Z S After two fully connected layers, the model's coiling temperature prediction value is obtained. Assume that the output dimension of the first fully connected layer is λ 3 , coiling temperature prediction value The calculation is as follows:

[0104]

[0105] In the formula, relu is the relu activation function; and is the weight; ε 1 and ε 2 It's a deviation.

[0106] This paper uses back propagation to train the model parameters, and uses the Adaptive Moment Estimation (Adam) optimization algorithm to train the model. The mean absolute error is used as the loss function, and the calculation formula is:

[0107]

[0108] Based on the above, the coiling temperature prediction method based on multi-level feature fusion of this embodiment can be summarized as follows:

[0109] Fully extract multi-level features. Most existing methods only consider the information of the manufacturing execution layer when predicting the coiling temperature, while the method of this embodiment considers not only the information of the manufacturing execution, but also the information of the process control layer. Specifically, the method of this embodiment extracts the process control layer features of the current strip process data of the process control layer to extract the underlying working condition information of the current slab, and extracts the upper-level production scheduling setting information of the next slab by performing feature extraction on the steel coil setting data of the manufacturing execution layer. Afterwards, feature fusion is performed on the multi-level features extracted by the above feature extraction process, and the multi-level information is used for CT prediction, which more comprehensively considers the factors affecting the slab temperature, and is conducive to achieving higher-precision predictions.

[0110] Distributed deployment of cloud-edge-end. Considering that there is a certain time interval between adjacent slabs in the actual industrial process, the feature extraction of the current strip of the process control layer and the feature extraction of the next steel coil of the manufacturing execution layer can be deployed on the edge and cloud side respectively, making full use of the advantages of cloud-edge computing. On the one hand, due to the flexibility and real-time performance of edge computing, the feature extraction of process data with high real-time requirements is deployed on the edge to extract the underlying working condition information. On the other hand, due to the high efficiency of cloud-side computing, the coiling temperature that requires multi-level feature fusion is deployed on the cloud side. The multi-level structure of the method is conducive to distributed deployment, thereby ensuring the effectiveness and real-time performance of the prediction. In addition, deploying data preprocessing on the real-time control layer on the end side is conducive to reducing the burden and overhead of edge computing, and is conducive to speeding up the extraction of multi-level features.

[0111] Consider the multi-scale information of process data. The method of this embodiment extracts the information of continuous variables and discrete variables through GRU and SlowFast respectively, aiming at the situation where there are continuous quantities and discrete quantities in the actual process. The GRU combined with the attention mechanism in the method can capture long-term dependencies and focus on key information at the same time. GRU reduces the number of parameters while capturing long-term dependencies by streamlining the gating mechanism. The attention mechanism is conducive to focusing on key information at a specific time point or time period, thereby improving the accuracy of feature extraction. The SlowFast network in the method adopts a dual-path structure, which simultaneously captures the static information of the field time scale and the dynamic changes of the short time scale in the data through fast and slow paths. Multi-scale features are fused through lateral connections, and the fusion mechanism is conducive to more comprehensive capture of multi-scale information in discrete data.

[0112] Below, a practical application example is used to describe the implementation process of this method and verify its effect.

[0113] 1. Data description

[0114] This paper takes the actual process data of 50 steel coils from a steel plant as an example, including data of 6 types of steel for training and testing, of which 40 steel coils are used for training and 10 steel coils are used for testing. Before this, effective data interception, data alignment, and sliding time window operations need to be performed.

[0115] The data exported from each steel coil is time-series. The rolling of different slabs is not continuous in time. Therefore, it is necessary to intercept the effective rolling data and pre-process it. The data interception and alignment process is similar to Figure 5 As shown in the figure, the data of the slab in the laminar cooling process is intercepted according to the time interval when the two key signals "FDT_ON" and "CT_ON" are 1, and the data of different slabs are distinguished and stored according to the slab number variable "COIL". Secondly, the effective rolling data of different header sections of each steel coil are not aligned in the time dimension, so it is necessary to align the data according to the length position. Specifically, the data processing process includes: first, it is necessary to intercept the data of a piece of steel according to the finishing exit and coiling mark, that is, the earliest time and the latest time of the variable corresponding to a piece of steel are included in the time interval of the intercepted data. Second, according to the different characteristics of the header sections where different variables are located, the variables of each header section are grouped, and the position of the slab in the laminar cooling process is calculated by the variables "FDT Strip Transfer Length" and "CT Strip Transfer Length", and the effective data of different header sections are intercepted according to the actual physical distribution of the laminar cooling header. Finally, according to the total length of the slab and the length of different header sections, the data from different header sections are aligned according to the length of each piece of steel at an interval of 1 meter. Repeat the above operation until the data of all slabs within the time range are processed.

[0116] In order to extract multi-scale features, it is also necessary to perform time window interception on discrete variables. The step size of the time window is set to 1, the width of the time window is w, and the number of time windows is n. The time window is then downsampled to obtain multi-scale discrete data as the input of the SlowFast network.

[0117] In order to establish a data-driven CT prediction model, it is necessary to select the input variables of the model. Combining the knowledge of heat transfer principles related to the CT mechanism and the temperature drop model of laminar cooling, the formulas are shown in (21) to (22).

[0118] T(t)=T h +(T FDT -T h ) -λt (twenty one)

[0119]

[0120] Where, T(t) is the real-time surface temperature of the strip, °C; T FDT is the finishing exit temperature, °C; λ is the model factor; t is the strip cooling time, s; T e is the coiling temperature, °C; T W is the cooling water temperature, °C; F h is the finishing exit thickness, mm; It is based on the empirical parameters given by the on-site equipment.

[0121] According to the existing process variables of the actual production process, the input variables are divided into two parts: process control level variables and manufacturing execution level variables. The continuous input variables of the process control layer include the actual value of the finishing rolling exit thickness and the set value of the finishing rolling exit thickness, the actual value of the finishing rolling exit width and the deviation value of the finishing rolling exit width; the finishing rolling exit speed, the intermediate speed 1 and the intermediate speed 2; the finishing rolling exit temperature, the intermediate temperature of laminar cooling, the water temperature of the water tank, the temperature of the actuator of the rapid cooling section, the water temperature of each actuator above and below the rapid cooling sections 1, 2, and 18; the valve position, the pressure and flow increment of the actuator of the rapid cooling section. Among the discrete input variables of the process control layer, the switch status of the top and bottom manifolds of zones 1 to 20 are taken into consideration. The setting variables of the manufacturing execution level include the finishing rolling temperature setting, thickness setting, width setting, coiling temperature setting, and the actual coiling temperature of the previous steel coil. Therefore, the input of the final CT model includes 110 continuous variables, recorded as There are 98 discrete variables, denoted as Setting variables The input variables of the multi-level prediction model are shown in Table 1.

[0122] Table 1 Input variables of multi-level prediction model

[0123]

[0124]

[0125] 2. Model training and parameter determination

[0126] 2.1 Evaluation indicators

[0127] This paper uses Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) as indicators to evaluate the prediction performance of the model:

[0128]

[0129] In the formula, is the predicted value of coiling temperature, y i is the actual value of coiling temperature, and n is the number of predicted samples.

[0130] 2.2 Model parameter determination

[0131] The main parameters of the proposed method include the learning rate γ of the feature extraction part of the side process control layer 1 , the output dimension λ of the GRU layer 1 , the width of the time window w, the number of time windows n; the learning rate of the feature extraction part of the cloud-side manufacturing execution layer γ 2 , the output dimension λ of the fully connected layer of the multi-level feature fusion part 3 These parameters all have an impact on the performance of the model. The values ​​of the parameters are determined through experiments, as shown in Tables 2 to 6.

[0132] Table 2 Different γ 1 The prediction results

[0133]

[0134]

[0135] After many experiments, the learning rate γ of the feature extraction part of the side process control layer 1 Taking [0.01, 0.005, 0.001, 0.0005, 0.0001], the results of the coiling temperature prediction model are shown in Table 2. When γ 1 When 0.0005 is taken, the performance indicators are the smallest, the prediction effect is better, and γ is finally determined. 1 =0.0005. The output dimension of the GRU layer is λ 1Taken from [32, 48, 64, 80, 96, 112, 128], the performance indicators of the prediction results are shown in Table 3. When λ 1 When 96 is selected, the prediction performance is the best, and λ is finally determined. 1 =96.

[0136] Table 3 Different λ 1 The prediction results

[0137]

[0138] The width w and number n of the time window determine the input shape of the multi-scale feature extraction of the side process control layer, which has a great influence on the effect of feature extraction. Table 4 gives the model performance indicators under different w and n, and finally determines n = 16 and w = 16.

[0139] Table 4 Prediction results of different n and w

[0140]

[0141] Learning rate γ of the feature extraction part of the cloud-side manufacturing execution layer 2 Taking [0.01, 0.005, 0.001, 0.0005, 0.0001], the results of the coiling temperature prediction model are shown in Table 5. 2 When 0.0001 is taken, the performance indicators are the smallest, the prediction effect is better, and γ is finally determined. 2 =0.0001. Output dimension λ of the multi-level feature fusion fully connected layer 3 Taken from [32, 48, 64, 80, 96, 112, 128], the performance indicators of the prediction results are shown in Table 6, and finally λ is determined 3 =80.

[0142] Table 5 Different γ 2 The prediction results

[0143]

[0144] Table 6 Different λ 3 The prediction results

[0145]

[0146] 2.3 Experimental results and analysis

[0147] The above model setting parameters are used for training. In this paper, MAE is selected as the loss function for back propagation training of the model. The coiling temperature prediction results are as follows: Figure 6As shown in the figure, it can be seen that under the current model parameter settings, the predicted value of the coiling temperature based on the fusion of multi-level features has a good fit with the actual value of the coiling temperature, the deviation between the two is small, and the predicted value is consistent with the change trend of the actual value to a certain extent. At the same time, the prediction effect based on the fusion of multi-level features is better than that without the fusion of process control layer features, which shows that the fusion of multi-level information is conducive to improving the prediction accuracy of the coiling temperature, verifying the effectiveness of the method described in this paper.

[0148] The prediction results of the method based on multi-level feature fusion proposed in this paper are compared with the traditional method. The comparison is shown in Table 7. The comparison methods in the table do not fuse multi-level features. The prediction results of the method proposed in this paper have higher prediction accuracy than the prediction results without fusion of multi-level features, and have lower MAE, RMSE and MAPE. The MAE of the prediction model proposed in this paper is 2.6560 lower than that of the model of the comparative experiment. This reflects that feature extraction of steel coil setting data and fusion of process control layer features can more fully extract multi-level information of process control layer and manufacturing execution layer, and improve the prediction accuracy of CT.

[0149] Table 7 Analysis of coiling temperature comparison test results

[0150]

[0151] 3. Real-time applications of cloud-edge-device collaboration

[0152] 3.1 Description of the cloud-edge-device framework

[0153] The information flow of the real data of the hot rolling process in the cloud-edge-end framework is as follows: Figure 7 As shown. The cloud-side server has higher resource allocation than the edge-side server. The edge-side server completes some tasks with high real-time requirements, and the cloud-side server completes some tasks with high server memory and computing performance requirements. Preprocessing operations such as data alignment are performed on the end side, and then the real-time process data is uploaded to the real-time database on the edge side. The edge side obtains the real-time process data of the strip uploaded by the edge side through the real-time database for feature extraction, and uploads the extracted features to the cloud side. The cloud side obtains the manufacturing execution layer data from the relational database and receives the features uploaded by the edge side for coiling temperature prediction.

[0154] Based on the flexibility and real-time performance of edge computing and the efficiency of cloud computing, the coiling temperature prediction model used in this paper is deployed on Figure 7In the cloud-edge-end collaboration framework shown in the figure, multi-level feature extraction is performed on the prediction model on the cloud side and the edge side respectively. On the edge side, first obtain the data uploaded in real time after data preprocessing, then load the pre-trained model to extract the multi-scale features of the current steel coil, obtain the features containing the underlying operating status information and upload them to the cloud side. The cloud side obtains the historical steel coil setting data through the relational database history database, then performs feature extraction to obtain the features containing the top-level scheduling information, and then performs multi-level feature fusion with the features uploaded on the edge side, and finally obtains the predicted value of the coiling temperature of the next coil.

[0155] 3.2 Real-time prediction of coiling temperature

[0156] The coiling temperature prediction method based on multi-level feature fusion is deployed in a cloud-edge-end collaborative framework. The framework uses PLC to upload data on the end side, receives data from PLC through the end-side server and uploads it to the real-time database. The end-side server uploads the real-time data to the real-time database of the edge server through TCP communication. Based on the efficiency of cloud-side computing and the flexibility and real-time performance of edge-side computing, manufacturing execution layer feature extraction and multi-level feature fusion prediction are performed on the cloud side, and process control layer feature extraction is performed on the edge.

[0157] Table 8 lists the sampling interval and average time cost of the cloud-edge-end collaborative framework. On the cloud side, the average time from the end of the previous steel coil to the completion of the coiling temperature prediction of the next steel coil is 11.9s, which is less than the 90s time interval between adjacent steel coils, which reflects the feasibility of real-time coiling temperature prediction. On the edge side, the feature extraction time of a sample is 7.2ms, which is less than the sampling interval of 500ms, ensuring real-time prediction capability. On the end side, the average data preprocessing time is 58.4ms. Based on the above, the feasibility and real-time performance of the method proposed in this paper under the framework of cloud-edge-end collaboration have been verified.

[0158] Table 8 Sampling interval and average time length under the cloud-edge-end collaboration framework

[0159]

[0160] 4. Conclusion

[0161] In order to improve the prediction accuracy of the coiling temperature in the hot rolling process, this paper proposes a coiling temperature prediction method based on multi-level feature fusion. In the feature extraction part of the process control layer of the method, GRU and SlowFast Networks are used for multi-scale feature extraction to capture the long-term dependency of continuous variables and the multi-scale information of discrete variables, and obtain the operating conditions of the current steel coil at the bottom equipment end. Then in the feature extraction part of the manufacturing execution layer of the method. The scheduling information of the next steel coil at the top manufacturing execution layer is extracted through SEnet, and then multi-level feature fusion prediction is performed with the features extracted from the process control layer. Finally, a high-precision prediction result of the coiling temperature of the next steel coil is achieved.

[0162] This paper uses 50 steel coil data for training and testing. After preprocessing the data such as data alignment and sliding time window, the samples required for model input are obtained. Then, the grid search algorithm is used to determine the hyperparameters required for the model. The prediction accuracy of the method described in this paper is compared with that of prediction methods such as SVM, RF, XGB, GBRT, TCN, LSTM and GRU. The performance of the model prediction is revealed by calculating performance indicators such as MAE, RMSE and MAPE, and the prediction error results of each method are analyzed and compared. The experimental results show that the prediction model based on multi-level feature fusion proposed in this paper has good prediction performance. In addition, the proposed prediction model is deployed in the cloud-edge-end collaborative framework for application verification, and the manufacturing execution layer, process control layer and real-time control layer are deployed on the cloud side, edge side and end side respectively. The end side implements preprocessing such as process data alignment and sliding time window; the edge side extracts the features of the underlying working condition information; the cloud side extracts the top-level scheduling information, and also receives the process control layer features extracted by the edge process control layer, and then integrates the multi-system level features of the manufacturing execution layer and the process control layer, and finally realizes the real-time prediction of the coiling temperature. The effectiveness and real-time performance of the proposed method were verified using actual hot rolling process data.

[0163] Second embodiment

[0164] This embodiment provides an electronic device, such as Figure 8 As shown, the electronic device includes: a processor and a memory; wherein the processor and the memory can be connected via a communication bus; the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method of the first embodiment. In addition, the electronic device may also include a transceiver, the processor and the transceiver can be connected via a communication bus, and the transceiver is used to communicate with other devices.

[0165] Next, combine Figure 8 The following is a detailed introduction to the various components of the electronic device:

[0166] Among them, the processor is the control center of the electronic device, and the electronic device may include multiple processors, each of which may be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). The processor here may be a processor or a general term for multiple processing elements. For example, the processor is one or more central processing units (CPUs), or other general-purpose processors, application specific integrated circuits (ASICs), or one or more integrated circuits configured to implement an embodiment of the present invention, such as one or more microprocessors (digital signal processors, DSPs), or one or more field programmable gate arrays (field programmable gate arrays, FPGAs), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor may execute various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0167] In a specific implementation, as an embodiment, the processor may include one or more CPUs, such as Figure 8 The CPU0 and CPU1 shown in the figure are, of course, only exemplary.

[0168] The memory is used to store the software program for executing the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can refer to the above method embodiment and will not be repeated here.

[0169] Optionally, the memory may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may be integrated with the processor or exist independently and accessed through the interface circuit ( Figure 8 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0170] The transceiver may include a receiver and a transmitter ( Figure 8 The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function. The transceiver can be integrated with the processor or exist independently and communicate with the electronic device through the interface circuit ( Figure 8 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0171] In addition, it should be noted that Figure 8 The structure of the electronic device shown in the figure does not constitute a limitation on the device, and the actual device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. In addition, the technical effects achieved by the electronic device when executing the method of the first embodiment above can refer to the technical effects described in the first embodiment above, so they are not repeated here.

[0172] Third embodiment

[0173] This embodiment provides a computer-readable storage medium, which stores at least one instruction, and the instruction is loaded and executed by a processor to implement the method of the first embodiment. The computer-readable storage medium may be a ROM, a random access memory, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc. The instructions stored therein may be loaded by a processor in a terminal to execute the method.

[0174] In addition, it should be noted that the present invention can be provided as a method, an apparatus or a computer program product. Therefore, the embodiment of the present invention can be in the form of a full or partial hardware embodiment, a full or partial software embodiment or an embodiment combining software and hardware. Moreover, when implemented using software, the embodiment of the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program codes. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center containing one or more available media sets. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium. The semiconductor medium may be a solid state hard disk.

[0175] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0176] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable terminal device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0177] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements. In addition, the term "and / or" is only an association relationship describing the associated objects, indicating that there can be three relationships, for example, A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone, wherein A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding. "At least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can be represented by: a, b, c, ab, ac, bc or abc, where a, b, c can be single or plural.

[0178] In addition, it can be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0179] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0180] In several embodiments provided by the present invention, it should be understood that the disclosed equipment, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of functional modules / units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point, the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0181] If the method is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0182] Finally, it should be noted that the above is only a preferred embodiment of the present invention. It should be pointed out that although the preferred embodiment of the present invention has been described, for ordinary technicians in this technical field, once the basic creative concept of the present invention is known, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. Therefore, the attached claims are intended to be interpreted as including the preferred embodiment and all changes and modifications that fall within the scope of the embodiments of the present invention.

Claims

1. A coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration is used for coiling temperature prediction in hot rolling industrial process, where: The hot rolling industrial process is divided into a manufacturing execution layer, a process control layer and a real-time control layer from top to bottom; characterized in that the method comprises: The manufacturing execution layer is deployed on the cloud side, the process control layer is deployed on the edge side, and the real-time control layer is deployed on the device side. Collect the current strip process data of the process control layer and the steel coil setting data of the manufacturing execution layer, and pre-process the collected current strip process data and steel coil setting data on the terminal side; On the side, feature extraction is performed on the pre-processed current strip process data to obtain the underlying working condition information of the current slab; On the cloud side, feature extraction is performed on the pre-processed coil setting data to obtain the upper-level production setting information for the next slab; On the cloud side, the underlying working condition information of the current slab is fused with the upper-level production scheduling setting information of the next slab to obtain multi-level fusion features, and the coiling temperature prediction is achieved based on the multi-level fusion features.

2. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 1 is characterized in that: The current strip process data includes: the actual value of the finishing rolling exit thickness and the set value of the finishing rolling exit thickness, the actual value of the finishing rolling exit width and the finishing rolling exit width deviation value, the finishing rolling exit speed, the first intermediate speed, the second intermediate speed, the finishing rolling exit temperature, the laminar cooling intermediate temperature, the water temperature of the water tank, the actuator temperature of the rapid cooling section, the water temperature of the rapid cooling first end, the rapid cooling second section, and the upper and lower actuators of the rapid cooling 18th section, the valve position, the rapid cooling section actuator pressure, and the flow increment.

3. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 1 is characterized in that: The steel coil setting data includes: a finishing temperature setting value, a thickness setting value, a width setting value, a coiling temperature setting value, and an actual coiling temperature of the previous steel coil.

4. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 1 is characterized in that: Pre-process the collected current strip process data and coil setting data, including: Perform data interception and alignment operations on the current strip process data and coil setting data; The discrete variables in the current strip process data are subjected to time window truncation and downsampling processing to obtain multi-scale discrete data; wherein the discrete variables include the switch status of the manifold at the top and bottom of each zone.

5. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 4 is characterized in that: The data interception and alignment operations include: The collected data is used for slab tracking, that is, the data is intercepted and processed according to different slab numbers, and the processed data is stored separately by slab numbers. Specifically, the time interval of the same slab is intercepted according to the finishing exit and coiling mark positions, and then the position and time of the manifold passing through different sections are calculated according to the speed and time of the slab. Finally, the data in the length direction with an interval of 1 meter is aligned based on the total length of the slab.

6. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 4 is characterized in that: The time window interception and downsampling processing includes: Perform time window interception on discrete variables, where the step size of the time window is set to 1, the width of the time window is w, the number of time windows is n, and w and n are both preset values; After completing the time window interception, the time window is downsampled to obtain multi-scale discrete data.

7. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 4 is characterized in that: Feature extraction is performed on the pre-processed current strip process data, including: The GRU layer with fusion attention mechanism is used to extract features of the current strip process data that has completed the alignment operation, so as to extract the continuous feature vector of the steel coil on the process control layer; and SlowFast Networks is used to extract features of multi-scale discrete data, so as to extract the discrete feature vector of the steel coil on the process control layer.

8. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 7 is characterized in that: When using SlowFast Networks to extract features from multi-scale discrete data, the lateral connection results of the second residual network in SlowFastNetworks are used as the extracted features.

9. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as described in claim 7 is characterized in that: Feature extraction is performed on the pre-processed coil setting data, including: SENet is used to extract features from the setting data of the steel coil that has completed the alignment operation, and the feature vector of the next steel coil in the manufacturing execution layer is obtained.

10. The coiling temperature prediction method based on multi-level feature fusion under cloud-edge-end collaboration as claimed in claim 9 is characterized in that: The bottom-layer working condition information of the current slab is integrated with the upper-layer production setting information of the next slab to obtain multi-level integrated features, and the coiling temperature prediction is realized based on the multi-level integrated features, including: The discrete feature vector of a steel coil on the process control layer is averaged and flattened, and then spliced ​​with the continuous feature vector of a steel coil on the process control layer to obtain the feature vector of the steel coil on the process control layer; The feature vector of a steel coil on the process control layer is averaged over time steps and then concatenated with the feature vector of a steel coil on the manufacturing execution layer to obtain a multi-level fusion feature. The multi-level fusion features are passed through two fully connected layers to obtain the predicted value of the coiling temperature.

Images