Metal plate strip shape recognition and control method and system based on model fusion, medium and processor

Abstract

The invention provides a metal plate strip shape recognition and control method based on model fusion. The method comprises the steps that plate shape data and time sequence data corresponding to the plate shape data are collected; carrying out statistics on plate shape defect conditions according to the plate shape data, and carrying out classification and label identification on defects; carrying out normalization processing and matrix processing on the plate shape data, and endowing a probability label value of a category to which the plate shape belongs, so as to form a first data set; constructing a CNN model, and dividing the first data set for training, verifying and testing the CNN model; marking corresponding plate shape state change time points and types for the time sequence data; segmenting the time sequence data into time windows with fixed lengths to form a second data set; constructing an LSTM model, and using the second data set to train, verify and test the LSTM model; and connecting the trained and tested CNN model and the trained and tested LSTM model in series so as to predict and control defects of the plate shape.

Description

Technical Field

[0001] The present application relates to the technical field of plate and strip rolling production, and in particular to a method, system, medium and processor for metal plate and strip shape recognition and control based on model fusion. Background Art

[0002] High-quality thin strips have high requirements for plate shape quality, so they need to be realized through the comprehensive coordinated control of multiple processes including rolling mills, continuous annealing units, leveling units and straightening units. The rolling mill presses the strip steel to a large extent and forms a strip steel with a certain plate shape. Then, annealing is performed to improve the strip steel structure performance and plate shape. The leveling unit is then used to further improve the flatness of the strip steel. Finally, the straightening and rewinding unit is used to improve the plate shape quality and pack it for delivery. Thin strips need to go through multiple processes from hot-rolled incoming materials to finished products. The plate shape is the result of the combined action of multiple processes, which makes it very difficult to control.

[0003] In the previous research on the whole process of cold rolling, Li Jun and others invented a full-process continuous production process and production line for stainless steel strip, using hot-rolled black steel coil as raw material, and sequentially carrying out continuous uncoiling, five-roll straightening, laser welding, continuous annealing of strip steel, descaling and rust removal, shot blasting and rust removal, mixed acid pickling and rust removal, cold rolling of 18-roll mill unit, high-pressure water brushing and flushing, continuous annealing of strip steel, electrolysis of neutral salt electrolyte, mixed acid pickling and passivation, as well as strip leveling, trimming, slitting and winding, so as to obtain the final product of stainless steel strip (patent number: CN202410027342.6). The above research and invention comprehensively consider the influence of multiple processes in the strip production process on the final finished product plate shape. However, there is a physical space distance between the detection unit and the rolling unit, the signal is delayed, the calculation is complex, and the prediction and compensation are difficult, especially the bad plate shape in the speed increase and speed decrease stages cannot be effectively identified and controlled, and the length of the head and tail defective parts is long.

[0004] In view of this, a method, system, medium and processor for metal strip shape recognition and control based on model fusion are needed. Summary of the invention

[0005] In view of the problems of signal lag, complex calculation, and difficulty in prediction and compensation in the prior art, this application provides a metal plate shape recognition and control method based on model fusion, which can accurately predict the shape defects. The characteristic parameter information involved in the control can produce high-quality and high-precision products, and effectively reduce the length of unqualified products when speeding up and speeding down. The specific technical solution is as follows:

[0006] A metal strip shape recognition and control method based on model fusion comprises the following steps:

[0007] Collect a large amount of plate shape data including different plate shapes and time series data corresponding to the plate shape data;

[0008] In the case of plate shape defects, the defects are classified and labeled; the plate shape data is normalized and matrixed, and a probability label value of the category to which the plate shape belongs is assigned, and finally a first data set composed of a matrix vector form is obtained;

[0009] Constructing a CNN model, dividing the first data set into a first training set, a first validation set, and a first test set, and using the first data set to train, validate, and test the CNN model;

[0010] The time series data are labeled with the corresponding time points and categories of the board shape state changes and preprocessed; the time series data are divided into time windows of fixed length, and each window is used as an input sample of the model to form the second data set;

[0011] Constructing an LSTM model, dividing the second data set into a second training set, a second validation set, and a second test set; and using the second data set to train, validate, and test the LSTM model;

[0012] The trained and tested CNN model and LSTM model are connected in series to form a complete fusion model, which can be used to predict and control the defects of the plate shape.

[0013] Furthermore, the plate shape data includes plate shape measurement data or plate strip image data.

[0014] Furthermore, the categories to which the plate shape belongs include normal plate shape and typical abnormal defects.

[0015] Furthermore, the process of normalizing and matrixing the plate shape data, assigning a probability label value to the category to which the plate shape belongs, and finally obtaining a first data set in the form of a matrix vector includes the following steps:

[0016] According to the number of flatness roller detection units, the maximum and minimum values ​​of all flatness data are counted respectively, and the integer rounded up to the maximum value and the integer rounded down to the minimum value are used as the extreme values ​​of the flatness range, and the extreme values ​​upward and downward are used as the flatness value range of each flatness roller detection unit;

[0017] Divide the flatness value variation interval of each flatness roller detection unit according to the same flatness accuracy to form a flatness grid of each flatness roller detection unit, and determine the normalized numerical representation of the flatness of each unit respectively based on this;

[0018] Generate a matrix for the i-th data record of the plate shape data, and fill the part without data in the matrix with 0;

[0019] According to the plate shape defect classification, the probability label value and the corresponding label name are assigned to the classification, and finally a first data set consisting of all plate shape data in the form of a matrix vector is obtained.

[0020] Furthermore, the preprocessing includes data cleaning, removing outliers and noise, and standardizing the data.

[0021] Furthermore, the method also includes the following steps: freezing and fine-tuning some parameters of the fusion model.

[0022] Furthermore, the following steps are included: the output of the LSTM model is returned to the control system of the cold rolling mill, and the cold rolling mill corrects the defects according to the corresponding plate shape defect control strategy model to achieve high-precision metal strip production output.

[0023] A metal strip shape recognition and control system based on model fusion, applied to the above-mentioned metal strip shape recognition and control method based on model fusion, comprises:

[0024] An acquisition module, which is used to collect a large amount of plate shape data including different plate shapes and time series data corresponding to the plate shape data;

[0025] The first processing module is used to count the board shape defects of the board shape data, classify the defects and identify labels; perform normalization and matrix processing on the board shape data, assign probability label values ​​to the categories to which the board shape belongs, and finally obtain a first data set in the form of matrix vectors;

[0026] A first construction module is used to construct a CNN model, divide the first data set into a first training set, a first validation set and a first test set, and is used to train, validate and test the CNN model;

[0027] The second processing module is used to mark the corresponding time points and categories of the plate shape state changes of the time series data and perform preprocessing; the time series data is divided into time windows of fixed length, each window is used as an input sample of the model, and constitutes the second data set;

[0028] A second construction module is used to construct an LSTM model, divide the second data set into a second training set, a second validation set, and a second test set; and is used to train, validate, and test the LSTM model;

[0029] The fusion module is used to connect the trained and tested CNN model and LSTM model in series to form a complete fusion model for predicting and controlling plate shape defects.

[0030] A computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned metal plate shape recognition and control method based on model fusion.

[0031] A processor is used to run a program, wherein the program executes the above-mentioned metal strip shape recognition and control method based on model fusion when running.

[0032] Compared with the prior art, the beneficial effects of this application are:

[0033] 1. This application uses a proprietary CNN training framework for learning and training to achieve plate shape anomaly recognition, accurately identify single-sided waves, two-rib waves, four-wave patterns, etc. at different positions, accurately identify and locate defects, broadband positions and confidence rates. According to the defect category recognition results of CNN, the LSTM network algorithm is used to identify periodic defects and quantitatively analyze the defect characteristic parameters: including defect category, defect type confidence, defect severity quantitative index, specific width position, defect occurrence time and duration, etc. According to the quantitative analysis results, a plate shape defect strategy model is established to participate in the adjustment and control of the tilting roll control, bending roll control, and regional cooling control to control the final plate shape. Moreover, according to the different widths of the incoming materials, different plate shape roller areas can be covered, and the unit plate shape vector scaling processing can be determined for the number of secondary plate shape roller detection units to meet the needs of generalization and versatility.

[0034] 2. This application is simple to deploy and quick to deploy, with no special configuration and environmental requirements. It can be used without changing the existing equipment configuration. The characteristic parameter information involved in the control can produce high-quality and high-precision products, and effectively reduce the length of unqualified products when speeding up and down, effectively improving the product yield. Even if it does not participate in the control and only uses the recognition mode, it can also be compared in real time and quickly, improve the actual level of operators, strengthen the action standards of operators, and improve product consistency. It is particularly suitable for various types of rolling mills widely used at home and abroad.

[0035] 3. The present application performs image recognition or signal analysis on plate shape measurement signals and video signals. Even in the presence of strong external interference signals, it is still possible to quickly and accurately identify plate shape defects, correctly calculate the location and credibility of the defects, and identify and display plate shape defects. Combined with a certain algorithm to predict and compensate for signal time lag, it is also possible to generate incremental signals involved in control and send them to the rolling mill control system, thereby achieving plate shape optimization control, or replacing and supplementing the original control system algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the prior art description. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0037] Figure 1It is a flow chart of a metal strip shape recognition and control method based on model fusion;

[0038] Figure 2 This is a schematic diagram for identifying the second rib wave defect;

[0039] Figure 3 This is a schematic diagram for identifying abnormal negative bending plate shape;

[0040] Figure 4 It is a structural schematic diagram of a metal strip shape recognition and control system based on model fusion. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0042] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0043] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include plural forms unless the context clearly indicates otherwise.

[0044] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0045] Embodiment 1

[0046] In this embodiment of the present application, the metal strip is an aluminum strip, which is used in the production of an aluminum strip cold rolling unit. The cold rolling unit is equipped with a plate shape roller after the last stand, and has 42 plate shape roller detection units. The cold rolling unit makes strategic control adjustments to the tilting roller, bending roller, etc. according to the results of each plate shape roller detection unit to improve the plate shape. The plate shape control capability of the cold rolling unit is divided into rolling force tilting roller control, bending roller control, and regional cooling control, so as to establish a corresponding plate shape defect control strategy model.

[0047] like Figures 1 to 3 The figure shows a flow chart of a metal strip shape recognition and control method based on model fusion, which includes the following steps:

[0048] S1: Collect a large amount of plate shape data including different plate shapes (normal plate shapes and various defective plate shapes) and time series data corresponding to the plate shape data. At the same time, collect time series data related to plate shape, such as measurement data of plate shape at different times during the production process, which can be obtained through sensors installed on production equipment.

[0049] Furthermore, each piece of plate shape data includes historical plate shape measurement data or plate shape image data based on the rolling of the plate and strip by the cold rolling mill.

[0050] Furthermore, the image data of each strip can be obtained from the camera on the production line to ensure clear image quality and accurately reflect the actual shape characteristics of the strip, while also including the corresponding material information obtained from the production system, such as coil number, alloy grade, width and thickness, coil weight, process data and other necessary information.

[0051] Furthermore, each piece of flatness measurement data includes the flatness measurement value of the aluminum strip measured by each flatness roller detection unit, and also includes the corresponding relevant material information, such as coil number, alloy grade, width thickness, coil weight, process data and other necessary information. The flatness measurement data can be derived from the iba data storage device at the production site of the cold rolling unit to obtain the flatness measurement data of the aluminum strip during continuous rolling, with a sampling interval of 100ms.

[0052] Furthermore, the process data includes rolling temperature, rolling force, rolling speed, etc.

[0053] Furthermore, the plate shape data includes geometric parameters of the aluminum strip, such as length, width, thickness distribution, etc.

[0054] S2: For the plate shape data, count the defects of each strip, select or use existing standards to mark each strip shape category, and classify and label the typical defects. The plate shape categories include normal plate shape and typical abnormal defects; the typical abnormal defects include edge waves, middle waves, scooped waves, two rib waves, abnormal negative bending and other defect types, so as to facilitate the subsequent establishment of a training data set of typical defects of 8 major types of cold-rolled plate shapes, that is, for the data at each sampling moment, select typical abnormal defects, and establish 8 major types of cold-rolled plate shape typical defects, so as to facilitate the subsequent formation of a training data set.

[0055] S3: For each piece of flatness data, data cleaning and normalization are performed within the flatness value range, and the pixel value of the plate image data or the flatness measurement data is normalized to a suitable range to calculate the flatness value at the corresponding grid position and distribution. For example, the flatness measurement data is normalized to the interval [-128, 128], and then it is matrixed and converted into a 640*480*3 matrix. Finally, the probability label value of the category to which the flatness belongs is assigned, and finally the first data set consisting of all flatness data in the form of matrix vectors is obtained.

[0056] Furthermore, the flatness value range refers to the range of extreme flatness values.

[0057] Further, step S3 includes the following detailed steps:

[0058] S31: according to the number of flatness roller detection units, the maximum and minimum values ​​of all flatness data are counted respectively, and the integer rounded up to the maximum value and the integer rounded down to the minimum value are used as the extreme values ​​of the flatness range, and the extreme values ​​upward and downward are used as the flatness value range of each flatness roller detection unit;

[0059] S32: dividing the flatness value variation intervals of each flatness roller detection unit according to the same flatness accuracy to form a flatness grid of each flatness roller detection unit, and determining the normalized numerical representation of the flatness of each unit respectively;

[0060] S33: Generate a matrix for the i-th data record of the plate shape data, and fill the part of the matrix without data with 0;

[0061] S34: According to the plate shape defect classification, a probability label value and a corresponding label name are assigned to the classification, and finally a first data set consisting of all plate shape data in the form of a matrix vector is obtained.

[0062] In this embodiment, the normalization interval of the unit plate shape measurement data is [-128, 128], the matrix vector is three-dimensional 640*480*3; the classification label code is 0-8.

[0063] In this embodiment, the unit plate image data can be normalized by dividing each pixel value by 255 (for 8-bit images), and the image can be cropped, rotated, scaled, etc. to ensure the consistency and diversity of the data.

[0064] S4: Divide the first data set into a first training set, a first validation set and a first test set. Generally, 70%-80% of the data is used as the first training set, 10%-15% of the data is used as the first validation set, and the remaining data is used as the first test set.

[0065] S5: Construct a CNN model, and use the plate shape data in the first training set, the first validation set, and the first test set to train, validate, and test the CNN model to obtain a plate shape recognition model for subsequent use in plate shape recognition.

[0066] In a specific implementation, the step S5 includes the following detailed steps:

[0067] S51: Establish a CNN model consisting of N convolutional layers and maximum pooling layers. The CNN model consists of several convolutional layers and maximum pooling layers to form the backbone network Backbone and Neck (feature enhancement network), which are used for feature extraction and feature fusion of matrix vectors, and then the detection layer network recognizes and classifies them.

[0068] S52: Take the historical plate shape data in the first training set as the input of the CNN model, take the class probability label value in the first training set as the target recognition expectation, train the CNN model offline, and obtain the feature vector value of the plate shape data and classify it through the Backbone (backbone network), Neck (feature enhancement network) and Head (detection head) of the CNN model. In this CNN model, the previous layer receives the matrix vector as input, and after grid division, a 3*3 kernel matrix is ​​used for scanning convolution. Each scanning convolution groups the information in a smaller window by applying a pooling layer to obtain the basic features of the input.

[0069] S53: Use the feature outputs of different convolutional layers to create a pyramid-like hierarchical representation to better examine the subtle differences in the detection areas of different plate-shaped rollers.

[0070] S54: Repeat the convolution and pooling process for multiple times, gradually extract important features, and output three features of different scales P3, P4, and P5. The data format is a four-dimensional vector with a shape of (N, C, H, W), where:

[0071] N is the batch size, which indicates the number of feature maps processed in one forward propagation.

[0072] C is the number of channels, which indicates the depth of the feature map.

[0073] H is the height, which indicates the height of the feature map.

[0074] W is the width, which indicates the width of the feature map.

[0075] S55: Use three detection networks to predict the location, category and confidence of the target from the feature matrix. The detection network consists of multiple convolutional layers (Conv) and normalization layers (BN), as well as activation functions (ReLU). The input is the output of the above step S52. The feature vector is converted into a prediction vector, the bounding box is decoded and generated, and the confidence is assigned. The non-maximum suppression technology is used to retain the most certain and accurate classification of typical defects of cold-rolled plate shape.

[0076] Further, repeat the above step S5, use each defect training set data for training, and generate a CNN model for predicting targets and categories. In this example, a weighted combination of three loss functions, namely, classification loss VFL (Varifocal Loss) and regression loss CIOU (Complete-IOU) + DFL (Deep Feature Loss), is used to form a target loss function to guide the optimization of the CNN model and adjust the hyperparameters to improve the convergence speed and performance of the CNN model.

[0077] Furthermore, if new data is added, it can be added to the first training set to repeat the training to update the CNN model. The CNN model trained in this step serves as the basis for subsequent prediction and recognition.

[0078] S56: Use the first validation set to evaluate the performance of the CNN model and observe indicators such as accuracy, recall, and F1-score. If the model is overfitting (shown as high accuracy on the first training set, but decreased accuracy on the first validation set), regularization methods such as L1 or L2 regularization, Dropout (randomly discarding some neurons in the fully connected layer), etc. can be used to optimize the model. According to the results of the first validation set, adjust the model's hyperparameters, such as network structure, learning rate, etc., until satisfactory performance is obtained.

[0079] The first validation set is another set of fully labeled data sets, which is used to evaluate the performance of the detection model and avoid overfitting. In each round of training, the first training set is first forward propagated and back-propagated, and then the first validation set is used for evaluation. For example, the overlap between the predicted bounding box and the true bounding box is measured, and the difference between the category distribution predicted by the model and the true label is measured.

[0080] S57: Input the first test set or new plate image data into the verified CNN model, the model outputs the plate shape category probability distribution of the plate, and selects the category with the highest probability as the recognition result, thereby realizing the plate shape recognition test.

[0081] After the above steps, a plate shape recognition model that has been trained, verified, tested and evaluated is generated for subsequent plate shape recognition; the plate shape defects of each rolling unit plate shape are determined based on the maximum probability value of each plate shape detection defect classifier, and multiple plate shape defects can be identified simultaneously.

[0082] S6: Label the corresponding time points and categories of the plate shape state changes of the time series data and perform preprocessing.

[0083] Furthermore, the preprocessing includes data cleaning, removing outliers and noise, and standardizing the data so that its mean is 0 and the standard deviation is 1.

[0084] S7: Divide the time series data into time windows of fixed length, each window is used as an input sample of the model, and each input sample is associated with a corresponding board shape category, together forming a second data set.

[0085] S8: Divide the second data set into a second training set, a second validation set, and a second test set.

[0086] S9: Construct an LSTM model, and use the time series data in the second training set, the second validation set and the second test set to train, validate and test the LSTM model, so as to facilitate subsequent use in plate shape prediction (mainly for time series data).

[0087] Furthermore, the construction of the LSTM model includes determining the number of LSTM layers, the activation function of the internal state, the cyclic activation function, the number of hidden units and other parameters. For example, 1-3 LSTM layers are set, and the number of hidden units in each LSTM layer can be selected according to the complexity of the data and the length of the sequence, such as 64-256 hidden units. The input layer receives the preprocessed time series data, and the output layer can be designed according to the prediction task, such as predicting the plate shape state at the next moment or predicting whether the plate shape will have defects in the future.

[0088] Furthermore, the time series data in the second training set is input into the LSTM model for training. For the loss function, a suitable function is selected according to the prediction task, such as the mean square error loss function for predicting numerical board shape parameters, and the cross entropy loss function for predicting board shape category changes. The model is also trained using an optimization algorithm (such as Adam) and a suitable learning rate, number of training rounds, and batch size. During the training process, the LSTM model adjusts the internal weight parameters through the back propagation through time algorithm (BPTT).

[0089] Furthermore, the performance of the LSTM model is evaluated using the time series data in the second validation set. The root mean square error (RMSE) can be used to measure the accuracy of the predicted value, or the accuracy rate and other indicators can be used to evaluate the accuracy of the prediction of the board shape state change. If overfitting occurs, regularization techniques are used, such as adding a Dropout layer or adjusting the weight decay coefficient of the LSTM. The hyperparameters of the model, such as the number of hidden units and the learning rate, are optimized according to the validation set results.

[0090] Furthermore, the second test set or new time series data is input into the trained LSTM model. The model predicts the future plate shape based on historical data, such as whether edge wave defects will appear in the plate at the next time point, or predicts the changing trend of the plate shape in the future, so as to provide early warning for quality control in the production process.

[0091] Furthermore, the trained model is used to make predictions for future time points to check its accuracy in predicting periodic changes.

[0092] S10: The trained and tested CNN model and LSTM model are connected in series to form a complete fusion model for predicting and controlling the periodic defects of the aluminum strip plate shape. That is, the plate shape data is first extracted by the CNN model for spatial features and category recognition, and then these spatial features and category information are input into the LSTM model as a time series for further processing, and finally the plate shape prediction output is obtained.

[0093] Furthermore, according to the structure of the fusion model, the loss function and optimization algorithm are determined, and the fusion model is trained at the same time. When applied, the corresponding data (image and time series data) are input into the fusion model to obtain more accurate plate shape recognition and prediction results.

[0094] Furthermore, in step S10, the trained and tested CNN model and LSTM model are connected in series to form a complete fusion model, including the following steps:

[0095] S101: Determine whether the CNN model output format matches the LSTM model input format.

[0096] In the tandem fusion, the output of the CNN model will be used as the input of the LSTM model. You need to carefully check the dimension, shape, and data type of the feature vector output by the last layer of the CNN model. For example, if the last layer of the CNN model is a fully connected layer, the output shape may be, where is the number of samples processed each time and is the feature dimension.

[0097] Make sure this output format matches the input format expected by the LSTM model. LSTM models usually expect input shape . If the CNN model outputs a feature vector for a single time step, you may need to adjust it appropriately. For example, convert it to a form by adding a dimension to make it meet the input requirements of the LSTM model.

[0098] Furthermore, data type conversion is required if necessary. Check the data type requirements of CNN and LSTM. For example, CNN may output data of floating point type (such as float32), while LSTM expects the input data type to be a specific tensor type (such as torch.Tensor type in PyTorch). If the data types do not match, conversion is required. In Python and deep learning frameworks (such as TensorFlow or PyTorch), corresponding functions can be used to convert data types. For example, in PyTorch, if the data output by the CNN model is a numpy array, it needs to be converted to the torch.Tensor type. The torch.from_numpy() function can be used for conversion.

[0099] S102: Model connection. Build connection logic in the code: In a deep learning framework (such as TensorFlow or PyTorch), define an integrated model class (if it is object-oriented programming) or function (if it is functional programming) to encapsulate the concatenated model (the fusion model in this example).

[0100] In this model definition, first instantiate the trained CNN and LSTM models. For example, in PyTorch, assuming that there are already two trained model objects, cnn_model and lstm_model, they can be introduced as member variables in the new model definition through self.cnn = cnn_model and self.lstm = lstm_model.

[0101] S103: Forward propagation connection. Define the forward propagation function of the fusion model. In this function, first pass the input data (plate shape data) into the CNN model for spatial feature extraction and category recognition. For example, in PyTorch, assuming that the input data is x, the output of CNN is obtained by cnn_output = self.cnn(x). Then use the output of the CNN model as the input of LSTM for time series processing. Continuing with the above example, pass cnn_output into the LSTM model, and obtain the output of the LSTM model through lstm_output = self.lstm(cnn_output).

[0102] S103: According to the requirements of the prediction task, the output of LSTM is appropriately processed. This application is a regression task, and the output of LSTM may be the final prediction result, that is, the periodic defects of the plate shape are identified.

[0103] The specific implementation also includes the following steps:

[0104] S11: The output (prediction result) of the LSTM model is returned to the control system of the cold rolling mill. The cold rolling mill uses the corresponding rolling force tilting roll control, bending roll control, and regional cooling control to correct the defects according to the corresponding plate shape defect control strategy model, so as to achieve high-precision aluminum strip production output. Combined with the LSTM model prediction and compensation signal time lag, the incremental signal involved in the control is generated and sent to the control system of the cold rolling mill to achieve plate shape optimization control or replace and supplement the original control system algorithm.

[0105] Furthermore, some parameters of the fusion model can be frozen and fine-tuned.

[0106] Parameter freezing decision: Consider whether to freeze some parameters of the CNN or LSTM model. If you think that the CNN has extracted good enough spatial features and do not want these spatial features to be excessively changed during the training process after the series fusion, you can freeze the parameters of the CNN. For example, in PyTorch, you can freeze the parameters by traversing the parameters of the CNN model and setting param.requires_grad = False. In this way, these parameters will not be updated in the subsequent training process.

[0107] Parameter fine-tuning: If you decide to fine-tune some model parameters, you can set the appropriate learning rate and optimizer. For example, use a smaller learning rate (such as the learning rate of the Adam optimizer is set to 0.0001) to fine-tune the parameters of the LSTM or the entire cascade model. During the fine-tuning process, the training data is also used for training, and the selected parameters are updated by calculating the loss function (such as the mean square error or cross entropy loss function, depending on the prediction task) and then back-propagating. Pay attention to the performance of the model on the validation set to prevent overfitting.

[0108] The beneficial effects of this application scheme are as follows:

[0109] This application uses a proprietary CNN training framework for learning and training to achieve board shape anomaly recognition and accurately identify single-sided waves and two-rib waves at different positions (such as Figure 2As shown), four-wave pattern, etc., accurately identify and locate defects, broadband position and confidence rate. According to the defect category recognition results of CNN, the LSTM network algorithm is used for periodic signals to identify periodic defects and quantitatively analyze the defect characteristic parameters: including defect category, defect type confidence, defect severity quantitative index, specific width position, defect occurrence time and duration, etc. According to the quantitative analysis results, a plate shape defect strategy model is established to participate in the adjustment and control of tilting roll control, bending roll control, and regional cooling control to control the final plate shape. Moreover, according to the different widths of incoming materials, different plate shape roller areas can be covered, and the number of secondary plate shape roller detection units can be determined to determine the unit plate shape vector scaling processing to meet the needs of generalization and versatility.

[0110] This application is simple to deploy and quick to deploy, and has no special configuration and environmental requirements. It can be used without changing the existing equipment configuration. The characteristic parameter information involved in the control can produce high-quality and high-precision products, and effectively reduce the length of unqualified products when speeding up and down, effectively improving the product yield. Even if it does not participate in the control and only uses the recognition mode, it can also be compared in real time and quickly, improve the actual combat level of operators, strengthen the action standards of operators, and improve product consistency. It is particularly suitable for various types of rolling mills widely used at home and abroad.

[0111] The present application performs image recognition or signal analysis on the plate shape measurement signal and the video signal. Even in the presence of strong external interference signals, the plate shape defects can still be quickly and accurately identified, and the location and credibility of the defects can be correctly calculated (such as Figure 2 The second rib wave defect recognition, the recognition confidence level is 0.99, Figure 3 Abnormal negative bend recognition) can identify and display plate shape defects; combined with a certain algorithm to predict and compensate for signal time lag, it can also generate incremental signals involved in control and send them to the rolling mill control system to achieve plate shape optimization control, or replace and supplement the original control system algorithm.

[0112] Therefore, the present application solution can solve the following deficiencies or unresolved problems in the prior art:

[0113] 1. Existing technical identification and control can meet the needs of ordinary products, but cannot meet the needs of high-precision products. The detection results are quite different from the actual plate shape. For example, the plate shape waves are obvious to the naked eye, but the system recognition is good.

[0114] 2. Existing control technologies lack anti-interference and robustness and are greatly affected by external interference.

[0115] 3. The existing control technology cannot effectively identify and control the bad plate shape during the speed increase and speed decrease stages, and the length of the bad parts at the head and tail is relatively long.

[0116] 4. The existing control technology cannot detect and identify the double-rib wave and quarter-wave plate shape defects, and cannot specifically locate the broadband position.

[0117] 5. Periodic defects in the length direction of the strip cannot be identified.

[0118] 6. There is a physical distance between the detection unit and the rolling unit, resulting in signal lag, making prediction and compensation difficult.

[0119] Embodiment 2

[0120] like Figure 4 The figure shows a schematic diagram of the structure of a metal strip shape recognition and control system based on model fusion, which is applied to the metal strip shape recognition and control method based on model fusion described above, and includes:

[0121] An acquisition module, which is used to collect a large amount of plate shape data including different plate shapes and time series data corresponding to the plate shape data;

[0122] The first processing module is used to count the board shape defects of the board shape data, classify the defects and identify labels; perform normalization and matrix processing on the board shape data, assign probability label values ​​to the categories to which the board shape belongs, and finally obtain a first data set in the form of matrix vectors;

[0123] A first construction module is used to construct a CNN model, divide the first data set into a first training set, a first validation set and a first test set, and is used to train, validate and test the CNN model;

[0124] The second processing module is used to mark the corresponding time points and categories of the plate shape state changes of the time series data and perform preprocessing; the time series data is divided into time windows of fixed length, each window is used as an input sample of the model, and constitutes the second data set;

[0125] A second construction module is used to construct an LSTM model, divide the second data set into a second training set, a second validation set, and a second test set; and is used to train, validate, and test the LSTM model;

[0126] The fusion module is used to connect the trained and tested CNN model and LSTM model in series to form a complete fusion model for predicting and controlling plate shape defects.

[0127] Embodiment 3

[0128] A computer-readable storage medium includes a stored program, wherein when the program is run, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned metal plate shape recognition and control method based on model fusion.

[0129] Embodiment 4

[0130] A processor is used to run a program, wherein the program executes the above-mentioned metal strip shape recognition and control method based on model fusion when running.

[0131] The present application provides a metal plate shape recognition and control method based on model fusion, comprising the following steps: collecting a large amount of plate shape data containing different plate shapes and time series data corresponding to the plate shape data; classifying and labeling the defects in the plate shape data; normalizing and matrixing the plate shape data, assigning probability label values ​​to the categories to which the plate shape belongs, and finally obtaining a first data set in the form of a matrix vector; constructing a CNN model, dividing the first data set into a first training set, a first verification set and a first test set, and using them to train, verify and test the CNN model; marking the corresponding plate shape state change time points and categories for the time series data and preprocessing them; dividing the time series data into time windows of fixed length, each window as an input sample of the model, to constitute a second data set; constructing an LSTM model, dividing the second data set into a second training set, a second verification set and a second test set; and using them to train, verify and test the LSTM model; connecting the trained and tested CNN model and the LSTM model in series to form a complete fusion model, so as to be used for predicting and controlling the defects of the plate shape.

[0132] Those of ordinary skill in the art will appreciate that the units of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. 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 this application.

[0133] In the embodiments provided in the present application, it should be understood that the division of units is merely a logical function division, and there may be other division methods in actual implementation, for example, multiple units may be combined into one unit, one unit may be split into multiple units, or some features may be ignored.

[0134] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0135] If the integrated unit 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 application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nlyMemory), random access memory (RAM, RandomAccessMemory), mobile hard disk, disk or CD-ROM and other media that can store program code.

[0136] 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 it. 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 replace some or all of the technical features therein by equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application, and they should all be included in the scope of the claims and specification of the present application.

Claims

1. A metal strip shape recognition and control method based on model fusion, characterized in that: The following steps are involved: Collect a large amount of plate shape data including different plate shapes and time series data corresponding to the plate shape data; In the case of plate shape defects, the defects are classified and labeled; the plate shape data is normalized and matrixed, and a probability label value of the category to which the plate shape belongs is assigned, and finally a first data set composed of a matrix vector form is obtained; Constructing a CNN model, dividing the first data set into a first training set, a first validation set, and a first test set, and using the first data set to train, validate, and test the CNN model; Label the time points and categories of the corresponding board shape state changes in the time series data and perform preprocessing; The time series data is divided into time windows of fixed length, and each window is used as an input sample of the model to form the second data set; Constructing an LSTM model, dividing the second data set into a second training set, a second validation set, and a second test set; and using the second data set to train, validate, and test the LSTM model; The trained and tested CNN model and LSTM model are connected in series to form a complete fusion model, which can be used to predict and control the defects of the plate shape.

2. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The flatness data includes flatness measurement data or strip image data.

3. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The categories to which the plate shape belongs include normal plate shape and typical abnormal defects.

4. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The process of normalizing and matrixing the plate shape data, assigning a probability label value to the category to which the plate shape belongs, and finally obtaining a first data set in the form of a matrix vector comprises the following steps: According to the number of flatness roller detection units, the maximum and minimum values ​​of all flatness data are counted respectively, and the integer rounded up to the maximum value and the integer rounded down to the minimum value are used as the extreme values ​​of the flatness range, and the extreme values ​​upward and downward are used as the flatness value range of each flatness roller detection unit; Divide the flatness value variation interval of each flatness roller detection unit according to the same flatness accuracy to form a flatness grid of each flatness roller detection unit, and determine the normalized numerical representation of the flatness of each unit respectively based on this; Generate a matrix for the i-th data record of the plate shape data, and fill the part without data in the matrix with 0; According to the plate shape defect classification, the probability label value and the corresponding label name are assigned to the classification, and finally a first data set consisting of all plate shape data in the form of a matrix vector is obtained.

5. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The preprocessing includes data cleaning, removing outliers and noise, and standardizing the data.

6. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The following steps are also included: Freeze and fine-tune some parameters of the fusion model.

7. The metal strip shape recognition and control method based on model fusion according to claim 1 is characterized in that: The following steps are also included: The output of the LSTM model is returned to the control system of the cold rolling mill. The cold rolling mill corrects the defects according to the corresponding plate shape defect control strategy model to achieve high-precision metal strip production output.

8. A metal strip shape recognition and control system based on model fusion, characterized in that: The metal strip shape recognition and control method based on model fusion applied to any one of claims 1 to 7 comprises: An acquisition module, which is used to collect a large amount of plate shape data including different plate shapes and time series data corresponding to the plate shape data; The first processing module is used to count the board shape defects of the board shape data, classify the defects and identify labels; perform normalization and matrix processing on the board shape data, assign probability label values ​​to the categories to which the board shape belongs, and finally obtain a first data set in the form of matrix vectors; A first construction module is used to construct a CNN model, divide the first data set into a first training set, a first validation set and a first test set, and is used to train, validate and test the CNN model; The second processing module is used to mark the corresponding time points and categories of the plate shape state changes of the time series data and perform preprocessing; the time series data is divided into time windows of fixed length, each window is used as an input sample of the model, and constitutes the second data set; A second construction module is used to construct an LSTM model, divide the second data set into a second training set, a second validation set, and a second test set; and is used to train, validate, and test the LSTM model; The fusion module is used to connect the trained and tested CNN model and LSTM model in series to form a complete fusion model for predicting and controlling plate shape defects.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the metal plate shape recognition and control method based on model fusion as described in any one of claims 1 to 7.

10. A processor, characterized in that: The processor is used to run a program, wherein the program, when running, executes the metal strip shape recognition and control method based on model fusion as described in any one of claims 1 to 7.

Images