A multi-objective hot rolling quality detection method based on progressive hierarchical extraction

Abstract

The present invention discloses a multi-objective hot rolling quality detection method based on progressive hierarchical extraction. The present invention includes the following steps: First, collect hot rolling production process data, determine multiple hot rolling quality detection parameters corresponding to the hot rolling production process data, and thus construct a hot rolling production data set; then perform data preprocessing on the hot rolling production data set to obtain a preprocessed data set; then, according to the preprocessed data set, train a multi-objective network based on progressive hierarchical extraction to obtain a multi-objective hot rolling quality detection model; finally, input the preprocessed hot rolling production data to be predicted into the multi-objective hot rolling quality detection model, output the multi-objective hot rolling quality detection result, and calculate the predicted values of multiple hot rolling quality detection parameters based on the multi-objective hot rolling quality detection result. The present invention realizes regularization by weighted fusion of the features of different quality detection parameters, and improves the accuracy and anti-interference ability of hot rolling quality detection.

Description

Technical Field

[0001] The present invention relates to a method for measuring product quality detection parameters in the fault prediction and diagnosis of the hot rolling production process, and specifically relates to a multi-objective hot rolling quality detection method based on progressive hierarchical extraction. Background Art

[0002] Hot rolling is widely used in the manufacture of general structural parts such as buildings, bridges, ships, and vehicles due to its low energy consumption, good plastic processing, low deformation resistance, insignificant work hardening, and easy rolling. However, since the hot rolling production process is affected by variables such as rolling force, rolling speed, roll gap value, bending roll force, and temperature, it has the characteristics of multi-variables, strong coupling, non-linearity, and time-variation. At the same time, since the quality of hot rolling products is determined by multiple parameters simultaneously, how to accurately evaluate various quality detection parameters in the hot rolling production process has become an important guarantee and challenge for ensuring product quality.

[0003] Currently, the evaluation and detection of the hot rolling production process can be based on two methods: mechanism models and data models respectively. For the mechanism model method, a large amount of prior knowledge is first required to establish the model. Secondly, these methods simplify and idealize the hot rolling production, and their robustness and practicability need to be tested. For the data-driven method, it does not require a large amount of prior knowledge. At the same time, in recent years, with the improvement of hardware and the growth of data volume, more and more researchers have turned their attention to the data-driven field, and the data-driven method has been widely applied in many scenarios and has high generalization and robustness.

[0004] However, the application of traditional data-driven methods in the industrial field has been greatly limited. Firstly, the data in the hot rolling production process has a large amount of noise. Secondly, due to the different demands of customers for products such as production specifications and quantities, the data in hot rolling production has a high degree of imbalance, and the performance of traditional data-driven methods on unbalanced data is often unsatisfactory. Thirdly, it is understood that the data-driven methods currently applied in the field of hot rolling production process fault prediction and diagnosis are all single-objective prediction models. Although they can achieve good prediction results on some parameters, their performance on other parameters often deteriorates. At the same time, the single-objective model also requires additional time for model training, and additional storage space and computing power for real-time prediction. Summary of the Invention

[0005] In view of this, in order to solve the problems in the background art, the object of the present invention is to propose a multi-objective hot rolling quality detection method based on progressive hierarchical extraction, enabling it to simultaneously predict multiple quality detection parameters on an unbalanced hot rolling production data set under noise interference and improving the prediction accuracy.

[0006] The technical solution of the present invention is as follows:

[0007] Step 1: Collect hot rolling production process data, determine multiple hot rolling quality detection parameters corresponding to the hot rolling production process data, and thus construct a hot rolling production data set;

[0008] Step 2: Perform data preprocessing on the hot rolling production data set to obtain a preprocessed data set;

[0009] Step 3: According to the preprocessed data set, train a multi-objective network based on progressive hierarchical extraction to obtain a multi-objective hot rolling quality detection model;

[0010] Step 4: Input the preprocessed hot rolling production data to be predicted into the multi-objective hot rolling quality detection model, output the multi-objective hot rolling quality detection result, and calculate the predicted values of multiple hot rolling quality detection parameters based on the multi-objective hot rolling quality detection result.

[0011] In the said Step 1, the hot rolling production process data includes rolling force, rolling speed, bending roll force, roll gap value, roll shift amount, rolling-in thickness, and inlet temperature.

[0012] In the said Step 1, the hot rolling quality detection parameters include crown, centerline offset, outlet temperature, wedge, width, and flatness.

[0013] The said Step 2 is specifically as follows:

[0014] First, eliminate the outliers in the hot rolling production data set according to the 3-sigma principle to obtain an eliminated data set; then, perform normalization processing on the eliminated data set to obtain a normalized data set; finally, use the sliding window method to process the data in the normalized data set to obtain a preprocessed data set.

[0015] The multi-objective network based on progressive hierarchical extraction is composed of a first-layer progressive hierarchical extraction network, a second-layer progressive hierarchical extraction network, and a multi-objective output network;

[0016] The first-layer progressive hierarchical extraction network is composed of k first private networks, 1 first shared network, and k + 1 first gating networks; the second-layer progressive hierarchical extraction network is composed of k second private networks, 1 second shared network, and k second gating networks; the multi-objective output network is composed of k output networks;

[0017] The inputs of the multi-objective network based on progressive hierarchical extraction are all used as the inputs of k first private networks and 1 first shared network. k first gating networks are respectively connected to the corresponding first private networks, and k first gating networks are also respectively connected to the first shared network. The remaining one first gating network is connected to both k first private networks and the first shared network. The inputs of the multi-objective network based on progressive hierarchical extraction are all connected to k + 1 first gating networks. The outputs of the k + 1 first gating networks are used as the outputs of the first-layer progressive hierarchical extraction network and are respectively connected to k second private networks and 1 second shared network of the second-layer progressive hierarchical extraction network;

[0018] The k second private networks are respectively connected to the corresponding second gating networks, the second shared network is connected to all k second gating networks, the outputs of the k first gating networks are also respectively connected to the corresponding second gating networks, and the k second gating networks are respectively connected to k output networks.

[0019] The structures of the k first private networks and 1 first shared network are the same, and they are all composed of a convolutional neural network, a gated recurrent network, and a long short-term memory network in parallel. Among them, the convolutional neural network is composed of 2 one-dimensional convolutional layers, 1 one-dimensional max pooling layer, and 1 fully connected layer connected in sequence. The gated recurrent network is composed of 1 gated recurrent unit and 1 fully connected layer connected in sequence. The long short-term memory network is composed of 1 long short-term memory unit and 1 fully connected layer connected in sequence;

[0020] The structures of the k second private networks and 1 second shared network are the same, and they are all composed of 3 fully connected layer networks in parallel.

[0021] The structures of the k + 1 first gating networks are the same, and they are all composed of 1 gated recurrent unit and 1 fully connected layer connected in sequence; the structures of the k second gating networks are the same, and they are all composed of 2 fully connected layers connected in sequence.

[0022] The single-objective loss function of the multi-objective network based on progressive hierarchical extraction during training is the mean square error, and the initial weight of each single-objective loss function is 1, and it is trained according to the dynamic weighted average rule.

[0023] In step 4, the preprocessed hot-rolling production data to be predicted is obtained after preprocessing the collected hot-rolling production data to be predicted in step 2.

[0024] In step 4, after anti-normalizing the multi-objective hot-rolling quality detection results, the predicted values of multiple hot-rolling quality detection parameters are obtained.

[0025] The beneficial effects of the present invention are:

[0026] 1. The present invention provides a method for predicting the quality inspection parameters of hot-rolled products with multiple objectives. Through a progressive hierarchical extraction method, the features between different quality inspection parameters can be fused to regularize the model and improve the model prediction accuracy.

[0027] 2. Compared with the single-objective data-driven model, the present invention can obtain relatively satisfactory prediction results with less training data, and can overcome the situation where there may be insufficient data of a single type of hot-rolled sample data in the hot-rolling production process.

[0028] 3. The present invention can simultaneously predict multiple quality inspection parameters, occupy less space and obtain more accurate prediction results with shorter prediction time. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is the flowchart of the method of the present invention.

[0030] Figure 2 is the network structure diagram of the present invention.

[0031] Figure 3 is the experimental result diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0032] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0033] As Figure 1 shown, the embodiments of the present invention are as follows:

[0034] Step 1: Collect the hot-rolling production process data, determine multiple hot-rolling quality inspection parameters corresponding to the hot-rolling production process data, and thus construct a hot-rolling production data set; the data set D can be expressed as: where x n represents the time series of the operating variables in the nth hot-rolling production process, and y k represents the time series of the kth quality inspection parameter, and k is the number of quality inspection parameters;

[0035] Among them, the hot-rolling production process data includes rolling force, rolling speed, bending roll force, roll gap value, roll shifting amount, rolling-in thickness, and inlet temperature. The hot-rolling quality inspection parameters include crown, centerline offset, outlet temperature, wedge, width, and flatness.

[0036] Step 2: Perform data preprocessing on the hot-rolling production data set to obtain the preprocessed data set;

[0037] Step 2 is specifically as follows:

[0038] First, outliers in the hot-rolling production dataset are removed according to the 3-sigma principle. Specifically, data satisfying [μ - 3×σ, μ + 3×σ] are selected, where μ is the mean of a single operating variable / parameter and σ is the standard deviation of a single operating variable / parameter, to obtain the dataset after outlier removal. Then, the dataset after outlier removal is normalized to obtain the normalized dataset. Finally, the data in the normalized dataset is processed using the sliding window method to obtain the preprocessed dataset, and the preprocessed dataset consists of multiple input matrices.

[0039] Among them, the normalization formula is:

[0040]

[0041] Among them, x is the original data, x max 、x min are the maximum and minimum values in the original data respectively, and x norm is the data after normalization processing;

[0042] Step 3: According to the preprocessed dataset, train the multi-objective network based on progressive hierarchical extraction to obtain the multi-objective hot-rolling quality detection model;

[0043] As Figure 2 shown, the multi-objective network based on progressive hierarchical extraction consists of the first-layer progressive hierarchical extraction network, the second-layer progressive hierarchical extraction network, and the multi-objective output network; the first-layer progressive hierarchical extraction network consists of k first private networks, 1 first shared network, and k + 1 first gating networks, where k is the number of hot-rolling quality detection parameters; the second-layer progressive hierarchical extraction network consists of k second private networks, 1 second shared network, and k second gating networks; the multi-objective output network consists of k output networks;

[0044] The inputs of the multi-objective network based on progressive hierarchical extraction are all used as the inputs of k first private networks and 1 first shared network. The k first gating networks are respectively connected to the corresponding first private networks, that is, one first gating network is connected to the corresponding one first private network one by one. The k first gating networks are also respectively connected to the first shared network. The remaining one first gating network (i.e., the (k + 1)th first gating network) is connected to both the k first private networks and the first shared network. The inputs of the multi-objective network based on progressive hierarchical extraction are all connected to the k + 1 first gating networks. The outputs of the k + 1 first gating networks are used as the outputs of the first-layer progressive hierarchical extraction network and are respectively connected to the k second private networks and 1 second shared network of the second-layer progressive hierarchical extraction network;

[0045] k second private networks are respectively connected to corresponding second gating networks, that is, one second private network is connected to one corresponding second gating network in a one-to-one manner. The second shared network is connected to k second gating networks. The outputs of the k first gating networks are also respectively connected to the corresponding second gating networks. The k second gating networks are respectively connected to k output networks. Each output network is composed of two fully connected layers connected in sequence. The output of the kth task is y k is: y k = t k (r k (x)).

[0046] The structures of the k first private networks and one first shared network are the same, and they are all composed of a convolutional neural network a gated recurrent network and a long short-term memory network in parallel. Parallel means that the input matrix is simultaneously input into three sub-networks respectively. Among them, the convolutional neural network is composed of two one-dimensional convolutional layers, one one-dimensional max-pooling layer and one fully connected layer connected in sequence. The gated recurrent network is composed of one gated recurrent unit and one fully connected layer connected in sequence. The long short-term memory network is composed of one long short-term memory unit and one fully connected layer connected in sequence. The structures of the k + 1 first gating networks g are the same, and they are all composed of one gated recurrent unit and one fully connected layer connected in sequence.

[0047] The structures of the k second private networks and one second shared network are the same, and they are all composed of three fully connected layer networks in parallel. The structures of the k second gating networks are the same, and they are all composed of two fully connected layers connected in sequence.

[0048] Among them, the process of the first-layer progressive hierarchical extraction network can be expressed as:

[0049]

[0050]

[0051]

[0052] Among them, r k is the output of the kth task in the first-layer progressive hierarchical extraction network.

[0053] In step 4, the input matrix is divided into a training set, a test set and a validation set, and the division ratio is 20% for the training set, 5% for the test set and 75% for the validation set. Construct the loss function of the multi-objective network based on progressive hierarchical extraction which is a single-objective loss function and the product of the corresponding weight λ k The sum can be expressed as:

[0054]

[0055] Among them, the single-objective loss function θ k is the mean squared error, and λ k (t) represents the weight of the k-th task during the t-th training, and θ s represents the parameters of the shared network, and y k represents the observed value of the k-th quality inspection parameter; represents the predicted value of the k-th quality inspection parameter; the initial weight of each single-objective loss function is 1, and it is trained according to the dynamic weighted average rule. The weight update rule is:

[0056]

[0057]

[0058] Among them, ω k (t) is the loss ratio of the previous round and the round before the previous round, and ω i (t) represents the loss ratio of the i-th task in the previous round and the round before the previous round. T is the smoothing coefficient, satisfying It calculates the loss reduction speed of each round, assigns higher weights to the tasks that are easy to decline, and thus realizes the minimization of the overall weight.

[0059] Step 5: Input the preprocessed hot-rolled production data to be predicted into the multi-objective hot-rolled quality inspection model, output the multi-objective hot-rolled quality inspection results, and calculate the predicted values of multiple hot-rolled quality inspection parameters based on the multi-objective hot-rolled quality inspection results. Specifically, after performing anti-normalization on the multi-objective hot-rolled quality inspection results (i.e., ), the predicted values of multiple hot-rolled quality inspection parameters are obtained.

[0060] Among them, the preprocessed hot-rolled production data to be predicted is obtained by preprocessing the collected hot-rolled production data to be predicted in Step 2.

[0061] In the simulation experiment, the data on the hot-rolled manufacturing site are collected, including 59 operating variables and 6 quality inspection parameters, including camber, centerline offset, exit temperature, wedge, width, and flatness. Each group has 60,000 data. Each group uses 12,000 data for training and 45,000 data for verification. The root mean square error (RMSE) and mean absolute error (MAE) are used as the prediction performance. Among them, the convolutional neural network (CNN), long short-term memory network (LSTM), gated recurrent unit (GRU), and deep neural network (DNN) are used as comparison algorithms. The prediction results are shown in Table 1.

[0062] Table 1 is a comparison table of the prediction results between the present invention and the existing methods

[0063]

[0064] As can be seen from the table, the PLE-MTL proposed by the present invention has achieved good prediction results in six quality inspection parameters. Figure 3 They are the frequency histograms of the prediction errors of six quality inspection parameters. Most of the errors are concentrated around 0, thus proving that the proposed method has a higher performance in predicting the product quality in the hot rolling production process.

[0065] The above embodiments are used to explain the present invention rather than limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A multi-objective hot rolling quality detection method based on progressive hierarchical extraction, characterized in that, It includes the following steps: Step 1: Collect the hot rolling production process data, determine multiple hot rolling quality detection parameters corresponding to the hot rolling production process data, and thus construct and obtain a hot rolling production data set; Step 2: Perform data preprocessing on the hot rolling production data set to obtain a preprocessed data set; Step 3: According to the preprocessed data set, train a multi-objective network based on progressive hierarchical extraction to obtain a multi-objective hot rolling quality detection model; Step 4: Input the preprocessed hot rolling production data to be predicted into the multi-objective hot rolling quality detection model, output and obtain the multi-objective hot rolling quality detection result, and calculate the predicted values of multiple hot rolling quality detection parameters based on the multi-objective hot rolling quality detection result; The multi-objective network based on progressive hierarchical extraction is composed of a first-layer progressive hierarchical extraction network, a second-layer progressive hierarchical extraction network, and a multi-objective output network; The first-layer progressive hierarchical extraction network is composed of k first private networks, 1 first shared network, and k + 1 first gating networks; the second-layer progressive hierarchical extraction network is composed of k second private networks, 1 second shared network, and k second gating networks; the multi-objective output network is composed of k output networks; The inputs of the multi-objective network based on progressive hierarchical extraction are all used as the inputs of k first private networks and 1 first shared network. The k first gating networks are respectively connected to the corresponding first private networks, and the k first gating networks are also respectively connected to the first shared network. The remaining one first gating network is connected to both the k first private networks and the first shared network. The inputs of the multi-objective network based on progressive hierarchical extraction are all connected to the k + 1 first gating networks. The outputs of the k + 1 first gating networks are used as the outputs of the first-layer progressive hierarchical extraction network and are respectively connected to the k second private networks and 1 second shared network of the second-layer progressive hierarchical extraction network; The k second private networks are respectively connected to the corresponding second gating networks, the second shared network is connected to all the k second gating networks, the outputs of the k first gating networks are also respectively connected to the corresponding second gating networks, and the k second gating networks are respectively connected to the k output networks; The structures of the k first private networks and 1 first shared network are the same, and they are all composed of a convolutional neural network, a gated recurrent unit network, and a long short-term memory network in parallel. Among them, the convolutional neural network is composed of 2 one-dimensional convolutional layers, 1 one-dimensional max pooling layer, and 1 fully connected layer connected in sequence. The gated recurrent unit network is composed of 1 gated recurrent unit layer and 1 fully connected layer connected in sequence. The long short-term memory network is composed of 1 long short-term memory unit layer and 1 fully connected layer connected in sequence; The structures of the k second private networks and 1 second shared network are the same, and they are all composed of 3 fully connected layer networks in parallel; The single-objective loss function of the multi-objective network based on progressive hierarchical extraction during training is the mean square error, and the initial weight of each single-objective loss function is 1, and it is trained according to the dynamic weighted average rule.

2. The multi-objective hot rolling quality detection method based on progressive hierarchical extraction according to claim 1, wherein In the above Step 1, the hot rolling production process data includes rolling force, rolling speed, bending roll force, roll gap value, roll shift amount, rolling-in thickness, and inlet temperature.

3. A multi-objective hot rolling quality detection method based on progressive hierarchical extraction according to claim 1, characterized in that, In the said step 1, the hot rolling quality detection parameters include convexity, center line offset, outlet temperature, wedge, width and flatness.

4. The multi-objective hot rolling quality detection method based on progressive hierarchical extraction according to claim 1, characterized in that, The said step 2 is specifically as follows: First, according to the 3-sigma principle, the outliers in the hot rolling production data set are removed to obtain the data set after removal; then, the data set after removal is normalized to obtain the normalized data set; finally, the data in the normalized data set is processed by the sliding window method to obtain the preprocessed data set.

5. The multi-objective hot rolling quality detection method based on progressive hierarchical extraction according to claim 1, characterized in that The structures of the said k + 1 first gating networks are the same, and each is composed of a 1-layer gated recurrent unit and a 1-layer fully connected layer connected in sequence; the structures of the said k second gating networks are the same, and each is composed of 2-layer fully connected layers connected in sequence.

6. The multi-objective hot rolling quality inspection method based on progressive hierarchical extraction according to claim 1, wherein, In the said step 4, the preprocessed hot rolling production data to be predicted is obtained by performing the preprocessing in step 2 on the collected hot rolling production data to be predicted.

7. A multi-objective hot rolling quality detection method based on progressive hierarchical extraction according to claim 1, characterized in that In the said step 4, after the multi-objective hot rolling quality detection results are denormalized, the predicted values of multiple hot rolling quality detection parameters are obtained.

Images