WO2025015753A1 - 一种基于渐进式分层提取的多目标热轧质量检测方法 - Google Patents

一种基于渐进式分层提取的多目标热轧质量检测方法 Download PDF

Info

Publication number
WO2025015753A1
WO2025015753A1 PCT/CN2023/128312 CN2023128312W WO2025015753A1 WO 2025015753 A1 WO2025015753 A1 WO 2025015753A1 CN 2023128312 W CN2023128312 W CN 2023128312W WO 2025015753 A1 WO2025015753 A1 WO 2025015753A1
Authority
WO
WIPO (PCT)
Prior art keywords
hot rolling
network
objective
quality detection
networks
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/128312
Other languages
English (en)
French (fr)
Inventor
刘之涛
张树信
苏宏业
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University
Original Assignee
Zhejiang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University filed Critical Zhejiang University
Publication of WO2025015753A1 publication Critical patent/WO2025015753A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Definitions

  • the invention relates to a product quality detection parameter determination method for hot rolling production process fault prediction and diagnosis, and in particular to a multi-target hot rolling quality detection method based on progressive hierarchical extraction.
  • the evaluation and detection of hot rolling production process can be based on two methods: mechanism model and data model.
  • mechanism model method firstly, a lot of prior knowledge is needed to establish the model. Secondly, these methods simplify and idealize the hot rolling production. Their robustness and practicality need to be tested.
  • data-driven method it does not require a lot of prior knowledge.
  • the data-driven method has been widely used in many scenarios and has high generalization and robustness.
  • the application of traditional data-driven methods in the industrial field is greatly restricted.
  • First, the data of the hot rolling production process has a lot of noise;
  • third, it is understood that the data-driven methods currently used in the field of fault prediction and diagnosis in the hot rolling production process are all single-objective prediction models. Although they can achieve good prediction results on some parameters, their performance on other parameters often deteriorates.
  • the single-objective model also requires extra time for model training, and also requires additional storage space and computing power for real-time prediction.
  • the purpose of the present invention is to propose a multi-objective hot rolling quality detection method based on progressive hierarchical extraction, so that it can realize the simultaneous prediction of multiple quality detection parameters on an unbalanced hot rolling production data set under noise interference, and improve the prediction accuracy.
  • Step 1 Collect hot rolling production process data and determine the multiple hot rolling quality indicators corresponding to the hot rolling production process data. Measure the detection parameters to construct a hot rolling production data set;
  • Step 2 preprocess the hot rolling production data set to obtain a preprocessed data set
  • Step 3 According to the preprocessed data set, a multi-objective network based on progressive hierarchical extraction is trained 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 the multi-objective hot rolling quality detection results, and calculate the predicted values of multiple hot rolling quality detection parameters based on the multi-objective hot rolling quality detection results.
  • the hot rolling production process data includes rolling force, rolling speed, bending roll force, roll gap value, roll shifting amount, rolling thickness and inlet temperature.
  • the hot rolling quality inspection parameters include convexity, centerline offset, outlet temperature, wedge shape, width and straightness.
  • the step 2 is specifically as follows:
  • the outliers in the hot rolling production data set are eliminated according to the 3-sigma principle to obtain the eliminated data set; then, the eliminated data set is normalized to obtain the normalized data set; finally, the data in the normalized data set is processed using the sliding window method to obtain the preprocessed data set.
  • 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-target output network is composed of k output networks;
  • the inputs of the multi-objective network based on progressive hierarchical extraction are all used as 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, the k first gating networks are also respectively connected to the first shared network, the remaining 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 k+1 first gating networks, the outputs of the k+1 first gating networks are used as the outputs of the first layer of progressive hierarchical extraction network and are respectively connected to the k second private networks and 1 second shared network of the second layer of progressive hierarchical extraction network;
  • the k second private networks are respectively connected to the corresponding second gating networks, the second shared networks are all connected to 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 k first private networks and the one first shared network have the same structure, and are both composed of a convolutional neural network, a gated recurrent network, and a long short-term memory network in parallel, wherein the convolutional neural network consists of two layers of The gated recurrent network consists of a layer of gated recurrent units and a layer of fully connected layers connected in sequence, and the long short-term memory network consists of a layer of long short-term memory units and a layer of fully connected layers connected in sequence;
  • the k second private networks and the one second shared network have the same structure, and are both composed of three fully connected layer networks in parallel.
  • the k+1 first gating networks have the same structure, and are both composed of 1 layer of gated recurrent unit and 1 layer of fully connected layer connected in sequence; the k second gating networks have the same structure, and are both composed of 2 layers of fully connected layers connected in sequence.
  • the single-objective loss function of the multi-objective network based on progressive hierarchical extraction during training is the mean square error, the initial weight of each single-objective loss function is 1, and training is performed according to the dynamic weighted average rule.
  • the pre-processed hot rolling production data to be predicted is obtained by performing the pre-processing in step 2 on the collected hot rolling production data to be predicted.
  • step 4 after the multi-objective hot rolling quality detection results are denormalized, predicted values of multiple hot rolling quality detection parameters are obtained.
  • the present invention proposes a multi-objective prediction method for hot-rolled product quality inspection parameters. Through a progressive hierarchical extraction method, the features between different quality inspection parameters can be integrated to perform model regularization and improve the model prediction accuracy.
  • the present invention can obtain relatively satisfactory prediction results using less training data, and can overcome the situation of insufficient sample data for a single type of hot rolling that may exist in the hot rolling production process.
  • the present invention can predict multiple quality detection parameters at the same time, occupying less space and using shorter prediction time to obtain more accurate prediction results.
  • FIG1 is a flow chart of the method of the present invention.
  • FIG. 2 is a network structure diagram of the present invention.
  • the embodiments of the present invention are as follows:
  • Step 1 Collect 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;
  • Data set It can be expressed as: Where x n represents the time series of the operating variables in the nth hot rolling production process, y k represents the time series of the kth quality inspection parameter, and k is the number of quality inspection parameters;
  • the hot rolling production process data includes rolling force, rolling speed, bending roll force, roll gap value, roll shifting amount, rolling thickness and inlet temperature.
  • the hot rolling quality inspection parameters include crown, centerline offset, outlet temperature, wedge, width and straightness.
  • Step 2 preprocess the hot rolling production data set to obtain a preprocessed data set
  • Step 2 is as follows:
  • outliers in the hot rolling production data set are eliminated according to the 3-sigma principle. Specifically, data satisfying [ ⁇ -3 ⁇ , ⁇ +3 ⁇ ] are selected, where ⁇ is the average value of a single operating variable/parameter and ⁇ is the standard deviation of a single operating variable/parameter, to obtain the eliminated data set; then, the eliminated data set is normalized to obtain a normalized data set; finally, the data in the normalized data set is processed using the sliding window method to obtain a preprocessed data set, which consists of multiple input matrices.
  • x is the original data
  • x max and x min are the maximum and minimum values in the original data respectively
  • x norm is the data after normalization
  • Step 3 According to the preprocessed data set, a multi-objective network based on progressive hierarchical extraction is trained to obtain a multi-objective hot rolling quality detection model;
  • the multi-objective network based on progressive hierarchical extraction consists 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 consists of k first private networks, 1 first shared network and k+1 first gated 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 gated networks;
  • the multi-objective output network consists of k output networks;
  • the input of the multi-target network based on progressive hierarchical extraction is used as the input 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 first private network one by one, the k first gating networks are also respectively connected to the first shared network, the remaining first gating network (that is, the k+1th first gating network) is connected to the k first private networks and the first shared network, the input of the multi-target network based on progressive hierarchical extraction is connected to the k+1 first gating networks, and the output of the k+1 first gating network is used as the first layer of progressive
  • the output of the hierarchical extraction network is respectively connected to k second private networks and 1 second shared network of the second progressive hierarchical extraction network;
  • the k second private networks are respectively connected to the corresponding second gating networks, that is, one second private network is connected one by one to a corresponding second gating network, and the second shared networks are all connected to the 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, and each output network is composed of 2 layers of fully connected layers connected in sequence.
  • the structures of the k first private networks and the 1 first shared network are the same, both of which are composed of convolutional neural networks.
  • Gated Recurrent Networks and long short-term memory networks Parallel construction, parallel means that the input matrix is input into the three sub-networks at the same time, among which the convolutional neural network is composed of two layers of one-dimensional convolutional layers, one layer of one-dimensional maximum pooling layer and one layer of fully connected layer connected in sequence, the gated recurrent network is composed of one layer of gated recurrent unit and one layer of fully connected layer connected in sequence, and the long short-term memory network is composed of one layer of long short-term memory unit and one layer of fully connected layer connected in sequence.
  • the structures of the k+1 first gated networks g are the same, and they are all composed of one layer of gated recurrent unit and one layer of fully connected layer connected in sequence.
  • the k second private networks and the 1 second shared network have the same structure, both consisting of 3 fully connected layer networks in parallel.
  • the k second gated networks have the same structure, both consisting of 2 fully connected layers connected in sequence.
  • the process of the first layer of progressive hierarchical extraction network can be expressed as:
  • r k is the output of the k-th task in the first layer of the progressive hierarchical extraction network.
  • step 4 the input matrix is divided into training set, test set and validation set, with the ratio of 20% training set, 5% test set and 75% validation set.
  • the sum of the products of and the corresponding weight ⁇ k can be expressed as:
  • the single-objective loss function ⁇ k is the mean square error
  • ⁇ k (t) represents the weight of the kth task during the tth training
  • ⁇ s represents the parameters of the shared network
  • yk represents the observed value of the kth quality detection parameter
  • surface shows the predicted value of the kth quality detection parameter
  • the initial weight of each single-objective loss function is 1, and it is trained according to the dynamic weighted average rule, and the weight update rule is:
  • ⁇ k (t) is the loss ratio of the previous round and the previous round
  • ⁇ i (t) represents the loss ratio of the i-th task in the previous round and the previous round
  • T is the smoothing coefficient, and satisfies It calculates the rate at which the loss decreases in each round and assigns higher weights to tasks that are prone to decrease, thereby minimizing the overall weight.
  • Step 5 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 results, and calculate the predicted values of multiple hot rolling quality detection parameters based on the multi-objective hot rolling quality detection results.
  • the multi-objective hot rolling quality detection results are denormalized (i.e. ), the predicted values of multiple hot rolling quality detection parameters are obtained.
  • 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.
  • hot rolling manufacturing site data were collected, including 59 operating variables and 6 quality inspection parameters, including crown, centerline offset, outlet temperature, wedge, width and straightness. Each group had 60,000 data, 12,000 data were used for training and 45,000 data were used for verification.
  • the root mean square error (RMSE) and mean absolute error (MAE) were used as the prediction performance.
  • Convolutional neural network (CNN), long short-term memory network (LSTM), gated recurrent network (GRU) and deep neural network (DNN) were used as comparison algorithms. The prediction results are shown in Table 1.
  • Table 1 is a comparison table of prediction results between the present invention and the existing method
  • FIG3 is a frequency histogram of the prediction errors of the six quality detection parameters. Most of the errors are concentrated around 0, which proves that the proposed method has a higher performance in predicting product quality in the hot rolling production process.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Control Of Metal Rolling (AREA)

Abstract

本发明公开了一种基于渐进式分层提取的多目标热轧质量检测方法。本发明包括以下步骤：首先，收集热轧生产过程数据，确定热轧生产过程数据对应的多个热轧质量检测参数，从而构建获得热轧生产数据集；接着对热轧生产数据集进行数据预处理，获得预处理后的数据集；再根据预处理后的数据集，对基于渐进式分层提取的多目标网络进行训练，获得多目标热轧质量检测模型；最后将待预测的预处理后的热轧生产数据输入到多目标热轧质量检测模型中，输出获得多目标热轧质量检测结果，基于多目标热轧质量检测结果计算获得多个热轧质量检测参数的预测值。本发明通过加权融合不同质量检测参数的特征，实现了正则化，提高了热轧质量检测的精确度和抗干扰性。

Description

一种基于渐进式分层提取的多目标热轧质量检测方法 技术领域
本发明涉及热轧生产过程故障预测与诊断的一种产品质量检测参数测定方法，具体涉及了一种基于渐进式分层提取的多目标热轧质量检测方法。
背景技术
热轧由于能耗低，塑性加工良好，变形抗力低，加工硬化不明显，易进行轧制等特点被广泛应用于建筑、桥梁、船舶、车辆等一般结构件的制造，但是，由于热轧生产过程受到轧制力、轧制速度、辊缝值、弯辊力、温度等变量的影响，其具备多变量、强耦合、非线性和时变等特点，同时由于热轧产品的质量由多个参数同时决定，因此如何准确评估热轧在生产过程中的各项质量检测参数成为确保产品质量的重要保障和挑战。
目前对于热轧生产流程的评估和检测可以分别基于机理模型和数据模型两种方法，对于机理模型的方法，首先需要大量的先验知识来建立模型，其次这些方法都是将热轧生产化成简化和理想化，其鲁棒性和实用性有待检验。而对于数据驱动的方法，其并不需要大量的先验知识，同时，近几年来随着硬件的改善和数据量的增长，越来越多的研究者将目光投向数据驱动领域，而基于数据驱动的方法在很多场景都有了大量的应用并具有较高的泛化性和鲁棒性。
但是传统数据驱动方法在工业领域的应用受到了很大的限制，首先热轧生产过程的数据具备大量的噪声；其次，由于客户对于产品的需求比如生产规格和数量各不相同，热轧生产的数据具备很高的不平衡性，而传统的数据驱动方法在不平衡数据上的表现往往不尽人意；再次，据了解，目前应用在热轧生产过程故障预测和诊断领域的数据驱动方法都是单目标预测模型，虽然在部分参数上其能取得不错的预测结果，但是在其他参数上其表现能力则往往会变差，同时，单目标模型还需要耗费额外的时间进行模型训练，同时也需要额外的存储空间和算力进行实时预测。
发明内容
有鉴于此，为了解决背景技术中存在的问题，本发明的目的是提出一种基于渐进式分层提取的多目标热轧质量检测方法，让其能够在噪声干扰下的不平衡热轧生产数据集上实现多质量检测参数的同时预测，并且提高预测的精确度。
本发明的技术方案如下：
步骤1：收集热轧生产过程数据，确定热轧生产过程数据对应的多个热轧质 量检测参数，从而构建获得热轧生产数据集；
步骤2：对热轧生产数据集进行数据预处理，获得预处理后的数据集；
步骤3：根据预处理后的数据集，对基于渐进式分层提取的多目标网络进行训练，获得多目标热轧质量检测模型；
步骤4：将待预测的预处理后的热轧生产数据输入到多目标热轧质量检测模型中，输出获得多目标热轧质量检测结果，基于多目标热轧质量检测结果计算获得多个热轧质量检测参数的预测值。
所述步骤1中，热轧生产过程数据包括轧制力、轧制速度、弯辊力、辊缝值、窜辊量、轧入厚度和入口温度。
所述步骤1中，热轧质量检测参数包括凸度、中心线偏移量、出口温度、楔形、宽度和平直度。
所述步骤2具体为：
首先，根据3-sigma原则剔除热轧生产数据集中的离群值，获得剔除后的数据集；接着，对剔除后的数据集进行归一化处理，获得归一化后的数据集，最后，利用滑动窗口方法对归一化后的数据集中的数据进行处理后，获得预处理后的数据集。
所述基于渐进式分层提取的多目标网络由第一层渐进式分层提取网络、第二层渐进式分层提取网络和多目标输出网络组成；
第一层渐进式分层提取网络由k个第一私有网络、1个第一共享网络和k+1个第一门控网络构成；第二层渐进式分层提取网络由k个第二私有网络、1个第二共享网络和k个第二门控网络构成；多目标输出网络由k个输出网络组成；
基于渐进式分层提取的多目标网络的输入均作为k个第一私有网络、1个第一共享网络的输入，k个第一门控网络分别与对应的第一私有网络相连，k个第一门控网络还分别与第一共享网络相连，剩余的一个第一门控网络与k个第一私有网络和第一共享网络均相连，基于渐进式分层提取的多目标网络的输入均与k+1个第一门控网络相连，k+1个第一门控网络的输出作为第一层渐进式分层提取网络的输出并分别与第二层渐进式分层提取网络的k个第二私有网络、1个第二共享网络相连；
k个第二私有网络分别与对应的第二门控网络相连，第二共享网络均与k个第二门控网络相连，k个第一门控网络的输出还分别与对应的第二门控网络相连，k个第二门控网络分别与k个输出网络相连。
所述k个第一私有网络和1个第一共享网络的结构相同，均由卷积神经网络、门控循环网络和长短期记忆网络并行构成，其中，卷积神经网络由2层一 维卷积层、1层一维最大池化层和1层全连接层依次相连构成，门控循环网络由1层门控循环单元和1层全连接层依次相连构成，长短期记忆网络由1层长短期记忆单元和1层全连接层依次相连构成；
所述k个第二私有网络和1个第二共享网络的结构相同，均由3个全连接层网络并行构成。
所述k+1个第一门控网络的结构相同，均由1层门控循环单元和1层全连接层依次相连构成；所述k个第二门控网络的结构相同，均由2层全连接层依次相连构成。
所述基于渐进式分层提取的多目标网络在训练时的单目标损失函数为均方误差，每个单目标损失函数的初始权重为1，根据动态加权平均法则进行训练。
所述步骤4中，待预测的预处理后的热轧生产数据是对采集的待预测热轧生产数据进行步骤2中的预处理后获得的。
所述步骤4中，对多目标热轧质量检测结果进行反归一化后，获得多个热轧质量检测参数的预测值。
本发明的有益效果为：
1.本发明提出了一种多目标的热轧产品质量检测参数的预测方法，通过渐进式分层提取的方法，可以融合不同质量检测参数之间的特征，进行模型正则化，提高模型预测精确度。
2.本发明与单目标数据驱动模型相比，本发明使用更少的训练数据就可以得到相对满意的预测结果，能够克服热轧生产过程中可能存在的对于单一类型的热轧样本数据不充足的情况。
3.本发明能够同时预测多个质量检测参数，占用更少的空间并用更短的预测时间就能得到更精确的预测结果。
附图说明
图1为本发明的方法流程图。
图2为本发明的网络结构图。
图3为本发明实验结果图。
具体实施方式
下面结合附图以及具体实施例对本发明作进一步的说明，但本发明的保护范围并不限于此。
如图1所示，本发明的实施例如下：
步骤1：收集热轧生产过程数据，确定热轧生产过程数据对应的多个热轧质量检测参数，从而构建获得热轧生产数据集；数据集可以表示为： 其中，xn表示第n个热轧生产过程中操作变量的时间序列，yk表示第k个质量检测参数的时间序列，k为质量检测参数的数量；
其中，热轧生产过程数据包括轧制力、轧制速度、弯辊力、辊缝值、窜辊量、轧入厚度和入口温度。热轧质量检测参数包括凸度、中心线偏移量、出口温度、楔形、宽度和平直度。
步骤2：对热轧生产数据集进行数据预处理，获得预处理后的数据集；
步骤2具体为：
首先，根据3-sigma原则剔除热轧生产数据集中的离群值，具体是挑选满足[μ-3×σ,μ+3×σ]的数据，其中μ为单个操作变量/参数的平均值，σ为单个操作变量/参数的标准差，获得剔除后的数据集；接着，对剔除后的数据集进行归一化处理，获得归一化后的数据集，最后，利用滑动窗口方法对归一化后的数据集中的数据进行处理后，获得预处理后的数据集，预处理后的数据集由多个输入矩阵组成。
其中，归一化公式为：
其中，x为原始数据，xmax、xmin分别为原始数据中的最大值和最小值，xnorm为归一化处理后的数据；
步骤3：根据预处理后的数据集，对基于渐进式分层提取的多目标网络进行训练，获得多目标热轧质量检测模型；
如图2所示，基于渐进式分层提取的多目标网络由第一层渐进式分层提取网络、第二层渐进式分层提取网络和多目标输出网络组成；第一层渐进式分层提取网络由k个第一私有网络、1个第一共享网络和k+1个第一门控网络构成，k为热轧质量检测参数的数量；第二层渐进式分层提取网络由k个第二私有网络、1个第二共享网络和k个第二门控网络构成；多目标输出网络由k个输出网络组成；
基于渐进式分层提取的多目标网络的输入均作为k个第一私有网络、1个第一共享网络的输入，k个第一门控网络分别与对应的第一私有网络相连，即一个第一门控网络与对应的一个第一私有网络一一相连，k个第一门控网络还分别与第一共享网络相连，剩余的一个第一门控网络(即第k+1个第一门控网络)与k个第一私有网络和第一共享网络均相连，基于渐进式分层提取的多目标网络的输入均与k+1个第一门控网络相连，k+1个第一门控网络的输出作为第一层渐进式 分层提取网络的输出并分别与第二层渐进式分层提取网络的k个第二私有网络、1个第二共享网络相连；
k个第二私有网络分别与对应的第二门控网络相连，即一个第二私有网络与对应的一个第二门控网络一一相连，第二共享网络均与k个第二门控网络相连，k个第一门控网络的输出还分别与对应的第二门控网络相连，k个第二门控网络分别与k个输出网络相连，每个输出网络由2层全连接层依次相连构成，第k个任务输出yk为：yk＝tk(rk(x))。
k个第一私有网络和1个第一共享网络的结构相同，均由卷积神经网络门控循环网络和长短期记忆网络并行构成，并行指输入矩阵同时分别输入3个子网络，其中，卷积神经网络由2层一维卷积层、1层一维最大池化层和1层全连接层依次相连构成，门控循环网络由1层门控循环单元和1层全连接层依次相连构成，长短期记忆网络由1层长短期记忆单元和1层全连接层依次相连构成。k+1个第一门控网络g的结构相同，均由1层门控循环单元和1层全连接层依次相连构成。
k个第二私有网络和1个第二共享网络的结构相同，均由3个全连接层网络并行构成。k个第二门控网络的结构相同，均由2层全连接层依次相连构成。
其中，第一层渐进式分层提取网络的过程可以表示为：


其中，rk为第k个任务在第一层渐进式分层提取网络的输出。
步骤4中，将输入矩阵划分成训练集，测试集和验证集，划分比例为20％训练集，5％测试集和75％验证集。构建基于渐进式分层提取的多目标网络的损失函数其为单目标损失函数与对应权重λk之积的总和，可以表示为：
其中，单目标损失函数θk为均方误差，λk(t)表示第t次训练时，第k个任务的权重，θs表示共享网络的参数，yk表示第k个质量检测参数的观测值；表 示第k个质量检测参数的预测值；每一个单目标损失函数的初始权重为1，根据动态加权平均法则进行训练，权重更新法则为：

其中，ωk(t)为上一轮和上上轮的损失比，ωi(t)表示第i个任务在上一轮和上上轮的损失比，T为平滑系数，满足其通过计算每一轮损失下降速度，对容易下降的任务赋予更高的权重，从而实现整体权重的最小化。
步骤5：将待预测的预处理后的热轧生产数据输入到多目标热轧质量检测模型中，输出获得多目标热轧质量检测结果，基于多目标热轧质量检测结果计算获得多个热轧质量检测参数的预测值，具体是对多目标热轧质量检测结果进行反归一化(即)后，获得多个热轧质量检测参数的预测值。
其中，待预测的预处理后的热轧生产数据是对采集的待预测热轧生产数据进行步骤2中的预处理后获得的。
仿真实验中，收集热轧制造现场数据，包括59个操作变量，6个质量检测参数，包括凸度、中心线偏移量、出口温度、楔形、宽度和平直度，每组有60000个数据，每组使用12000个数据进行训练，45000个数据进行验证，采用均方根误差(RMSE)和平均绝对误差(MAE)作为预测性能，其中卷积神经网络(CNN)，长短期记忆网络(LSTM)，门控循环网络(GRU)，深度神经网络(DNN)作为对比算法，预测结果如表1所示。
表1为本发明与现有方法之间的预测结果对比表
从表中可以看到，本发明提出的PLE-MTL在6个质量检测参数上都取得了很好的预测结果，图3分别为6个质量检测参数预测误差的频率直方图，大部分误差都集中在0附近，因此证明提出的方法在热轧生产过程产品质量预测上有更高的表现。
上述实施例用来解释说明本发明，而不是对本发明进行限制，在本发明的精神和权利要求的保护范围内，对本发明做出的任何修改和改变，都落入本发明的保护范围。

Claims (10)

  1. 一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，包括以下步骤：
    步骤1：收集热轧生产过程数据，确定热轧生产过程数据对应的多个热轧质量检测参数，从而构建获得热轧生产数据集；
    步骤2：对热轧生产数据集进行数据预处理，获得预处理后的数据集；
    步骤3：根据预处理后的数据集，对基于渐进式分层提取的多目标网络进行训练，获得多目标热轧质量检测模型；
    步骤4：将待预测的预处理后的热轧生产数据输入到多目标热轧质量检测模型中，输出获得多目标热轧质量检测结果，基于多目标热轧质量检测结果计算获得多个热轧质量检测参数的预测值。
  2. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述步骤1中，热轧生产过程数据包括轧制力、轧制速度、弯辊力、辊缝值、窜辊量、轧入厚度和入口温度。
  3. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述步骤1中，热轧质量检测参数包括凸度、中心线偏移量、出口温度、楔形、宽度和平直度。
  4. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述步骤2具体为：
    首先，根据3-sigma原则剔除热轧生产数据集中的离群值，获得剔除后的数据集；接着，对剔除后的数据集进行归一化处理，获得归一化后的数据集，最后，利用滑动窗口方法对归一化后的数据集中的数据进行处理后，获得预处理后的数据集。
  5. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述基于渐进式分层提取的多目标网络由第一层渐进式分层提取网络、第二层渐进式分层提取网络和多目标输出网络组成；
    第一层渐进式分层提取网络由k个第一私有网络、1个第一共享网络和k+1个第一门控网络构成；第二层渐进式分层提取网络由k个第二私有网络、1个第二共享网络和k个第二门控网络构成；多目标输出网络由k个输出网络组成；
    基于渐进式分层提取的多目标网络的输入均作为k个第一私有网络、1个第一共享网络的输入，k个第一门控网络分别与对应的第一私有网络相连，k个第一门控网络还分别与第一共享网络相连，剩余的一个第一门控网络与k个第一 私有网络和第一共享网络均相连，基于渐进式分层提取的多目标网络的输入均与k+1个第一门控网络相连，k+1个第一门控网络的输出作为第一层渐进式分层提取网络的输出并分别与第二层渐进式分层提取网络的k个第二私有网络、1个第二共享网络相连；
    k个第二私有网络分别与对应的第二门控网络相连，第二共享网络均与k个第二门控网络相连，k个第一门控网络的输出还分别与对应的第二门控网络相连，k个第二门控网络分别与k个输出网络相连。
  6. 根据权利要求5所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述k个第一私有网络和1个第一共享网络的结构相同，均由卷积神经网络、门控循环网络和长短期记忆网络并行构成，其中，卷积神经网络由2层一维卷积层、1层一维最大池化层和1层全连接层依次相连构成，门控循环网络由1层门控循环单元和1层全连接层依次相连构成，长短期记忆网络由1层长短期记忆单元和1层全连接层依次相连构成；
    所述k个第二私有网络和1个第二共享网络的结构相同，均由3个全连接层网络并行构成。
  7. 根据权利要求5所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述k+1个第一门控网络的结构相同，均由1层门控循环单元和1层全连接层依次相连构成；所述k个第二门控网络的结构相同，均由2层全连接层依次相连构成。
  8. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述基于渐进式分层提取的多目标网络在训练时的单目标损失函数为均方误差，每个单目标损失函数的初始权重为1，根据动态加权平均法则进行训练。
  9. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述步骤4中，待预测的预处理后的热轧生产数据是对采集的待预测热轧生产数据进行步骤2中的预处理后获得的。
  10. 根据权利要求1所述的一种基于渐进式分层提取的多目标热轧质量检测方法，其特征在于，所述步骤4中，对多目标热轧质量检测结果进行反归一化后，获得多个热轧质量检测参数的预测值。
PCT/CN2023/128312 2023-07-18 2023-10-31 一种基于渐进式分层提取的多目标热轧质量检测方法 Ceased WO2025015753A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310880343.0 2023-07-18
CN202310880343.0A CN117009815B (zh) 2023-07-18 2023-07-18 一种基于渐进式分层提取的多目标热轧质量检测方法

Publications (1)

Publication Number Publication Date
WO2025015753A1 true WO2025015753A1 (zh) 2025-01-23

Family

ID=88562919

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/128312 Ceased WO2025015753A1 (zh) 2023-07-18 2023-10-31 一种基于渐进式分层提取的多目标热轧质量检测方法

Country Status (2)

Country Link
CN (1) CN117009815B (zh)
WO (1) WO2025015753A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120196565A (zh) * 2025-03-06 2025-06-24 齐鲁工业大学(山东省科学院) 基于多头注意力机制和rnn-lstm网络的数据预取方法
CN120790676A (zh) * 2025-06-12 2025-10-17 上海宝信软件股份有限公司 中厚板自动厚度控制方法、系统及介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119378756B (zh) * 2024-11-01 2025-09-26 中国民航大学 一种目标时段的时长预测方法、设备及介质

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110428175A (zh) * 2019-08-05 2019-11-08 东北大学秦皇岛分校 一种基于深度学习的热轧带钢板凸度预报方法
CN112001527A (zh) * 2020-07-29 2020-11-27 中国计量大学 多特征融合深度神经网络的工业生产过程目标数据预测方法
US20230004781A1 (en) * 2020-12-04 2023-01-05 Northeastern University Lstm-based hot-rolling roll-bending force predicting method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115034803B (zh) * 2022-04-13 2025-06-24 北京京东尚科信息技术有限公司 新物品挖掘方法和装置及存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110428175A (zh) * 2019-08-05 2019-11-08 东北大学秦皇岛分校 一种基于深度学习的热轧带钢板凸度预报方法
CN112001527A (zh) * 2020-07-29 2020-11-27 中国计量大学 多特征融合深度神经网络的工业生产过程目标数据预测方法
US20230004781A1 (en) * 2020-12-04 2023-01-05 Northeastern University Lstm-based hot-rolling roll-bending force predicting method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
张大海 等 (ZHANG, DAHAI ET AL.): "基于耦合关系挖掘及渐进式分层提取多任务学习网络的风-光-荷短期预测 (Wind-photovoltaic-load Short-term Forecast Based on Coupling Relation Mining and Progressive Layered Extraction of Multi-task Learning network)", 电网技术 (POWER SYSTEM TECHNOLOGY), 1 September 2022 (2022-09-01), pages 3538 - 3542 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120196565A (zh) * 2025-03-06 2025-06-24 齐鲁工业大学(山东省科学院) 基于多头注意力机制和rnn-lstm网络的数据预取方法
CN120790676A (zh) * 2025-06-12 2025-10-17 上海宝信软件股份有限公司 中厚板自动厚度控制方法、系统及介质

Also Published As

Publication number Publication date
CN117009815B (zh) 2025-07-18
CN117009815A (zh) 2023-11-07

Similar Documents

Publication Publication Date Title
CN111738512B (zh) 一种基于cnn-ipso-gru混合模型的短期电力负荷预测方法
CN117009815B (zh) 一种基于渐进式分层提取的多目标热轧质量检测方法
CN101604144B (zh) 一种板材轧制在线控制模型的建模方法
CN104200268B (zh) 一种基于粒子群优化极限学习机的带钢出口厚度预测方法
CN110096810B (zh) 一种基于逐层数据扩充深度学习的工业过程软测量方法
CN112990556A (zh) 一种基于Prophet-LSTM模型的用户用电能耗预测方法
WO2024174300A1 (zh) 一种基于图神经网络的工艺质量时间序列预测方法
CN114118789B (zh) 基于模糊综合评判和综合赋权的雷达发射机状态评估方法
CN112861436A (zh) 一种发动机排放实时预测方法
CN110428175A (zh) 一种基于深度学习的热轧带钢板凸度预报方法
CN110222387B (zh) 基于混合漏积分crj网络的多元钻井时间序列预测方法
CN112836884A (zh) 基于Copula-DBiLSTM的综合能源系统多元负荷精确预测方法
CN110516391A (zh) 一种基于神经网络的航空发动机动态模型建模方法
CN106991212A (zh) 一种基于ga_pso优化grnn网络算法的叶根强度安全特性预测方法
CN112434848A (zh) 基于深度信念网络的非线性加权组合风电功率预测方法
CN118157162B (zh) 基于神经网络的新型电力系统等效惯量评估方法及装置
CN118504391B (zh) 一种夹心波纹钢管涵洞的稳定性分析方法
CN113609772A (zh) 基于卷积准循环神经网络的水泥成品比表面积预测方法
CN106296434B (zh) 一种基于pso-lssvm算法的粮食产量预测方法
CN107274009A (zh) 一种基于相关性的时间序列数据多步预测方法和系统
CN112035970A (zh) 一种基于船用柴油机关键性能的参数优化方法
CN115526308A (zh) 基于宽度剪枝回声状态网络的时间序列预测方法
CN116662925A (zh) 一种基于加权稀疏神经网络工业过程软测量方法
CN118410702A (zh) 一种基于EnKF滤波修正RBF神经网络的高维气动力建模方法
CN117688376A (zh) 一种海洋工程结构物外腐蚀预测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23945633

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 23945633

Country of ref document: EP

Kind code of ref document: A1