Strip steel hot continuous rolling industrial process monitoring method and device
Through the DDIM data enhancement and random occlusion feature focus mechanism, the unsupervised monitoring model is constructed, which solves the problem of insufficient data during the hot continuous rolling of strip steel, realizes efficient abnormal working conditions detection, and improves monitoring accuracy and sensitivity.
Patent Information
- Application Number
- CN202510343086.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-21
- Publication Date
- 2025-07-18
AI Technical Summary
The prior art lacks data samples during the hot continuous rolling process of strip steel, insufficient extraction of dynamic timing features, and difficult to achieve effective abnormal working conditions monitoring. Traditional methods lack applicability and accuracy when facing complex working conditions.
DDIM is used for data enhancement, combined with random occlusion and feature focus collaboration mechanism, an unsupervised process monitoring model is built, and the control limit is determined using kernel density estimation to realize abnormal working conditions detection.
It improves the accuracy and sensitivity of thermal continuous rolling process monitoring of strip steel, reduces the risk of overfitting, enhances the generalization ability and robustness of the model, and provides a reliable process monitoring solution.
Smart Images

Figure CN120325685A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of process monitoring in the iron and steel metallurgy industry, and particularly to a method and device for industrial process monitoring in hot strip continuous rolling. Background Art
[0002] Hot strip continuous rolling is an important process in iron and steel production, widely used in the production of steel products such as steel plates and steel strips. Its main process includes feeding the preheated billet through high-temperature roller tables into a series of rolling mills, and gradually compressing the billet into steel with the required thickness and width through multiple rolling processes. The hot strip continuous rolling process usually involves multiple links, including heating, rolling, cooling, and final cutting and coiling operations. Each link requires precise process control and real-time monitoring to ensure product quality and production efficiency. Traditional methods for monitoring the hot strip continuous rolling process rely on process models and sensor data. These methods usually adopt control technologies based on analytical models, such as state estimation and parameter estimation methods based on thermodynamics and physical process models. However, due to the high complexity of the hot strip continuous rolling process and large fluctuations in working conditions, the applicability and accuracy of these methods are poor. Especially when faced with a large amount of high-dimensional, complex, and ever-changing data, it is often difficult to achieve effective process diagnosis and fault prediction.
[0003] In the era of Industry 4.0, artificial intelligence technology and big data processing technology have developed rapidly. The hot strip continuous rolling industry has gradually introduced emerging technologies such as the Internet of Things, cloud computing, artificial intelligence, and machine learning, and integrated them into production facilities and the entire operation process. Modern hot strip continuous rolling production lines are equipped with advanced sensors, embedded software, and automation equipment. By collecting and analyzing process data in real time, the rolling process is optimized, product quality is improved, and intelligent decision-making is achieved. The application of these technologies makes the production process more efficient and flexible, and can detect potential problems in a timely manner, thus enhancing the production stability and quality control level. In addition, with the large-scale development of production enterprises and the continuous improvement of product quality requirements, modern industrial processes have characteristics such as unclear process mechanisms, large fluctuations in working conditions, and strong variable coupling and correlation. These characteristics pose major challenges to traditional process monitoring methods.
[0004] Currently, for the problem of insufficient industrial process data in hot strip continuous rolling, there are mainly transfer learning methods and data generation methods based on generative adversarial networks. However, these methods all have deficiencies:
[0005] Firstly, transfer learning has certain limitations. Especially when there are large differences between the source domain and the target domain, transfer learning may introduce irrelevant or even harmful information, which instead reduces the diagnostic performance;
[0006] Secondly, generative adversarial networks are prone to mode collapse during the training process, resulting in a lack of diversity in the generated samples, especially in the generation of high-dimensional data, where there are still certain limitations.
[0007] Therefore, there is no efficient data generation and process monitoring method for hot strip continuous rolling process monitoring at home and abroad, especially when facing the problem of lack of high-quality and large-scale data under complex working conditions. Summary of the Invention
[0008] In order to solve the technical problems existing in the prior art, such as insufficient data samples in the hot strip continuous rolling industrial process, insufficient extraction of dynamic time series features, and difficulty in realizing effective abnormal working condition monitoring, the embodiments of the present invention provide a hot strip continuous rolling industrial process monitoring method and device. The technical solutions are as follows:
[0009] On the one hand, a hot strip continuous rolling industrial process monitoring method is provided. This method is implemented by a hot strip continuous rolling industrial process monitoring device, and the method includes:
[0010] Offline Modeling Stage:
[0011] S1, Collect multi-dimensional sensor data of the hot strip continuous rolling industrial process for standard cleaning;
[0012] S2, After using DDIM to perform data augmentation on the cleaned data, introduce a sliding window to extract a time window of a fixed length; where DDIM represents the Denoising Diffusion Implicit Model;
[0013] S3, Use a feature focusing collaboration mechanism based on random occlusion to randomly occlude the data extracted by the sliding window;
[0014] S4, Use the time and space information of the hot strip continuous rolling industrial process data to construct an unsupervised process monitoring model, and use the occluded data to train the constructed unsupervised process monitoring model;
[0015] S5, Use the kernel density estimation method to process the data reconstructed by the trained unsupervised process monitoring model to determine the control limit;
[0016] Online Monitoring Stage:
[0017] S6, Calculate the reconstruction error based on the multi-dimensional sensor data of the hot strip continuous rolling industrial process collected in real time, and compare the reconstruction error with the control limit to realize the detection of abnormal working conditions in the hot strip continuous rolling industrial process.
[0018] On the other hand, a hot strip continuous rolling industrial process monitoring device is provided. This device is applied to the hot strip continuous rolling industrial process monitoring method, and the device includes:
[0019] The first processing module is used to collect multi-dimensional sensor data of the strip hot continuous rolling industrial process for standard cleaning;
[0020] The data enhancement module is used to perform data enhancement on the cleaned data by using DDIM (Denoising Diffusion Implicit Model), and then introduce a sliding window to extract a time window of a fixed length; where DDIM represents the denoising diffusion implicit model;
[0021] The random occlusion module is used to randomly occlude the data extracted by the sliding window by using a feature focusing cooperation mechanism based on random occlusion;
[0022] The model training module is used to construct an unsupervised process monitoring model by using the time and space information of the strip hot continuous rolling industrial process data, and use the data after the occlusion operation to train the constructed unsupervised process monitoring model;
[0023] The second processing module is used to process the data reconstructed by the trained unsupervised process monitoring model by using the kernel density estimation method to determine the control limit;
[0024] The data evaluation module is used to calculate the reconstruction error according to the multi-dimensional sensor data of the strip hot continuous rolling industrial process collected in real time, and compare the reconstruction error with the control limit to realize the detection of abnormal working conditions in the strip hot continuous rolling industrial process.
[0025] On the other hand, a strip hot continuous rolling industrial process monitoring device is provided, and the strip hot continuous rolling industrial process monitoring device includes: a processor; a memory, and a computer-readable instruction is stored on the memory, and when the computer-readable instruction is executed by the processor, any one of the methods in the above strip hot continuous rolling industrial process monitoring method is implemented.
[0026] On the other hand, a computer-readable storage medium is provided, and at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement any one of the methods in the above strip hot continuous rolling industrial process monitoring method.
[0027] The beneficial effects brought by the technical solution provided by the embodiments of the present invention at least include:
[0028] (1) The denoising diffusion implicit model is introduced to fully mine and utilize the information in the original data, generate more enhanced data that conforms to the distribution characteristics of the strip hot continuous rolling industrial process, improve the generalization ability of the subsequent unsupervised process monitoring model, and thus improve the accuracy of the strip hot continuous rolling industrial process monitoring; and the U-net structure therein is improved, which can more effectively learn the spatio-temporal characteristics of the original data and generate higher-quality industrial process data;
[0029] (2) The feature focusing collaborative mechanism based on random occlusion is introduced, where elements at some time steps are randomly replaced with 0, significantly reducing the overfitting risk during the training of the monitoring model (referring to the unsupervised process monitoring model) in the hot strip continuous rolling industrial process, improving the generalization ability of the model, and enhancing the model robustness;
[0030] (3) An unsupervised process monitoring model is proposed, which combines the structures of an autoencoder and a gated recurrent unit. The encoder layer uses the gated recurrent unit to extract temporal dynamic features; the gated recurrent unit is used to solve the problems of inability to long-term remember and gradient vanishing and gradient explosion in backpropagation in the recurrent neural network, and is used to further focus on the temporal features in the collected sensor data;
[0031] (4) A method for monitoring the hot strip continuous rolling industrial process based on an improved implicit diffusion model and unsupervised learning is constructed, effectively improving the generalization ability of the model under small sample conditions. The combined spatio-temporal feature joint extraction mechanism significantly improves the monitoring sensitivity of the hot strip continuous rolling industrial process, providing a reliable process monitoring scheme for the hot rolling process. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0033] Figure 1 It is a flowchart of a method for monitoring the hot strip continuous rolling industrial process provided by an embodiment of the present invention;
[0034] Figure 2 It is a schematic diagram of the online monitoring and offline training process provided by an embodiment of the present invention;
[0035] Figure 3 It is a schematic diagram of the standardized cleaning process provided by an embodiment of the present invention;
[0036] Figure 4 It is a schematic diagram of the data augmentation process provided by an embodiment of the present invention;
[0037] Figure 5 It is a block diagram of a device for monitoring the hot strip continuous rolling industrial process provided by an embodiment of the present invention;
[0038] Figure 6 It is a schematic diagram of the structure of a device for monitoring the hot strip continuous rolling industrial process provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0039] The technical solutions in the present invention will be described below with reference to the accompanying drawings.
[0040] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two.
[0041] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference is not emphasized, the meanings they express are the same.
[0042] In the embodiments of the present invention, sometimes subscripts such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meanings they express are the same.
[0043] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.
[0044] The embodiments of the present invention provide a method for monitoring the hot strip continuous rolling industrial process. This method can be implemented by a monitoring device for the hot strip continuous rolling industrial process, and the monitoring device for the hot strip continuous rolling industrial process can be a terminal or a server. As Figure 1 and Figure 2 shown in the flowchart of the method for monitoring the hot strip continuous rolling industrial process, the processing flow of this method can include the following steps:
[0045] Offline modeling stage:
[0046] S1. Collect multi-dimensional sensor data (which can also be called: multi-source heterogeneous data) of the hot strip continuous rolling industrial process for standard cleaning.
[0047] S2. After using DDIM (Denoising Diffusion Implicit Model, DDIM) to perform data enhancement on the cleaned data, introduce a sliding window to extract a time window with a fixed length; where DDIM represents a denoising diffusion implicit model.
[0048] S3. Use a feature focusing collaboration mechanism based on random occlusion to perform random occlusion on the data extracted by the sliding window.
[0049] S4. Construct an unsupervised process monitoring model, and use the data after the occlusion operation to train the constructed unsupervised process monitoring model. During the training process, the unsupervised process monitoring model learns the temporal features (such as the dynamic evolution trends of temperature, pressure, speed, and thickness) and spatial features (the physicochemical coupling relationships between different sensors) in the strip hot rolling industrial process data to reconstruct the data;
[0050] S5. Use the Kernel Density Estimation (KDE) method to process the data reconstructed by the trained unsupervised process monitoring model to determine the control limits;
[0051] Online monitoring stage:
[0052] S6. Calculate the reconstruction error based on the multi-dimensional sensor data of the strip hot rolling industrial process collected in real time, and compare the reconstruction error with the control limits to achieve the detection of abnormal working conditions in the strip hot rolling industrial process.
[0053] In this embodiment, as Figure 3 shown, the standardization cleaning (S1) of collecting the multi-dimensional sensor data of the strip hot rolling industrial process includes:
[0054] S11. Collect the multi-dimensional sensor data of the entire process of the strip hot rolling industrial process, and perform a preliminary screening on the collected multi-dimensional sensor data to obtain the data that has a key impact on the monitoring of the strip hot rolling industrial process; among them, the data collected in the offline modeling stage is the data under normal working conditions;
[0055] In this embodiment, using the advanced sensors and industrial Internet technology on the strip hot rolling site, collect the multi-dimensional real-time data of the entire process of the strip hot rolling industrial process (including: temperature, pressure, tension, flow rate, rolling force, speed, thickness, roll gap, etc.) and upload it to the data platform. Considering that the strip rolling process is affected by many aspects such as temperature and rolling force, and there is a large amount of redundant data on site, the data platform conducts a preliminary screening on the multi-dimensional sensor data directly collected on site, eliminates redundant and invalid data, obtains the data that has a key impact on the monitoring of the strip hot rolling industrial process and saves it to the data platform, avoiding the interference of data redundancy or irrelevant data on the system performance, and providing data support for subsequent operations.
[0056] In this embodiment, the data module at the hot strip continuous rolling production site is used as the digital base, which includes the field data bus and the high-performance programmable logic controller. The data collected by the field sensors is uploaded to the data platform in real time through the industrial Internet, field bus, etc. The data platform takes the Pspace database as the core, and combines the time series database influxDB and the relational database mySQL to build an efficient and reliable data storage and processing system. The data platform supports the access and integration of multi-source heterogeneous data, and can clean and standardize the data uploaded in real time to ensure data quality.
[0057] In this embodiment, the data platform combines the strip steel mechanism model to establish a multi-variable correlation analysis mechanism. Since metal is ductile, during the rolling process, the length of the strip steel becomes longer with the process. Therefore, the data platform adopts a data alignment technology to align the strip steel of different processes according to the percentile points, achieving the unity of data in the time and space dimensions.
[0058] S12, preprocess the filtered data; among them, the preprocessing includes: normalization and missing value filling.
[0059] In this embodiment, since there are data of different magnitudes in the hot strip continuous rolling production line, the data collected by the sensors covers various types from physical quantities to control parameters. For example, the data ranges of physical quantities such as temperature, rolling force, and speed may vary greatly. Therefore, data cleaning processing operations such as normalization and missing value filling need to be performed on the filtered data to ensure the consistency of data magnitude, improve data quality, ensure the reliability of subsequent model training, and effectively avoid problems such as data loss and data pollution caused by the complex industrial site environment.
[0060] In this embodiment, for the convenience of subsequent random occlusion, the filtered data needs to be reconstructed into data with a mean of 0, that is, the random occlusion can be directly filled with a mean of 0. The following normalization processing is performed on the filtered data:
[0061]
[0062] Among them, x ori is the filtered data, x is the data obtained by normalization processing, μ is the overall mean of the input data, and σ is the standard deviation.
[0063] In this embodiment, as Figure 4 shown, after data augmentation of the cleaned data using DDIM, introducing a sliding window to extract a fixed-length time window (S2) includes:
[0064] S21. Use DDIM to perform implicit space modeling on the cleaned data, generate more enhanced data that conforms to the distribution characteristics of the hot strip continuous rolling industrial process, and upload it to the data platform to achieve data enhancement in the case of insufficient samples, which can improve the generalization ability of subsequent unsupervised process monitoring models.
[0065] In the existing DDPM (Denoising Diffusion Probabilistic Models), the denoising process is a Markov chain process, that is, there is a tight logical relationship between the front and back data, so interval derivation cannot be performed, resulting in too slow inference speed. To solve the above problems, in this embodiment, a data enhancement method based on DDIM is constructed, which breaks the limitation of the Markov chain and modifies the sampler's reverse denoising process, which can greatly increase the sampling speed.
[0066] The basic principle of DDIM is to construct a forward diffusion - reverse denoising, use DDIM to learn the noise distribution therein, so as to realize the generation from noise to data. DDIM includes two key stages: forward diffusion and reverse denoising.
[0067] Forward diffusion is to gradually add noise with a normal distribution to the original data until the original data becomes blurred and approaches pure noise. Pure noise means the diversity of data distribution, that is, when reverse denoising, the model can generate more diverse data to achieve sample expansion. Specifically expressed as:
[0068] Assume that the original data x0 input to DDIM satisfies the distribution q(x0), denoted as x0 ∼ q(x0). For the time step t ∈ [1, T], use x0, x1, …, x t , …, x T to represent the values after each step of diffusion. q(x0) represents the function distribution satisfied by the original data, T represents the time step when the diffusion process terminates, x t represents the value of the original data x0 after diffusion at the t time step, and x t and x t-1 satisfy the following formula:
[0069]
[0070] where ε ∼ N(0, 1), β t is a hyperparameter and approaches 1 as time t increases. ε is the noise term, and ε ∼ N(0, 1) means that ε follows the standard normal distribution with a mean of 0 and a variance of 1. For convenience of representation, let α t = 1 - β t , and by combining mathematical induction, the mathematical expressions of x t and x0 are derived:
[0071]
[0072] For the convenience of subsequent calculations, let After further simplification, we have:
[0073]
[0074] Finally, x t can be directly obtained from x0 without a multi-step iteration process, which greatly reduces the computational complexity of forward diffusion. Since β t increases, then α t decreases. It is not difficult to deduce that when t → T, then x T → ε, that is, in the process of forward diffusion, after a sufficient number of iterations, the finally generated data approaches a Gaussian distribution.
[0075] The core idea of reverse denoising is to denoise the noisy data so that it gradually returns to the distribution of the original data. In the process of denoising, it is necessary to predict the data of the previous time step based on the data of the current time step, which means removing a part of the noise and restoring it to the data of the previous time step. Specifically expressed as:
[0076] Given P(x t-1 |x t , x0), assuming it satisfies the following normal distribution, P(x t-1 |x t , x0) represents the conditional probability distribution for the convenience of subsequent calculations:
[0077] P(x t-1 |x t , x0) ~ N(kx0 + mx t , σ 2 )
[0078] x t-1 = kx0 + mx1 + σε, ε ~ N(0, 1)
[0079] Among them, k, m, and σ are intermediate variables. P(x t-1 |x t , x0) ~ N(kx0 + mx t , σ 2 ) means that P(x t-1 |x t , x0) follows a normal distribution with a mean of kx0 + mx t and a variance of σ 2 ;
[0080] It is known that during the noise addition process:
[0081]
[0082] Then it can be deduced that:
[0083]
[0084] Among them, indicates that ε′ follows a normal distribution with a mean of 0 and a variance of ;
[0085] According to the same corresponding coefficients, k and m are obtained, and the conditional probability distribution satisfied by x is given again:
[0086]
[0087] Replace ε in the above distribution with ε t to obtain:
[0088]
[0089] Among them, x prev represents the data generated in reverse denoising, represents the cumulative noise attenuation factor;
[0090] Since the above derivation process does not use the Markov chain process, from x t to x0, strict step-by-step reasoning may not be required, and x prev can be used to replace x t-1 , that is, there can be any number of iteration steps between x prev and x t .
[0091] In this embodiment, by reconstructing the noise characteristics, DDIM generates an output similar to the input. A loss function is constructed by measuring the difference between the predicted noise and the actual noise, thereby guiding the optimization of DDIM. DDIM predicts the noise ε θ (x t , t, c) at each time step t, where x t is the value after diffusion of the original data x0 at time step t, c is a constant term, and the data is gradually denoised. The reconstruction loss is used to measure the difference between the predicted noise and the actual noise, thereby guiding the optimization of DDIM, is expressed as:
[0092]
[0093] Among them, x t is generated by forward diffusion, ε is a noise term that follows a standard normal distribution, represents the mean squared error (MSE, Mean Squared Error), ‖ε - ε θ (x t , t)‖2 It represents the gap between the predicted noise of DDIM and the true noise. Finally, DDIM fits the noise at each time step, and after the reverse denoising process, the generated data is obtained.
[0094] In this embodiment, DDIM uses an improved U-net structure for feature learning. As a network architecture composed of an encoder and a decoder, the traditional U-net can effectively capture multi-scale features and accurately estimate noise. To fully exploit the spatio-temporal characteristics of the original data and improve the computational efficiency, the traditional U-net structure is optimized as follows: smaller convolutional kernels are used, and the number of convolutional layers is streamlined. The specific improvement points are as follows:
[0095] 1. The traditional U-net uses larger convolutional kernels for feature extraction, while the improved U-net uses smaller convolutional kernels, thereby reducing the computational complexity, accelerating the training and inference speed, enhancing the ability to capture the features of the original data, and avoiding information redundancy.
[0096] 2. The traditional U-net uses multiple convolutional layers for feature extraction, while the improved one streamlines the number of convolutional layers, making the model more lightweight, suitable for running in resource-constrained environments, and preventing over-compression of information, thereby maintaining key information and improving the quality of the generated data.
[0097] S22, fuse the cleaned data and the augmented data, and extract fixed-length time windows from the fused data using a sliding window;
[0098] In this embodiment, due to the insufficient original data samples, it is difficult to fully extract the sample features of a normal distribution, so the method of sliding window is introduced. This method extracts fixed-length time windows from the time series data, enabling the subsequent unsupervised process monitoring model to capture the time series features, and at the same time combining the spatial dimension features, such as the coupling relationship between sensor data such as temperature, pressure, and flow rate, to enhance the modeling ability of the hot strip continuous rolling industrial process.
[0099] In this embodiment, the cleaned data and the augmented data are fused to obtain X z , Extract a fixed-length subsequence X i through the sliding window, which is expressed as follows:
[0100] X i = X z [i·s:i·s+w,:]
[0101] where N represents the number of samples, M represents the number of variables, w is the length of the sliding window, s is the step size of the sliding window, i is the index of the sliding window, and i ∈ {0,1,2,…,n}. The finally generated results are stacked as a three-dimensional tensor, expressed as follows:
[0102]
[0103] where X window is the data extracted by the sliding window.
[0104] In this embodiment, the use of the feature focusing and collaborative mechanism based on random occlusion to perform random occlusion on the data extracted by the sliding window (S3) includes:
[0105] For the data X window extracted by the sliding window, using the feature focusing and collaborative mechanism based on random occlusion to randomly replace the elements of some time steps with 0 to obtain the data X m after the occlusion operation:
[0106] X m = X window ⊙ M
[0107]
[0108] where M is the occlusion matrix, and M i,j is the element in the i-th row and j-th column of the matrix M; p ∈ (0, 1) represents the retention rate, which is a hyperparameter related to the size of the occlusion area; ⊙ represents the Hadamard multiplication of matrices, that is, multiplying the elements at the same position one by one.
[0109] In this embodiment, a feature focusing and collaborative mechanism based on random occlusion is designed to perform random occlusion on the data extracted by the sliding window, thereby automatically identifying the key feature information and forcing the subsequent unsupervised process monitoring model in S4 to focus on the core information in the data, greatly reducing the overfitting risk when training the monitoring model (referring to the unsupervised process monitoring model) in the hot strip mill industrial process monitoring, improving the generalization ability of the unsupervised process monitoring model, and improving the robustness of the model.
[0110] In this embodiment, in step S4, an unsupervised process monitoring model is constructed, and the constructed unsupervised process monitoring model is trained using the data after the occlusion operation. During the training process, the unsupervised process monitoring model learns the time features (such as the dynamic evolution trends of temperature, pressure, speed, and thickness) and spatial features (the physical and chemical coupling relationships between different sensors (such as temperature, pressure, tension, and thickness sensors)) in the hot strip mill industrial process data to reconstruct the data, providing a basis for subsequent control limit calculation;
[0111] In this embodiment, the unsupervised process monitoring model adopts an autoencoder-gated recurrent unit unsupervised learning framework, that is: the AE-GRU model (AutoEncoder-Gate Recurrent Unit, autoencoder-gated recurrent unit);
[0112] The autoencoder-gated recurrent unit unsupervised learning framework includes: an encoder layer, a fully connected layer connected to the encoder layer, and a decoder layer connected to the fully connected layer;
[0113] The encoder layer uses a gated recurrent unit for feature extraction.
[0114] In this embodiment, the autoencoder (AE) is a typical structure of unsupervised learning and is often used to compress high-dimensional data. It is internally divided into two major structures: an encoder and a decoder. Encoding and decoding are mainly achieved through neural networks. The encoder is used to reduce the dimension of the original input data and give the low-dimensional feature distribution of the original data in the latent space. The decoder is used to expand the low-dimensional features and reconstruct the original data.
[0115] The encoding-decoding process of the autoencoder is as follows: In the encoding operation, the input sample x is mapped into the latent space (Code), which is represented by the following formula:
[0116] y = f(wx + b)
[0117] where f is the activation function of the encoding layer, w is the weight parameter in the encoder, and b is the bias vector.
[0118] The decoding process is opposite to the encoding process, that is, mapping from low dimension to high dimension to reconstruct the input process, which is specifically represented by the following formula:
[0119]
[0120] where w′, b′ are the parameters inside the decoder, and g represents the mapping process of the decoder, is the reconstructed data.
[0121] In the encoding-decoding process, the parameters D(w; w′; b; b′) of the autoencoder are adjusted by minimizing the reconstruction error, and the reconstruction loss Loss is calculated using a non-linear process. To avoid the overfitting problem of the autoencoder, an L1 regularization term is added, and the final loss function is as follows:
[0122]
[0123] where x i is the actual input sample, is the reconstructed data, N is the number of samples, p is the total number of autoencoder parameters, λ is the hyperparameter controlling the regularization strength, wj The weight parameters of the autoencoder.
[0124] In the strip hot continuous rolling industrial process, the data has significant time dependence. By integrating GRU into AE, GRU can effectively capture the temporal dynamic characteristics of the input data, thus more accurately describing the variation law of the dynamic process, predicting the evolution trend of features in the time dimension, and providing a basis for subsequent control limit calculation.
[0125] GRU (Gate Recurrent Unit) is a variant of the Recurrent Neural Network (RNN), which solves the problems of inability to long-term memory and gradient vanishing and gradient explosion in backpropagation in RNN. Compared with its predecessor LSTM, GRU combines the input gate and the forget gate into an update gate, with a simpler network structure. While maintaining computational efficiency, it has high performance and has more advantages than LSTM in medium and short-term dependence tasks, and is more suitable for industrial process monitoring.
[0126] In order to make full use of the unsupervised learning ability of the autoencoder and the temporal feature learning ability of GRU, this embodiment constructs an AE-GRU unsupervised learning framework suitable for strip hot continuous rolling industrial process monitoring. This framework consists of three parts: an encoder layer, a fully connected layer connected to the encoder layer, and a decoder layer connected to the fully connected layer; among them, the low-dimensional features extracted by the encoder layer usually still contain redundant information, and the fully connected layer connected to the encoder layer can further reduce the dimension, retain the main information and remove noise, so that the subsequent decoder can process the features more efficiently.
[0127] In this embodiment, the encoder layer uses a gated recurrent unit to extract temporal dynamic features and maps the input to a low-dimensional latent space. For the input data X m that has been randomly occluded, the hidden state of GRU is calculated by the following recursive formula:
[0128] h t = GRU(x mt , h t-1 )
[0129] Among them, h t represents the hidden state at the t-th time step in the encoder layer, and x mt represents the data of X m at time step t.
[0130] In order to further compress the features in the latent space, the last hidden state h T of the encoder layer is input into the fully connected layer of the AE-GRU model and compressed into a feature representation z:
[0131] z = σ(Wh T + b)
[0132] Among them, w and b are the weights and biases of the fully connected layer of the AE-GRU model, σ is the ReLU activation function. After the above operations, the fully connected layer of the AE-GRU model learns the joint distribution between data and extracts the key spatial features of the data.
[0133] Based on the decoder layer, perform time series reconstruction on the feature z, and use z as the initial hidden state of the decoder layer Gradually generate the reconstructed values at each time step represents the hidden state at the t-th time step in the decoder layer. For the convenience of distinction, the decoder layer function is generally represented as GRU′, σ is the ReLU activation function, W dec and b dec are the weights and biases of the fully connected layer of the decoder layer:
[0134]
[0135] The AE-GRU model optimizes the model parameters by maintaining the complexity of the model and minimizing the error between the input data and the reconstructed data. The loss function is defined as:
[0136]
[0137] In the above loss function, represents the neuron parameters in the encoder layer, the fully connected layer connected to the encoder layer, and the decoder layer connected to the fully connected layer. T is the normalization factor, a is the total number of time steps, and b is the total number of neuron parameters in the encoder layer, the fully connected layer connected to the encoder layer, and the decoder layer connected to the fully connected layer; the first part is the reconstruction loss, represents the input sample x t the reconstructed data after encoding-decoding, is the square of the L2 norm of the reconstruction error; the second part is the regularization term, represents the square of the L2 norm of the parameters of the AE-GRU model, reflecting the complexity of the AE-GRU model. λ represents the hyperparameter of the regularization strength, which is used to adjust the weight of the regularization term.
[0138] In summary, the encoder layer of the AE-GRU model uses a gated recurrent unit to extract temporal dynamic features, and uses a fully connected layer to learn the joint distribution between data and capture spatial correlations, so as to realize unsupervised representation learning of the process state.
[0139] In this embodiment, the method of using kernel density estimation to process the data reconstructed by the trained unsupervised process monitoring model to determine the control limit (S5) includes:
[0140] S51. For the data reconstructed by the unsupervised process monitoring model, calculate the reconstruction error between it and the original data as the statistic e i :
[0141]
[0142] E = [e1, e2, …, e k
[0143] where x i is the input sample of the unsupervised process monitoring model, is the data reconstructed by the unsupervised process monitoring model, e i is the reconstruction error of the i-th sample, k is the number of samples input to the unsupervised process monitoring model, and E is the reconstruction error matrix;
[0144] S52. Use the KDE to estimate the probability density function f(e) to model the obtained reconstruction error and estimate the probability density of each sample appearing, so as to comprehensively grasp the system operation state. KDE is a probability density estimation method based on non-parametric statistical theory. Different from traditional parametric methods, it does not require prior assumptions about the data distribution, so it has high flexibility and adaptability, and is especially suitable for processing data with complex, high-dimensional, and non-linear distributions.
[0145] In this embodiment, use the KDE to estimate the probability density function f(e), and its calculation formula is as follows:
[0146]
[0147] where f(e) represents the kernel density estimation probability density function, that is, the probability density of each sample appearing; h is the bandwidth parameter; K(·) is the kernel function; e is the intermediate variable. Similar to S51, k is the number of samples input to the unsupervised process monitoring model, and e i is the reconstruction error of the i-th sample;
[0148] After the KDE estimates the probability density function, the cumulative distribution function F(e) can be further calculated:
[0149]
[0150] where the cumulative distribution function F(e) represents the probability that the intermediate variable e is less than or equal to a certain value.
[0151] Through the KDE method, the data reconstructed by the unsupervised process monitoring model can be modeled to estimate the probability density of each sample. The magnitude of the probability density can reflect the rarity of the sample in the feature space, thereby revealing potential abnormal samples or state changes. Compared with traditional parametric methods, KDE can capture more details and can better handle the potential multimodal characteristics in the data, making it an effective tool in industrial process monitoring and anomaly detection.
[0152] In this embodiment, after utilizing the advantages of KDE for the data reconstructed by the unsupervised process monitoring model, the operating state of the system can be further analyzed through the dynamic change of the probability density value. The low probability density region usually corresponds to potential anomalies or unseen states. Therefore, by setting appropriate thresholds to distinguish normal states from abnormal states, a reliable monitoring mechanism is provided for the industrial process. In summary, without making any assumptions about the data distribution, KDE can flexibly adapt to a variety of complex scenarios, significantly improving the accuracy and robustness of process monitoring.
[0153] S53, find the first F(e) that satisfies the following equation to determine the control limit C:
[0154] F(e)≥τ
[0155] C=min{x|F(e)≥τ}
[0156] where τ represents the pre-specified confidence level.
[0157] In this embodiment, in step S6, multi-dimensional sensor data of the strip hot continuous rolling industrial process is collected in real time. Standardized cleaning, sliding window extraction of a fixed-length time window, and random occlusion operations are performed on the real-time collected sensor data. For the data after the occlusion operation, the trained unsupervised process monitoring model is used to reconstruct the data, calculate the reconstruction error, and compare the reconstruction error with the control limit to achieve anomaly condition detection in the strip hot continuous rolling industrial process, where the anomaly condition detection is expressed as:
[0158]
[0159] where e current represents the reconstruction error between the real-time data after the occlusion operation and its reconstructed data.
[0160] In this implementation, during online monitoring, the control limit is used to determine whether the current sample is normal, realizing anomaly condition detection, aiming to reduce the risk of production accidents and ensure the safe and stable operation of the industrial production process.
[0161] In this embodiment, a method for monitoring the hot strip continuous rolling industrial process based on the denoising diffusion implicit model and unsupervised learning is constructed, which is mainly divided into two parts: data generation and process monitoring, covering raw data collection, data cleaning, data augmentation, training of the unsupervised process monitoring model, calculation of offline control limits, and online monitoring of samples, realizing dynamic real-time data augmentation and process monitoring of the production process, and providing an efficient and reliable solution for the intelligent monitoring and control of industrial processes.
[0162] In summary, the method for monitoring the hot strip continuous rolling industrial process provided in this embodiment has at least the following beneficial effects:
[0163] (1) The denoising diffusion implicit model is introduced to fully mine and utilize the information in the raw data, generate more enhanced data that conform to the distribution characteristics of the hot strip continuous rolling industrial process, improve the generalization ability of the subsequent unsupervised process monitoring model, and thus improve the accuracy of the hot strip continuous rolling industrial process monitoring; and the U-net structure therein is improved, which can more effectively learn the spatio-temporal characteristics of the raw data and generate higher-quality industrial process data;
[0164] (2) The feature focusing collaboration mechanism based on random occlusion is introduced, and the elements at some time steps are randomly replaced with 0, which greatly reduces the overfitting risk when training the monitoring model (referring to the unsupervised process monitoring model) in the hot strip continuous rolling industrial process monitoring, improves the generalization ability of the model, and enhances the robustness of the model;
[0165] (3) An unsupervised process monitoring model is proposed, which combines the structures of the autoencoder and the gated recurrent unit. The encoder layer uses the gated recurrent unit to extract the temporal dynamic features; the gated recurrent unit solves the problems of inability to long-term remember and gradient disappearance and gradient explosion in the backpropagation in the recurrent neural network, and is used to further focus on the temporal features in the collected sensor data;
[0166] (4) A method for monitoring the hot strip continuous rolling industrial process based on the improved implicit diffusion model and unsupervised learning is constructed, which effectively improves the generalization ability of the model under the small sample condition, and significantly improves the monitoring sensitivity of the hot strip continuous rolling industrial process by combining the spatio-temporal feature joint extraction mechanism, providing a reliable process monitoring solution for the hot continuous rolling process.
[0167] Figure 5 It is a block diagram of a device for monitoring the hot strip continuous rolling industrial process shown according to an exemplary embodiment. This device is used for the method for monitoring the hot strip continuous rolling industrial process. Referring to Figure 5 , this device includes a first processing module 110, a data augmentation module 120, a random occlusion module 130, a model training module 140, a second processing module 150, and a data evaluation module 160. Among them:
[0168] The first processing module 110 is configured to collect multi-dimensional sensor data of the hot strip continuous rolling industrial process for standard cleaning;
[0169] The data enhancement module 120 is configured to perform data enhancement on the cleaned data using DDIM (Denoising Diffusion Implicit Model), and then introduce a sliding window to extract a fixed-length time window;
[0170] The random occlusion module 130 is configured to perform random occlusion on the data extracted by the sliding window by using a feature focusing cooperation mechanism based on random occlusion;
[0171] The model training module 140 is configured to construct an unsupervised process monitoring model by using the time and space information of the hot strip continuous rolling industrial process data, and train the constructed unsupervised process monitoring model by using the data after the occlusion operation;
[0172] The second processing module 150 is configured to process the data reconstructed by the trained unsupervised process monitoring model by using the kernel density estimation method to determine the control limit;
[0173] The data evaluation module 160 is configured to calculate the reconstruction error according to the multi-dimensional sensor data of the hot strip continuous rolling industrial process collected in real time, and compare the reconstruction error with the control limit to realize the detection of abnormal working conditions in the hot strip continuous rolling industrial process.
[0174] Figure 6 FIG. is a schematic structural diagram of a hot strip continuous rolling industrial process monitoring device provided by an embodiment of the present invention. As Figure 6 shown, the hot strip continuous rolling industrial process monitoring device may include the above-mentioned Figure 3 shown hot strip continuous rolling industrial process monitoring device. Optionally, the hot strip continuous rolling industrial process monitoring device 410 may include a first processor 2001.
[0175] Optionally, the hot strip continuous rolling industrial process monitoring device 410 may further include a memory 2002 and a transceiver 2003.
[0176] Wherein, the first processor 2001 is connected to the memory 2002 and the transceiver 2003, for example, through a communication bus.
[0177] Next, in combination with Figure 6 each component of the hot strip continuous rolling industrial process monitoring device 410 will be specifically introduced:
[0178] Among them, the first processor 2001 is the control center of the strip hot continuous rolling industrial process monitoring device 410, which can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or can be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention, such as: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).
[0179] Optionally, the first processor 2001 can execute various functions of the strip hot continuous rolling industrial process monitoring device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.
[0180] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 6 the CPU0 and CPU1 shown in
[0181] In a specific implementation, as an embodiment, the strip hot continuous rolling industrial process monitoring device 410 may also include multiple processors, such as Figure 6 the first processor 2001 and the second processor 2004 shown in. Each of these processors can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, the processor may refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions).
[0182] Among them, the memory 2002 is used to store software programs for executing the solution of the present invention and is controlled by the first processor 2001 for execution. The specific implementation manner can refer to the above method embodiments and will not be elaborated here.
[0183] Optionally, the memory 2002 can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 can be integrated with the first processor 2001 or can exist independently and is coupled to the first processor 2001 through an interface circuit ( Figure 6 not shown) of the hot strip continuous rolling industrial process monitoring device 410. The embodiments of the present invention do not make specific limitations on this.
[0184] The transceiver 2003 is used to communicate with a network device or with a terminal device.
[0185] Optionally, the transceiver 2003 can include a receiver and a transmitter ( Figure 6 not shown separately). Among them, the receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.
[0186] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or can exist independently and is coupled to the first processor 2001 through an interface circuit ( Figure 6 not shown) of the hot strip continuous rolling industrial process monitoring device 410. The embodiments of the present invention do not make specific limitations on this.
[0187] It should be noted that Figure 6 the structure of the hot strip continuous rolling industrial process monitoring device 410 shown does not constitute a limitation on the router. The actual knowledge structure recognition device can include more or fewer components than shown, or combine certain components, or have different component arrangements.
[0188] In addition, the technical effects of the hot strip continuous rolling industrial process monitoring device 410 can refer to the technical effects of the hot strip continuous rolling industrial process monitoring method described in the above method embodiments, and will not be elaborated here.
[0189] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.
[0190] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories.
[0191] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims
1. A method for monitoring the hot continuous rolling industrial process of strip steel, characterized in that, The method includes: Offline modeling stage: S1. Collect multi-dimensional sensor data of the strip hot continuous rolling industrial process and perform standardized cleaning; S2. After using DDIM to perform data augmentation on the cleaned data, introduce a sliding window to extract a time window of a fixed length; where DDIM represents a denoising diffusion implicit model; S3. Use a feature focusing cooperation mechanism based on random occlusion to randomly occlude the data extracted by the sliding window; S4. Construct an unsupervised process monitoring model, and use the occluded data to train the constructed unsupervised process monitoring model. During the training process, the unsupervised process monitoring model learns the temporal features and spatial features in the strip hot continuous rolling industrial process data to reconstruct the data; S5. Use the kernel density estimation method to process the data reconstructed by the trained unsupervised process monitoring model to determine the control limit; Online monitoring stage: S6. Calculate the reconstruction error based on the real-time collected multi-dimensional sensor data of the strip hot continuous rolling industrial process, and compare the reconstruction error with the control limit to achieve the detection of abnormal working conditions in the strip hot continuous rolling industrial process.
2. The strip hot continuous rolling industrial process monitoring method according to claim 1, characterized in that, The collection of multi-dimensional sensor data of the strip hot continuous rolling industrial process and performing standardized cleaning includes: Collect multi-dimensional sensor data of the entire process of the strip hot continuous rolling industrial process, and perform preliminary screening on the collected multi-dimensional sensor data to obtain data that has a key impact on the monitoring of the strip hot continuous rolling industrial process; Perform preprocessing on the screened data; where the preprocessing includes: normalization and missing value filling.
3. The strip hot continuous rolling industrial process monitoring method according to claim 1, wherein DDIM uses a U-net structure with a reduced number of convolutional layers for feature extraction. During the feature extraction process, the U-net structure uses a smaller convolutional kernel.
4. The strip hot continuous rolling industrial process monitoring method according to claim 1, characterized in that, The use of a feature focusing cooperation mechanism based on random occlusion to randomly occlude the data extracted by the sliding window includes: For the data X extracted by the sliding window window , the elements of some time steps are randomly replaced with 0 by using the feature focusing collaborative mechanism based on random occlusion, and the data X after the occlusion operation is obtained m : X m = X window ⊙M where M is the occlusion matrix, and M i,j is the element in the i-th row and j-th column of the matrix M; p ∈ (0, 1) represents the retention rate; ⊙ represents the Hadamard multiplication of matrices, that is, the elements in the same position are multiplied one by one.
5. The strip hot continuous rolling industrial process monitoring method according to claim 3, characterized in that The unsupervised process monitoring model uses an autoencoder-gated recurrent unit unsupervised learning framework; The autoencoder-gated recurrent unit unsupervised learning framework includes: an encoder layer, a fully connected layer connected to the encoder layer, and a decoder layer connected to the fully connected layer; The encoder layer uses a gated recurrent unit to extract temporal dynamic features; The fully connected layer is used to learn the joint distribution between data and capture spatial correlation.
6. The strip hot continuous rolling industrial process monitoring method according to claim 4, characterized in that, The use of the kernel density estimation method to process the data reconstructed by the trained unsupervised process monitoring model to determine the control limit includes: For the data reconstructed by the unsupervised process monitoring model, calculate the reconstruction error between it and the original data as the statistic e i : E = [e1, e2, …, e k where x i is the input sample of the unsupervised process monitoring model, is the data reconstructed by the unsupervised process monitoring model, e i is the reconstruction error of the i-th sample, k is the number of samples input to the unsupervised process monitoring model, and E is the reconstruction error matrix; Use KDE to estimate the probability density function f(e) to model the obtained reconstruction error, and estimate the probability density of each sample occurring. Where f(e) is expressed as: Among them, f(e) represents the kernel density estimation probability density function, that is, the probability density of each sample occurring; h is the bandwidth parameter; K(·) is the kernel function; e is an intermediate variable; According to the obtained f(e), calculate the cumulative distribution function F(e): Among them, F(e) represents the probability that the intermediate variable e is less than or equal to a certain value; Find the first F(e) that satisfies the following equation to determine the control limit C: F(e)≥τ C=min{x|F(e)≥τ} Among them, τ represents a pre-specified confidence level.
7. The strip hot continuous rolling industrial process monitoring method according to claim 4, characterized in that Calculating the reconstruction error based on the multi - dimensional sensor data collected in real - time during the strip hot - continuous rolling industrial process, and comparing the reconstruction error with the control limit to achieve the detection of abnormal working conditions in the strip hot - continuous rolling industrial process includes: Collecting multi - dimensional sensor data in real - time during the strip hot - continuous rolling industrial process, and performing standardized cleaning, sliding window extraction of a fixed - length time window, and random occlusion operation on the sensor data collected in real - time; For the data after the occlusion operation, using the trained unsupervised process monitoring model to reconstruct the data, calculating the reconstruction error, and comparing the reconstruction error with the control limit to achieve the detection of abnormal working conditions in the strip hot - continuous rolling industrial process, where the detection of abnormal working conditions is expressed as: Among them, e current represents the reconstruction error between the real-time data after the occlusion operation and its reconstructed data.
8. A strip hot continuous rolling industrial process monitoring device, wherein the strip hot continuous rolling industrial process monitoring device is used to implement the strip hot continuous rolling industrial process monitoring method according to any one of claims 1-7, and is characterized in that, The device includes: A first processing module, configured to collect multi - dimensional sensor data in the strip hot - continuous rolling industrial process for standardized cleaning; A data enhancement module, configured to perform data enhancement on the cleaned data using DDIM (Denoising Diffusion Implicit Model), and then introduce a sliding window to extract a fixed - length time window; A random occlusion module, configured to perform random occlusion on the data extracted by the sliding window using the feature - focusing collaborative mechanism based on random occlusion; A model training module, configured to construct an unsupervised process monitoring model using the temporal and spatial information of the strip hot - continuous rolling industrial process data, and train the constructed unsupervised process monitoring model using the data after the occlusion operation; A second processing module, configured to process the data reconstructed by the trained unsupervised process monitoring model using the kernel density estimation method to determine the control limit; A data evaluation module, configured to calculate the reconstruction error based on the multi - dimensional sensor data collected in real - time during the strip hot - continuous rolling industrial process, and compare the reconstruction error with the control limit to achieve the detection of abnormal working conditions in the strip hot - continuous rolling industrial process.
9. A strip hot continuous rolling industrial process monitoring device, characterized in that, The strip hot - continuous rolling industrial process monitoring device includes: A processor; A memory, on which computer - readable instructions are stored. When the computer - readable instructions are executed by the processor, the method described in any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, characterized in that, The computer - readable storage medium stores program code, and the program code can be called by the processor to execute the method described in any one of claims 1 to 7.
Citation Information
Cited By
Construction method and device of training data set, electronic equipment and program product
CN121614862A
