Sinter Feo content cross-domain identification method

Abstract

The application provides a sinter FeO content cross-domain recognition method, comprising the following steps: collecting sinter machine tail cross-section images under different working conditions, and dividing the image samples into source domain data sets and target domain data sets; converting the source domain and target domain images into Token sequences through a patch embedding module, and mixing the source domain Token and the target domain Token at the same spatial position in a linear interpolation manner to obtain an intermediate domain Token sequence; setting a local domain adaptation module in a feature extractor, performing fine-grained feature alignment on the source domain Token, the target domain Token and the intermediate domain Token, and adding disturbance to the Token sequence in the randomly selected Transformer Block hidden layer; adopting a game strategy, performing coarse-grained feature alignment through a domain mixing module, and iteratively optimizing the pseudo label by adopting a self-refinement strategy; inputting the unlabeled target domain sample into the trained model, and outputting the sinter FeO content category. The application can effectively solve the problems of inconsistent data distribution and label data scarcity commonly encountered in actual industrial process working condition recognition.

Description

Technical Field

[0001] This invention relates to the field of process monitoring technology in sintering, and more particularly to a cross-domain identification method for FeO content in sintered ore based on visual Transformer and unsupervised domain adaptation. Background Technology

[0002] The sintering process involves mixing various powdered iron-containing raw materials with appropriate amounts of fuel and flux, along with water. After mixing and pelletizing, the materials undergo a series of physicochemical changes on sintering equipment, agglomerating the mineral particles into sinter, thus forming sinter, the raw material for blast furnace ironmaking. Numerous indicators measure the quality of sinter, primarily including grade, basicity, and ferrous oxide (FeO) content. Among these, the ferrous oxide (FeO) content directly affects the reducibility and strength of the sinter, ultimately determining the quality level of blast furnace smelting and steel production. Therefore, accurate detection of the FeO content in sinter is crucial for ensuring high-quality steelmaking.

[0003] Currently, methods for detecting FeO content are mainly divided into three categories: chemical analysis, mechanistic modeling, and data-driven methods. Chemical analysis is the most accurate but time-consuming and has a large time lag. Mechanistic modeling predicts FeO content by establishing a mathematical model of the sintering process. However, the sintering process is extremely complex, making it difficult for the aforementioned two methods to obtain accurate sintering mechanism models, thus resulting in insufficient accuracy and practicality. Data-driven methods construct predictive models using sintering process variables, but traditional data-driven methods rely heavily on data quality and feature selection, leading to low model interpretability. Chinese patent CN104977316A discloses "a method for determining the trend of FeO content in sintered ore," which establishes a mapping relationship between FeO content and temperature difference ranges by relating the FeO content of sintered ore to the temperature difference between the flue gas temperature and the ambient cooling exhaust gas temperature. However, this method overly relies on data quality and ignores the influence of other process parameters.

[0004] In actual sintering processes, the FeO content and sintering state can be intuitively reflected by cross-sectional images of the sintering mill tail, without the need for manual selection of feature variables. Therefore, image data-driven FeO content detection methods have received increasing attention in recent years. Chinese patent CN105276988A discloses "A method for controlling FeO content in the cross-section of the sintering mill tail," which addresses the complex, multivariate, and time-dependent sintering process involving numerous physicochemical changes. It employs a dual-CCD image monitoring system to acquire cross-sectional images of the sintering mill tail, including visible light and infrared images. Finally, a fuzzy clustering system and a neural network system are used to identify the FeO content of the sinter. However, existing image data-based FeO content detection methods generally suffer from two drawbacks: inconsistent data distribution and reliance on a large amount of labeled data. To address these challenges, selecting a suitable processing method has become the focus of current research.

[0005] In recent years, transfer learning technology has developed rapidly and has been widely applied in various engineering fields. Unsupervised Domain Adaptation (UDA), as an important branch of transfer learning, is specifically applied to scenarios where there is no available labeled data in the target domain. Its core idea is to project labeled source domain data and unlabeled target domain data into a common feature space, making the projected features domain invariant, thereby reducing the distribution difference between the source and target domains. This technique not only avoids the time-consuming and labor-intensive collection of target domain labeled data, but also effectively addresses the problem of inconsistent data distribution.

[0006] It should be noted that existing data-driven methods in the literature all have two drawbacks: First, these methods typically assume that the training and testing data have the same distribution. However, in actual industrial scenarios, the sintering process is greatly affected by operating conditions, such as variations in sintering machine speed and average material layer thickness, leading to significant differences in the distribution of training and testing data. Second, most existing methods rely on large amounts of labeled data, making data collection costly and time-consuming. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a cross-domain identification method for FeO content in sintered ores. This invention achieves intelligent identification of ferrous oxide content during the sintering process by integrating a visual converter with unsupervised domain adaptation technology. First, the input sample is converted into a token sequence, and an intermediate domain is constructed using the source and target domains. Then, multi-level feature alignment is achieved at both fine-grained and coarse-grained levels. Finally, a classifier is used to obtain the category of ferrous oxide content in the target domain.

[0008] The technical means employed in this invention are as follows:

[0009] A method for cross-domain identification of FeO content in sintered ore, comprising:

[0010] S1. Collect cross-sectional images of the sintering machine tail under different working conditions, and divide the image samples into labeled source domain datasets and unlabeled target domain datasets.

[0011] S2. The source and target domain images are converted into token sequences by the patch embedding module, and the source domain tokens and target domain tokens are mixed at the same spatial location by linear interpolation to obtain the intermediate domain token sequence.

[0012] S3. Set up a local domain adaptation module in the feature extractor to perform fine-grained feature alignment on the source domain token, target domain token and intermediate domain token, and add perturbation to the token sequence in the randomly selected Transformer Block hidden layer to destroy the domain discriminative features.

[0013] S4. A game theory strategy is adopted to perform coarse-grained feature alignment through the domain mixing module, and a self-refinement strategy is used to iteratively optimize the pseudo-labels.

[0014] S5. Input the unlabeled target domain samples into the model trained by steps S1-S4, and output the FeO content category of sintered ore.

[0015] Furthermore, in step S1, the labeled source domain dataset is... , The unlabeled target domain dataset is , .

[0016] Furthermore, the patch embedding module is implemented using a visual Transformer architecture, wherein:

[0017] The image must be converted into a token sequence using patch embedding; the patch size and stride can be adjusted from 8×8 to 32×32 pixels; the ViT structure cannot be replaced by a convolutional neural network (CNN) because the self-attention mechanism of ViT is the basis for the subsequent local domain adaptation module.

[0018] Furthermore, the intermediate domain Token sequence is as follows:

[0019]

[0020] in, This represents two randomly selected pairs of samples. and Linear interpolation operation, , , These represent the first, second, and third samples of the intermediate domain, source domain, and target domain, respectively. k Each Token Derived from a learnable Beta distribution, it is a mixture of weights for sampled tokens at the same spatial location. k Linear interpolation is performed on the source and target domain tokens to preserve the topological structure of the token sequence, and updates are performed using gradient descent. This indicates the number of tokens in a single sample.

[0021] Furthermore, the loss function of the local domain adaptation module is:

[0022]

[0023] in, , , It is the cross-entropy loss function; yes P Feature extractor composed of TransformerBlocks This represents the output of the local discriminator; Indicates the first i The first sample k Each token, labeled s or t, represents the source domain or the target domain; It's a domain tag. Represents the source domain. Indicates the target domain.

[0024] Furthermore, in step S3, the perturbation of the token sequence in the randomly selected Transformer Block hidden layer is added as follows:

[0025]

[0026] in, Indicates sample x In the The token sequence in each Transformer Block Represents another sample Token sequence, It is a scalar quantity used to control the magnitude of the disturbance.

[0027] Further, in step S4, the domain blending module performs coarse-grained feature alignment, including:

[0028] Calculate the feature space cosine similarity between the intermediate domain and the source and target domains;

[0029] Calculate the label space similarity between the intermediate domain and the source domain;

[0030] The quality of pseudo-labels is optimized through a self-refinement module, and prediction confidence scores higher than a threshold are filtered out. The target domain samples are used to improve the model by using high-confidence prediction vectors as supervision signals.

[0031] Furthermore, in step S4, the loss function of the game strategy is as follows:

[0032]

[0033]

[0034]

[0035] in, These are the parameters for the feature extractor, classifier, and domain mixing module, respectively. It refers to the similarity between the tags in the intermediate domain and the source domain in the tag space. For a joint parameter set, These are pseudo-labels for the target domain data; and These represent the cosine similarity between the intermediate domain and the source and target domains in the feature space, respectively. It is a unit matrix used as the label similarity between the intermediate and target domains.

[0036] Furthermore, in step S4, the loss function of the self-refinement strategy is as follows:

[0037]

[0038] in, and They are and The predicted probability vector, It comes from the Bernoulli distribution. random variables, It is a confidence filter. It uses a predefined threshold to filter high-confidence samples. , Use credibility greater than The prediction vector improves the model.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. The present invention provides a cross-domain identification method for FeO content in sinter, which can effectively solve the problems of inconsistent data distribution and scarce label data that are common in the identification of actual industrial process conditions. Through visual converter and unsupervised domain adaptation technology, the ferrous oxide content of sinter can be accurately detected.

[0041] 2. The present invention provides a cross-domain identification method for FeO content in sintered ore, which can significantly reduce the dependence on target domain label data and reduce data collection costs, and is particularly suitable for scenarios where there is no available label data for the target domain.

[0042] 3. The present invention provides a cross-domain identification method for FeO content in sinter, which can effectively extract domain-invariant features, reduce domain differences, and realize cross-domain knowledge transfer. It is particularly suitable for identifying the ferrous oxide content of sinter when the data distributions of the source domain and the target domain are inconsistent, ensuring the performance of the model under different working conditions, thereby effectively improving production efficiency and product quality.

[0043] Based on the above reasons, this invention can be widely applied in fields such as process monitoring during sintering. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the method of the present invention.

[0046] Figure 2 This is a flowchart of the process structure of the present invention.

[0047] Figure 3 This is a schematic diagram of the intermediate domain generation mechanism of the present invention.

[0048] Figure 4 This is a confusion matrix diagram provided for an embodiment of the present invention.

[0049] Figure 5 A t-SNE visualization diagram provided for an embodiment of the present invention.

[0050] Figure 6 Attention map of sintering machine tail image of category 0 provided in an embodiment of the present invention.

[0051] Figure 7 Attention map of sintering machine tail image of category 1 provided in the embodiments of the present invention.

[0052] Figure 8 Attention map of sintering machine tail image of category 2 provided in the embodiments of the present invention. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0055] like Figure 1 As shown, this invention provides a method for cross-domain identification of FeO content in sintered ore, comprising:

[0056] S1. Collect cross-sectional images of the sintering machine tail under different working conditions, and divide the image samples into labeled source domain datasets and unlabeled target domain datasets.

[0057] S2. The source and target domain images are converted into token sequences by the patch embedding module, and the source domain tokens and target domain tokens are mixed at the same spatial location by linear interpolation to obtain the intermediate domain token sequence.

[0058] S3. Set up a local domain adaptation module in the feature extractor to perform fine-grained feature alignment on the source domain token, target domain token and intermediate domain token, and add perturbation to the token sequence in the randomly selected Transformer Block hidden layer to destroy the domain discriminative features.

[0059] S4. A game theory strategy is adopted to perform coarse-grained feature alignment through the domain mixing module, and a self-refinement strategy is used to iteratively optimize the pseudo-labels.

[0060] S5. Input the unlabeled target domain samples into the model trained in steps S1-S4, and output the FeO content category of the sinter. In this embodiment, the batch size is 8, the optimizer is SGD, the learning rate is 0.03, and Dropout is 0.1.

[0061] In a specific implementation, as a preferred embodiment of the present invention, in step S1, the labeled source domain dataset is... , The unlabeled target domain dataset is , .

[0062] In this embodiment, the image data of both the source and target domains contain three categories, with their FeO contents falling within the range of... , and The corresponding labels are 0, 1, and 2, and a total of 1500 cross-sectional images of actual industrial sintering machine tail sections were collected for each category. .

[0063] In a specific implementation, as a preferred embodiment of the present invention, the patch embedding module is implemented using a visual Transformer structure, wherein:

[0064] The image must be converted into a token sequence using patch embedding; the patch size and stride can be adjusted from 8×8 to 32×32 pixels; the ViT structure cannot be replaced by a convolutional neural network (CNN) because the self-attention mechanism of ViT is the basis for the subsequent local domain adaptation module.

[0065] In this embodiment, the patch embedding uses a two-dimensional convolution with 3 input channels and 384 output channels. The kernel size and stride are both 16. .

[0066] In a specific implementation, as a preferred embodiment of the present invention, the intermediate domain Token sequence in step S2 is as follows:

[0067]

[0068] in, This represents two randomly selected pairs of samples. and Linear interpolation operation, , , These represent the first, second, and third samples of the intermediate domain, source domain, and target domain, respectively. k Each Token Derived from a learnable Beta distribution, it is a mixture of weights for sampled tokens at the same spatial location. k Linear interpolation is performed on the source and target domain tokens to preserve the topological structure of the token sequence, and updates are performed using gradient descent. This indicates the number of tokens in a single sample. For example... Figure 3 The diagram shown is a schematic of the intermediate domain generation mechanism.

[0069] In a specific implementation, as a preferred embodiment of the present invention, in step S3, the loss function of the local domain adaptation module is:

[0070]

[0071] in, , , It is the cross-entropy loss function; yes P Feature extractor composed of TransformerBlocks This represents the output of the local discriminator; Indicates the first i The first sample k Each token, labeled s or t, represents the source domain or the target domain; It's a domain tag. Represents the source domain. Indicates the target domain.

[0072] In a specific implementation, as a preferred embodiment of the present invention, the perturbation of the token sequence in the randomly selected Transformer Block hidden layer is added as follows:

[0073]

[0074] in, Indicates sample x In the The token sequence in each Transformer Block Represents another sample Token sequence, It is a scalar quantity used to control the magnitude of the disturbance.

[0075] In this embodiment, each Transformer Block's attention mechanism has 6 heads, the hidden layer dimension is 384, and the local domain discriminator module consists of three fully connected layers with output dimensions of 384, 384, and 1 respectively. Perturbations are randomly added to a specific set of hidden layers. P=12 This embodiment also provides a comparative analysis table between the perturbation mechanism of the present invention and existing methods, as follows:

[0076] Comparative analysis table of perturbation mechanisms and existing methods

[0077]

[0078] In a specific implementation, as a preferred embodiment of the present invention, in step S4, the domain mixing module performs coarse-grained feature alignment, including:

[0079] Calculate the feature space cosine similarity between the intermediate domain and the source and target domains;

[0080] Calculate the label space similarity between the intermediate domain and the source domain;

[0081] The quality of pseudo-labels is optimized through a self-refinement module, and prediction confidence scores higher than a threshold are filtered out. The target domain samples are used to improve the model by using high-confidence prediction vectors as supervision signals.

[0082] In a specific implementation, as a preferred embodiment of the present invention, the loss function of the game strategy in step S4 is as follows:

[0083]

[0084]

[0085]

[0086] in, These are the parameters for the feature extractor, classifier, and domain mixing module, respectively. It refers to the similarity between the tags in the intermediate domain and the source domain in the tag space. For a joint parameter set, These are pseudo-labels for the target domain data; and These represent the cosine similarity between the intermediate domain and the source and target domains in the feature space, respectively. It is a unit matrix used as the label similarity between the intermediate and target domains.

[0087] In a specific implementation, as a preferred embodiment of the present invention, the loss function of the self-refinement strategy in step S4 is as follows:

[0088]

[0089] in, and They are and The predicted probability vector, It comes from the Bernoulli distribution. random variables, It is a confidence filter. It uses a predefined threshold to filter high-confidence samples. , Use credibility greater than The predicted vectors are used to refine the model. A self-refinement strategy is employed to further improve the model's accuracy in the target domain.

[0090] In this embodiment, the classifier includes two fully connected layers with input and output dimensions of 384→1536 and 1536→384, respectively. GELU activation is used and Dropout is added to prevent overfitting. The experimental results are as follows: Figures 4 to 8 As shown, where, Figure 4 This is a confusion matrix. The numbers in this graph represent the percentage of predicted labels and true labels that are equal. For example, 0.86 represents the percentage where both the true label and the predicted label are 0, and 0.04 represents the percentage where the true label is 0 but the predicted label is 1. Figure 5 This is a visualization result of t-SNE. Principal component 1 and principal component 2 are two dimensions when high-dimensional features are mapped to two-dimensional space, and they have no specific physical meaning. Figures 6 to 8 These are attention visualization results for three different types of labeled data.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of sinter FeO content cross-domain identification, characterized in that, include: S1. Collect cross-sectional images of the sintering machine tail under different working conditions, and divide the image samples into labeled source domain datasets and unlabeled target domain datasets. S2. The source and target domain images are converted into token sequences by the patch embedding module, and the source and target domain tokens are mixed at the same spatial location by linear interpolation to obtain the intermediate domain token sequence. S3. Set up a local domain adaptation module within the feature extractor to perform fine-grained feature alignment on the source domain token, target domain token, and intermediate domain token, and add perturbation to the token sequence in a randomly selected Transformer Block hidden layer; fine-grained feature alignment is achieved through the following loss function: wherein, , , is a cross-entropy loss function; is P a feature extractor consisting of denotes the output of the local domain discriminator; denotes the i th Token of the k th sample, with superscript s or t denoting the source domain or the target domain; is a domain label, denotes the source domain, denotes the target domain; S4. A game-theoretic strategy is adopted, using the domain fusion module to perform coarse-grained feature alignment, and a self-refining strategy to iteratively optimize the pseudo-labels; the loss function of the game-theoretic strategy is as follows: wherein, are parameters of the feature extractor, the classifier and the domain mixing module, respectively, is the label similarity between the intermediate domain and the source domain in the label space, is the joint parameter set, is the pseudo label of the target domain data; and are the cosine similarities between the intermediate domain and the source domain, and between the target domain, respectively, in the feature space; is a unit matrix as the label similarity between the intermediate domain and the target domain. Coarse-grained feature alignment is performed through the domain blending module, including: Calculate the feature space cosine similarity between the intermediate domain and the source and target domains; Calculate the label space similarity between the intermediate domain and the source domain; The quality of pseudo-labels is optimized through a self-refinement module, and prediction confidence scores higher than a threshold are filtered out. The target domain samples are used as supervision signals to improve the model, and the high-confidence prediction vectors are used as supervision signals. S5. Input the unlabeled target domain samples into the model trained by steps S1-S4, and output the FeO content category of sintered ore.

2. The method for cross-domain identification of FeO content in sintered ore according to claim 1, characterized in that, In step S1, the labeled source domain dataset is , The unlabeled target domain dataset is , .

3. The method for cross-domain identification of FeO content in sintered ore according to claim 1, characterized in that, The patch embedding module is implemented using a visual Transformer architecture, wherein: The image must be converted into a token sequence using patch embedding; the patch size and stride can be adjusted from 8×8 to 32×32 pixels; the ViT structure cannot be replaced by a convolutional neural network.

4. The method for cross-domain identification of FeO content in sintered ore according to claim 1, characterized in that, In step S2, the intermediate domain Token sequence is as follows: in, This represents two randomly selected pairs of samples. and Linear interpolation operation, , , These represent the first, second, and third samples of the intermediate domain, source domain, and target domain, respectively. k Each Token Derived from a learnable Beta distribution, it is a mixture of weights for sampled tokens at the same spatial location. k Linear interpolation is performed on the source and target domain tokens to preserve the topological structure of the token sequence, and updates are performed using gradient descent. This indicates the number of tokens in a single sample.

5. The method for cross-domain identification of FeO content in sintered ore according to claim 1, characterized in that, In step S3, the perturbation of the token sequence in the randomly selected Transformer Block hidden layer is added as follows: in, Indicates sample x In the The token sequence in each Transformer Block Represents another sample Token sequence, It is a scalar quantity used to control the magnitude of the disturbance.

6. The method for cross-domain identification of FeO content in sintered ore according to claim 1, characterized in that, In step S4, the loss function of the self-refinement strategy is as follows: in, and They are and The predicted probability vector, It comes from the Bernoulli distribution. random variables, It is a confidence filter. It uses a predefined threshold to filter high-confidence samples. , Use credibility greater than The prediction vector improves the model.

Images