Soft measurement method for ferrous oxide content of sintered ore and related device

Abstract

The invention discloses a soft measurement method for the content of ferrous oxide in sintered ore and a related device. The method comprises the following steps: acquiring time sequence data and image data of different sinter samples under different sintering stage conditions, constructing a labeled data set and a non-labeled data set, dividing the labeled data set into a training set, a verification set and a test set according to a preset proportion, and adding the non-labeled data set as extension into the training set; performing iterative training and evaluation on the MTFIF-STMM model based on the training set, the verification set and the test set to obtain a final MTFIF-STMM model; and inputting multi-source data, collected in real time, of the sinter into the final MTFIF-STMM model for soft measurement to obtain a real-time evaluation result of the ferrous oxide content of the sinter. According to the method, real-time soft measurement of the ferrous oxide content in the ore in the sintering process is realized, and the modeling precision is improved.

Description

Technical Field

[0001] This invention relates to the field of soft measurement technology, and in particular to a soft measurement method and related apparatus for the ferrous oxide content of sintered ore. Background Technology

[0002] The sintering process, as the iron ore pretreatment unit in the blast furnace ironmaking process, provides the main raw materials for the blast furnace and is a core link in the steel production chain. Its sintering quality directly affects iron ore reduction efficiency, fuel consumption, and molten iron composition. Low-quality sinter leads to increased return ore rates, reduced production, increased energy consumption, and accelerated equipment wear. Therefore, real-time measurement and effective estimation of sinter quality indicators are crucial for optimizing sintering process parameters and implementing control decisions. This is of great significance and value for stabilizing blast furnace operation, improving production efficiency, reducing raw material and energy consumption, and enhancing product quality and economic benefits.

[0003] However, the various important quality indicators of sinter, including chemical composition, physical properties and metallurgical properties, are diverse and complex. Traditional sinter quality testing relies on a series of manual physicochemical experiments, which is slow, has a strong lag, cannot meet the timeliness requirements of process strategy adjustment, and is costly.

[0004] In recent years, data-driven soft measurement models have been widely applied to complex problems. Researchers have achieved satisfactory results in various industrial soft measurement fields through mechanistic mathematical models, machine learning, and deep learning methods. However, high-performance deep learning models heavily rely on the quality, richness, distribution consistency, and correlation of training and testing data. To ensure the reliability of sinter quality evaluation, it is necessary to consider multiple measurement data, even multi-source heterogeneous data, simultaneously. Failure to properly handle the relationships between various and heterogeneous data can lead to poor model performance or even model failure. Furthermore, the high cost and long lead time of creating labels for sinter quality indicators result in the underutilization of a large amount of unlabeled, multi-source data.

[0005] Therefore, how to effectively integrate various types of data and heterogeneous data from multiple sources, and make full use of the information contained in rich unlabeled data, is crucial for refined modeling and improving the soft measurement performance of models. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes a soft measurement method and related device for ferrous oxide content in sintered ore. This method integrates complementary and multi-dimensional information from multi-source heterogeneous data during the sintering process, enabling the full exploration of the relationship between different dimensional data and sintered ore quality indicators based on multimodal fusion, and extending the learning of unlabeled time-series feature similarity.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A soft measurement method for ferrous oxide content in sintered ore includes the following steps:

[0009] Time series and image data of different sintered ore samples under different sintering stage conditions were obtained, and labeled and unlabeled datasets were constructed. The labeled dataset was divided into training set, validation set and test set according to a preset ratio, and the unlabeled dataset was added to the training set as an extension.

[0010] The training set is input into the MTFIF-STMM model for iterative training. For unlabeled data, the model is optimized by minimizing the similarity loss of the two-branch time series features. For labeled data, the model is optimized by combining the mean squared error loss and the similarity loss of the two-branch sequence features into an overall loss function. The model parameters are updated using the backpropagation mechanism. The optimal MTFIF-STMM model is selected using the validation set. The performance of the optimal MTFIF-STMM model is evaluated using the test set until the performance evaluation meets the set threshold, at which point the final MTFIF-STMM model is obtained.

[0011] The real-time collected multi-source data of sinter is input into the final MTFIF-STMM model for soft measurement to obtain the real-time assessment results of the ferrous oxide content of sinter.

[0012] Preferably, the MTFIF-STMM model includes a multi-source data feature encoder, a feature interactor, and a decoder and predictor, wherein,

[0013] In the multi-source data feature encoder, a Transformer encoder is used to encode time-series data to obtain features. The MobileViT and Semi-MobileViT modules are used to extract and encode features from image data to obtain features. ;

[0014] In the feature interactor, fully connected layers are used to interact with the features respectively. and characteristics After downsampling, cosine similarity is used for feature interaction to obtain features. and characteristics The features and characteristics Each feature is reconstructed using a fully connected layer to restore it to its original dimensions, thus obtaining the features. and characteristics The features and characteristics Separate and characteristic and characteristics Residual connections yield features and characteristics ; the features and characteristics Feature fusion is performed to obtain multimodal fused features;

[0015] In the decoder and predictor, the multimodal fusion features are decoded and predicted to obtain the prediction result.

[0016] Preferably, the step involves using the MobileViT module and the Semi-MobileViT module to extract and encode features from the image data to obtain features.

[0017] Initial features are obtained by extracting features from image data using convolutional layers.

[0018] Multi-level local feature extraction is performed on the initial features using MobileNetV2-style inverted residual blocks to obtain local features. These local features are then expanded into a sequence using the MobileViT module, input into a lightweight Transformer for global relational modeling, and finally folded back into spatial features to obtain the final features. , , and These represent the feature map height, width, and number of channels input to the Semi-MobileViT module, respectively.

[0019] The Semi-MobileViT module follows the MobileViT module's design from local representation to global representation, and incorporates the features... Transform into , , and These represent the number of pixels, the number of patches, and the number of channels within each patch, respectively; a compression-excitation module is used to... Transform into .

[0020] Preferably, the decoder and predictor comprises several stacked decoding layers with identical structures, wherein the decoding layer comprises two multi-head attention sublayers and a point-by-point feedforward network sublayer.

[0021] Preferably, mean square error loss Defined as:

[0022]

[0023] In the formula, Let i be the label value of sample i. Let be the predicted value for sample i, and n be the number of samples.

[0024] Preferably, the similarity loss of the temporal features of the two branches Defined as:

[0025]

[0026] In the formula, the cosine similarity between the temporal features of the two branches is expressed as: Entropy regularization constraint is expressed as ,in, The regularization coefficient is used. This represents the probability distribution of cosine similarity. for The Each element.

[0027] The overall loss function is:

[0028]

[0029] In the formula, for The weights are used to adjust and The ratio between them.

[0030] Based on the above, the present invention also discloses a soft measurement system for ferrous oxide content in sintered ore, comprising:

[0031] The acquisition module is used to acquire time series data and image data of different sintered ore samples under different sintering stage conditions, construct labeled datasets and unlabeled datasets, divide the labeled datasets into training set, validation set and test set according to a preset ratio, and add the unlabeled datasets as an extension to the training set.

[0032] The training module is used to input the training set into the MTFIF-STMM model for iterative training. For unlabeled data, the model is optimized by minimizing the similarity loss of the two-branch temporal features. For labeled data, the model is optimized by a combination of mean squared error loss and similarity loss of the two-branch sequence features. The model parameters are updated using a backpropagation mechanism. The optimal MTFIF-STMM model is selected using the validation set. The performance of the optimal MTFIF-STMM model is evaluated using the test set until the performance evaluation meets the set threshold, at which point the final MTFIF-STMM model is obtained.

[0033] The prediction module is used to input the real-time collected multi-source data of sinter into the final MTFIF-STMM model for soft measurement to obtain the real-time assessment results of the ferrous oxide content of sinter.

[0034] Based on the foregoing, the present invention also discloses a computer device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement any of the methods described above.

[0035] Based on the above, the present invention also discloses a readable storage medium storing a computer program, which, when executed by a processor, implements any of the methods described above.

[0036] Based on the above technical solution, the beneficial effects of the present invention are as follows: The present invention provides a soft measurement method and related apparatus for ferrous oxide content in sintered ore. This method acquires time-series data and image data of different sintered ore samples under different sintering stage conditions, constructs labeled and unlabeled datasets, divides the labeled dataset into training, validation, and test sets according to a preset ratio, and adds the unlabeled dataset as an extension to the training set; inputs the training set into the MTFIF-STMM model for iterative training, optimizes the model by minimizing the similarity loss of the two-branch time-series features for unlabeled data, and optimizes the model by using an overall loss function combining mean squared error loss and the similarity loss of the two-branch sequence features for labeled data, updates the model parameters using a backpropagation mechanism, selects the optimal MTFIF-STMM model using the validation set, and evaluates the performance of the optimal MTFIF-STMM model using the test set until the performance evaluation meets a set threshold, thus obtaining the final MTFIF-STMM model; inputs real-time collected multi-source data of sintered ore into the final MTFIF-STMM model for soft measurement to obtain real-time evaluation results of the ferrous oxide content in the sintered ore. This invention can learn the complex feature mapping relationship between ferrous oxide content in sinter and various types and modalities by using time-series data, image data, and physicochemical performance indicators of samples corresponding to some data. Simultaneously, it mines the temporal feature similarity between time-series and image branches, effectively utilizing rich unlabeled data information to further improve the soft measurement modeling accuracy of supervised models. Under the conditions of high label cost, strong lag, and scarcity, it achieves real-time soft measurement of ferrous oxide content in sinter and improves modeling accuracy. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the structure of the MTFIF-STMM model in one embodiment;

[0038] Figure 2 This is a schematic flowchart of a soft measurement method for ferrous oxide content in sintered ore in one embodiment.

[0039] Figure 3 A schematic diagram of the MV2 module structure in image feature encoding in one embodiment.

[0040] Figure 4 A schematic diagram of the MobileViT module structure in image feature encoding in one embodiment.

[0041] Figure 5 A schematic diagram of the Semi-MobileViT module structure in image feature encoding in one embodiment.

[0042] Figure 6 This is a schematic diagram of the Transformer encoder and decoder structure of the MTFIF-STMM model in one embodiment. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0044] like Figure 1 , 2 As shown, this embodiment provides a soft measurement method for the ferrous oxide content of sintered ore, including the following steps:

[0045] Step 100: Obtain time series data and image data of different sintered ore samples under different sintering stage conditions, construct labeled datasets and unlabeled datasets, divide the labeled datasets into training set, validation set and test set according to a preset ratio, and add the unlabeled datasets as an extension to the training set.

[0046] In this embodiment, under different sintering stages and conditions, various sensor data (time series data) and image data reflecting the quality of sintered ore, such as temperature and pressure, are collected from different sintered ore samples. Physicochemical and other performance experiments are conducted on selected samples, and the combined experimental results serve as labels for the corresponding data. Labeled data is divided into training and validation sets proportionally, and unlabeled data is added to the training set as an extension.

[0047] Step 200: Input the training set into the MTFIF-STMM model for iterative training. For unlabeled data, optimize the model by minimizing the similarity loss of the two-branch time series features. For labeled data, optimize the model by combining the mean squared error loss and the similarity loss of the two-branch sequence features into an overall loss function. Update the model parameters using the backpropagation mechanism, select the optimal MTFIF-STMM model using the validation set, and evaluate the performance of the optimal MTFIF-STMM model using the test set until the performance evaluation meets the set threshold, thus obtaining the final MTFIF-STMM model.

[0048] In this embodiment, a semi-supervised Transformer model (MTFIF-STMM) for multi-source temporal feature interaction and fusion is constructed. This MTFIF-STMM model comprises three parts: a multi-source data feature encoder, a feature interactor, and a decoder and predictor. (See [link to relevant documentation]). Figure 2 .

[0049] The multi-source data feature encoder processes data as follows:

[0050] 1) The image feature encoding part uses MobileViT information coding blocks for encoding. It extracts the initial features of the image through convolutional layers, and then uses the MobileNetV2 style (MV2 module, see MobileNetV2). Figure 3 The inverted residual block enhances inter-channel information interaction and gradient flow, efficiently capturing local features while reducing computation. Then, the MobileViT module expands the local features into a sequence input lightweight Transformer for global relationship modeling before folding them back into spatial features. This preserves the local inductive ability of CNNs while introducing the global interactive capabilities of Transformers. (See [link to relevant documentation]). Figure 4 Finally, the Semi-MobileViT module, which follows the MobileViT module's design from local representation to global representation, processes the input image... Transform into , , and These represent the feature map height, width, and number of channels input to the Semi-MobileViT module, respectively. , and These represent the number of pixels in each patch, the number of patches (i.e., the length of the image keyframe sequence), and the number of channels, respectively. Subsequently, a compression-excitation (SE) module is used to... Transformed into encoded image features See Figure 5 .

[0051] 2) The time series feature encoding part processes the input time series samples through the Transformer encoder to achieve global information interaction and feature extraction.

[0052] See Figure 6 This Transformer encoder, for lengths of Dimensions Input time series features The coding implementation steps are as follows:

[0053] First, time series features The input features of the encoding layer are obtained after embedding and positional encoding. ,Right now:

[0054]

[0055] in, For embedding matrix, It is a positional encoding matrix, implemented using a periodic function, and then... It goes through N coding layers in sequence.

[0056] Each coding layer consists of a multi-head attention (MHA) sublayer and a pointwise feedforward (FFN) sublayer:

[0057]

[0058]

[0059] in, For inter-layer coding features, The output features of the coding layer Representation layer normalization; the multi-head attention layer can be represented as:

[0060]

[0061] In the formula, This is a dimensional splicing operation, where h is the number of attention heads. The single attention head is represented and its calculation method is as follows:

[0062]

[0063] Among them, query ,key ,value From the input matrix respectively Obtained through the embedding layer; , , , They are respectively for , , and The projection matrix of each parallel attention head; where... ; The attention function is represented as:

[0064]

[0065] The FFN layer consists of two fully connected layers and an intermediate ReLU activation function. For the input inter-layer encoded features... The FFN layer can be represented as:

[0066]

[0067] in, , , , These are learnable parameters in linear transformations.

[0068] Finally, after completing the stacking operation of N encoding layers, the output of the last encoding layer is taken as the time series encoded feature. .

[0069] Since both types of data are acquired from the same sintering process and have the same sampling frequency, the length of the image keyframe sequence is equal to the length of the time series, i.e. .

[0070] Based on the consistency of time stamps in multi-source data, image features and time series features are aligned by aligning timestamps.

[0071] Feature interactors process as follows:

[0072] Time series data and image data are obtained after feature encoding. and After passing through the feature interactor, the feature dimensionality is first reduced using a fully connected layer (FC) to obtain... and To reduce computational cost, the cosine similarity between the temporal features of the image branch and the time series branch is then calculated to measure the similarity of the temporal features of the two branches and to perform feature interaction. and After passing through the FC layer, the original dimensions are restored to obtain... and ;final, and Respectively with original features and The residual connection yields two interactive features. and Where L represents the length of the time series, The dimension represents the time series features, and P represents the number of patches (i.e., the length of the image keyframe sequence). 'r' represents the number of channels, and 'r' represents the dimensional scaling ratio.

[0073] The decoder and predictor process is as follows:

[0074] After completing the interaction of time series features, time series branch features and image branch features Fusion features can be obtained by splicing. Subsequently Convolutional layers adjust the feature dimensions before model decoding and prediction. Similar to the encoder structure, the decoder consists of multiple identical decoding layers stacked together. Each decoding layer introduces an additional cross-MHA sublayer on top of the encoder layer structure to receive fused features. In the decoder This is the input to the decoder, and its sequence length is... , dimension , = ;

[0075] The initial input to the decoding layer in the decoder is , can be represented as:

[0076]

[0077]

[0078] in, and They are respectively Embedding matrix and position encoding matrix; express At time step to Historical data was used, and initial values ​​were adopted at the start of training. It is a zero matrix. and After splicing, the result is ; This indicates the prediction time step.

[0079] The decoder output is mapped and predicted by a fully connected layer to produce the corresponding soft measurement results.

[0080] The objective function of the MTFIF-STMM model consists of two parts: one is the regression loss represented by the mean squared error, and the other is the similarity loss of the time series features of the two branches.

[0081] The regression loss of soft measurement is the mean square error between the predicted and labeled values ​​of all labeled data. Defined as:

[0082]

[0083] In the formula, Let i be the label value of sample i. Let be the predicted value for sample i, and n be the number of samples.

[0084] Similarity loss of temporal features of two branches The situation that reflects consistent changes over time is defined as:

[0085]

[0086] In the formula, Let c represent the cosine similarity of the i-th feature in the two-branch time series, where If two feature branches are too similar in the time dimension, it will weaken the multi-source nature of the sintering process information. Therefore, an entropy regularization constraint term is introduced. In order to avoid extreme situations, ,in, This is the regularization coefficient, which controls the strength of regularization. This represents the probability distribution of cosine similarity. for The Each element.

[0087] The overall loss function of the soft measurement model constructed based on the above two parts of loss is:

[0088]

[0089] In the formula, for The weights are used to adjust and The ratio between them. During model training, if data labels exist, then... These are standard hyperparameters; if no data labels exist, then... Set to 0.

[0090] The MTFIF-STMM model, as a semi-supervised soft measurement model, uses training data containing a small portion of labeled data and a large portion of unlabeled data. The model automatically selects the appropriate loss strategy based on the labeling status of the current training data, fully utilizing both labeled and unlabeled information to complete the semi-supervised soft measurement. For unlabeled data, it minimizes the similarity loss between the temporal features of the two branches and introduces an entropy regularization term to ensure that the temporal features of the two branches retain some differences while narrowing their differences, which better reflects the situation where the temporal features of time-series data and image data are correlated but not identical during the sintering process. For labeled data, supervised training is used, employing the mean squared error loss composed of the decoder's predicted values ​​and the label values. By fully utilizing labeled data to supervise model performance, and combining it with the similarity loss of two-branch sequence features, the model can extract additional unlabeled data information based on the similarity of time-series features, thereby improving the model's ability to learn from unlabeled data.

[0091] By dynamically achieving this optimization goal, the model learns expert knowledge while utilizing rich sensor and image data to enhance the mapping relationship between two features and sintering quality indicators under unlabeled conditions. This enables a high-performance soft measurement target that achieves multimodal feature interaction and fusion, suitable for real-time assessment of ferrous oxide content in sintered ore.

[0092] Step 300: Input the real-time collected multi-source data of sinter into the final MTFIF-STMM model to obtain the real-time evaluation results of the ferrous oxide content of the sinter.

[0093] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0094] Based on the same inventive concept, this application also provides a system for implementing the soft measurement method for ferrous oxide content in sintered ore as described above. The solution provided by this system is similar to the solution described in the above method, and therefore will not be repeated here.

[0095] In one embodiment, a soft measurement system for ferrous oxide content in sintered ore is also provided, comprising:

[0096] The acquisition module is used to acquire time series data and image data of different sintered ore samples under different sintering stage conditions, construct labeled datasets and unlabeled datasets, divide the labeled datasets into training set, validation set and test set according to a preset ratio, and add the unlabeled datasets as an extension to the training set.

[0097] The training module is used to input the training set into the MTFIF-STMM model for iterative training. For unlabeled data, the model is optimized by minimizing the similarity loss of the two-branch temporal features. For labeled data, the model is optimized by a combination of mean squared error loss and similarity loss of the two-branch sequence features. The model parameters are updated using a backpropagation mechanism. The optimal MTFIF-STMM model is selected using the validation set. The performance of the optimal MTFIF-STMM model is evaluated using the test set until the performance evaluation meets the set threshold, at which point the final MTFIF-STMM model is obtained.

[0098] The prediction module is used to input the real-time collected multi-source data of sinter into the final MTFIF-STMM model for soft measurement to obtain the real-time assessment results of the ferrous oxide content of sinter.

[0099] In the above embodiments, each module of the soft measurement system for ferrous oxide content in sintered ore can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0100] In one embodiment, a computer device is also provided, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps as described in all the above method embodiments.

[0101] In one embodiment, a readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps as described in all the above method embodiments.

[0102] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0103] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. A soft measurement method for ferrous oxide content in sintered ore, characterized in that, Includes the following steps: Time series and image data of different sintered ore samples under different sintering stage conditions were obtained, and labeled and unlabeled datasets were constructed. The labeled dataset was divided into training set, validation set and test set according to a preset ratio, and the unlabeled dataset was added to the training set as an extension. The training set is input into the MTFIF-STMM model for iterative training. For unlabeled data, the model is optimized by minimizing the similarity loss of the two-branch time series features. For labeled data, the model is optimized by combining the mean squared error loss and the similarity loss of the two-branch sequence features into an overall loss function. The model parameters are updated using the backpropagation mechanism. The optimal MTFIF-STMM model is selected using the validation set. The performance of the optimal MTFIF-STMM model is evaluated using the test set until the performance evaluation meets the set threshold, at which point the final MTFIF-STMM model is obtained. The real-time collected multi-source data of sinter is input into the final MTFIF-STMM model for soft measurement to obtain the real-time assessment results of the ferrous oxide content of sinter.

2. The soft measurement method for ferrous oxide content in sintered ore according to claim 1, characterized in that, The MTFIF-STMM model includes a multi-source data feature encoder, a feature interactor, and a decoder and predictor, wherein... In the multi-source data feature encoder, a Transformer encoder is used to encode time-series data to obtain features. The MobileViT and Semi-MobileViT modules are used to extract and encode features from image data to obtain features. ; In the feature interactor, fully connected layers are used to interact with the features respectively. and characteristics After downsampling, cosine similarity is used for feature interaction to obtain features. and characteristics The features and characteristics Each feature is reconstructed using a fully connected layer to restore it to its original dimensions, thus obtaining the features. and characteristics The features and characteristics Separate and characteristic and characteristics Residual connections yield features and characteristics ; the features and characteristics Feature fusion is performed to obtain multimodal fused features; In the decoder and predictor, the multimodal fusion features are decoded and predicted to obtain the prediction result.

3. The soft measurement method for ferrous oxide content in sintered ore according to claim 2, characterized in that, The method utilizes the MobileViT module and the Semi-MobileViT module to extract and encode features from image data, obtaining features. Initial features are obtained by extracting features from image data using convolutional layers. Multi-level local feature extraction is performed on the initial features using MobileNetV2-style inverted residual blocks to obtain local features. These local features are then expanded into a sequence using the MobileViT module, input into a lightweight Transformer for global relational modeling, and finally folded back into spatial features to obtain the final features. , , and These represent the feature map height, width, and number of channels input to the Semi-MobileViT module, respectively. The Semi-MobileViT module follows the MobileViT module's design from local representation to global representation, and incorporates the features... Transform into , , and These represent the number of pixels, the number of patches, and the number of channels within each patch, respectively; a compression-excitation module is used to... Transform into .

4. The soft measurement method for ferrous oxide content in sintered ore according to claim 2, characterized in that, The decoder and predictor comprises several stacked, identical decoding layers, each including two multi-head attention sublayers and a point-by-point feedforward network sublayer.

5. The soft measurement method for ferrous oxide content in sintered ore according to claim 2, characterized in that, Mean square error loss Defined as: In the formula, Let i be the label value of sample i. Let be the predicted value for sample i, and n be the number of samples.

6. The soft measurement method for ferrous oxide content in sintered ore according to claim 5, characterized in that, Similarity loss of temporal features of two branches Defined as: In the formula, the cosine similarity between the temporal features of the two branches is expressed as: Entropy regularization constraint is expressed as ,in, The regularization coefficient is used. This represents the probability distribution of cosine similarity. for The Each element. The overall loss function is: In the formula, for The weights are used to adjust and The ratio between them.

7. A soft measurement system for ferrous oxide content in sintered ore, characterized in that, include: The acquisition module is used to acquire time series data and image data of different sintered ore samples under different sintering stage conditions, construct labeled datasets and unlabeled datasets, divide the labeled datasets into training set, validation set and test set according to a preset ratio, and add the unlabeled datasets as an extension to the training set. The training module is used to input the training set into the MTFIF-STMM model for iterative training. For unlabeled data, the model is optimized by minimizing the similarity loss of the two-branch temporal features. For labeled data, the model is optimized by a combination of mean squared error loss and similarity loss of the two-branch sequence features. The model parameters are updated using a backpropagation mechanism. The optimal MTFIF-STMM model is selected using the validation set. The performance of the optimal MTFIF-STMM model is evaluated using the test set until the performance evaluation meets the set threshold, at which point the final MTFIF-STMM model is obtained. The prediction module is used to input the real-time collected multi-source data of sinter into the final MTFIF-STMM model for soft measurement to obtain the real-time assessment results of the ferrous oxide content of sinter.

8. A computer device, characterized in that, Includes: memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 6.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images