A sinter ferrous oxide content real-time detection method based on a bias correction formal recursive fusion framework

Abstract

The application discloses a sinter ferrous oxide content real-time detection method based on a bias correction formal recursive fusion framework. Process variable data and infrared image data are collected and preprocessed, shallow feature data is extracted from the preprocessed infrared image, and the shallow feature data is fused with the preprocessed process variable data. The fused variable data and the preprocessed sinter cross-section infrared image are respectively input into corresponding branches to obtain two modal original features. The two modal original features are input into the bias correction formal recursive fusion framework, and the output of the bias correction formal recursive fusion framework is processed to obtain the ferrous oxide content. The application can explicitly utilize inconsistent information between modes to correct state evolution, improve noise resistance, drift resistance and mismatch resistance under complex conditions, and maintain the real-time reasoning advantage brought by the Mamba linear complexity, and is suitable for sinter process ferrous oxide content online real-time detection and intelligent monitoring.

Description

Technical Field

[0001] This invention belongs to the field of industrial process soft measurement and intelligent sensing technology, specifically involving a real-time detection method for ferrous oxide content in sintered ore based on a deviation correction recursive fusion framework. Background Technology

[0002] Sintering is a crucial upstream process in iron and steel smelting, and its operational status directly impacts the stability, fuel consumption, and production efficiency of subsequent blast furnace ironmaking. Ferrous oxide (FeO) content is a key quality indicator characterizing the thermal state and metallurgical properties of sintered ore. Excessive FeO content typically indicates over-burning during sintering, leading to decreased reducibility; conversely, insufficient FeO content may indicate under-burning, resulting in decreased sinter strength and increased pulverization rate. Therefore, accurate and real-time acquisition of FeO content is essential for controlling the sintering process.

[0003] Current industrial methods primarily rely on manual sampling and laboratory testing to obtain FeO content, which suffers from long detection cycles and significant feedback lags, making it difficult to meet the needs of online optimization control. Although soft sensing methods can predict FeO content using process variables and image information, existing technologies still have the following shortcomings:

[0004] 1. Insufficient representation using a single mode. Using only process variable data is insufficient to reflect the hot structure of the sintered cross section, while using only image data is insufficient to fully characterize the dynamic changes of the process, resulting in limited ability of the model to represent complex sintering states.

[0005] 2. Cross-modal alignment and enhancement methods are relatively crude. Existing multimodal soft measurement methods usually achieve information fusion only through simple splicing, weighting, or static alignment, which makes it difficult to simultaneously retain cross-modal shared information and bias information, resulting in semantic inconsistencies between modes and insufficient utilization of anomalous information.

[0006] 3. Existing coupled state update mechanisms lack explicit utilization of modal inconsistencies. Existing multimodal methods based on state-space models or Mamba mostly adopt direct state coupling or feature-weighted fusion. Inconsistencies between modes are usually treated as noise and weakened, making it difficult to transform them into state correction signals. As a result, they are unable to cope with drift, mismatch, occlusion and abnormal disturbances under complex operating conditions.

[0007] 4. It is difficult to balance real-time performance and robustness. Models based on self-attention mechanisms have high computational complexity in long sequence modeling, while ordinary recursive models struggle to maintain sensitivity to abnormal states under complex operating conditions, limiting their application in industrial online real-time detection scenarios.

[0008] Therefore, there is an urgent need to propose a new multimodal real-time detection method that, while maintaining the linear complexity advantage of the state-space model, enhances the cross-modal alignment capability and can explicitly utilize intermodal inconsistency information to correct state evolution, thereby improving the accuracy, stability and robustness of FeO content detection in sinter. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention provides a real-time detection method for ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework. This method solves the technical problems in the prior art, such as the difficulty in real-time and accurate detection of FeO content, insufficient utilization of abnormal information by cross-modal alignment and enhancement methods, difficulty in explicitly utilizing modal inconsistencies through direct coupling mechanism of state-space model, and difficulty in balancing real-time performance and robustness.

[0010] The technical solution adopted in this invention is:

[0011] I. A real-time detection method for ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework:

[0012] S1. Real-time acquisition and preprocessing of process variable data and infrared image data of sinter cross sections during the sintering process.

[0013] The process variable data include the temperature of the annular cooler shell, the temperature of the exhaust gas from the air box, the frequency of the main fan, the negative pressure of the main flue, the temperature of the exhaust gas from the main flue, the thickness of the material layer, the amount of water added to the primary mixing and granulation cylinder, the amount of water added to the secondary mixing and granulation cylinder, the quicklime ratio, and the return ore rate.

[0014] S2. Based on expert knowledge, shallow feature data containing cross-sectional temperature, cross-sectional thickness and sintering uniformity are extracted from the pre-processed infrared image of the sinter cross-section. The shallow feature data is then fused with the pre-processed process variable data to obtain the fused variable data.

[0015] S3. Input the fused variable data and the preprocessed infrared image of the sinter section into the time-series coding branch and the image coding branch respectively for processing, and obtain the original features of the process mode and the original features of the image mode respectively.

[0016] S4. Input the original features of the process modality and the original features of the image modality together into a deviation correction recursive fusion framework that includes several deviation correction modules connected in series for processing, to obtain the corrected process modality hidden state of each deviation correction module at each time step and the corrected image modality hidden state of each deviation correction module at each time step.

[0017] Each of the deviation correction modules includes a consistent difference feature alignment enhancement module and a deviation correction coupling Mamba module connected in series.

[0018] S5. Pool the corrected process modal hidden state and the corrected image modal hidden state output by each deviation correction module together. Then, perform weighted fusion and regression processing on the pooled outputs of each deviation correction module in sequence to finally obtain the real-time detection result of ferrous oxide content in sinter.

[0019] The cross-sectional temperature, cross-sectional thickness, and sintering uniformity are obtained by processing according to the following formula:

[0020] ; ;

[0021] in, The cross-sectional temperature; The thickness of the cross section; To ensure sintering uniformity; and These represent the height and width of the infrared image, respectively. and All are indexes; Represents pixels in an infrared image The brightness value at that location; Indicating the first in an infrared image The height of the flames in the column; Indicating the first in an infrared image The height of the highest temperature point; This represents the average height of the highest temperature points across all columns in the infrared image.

[0022] The temporal coding branch employs a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU) network, or other temporal neural networks; the image coding branch employs an SE-ResNet50 network or other image feature extraction networks.

[0023] The specific process of the consistent and differential feature alignment enhancement module is as follows: First, cross-modal alignment of the two input modalities is achieved through nonlinear mapping and deep canonical correlation analysis. Then, the aligned dual-modal features are decoupled into consistent features and differential features. Consistent perception enhancement weights and differential perception enhancement weights are generated based on the consistent features and differential features, and finally, the observation features of the two modalities are obtained.

[0024] Each of the aforementioned consistent difference feature alignment enhancement modules is configured according to the following formula:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] in, This is a time index, representing the current moment. ; and These represent process mode and image mode, respectively. and These represent process modal observation features and image modal observation features, respectively. This represents the process modal input of the consistent difference feature alignment enhancement module, where Represents the original characteristics of the process mode. The corrected process modal hidden state is the output of the Mamba module coupled with the deviation correction of the previous deviation correction module. The image modality input represents the consistent difference feature alignment enhancement module, where Represents the original features of the image modality. The corrected image modality hiding state is the output of the Mamba module coupled with the deviation correction of the previous deviation correction module. and These are the first and second characteristic projection operators of the process mode, respectively; and These are the first and second feature projection operators for the image modality, respectively; This represents element-wise multiplication; and These represent the consistency perception enhancement weight and the difference perception enhancement weight, respectively. and These represent consistent features and dissimilar features, respectively. express Activation function; and These represent the enhancement weight mapping matrices for the process mode and the image mode, respectively; and These represent the bias terms for the process mode and the image mode, respectively. The intermediate features of the process and the image are aligned by deep canonical correlation analysis to form bimodal features; and These represent the process features and image features in the bimodal features, respectively; and These represent intermediate features during the process and intermediate features in the image, respectively. Indicates intermediate features and Conduct in-depth canonical correlation analysis; and These represent the nonlinear mappings of the process mode and the image mode, respectively; This indicates a feature splicing operation.

[0031] The specific processing procedure of the deviation correction coupled Mamba module is as follows:

[0032] D1. Perform cross-modal state prediction on the process modal observation features and image modal observation features output by the consistent difference feature alignment enhancement module to obtain the process modal prediction state and image modal prediction state corresponding to the two modalities respectively.

[0033] D2. Reconstruct the current observation based on the process modality prediction state and the image modality prediction state, construct the prediction bias of each modality, and use the prediction bias of this modality and the cross-modality prediction bias to perform coupled correction and update of the prediction state, so as to obtain the corrected process modality hidden state and the corrected image modality hidden state respectively.

[0034] The process modality prediction state and the image modality prediction state are obtained by processing according to the following formula:

[0035] ;

[0036] ;

[0037] ; ;

[0038] in, This is a time index, representing the current moment. ; and These represent the process mode prediction state and the image mode prediction state, respectively. and They represent the previous moment respectively. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and These represent process modal observation features and image modal observation features, respectively. and These represent cross-modal state prediction operators from image mode to process mode and from process mode to image mode, respectively. and These represent the discretized state transition parameters for the process mode and the image mode, respectively; and These represent the discretized input mapping parameters for the process mode and the image mode, respectively; Represents process mode or image mode, where Representing process modes, Represents image modality; and They represent Discretized state transition parameters and discretized input mapping parameters of the mode; , and They represent The modal's selective step size parameter, selective input mapping parameter, and selective output mapping parameter; express Modal observation characteristics; Represents the identity matrix; and They represent The continuous state transition matrix and continuous input mapping matrix of the modality; , and These represent the mapping functions that generate the corresponding selective parameters.

[0039] The corrected modal hidden state and the corrected image modal hidden state are obtained by processing according to the following formula:

[0040] ;

[0041] ;

[0042] ;

[0043] in, This is a time index, representing the current moment. ; and They represent the current time. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and These represent the process mode prediction state and the image mode prediction state, respectively. and These represent the deviation correction operators for the process mode and the image mode, respectively; and These represent the deviation correction operators from image mode to process mode and from process mode to image mode, respectively; and These represent the process mode prediction bias and the image mode prediction bias, respectively. and These represent process modal observation features and image modal observation features, respectively. and These represent process mode prediction observations and image mode prediction observations, respectively. and These represent the selective output mapping parameters for the process mode and the image mode, respectively.

[0044] The real-time detection results of the ferrous oxide content in the sintered ore are obtained by processing it according to the following formula:

[0045] ;

[0046]

[0047] in, This indicates the test results for the ferrous oxide content in sintered ore; This represents a multilayer perceptron regressor; For index, indicating the first One deviation correction module; This represents the total number of deviation correction modules; For the first Inter-layer fusion weights of a deviation correction module; For the first The joint characterization of the outputs of the deviation correction modules; Indicates the pooling layer; and They represent the first Individual deviation correction modules Time's up The length of time is The corrected process modal hidden state sequence and the corrected image modal hidden state sequence obtained within the time window.

[0048] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] 1. Improved completeness of state representation through collaborative modeling of shallow expert knowledge and deep dual-modal modeling. By fusing three shallow expert features—section temperature, section thickness, and sintering uniformity—with process variable data, and combining deep image features and deep process features, the comprehensive ability to describe the sintering state is improved.

[0051] 2. Enhanced alignment of consistent and dissimilar features improves cross-modal alignment and anomaly detection capabilities. This invention aligns bimodal features in a shared subspace using the CD-FAE module and further decomposes consistent and dissimilar features, enabling the model to utilize both cross-modal consistent and dissimilar information simultaneously, rather than retaining only the shared portion, thereby improving anomaly sensitivity under complex operating conditions.

[0052] 3. Deviation correction coupled with state update enhances robustness against mismatch and drift. Unlike existing direct state coupling methods, the DCC-Mamba module proposed in this invention explicitly transforms intermodal inconsistencies into state correction signals through a prediction-deviation-correction mechanism, thereby enhancing the model's robustness in scenarios such as image occlusion, sensor fluctuations, and operating condition drift.

[0053] 4. Balancing robustness and real-time performance. A bias-corrected recursive fusion framework is used to progressively pass the corrected fusion state layer by layer, achieving real-time online detection and intelligent monitoring of FeO content while maintaining the linear complexity advantage of the Mamba state-space model. Attached Figure Description

[0054] Figure 1 This is an overall flowchart of the method of the present invention.

[0055] Figure 2 This is a schematic diagram of the deviation correction recursive fusion framework of the present invention.

[0056] Figure 3 The diagram shows shallow feature data, where (a) is the cross-sectional temperature diagram, (b) is the cross-sectional thickness diagram, and (c) is the sintering uniformity diagram. Detailed Implementation

[0057] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0058] like Figure 1 As shown, the real-time detection method for ferrous oxide content in sintered ore in this embodiment is implemented according to the following steps:

[0059] S1. Real-time acquisition and preprocessing of process variable data and infrared image data of sinter cross sections during the sintering process.

[0060] Process variable data include the outer shell temperature of the annular cooler, the exhaust gas temperature of the air box, the frequency of the main fan, the negative pressure of the main flue, the exhaust gas temperature of the main flue, the thickness of the material layer, the water addition amount of the primary mixing and granulation cylinder, the water addition amount of the secondary mixing and granulation cylinder, the quicklime ratio, and the return ore rate.

[0061] Specifically, process variable data and infrared image data of sinter cross sections are collected in real time during the sintering process, and the process variable data and infrared image data of sinter cross sections are preprocessed to obtain preprocessed process variable data and infrared image data of sinter cross sections.

[0062] In this embodiment, process variable data, infrared images of sinter cross-sections, and corresponding FeO content test values ​​of the sinter are collected at the sintering production site. Preferably, the variables among the 17 process variable data used for soft measurement modeling after correlation screening are: annular cooler shell temperature, exhaust gas temperature of south side of #21 wind box, exhaust gas temperature of south side of #20 wind box, exhaust gas temperature of north side of #24 wind box, exhaust gas temperature of north side of #23 wind box, exhaust gas temperature of north side of #22 wind box, exhaust gas temperature of north side of #21 wind box, exhaust gas temperature of north side of #20 wind box, main fan frequency, negative pressure of south side of main flue, negative pressure of north side of main flue, exhaust gas temperature of main flue, material layer thickness, water addition amount of primary mixing granulation cylinder, water addition amount of secondary mixing granulation cylinder, quicklime ratio, and return ore rate. Infrared images can be collected by an infrared thermal imager installed at the tail of the sintering machine. Tag data are the FeO content test values ​​of the sinter at the corresponding time or time window. The # symbol is an abbreviation for the equipment number. For example, the exhaust gas temperature in the south of the 21# wind box is the exhaust gas temperature in the south of the 21# wind box.

[0063] In this embodiment, the preprocessing of process variable data specifically involves: normalizing, segmenting by sliding time windows, and aligning the process variable data sequentially. The preprocessing of sinter cross-sectional infrared image data specifically involves: cropping, standardizing, and unifying the dimensions of the sinter cross-sectional infrared image data sequentially.

[0064] Expert feature extraction + feature stitching: S2, Based on expert knowledge, shallow feature data containing cross-sectional temperature, cross-sectional thickness and sintering uniformity are extracted from the pre-processed infrared image of the sinter cross-section, and the shallow feature data is fused with the pre-processed process variable data to obtain fused variable data.

[0065] The cross-sectional temperature, cross-sectional thickness, and sintering uniformity are obtained by processing according to the following formula:

[0066] ; ;

[0067] in, The cross-sectional temperature; The thickness of the cross section; To ensure sintering uniformity; and These represent the height and width of the infrared image of the sinter cross section, respectively. and All are indexes; Represents the pixels in an infrared image of a sintered ore cross section. The brightness value at that location; The first part of the infrared image of the cross-section of sintered ore The height of the flames in the column; The first part of the infrared image of the cross-section of sintered ore The height of the highest temperature point; This represents the average height of the highest temperature points in all columns of the infrared image of the sinter cross-section.

[0068] like Figure 3 This is a schematic diagram of shallow feature data extracted from a preprocessed infrared image of a sinter cross-section, where... Figure 3 (a) represents the cross-sectional temperature. Figure 3 (b) represents the cross-sectional thickness. Figure 3 (c) represents sintering uniformity.

[0069] The shallow feature data is merged with the preprocessed process variable data by directly combining the shallow feature data with the preprocessed process variable data.

[0070] S3. Input the fused variable data and the preprocessed infrared image of the sinter section into the time-series coding branch (process modal encoder) and the image coding branch (image modal encoder) for processing, respectively, to obtain the original process modal features and the original image modal features.

[0071] The temporal coding branch uses a Long Short-Term Memory (LSTM) network, a gated recurrent unit (GRU) or other temporal neural networks; the image coding branch uses an SE-ResNet50 or other image feature extraction network.

[0072] In this embodiment, the temporal coding branch uses a Long Short-Term Memory (LSTM) network, and the image coding branch uses an SE-ResNet50 network.

[0073] like Figure 2 As shown in step S4, the original features of the process modality and the original features of the image modality are input together into a deviation correction recursive fusion framework consisting of several deviation correction modules connected in series for processing, so as to obtain the corrected process modality hidden state of each deviation correction module at each time step and the corrected image modality hidden state of each deviation correction module at each time step.

[0074] Each deviation correction module includes a Consistency-Difference Feature Alignment and Enhancement (CD-FAE) module and a deviation correction coupled Mamba (DCC-Mamba) module connected in series.

[0075] The input of the consistent difference feature alignment enhancement module is used as the input of the deviation correction module. The output of the consistent difference feature alignment enhancement module is connected to the input of the deviation correction coupled Mamba module. The output of the deviation correction coupled Mamba module is used as the output of the deviation correction module.

[0076] Specifically, the original features of the process modality and the original features of the image modality are input together into the input of the first deviation correction module, and then processed by multiple subsequent deviation correction modules in sequence. Each deviation correction module outputs the hidden state of the process modality and the hidden state of the image modality at each time step.

[0077] The consistency-difference feature alignment enhancement module first achieves cross-modal alignment of the two input modalities through nonlinear mapping and deep canonical correlation analysis. Then, it decouples the aligned dual-modal features into consistency features and difference features, and generates consistency-aware enhancement weights and difference-aware enhancement weights based on the consistency features and difference features, finally obtaining the observation features of the two modalities.

[0078] Specifically, the consistency difference feature alignment enhancement module includes a nonlinear mapping submodule, a deep canonical correlation analysis alignment submodule, a consistency / difference decoupling submodule, and a consistency difference perception enhancement submodule.

[0079] In this implementation, the processing procedure for each consistent difference feature alignment enhancement module is as follows:

[0080] Nonlinear mapping submodule: First, intermediate features of the two modes are obtained via nonlinear mapping:

[0081] ;

[0082] Deep Canonical Correlation Analysis Alignment Submodule: Then, aligned bimodal features are obtained through deep canonical correlation analysis (DCCA).

[0083]

[0084] Consistent / Difference Decoupling Submodule: Further, the aligned bimodal features are decoupled into consistent features and difference features.

[0085] ;

[0086] Generate consistency-aware enhancement weights and difference-aware enhancement weights based on consistency and difference features:

[0087] ;

[0088] Consistent difference perception enhancement submodule: Obtains enhanced process modal observation features and image modal observation features respectively.

[0089]

[0090]

[0091] in, This is a time index, representing the current moment. ; and These represent process mode and image mode, respectively. and They represent the current time. Process modal observation features and image modal observation features; Indicates the current time The process modal input of the consistent difference feature alignment enhancement module, where Indicates the current time Original characteristics of process modes For the current moment The deviation correction of the previous deviation correction module is coupled with the correction of the Mamba module output, resulting in the post-correction process modal hidden state. Indicates the current time The consistent difference feature alignment enhancement module's image modality input, where Indicates the current time Image modal primitive features, For the current moment The deviation correction of the previous deviation correction module is coupled with the Mamba module output of the corrected image modality hiding state; and These are the first and second characteristic projection operators of the process mode, respectively; and These are the first and second feature projection operators for the image modality, respectively; This represents element-wise multiplication; and They represent the current time. Consistency perception enhancement weights and difference perception enhancement weights; and They represent the current time. Consistent and differing characteristics; express Activation function; and These represent the enhancement weight mapping matrices for the process mode and the image mode, respectively; and These represent the bias terms for the process mode and the image mode, respectively. Indicates the current time The process and intermediate features of the image are aligned by deep canonical correlation analysis to form bimodal features; and They represent the current time. Process features and image features in dual-modal features; and They represent the current time. Intermediate features in the process and intermediate features in the image; Indicates the current time Lower pair of intermediate features and Conduct in-depth canonical correlation analysis; and These represent the nonlinear mappings of the process mode and the image mode, respectively; This indicates a feature splicing operation.

[0092] Specifically, the first feature projection operator of the process mode The second characteristic projection operator of the process mode First feature projection operator of image modality and the second feature projection operator of the image modality All are learnable matrices.

[0093] Specifically, the process modal input in the first consistent difference feature alignment enhancement module. Representing the original features of process modes Image modality input in the first consistent difference feature alignment enhancement module Represents the original features of image modalities The process modal inputs in the alignment enhancement module for each consistent difference feature except the first one. Both represent the corrected modal hidden state of the previous deviation correction module's output coupled with the Mamba module. Image modality input in each consistent difference feature alignment enhancement module except the first one. Both represent the corrected image modality hidden state output by the deviation correction coupled Mamba module of the previous deviation correction module. .

[0094] The specific processing procedure of the deviation correction coupled Mamba module is as follows:

[0095] D1. Perform cross-modal state prediction on the process modal observation features and image modal observation features output by the consistent difference feature alignment enhancement module to obtain the process modal prediction state and image modal prediction state corresponding to the two modalities respectively.

[0096] In this embodiment, the process of obtaining the process modality prediction state and the image modality prediction state is as follows:

[0097] Selective step size parameters, selective input mapping parameters, and selective output mapping parameters are generated based on process modal observation features and image modal observation features:

[0098] ; ;

[0099] Discretized state transition parameters and discretized input mapping parameters are obtained through discretization:

[0100] ;

[0101] Subsequently, cross-modal state prediction was performed on the process mode and the image mode respectively, resulting in the predicted states for the process mode and the image mode:

[0102] ;

[0103] in, and They represent the current time. The process modal prediction state and the image modal prediction state; and They represent the previous moment respectively. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and They represent the current time. Process modal observation features and image modal observation features; and They represent the current time. Cross-modal state prediction operators from image mode to process mode and from process mode to image mode; and They represent the current time. Discretized state transition parameters for process modes and image modes; and They represent the current time. Discretized input mapping parameters for process modes and image modes; Represents process mode or image mode, where Representing process modes, Represents image modality; and They represent the current time. of Discretized state transition parameters and discretized input mapping parameters of the mode; , and They represent the current time. of Selective step size parameters, selective input mapping parameters, and selective output mapping parameters of the modality, which are related to the current observed features; Indicates the current time of Modal observation characteristics; Represents the identity matrix; and They represent The continuous state transition matrix and continuous input mapping matrix of the modality; , and These represent the mapping functions that generate the corresponding selective parameters.

[0104] In practical implementation, a cross-modal state prediction operator from image mode to process mode. and cross-modal state prediction operators from process mode to image mode Essentially, it is a learnable matrix.

[0105] D2. Reconstruct the current observation based on the process modality prediction state and the image modality prediction state, construct the prediction bias of each modality, and use the prediction bias of this modality and the cross-modality prediction bias to perform coupled correction and update of the prediction state, so as to obtain the corrected process modality hidden state and the corrected image modality hidden state respectively.

[0106] In this embodiment, the process of obtaining the corrected process modal hidden state and the corrected image modal hidden state is as follows:

[0107] First, the current observation is reconstructed from the process mode prediction state and the image mode prediction state, resulting in process mode prediction observations and image mode prediction observations, respectively:

[0108] ;

[0109] Then, the construction process modal prediction bias and the image modal prediction bias:

[0110] ;

[0111] The state is then corrected using the local modality prediction bias and the cross-modality prediction bias, resulting in the corrected process modality hidden state and the corrected image modality hidden state, respectively:

[0112] ;

[0113] in, and They represent the current time. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and They represent the current time. The process modal prediction state and the image modal prediction state; and They represent the current time. The deviation correction operator for process mode and image mode; and They represent the current time. Deviation correction operators from image mode to process mode and from process mode to image mode; and They represent the current time. Process mode prediction bias and image mode prediction bias; and They represent the current time. Process modal observation features and image modal observation features; and They represent the current time. Process modal prediction observation and image modal prediction observation; and They represent the current time. Selective output mapping parameters for process mode and image mode.

[0114] Specifically, within this framework, inconsistencies between modes are no longer simply viewed as noise, but are transformed into state correction signals. Deviation correction and deep fusion are gradually achieved through the recursive stacking of multiple deviation correction modules (multiple CD-FAE modules and multiple DCC-Mamba modules are alternately stacked).

[0115] S5. Pool the corrected process modal hidden state and the corrected image modal hidden state output by each deviation correction module together. Then, perform weighted fusion and regression processing on the pooled outputs of each deviation correction module in sequence to finally obtain the real-time detection result of ferrous oxide content in sinter.

[0116] The specific process for real-time detection of ferrous oxide content in sintered ore is as follows:

[0117] Pool the output of each deviation correction module. The output of the deviation correction module is:

[0118]

[0119] The outputs of each bias correction module after pooling are then weighted, fused, and regressed sequentially to obtain the final detection prediction value:

[0120] ;

[0121] in, This indicates the predicted results of the detection of ferrous oxide content in sintered ore; This represents a multilayer perceptron regressor; For index, indicating the first One deviation correction module; This represents the total number of deviation correction modules; For the first Inter-layer fusion weights of a deviation correction module; For the first The joint characterization of the outputs of the deviation correction modules; Indicates the pooling layer; and They represent the first Individual deviation correction modules Time's up The length of time is The corrected process modal hidden state sequence and the corrected image modal hidden state sequence obtained within the time window.

[0122] Further, steps S1-S5 constitute the reasoning process when the method of the present invention is used directly. Training is required before using the reasoning process. Process variable data and infrared image data of sinter cross-sections at each moment in the historical process are collected as supervisory labels to form multi-source heterogeneous samples, thereby constructing a ferrous oxide content detection dataset. Steps S2-S5 are repeated based on the constructed ferrous oxide content detection dataset for continuous training until training is complete.

[0123] To verify the effectiveness of the method of the present invention, this embodiment also selects SF-Pyraformer, MCFF and MRFF in the prior art as comparative methods, and uses RMSE, MAE and R... 2HR is used as the evaluation metric. Among them, SF-Pyraformer is a single-level fusion method based on process data and three shallow image features; MCFF is a previously proposed multi-level correlational fusion method; MRFF is the original multi-level recursive fusion framework based on FAE and Coupled Mamba, which can be regarded as a direct predecessor comparison framework of the method of this invention. The method of this invention further introduces CD-FAE and DCC-Mamba on the basis of MRFF to achieve consistent difference perception enhancement and prediction-bias-correction coupled state update. The performance comparison of each method under the same dataset and evaluation metrics is shown in Table 1 below:

[0124] Table 1:

[0125]

[0126] As shown in the table above, the method of this invention outperforms the existing SF-Pyraformer, MCFF, and MRFF in all four indicators. Specifically, compared to the original MRFF framework, the method of this invention further improves the accuracy and robustness of ferrous oxide (FeO) content detection by replacing the original direct state coupling mechanism with a bias-corrected coupling Mamba module (DCC-Mamba) and introducing a consistent / differential decoupling consistent difference feature alignment enhancement module (CD-FAE).

[0127] This invention can explicitly utilize intermodal inconsistency information to correct state evolution, improve noise resistance, drift resistance and mismatch resistance under complex working conditions, and maintain the real-time inference advantage brought by Mamba linear complexity. It is suitable for online real-time detection and intelligent monitoring of ferrous oxide (FeO) content in the sintering process.

[0128] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for real-time detection of ferrous oxide content in sintered ore based on a bias-corrected recursive fusion framework, characterized in that, Includes the following steps: S1. Real-time acquisition and preprocessing of process variable data and infrared image data of sinter cross section during sintering; S2. Based on expert knowledge, shallow feature data containing cross-sectional temperature, cross-sectional thickness and sintering uniformity are extracted from the pre-processed infrared image of the sinter cross-section. The shallow feature data is then fused with the pre-processed process variable data to obtain the fused variable data. S3. Input the fused variable data and the preprocessed infrared image of the sinter section into the time-series coding branch and the image coding branch respectively for processing, and obtain the original features of the process mode and the original features of the image mode respectively. S4. Input the original features of the process modality and the original features of the image modality together into a deviation correction recursive fusion framework that includes several deviation correction modules connected in series for processing, to obtain the corrected process modality hidden state of each deviation correction module at each time step and the corrected image modality hidden state of each deviation correction module at each time step. Each of the deviation correction modules includes a consistent difference feature alignment enhancement module and a deviation correction coupling Mamba module connected in series. The specific process of the consistent and differential feature alignment enhancement module is as follows: First, cross-modal alignment of the two input modalities is achieved through nonlinear mapping and deep canonical correlation analysis. Then, the aligned dual-modal features are decoupled into consistent features and differential features. Consistent perception enhancement weights and differential perception enhancement weights are generated based on the consistent features and differential features, and finally, the observation features of the two modalities are obtained. The specific processing procedure of the deviation correction coupled Mamba module is as follows: D1. Perform cross-modal state prediction on the process modal observation features and image modal observation features output by the consistent difference feature alignment enhancement module to obtain the process modal prediction state and image modal prediction state corresponding to the two modes respectively; D2. Reconstruct the current observation based on the process modality prediction state and the image modality prediction state, construct the prediction bias of each modality, and use the prediction bias of this modality and the cross-modality prediction bias to perform coupled correction and update of the prediction state, so as to obtain the corrected process modality hidden state and the corrected image modality hidden state respectively. S5. Pool the corrected process modal hidden state and the corrected image modal hidden state output by each deviation correction module together. Then, perform weighted fusion and regression processing on the pooled outputs of each deviation correction module in sequence to finally obtain the real-time detection result of ferrous oxide content in sinter.

2. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: The cross-sectional temperature, cross-sectional thickness, and sintering uniformity are obtained by processing according to the following formula: ； ； in, The cross-sectional temperature; The thickness of the cross section; To ensure sintering uniformity; and These represent the height and width of the infrared image, respectively. and All are indexes; Represents pixels in an infrared image The brightness value at that location; Indicating the first in an infrared image The height of the flames in the column; Indicating the first in an infrared image The height of the highest temperature point; This represents the average height of the highest temperature points across all columns in the infrared image.

3. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: The temporal coding branch uses a Long Short-Term Memory (LSTM) network and a gated recurrent unit (GRU); the image coding branch uses SE-ResNet50.

4. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: Each of the aforementioned consistent difference feature alignment enhancement modules is configured according to the following formula: ； ； ； ； ； in, This is a time index, representing the current moment. ; and These represent process mode and image mode, respectively. and These represent process modal observation features and image modal observation features, respectively. This represents the process modal input of the consistent difference feature alignment enhancement module, where Represents the original characteristics of the process mode. The corrected process modal hidden state is the output of the Mamba module of the deviation correction coupling module of the previous deviation correction module. The image modality input represents the consistent difference feature alignment enhancement module, where Represents the original features of the image modality. The corrected image modality hiding state is the output of the Mamba module coupled with the deviation correction of the previous deviation correction module. and These are the first and second characteristic projection operators of the process mode, respectively; and These are the first and second feature projection operators for the image modality, respectively; This represents element-wise multiplication; and These represent the consistency perception enhancement weight and the difference perception enhancement weight, respectively. and These represent consistent features and dissimilar features, respectively. express Activation function; and These represent the enhancement weight mapping matrices for the process mode and the image mode, respectively; and These represent the bias terms for the process mode and the image mode, respectively. The intermediate features of the process and the image are aligned by deep canonical correlation analysis to form bimodal features; and These represent the process features and image features in the bimodal features, respectively; and These represent intermediate features during the process and intermediate features in the image, respectively. Indicates intermediate features and Conduct in-depth canonical correlation analysis; and These represent the nonlinear mappings of the process mode and the image mode, respectively; This indicates a feature splicing operation.

5. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: The process modality prediction state and the image modality prediction state are obtained by processing according to the following formula: ； ； ； ； in, This is a time index, representing the current moment. ; and These represent the process mode prediction state and the image mode prediction state, respectively. and They represent the previous moment respectively. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and These represent process modal observation features and image modal observation features, respectively. and These represent cross-modal state prediction operators from image mode to process mode and from process mode to image mode, respectively. and These represent the discretized state transition parameters for the process mode and the image mode, respectively. and These represent the discretized input mapping parameters for the process mode and the image mode, respectively; Represents process mode or image mode, where Representing process modes, Represents image modality; and They represent Discretized state transition parameters and discretized input mapping parameters of the mode; , and They represent The modal's selective step size parameter, selective input mapping parameter, and selective output mapping parameter; express Modal observation characteristics; Represents the identity matrix; and They represent The continuous state transition matrix and continuous input mapping matrix of the modality; , and These represent the mapping functions that generate the corresponding selective parameters.

6. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: The corrected modal hidden state and the corrected image modal hidden state are obtained by processing according to the following formula: ； ； ； in, This is a time index, representing the current moment. ; and They represent the current time. The modal hidden state of the corrected process and the modal hidden state of the corrected image; and These represent the process mode prediction state and the image mode prediction state, respectively. and These represent the deviation correction operators for the process mode and the image mode, respectively; and These represent the deviation correction operators from image mode to process mode and from process mode to image mode, respectively; and These represent the process mode prediction bias and the image mode prediction bias, respectively. and These represent process modal observation features and image modal observation features, respectively. and These represent process mode prediction observations and image mode prediction observations, respectively. and These represent the selective output mapping parameters for the process mode and the image mode, respectively.

7. The method for real-time detection of ferrous oxide content in sintered ore based on a deviation-corrected recursive fusion framework according to claim 1, characterized in that: The real-time detection results of the ferrous oxide content in the sintered ore are obtained by processing the data using the following formula: ； in, This indicates the test results for the ferrous oxide content in sintered ore; This represents a multilayer perceptron regressor; For index, indicating the first One deviation correction module; This represents the total number of deviation correction modules; For the first Inter-layer fusion weights of a deviation correction module; For the first The joint characterization of the outputs of the deviation correction modules; Indicates the pooling layer; and They represent the first Individual deviation correction modules in Time's up The length of time is The corrected process modal hidden state sequence and the corrected image modal hidden state sequence obtained within the time window.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images