A blast furnace working condition identification method and system fusing process variables and video images

Abstract

The application relates to the technical field of industrial working condition recognition, and discloses a blast furnace working condition recognition method and system fusing process variables and video images. The method performs interpolation processing on the process variables, performs averaging on the video image frames, and aligns the multi-source heterogeneous information to the same time step. In this way, the multi-source heterogeneous information with inconsistent sampling frequencies is aligned, a deep neural network model jointly driven by the process variables and the video images is constructed, representative working condition features in the multi-source heterogeneous information are extracted, the problem that working condition information is not completely utilized when only a single information source is used is overcome, a double-layer residual connection module and a model stability evaluation index are designed, and accurate recognition of the blast furnace working condition is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial working condition recognition, and in particular to a blast furnace working condition recognition method and system fusing process variables and video images. BACKGROUND

[0002] As the core process of the steel metallurgical industry, the stable operation of blast furnace ironmaking is an important prerequisite for high-quality and efficient steel production. The blast furnace working condition is a direct reflection of the blast furnace operation state, and can provide key feedback information for closed-loop control of blast furnace ironmaking. Therefore, accurate identification of the blast furnace working condition is a mandatory requirement in the blast furnace ironmaking process.

[0003] With the continuous progress of key technologies such as sensor devices and distributed control systems, a large amount of multi-source heterogeneous information has been accumulated in the blast furnace ironmaking process, including but not limited to process variables, real-time video streams, and static images. According to the information source used for working condition recognition, existing blast furnace working condition recognition methods can be divided into two categories: process variable-based working condition recognition methods and video image-based working condition recognition methods. The process variable-based working condition recognition method uses statistical analysis of process variables collected by various sensors during blast furnace operation to establish a data-driven model, but this method requires a large number of samples for model training, and in actual situations, the number of abnormal working condition samples is often scarce, making it difficult to train a data-driven model with good generalization performance. At present, the video image-based working condition recognition method often uses image enhancement, semantic segmentation, and other technologies to establish an image feature extraction model, but this method has high complexity, long training time, and requires a large amount of computing resources, making it face adaptability challenges in industrial practical applications. In addition, existing technologies often use only a single information source, resulting in incomplete utilization of working condition information and low recognition accuracy. SUMMARY

[0004] The present application provides a blast furnace working condition recognition method and system fusing process variables and video images to solve the problems existing in the prior art.

[0005] In order to achieve the above-mentioned purpose, the technical scheme is as follows:

[0006] In a first aspect, the present application provides a blast furnace working condition recognition method fusing process variables and video images, comprising:

[0007] S1: collecting process variables and material surface video images of the blast furnace working condition to obtain multi-source heterogeneous information;

[0008] S2: interpolating the process variables and averaging the video image frames to align the multi-source heterogeneous information to the same time step;

[0009] S3: constructing a single-layer fusion network architecture;

[0010] S4: extracting a preliminary feature set of the aligned multi-source heterogeneous information based on the single-layer fusion network architecture;

[0011] S5: training and optimizing a double-layer fusion network model based on the preliminary feature set and a pre-constructed double-layer fusion network architecture;

[0012] S6: identifying a blast furnace working condition type based on the double-layer fusion network model.

[0013] In a second aspect, the present application provides a blast furnace working condition identification system fusing process variables and video images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the method of the first aspect when executing the computer program.

[0014] Advantages:

[0015] The blast furnace working condition identification method fusing process variables and video images provided by the present application performs interpolation processing on process variables, and performs averaging on video image frames, aligns multi-source heterogeneous information to the same time step; in this way, the multi-source heterogeneous information with inconsistent sampling frequencies is aligned, a deep neural network model jointly driven by process variables and video images is constructed, representative working condition features in the multi-source heterogeneous information are extracted, the problem of incomplete utilization of working condition information when only a single information source is used is overcome, a double-layer residual connection module and a model stability evaluation index are designed, and accurate identification of the blast furnace working condition is realized.

[0016] In further technical solutions, the present application proposes a deep neural network model jointly driven by process variables and video images, extracts process variable and video image features based on LSTM and CNN respectively, performs feature simplification through CCA, and realizes fusion of multi-source heterogeneous information;

[0017] In further technical solutions, the present application designs a double-layer residual connection module, splices the preliminary features extracted by the single-layer network and the top-layer input data after data alignment, realizes effective extraction and full utilization of important features of multi-source heterogeneous data;

[0018] In further technical solutions, the present application designs a model training stability evaluation index, analyzes the time consumption of the training process, the accuracy of the training set, the loss of the training set, the accuracy of the test set, and the correlation of the test set, and comprehensively evaluates the optimization speed, accuracy, and fluctuation of the model; BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1One of the flowcharts of a blast furnace working condition recognition method fusing process variables and video images according to a preferred embodiment of the present application;

[0020] Figure 2 One of the flowcharts of a blast furnace working condition recognition method fusing process variables and video images according to a preferred embodiment of the present application;

[0021] Figure 3 (a) A blast furnace top temperature change curve according to a preferred embodiment of the present application; Figure 3 (b) A charge surface image of each working condition according to a preferred embodiment of the present application;

[0022] Figure 4 A model training process index change curve schematic diagram according to a preferred embodiment of the present application;

[0023] Figure 5 A working condition recognition confusion matrix schematic diagram according to a preferred embodiment of the present application. DETAILED DESCRIPTION

[0024] The technical solutions of the present application will be described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] Unless otherwise defined, the technical terms or scientific terms used in the present application shall be understood as the usual meanings understood by those skilled in the art in the field to which the present application belongs. The terms "first", "second", and similar terms used in the present application do not represent any order, quantity, or importance, but are only used to distinguish different components. Similarly, "one" or "a" and similar terms do not represent a quantity limitation, but represent the existence of at least one. The terms "connected" or "connected" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right", and the like are only used to represent relative positional relationships, and when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0026] Please refer to Figure 1 The blast furnace working condition recognition method fusing process variables and video images provided by the present application comprises:

[0027] S1: Collecting process variables and charge surface video images of a blast furnace working condition to obtain multi-source heterogeneous information;

[0028] S2: Interpolating the process variables and averaging the video image frames to align the multi-source heterogeneous information to the same time step;

[0029] S3: constructing a single-layer fusion network architecture;

[0030] S4: extracting a preliminary feature set of the aligned multi-source heterogeneous information based on the single-layer fusion network architecture;

[0031] S5: training and optimizing a double-layer fusion network model based on the preliminary feature set and a pre-constructed double-layer fusion network architecture;

[0032] S6: identifying a blast furnace working condition type based on the double-layer fusion network model.

[0033] In this embodiment, process variables are collected by a self-deployed sensor cluster inside and outside the blast furnace, and a material surface video image is obtained by a high-temperature industrial endoscope on the blast furnace top. Specifically, the process variables include parameters such as top temperature, hot blast pressure, oxygen-enriched flow, gas composition, and permeability index, which describe the state data of the blast furnace ironmaking process in detail. The video image is a continuous information stream, which intuitively shows the changes of the material morphology and movement state in the blast furnace.

[0034] Next, a complete example of the specific implementation scheme of the above-mentioned blast furnace working condition identification method of fusing process variables and video images is described in detail as follows:

[0035] (1) Multi-source heterogeneous data alignment

[0036] It is worth noting that, in the blast furnace ironmaking process, the sampling methods of multi-source heterogeneous information are different, and the information collection frequencies show significant differences. Specifically, the blast furnace ironmaking process variables exist in discrete form, with a sampling interval of ΔT seconds, including parameters such as top temperature, hot blast pressure, oxygen-enriched flow, gas composition, and permeability index, which describe the state data of the blast furnace ironmaking process in detail. The video image is a continuous information stream, which intuitively shows the changes of the material morphology and movement state in the blast furnace. In order to synchronize the two in the time dimension, the data alignment method based on the Gaussian function is used to interpolate the process variables and average the video image frames, which specifically includes the following steps:

[0037] Step 1: Let Ψ represent the original process variable set, which contains p process variable types and m samples. Ψ c z Ψc(z) represents the zth time window sequence of the cth process variable, where c = 1, 2,..., p and z = 1, 2,..., s. By setting a certain window width, the m samples are divided into s time window sequences, so the number of samples contained in the time window sequence Ψ c z Ψc(z) is s.

[0038] Step 2: Interpolate the sample points in each time window sequence sequentially. Let n be the value of n in the time window sequence Ψ. c z The number of new samples interpolated between any two original sample points, point (t) i ,x i Let ) represent the i-th original sample point, i = 1, 2, ..., q, and let point Let represent the new sample point of the j-th interpolation between the i-th original sample point and its next adjacent original sample point, where Δt = 1 and j = 1,...,n. Let n = ΔT-1, and add a new sample at each second point. Step 3: Calculate the time window sequence Ψ c z The Gaussian function used for interpolation is modified to K(t), satisfying the following relationship:

[0039]

[0040] Where σ is the bandwidth, t is the time, and μ is the mean;

[0041] Calculate the value at time point based on the Gaussian function K(t). The size of the new sample point for interpolation at a given location satisfies the following relationship:

[0042]

[0043] In the formula, x m Let t be the true value of the m-th original data point. m Let K represent the time of the m-th original data point, and let K denote the Gaussian function.

[0044] Step 4: As a continuous information stream, video contains frames per second. To align this with the process variable set of one sample per second after interpolation, this application averages all frames per second to obtain an image representing the time per second. Let each second of video contain *n* frames. The average value of these *n* frames per second at each pixel location is calculated to obtain an image *F* representing the time per second. η The image F η It contains information about all video frame images within that second.

[0045] This data alignment method can effectively align process variables and video images with inconsistent sampling frequencies. Let the aligned process variable be X and the aligned video image be Y, which provides a solid and reliable data foundation for subsequent modeling and blast furnace condition identification tasks.

[0046] (2) Design a single-layer converged network architecture

[0047] The application proposes a deep neural network model jointly driven based on process variables and video images, wherein a single-layer fusion network architecture is shown as "the first layer" in Figure 2 .

[0048] In the single-layer fusion network architecture, the data-aligned process variables X and the data-aligned video images Y are input, and the first deep neural network model LSTM and the second deep neural network model CNN are used to extract the features of the data-aligned process variables X and the data-aligned video images Y respectively, so as to obtain two sets of feature sets F1 and F2 of heterogeneous information.

[0049] For the data-aligned process variables X and the data-aligned video images Y, the product of the covariance matrix is calculated:

[0050]

[0051] wherein Σ XX , Σ YY and Σ XY are the covariance matrix and the cross-covariance matrix of the data-aligned process variables X and the data-aligned video images Y, and Y is the product of the covariance matrix, and Y is singular value decomposed, satisfying the following relationship:

[0052] Y = ΓΣR T ; (4)

[0053]

[0054] wherein Γ, Σ and R are the matrix factors after singular value decomposition, J and L are weight parameters, is a vector space with the dimension size of lxl, is a vector space with the dimension size of mxm, and T represents transposition.

[0055] The feature sets u1 and v1 after the information simplifier are calculated, satisfying the following relationship:

[0056]

[0057] wherein the parameters J1, L1 and Σ1 are the weight parameters calculated by the first deep neural network model.

[0058] The function of the information simplifier is to reduce the typical correlation of the two sets of heterogeneous information features, so as to reduce the common features of the two sets of heterogeneous information features and retain most of the unique features of the two sets of heterogeneous information features.

[0059] (3) A double-layer residual connection module is proposed

[0060] The application proposes a double-layer residual connection module, which splices the feature sets u1 and v1 extracted by the single-layer fusion network architecture with the data-aligned process variable X and the data-aligned video image Y to obtain data sets X' and Y'

[0061]

[0062] wherein, is a data splicing operation.

[0063] Then, a new layer of fusion network architecture with the same structure but different parameters is built, as shown in the "second layer" in Figure 2 , and the data sets X' and Y' are input into the new fusion network architecture to extract features, obtaining feature sets F3 and F4 of two heterogeneous information:

[0064]

[0065] wherein, J2, L2 and Σ2 are weight parameters calculated by the second deep neural network model;

[0066] Next, the feature sets u2 and v2 are spliced and input into the full connection layer and the Softmax output layer, and the blast furnace working condition recognition result is calculated and the model is trained and the parameters are optimized through the cross-entropy loss function of the following model

[0067]

[0068] wherein, θ1 and θ2 are the model parameter sets of LSTM and CNN in the first layer, respectively, θ3 and θ4 are the model parameter sets of LSTM and CNN in the second layer, respectively, and θ5 is the parameter set of the full connection layer, is the parameter combination when the network reaches the optimal effect, y i represents the true label of the i-th sample, represents the recognition result of the i-th sample.

[0069] The double-layer residual connection module makes the output of the previous layer bypass the intermediate layer and directly connect with the subsequent layer. This architecture promotes the learning of residual mapping and ensures the effective propagation of input information within the network. It is worth noting that the network contains two independent feature extraction models, each of which is trained on different data sets, allowing relatively independent training processes. Despite this independence, both feature extraction models use the same optimization strategy and loss function, thereby improving overall performance while maintaining model autonomy.

[0070] (4) Design model training stability evaluation index

[0071] It is worth pointing out that in order to evaluate the stability of model training, the present application designs a model training stability evaluation index to evaluate the optimization speed, accuracy and fluctuation of the model. For H training processes of the model, it is assumed that each training process includes h training stages, and the following statistical information of each training process is calculated:

[0072]

[0073] wherein, The values are [t, d, l, e, g], respectively representing the five indicators of training process time consumption, training set accuracy, training set loss, test set accuracy and test set correlation, list i is the data sequence of the corresponding indicator, max() and min() represent the maximum and minimum values in the sequence list i , respectively. When represents the training set accuracy or test set accuracy, formula (14) takes max(), when it represents the training set loss or test set correlation, formula (14) takes min(), represents the value of the corresponding element in the jth training stage of the ith training process, and respectively represent the mean and standard deviation of the corresponding element in the ith training process, wherein i = 1, 2,..., H and j = 1, 2,..., h. The present application evaluates the accuracy of model training by calculating and

[0074] In order to evaluate the fluctuation of model training, the present application defines a fluctuation coefficient CV i γ (γ = 1, 2, 3, 4, 5), respectively representing the fluctuation coefficients of the five indicators t, d, l, e, g, see formula (17), and the smaller the value, the smaller the fluctuation of the model relative to the mean.

[0075] In order to quantify the optimization speed of model training, the present application defines the training stage number of d, l, e, g when the four indicators reach the set threshold of the optimal value of each indicator, i.e. For training set accuracy and test set accuracy, the threshold settings are [85%, 90%, 95%] respectively, and for training set loss and test set correlation, the threshold settings are [115%, 110%, 105%] respectively. The smaller the value, the faster the optimization speed of model training.

[0076] In summary, for the ith training process, the present application calculates the following statistical information for the five indicators t, d, l, e, g

[0077] ​

[0078] wherein, γ = 2, 3, 4, 5, respectively represent the training stage number when each index reaches the set threshold, and CV i γ respectively represent the extreme value, mean value, standard deviation and coefficient of variation of each index.

[0079] In order to evaluate the stability of the H training processes of the model, the above statistical information calculated for each training process is sorted and normalized, and the normalized result of the training process time t is Then, the weight is used to fuse the normalized data of the four indexes d, l, e, and g, and the fusion result r i γ The calculation formula of r is:

[0080]

[0081] wherein, γ = 2, 3, 4, 5, respectively represent the four indexes d, l, e, and g. represents the normalized data vector of the four indexes d, l, e, and g, and W1 represents the fusion weight vector.

[0082] Finally, the fusion result of the five indexes t, d, l, e, and g in the i-th training process is and the weight is used to fuse R i , to obtain the final model stability evaluation index SE.

[0083] SE i = W2·R i (20)

[0084] wherein, W2 is the fusion weight vector of the stability evaluation index calculation process, R i is the fusion result of the five indexes t, d, l, e, and g, and SE iThe SE is a stability evaluation index for the i-th training process, and the value range of SE is [0, 1], and the greater the value, the more stable the training process. In summary, the SE analyzes the time consumption of the training process, the accuracy of the training set, the loss of the training set, the accuracy of the test set and the correlation of the test set, and comprehensively evaluates the optimization speed, accuracy and fluctuation of the model training, and objectively and accurately describes the stability of the model training. In the preferred example, the training steps of the model can be optimized according to the evaluation index SE, for example, when the accuracy embodied by the evaluation index SE is low, the training process of the model is updated to obtain new parameters to improve the accuracy of the model. Herein, only examples are given, and no limitation is made.

[0085] In summary, the application provides a blast furnace working condition recognition method fusing process variables and video images. In view of the problem that process variables and video images are difficult to be synchronized in time sequence due to different sampling frequencies, the application designs a data alignment strategy based on a Gaussian function, realizes the alignment of heterogeneous data with inconsistent sampling frequencies in the time dimension by interpolating the process variables and averaging the video frames. On this basis, the application establishes a deep neural network model fusing multi-source heterogeneous information, uses LSTM and CNN to extract features of process variables and video images respectively, and uses CCA to simplify information. In order to more effectively extract multi-source data features, the application designs a double-layer residual connection module to combine the preliminary features with the top-layer input data after data alignment, which improves the extraction and utilization of important features of multi-source heterogeneous data. In addition, the application also provides a model training stability evaluation index, which analyzes the time consumption of the training process, the accuracy of the training set, the loss of the training set, the accuracy of the test set and the correlation of the test set, and comprehensively evaluates the optimization speed, accuracy and fluctuation of the model. The method provided by the application can accurately recognize eight types of blast furnace working conditions, solves the problem of incomplete utilization of working condition information and inaccurate blast furnace working condition recognition caused by using a single information source, and provides a scientific and reliable working condition recognition result for field workers.

[0086] In an example, taking a 2650m 3 high furnace of a domestic steel plant as an example, the blast furnace working condition recognition method fusing process variables and video images provided by the application is verified, specifically, the data samples used contain process variables and furnace top material surface videos in the same time period, and the data samples are as shown in Figure 3 . Figure 3(a) shows the change curve of the blast furnace top temperature, the data size of the original process variable is (1, 3899), and the data size after alignment is (1, 38987). In order to fully capture the data characteristics, the application adopts a sliding window processing, and the data size after sliding window processing is (1, 11, 38987). Similarly, 35 process variables are selected, and after transposing the data set, the data size of X is (38987, 11, 35). Figure 3 (b) shows the video images of the material surface of 8 blast furnace working conditions, which have been processed by image enhancement, and after data alignment, the data size of the material surface video image Y is (38987, 256, 256). Finally, X and Y are input into the model to perform the experiment of blast furnace working condition recognition task.

[0087] The index change curve of the model training process and the working condition recognition experiment result (recognition accuracy) are shown in Figure 4 and Figure 5 It can be seen that the blast furnace working condition recognition method provided by the application has fast optimization speed, good information extraction effect and high accuracy during model training. The average accuracy is 96.99%, and the recognition accuracy of each working condition is above 95%. At the same time, for the 20 training processes of the model, the average value of the model stability evaluation index calculation result is 0.752, so the stability during model training is good. In summary, the experimental results show that the method provided by the application achieves very high accuracy in the blast furnace working condition recognition task, and can meet the needs of field work, and provides reliable decision support for real-time monitoring of the blast furnace running state for field workers.

[0088] The application embodiment further provides a blast furnace working condition recognition system fusing process variables and video images, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the above method when executing the computer program.

[0089] The blast furnace working condition recognition system fusing process variables and video images can implement each embodiment of the above method and achieve the same beneficial effects, which will not be described here.

[0090] The above describes the preferred embodiments of the application in detail. It should be understood that those skilled in the art can make many modifications and changes without creative labor based on the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment based on the prior art according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A blast furnace operating condition recognition method that fuses process variables and video images, characterized by, The method comprises the following steps: S1: collecting process variables and video images of the material surface of the blast furnace to obtain multi-source heterogeneous information; S2: interpolating the process variables and averaging the video image frames to align the multi-source heterogeneous information to the same time step; S3: constructing a single-layer fusion network architecture; S4: extracting a preliminary feature set of the aligned multi-source heterogeneous information based on the single-layer fusion network architecture; S5: training and optimizing a double-layer fusion network model based on the preliminary feature set and a pre-constructed double-layer fusion network architecture; S6: identifying the blast furnace working condition type based on the double-layer fusion network model. The single-layer fusion network architecture comprises a first deep neural network model, a second deep neural network model, and an information simplifier, the input end of the information simplifier being connected with the output ends of the first deep neural network model and the second deep neural network model respectively. The S4 comprises: inputting the data-aligned process variables X into the first deep neural network model and the data-aligned video images Y into the second deep neural network model to extract features of the data-aligned process variables X and the data-aligned video images Y respectively, and obtaining feature sets F1 and F2 of the two heterogeneous information; calculating the product of the covariance matrix for the data-aligned process variables X and the data-aligned video images Y: wherein∑ XX ,∑ YY and∑ XY are the covariance matrix of the data-aligned process variable X and the cross-covariance matrix of the data-aligned video image Y, respectively, and Y is the product of the covariance matrices, and Y is singular value decomposed, satisfying the following relationship: Y = ΓΣR T ; (2) where Γ,∑ and R are matrix factors after singular value decomposition, J and L are weight parameters, are vector spaces with dimension size lxl, are vector spaces with dimension size mxm, T denotes transpose; calculating the features u1 and v1 simplified by the information simplifier, which satisfy the following relationship: wherein, parameters J1, L1 and Σ1 are weight parameters calculated by the first deep neural network model.

2. The blast furnace condition identification method of claim 1, wherein The interpolation of the process variables comprises: Let Ψ represent the original set of process variables, containing p process variable types and m samples, represent the zth time window sequence of the cth process variable, where c = 1, 2,..., p, and z = 1, 2,..., s, m samples are divided into s time window sequences by a preset window width length; The sample points in each time window sequence are interpolated in sequence, n is the number of new samples interpolated between any two original sample points in the time window sequence i , x i ) represents the i-th original sample point, i = 1, 2,..., q, q represents the total number of original sample points, let point represents the j-th interpolated new sample point between the i-th original sample point and its next adjacent original sample point, where Δt = 1 and j = 1,..., n, and n = ΔT - 1, a new sample is added every second, and ΔT is the sampling interval;​ Computing the time window sequence The Gaussian function that performs the interpolation is modified to K(t) that satisfies the following relation: wherein σ is the bandwidth of the Gaussian function K(t) for interpolation, t is time, and μ is the mean of the Gaussian function K(t) for interpolation. Calculate the value at time point based on the Gaussian function K(t). The size of the new sample point for interpolation at a given location satisfies the following relationship: In the formula, x m is the true value of the mth original data point, t m is the time at which the mth original data point is located, and K is a Gaussian function.

3. The blast furnace condition identification method of claim 1, wherein The averaging of the video image frames comprises: Let num be the number of frames per second, calculate the average value of num frames per second at each pixel position, and get an image F representing one second of time η , which contains the information of all video frame images in the second η .

4. The blast furnace condition identification method of claim 1, wherein The S5 comprises: splicing the features u1 and v1 extracted by the single-layer fusion network architecture with the data-aligned process variables X and the data-aligned video images Y respectively to obtain data sets X' and Y', which satisfy the following relationship: wherein is a data stitching operation; building a new fusion network architecture, which has the same structure as the single-layer fusion network architecture but different parameters, and constructing a double-layer fusion network architecture based on the new fusion network architecture and the single-layer fusion network architecture; extracting features of the data sets X' and Y' in the new fusion network architecture to obtain feature sets F3 and F4 of the two heterogeneous information: wherein, J2, L2 and Σ2 are weight parameters calculated by the second deep neural network model. The feature sets u2 and v2 are spliced, and a blast working condition recognition result is calculated And through the cross-entropy loss function of the following model, the double-layer fusion network model is obtained by model training and parameter optimization: Wherein θ1 and θ2 are model parameter sets of the single-layer fusion network architecture respectively, θ3 and θ4 are model parameter sets of the new fusion network architecture respectively, θ5 is a full connection layer parameter set, The parameter combination when the network achieves the optimal effect, y i represents the true label of the i-th sample, represents the recognition result of the i-th sample.

5. The blast furnace condition identification method of claim 1, wherein The method further comprises: constructing a model training stability evaluation index to determine the optimization speed, accuracy and fluctuation of the double-layer fusion network model based on the model training stability evaluation index; optimizing the training model based on the evaluation result.

6. A blast furnace operating condition identification system that fuses process variables and video images, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the method in any one of claims 1 to 5.

Images