An intelligent detection method for sinter drum strength integrating knowledge and data

Abstract

The present invention discloses an intelligent detection method for sinter drum strength that integrates knowledge and data, and belongs to the field of soft measurement modeling for industrial processes. The present invention utilizes a gated spatiotemporal convolution module to extract the spatiotemporal features of the sintering process from the perspectives of knowledge and data, respectively, and adopts a local preservation projection method to fuse these two types of features, thereby realizing intelligent detection of drum strength driven by knowledge and data. First, expert knowledge and the Gaussian kernel distance formula are used to construct adjacency matrices based on knowledge and data, respectively. Then, the graph convolution and time convolution are fused using the gating idea to build a gated spatiotemporal convolution module; then, a dual-branch network is constructed to extract knowledge-based features and data-based features, respectively. Finally, the local preservation projection method is used to fuse these two types of features and input them to a fully connected layer to realize intelligent detection of drum strength. Data on sinter drum strength were collected in a sintering plant to verify the effectiveness and feasibility of this method.

Description

Technical Field

[0001] The present invention belongs to a soft measurement method for key performance indicators in the process industry field, and specifically relates to an intelligent detection method for sinter drum strength by integrating knowledge and data. Background Art

[0002] The steel industry, a crucial sector for measuring a country's comprehensive strength and economic development, not only underpins its survival and development but also serves as a pillar of the national economy. In the blast furnace ironmaking process, sintered ore is the primary raw material, accounting for approximately 70% of the input. The quality of the molten iron is closely linked to its output and quality. Stable and high-quality sintered ore production also ensures smooth ironmaking. Therefore, as the primary method for obtaining sintered ore, the sintering process is a critical step in the blast furnace ironmaking process.

[0003] As a heavy industry with high energy consumption and emissions, the metallurgical industry faces urgent structural reform and transformation. Reforms in steel companies are progressing rapidly, and advancements in blast furnace ironmaking technology primarily depend on improving the properties of incoming raw materials. Therefore, improving sinter production is an irreversible trend.

[0004] Drum strength, a key physical property of sintered ore, is crucial for the quality and yield of subsequent blast furnace ironmaking. Current approaches to predicting drum strength primarily include numerical simulation models and data-driven models. Numerical simulation models typically establish mathematical models based on thermochemical equations from the sintering process. However, complex thermochemical reactions present significant nonlinearity and strongly coupled uncertainties. These characteristics make mathematical models with strict constraints and assumptions unsuitable for practical production, resulting in low prediction accuracy. With the rise of machine learning and big data, a growing number of researchers are approaching modeling from a data perspective. Umadevi et al. proposed a neural network-based model that correlates drum strength with nine process parameters for prediction. Wang et al. established an online drum strength prediction model based on an Elman neural network. Li et al. selected a subset of operating parameters through clustering and constructed an online sequential extreme value learning machine for prediction. Zou et al. proposed a hybrid ensemble prediction model for drum strength that combines multiple machine learning algorithms. Compared with numerical simulation models, data-driven models improve prediction accuracy and simplify the modeling process.

[0005] Through the analysis of the drum strength prediction problem, it can be seen that this type of prediction problem has two difficulties: (1) The sintering process has the characteristics of multivariate coupling, and the process data presents a non-Euclidean topological structure in space. Most Euclidean space models have no way to explicitly represent the interdependence between variables, which hinders the extraction of potential spatial correlation features. (2) The relationship itself may be diversified, and there is not only one relationship between two variables. The graph structure constructed based on Gaussian kernel distance has limitations. The distance cannot fully reflect the relationship between nodes. It is necessary to integrate expert knowledge to further improve the performance of the model. Summary of the Invention

[0006] In response to the problem of difficulty in detecting drum strength during the sintering process, the present invention proposes an intelligent detection method for sinter drum strength that integrates knowledge and data. The goal of this method is to integrate expert knowledge into graph neural networks to achieve the fusion of knowledge and data. It mainly includes the following four steps: First, based on expert knowledge, a calculation method for the interaction between variables is proposed to capture the complex coupling relationship between variables and construct a knowledge-based adjacency matrix; then, a gated spatiotemporal convolution module is constructed using graph convolution and time convolution to extract potential spatiotemporal features in the data; then, the knowledge-based features and data-based features are fused using the local preserving projection method; finally, the fused features are input to the fully connected layer to realize the prediction of drum strength, and the root mean square loss function is used for gradient back propagation to continuously update the network parameters; the invented method is verified on real data from an actual sintering plant, and the results show that this method has higher accuracy and robustness than other methods.

[0007] The present invention is achieved by adopting the following technical solutions:

[0008] The present invention first provides an intelligent detection method for sinter drum strength by integrating knowledge and data, which comprises the following steps:

[0009] 1) Auxiliary variables related to sinter drum strength were selected, and the sliding window method was used to extract time segments to construct knowledge-based and data-based adjacency matrices.

[0010] 2) Constructing a gated spatiotemporal convolution module to extract spatiotemporal features from the data; the gated spatiotemporal convolution module includes three components: graph convolution, temporal convolution, and gated fusion, and uses the gating concept to fuse graph convolution and temporal convolution;

[0011] 3) A dual-branch network is constructed using a gated spatiotemporal convolutional module. The inputs of the dual-branch network are a knowledge-based adjacency matrix and a data-based adjacency matrix, respectively, to extract knowledge-based features and data-based features.

[0012] 4) The two types of features are fused using the locality-preserving projection method and the fused features are fed into a fully connected layer to build an intelligent drum strength detection model.

[0013] 5) The intelligent detection model is trained using real data from sintering plants and deployed to actual industrial sites to achieve intelligent detection of drum strength.

[0014] As a preferred embodiment of the present invention, in step 1), the knowledge-based adjacency matrix construction method is as follows: first, m auxiliary variables related to the sinter drum strength are selected, and the auxiliary variable time series at any time can be expressed as ; Then the sliding window method is used to extract the time segment. The input sequence of time steps can be expressed as ,in represents the length of the time series; assuming and Represent the time series of the i-th and j-th auxiliary variables, i, j = 1, 2, 3, …, m, respectively. and represents the information entropy of the i-th and j-th variables, represents the distribution distance between two variables, then Represents the attraction between two auxiliary variables, and the specific formula is as follows:

[0015] (1)

[0016] (2)

[0017] (3)

[0018] (4)

[0019] in, represents the probability density distribution of the i-th variable, Represents a regulating factor, which is used to control the attraction between two variables. is a hyperparameter, Indicates the magnitude of the correlation between two variables guided by prior knowledge; Represents the distance between the maximum means of two variables in the Hilbert space. The specific calculation formula is as follows:

[0020] (5)

[0021] Where n represents the number of samples, represents the Gaussian kernel mapping function; then, the final knowledge-based adjacency matrix It can be expressed as:

[0022] (6)

[0023] in, is a threshold;

[0024] Similarly, the Gaussian kernel distance formula can be used to calculate the adjacency matrix based on the data , as follows:

[0025] (7)

[0026] in, Indicates the Euclidean distance between two variables. is the Gaussian kernel hyperparameter.

[0027] As a preferred solution of the present invention, the step 2) is: first, the auxiliary variable time series and the adjacency matrix are input into the graph convolution GCN, then the first Output corresponding to the layer graph convolution :

[0028] (8)

[0029] in, represents the sigmoid activation function, represents the degree matrix, Representative The output of the layer graph convolution, Representative The weight coefficient matrix of the layer graph convolution, Represents the adjacency matrix, which can be a knowledge-based adjacency matrix or an adjacency matrix based on the data ;

[0030] Then through a 1*1 convolution operation, Mapped to two different channels of sub-output and , and input them to two time convolution TCN-a and TCN-b respectively;

[0031] Then the gated fusion method is used to connect GCN and TCN, as shown in the following formula:

[0032] (9)

[0033] (10)

[0034] in, represents the mapping function of TCN-a, represents the mapping function of TCN-b, , and are all learnable parameters. is the gating coefficient, is the matrix dot product, It is The output of the gated convolution module at the layer.

[0035] As a preferred embodiment of the present invention, in step 3), the steps of constructing and extracting features of the dual-branch network are as follows: the two branches of the dual-branch network are composed of a gated spatiotemporal convolution module GSTC, and the input of one branch is based on the knowledge adjacency matrix , the input of the other branch is based on the adjacency matrix of the data , respectively extract knowledge-based features and data-based features , which can be specifically expressed as:

[0036] (11)

[0037] (12)

[0038] Represents the mapping function of the gated spatiotemporal convolution module.

[0039] As a preferred solution of the present invention, in step 4), the calculation steps of the local preservation projection fusion method are as follows: First, the knowledge-based features and data-based features are connected in series to form a large matrix ,in Represents the dimension of the original feature; the local preservation method is to find a projection matrix The original features Mapped into a Matrix, where , Represents the fused dimension, namely:

[0040] (13)

[0041] To solve the projection matrix , the following optimization function needs to be optimized:

[0042] (14)

[0043] (15)

[0044] in, Represents the original features The adjacency matrix of is a hyperparameter;

[0045] Then, the minimum optimization problem can be transformed into the following equation:

[0046] (16)

[0047] in, is the Laplace matrix, and then the singular value decomposition is used to solve the above formula to get the projection matrix ; Finally, the fused features The input is given to a fully connected layer to predict the drum strength.

[0048] The beneficial effects of the present invention are:

[0049] 1. This invention utilizes a graph neural network model to develop an intelligent drum strength detection method that integrates knowledge and data. Previous studies have mostly used purely data-driven models, neglecting the extraction and embedding of knowledge. However, for actual industrial processes, pure data models cannot effectively represent the actual physical and chemical processes. This model combines the domain knowledge of field experts and effectively integrates this knowledge into the graph neural network, resulting in a model that is more consistent with actual industrial production.

[0050] 2. This paper uses a gated spatiotemporal convolution module to extract the complex spatiotemporal coupling characteristics between auxiliary variables. Previous studies have ignored the non-Euclidean topological structure between auxiliary variables in industry. The graph neural network used in this paper excels at processing the complex coupling of non-Euclidean structured data and uses gated connections to fuse graph convolution and temporal convolution to simultaneously capture temporal and spatial correlations in the data.

[0051] 3. The present invention uses a locality-preserving projection method to fuse knowledge-based and data-based features. Previous studies have typically used direct splicing to fuse features, which introduces a lot of redundant information. The present invention uses a locality-preserving projection method to fuse these two types of features, eliminating excess information between them. This allows for efficient fusion and improves the accuracy of drum strength prediction. This method can provide precise intelligent detection strategies for key indicators in other industrial processes, improving the stability and safety of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of an intelligent detection method for sinter drum strength that integrates knowledge and data;

[0053] Figure 2 Schematic diagram of the spatiotemporal characteristics analysis of the sintering process;

[0054] Figure 3Schematic diagram of the gated spatiotemporal convolution module;

[0055] Figure 4 Comparison chart of drum strength prediction results. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are intended to facilitate understanding of the present invention and do not have any limiting effect on the present invention.

[0057] Figure 1 The steps of a drum strength prediction method based on a graph network that integrates knowledge and data are provided, specifically including:

[0058] 1) Use expert knowledge and Gaussian kernel distance formula to construct knowledge-based adjacency matrix and data-based adjacency matrix respectively. The calculation method of knowledge-based adjacency matrix is ​​as follows: First, select m auxiliary variables related to the sinter drum strength. The time series of auxiliary variables at any time can be expressed as ; Then the sliding window method is used to extract the time segment. The input sequence of time steps can be expressed as ,in Represents the length of the time series. Assume and denote the time series of the i-th and j-th auxiliary variables respectively (i, j=1, 2, 3, …, m), and represents the information entropy of the i-th and j-th variables, represents the distribution distance between two variables, then Represents the attraction between two auxiliary variables, and the specific formula is as follows:

[0059] (1)

[0060] (2)

[0061] (3)

[0062] (4)

[0063] in, represents the probability density distribution of the i-th variable, Represents a regulating factor, which is used to control the attraction between two variables. is a hyperparameter, It indicates the size of the correlation between two variables under the guidance of prior knowledge. Represents the distance between the maximum means of two variables in the Hilbert space. The specific calculation formula is as follows:

[0064] (5)

[0065] Where n represents the number of samples, represents the Gaussian kernel mapping function. Then, the final knowledge-based adjacency matrix It can be expressed as:

[0066] (6)

[0067] in, is a threshold value, which is 0.8 in this embodiment. Similarly, the Gaussian kernel distance formula can be used to calculate the adjacency matrix based on the data. , as follows:

[0068] (7)

[0069] in, Indicates the Euclidean distance between two variables. is the Gaussian kernel hyperparameter, In this embodiment, it is taken as 0.8.

[0070] 2) Use the gating idea to fuse graph convolution and temporal convolution to build a gated spatiotemporal convolution module to extract spatiotemporal features from the data. The gated convolution module includes three components: graph convolution, temporal convolution, and gated fusion. First, input the auxiliary variable time series and adjacency matrix into the graph convolution GCN, then we can get the first Output corresponding to the layer graph convolution :

[0071] (8)

[0072] in, represents the sigmoid activation function, represents the degree matrix, Representative The output of the layer graph convolution, Representative The weight coefficient matrix of the layer graph convolution, Represents the adjacency matrix, which can be a knowledge-based adjacency matrix or an adjacency matrix based on the data Then through a 1*1 convolution operation, Mapped to two different channels of sub-output and , and input them to two time convolution TCN-a and TCN-b respectively. Then the gated fusion method is used to connect GCN and TCN, as shown in the following formula:

[0073] (9)

[0074] (10)

[0075] in, represents the mapping function of TCN-a, represents the mapping function of TCN-b, , and are all learnable parameters. is the gating coefficient, is the matrix dot product, It is The output of the gated convolution module at the layer.

[0076] 3) Use the gated spatiotemporal convolution module to construct a dual-branch network. The input of the network is the knowledge-based adjacency matrix and the data-based adjacency matrix, which are used to extract knowledge-based features and data-based features respectively. The steps of constructing and extracting features of the dual-branch network are as follows: Both branches of the dual-branch network are composed of a gated spatiotemporal convolution module GSTC, and the input of one branch is the knowledge-based adjacency matrix , the input of the other branch is based on the adjacency matrix of the data , respectively extract knowledge-based features and data-based features , which can be specifically expressed as:

[0077] (11)

[0078] (12)

[0079] Represents the mapping function of the gated spatiotemporal convolution module.

[0080] 4) The two types of features are fused using the locality preserving projection method and the fused features are fed into a fully connected layer to predict the drum strength. The calculation steps of the locality preserving projection fusion method are as follows: First, the knowledge-based features and data-based features are connected in series to form a large matrix ,in Represents the dimension of the original feature. The local preservation method is to find a projection matrix The original features Mapped into a Matrix, where , Represents the fused dimension, namely:

[0081] (13)

[0082] To solve the projection matrix , the following optimization function needs to be optimized:

[0083] (14)

[0084] (15)

[0085] in, Represents the original features The adjacency matrix of is a hyperparameter.

[0086] Then, the minimum optimization problem can be transformed into the following equation:

[0087] (16)

[0088] in, is the Laplace matrix. Then use singular value decomposition to solve the above formula to get the projection matrix Finally, the fused features The input is given to a fully connected layer to build an intelligent drum strength detection model for predicting drum strength.

[0089] 5) The intelligent detection model is trained using real data from sintering plants and deployed to actual industrial sites to achieve intelligent detection of drum strength.

[0090] The present invention is further described below with reference to specific cases.

[0091] (1) Sintering process description and auxiliary variable determination

[0092] The sintering machine used in this experiment is a bag-type sintering machine with 24 bellows. The machine is divided into north and south sides (for example, south side of bellows No. 1 and north side of bellows No. 1). The specific process includes five steps: batching, mixing, ignition, ventilation sintering, cooling, and screening. First, the raw materials (such as iron ore mixture, coke, and limestone) are mixed in a predetermined ratio. Water is then added to the mixing drum to moisten the material layer and enhance the permeability of the material layer. Next, the mixture is loaded onto a mobile cart by a feeder, and the surface of the mixture is ignited by an ignition system. As the cart moves forward, the mixture gradually burns, with the air in the bellows providing the combustion power. After the sintering process is completed, the sinter is cooled, crushed, and screened. The drum strength of the sintered ore significantly affects the blast furnace ironmaking process. To improve the quality of blast furnace ironmaking, real-time drum strength monitoring is necessary. Because laboratory drum strength testing is very time-consuming, soft sensing methods based on machine learning are generally used in industry for intelligent prediction. Through expert knowledge and Pearson correlation analysis, 23 relevant variables were selected as the input of the model, such as the temperature and negative pressure (south and north sides) of wind box No. 1, as shown in Table 1.

[0093] Table 1 Input variables and output variables of sintering process

[0094]

[0095] (2) Analysis of spatiotemporal characteristics of sintering process

[0096] The sintering process involves many process variables, such as material layer thickness, ignition temperature, trolley speed, etc., and the coupling relationship between them is very complex. Figure 2 As shown in , these variables interact with each other, and the change of a certain variable at the same time will have a direct impact on other variables. For example, the trolley speed is related to the thickness of the material layer, and the thickness of the material layer also affects the trolley speed. In addition, the trolley speed is also affected by the ignition temperature. Based on these observations and analyses, we can find that these process variables have spatial coupling correlation in the sintering process. At the same time, there is also a temporal correlation between the variables at different times, because iron ore sintering is a continuous physical and chemical reaction process. As Figure 2 As shown, adjacent samples are closely correlated, leading to the mutual influence of the same variable at different times. For example, the current trolley speed is related to its previous state, meaning that the combustion process continuously changes as the trolley gradually moves forward. This is because heat transfer and temperature changes are gradual processes that take time. This physical significance can be understood as the time-varying characteristics of chemical reactions during sintering. Therefore, it is necessary to capture the intra-variable temporal correlation in sintering data.

[0097] (1) Dataset construction and experimental setup

[0098] In order to verify the accuracy of the method of the present invention, relevant data of sintered ore were collected from a sintering plant in South China. After data preprocessing and sliding window method, 5000 segments were constructed for training, 1000 segments for validation, and 1000 segments for testing. The model results were calculated using the root mean square error (RMSE). , mean absolute error and hit rate HR (k=0.03) are used to measure the three evaluation indicators, among which, represents the predicted value, Represents the true value, and k represents the relative percentage error between the true and predicted values. To reduce contingency, the evaluation index of all models is the average of 10 results.

[0099] (17)

[0100] (18)

[0101] (19)

[0102] (20)

[0103] (2) Offline model establishment and comparison

[0104] To compare the superiority of the knowledge and data fusion graph network model (KDGNN) developed in this paper, several typical machine learning models were selected for comparison, including partial least squares regression (PLS), random forest (RF), autoencoder (SAE), variable weighted autoencoder (VWSAE), homogeneous autoencoder (SAIE), long short-term memory (LSTM), and spatiotemporal long short-term memory (STA-LSTM). Table 2 shows the comparative results of different prediction models, with the bolded data representing the performance of the proposed KDGNN method. Overall, the proposed KDGNN model achieved the best performance across all evaluation metrics, with RMSE, MAE, and HR values ​​of 0.4351, 0.3356, and 92.40%, respectively. Specifically, the traditional model (PLS) performed the worst due to its linear nature and inability to learn the severe nonlinear characteristics of the sintering process. In contrast, RF and SAE can handle nonlinear data and therefore achieve better performance than PLS. However, the SAE model cannot guarantee the correlation between latent features and the target variable, which negatively impacts prediction performance. VWSAE can extract high-level features that are very relevant to the target by assigning different weights to features, thereby achieving better performance. In addition, the accuracy of SIAE also exceeds that of VWSAE because SIAE can learn excellent features from raw data by stacking layered isomorphic autoencoders. In fact, sintering is a continuous dynamic process, and there is temporal correlation between adjacent samples. Therefore, dynamic networks such as LSTM and VWLSTM perform better than the static models above. However, the limitation of these models is that it is difficult for them to capture potential features from non-Euclidean spaces. That is, these models cannot explicitly model the dependencies between variables. The sintering process has a strong coupling characteristic, and the graph-based KDGNN model of the present invention is good at extracting the potential mutual dependencies between process variables, achieving the best prediction performance.

[0105] Table 2 Comparison of model prediction results

[0106]

[0107] (3) Online model results

[0108] Finally, the model established by the method of the present invention was tested online on an actual sintering machine. Figure 4 The curves showing the actual value, predicted value and error percentage of sinter drum strength are shown respectively. According to the definition of HR, when the error percentage falls within the range of [-1%, 1%], the prediction accuracy of drum strength meets the engineering requirements. Figure 4As shown in (c), the absolute value of the percentage error for most samples is within 1%, and only a few samples are around 2%. Therefore, the online detection experiment shows that the proposed KDGNN model can accurately predict drum strength and meet the needs of practical engineering applications.

[0109] The above-described embodiments merely represent several implementation methods of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. An intelligent detection method for sinter drum strength integrating knowledge and data, characterized in that: The steps include: 1) Auxiliary variables related to sinter drum strength were selected, and the sliding window method was used to extract time segments to construct knowledge-based and data-based adjacency matrices. In step 1), the knowledge-based adjacency matrix construction method is as follows: first, m auxiliary variables related to the sinter drum strength are selected, and the auxiliary variable time series at any time is expressed as ; Then the sliding window method is used to extract the time segment. The input sequence of time steps is represented as ,in represents the length of the time series; assuming and Represent the time series of the i-th and j-th auxiliary variables, i, j = 1, 2, 3, ..., m, respectively. and represents the information entropy of the i-th and j-th variables, represents the distribution distance between two variables, then Represents the attraction between two auxiliary variables, and the specific formula is as follows: （1） （2） （3） （4） in, represents the probability density distribution of the i-th variable, Represents a regulating factor, which is used to control the attraction between two variables. is a hyperparameter, Indicates the magnitude of the correlation between two variables guided by prior knowledge; Represents the distance between the maximum means of two variables in the Hilbert space. The specific calculation formula is as follows: （5） Where n represents the number of samples, represents the Gaussian kernel mapping function; then, the final knowledge-based adjacency matrix Expressed as: （6） in, is a threshold; Similarly, the Gaussian kernel distance formula is used to calculate the adjacency matrix based on the data , as follows: （7） in, Indicates the Euclidean distance between two variables. is the Gaussian kernel hyperparameter; 2) Constructing a gated spatiotemporal convolution module to extract spatiotemporal features from the data; the gated spatiotemporal convolution module includes three components: graph convolution, temporal convolution, and gated fusion, and uses the gating concept to fuse graph convolution and temporal convolution; 3) A dual-branch network is constructed using a gated spatiotemporal convolutional module. The inputs of the dual-branch network are a knowledge-based adjacency matrix and a data-based adjacency matrix, respectively, to extract knowledge-based features and data-based features. 4) The two types of features are fused using the locality-preserving projection method and the fused features are fed into a fully connected layer to build an intelligent drum strength detection model. 5) The intelligent detection model is trained using real data from sintering plants and deployed to actual industrial sites to achieve intelligent detection of drum strength.

2. The intelligent detection method for sinter drum strength integrating knowledge and data according to claim 1 is characterized in that: The step 2) is: first, the auxiliary variable time series and adjacency matrix are input to the graph convolution GCN, then the first Output corresponding to the layer graph convolution : （8） in, represents the sigmoid activation function, represents the degree matrix, Representative The output of the layer graph convolution, Representative The weight coefficient matrix of the layer graph convolution, Represents the adjacency matrix, which is a knowledge-based adjacency matrix or an adjacency matrix based on the data ; Then through a 1*1 convolution operation, Mapped to two different channels of sub-output and , and input them to two time convolution TCN-a and TCN-b respectively; Then the gated fusion method is used to connect GCN and TCN, as shown in the following formula: （9） （10） in, represents the mapping function of TCN-a, represents the mapping function of TCN-b, , and are all learnable parameters. is the gating coefficient, is the matrix dot product, It is The output of the gated convolution module at the layer.

3. The intelligent detection method for sinter drum strength integrating knowledge and data according to claim 1 is characterized in that: In step 3), the steps of building and extracting features of the dual-branch network are as follows: Both branches of the dual-branch network are composed of a gated spatiotemporal convolution module GSTC, and the input of one branch is the knowledge-based adjacency matrix , the input of the other branch is based on the adjacency matrix of the data , respectively extract knowledge-based features and data-based features , specifically expressed as: （11） （12） Represents the mapping function of the gated spatiotemporal convolution module.

4. The intelligent detection method for sinter drum strength integrating knowledge and data according to claim 1 is characterized in that: In step 4), the calculation steps of the local preserving projection fusion method are as follows: First, the knowledge-based features and data-based features are connected in series to form a large matrix ,in Represents the dimension of the original feature; the local preservation method is to find a projection matrix The original features Mapped into a Matrix, where , Represents the fused dimension, namely: （13） To solve the projection matrix , the following optimization function needs to be optimized: （14） （15） in, Represents the original features The adjacency matrix of is a hyperparameter; Then, the minimum optimization problem is transformed into the following equation: （16） in, is the Laplace matrix, and then the singular value decomposition is used to solve the above formula to obtain the projection matrix ; Finally, the fused features The input is given to a fully connected layer to predict the drum strength.

Images