A sintering endpoint intelligent perception method based on a 3D convolution network

Abstract

The application discloses a sintering endpoint intelligent sensing method based on a 3D convolution network and belongs to the field of industrial process soft measurement modeling. The prediction model utilizes a space-time 3D convolution network to simultaneously learn space-time features hidden in data, and adopts an encoding-decoding network to perform multi-step prediction on a present sintering endpoint. Firstly, a random permutation and combination method is used to convert a two-dimensional data structure into a three-dimensional format. Then, a 3D convolution is used to extract time features and space features in the data. Next, a space-time feature calibration module is proposed to learn the importance of different features and finely process the features. Finally, the fine space-time features are input into the encoding-decoding network, and a dynamic multi-step prediction loss function is used to realize multi-step prediction on the sintering endpoint, so that the purpose of intelligent sensing is achieved. Related data collected in a sintering plant in South China verifies the effectiveness and feasibility of the method.

Description

Technical Field

[0001] This invention relates to a soft measurement method for intelligent sensing of the sintering endpoint in the sintering process, specifically a method for intelligent sensing of the sintering endpoint based on a 3D convolutional network. Background Technology

[0002] Sintering is one of the main production methods for artificial bulk raw materials. It involves heating powdered materials (such as ore powder, fine granules, and concentrates) at high temperatures and sintering them into lumps under conditions of incomplete melting. The product obtained from sintering is called sintered ore. Sintering technology is widely used in the metal smelting industries of iron and steel, lead, lead-zinc, and copper. As an important process in pyrometallurgical processes, it directly affects the success of subsequent smelting processes and influences yield and quality indicators. Currently, competition among steel enterprises is increasingly fierce, and improvements in the technical and economic indicators and technological advancements in blast furnace ironmaking mainly rely on improvements in the properties of the raw materials fed into the furnace.

[0003] The burning through point (BTP) is the location of the bellows corresponding to the point where the sintered material layer burns through from top to bottom. The BTP reflects the quality of the current sintering process. If the BTP deviates from the preset value, incomplete burning or over-burning will occur, affecting not only the yield of sintered ore but also damaging the sintering machine, increasing maintenance costs and reducing its operating efficiency. Controlling the sintering process essentially involves controlling the BTP, ensuring it stabilizes at the second-to-last bellows position. However, due to the influence of various substances during the reaction, parameters fluctuate greatly, and random disturbances are difficult to predict, making predictive control of the BTP a major challenge in the sintering process. Currently, sintering plants still rely heavily on manual monitoring to control the BTP. However, manual operation depends excessively on individual experience, and employee changes can easily lead to problems. Furthermore, individual operation is inherently unpredictable, easily resulting in unsatisfactory control of the BTP.

[0004] Therefore, developing an intelligent sensing method for the sintering endpoint is of great significance for intelligent control of the sintering endpoint. Intelligent sensing of the sintering endpoint position not only ensures the stable operation of each stage of the sintering process, preventing interference between upstream and downstream processes and ensuring efficient operation, but also ensures effective utilization of the sintering area, increasing sinter yield and saving energy. Based on real-time process parameters, status parameters, and operating parameters of the sintering process, the sintering process can be rationally judged, and the sintering endpoint position can be accurately sensed to adjust the speed of the sintering machine trolley, achieving the goal of stabilizing the sintering endpoint, reducing fluctuations in the endpoint position, and improving the control of sinter yield and quality. Timely sensing of the sintering endpoint is of great importance for the rational use of existing sintering equipment, stabilizing sintering production, promoting the improvement of sintering control levels, and increasing the economic benefits of the sintering plant. Summary of the Invention

[0005] This invention addresses the challenge of intelligently sensing the endpoint of the sintering process by introducing the currently popular 3D convolutional neural network into the sintering field, proposing an intelligent sensing method for the sintering endpoint based on a 3D convolutional network. The method mainly includes the following four steps: First, auxiliary variables related to BTP are identified through mechanistic analysis, and a random permutation and combination method is used to convert the two-dimensional time series into a three-dimensional data format; then, a 3D convolutional layer is constructed to simultaneously extract temporal and spatial features from the data, and a spatiotemporal feature calibration module is used to refine each feature; subsequently, the refined spatiotemporal features are input into an encoding and decoding network to achieve multi-step prediction of BTP; then, a dynamic multi-step prediction loss function is constructed to train the entire network; finally, the model parameters are updated and uploaded to the sintering plant system platform for operation.

[0006] This invention is achieved using the following technical solution:

[0007] This invention first provides a method for intelligent sensing of sintering endpoint based on 3D convolutional networks, which includes the following steps:

[0008] 1) Based on the mechanism analysis, the auxiliary variables related to the sintering endpoint were determined. The waste gas temperature fitting method was used to calculate BTP, and the sliding window method was used to divide the data to construct a two-dimensional time series. The two-dimensional time series was converted into a three-dimensional data format by random permutation and combination method.

[0009] 2) Use 3D convolutional networks to learn the temporal and spatial features of 3D data, and use the spatiotemporal feature calibration module to calculate the importance of different spatiotemporal features;

[0010] 3) Construct an encoding and decoding network based on a gated recurrent neural network to perform multi-step prediction of the sintering endpoint;

[0011] 4) A BTP prediction model is constructed using a 3D convolutional network, a spatiotemporal feature calibration module, and an encoder-decoder network. A dynamic multi-step prediction loss function is used to measure the difference between the predicted sequence and the real sequence, and the Adam optimizer is used to train the loss function.

[0012] 5) Collect auxiliary variable data related to the sintering endpoint from the sintering plant and preprocess them, then use the established BTP prediction model to intelligently sense the sintering endpoint.

[0013] As a preferred embodiment of the present invention, in step 1), the auxiliary variables are selected as: neutralized ore ratio, quicklime ratio, limestone ratio, dolomite water ratio, secondary mixed water content, material thickness, ignition temperature, main extraction negative pressure, trolley speed, flue gas temperature and the position of temperature rise point.

[0014] As a preferred embodiment of the present invention, in step 1), the method of converting the two-dimensional time series into a three-dimensional data format using a random permutation and combination method is as follows: First, assume that there are m auxiliary variables related to the sintering endpoint, and use these auxiliary variables as input variables; then, the input variables at any time t are expressed as follows: Then, these auxiliary variables are randomly permuted to obtain m! combinations; next, m combinations are randomly selected to construct a 2D tensor; finally, the 2D tensor is converted into a 3D data format along the time axis, that is, the two-dimensional time series is converted into a 3D data format.

[0015] As a preferred embodiment of the present invention, in step 3), the spatiotemporal feature calibration module is constructed as follows: the module comprises two parts: a time calibration module and a spatial calibration module. The time calibration module decomposes the original multi-channel spatiotemporal features into single-channel feature maps, employs a multi-channel attention mechanism to learn the feature importance at different times, and finally aggregates the attention scores of different channels. The spatial feature calibration module first uses a special 3D convolution operation to compress the spatiotemporal feature map to a single-channel dimension, then aggregates the features using average pooling and max pooling respectively, and finally fuses them using 2D convolution. The refined features are then input into the encoder-decoder network to achieve multi-step prediction.

[0016] As a preferred embodiment of the present invention, in step 4), the distance calculation method based on the autocorrelation coefficient is specifically as follows: firstly, the mean and variance of the predicted sequence and the true value sequence are calculated respectively, the autocorrelation coefficient of each sequence is obtained, a vector of autocorrelation coefficients is obtained for each time series, and then the Euclidean distance between the autocorrelation coefficient vectors of the two time series is calculated.

[0017] Preferably, the parameters of the BTP prediction model are adjusted in real time based on real-time data during the sintering process, and continuously optimized and iterated, so that the model has strong robustness.

[0018] The beneficial effects of this invention are as follows:

[0019] 1. This invention utilizes 3D convolutional networks to simultaneously extract temporal and spatial features from data, overcoming the information interaction loss problem caused by existing technologies that use two separate modules to learn temporal and spatial features separately, thus enabling more effective extraction of the spatiotemporal features of data.

[0020] 2. This invention uses a spatiotemporal feature calibration module to calculate the importance of spatiotemporal features extracted by 3D convolution. A set of weight coefficients is learned using both the temporal and spatial feature calibration modules to characterize the importance of these features. Since different spatiotemporal features have varying degrees of importance in predicting the sintering endpoint, this method can adaptively learn the importance of spatiotemporal features using a neural model, accelerating model convergence and improving model accuracy.

[0021] 3. This invention employs a dynamic multi-step prediction loss function to measure the shape and temporal differences between the predicted and true sequences. Traditional Euclidean distance only considers the shape differences between the two sequences and cannot calculate temporal differences. Therefore, compared to traditional Euclidean distance, the dynamic multi-step prediction loss function can better learn the distributional differences between the predicted and true sequences and continuously update the model parameters through the backpropagation method of the neural network, thereby continuously reducing the error between the predicted and true sequences. Attached Figure Description

[0022] Figure 1 This is a diagram of a sintering endpoint intelligent sensing model based on a 3D convolutional network.

[0023] Figure 2 The graph shows the temporal and spatial correlation analysis of the sintering data;

[0024] Figure 3 This is a schematic diagram of 3D convolution calculation;

[0025] Figure 4 This is a schematic diagram of the spatiotemporal feature calibration module;

[0026] Figure 5 This is a comparison chart of sintering endpoint prediction results based on 3D convolutional networks. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0028] Figure 1 The paper provides specific steps for constructing and applying a sintering endpoint prediction model based on 3D convolutional networks.

[0029] 1) Auxiliary variables related to the sintering endpoint were selected through mechanistic analysis, and the data were preprocessed. Then, the waste gas temperature fitting method was used to calculate BTP, and a dataset was constructed using the sliding window method, where the length of the historical time segment of the sliding window is t. h The length of the future time segment is t f ;

[0030] 2) Construct a 3D data format using random permutation and combination, build a 3D convolutional network model to extract spatiotemporal features from the data, and use a spatiotemporal feature extraction module to learn the importance of each feature, that is, refine the features and pass them to the encoding and decoding network for multi-step prediction.

[0031] 3) Construct a GRU-based encoding and decoding network model to perform multi-step prediction of the sintering endpoint;

[0032] 4) A BTP prediction model is constructed using a 3D convolutional network, a spatiotemporal feature calibration module, and an encoder-decoder network. A dynamic multi-step prediction loss function is used to measure the difference between the predicted sequence and the real sequence, and the Adam optimizer is used to train the loss function.

[0033] 5) Relevant sintering data were collected from a steel plant in South China, and the BTP prediction model was trained using historical data and then tested online on new data.

[0034] The present invention will be further explained below with reference to specific examples.

[0035] (1) Analysis of sintering mechanism and characteristics

[0036] Based on the sintering state of the material, the material on the sintering machine trolley can be divided into four layers: the wet layer (W layer), the preheating layer (P layer), the burning layer (B layer), and the sintered product layer (S layer). During the sintering process, considering a single strip of sintered material on a single trolley, the temperature change is not significant from the bottom to the surface of the wet layer. Upon entering the preheating layer, the temperature rise rate accelerates, resulting in an inflection point between the wet and preheating layers. Upon entering the burning layer, the material composition changes significantly; as the material moves away from the burning surface, its temperature decreases, and it cools to become the sintered layer. This temperature change pattern is largely consistent with the side plate temperature measurement curve; therefore, the calculated temperature curve of the trolley side plate is used to divide the material on the sintering machine trolley into layers.

[0037] The sintering process has the following characteristics due to its large number of complex steps.

[0038] (1.1) Strong nonlinearity

[0039] The sintering process involves complex physicochemical reactions in the mixture, such as the evaporation and condensation of moisture, fuel combustion loss, solvent decomposition, and redox reactions of iron-containing raw materials. The various parameters in the sintering process are highly coupled, exhibiting strong nonlinear relationships with operating parameters, state parameters, batching parameters, and index parameters. These nonlinear relationships are difficult to describe using traditional linear models, necessitating the construction of sinter quality prediction models using modeling methods with strong nonlinear approximation capabilities.

[0040] (1.2) Dynamism

[0041] Due to the unique physicochemical reactions involved in the sintering process, sintering process data exhibits significant dynamic characteristics. This dynamic nature is primarily caused by factors such as changes in sintering raw materials, equipment malfunctions, the inherent properties of the chemical reactions, and feedback control systems. Taking the input and output variables of the model as an example, the dynamic characteristics of the data are mainly manifested in the following ways: the output variable value at a certain moment is not only related to the current input variable value but also to the input variable values ​​at historical moments; for highly dynamic industrial processes, the output variable also shows a strong correlation with its own historical data.

[0042] (1.3) Many factors affect the quality of sintered ore.

[0043] The sintering process involves numerous variables, including batching parameters such as the iron grade and proportion of the neutralizing powder, the chemical composition of the mixture, and the solvent-fuel ratio; operational parameters such as counterweight, primary mixing water addition, secondary mixing water addition, trolley speed, ignition temperature, and bed thickness; equipment parameters such as main pipe negative pressure, exhaust gas temperature, sintering endpoint location, sintering endpoint temperature, and bed permeability; index parameters such as exhaust area and trolley leakage rate; and parameters such as sinter iron grade, basicity, drum index, sinter reducibility, screening index, and trolley utilization coefficient. All these variables directly or indirectly affect the quality of the sinter.

[0044] (1.4) The coupling relationship between influencing factors is complex.

[0045] The sintering process involves numerous variables with complex and highly coupled relationships. For example, the iron grade of the sintered ore is related to the iron grade of the mixture, which in turn is determined by the iron grade of the neutralizing powder. The iron grade of the sintered ore is also related to process states such as the sintering endpoint, ignition temperature, bed thickness, and bed permeability. Bed permeability is a state parameter, which is also related to the properties of each individual iron ore powder, the chemical composition of the neutralizing powder, and the moisture content of the mixture. The amount of water added in the primary and secondary mixing processes are operational parameters, which are influenced by the composition and moisture content of various raw materials. These complex coupling relationships between sintering process parameters are difficult to describe qualitatively or quantitatively using simple mathematical models. Therefore, suitable modeling methods are needed to describe the relationships between index parameters, process state parameters, operational parameters, and batching parameters.

[0046] (2) Construction of variables in the sintering process

[0047] Based on mechanistic analysis and expert knowledge, 12 auxiliary variables related to the sintering endpoint BTP were selected, as shown in Table 1.

[0048] Table 1 Model Input Parameters and Sintering Endpoint Location

[0049]

[0050] (3) Calculation of BTP

[0051] The three-point method from the wind box exhaust gas temperature method is used to achieve soft measurement of the sintering endpoint. The specific steps of the three-point method are as follows: Substitute the positions and temperature points (x1, T1), (x2, T2), and (x3, T3) of three adjacent wind boxes, including the highest temperature point, into the formula:

[0052] T = a + bx + cx 2 (1)

[0053] Considering that x2 = x1 + 1 and x3 = x2 + 1, we can obtain...

[0054]

[0055] Organize, get

[0056]

[0057] Subtracting the two equations, we get the solution.

[0058]

[0059] Substituting c into the equation to find b, we get...

[0060] b = T2 - T1 - 2cx² + c (5)

[0061]

[0062] Substitute the obtained a, b, and c into the above formula to find the peak value, and let

[0063]

[0064] Solving

[0065]

[0066] The sintering machine in this embodiment has 24 air boxes. To calculate the endpoint, the highest temperature point in each air box and its two adjacent points need to be identified and curve-fitted. The extreme point is the sintering endpoint. If the highest measured temperature point falls in air box number 23, the exhaust gas temperatures of the last three air boxes are taken, and the sintering endpoint is calculated using the three-point method. The sintering endpoint position x is then determined. BTP The calculation formula is:

[0067]

[0068] In industrial settings, severe air leakage at the tail end of sintering machinery can sometimes cause the measured exhaust gas temperature in the last air box to be lower than the actual value. This results in an extreme value after curve fitting, even though the material layer on the trolley may not be fully burned. To address this, this study introduces the exhaust gas temperature from the main flue to provide feedback correction for the BTP (Body Pressure Measurement) judgment value, ensuring the accuracy of the soft measurement model. Using the exhaust gas temperature from the main flue under normal thermal conditions and at a suitable sintering endpoint as the standard value, the deviation ΔT between the measured exhaust gas temperature and the standard value of the exhaust gas humidity in the main flue is calculated, weighted, and then expressed using the following formula:

[0069] BTP m =BTP0-αΔT (10)

[0070] The sintering endpoint determination value BTP0 is corrected to obtain the corrected sintering endpoint value BTP. m , where α is the weighting factor, which is usually taken as 0.02.

[0071] (4) Building a 3D convolutional network model

[0072] First, a permutation and combination approach is used to construct the 3D data format. Specifically, assuming m input variables are sampled at time t, these variables are randomly permuted, and m combinations are randomly selected to construct a 2D tensor. Finally, the 2D tensor is converted into a 3D data format along the time axis. During the data construction process, two types of correlations need to be considered in the modeling task: the temporal correlation between different samples and the spatial correlation between different variables, such as... Figure 2 As shown. How to simultaneously capture spatiotemporal correlations is a key issue. Recently, cutting-edge research has developed 3D convolutional neural networks (3DCNNs) to learn the spatiotemporal features of videos. Furthermore, the 3D tensor data constructed by this method exhibits strong similarity to video frames. Therefore, 3D convolution is used to model the complex spatiotemporal correlations of the sintering process. A 3D tensor can be understood as a 4D input with one channel, such as... Figure 3 As shown. Assume Let C represent the input feature map of the k-th 3D convolutional layer. (k-1) Let be the number of channels in the convolution, then the corresponding output is expressed as:

[0073]

[0074] Where * represents a 3D convolution operation, and φ is the sigmoid activation function. and These are the weights and biases that this convolutional layer needs to learn, where L represents the number of 3D convolutional layers, and T... h W represents the length of the historical sliding window. (k) and H (k)This represents the length and width of the feature map at layer k. Following this method, L convolutional layers are stacked to extract the spatiotemporal features of the data. The specific construction method is as follows: Figure 1 As shown.

[0075] After 3D convolution operations, the latent spatiotemporal features have been fully learned. However, the importance of these features differs in predictive modeling across the temporal and spatial domains. Therefore, to improve the expressive power of feature importance, a spatiotemporal feature calibration module was designed, enabling the entire network to explicitly learn the interdependencies of these features. Figure 4 As shown in (a), the spatiotemporal feature calibration module designed in this invention includes a temporal feature calibration module and a spatial feature calibration module. These two modules learn the importance of features from the temporal and spatial dimensions, respectively. Given a feature map... As input, R t and R s These represent the time feature calibration module and the spatial feature calibration module, respectively. The overall learning process is as follows:

[0076]

[0077] Where ⊙ represents the Hadama product. and F out These represent the outputs of the time feature calibration module and the spatial feature calibration module, respectively.

[0078] Temporal Feature Calibration Model: This module primarily helps the network learn which features are more important at which moment and assign them greater weight. For example... Figure 4 As shown in (b), the original multi-channel feature map is first... Decomposed into single-channel feature maps along the channel dimension Where W and H represent the length and width of a single-channel feature map, respectively, and then a dot product attention mechanism is used to calculate the importance of each feature. Assume Q... i K i and V i These represent the projection transformations of the i-th channel,

[0079]

[0080] Where, ω Q ω k and ω v These represent the weight parameters of the three projection transformations. The Attention score mechanism formula is as follows:

[0081]

[0082] in, This is the scaling factor, and softmax represents the exponential normalization function. Finally, the attention scores of all channels are combined, which can be obtained through another projection transformation.

[0083]

[0084] Here, Concat represents a merge operation, and channel represents a channel.

[0085] Spatial Feature Calibration Model: This module primarily generates a feature importance map to measure the dependencies between different variables. For example... Figure 4 As shown in (c), the original feature map is first converted into a single-channel feature map through a special 3D convolution transformation. Then, these features are aggregated using both max pooling and average pooling methods. Finally, the results of the two aggregation methods are fused, and the importance of different features is calculated using a 2D convolution. The final output can be expressed as:

[0086]

[0087] Where σ represents the Sigmoid activation function, f 1×1 It is a 2D convolution operation. This represents the result after average pooling. This represents the result after the max pooling operation. This represents the final refined feature map.

[0088] (5) Building a multi-step prediction model

[0089] Next, the refined features are fed into an encoder-decoder network based on a gated recurrent neural network. This network consists of two layers of gated recurrent units and uses an attention mechanism to pass the output of the encoding part to the decoding part, achieving multi-step prediction of the sintering endpoint. To better measure the difference between the predicted sequence and the real sequence, this invention employs a dynamic multi-step prediction loss function. It is assumed that the real sequence can be represented as... Then the autocorrelation coefficient ρ of the sequence at order h is... h,y It can be formalized as:

[0090]

[0091] Where μ and θ represent the mean and variance of the true sequence, respectively, h represents the order of the autocorrelation coefficient, and T f This represents the length of the predicted segment. Similarly, the autocorrelation coefficient vector of the predicted sequence can be obtained. Therefore, the autocorrelation coefficient ρ of the true value sequence... Y And the autocorrelation coefficient vector of the predicted sequence They can be represented as follows:

[0092] ρ Y =(ρ 1,y ,ρ 2,y ,...,ρ δ,y ) T (18)

[0093]

[0094] Where δ represents the length of the vector. Therefore, the distance formula based on the autocorrelation coefficient can be expressed as:

[0095]

[0096] in, This represents the autocorrelation distance between the true value sequence and the predicted sequence. Therefore, the final dynamic multi-step prediction loss function, Loss, can be expressed as:

[0097]

[0098] in, The distance between the true and predicted sequences is represented by λ, which is the balance coefficient used to measure the dynamics.

[0099] (6) Model performance verification

[0100] To verify the effectiveness of the model, 8500 samples were collected from a sintering plant at a 1-minute interval. After obtaining sample fragments using a sliding window, the data fragments were preprocessed and divided into 7000 training samples, 1000 validation samples, and 500 test samples. Through repeated experiments, the key parameters of the deep learning model were determined as follows: batch size of 20 samples, learning rate of 0.001, input stride and output stride of 40 and 5 respectively, number of channels in the 3D convolutional network of 3, 6, 9, and 12, and kernel size of 3×3×3.

[0101] To verify the effectiveness of the proposed method, Integrated Autoregressive Model (ARIMA), Gated Recurrent Network (GRU), Temporal Convolutional Network (TCN), and Long Short-Term Memory Network (LSTM) were selected as comparison models. The evaluation metrics were: hit rate (HR), root mean square error (RMSE), and mean absolute error (MAE). Table 3 shows the prediction results of different models. We can observe that the proposed method achieves the best performance in BTP prediction, outperforming other baseline models in all three metrics. The traditional ARIMA method, relying solely on historical target variable values ​​for prediction without considering external factors, fails to achieve good performance. In contrast, the three deep learning-based time series models (LSTM, GRU, and TCN) can learn temporal correlations, thus performing better, especially LSTM. However, these three models can only learn temporal correlations in the data and cannot extract spatial correlations. The proposed 3D convolutional network can capture both temporal and spatial correlations simultaneously, therefore outperforming other models in BTP prediction. Figure 4 The prediction results also show that the proposed method (CBMP) matches the actual BTP values ​​very well, which meets the needs of practical engineering and provides guidance for on-site sintering operators. If the sintering endpoint is predicted earlier, the operator should adjust the sintering machine speed appropriately to slow down the sintering process; if the endpoint is predicted later, the sintering machine speed should be increased to bring the endpoint position as close as possible to the penultimate bellows position. Therefore, accurate prediction of the sintering endpoint helps maintain the efficient and stable operation of the sintering process.

[0102] Table 3 Comparison of Model Prediction Results

[0103]

[0104] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for intelligent sensing of sintering endpoint based on 3D convolutional networks, characterized in that, Includes the following steps: 1) Based on the mechanism analysis, auxiliary variables related to the sintering endpoint were determined. The waste gas temperature fitting method was used to calculate BTP, and the sliding window method was used to divide the data to construct a two-dimensional time series. The two-dimensional time series was converted into a three-dimensional data format using a random permutation and combination method. The auxiliary variables related to the sintering endpoint include: neutralized ore ratio, quicklime ratio, limestone ratio, dolomite water ratio, secondary mixed water content, material thickness, ignition temperature, main extraction negative pressure, trolley speed, flue gas temperature, and the location of the temperature rise point. 2) Utilize 3D convolutional networks to learn the temporal and spatial features of 3D data, and use a spatiotemporal feature calibration module to calculate the importance of different spatiotemporal features; specifically: First, a 4-layer 3D convolutional network is constructed, with each layer having 3, 6, 9, and 12 channels, and the kernel size is set to [size missing]. ; Then, a spatiotemporal feature calibration module is constructed to learn the importance of the proposed spatiotemporal features; assuming This represents the input feature map of the k-th 3D convolutional layer. Let be the number of channels in the convolution, then the corresponding output is expressed as: ； in, This is a 3D convolution operation. It is the Sigmoid activation function. and These are the weights and biases that the convolutional layer needs to learn, where L represents the number of 3D convolutional layers. Represents the length of the historical sliding window. and This represents the length and width of the feature map at layer k; in this way, L convolutional layers are stacked to extract the spatiotemporal features of the data; The spatiotemporal feature calibration module includes a time feature calibration module and a spatial feature calibration module. Learning the importance of features from both temporal and spatial dimensions; given a feature map As input, and These represent the temporal feature calibration module and the spatial feature calibration module, respectively; the overall learning process is as follows: ， ； in, Represents the Hadama product. and These represent the outputs of the time feature calibration module and the spatial feature calibration module, respectively. 3) Construct an encoding and decoding network based on a gated recurrent neural network to perform multi-step prediction of the sintering endpoint; the encoding and decoding network based on the gated recurrent neural network consists of two layers of gated recurrent units, and the output of the encoding part is passed to the decoding part through an attention mechanism to realize multi-step prediction of the sintering endpoint. 4) A BTP prediction model is constructed using a 3D convolutional network, a spatiotemporal feature calibration module, and an encoder-decoder network. A dynamic multi-step prediction loss function is used to measure the difference between the predicted sequence and the real sequence, and the Adam optimizer is used to train the loss function. 5) Collect auxiliary variable data related to the sintering endpoint from the sintering plant and preprocess them, then use the established BTP prediction model to intelligently sense the sintering endpoint.

2. The intelligent sensing method for sintering endpoint based on 3D convolutional networks according to claim 1, characterized in that, In step 1), the two-dimensional time series is converted into a three-dimensional data format using a random permutation and combination method as follows: First, assume there are m auxiliary variables related to the sintering endpoint, and use these auxiliary variables as input variables; then, the input variables at any time t are expressed as follows: Then, these auxiliary variables are randomly permuted to obtain... There are m possible combinations; then, m combinations are randomly selected to construct a 2D tensor; finally, the 2D tensor is converted into a 3D data format along the time axis, that is, the two-dimensional time series is converted into a 3D data format.

3. The intelligent sensing method for sintering endpoint based on 3D convolutional networks according to claim 1, characterized in that, The aforementioned time feature calibration model first converts the original multi-channel feature map... Decomposed into single-channel feature maps along the channel dimension ,in and Let x and y represent the length and width of the single-channel feature map, respectively. Then, a dot product attention mechanism is used to calculate the importance of each feature; assuming... , and These represent the projection transformations of the i-th channel, i.e. ； in, , and These represent the weight parameters of the three projection transformations; the Attention score mechanism formula is as follows: ； in, This is the scaling factor, and softmax represents the exponential normalization function; finally, the attention scores of all channels are combined and obtained through another projection transformation. ； Here, Concat represents the merge operation, channel represents the channel, and c represents the number of channels.

4. The intelligent sensing method for sintering endpoint based on 3D convolutional networks according to claim 1, characterized in that, The spatial feature calibration model described above first uses a convolution kernel with a size of The 3D convolution operation converts the original feature map into a single-channel feature map. Then, max pooling and average pooling are used to aggregate these single-channel feature maps. Finally, the results of the two aggregation methods are fused, and the importance of different features is calculated using a 2D convolution. The final output is represented as follows: ； in, Represents the Sigmoid activation function. It is a 2D convolution operation. This represents the output after average pooling. This represents the output after the max pooling operation. This represents the final refined feature map.

5. The intelligent sensing method for sintering endpoint based on 3D convolutional networks according to claim 1, characterized in that, The specific construction method of the dynamic multi-step prediction loss function in step 4) is as follows: Assume the true sequence is represented as Then the autocorrelation coefficient of the sequence at order h is... Formulated as: ； in, and Let represent the mean and variance of the true sequence, respectively, and h represent the order of the autocorrelation coefficient. This represents the length of the predicted segment; similarly, the autocorrelation coefficient vector of the predicted sequence is obtained; therefore, the autocorrelation coefficient of the true value sequence... And the autocorrelation coefficient vector of the predicted sequence They are represented as follows: ； ； in, Let represent the length of the vector; then the distance formula based on the autocorrelation coefficient is expressed as: ； in, The autocorrelation distance between the true sequence and the predicted sequence is represented by: The final dynamic multi-step prediction loss function, Loss, is expressed as: ； in, Represents the Euclidean distance between the true sequence and the predicted sequence. This represents the balance coefficient, used to measure the magnitude of dynamics.

Images