Gated circulation unit and attention mechanism-based emission spectrum tomography reconstruction prediction method

Abstract

The invention relates to an emission spectrum chromatography reconstruction prediction method based on a gated circulation unit and an attention mechanism, which comprises the following steps: taking three-dimensional structure distribution of CH * free radicals in a combustion field reconstructed by a traditional chemiluminescence calculation chromatography algorithm as a true value of a deep learning neural network for comparison; a novel neural network model is designed by combining the advantages that a gating circulation unit processes a long-distance dependency relationship and an attention mechanism adjusts importance weights, and three-dimensional reconstruction and prediction of a combustion field at future moments are achieved according to multi-direction two-dimensional projection at historical moments. The method does not need to design a complex acquisition device and does not need to solve a partial differential equation set. According to the multi-direction projection data at the historical moment, the combustion evolution at the future moment can be quickly and accurately predicted within a certain range. And when a frame insertion problem between continuous time sequences is processed, a multi-frame three-dimensional structure can be accurately predicted by means of multi-direction projection at a front moment and a back moment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computational optical imaging, in particular to a method for predicting emission spectrum tomography reconstruction based on a gated recurrent unit and an attention mechanism. BACKGROUND

[0002] Emission spectrum tomography is often used to reconstruct the three-dimensional structure of the combustion field. By capturing the chemiluminescence signal, the three-dimensional data is reconstructed by detecting multiple perspectives around the target for projection measurement using tomography technology, which is of great significance in the fields of combustion process research and pollutant generation mechanism analysis. However, the traditional iterative reconstruction algorithm has large amount of calculation data and low calculation efficiency, and many studies aim to improve the efficiency of combustion field reconstruction. With the rapid development of deep learning, it brings more convenience to the field of three-dimensional reconstruction, significantly improves the efficiency of data processing, and reduces the experimental cost. The improvement of computational efficiency also makes real-time prediction of three-dimensional reconstruction possible.

[0003] In 2019, a research team from Shanghai Jiaotong University combined time-resolved volume tomography with deep learning and designed a hybrid CNN-LSTM model to achieve online prediction of three-dimensional flame evolution in two steps. The reconstruction prediction method uses a convolutional neural network to extract features, predicts future time features based on historical time features, and reconstructs three-dimensional data results at future time through other neural networks. Through multiple steps, the prediction of simulation and experimental data is realized, the model is relatively complex, and the calculation cost is high. SUMMARY

[0004] In view of the problems of high model complexity and high calculation cost in the three-dimensional reconstruction of the combustion field, the present application provides a method for predicting emission spectrum tomography reconstruction based on a gated recurrent unit and an attention mechanism, which can realize fast and high-precision reconstruction prediction of the combustion field in the future time and inter-frame prediction.

[0005] The technical solution of the present application is as follows:

[0006] A method for predicting emission spectrum tomography reconstruction based on a gated recurrent unit and an attention mechanism, characterized in that it comprises:

[0007] Obtaining continuous time multi-directional two-dimensional projection data of a detection area collected by a flame emission device;

[0008] Obtaining the structure distribution of CH* free radicals in the three-dimensional combustion field at the target time through a traditional reconstruction algorithm;

[0009] The reconstruction prediction model is constructed, and the model comprises four encoders, an attention module, a time processing module, three decoders and a convolution output layer; wherein each decoder comprises two convolution layers and a pooling layer, each decoder comprises a connection layer, two convolution layers and a deconvolution layer, the encoders and the decoders enable the model to fuse multi-scale information in space, the attention module is used for updating the key information weight in the detail feature, the time processing module is used for capturing the relationship between the data at each time of the input, so as to learn the evolution characteristics of the combustion, and the convolution output layer is used for adjusting the final output shape of the model;

[0010] Combining the continuous time multi-direction two-dimensional projection data of the combustion field, a loss function is constructed and the reconstruction prediction model is trained;

[0011] The historical time multi-direction two-dimensional projection data to be reconstructed and predicted is input into the trained reconstruction prediction model, and a three-dimensional combustion field reconstruction prediction result at a target time is obtained.

[0012] Further, the spatial multi-scale information features between the multi-direction two-dimensional projection data at the continuous historical times are extracted according to the encoders in the reconstruction prediction model, and specifically:

[0013] The multi-direction two-dimensional projection data is subjected to two times of 2*2 convolution and maximum pooling to obtain a first-level feature mark F 32 ;

[0014] The first-level feature mark F 32 is subjected to two times of 2*2 convolution and maximum pooling to obtain a second-level feature mark F 64 ;

[0015] The second-level feature mark F 64 is subjected to two times of 2*2 convolution and maximum pooling to obtain a third-level feature mark F 128 ;

[0016] The third-level feature mark F 128 is subjected to two times of 2*2 convolution and maximum pooling to obtain a fourth-level feature mark F 256 .

[0017] Further, the attention module comprises channel attention and spatial attention:

[0018] The channel attention generates spatial information through average pooling and maximum pooling, and infers the importance of different channels and distributes specific weight coefficients through shared multi-layer perceptron processing, wherein the selection of the channel number reduction ratio is related to the calculation efficiency and the feature expression ability of the model, and the channel number reduction ratio is selected as 4;

[0019] The spatial attention finds more important parts in the space by splicing the maximum pooling and average pooling features and generating spatial attention features through a 7*7 convolution kernel, thereby improving the expression ability of the key area of the combustion field in the network.

[0020] Further, the detail features of multiple continuous time steps output by the attention module are predicted to the detail features of the target time step by the time processing module, specifically:

[0021] The detail features (F 32 , F 64 , F 128 , F 256 ) of multiple known time steps are input into the time processing module after being updated by the attention module, to obtain the predicted detail features (F′ 32 , F′ 64 , F′ 128 , F′ 256 ) of the target time step. The size of the hidden layer of the gated recurrent unit is set to 256, and the number of hidden layers is 1 layer.

[0022] Further, the predicted detail features obtained by the time processing module are reconstructed by the decoding layer, and finally adjusted in shape and output by the convolution output layer, specifically:

[0023] The F′ 256 output by the time processing module is executed twice 2*2 convolution and 2*2 deconvolution, and F′ 128 is fused through the connection layer to obtain F″ 128 ;

[0024] The F″ 128 is executed twice 2*2 convolution and 2*2 deconvolution, and F′ 64 is fused through the connection layer to obtain F″ 64 ;

[0025] The F″ 64 is executed twice 2*2 convolution and 2*2 deconvolution, and F′ 32 is fused through the connection layer to obtain F″ 32 ;

[0026] The F″ 32 is executed once 1*1 convolution to obtain the output F″′ 96 , which is (t, 96, 48, 48), where t is the predicted target time step.

[0027] Further, the loss function calculation method of the prediction reconstruction model is:

[0028]

[0029] Where N is the number of samples, vi and represent the real and predicted three-dimensional data, respectively.

[0030] The above technical solutions of the present application have the following beneficial effects compared with the prior art:

[0031] The present application obtains the two-dimensional projection of the combustion field in multiple directions through the flame emission spectrum acquisition device, obtains the three-dimensional structure distribution of CH* free radicals in the three-dimensional combustion field at the historical moment by using the chemiluminescence tomography technology and the multiplication algebra iterative algorithm, processes the long-distance dependence relationship through the gate recurrent unit, and adjusts the feature importance weight through the attention mechanism. According to the multi-directional two-dimensional projection at the historical moment, the three-dimensional structure of CH*

[0032] The three-dimensional reconstruction of free radicals is realized. The present application does not need to design a complex acquisition device, nor to solve a complex partial differential equation set, and can realize the millisecond-level fast reconstruction and prediction of the free radical structure distribution in the combustion field under the condition of reducing the experimental and calculation costs. The method used in the present application has great potential in processing the three-dimensional reconstruction of time series, and also provides a new idea for solving other problems of long-distance dependence relationship. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is the reconstruction and prediction flowchart of the present application.

[0034] Figure 2 is a schematic diagram of a chemiluminescence tomography imaging system.

[0035] Figure 3 is a schematic diagram of the structure of the reconstruction and prediction network model.

[0036] Figure 4 is a comparison of the present application, the prior art and the real reconstructed structure.

[0037] Figure 5 is a comparison of the prediction results of the present application and the prior art at multiple future moments.

[0038] Figure 6 is a comparison of the cross-section and difference of the prediction results of the present application and the real reconstructed structure.

[0039] Figure 7 is a comparison of the three-dimensional structure of the prediction results and the real reconstructed results of the present application in the inter-frame prediction experiment.

[0040] Figure 8 is a comparison of the cross-section and difference of the prediction results and the real data of the present application in the inter-frame prediction experiment. DETAILED DESCRIPTION

[0041] The application will be further described in conjunction with the drawings and specific embodiments so that those skilled in the art can better understand the application and implement it, but the embodiments are not limiting to the application.

[0042] The application uses two burning candle flames as a three-dimensional combustion field of a to-be-measured region, collects multi-directional two-dimensional projections, and trains a network model to quickly predict the flame structure of the three-dimensional combustion field. The specific process is shown in Figure 1 First, multiple industrial cameras are used to collect projection information of the combustion field in different directions, and a computer is connected to a trigger board to control the cameras to realize simultaneous collection of the projection information, and the combustion field is continuously collected at the same time interval. The collected two-dimensional projections are arranged and processed in chronological order, and the three-dimensional field data obtained by multiplying the algebraic iterative algorithm on the future projection data is taken as the true value, which is compared with the three-dimensional field data predicted by the neural network from the historical projection, and the prediction performance of the model is determined according to different quality evaluation indexes.

[0043] The technical solutions realized by each step in the application are as follows:

[0044] First, a flame emission spectrum collection device is built to obtain two-dimensional projections in each direction. The device is composed of 12 color CCD cameras, the target surface pixel size of the camera is 3.75pm, and the resolution is 1294x964. The focal length of the camera lens is 12mm, and the maximum numerical aperture is 8.6mm. A band-pass filter of a specific wavelength is placed in the camera to obtain the free radical intensity signal of the burning flame. The center wavelength of the filter is 431.5nm, and the bandwidth is 20nm. The center of the device is the measured region, and 12 cameras are placed around the device to form a semicircle, and the interval between each camera is 15°. All cameras are connected to a computer through a trigger board, and at each time instant, the flame is captured by the cameras from 12 projections at different angles. In order to ensure the accuracy of the reconstruction, it is necessary to calibrate each camera in the collection device to ensure that the information of the combustion field is placed in the same position of the projection. 10 frames of continuous projection data are captured for 50s to obtain 500 continuous instantaneous projection data in 12 directions.

[0045] According to the prediction requirements, the data set is divided, the input of the neural network model is the multi-directional two-dimensional projection at the known time, and the output end is the three-dimensional combustion field structure distribution at the predicted target time, which is obtained by a traditional multiplying algebraic iterative algorithm.

[0046] A suitable neural network framework is constructed, and the processed dataset is fed into the neural network for training. This invention designs a novel hybrid neural network based on gated recurrent units and an attention mechanism. Convolutional neural networks are used to capture the complex spatial structure features of the data, combined with the gated recurrent units' ability to resolve long-range dependencies. The temporal processing module is placed at the skip connections of the network, enhancing the model's ability to model temporal information by fusing spatiotemporal features. Simultaneously, skip connections alleviate the gradient vanishing problem in deep network training. A hybrid attention mechanism is introduced into the temporal processing module to process important temporal and spatial information in each layer's features, achieving deep information fusion and efficient utilization, significantly improving the accuracy of prediction and reconstruction.

[0047] Neural network model architecture such as Figure 2 As shown, the U-Net model from the convolutional neural network is used as the backbone of this network, consisting of 4 encoding layers, 3 decoding layers, and 1 convolutional output layer. Each encoding layer contains two convolutional layers and one pooling layer, and each decoding layer contains two convolutional layers and one deconvolutional layer. The U-Net network can fuse multi-scale information. The encoder extracts features through convolution and downsampling, and then the decoder upsampling gradually recovers the spatial information to reconstruct three-dimensional data. The model's input consists of 2D projections from cameras in 12 directions at multiple consecutive instants, with a size of 96×96 pixels. The output is the 3D structural distribution of CH* components in the combustion field, with a size of 48×48×96. The number of sampling channels is set to 32. Too few channels will prevent the network from fully learning useful information in the combustion field, while setting too many channels will increase computational cost and the risk of model overfitting.

[0048] Gated Recurrent Units (GRUs) are placed in the skip links to accomplish time-series prediction tasks. Single reconstruction methods cannot capture the dynamic characteristics of combustion. GRUs, as a variant of recurrent neural networks, possess a strong ability to capture temporal dependencies, handle the connections between multi-directional projections at each time step, and avoid the gradient explosion and vanishing problems inherent in traditional recurrent neural networks, thus learning characteristics such as combustion shape and temperature distribution. GRUs include update and reset gates. The reset gate's value represents the correlation between the current input information vector and the information passed from the previous GRU unit. The reset gate determines the current candidate state value based on the previous unit's state and the current GRU input value. The update gate calculates the current unit's state value and passes it to the next GRU to achieve information transfer. Compared to other recurrent convolutional network variants (such as Long Short-Term Memory networks, LSTM), GRUs have a simpler structure, higher computational efficiency, and are more suitable for time-series prediction tasks. Embedding the fewer-parameter GRUs into the skip links of the convolutional neural network allows the network to consider temporal information when extracting and fusing combustion field features, enhancing feature representation capabilities, eliminating the need for complex time-series post-processing modules, and avoiding a significant increase in model complexity. The size of the hidden layer in the gated recurrent unit determines the expressive power of the model, which is set to 256 in this model.

[0049] During combustion, certain regions, such as the flame center and high-temperature areas, may be more important for the reconstruction task. A hybrid attention module is added before the gated recurrent unit to learn the relational weights in the features. For example... Figure 2 As shown, the hybrid attention mechanism includes channel attention and spatial attention, where C, M, and N represent the number of channels, spatial width, and spatial height of the input features, respectively. Channel attention uses average pooling and max pooling to generate spatial information. It infers the importance of different channels and assigns specific weight coefficients to them through a shared multilayer perceptron. The choice of the channel reduction ratio affects the model's computational efficiency and feature expressiveness; in lightweight networks (less than 256 channels), a channel reduction ratio of 4 is often chosen. Spatial attention, on the other hand, concatenates max-pooling and average-pooling features and generates spatial attention features using a 7×7 convolutional kernel. This identifies more important parts in space, improving the expressiveness of key combustion field regions in the network. Combining channel attention and spatial attention mechanisms can retain key information while suppressing noise and other useless information, thus improving model performance. Adding an attention module allows for better handling of local variations in the combustion field.

[0050] The specific implementation process of this invention is as follows: Figure 1As shown, multiple industrial cameras are first used to collect projection information of the combustion field from different directions. A computer connected to a trigger board controls the cameras to simultaneously acquire this projection information, continuously collecting data on the combustion field at equal time intervals. The acquired two-dimensional projections are then processed in chronological order. The three-dimensional field data reconstructed from future projection data using a multiplicative algebraic iterative algorithm is used as the true value and compared with the three-dimensional field data predicted by a neural network from historical projections. The predictive performance of the model is then determined based on different quality evaluation indicators.

[0051] The flame emission spectrum acquisition device constructed in this invention consists of 12 color CCD cameras (Allied Vision Technologies, Guppy Pro F-125C). The camera's target surface pixel size is 3.75 μm, and its resolution is 1294 × 964. The camera lens (CBC, Computar M1214) has a focal length of 12 mm and a maximum numerical aperture of 8.6 mm. A bandpass filter of a specific wavelength is placed in the camera to obtain the free radical intensity signal of the combustion flame. The filter's center wavelength is 431.5 nm, and its bandwidth is 20 nm. The measured area is at the center of the device, and the 12 cameras are placed around the device in a semicircle, with a 15° interval between each camera.

[0052] All cameras were connected to a computer via trigger boards. At each instant, the flame was simultaneously captured by the cameras from 12 projections at different angles. To ensure the accuracy of the reconstruction, calibration was required to ensure that each camera in the acquisition device placed the combustion field information in the same position on the projection. The projections were continuously captured at a rate of 10 frames per second for 50 seconds, resulting in 500 continuous instantaneous projection data sets in 12 directions.

[0053] Chemiluminescence chromatography techniques used in reconstructing three-dimensional combustion fields, such as... Figure 2 As shown. In chemiluminescence chromatography, the three-dimensional region of interest is discretized into several small voxels, and each voxel forms a light spot on the target surface of the camera after being imaged by a lens. For example... Figure 1 As shown, by constructing the physical correlation between each voxel and pixel during the imaging process, the weight matrix between the measured three-dimensional field and the imaging system can be obtained.

[0054] Assuming a multi-directional acquisition system consists of g cameras, each containing h pixels, the relationship between the 3D measured field and the 2D camera projection can be expressed as:

[0055]

[0056] Where i is the voxel index of the three-dimensional test field, j represents the index of the projection ray, X, Y, Z are the number of voxels in the x, y, z directions of the three-dimensional field, and j is the luminous intensity of the i-th voxel in the three-dimensional test field. The weight matrix can be regarded as the contribution rate of voxel i pairs. The imaging process can be summarized as follows:

[0057] I = WF

[0058] In the formula, I represents the light intensity information acquired by the camera; F is the three-dimensional field information to be measured; and W is the weight matrix of voxels in the three-dimensional field and their corresponding camera pixels. Therefore, the reconstruction calculation process can be equivalent to solving the formula.

[0059] The weight matrix and three-dimensional field information can be solved using many mature algorithms, among which algebraic iteration algorithms are widely used in solving emission tomography problems. The basic idea of ​​algebraic iteration algorithms is to treat the formula as a system of equations, set an initial value, project it onto the equations, and iterate step by step based on the error between the projected value and the actual value, eventually reaching a predetermined termination condition and ending the iteration. Commonly used algebraic iteration algorithms include additive algebraic iteration algorithms and multiplicative algebraic iteration algorithms. The multiplicative algebraic reconstruction technique (MART) optimizes the method of correcting the initial value, improves the speed of iteration convergence, and selects the solution with the greatest uncertainty (maximum entropy) under known constraints, ensuring the smoothness and rationality of the solution and improving the reconstruction effect of the iterative algorithm. The MART algorithm can be expressed by the following formula:

[0060]

[0061] × indicates element-wise multiplication. Correction factor in MART. The choice of correction factor has a significant impact on the speed and accuracy of iteration. A smaller correction factor results in slower convergence but higher accuracy, while a larger correction factor leads to faster convergence but lower accuracy. Therefore, the choice of correction factor is crucial in the iterative reconstruction process. Many studies have demonstrated that algebraic iterative reconstruction methods can be used for ground truth comparison in network training.

[0062] Figure 3 The reconstruction prediction neural network model architecture used in this invention is illustrated. Figure 3(a) illustrates the overall framework of the model, which uses the U-Net model from convolutional neural networks as its backbone, consisting of four encoding layers, three decoding layers, and one convolutional output layer. Each encoding layer contains two convolutional layers and one pooling layer, while each decoding layer contains two convolutional layers and one deconvolutional layer. The U-Net network can fuse multi-scale information. The encoder extracts features through convolution and downsampling, and then the decoder upsampling gradually recovers the spatial information, reconstructing 3D data. The model's input consists of 2D projections from 12 directions captured by the camera at 8 consecutive instants, with a size of 96×96 pixels. The output is the 3D structural distribution of CH* components in the combustion field, with a size of 48×48×96. The number of sampling channels is set to 32. Too few channels would prevent the network from fully learning useful information from the combustion field, while setting too many channels would increase computational cost and the risk of overfitting.

[0063] Gated recurrent units (GRUs) are placed in the skip link section to complete the time series prediction task. Single reconstruction methods cannot capture the dynamic characteristics of combustion. As a variant of recurrent neural networks, GRUs have a strong ability to capture temporal dependencies, handle the connections between multi-directional projections at each time step, and avoid the gradient explosion and vanishing problems of traditional recurrent neural networks, thus learning characteristics such as the shape and temperature distribution of combustion. Figure 3 As shown in (b), the GRU includes an update gate and a reset gate. The value of the reset gate represents the correlation between the current input information vector and the information passed by the previous GRU unit. The reset gate determines the current candidate state value based on the state of the previous unit and the current GRU input value. The update gate calculates the current unit state value and passes it to the next GRU to realize the information transmission. The prediction process of GRU can be represented as:

[0064]

[0065] In the formula, tanh is the hyperbolic tangent activation function, ⊙ represents the successive multiplication of each element, [] represents vector concatenation, and w,w z ,w r Here, b, b are the weight matrices for the update gate and reset gate, respectively. z ,b r These are the bias matrices for the bias matrix, update gate, and reset gate, respectively. The size of the hidden layer in the gated recurrent unit determines the expressive power of the model; in this model, it is set to 256. Figure 3(c) shows the hybrid attention module in the network model. The hybrid attention mechanism includes channel attention and spatial attention, where C, M, and N represent the number of channels, spatial width, and spatial height of the input features, respectively. Channel attention uses average pooling and max pooling to generate spatial information. It infers the importance of different channels through a shared multilayer perceptron and assigns specific weight coefficients to them. The choice of the channel reduction ratio affects the model's computational efficiency and feature expressiveness; in lightweight networks (less than 256 channels), a channel reduction ratio of 4 is often chosen. Spatial attention, on the other hand, concatenates max pooling and average pooling features and generates spatial attention features using a 7×7 convolutional kernel to identify more important parts in space, improving the expressiveness of key combustion field regions in the network. Combining channel attention and spatial attention mechanisms can retain key information while suppressing noise and other useless information, thus improving model performance. Adding an attention module allows for better handling of local variations in the combustion field.

[0066] A neural network framework was built using PyTorch version 2.4. This network model was accelerated and optimized on a computer equipped with an Intel Core i9-10900X CPU, 128GB RAM, and an Nvidia GeForce RTX 3090 GPU. During network training, the Adam optimizer was used to optimize the parameters, and the mean squared error (MSE) was chosen as the loss function, with the following expression:

[0067]

[0068] Where N is the number of samples, and and represent the real and predicted 3D data, respectively. The network is trained 50 times, with an input time step of 8 and an output of 5. To maintain a balance between computational efficiency and computational complexity and prevent overfitting, the number of channels in the network model is set to 32, and the number of features in the GRU hidden layer is 256. The initial learning rate is set to 0.001, and it is gradually reduced based on the decrease in the validation set loss function during training. If the validation loss does not decrease after 5 training iterations, the learning rate is reduced to half of the original value. The minimum learning rate is set to 10 to the power of -6.

[0069] This invention integrates chemiluminescence tomography reconstruction and the ability of deep learning to process spatial-temporal information and long-range dependencies, enabling rapid reconstruction and prediction of the three-dimensional combustion field spatial structure at future moments from multi-directional two-dimensional projections of historical moments. Compared with existing technologies, this invention has the following advantages:

[0070] 1. This invention obtains multi-directional two-dimensional projections of the detection area at continuous time by using an emission spectrum acquisition device, and combines deep learning to quickly and accurately reconstruct and predict the three-dimensional structural distribution of CH* radicals in the combustion field between future and current frames. This overcomes the shortcomings of traditional reconstruction methods, which require solving complex partial differential equations and have high computational costs.

[0071] 2. In the neural network model, the U-Net framework is used as the backbone to fuse multi-scale information, capture complex spatial structure features in the input data, and combine gated recurrent units to solve long-distance dependencies. The time processing module is placed in the skip connection part of the network. By fusing spatiotemporal features, the model's ability to model temporal information is enhanced. At the same time, skip connections are used to alleviate the gradient vanishing problem in deep network training.

[0072] 3. This invention introduces a hybrid attention mechanism in the time processing module to process important temporal and spatial information in each layer of features, thereby achieving deep fusion and efficient utilization of information and significantly improving the accuracy of prediction and reconstruction.

[0073] To further demonstrate the advantages of this invention, this embodiment uses the present invention and a CNN-LSTM model (see the paper "Huang J, Liu H, Cai W, et al. Online in situ prediction of 3-D flameevolution from its history 2-D projections via deep learning[J]. Journal of Fluid Mechanics, 2019, 875.") as the combustion field for experimental comparison. Structural Similarity Index Measure (SSIM), Pearson Correlation Coefficient (CORR), Peak Signal-to-Noise Ratio (PSNR), and Root Mean Squared Error (RMSE) are used as the quality evaluation functions for the experimental results.

[0074] The qualitative comparison results of reconstruction prediction between the present invention and the prior art are as follows: Figure 4 As shown. By comparing the predicted results, the predicted results of existing methods, and the actual data with the 3D view and 2D cross-section, it can be demonstrated that the model can predict the 3D structure of future times based on multi-directional projections of historical times, and the model's prediction performance is superior to existing reconstruction prediction methods. Figure 4(g) lists the quality evaluation functions for the predicted and actual structures, further confirming the reliability of the model's predictions.

[0075] Table 1 and Figure 5 This paper presents quantitative and qualitative comparisons between the present invention and existing technologies for reconstructing predictions at multiple future time points. The results show that the prediction accuracy of the model decreases with increasing prediction time intervals. This phenomenon aligns with the evolution of the combustion field: as the time span between the target prediction time and the input time increases, the combustion field is more significantly affected by external factors such as temperature and airflow, leading to increased uncertainty in the combustion field evolution process. The present invention demonstrates higher reconstruction prediction accuracy compared to the CNN-LSTM method, and maintains a high level of accuracy even with increasing prediction time intervals.

[0076] Table 1. Quantitative comparison of the predictive performance of the present invention and prior art.

[0077]

[0078] Figure 6 This invention illustrates the difference between the reconstruction prediction results of this invention and the internal structure of the combustion field reconstructed by traditional methods. This invention utilizes deep learning to extract crucial input information during the training process of the neural network, continuously optimizing parameters and updating weights. This reduces the impact of irrelevant information on the output results and significantly lowers the noise caused by traditional reconstruction methods.

[0079] In addition, this invention also exhibits excellent performance in interpolation prediction between time series. Experimental results are as follows: Figure 7 and Figure 8 As shown. Inter-frame prediction aims to predict multiple intermediate frames between two consecutive frames. Therefore, the input to the network model is the projection data of 12 directions of two consecutive frames in the sample, and the output of the model is the three-dimensional structural distribution of the CH* components of multiple intermediate frames between the input two consecutive frames. Without changing the overall structure of the neural network, the parameters such as the network's input and output are fine-tuned. The Bidirectional parameter in the GRU unit is changed to True, which means that the input information is processed bidirectionally, and the contextual information can be utilized simultaneously to improve the understanding of the global input.

Claims

1. A method for emission spectral tomography reconstruction and prediction based on gated recurrent units and attention mechanisms, characterized in that, include: Acquire continuous, multi-directional two-dimensional projection data of the detection area collected by the flame-emitting device; The structural distribution of CH* free radicals in the three-dimensional combustion field at the target time was obtained using a traditional reconstruction algorithm. A reconstruction prediction model is constructed, comprising four encoders, an attention module, a temporal processing module, three decoders, and a convolutional output layer. Each decoder contains two convolutional layers and one pooling layer, and each decoder also contains one connection layer, two convolutional layers, and one deconvolutional layer. The encoders and decoders enable the model to fuse multi-scale spatial information. The attention module is used to update the weights of key information in the detailed features. The temporal processing module is used to capture the relationships between the input data at each time step, thereby learning the evolutionary characteristics of combustion. The convolutional output layer is used to adjust the final output shape of the model. By combining the continuous multi-directional two-dimensional projection data of the combustion field, a loss function is constructed and the reconstruction prediction model is trained. By inputting the multi-directional two-dimensional projection data of the historical time to be reconstructed into the trained reconstruction prediction model, the three-dimensional combustion field reconstruction prediction result at the target time is obtained.

2. The emission spectral tomography reconstruction prediction method based on gated recurrent units and attention mechanisms according to claim 1, characterized in that: Based on the encoder in the reconstruction prediction model, spatial multi-scale information features between the multi-directional two-dimensional projection data of the continuous historical moments are extracted, specifically as follows: Perform two 2×2 convolutions and max pooling on the multi-directional two-dimensional projection data to obtain the first-level feature label F. 32 ; For the first-level feature label F 32 Perform two 2×2 convolutions and max pooling to obtain the second-level feature labeled F. 64 ; The second-level feature is labeled F 64 Perform two 2×2 convolutions and max pooling to obtain the third-level feature label F. 128 ; For the third-level feature label F 128 Perform two 2×2 convolutions and max pooling to obtain the fourth-level feature labeled F. 256 .

3. The emission spectral tomography reconstruction prediction method based on gated recurrent units and attention mechanisms according to claim 2, characterized in that: The attention module includes channel attention and spatial attention: The channel attention generates spatial information through average pooling and max pooling, and infers the importance of different channels through a shared multilayer perceptron and assigns specific weight coefficients to them. The choice of the channel number reduction ratio is related to the computational efficiency of the model and the expressive power of the features. The channel number reduction ratio is selected as 4. The spatial attention feature is generated by splicing max pooling and average pooling features and generating spatial attention features through a 7×7 convolution kernel to find more important parts in space and improve the expressive power of key areas of the combustion field in the network.

4. The emission spectral tomography reconstruction prediction method based on gated recurrent units and attention mechanisms according to claim 1, characterized in that: The detailed features from multiple consecutive time points output by the attention module are used by the time processing module to predict the detailed features at the target time point. Specifically: Detailed features at multiple known times (F) 32 F 64 F 128 F 256 After the attention module updates the weights, the data is input into the time processing module to obtain the detailed features (F′) of the predicted target time step. 32 F′ 64 F′ 128 F′ 256 The hidden layer size of the gated loop unit is set to 256, and the number of hidden layers is 1.

5. The emission spectral tomography reconstruction prediction method based on gated recurrent units and attention mechanisms according to claim 1, characterized in that: The predicted detailed features obtained from the temporal processing module are reconstructed through the decoding layer, and finally the data shape is adjusted and output through the convolutional output layer, specifically: F′ is obtained from the output of the time processing module 256 Perform two 2×2 convolutions and a 2×2 deconvolution, connected to F′ through a connecting layer. 128 F″ is obtained by fusion 128 ; For F″ 128 Perform two 2×2 convolutions and a 2×2 deconvolution, connected to F′ through a connecting layer. 64 F″ is obtained by fusion 64 ; For F″ 64 Perform two 2×2 convolutions and a 2×2 deconvolution, connected to F′ through a connecting layer. 32 F″ is obtained by fusion 32 ; For F″ 32 Performing a 1×1 convolution once yields the output F″′ 96 The shape is (t, 96, 48, 48), where t is the time step of the predicted target.

6. The emission spectral tomography reconstruction prediction method based on gated recurrent units and attention mechanisms according to claim 1, characterized in that: The loss function of the prediction and reconstruction model is calculated as follows: Where N is the number of samples, v i and These represent real and predicted 3D data, respectively.

Images