Method and system for simultaneous measurement of multi-parameter field of soot in flame based on Trident-net

Abstract

The present invention discloses a method for simultaneously measuring multi-parameter fields of soot in flames based on Trident-net, including: collecting experimental data and performing preprocessing, and randomly dividing the preprocessed experimental data into training set data and test set data; constructing a measurement model based on Trident-net and initializing the parameters of the measurement model; training the measurement model using the training set data, and testing the trained measurement model using the test set data according to a preset test index until the test result meets the preset conditions to obtain a measurement model with optimized parameters; inputting the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the multi-parameter fields of soot of the flame to be measured. The present invention also discloses a system for simultaneously measuring multi-parameter fields of soot in Trident-net. The measurement method and system provided by the present invention can effectively reduce the measurement cost and time cost while ensuring the measurement accuracy, and have a wide range of application scenarios.

Description

Technical Field

[0001] The present invention relates to the technical fields of optical diagnosis of combustion and artificial intelligence, and particularly relates to a method and system for simultaneously measuring multiple soot parameters in a flame based on Trident-net. Background Art

[0002] As a common combustion product, soot particles have always been a hot research topic. They will have an adverse impact on people's production and life, such as causing the greenhouse effect to the environment, damaging public health, and reducing energy utilization efficiency. However, soot particles play an active role in modern industries such as tires, inks, coatings, plastics, and electrical applications. The morphological parameters of soot play an important decisive role in the functionality of soot.

[0003] Artificial intelligence technology has been applied in the optical diagnosis of combustion. For example, due to the high non-linear fitting ability of machine learning, relevant scholars have begun to study the soot parameters of flames based on machine learning. However, in the existing laser diagnosis technology of combustion, it is difficult to achieve simultaneous measurement of multiple parameter fields, and it is also difficult to be widely used in actual industrial combustion equipment. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method and system for simultaneously measuring multiple soot parameters in a flame based on Trident-net, in order to solve one of the above problems.

[0005] According to the first aspect of the present invention, a method for simultaneously measuring multiple soot parameters in a flame based on Trident-net is provided, including:

[0006] Collecting experimental data and performing preprocessing, and randomly dividing the preprocessed experimental data into training set data and test set data, wherein the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data;

[0007] Constructing a measurement model based on the Trident-net convolutional neural network and initializing the parameters of the measurement model, wherein the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder;

[0008] Training the measurement model with the training set data, and testing the trained measurement model with the test set data according to a preset test index until the test result meets the preset condition, to obtain a measurement model with optimized parameters;

[0009] Inputting the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the multiple soot parameter fields of the flame to be measured.

[0010] According to an embodiment of the present invention, collecting the experimental data, performing preprocessing, and randomly dividing the preprocessed experimental data into training set data and test set data includes:

[0011] Using an image acquisition device to acquire an image of the flame to obtain a flame image, wherein the image acquisition device is communicatively connected to a dedicated image processing device;

[0012] Processing the flame image using an optical diagnostic method to obtain flame radiation field data;

[0013] Using a laser diagnostic method to obtain soot parameter field data of the flame, wherein the soot parameter field data of the flame includes soot temperature field data, soot particle size field data, and soot volume fraction field data;

[0014] Performing data normalization processing on the flame radiation field data and the soot parameter field data of the flame to obtain preprocessed flame radiation field data and preprocessed soot parameter field data of the flame;

[0015] Randomly dividing the preprocessed flame radiation field data and the preprocessed soot parameter field data of the flame into training set data and test set data.

[0016] According to an embodiment of the present invention, using the laser diagnostic method to process the flame radiation field data respectively to obtain the soot parameter field data of the flame includes:

[0017] Using TALF (two-atom laser-induced fluorescence) to obtain soot temperature field data in the flame;

[0018] Using TiRe-LII (time-resolved laser-induced incandescence) to obtain soot particle size field data in the flame;

[0019] Using the LII (laser-induced incandescence) method to obtain soot volume fraction field data in the flame.

[0020] According to an embodiment of the present invention, the above-mentioned shared encoder includes a plurality of convolutional layers, a plurality of downsampling layers, and a rectified linear unit;

[0021] Wherein, each convolutional layer processes the flame radiation field data using a batch normalization layer to obtain a flame feature map;

[0022] Wherein, the convolutional layer and the downsampling layer are overlapped and connected.

[0023] According to an embodiment of the present invention, the above-mentioned soot temperature decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, and a rectified linear unit;

[0024] The soot particle size decoder includes multiple upsampling layers, multiple convolutional layers, multiple fully connected layers, multiple skip connection layers, and ReLu (Rectified Linear Unit);

[0025] The soot volume fraction decoder includes multiple upsampling layers, multiple convolutional layers, multiple fully connected layers, multiple skip connection layers, multiple dropout layers, and ELU (Exponential Linear Unit) functions;

[0026] Among them, the soot temperature decoder, the soot particle size decoder, and the soot volume fraction decoder are connected to the shared encoder through their respective skip connection layers.

[0027] According to an embodiment of the present invention, the above-mentioned method of training the measurement model using the training set data and testing the trained measurement model using the test set data according to a preset test index until the test result meets the preset condition to obtain a measurement model with optimized parameters includes:

[0028] Set the hyperparameters of the measurement model according to the sample size of the training set data;

[0029] According to the preset weight coefficients, adjust the weight coefficients of the loss function of the soot temperature decoder, the weight coefficients of the loss function of the soot particle size decoder, and the weight coefficients of the loss function of the soot volume fraction decoder respectively;

[0030] Use the training set data to train the adjusted soot temperature decoder, soot particle size decoder, and soot volume fraction decoder, and obtain a training score;

[0031] According to the preset test index, use the test set data to test the trained measurement model to obtain a test result;

[0032] Iteratively perform the training operation and the testing operation, and analyze the test result and the number of iterations until the test result meets the preset condition to obtain a measurement model with optimized parameters.

[0033] According to an embodiment of the present invention, the above-mentioned method of iteratively performing the training operation and the testing operation, and analyzing the test result and the number of iterations until the test result meets the preset condition to obtain a measurement model with optimized parameters includes:

[0034] Judge the prediction situation of the measurement model using the test set data according to the training score and the test result to obtain a judgment result;

[0035] Plot the test result and the number of iterations generated in each iteration into a test result - number of iterations distribution diagram, and analyze the distribution diagram to obtain an analysis result;

[0036] Determine the training iteration number of the measurement model according to the judgment result and the analysis result.

[0037] According to an embodiment of the present invention, the above test indexes include the mean absolute errors of soot temperature field data, soot particle size field data, and soot volume fraction field data, and the total score of the measurement model.

[0038] According to an embodiment of the present invention, the above mean absolute error is represented by formula (1):

[0039] (1),

[0040] Wherein, the total score of the measurement model is represented by formula (2):

[0041] (2),

[0042] Wherein, is the experimental true value corresponding to the i-th index in the flame field data, is the model prediction value corresponding to the i-th index in the flame field data, and n is the total value of the experimental data.

[0043] According to a second aspect of the present invention, there is provided a system for simultaneous measurement of multiple soot parameter fields in a flame based on Trident-net, including:

[0044] A data acquisition and processing module, configured to acquire experimental data, perform preprocessing, and randomly divide the preprocessed experimental data into training set data and test set data, wherein the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data;

[0045] A model construction module, configured to construct a measurement model based on the Trident-net convolutional neural network and initialize the parameters of the measurement model, wherein the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder;

[0046] A training and testing module, configured to train the measurement model using the training set data, and test the trained measurement model using the test set data according to preset test indexes until the test result meets the preset conditions, so as to obtain a measurement model with optimized parameters;

[0047] A measurement module, configured to input the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the multiple soot parameter fields of the flame to be measured.

[0048] The method and system for simultaneous measurement of multiple soot parameter fields in a flame based on Trident-net provided by the present invention use experimental data to train and test a measurement model based on Trident-net, so as to obtain a measurement model with high prediction accuracy, strong fitting ability, and wide application scenarios. Brief Description of the Drawings

[0049] Figure 1 is a flowchart of a method for simultaneously measuring multiple soot parameters in a flame based on Trident-net according to an embodiment of the present invention;

[0050] Figure 2 is a flowchart of obtaining training set data and test set data according to an embodiment of the present invention;

[0051] Figure 3 is a schematic diagram of obtaining experimental data according to an embodiment of the present invention;

[0052] Figure 4 is a schematic structural diagram of a measurement model based on Trident-net according to an embodiment of the present invention;

[0053] Figure 5 is a flowchart of obtaining a measurement model with optimized parameters according to an embodiment of the present invention;

[0054] Figure 6 is a structural diagram of a system for simultaneously measuring multiple soot parameters in a flame based on Trident-net according to an embodiment of the present invention. Detailed Description of the Embodiments

[0055] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to specific embodiments and the accompanying drawings.

[0056] Currently, machine learning and neural networks have been applied in the field of optical diagnostics of combustion. Wang Qianlong et al. first developed and verified the BackPropagation (BP) neural network model, which is a two-parameter distribution model that can simultaneously predict the soot temperature and volume fraction distribution based on the flame radiation field, and a three-parameter model that can simultaneously predict the soot temperature, volume fraction, and particle size distribution. However, in the BP neural network, the flame image information is segmented into input neurons at the pixel size in the height direction for training. Therefore, there is a tendency to lose the parameter information in the height direction. Thus, the two-dimensional convolutional neural network (U-net) model was further developed and verified to be able to simultaneously predict the soot temperature and volume fraction field based on the flame radiation field. By comparing the prediction results of the U-net neural network with those of the BP neural network, it was found that the U-net model can indeed obtain a smoother soot parameter field in the height direction. In this model architecture and training process, there are certain limitations, mainly the following two points: (1) There is a trade-off relationship among the three soot parameters, which is likely to cause some parameters to dominate while some parameters are not fully optimized, thus affecting the prediction accuracy. (2) The prediction accuracy of the model is limited by the training sample size. When the training sample size is relatively small, the BP model and the U-net model can learn fewer feature quantities, and it is impossible to ensure the fitting ability and prediction accuracy of the model.

[0057] In order to measure each soot parameter with high precision, reduce the dependence on the sample size, and be able to flexibly set the network structure of each soot parameter branch according to research requirements during the measurement process, a new method is urgently needed.

[0058] The method and system for measuring soot parameters of a flame based on Trident-net (Trident fork network model) provided by the present invention can perform real-time monitoring of the soot parameters in the flame, and can flexibly set the network structure of each soot parameter branch decoder according to research requirements during the measurement process, reducing the model's demand for the training sample size and simultaneously achieving high-precision measurement of the three soot parameters.

[0059] Figure 1 It is a flowchart of a method for simultaneously measuring multiple soot parameter fields in a flame based on Trident-net according to an embodiment of the present invention.

[0060] As Figure 1 shown, the above method for simultaneously measuring multiple soot parameter fields in a flame based on Trident-net includes operations S110 to S140.

[0061] In operation S110, experimental data is collected and preprocessed, and the preprocessed experimental data is randomly divided into training set data and test set data. Among them, the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data.

[0062] In operation S120, a measurement model based on the Trident-net convolutional neural network is constructed and the parameters of the measurement model are initialized. Among them, the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder.

[0063] The soot temperature decoder, the soot particle size decoder, and the soot volume fraction decoder are respectively connected to the shared encoder for data transfer.

[0064] In operation S130, the measurement model is trained using the training set data, and the trained measurement model is tested using the test set data according to preset test metrics until the test results meet the preset conditions, and a measurement model with optimized parameters is obtained.

[0065] In operation S140, the flame radiation field data of the flame to be measured is input into the measurement model with optimized parameters to obtain the multi-parameter soot field of the flame to be measured.

[0066] The method and system for simultaneous measurement of multi-parameter soot fields in a flame based on Trident-net provided by the present invention uses experimental data to train and test a measurement model based on Trident-net, and obtains a measurement model with high prediction accuracy, strong fitting ability, and wide application scenarios.

[0067] Figure 2 It is a flowchart for obtaining training set data and test set data according to an embodiment of the present invention.

[0068] Figure 3 It is a schematic diagram for obtaining experimental data according to an embodiment of the present invention.

[0069] The following combines Figure 2 and Figure 3 to further elaborate in detail on the process of the present invention for collecting experimental data, preprocessing it, and randomly dividing the preprocessed experimental data into training set data and test set data.

[0070] As Figure 2 shown, the above process of collecting experimental data, preprocessing it, and randomly dividing the preprocessed experimental data into training set data and test set data includes operations S210 to S250.

[0071] In operation S210, an image acquisition device is used to acquire an image of the flame to obtain a flame image. Herein, the image acquisition device is communicatively connected to the image processing dedicated device.

[0072] In operation S220, an optical diagnostic method is used to process the flame image to obtain flame radiation field data.

[0073] In operation S230, a laser diagnostic method is used to obtain soot parameter field data of the flame. Herein, the soot parameter field data of the flame includes soot temperature field data, soot particle size field data, and soot volume fraction field data.

[0074] In operation S240, the flame radiation field data and the soot parameter field data of the flame are subjected to data normalization processing to obtain the preprocessed flame radiation field data and the preprocessed soot parameter field data of the flame.

[0075] The normalization processing aims to process the soot parameter field data, i.e., the soot temperature field data, the soot particle size field data, and the soot volume fraction field data, into data between 0 and 1.

[0076] In operation S250, the preprocessed flame radiation field data and the preprocessed soot parameter field data of the flame are randomly divided into training set data and test set data.

[0077] Both the training set data and the test set data include the preprocessed flame radiation field data and the preprocessed soot parameter field data of the flame.

[0078] The acquisition process of the experimental data is as Figure 3 shown. The flame generator is turned on, and the camera lens (i.e., the image acquisition device) connected to the CDD (charge coupled device) camera takes pictures of the flame generated by the flame generator to obtain a flame image. The CCD camera or the high-speed camera is connected to the data interface of the computer (i.e., the image processing device), and through Figure 3 the settings shown, a large amount of reliable experimental data can be acquired.

[0079] According to an embodiment of the present invention, the above-mentioned use of the laser diagnostic method to process the flame radiation field data respectively to obtain the soot parameter field data of the flame includes:

[0080] Using TALF to obtain the soot temperature field data in the flame.

[0081] Using TiRe-LII to obtain the soot particle size field data in the flame.

[0082] Using LII processing to obtain the soot volume fraction field data in the flame.

[0083] According to an embodiment of the present invention, the above-mentioned shared encoder includes a plurality of convolutional layers, a plurality of downsampling layers, and a rectified linear unit;

[0084] Wherein, each convolutional layer processes the flame radiation field data by using a batch normalization layer to obtain a flame feature map;

[0085] Wherein, the convolutional layers and the downsampling layers are connected in an overlapping manner.

[0086] According to an embodiment of the present invention, the above-mentioned soot temperature decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, and a rectified linear unit.

[0087] The soot particle size decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, and a ReLu (rectified linear unit).

[0088] The soot volume fraction decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, a plurality of dropout layers, and an ELU (exponential linear unit) function.

[0089] Wherein, the soot temperature decoder, the soot particle size decoder, and the soot volume fraction decoder are connected to the shared encoder through their respective skip connection layers.

[0090] Figure 4 It is a schematic structural diagram of a measurement model based on Trident-net according to an embodiment of the present invention.

[0091] The following Figure 4 will further elaborate on the above-mentioned measurement model based on Trident-net.

[0092] The shared encoder is composed of 4 convolutional layers and 3 downsampling layers. The kernel sizes of the first three convolutional layers are 3x3, and the kernel size of the last convolutional layer is 1x1. The convolutional layer with a kernel size of 3x3 and a convolutional stride of 2 serves as a downsampling layer to reduce the size of the feature map to half. After each convolutional layer is subjected to feature normalization by a BN layer, it is activated by the function ReLU and then connected to the downsampling layer.

[0093] The soot temperature decoder is composed of 3 upsampling layers, 3 convolutional layers, and 4 fully connected layers. After each convolutional layer is subjected to feature normalization by a BN layer, it is activated by the function ReLU, and 3 skip connection layers are used to connect the soot temperature decoder to the shared encoder.

[0094] The soot volume fraction decoder consists of 2 upsampling layers, 2 convolutional layers, and 2 fully connected layers. After the feature normalization of each convolutional layer by the BN layer (Batch Normalization), it is activated by the ELU function. And 2 skip connection layers are used to connect the soot volume fraction decoder and the shared encoder. Additionally, two dropout layers are added for regularization to prevent overfitting and loss of feature information.

[0095] The soot particle size decoder consists of 3 upsampling layers, 3 convolutional layers, and 3 fully connected layers. After the feature normalization of each convolutional layer by the BN layer, it is activated by the ReLU function. And three skip connection layers are used to connect the soot particle size decoder and the shared encoder.

[0096] The structures of the soot temperature decoder, the soot volume fraction decoder, and the soot particle size decoder are not fixed and can be flexibly set according to the results of real-time measurements. For example, when it is observed that a certain encoder has underfitting, the network structure of the encoder can be tried to be deepened. When there is overfitting, the number of neural network layers can be appropriately reduced or dropout layers can be added. When the prediction accuracy of the encoder is not ideal, methods such as increasing the number of skip connection layers can be tried.

[0097] As Figure 4 shown, the flame radiation field data is input into the measurement model based on the T-net (i.e., Trident-net, the same below) convolutional neural network. After being processed by a convolutional layer with a convolutional kernel of 3x3, a convolutional stride of 1, and an output channel of 32, the data channel is converted from 1 to 32. Then, it is subjected to feature normalization by the BN layer and activated by ReLU before being connected to the downsampling layer. After each processing by the downsampling layer, the data size is reduced to half of the previous convolutional layer. After each processing by a neural network layer, the data channel will become twice that of the previous neural network layer. The shared encoder is connected by 3 convolutional layers with a kernel size of 3x3 and 3 downsampling layers. The 3rd downsampling layer is connected to the 1st convolutional layer with a kernel size of 1x1.

[0098] The 3 skip connection layers of the soot temperature decoder are used to splice the 3rd downsampling layer of the shared encoder with the 1st upsampling layer of the soot temperature branch, splice the 2nd downsampling layer of the shared encoder with the 2nd upsampling layer of the soot temperature branch, splice the 1st downsampling layer of the shared part with the 3rd upsampling layer of the soot temperature branch. Then, the 3rd skip connection layer is connected to the 3rd convolutional layer. After the output of the 3rd convolutional layer passes through Flatten (flattened) and serves as the input of the fully connected layer, the data is output after passing through 4 fully connected layers.

[0099] After the soot temperature decoder, the soot volume fraction decoder, and the soot particle size decoder are processed by each upsampling layer, the data size will be doubled, and each neural network layer will reduce the number of data channels to half of the original.

[0100] Figure 5 It is a flowchart of obtaining a measurement model with optimized parameters according to an embodiment of the present invention.

[0101] As Figure 5 shown, the above-mentioned training of the measurement model using the training set data and testing the trained measurement model using the test set data according to the preset test indicators until the test results meet the preset conditions to obtain the measurement model with optimized parameters includes operations S510 to S550.

[0102] In operation S510, the hyperparameters of the measurement model are set according to the sample size of the training set data.

[0103] For example, when the training set data is small, the learning rate can be set to decay exponentially. When the effect of the exponentially decaying learning rate is not good, empirical numbers such as 0.0003 and 0.001 can be set manually. When the training sample size is small, batch_size can be set to a larger value such as 64, and when the sample size is sufficient, it can be set to a smaller value such as 8.

[0104] In operation S520, according to the preset weight coefficients, the weight coefficients of the loss function of the soot temperature decoder, the weight coefficients of the loss function of the soot particle size decoder, and the weight coefficients of the loss function of the soot volume fraction decoder are adjusted respectively.

[0105] In the training function of the T-net convolutional neural network model, the weights of the loss functions of the three parameters of soot can be set by the user himself. When it is necessary to measure a certain parameter emphatically or the measurement effect of this parameter is poor, the weight of its loss function can be appropriately increased.

[0106] In operation S530, the adjusted soot temperature decoder, soot particle size decoder, and soot volume fraction decoder are trained using the training set data, and a training score is obtained;

[0107] In operation S540, according to the preset test indicators, the trained measurement model is tested using the test set data to obtain the test results.

[0108] In operation S550, the training operation and the test operation are iteratively performed, and the test results are analyzed with the number of iterations until the test results meet the preset conditions to obtain the measurement model with optimized parameters.

[0109] According to an embodiment of the present invention, the above-mentioned iterative training operation and test operation are performed, and the test results are analyzed together with the number of iterations until the test results meet the preset conditions. The measurement model with optimized parameters obtained includes:

[0110] Based on the training score and the test results, the prediction situation of the measurement model is judged using the test set data to obtain a judgment result.

[0111] The test results generated in each iteration are plotted against the number of iterations to form a test result - number of iterations distribution diagram, and the distribution diagram is analyzed to obtain an analysis result.

[0112] Based on the judgment result and the analysis result, the training iteration number of the measurement model is determined.

[0113] During the process of training the model, whether the model has underfitting or overfitting is judged by the training score and the test score of each iteration of the model. And by plotting the distribution diagram of the test score and the number of iterations, it is judged whether the iteration number of the model training meets the requirements according to the convergence situation of the test score.

[0114] Through the above method, the test set data can be fully utilized to test and evaluate the measurement model, and the weight coefficient of the encoder can be adjusted in a timely manner according to the actual requirements, so as to obtain a measurement model with good fitting effect and prediction effect.

[0115] According to an embodiment of the present invention, the above-mentioned test indexes include the mean absolute error of the soot temperature field data, the soot particle size field data, and the soot volume fraction field data, and the total score of the measurement model.

[0116] According to an embodiment of the present invention, the above-mentioned mean absolute error is represented by formula (1):

[0117] (1),

[0118] Wherein, the total score of the measurement model is represented by formula (2):

[0119] (2),

[0120] Wherein, is the experimental true value corresponding to the i-th index in the flame field data, is the model prediction value corresponding to the i-th index in the flame field data, and n is the total value of the experimental data.

[0121] Using the total score of the test model can better judge the fitting performance and prediction effect of the test model.

[0122] Figure 6It is a structural diagram of a system for simultaneous measurement of multiple soot parameters in a flame based on Trident-net according to an embodiment of the present invention.

[0123] As Figure 6 shown, the above-mentioned system 600 for simultaneous measurement of multiple soot parameters in a flame based on Trident-net includes a data acquisition and processing module 610, a model construction module 620, a training and testing module 630, and a measurement module 640.

[0124] The data acquisition and processing module 610 is used to acquire experimental data, perform preprocessing, and randomly divide the preprocessed experimental data into training set data and test set data. Among them, the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data.

[0125] The model construction module 620 is used to construct a measurement model based on the Trident-net convolutional neural network and initialize the parameters of the measurement model. Among them, the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder.

[0126] The training and testing module 630 is used to train the measurement model using the training set data, and test the trained measurement model using the test set data according to preset test metrics until the test results meet the preset conditions to obtain a measurement model with optimized parameters.

[0127] The measurement module 640 inputs the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the multi-parameter field of soot in the flame to be measured.

[0128] The above-mentioned method and system for simultaneous measurement of multiple soot parameters in a flame based on Trident-net provided by the present invention can overcome the problem that due to the trade-off relationship between three soot parameters, some parameters dominate while some parameters cannot be fully optimized. By flexibly setting the network structure of the three parameters and the loss function weights of the three parameters, the adverse effects between the three parameters are reduced, the problem of poor prediction effect of a certain parameter is avoided, and the overall prediction effect of the three parameters is improved; at the same time, the requirement for the training sample size is reduced. The above-mentioned method and system for simultaneous measurement of multiple soot parameters in a flame based on Trident-net proposed by the present invention can process small data sample sizes using the multi-task learning method, and can effectively reduce the requirement for the training sample size. In addition, the above-mentioned method and system for simultaneous measurement of multiple soot parameters in a flame based on Trident-net provided by the present invention can realize real-time dynamic observation of the formation process of soot parameters in the flame, and solve the difficulty of unable to monitor the soot parameters in the flame in industrial production in real time.

[0129] In the specific embodiments described above, the object, technical solution and beneficial effects of the present invention have been further described in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for simultaneous measurement of multi-parameter fields of soot in flames based on Trident-net, comprising: Collecting experimental data and performing preprocessing, and randomly dividing the preprocessed experimental data into training set data and test set data, wherein the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data; Constructing a measurement model based on the Trident-net convolutional neural network and initializing the parameters of the measurement model, wherein the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder; Training the measurement model using the training set data, and testing the trained measurement model using the test set data according to a preset test index until the test result meets the preset conditions to obtain a measurement model with optimized parameters; Inputting the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the multi-parameter field of soot in the flame to be measured; Wherein, the soot temperature decoder, the soot volume fraction decoder, and the soot particle size decoder have different neural network structures and can dynamically set the neural network structure according to the fitting situation; Wherein, the soot temperature decoder, the soot particle size decoder, and the soot volume fraction decoder are connected to the shared encoder through their respective skip connection layers; Wherein, according to the preset weight coefficients, the weight coefficients of the loss function of the soot temperature decoder, the weight coefficients of the loss function of the soot particle size decoder, and the weight coefficients of the loss function of the soot volume fraction decoder are respectively adjusted.

2. The method according to claim 1, wherein, Collecting experimental data and performing preprocessing, and randomly dividing the preprocessed experimental data into training set data and test set data includes: Using an image acquisition device to acquire images of the flame to obtain flame images, wherein the image acquisition device is communicatively connected to a special image processing device; Processing the flame images using an optical diagnostic method to obtain flame radiation field data; Obtaining the soot parameter field data of the flame using a laser diagnostic method, wherein the soot parameter field data of the flame includes soot temperature field data, soot particle size field data, and soot volume fraction field data; Performing data normalization processing on the flame radiation field data and the soot parameter field data of the flame to obtain preprocessed flame radiation field data and preprocessed soot parameter field data of the flame; Randomly dividing the preprocessed flame radiation field data and the preprocessed soot parameter field data of the flame into training set data and test set data.

3. The method according to claim 2, wherein, The obtaining the soot parameter field data of the flame using a laser diagnostic method includes: Obtaining the soot temperature field data in the flame using two-atom laser-induced fluorescence; Obtaining the soot particle size field data in the flame using time-resolved laser-induced incandescence; Obtaining the soot volume fraction field data in the flame using laser-induced incandescence.

4. The method according to claim 1, wherein The shared encoder includes a plurality of convolutional layers, a plurality of downsampling layers, and a rectified linear unit; Among them, each of the convolutional layers processes the flame radiation field data using a batch normalization layer to obtain a flame feature map; Among them, the convolutional layer is overlappingly connected to the downsampling layer.

5. The method according to claim 1, wherein The soot temperature decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, and a rectified linear unit function; The soot particle size decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, and a rectified linear unit function; The soot volume fraction decoder includes a plurality of upsampling layers, a plurality of convolutional layers, a plurality of fully connected layers, a plurality of skip connection layers, a plurality of dropout layers, and an exponential linear unit function.

6. The method according to claim 1, wherein, Training the measurement model using the training set data, and testing the trained measurement model using the test set data according to a preset test metric until the test result meets the preset condition to obtain a measurement model with optimized parameters includes: Setting the hyperparameters of the measurement model according to the sample size of the training set data; Using the training set data to train the soot temperature decoder, the soot particle size decoder, and the soot volume fraction decoder that have been adjusted, and obtaining a training score; Testing the trained measurement model using the test set data according to the preset test metric to obtain a test result; Iteratively performing the training operation and the testing operation, and analyzing the test result and the iteration number until the test result meets the preset condition to obtain a measurement model with optimized parameters.

7. The method according to claim 6, wherein, The iteratively performing the training operation and the testing operation, and analyzing the test result and the iteration number until the test result meets the preset condition to obtain a measurement model with optimized parameters includes: Judging the prediction situation of the measurement model using the test set data according to the training score and the test result to obtain a judgment result; Plotting the test result generated in each iteration and the iteration number into a test result - iteration number distribution diagram, and analyzing the distribution diagram to obtain an analysis result; Determining the training iteration number of the measurement model according to the judgment result and the analysis result.

8. The method according to claim 1 or 7, wherein, The test metrics include the mean absolute error of the soot temperature field data, the soot particle size field data, and the soot volume fraction field data, and the total score of the measurement model.

9. The method according to claim 8, wherein, The mean absolute error is represented by formula (1): （1）， Among them, the total score of the measurement model is represented by formula (2): （2）， Among them, is the experimental true value corresponding to the i-th index in the flame field data, is the model prediction value corresponding to the i-th index in the flame field data, and n is the total value of the experimental data.

10. A system for simultaneously measuring multiple soot parameter fields in a flame based on Trident-net, comprising: A data acquisition and processing module, configured to acquire experimental data, perform preprocessing, and randomly divide the preprocessed experimental data into training set data and test set data, where the experimental data includes flame radiation field data, soot temperature field data, soot particle size field data, and soot volume fraction field data; A model construction module, configured to construct a measurement model based on a Trident-net convolutional neural network and initialize the parameters of the measurement model, where the measurement model includes a shared encoder, a soot temperature decoder, a soot particle size decoder, and a soot volume fraction decoder; A training and testing module, which is used to train the measurement model by using the training set data, and test the trained measurement model by using the test set data according to preset test metrics until the test results meet the preset conditions, so as to obtain a measurement model with optimized parameters; A measurement module, which is used to input the flame radiation field data of the flame to be measured into the measurement model with optimized parameters to obtain the soot multi-parameter field of the flame to be measured; Wherein, the soot temperature decoder, the soot volume fraction decoder and the soot particle size decoder have different neural network structures and can dynamically set the neural network structures according to the fitting situation; Wherein, the soot temperature decoder, the soot particle size decoder and the soot volume fraction decoder are connected to the shared encoder through their respective skip connection layers; Wherein, according to the preset weight coefficients, the weight coefficients of the loss function of the soot temperature decoder, the weight coefficients of the loss function of the soot particle size decoder and the weight coefficients of the loss function of the soot volume fraction decoder are respectively adjusted.

Images