Laser-induced plasma image fluid velocity inversion method and system based on deep learning

Abstract

The invention discloses a laser-induced plasma image fluid velocity inversion method and system based on deep learning, and belongs to the field of cross combination of an artificial intelligence method and a laser-induced plasma velocity measurement method. The method comprises the following steps: for n continuous plasma images, removing pump light noise and global background interference in each plasma image; the plasma velocity between the plasma images is obtained by estimating the mass center position of each denoised image; generating a feature matrix based on all the de-noised images, and distributing the feature matrix on n-1 attention heads by using a multi-head attention mechanism to obtain the weight of each plasma velocity; and obtaining the movement speed of the plasma in the flow field according to the plasma speed and the weight of the plasma speed. According to the invention, the fluid velocity can be accurately estimated.

Description

Technical Field

[0001] The present invention relates to the cross-combination field of artificial intelligence methods and laser induced plasma velocity measurement methods, and specifically to a laser induced plasma image fluid velocity inversion method and system based on deep learning. Background Art

[0002] In the field of industrial monitoring, accurate flow rate measurement is the key to ensuring efficient and safe production processes. In key industries such as coal mining, metal smelting and chemical industry, accurate monitoring of flow rate can achieve refined control of production processes. For example, in the process of gas transportation, by accurately controlling the flow rate, the stability and efficiency of logistics transmission can be ensured; in mines, by adjusting the emission rate of toxic gases and the supply of fresh air, the safety of miners can be guaranteed; and in the process of metal smelting, by adjusting the gas flow rate, the combustion process can be optimized and energy utilization efficiency can be improved. However, traditional gas flow rate measurement methods are usually only applicable to specific scenarios. Under extreme conditions, such as high temperature, high pressure and high humidity environments, higher requirements are placed on measurement equipment, and the equipment needs to have stronger temperature and pressure resistance. At the same time, the gas flow rate under extreme conditions may be unstable, showing complex flow states such as pulsation and turbulence, which increases the difficulty of measurement. Therefore, it is of great scientific significance to carry out research on flow rate measurement and data processing under extreme environments.

[0003] Laser induced breakdown spectroscopy is a new non-contact gas flow rate measurement method. The laser pulse emitted by the laser is focused on the surface of the gas particles to generate free electrons, which are accelerated in the subsequent excitation of the laser pulse, and collide with atoms to produce an avalanche effect, resulting in a sharp increase in the electron density in the region and the formation of plasma. The gas flow rate of its surrounding environment can be measured by taking an image of the plasma moving in the flow field with a high-speed camera. However, in addition to the plasma, the image captured by the high-speed camera also contains the pump light noise interference in the shape of "melon seeds", which will cause great interference to the estimation of the center of mass position of the plasma. In addition, due to the expansion and annihilation process of the plasma itself, predecessors often use the results measured when the gas flow rate is 0m / s for velocity correction processing, but because the motion characteristics of the plasma are very small and the self-disturbance at different flow rates is also different and difficult to quantify, a more adaptive correction algorithm is needed to suppress the influence of the plasma self-disturbance on the fluid velocity estimation. In recent years, deep learning has made great progress in tasks such as real-time target detection and target tracking, but there has been no relevant results reported in the field of fluid velocity inversion that the present invention focuses on. Summary of the invention

[0004] The present invention overcomes the shortcomings of the prior art and proposes a deep learning-based laser induced plasma image fluid velocity inversion method and system, which utilizes a deep learning-based laser induced plasma image fluid velocity estimation network (Fluid Velocity Estimation Network, FVEN) to achieve accurate estimation of fluid velocity.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for inverting fluid velocity from laser-induced plasma images based on deep learning, the method comprising:

[0007] For n consecutive plasma images, removing pump light noise and global background interference in each plasma image; wherein the n consecutive plasma images are generated by photographing the movement of plasma in a flow field using a high-speed camera;

[0008] The plasma velocity between plasma images is obtained by estimating the centroid position of each denoised image;

[0009] Generate a feature matrix based on all denoised images, and use a multi-head attention mechanism to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity;

[0010] The moving speed of the plasma in the flow field is obtained according to the plasma speed and the weight of the plasma speed.

[0011] Furthermore, the removing of local pump light noise and global background interference in each plasma image includes:

[0012] Among the consecutive n plasma images, the last m plasma images in the time dimension are obtained, m <n;

[0013] extracting pump noise features from the last m plasma images based on the image segmentation model, and suppressing pump light noise of all plasma images based on the pump noise features;

[0014] A multi-head filter integrating Gaussian and OTSU filters was used to suppress global noise in all plasma images.

[0015] Further, the method of estimating the centroid position of each denoised image to obtain the plasma velocity between plasma images includes:

[0016] Generate a feature vector for each denoised image and normalize the feature vector;

[0017] The normalized feature vector is used as the input of the neural network to obtain the centroid position x of the plasma image. i ; Wherein, the neural network is based on a plurality of modules consisting of a fully connected layer, a batch normalization layer and a ReLU activation function;

[0018] Calculate the actual distance corresponding to each pixel in the plasma image and obtaining a time interval Δt between consecutive frames corresponding to the high-speed camera;

[0019] Based on the centroid position x i , the actual distance and the time interval Δt, obtaining the plasma velocity between every two plasma images.

[0020] Furthermore, a feature vector of each denoised image is generated, including:

[0021] Calculate the sum of pixel values ​​S in each column of the denoised image j ;

[0022] The pixel value sum S j The denoised images are combined into a one-dimensional feature vector with a dimension of 1×W, wherein W is the width of the denoised image.

[0023] Furthermore, the multi-head attention mechanism is used to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity, including:

[0024] The feature matrix is ​​distributed to n-1 attention heads for multi-head attention, and a multi-head attention calculation result of n×H dimensions is obtained, where H represents the height of the denoised image;

[0025] Using a fully connected layer to reduce the dimension of the multi-head attention calculation result to obtain an n×(n-1)-dimensional fully connected layer output;

[0026] Performing a transpose-sum operation on the n×(n-1)-dimensional fully connected layer output to obtain a (n-1)×n-dimensional transposed result;

[0027] The (n-1)×n dimensional transposed result is processed using a batch normalization layer to obtain a (n-1)×1 dimensional weight matrix.

[0028] A laser-induced plasma image fluid velocity inversion system based on deep learning, the system comprising:

[0029] A preprocessing module, for removing pump light noise and global background interference in each of n consecutive plasma images; wherein the n consecutive plasma images are generated by photographing the movement of plasma in a flow field using a high-speed camera;

[0030] A plasma state estimation module, used for obtaining the plasma velocity between plasma images by estimating the centroid position of each denoised image;

[0031] The fluid velocity inversion module is used to generate a feature matrix based on all denoised images, and use a multi-head attention mechanism to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity; according to the plasma velocity and the weight of the plasma velocity, the velocity of the plasma in the flow field is obtained.

[0032] An electronic device, characterized in that the electronic device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the laser induced plasma image fluid velocity inversion method based on deep learning as described in any one of the above items is implemented.

[0033] A computer-readable storage medium, characterized in that computer program instructions are stored on the computer-readable storage medium, and when the computer program instructions are executed by a processor, any of the above-mentioned deep learning-based laser-induced plasma image fluid velocity inversion methods is implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention proposes a novel deep learning-based network for estimating fluid velocity from laser-induced plasma images. The network estimates the center of mass position of plasma based on a deep neural network by designing a plasma center of mass position estimation module, and shows higher accuracy compared with the traditional bounding box center method, center of mass method, and weighted center of mass method. The network designs a multi-image perception attention module, comprehensively considers the position information and image quality weights of multiple plasma photos, rather than simply using the average velocity, further improving the accuracy and stability of fluid velocity estimation. While obtaining a more accurate center of mass position of plasma, the network can effectively reduce the uncertainty caused by the interference of plasma itself in fluid velocity measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the overall framework diagram of the laser induced plasma image fluid velocity estimation network based on deep neural network proposed in the present invention.

[0037] Figure 2 Comparison of the plasma center of mass position estimated by different methods: (a) Bounding Box Center Method (BBCM), (b) Geometric Center Method (GCM), (c) Weighted Center of Mass Method (WCM), and (d) Center of Mass Estimation Module (CEM).

[0038] Figure 3 It is the fluid velocity inversion weight information diagram. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below through specific implementations and in conjunction with the accompanying drawings.

[0040] The present invention provides a novel deep learning-based laser-induced plasma image flow velocity estimation network (FVEN), which consists of three stages: plasma image preprocessing, plasma state estimation, and flow velocity inversion. By tracking the changes in the centroid position, FVEN can accurately derive the flow velocity at different time intervals, and is significantly better than existing algorithms in terms of flow velocity estimation accuracy.

[0041] like Figure 1 As shown, the following takes 6 plasma images at each speed as an example to specifically illustrate the laser induced plasma image fluid velocity inversion method based on deep learning provided by the present invention.

[0042] Step 1: For n consecutive plasma images, remove the pump light noise and global background interference in each plasma image.

[0043] The present invention designs a local-global denoising module, which removes local pump light noise and global background interference respectively through a local noise segmenter and a multi-head filter, so as to realize the preprocessing of plasma images.

[0044] In one embodiment, the pre-processing includes the following steps 1.1 to 1.3.

[0045] Step 1.1: Among the six consecutive plasma images, acquire the last three plasma images.

[0046] After research, it is found that in the six consecutive plasma images, the pump noise images and plasma images in the last three plasma images do not overlap. Therefore, the present invention first obtains the last three plasma images of the consecutive plasma images.

[0047] Step 1.2: Based on the image segmentation model, extract the pump noise features from the last three plasma images, and suppress the pump light noise of all plasma images based on the pump noise features.

[0048] This embodiment uses the Segment Anything Model (SAM) as a local noise segmentor to extract the complete pump noise features from the last three plasma images at each speed. SAM can accurately segment the target object in various complex backgrounds through self-supervised learning and multi-scale feature fusion. The core of this model is to use deep learning technology to learn feature representation from large-scale annotated data, thereby realizing automatic segmentation of objects in new images.

[0049] Then, the pump noise feature is used to suppress the pump noise in the remaining images, thereby achieving the removal of pump light noise in the plasma image.

[0050] Step 1.3: Use a multi-head filter that integrates a Gaussian filter and an OTSU filter to suppress global noise in all plasma images.

[0051] In order to further obtain better denoised images, this embodiment also uses Gaussian filters and OTSU filters to suppress the global noise of all images to accurately segment the areas of the plasma images.

[0052] Step 2: The plasma velocity between plasma images is obtained by estimating the centroid position of each denoised image.

[0053] The present invention first introduces a multi-layer perceptron to estimate the centroid position, and then uses the change of the centroid position between consecutive images to calculate the plasma speed between consecutive images.

[0054] In one embodiment, the calculation process of the plasma velocity between plasma images includes the following steps 2.1 to 2.4.

[0055] Step 2.1: Generate a feature vector for each denoised image and normalize the feature vector.

[0056] In this embodiment, the sum of the pixel values ​​of each column j of the denoised image I of size H×W (where H represents the image height, W represents the image width, and I(i,j) represents the pixel value of the i-th row and j-th column of the image) is calculated. And sum S j They are combined into a one-dimensional feature vector f with a dimension of 1×W, which is normalized and used as the input of the neural network.

[0057] Step 2.2: Use the normalized feature vector as the input of the neural network to obtain the centroid position x of the plasma image i .

[0058] In this embodiment, a neural network is designed to estimate the centroid position of the plasma image. The first layer of the neural network is a fully connected layer, which maps the input from 320 dimensions to 560 dimensions, and uses a ReLU activation function and a batch normalization layer. The second to fifth layers are all fully connected layers, and batch normalization and ReLU activation functions are applied thereafter.

[0059] The x-coordinate of the actual center of mass is used as the label, and the x-coordinate of the center of mass predicted in this stage is used as the output of the output layer to obtain the center of mass position of the plasma image.

[0060] Step 2.3: Calculate the actual distance corresponding to each pixel in the plasma image And obtain the time interval Δt between consecutive frames corresponding to the high-speed camera.

[0061] The calculation of plasma velocity between plasma images also requires the actual distance corresponding to the pixel. and the time interval Δt between consecutive frames, so this embodiment acquires relevant parameters.

[0062] Step 2.4: Based on the centroid position x i , the actual distance and the time interval Δt, obtaining the plasma velocity between every two plasma images.

[0063] The present invention uses the change in the centroid position between consecutive images to calculate the plasma velocity between consecutive images, and the calculation method is as follows:

[0064]

[0065] Among them, x i and x i+1 are the estimated x coordinates in the i-th and i+1-th images respectively, represents the actual distance corresponding to each pixel, and Δt=8.33μs represents the time interval between consecutive frames.

[0066] Step 3: Generate a feature matrix based on all denoised images, and use the multi-head attention mechanism to distribute the feature matrix to 5 attention heads to obtain the weight of each plasma velocity; according to the plasma velocity and the weight of the plasma velocity, obtain the speed of plasma movement in the flow field.

[0067] In the fluid velocity inversion stage, the present invention utilizes a multi-image perception attention module and dynamically adjusts the weights of multiple plasma images in the velocity estimation task through an attention mechanism to perform fluid velocity inversion.

[0068] In one embodiment, the present invention sequentially splices the input of the plasma center of mass position estimation module under each flow rate condition into a 6×W feature matrix, and sequentially connects them into a 6×W feature matrix as the input of the multi-image attention module (MIAAM). Then, a multi-head attention mechanism is introduced to distribute the matrix on five attention heads. And the output of the attention head is linearly transformed for dimensional adjustment. Then, a transpose-sum operation is performed for integration, and then a batch normalization layer is processed to obtain five weight values. Finally, the five weights are multiplied by the calculated plasma velocity between consecutive images to obtain the final fluid velocity estimate.

[0069] Below, a specific experiment is used to illustrate the laser induced plasma image fluid velocity inversion method based on deep learning provided by the present invention.

[0070] This study uses a plasma image dataset captured by a high-speed camera combined with laser induced breakdown spectroscopy (LIBS) technology. The gas flow rate ranges from 10 to 120m / s and is subdivided into 12 discrete velocity groups, each with a velocity interval of 10m / s. These velocity groups include 10m / s, 20m / s, 30m / s, 40m / s, 50m / s, 60m / s, 70m / s, 80m / s, 90m / s, 100m / s, 110m / s, and 120m / s. For each velocity group, 100 experiments were performed, and the LIBS measurement system captured 6 clear images per experiment for fluid velocity inversion. The complete dataset contains 7,200 images.

[0071] like Figure 2As shown in the figure, the traditional methods for estimating the center of mass of plasma mainly include the bounding box center method, the geometric center method and the weighted center of mass method. The purpose of these methods is to extract the precise center of mass position from the preprocessed plasma image, and then estimate the fluid velocity based on the change of the center of mass position between consecutive images. The bounding box center method (BBCM) determines the minimum rectangular boundary that completely surrounds the plasma and takes the intersection of the diagonals of the rectangle as the center of mass. The effectiveness of this method depends on whether the plasma boundary can be accurately identified and whether the size of the bounding box can accurately reflect the position and shape of the plasma in the image. However, due to the different sizes of plasma and the unclear boundaries caused by extinguishing at different times, the center of mass position estimated by BBCM often has large errors. The geometric center method (GCM) directly regards the geometric center of the plasma in the image as the center of mass. This method only considers the geometric position of the plasma, not the pixel brightness. The weighted center of mass method (WCM) calculates the center of mass by analyzing the pixel values ​​of the plasma, and uses these values ​​as weight factors to determine the weighted average position. This method takes into account the self-evolution process of the plasma to a certain extent. However, none of the above traditional methods considers the influence of fluid velocity on the evolution of plasma morphology. To address this limitation, the present invention introduces a plasma centroid estimation module (CEM), which uses a multi-layer perceptron (MLP) to exploit the nonlinear modeling capability of neural networks, thereby improving the robustness of plasma image centroid estimation.

[0072] In addition, the FVEN proposed in the present invention adopts a phased training strategy. First, in the plasma state estimation stage, the plasma position estimation module (PPEM) is pre-trained using a multilayer perceptron. The training process uses the theoretical x-coordinate of the plasma center of mass derived under the real fluid velocity conditions as the label and minimizes the mean square error (MSE) as the loss function. In the first 150 training cycles, an initial learning rate of 0.001 and a batch size of 64 are used. Subsequently, in the fluid velocity inversion stage, the actual gas flow rate is used as the label, the mean absolute error (MAE) is used as the loss function, and the batch size is 32 to train the multi-input attention adaptation module (MIAAM, which is assumed to be the name of the relevant module and is not clear in the original text). A phased learning rate adjustment strategy is adopted: in the first 150 training cycles, a learning rate of 0.03 is used to promote rapid model learning; in the next 150 training cycles, the learning rate is reduced to 0.003 to fine-tune the model parameters; in the last 600 training cycles, the learning rate is further reduced to 0.0003 to stabilize the model performance and reduce overfitting. In both stages, the ratio of training set to test set is kept at 4:1. All experiments are performed on NVIDIA RTX 4060 GPU using PyTorch framework. This study involves velocity inversion under different flow velocity scenarios. In order to evaluate the performance of different fluid velocity inversion methods, mean absolute error (MAE) and mean absolute percentage error (PMAE) are used as evaluation indicators.

[0073] For the preprocessed images, the bounding box center method (BBCM), geometric center method (GCM), weighted centroid method (WCM) and centroid estimation module (CEM) were applied to estimate the plasma center of mass position. Figure 3 The plasma center of mass estimated by each method when the gas flow rate is 70m / s is shown. The white dotted line represents the ideal plasma center of mass position obtained based on the gas flow rate. It can be observed that at time = Δt, the center of mass positions estimated by the four methods are relatively close to the ideal position. However, as time develops to time = 2Δt and time = 3Δt, the deviation between BBCM and the ideal center of mass is significant. The results obtained by GCM and WCM are similar, and have higher accuracy than BBCM. In contrast, the plasma center of mass estimate provided by the CEM proposed in the present invention is closest to the ideal position.

[0074] Table 1 is a comparison of different fluid velocity inversion methods. Experimental results show that the deep learning-based laser induced plasma image fluid velocity inversion method provided by the present invention is significantly better than the existing algorithms in terms of fluid velocity estimation accuracy.

[0075] TABLE I:Comparison of different fluid velocity inversion methods.

[0076]

[0077] Table 1 Comparison results of different fluid velocity inversion methods.

[0078] Based on the same concept, the present invention also provides a laser induced plasma image fluid velocity inversion system based on deep learning, the system comprising:

[0079] A preprocessing module, for removing pump light noise and global background interference in each of n consecutive plasma images; wherein the n consecutive plasma images are generated by photographing the movement of plasma in a flow field using a high-speed camera;

[0080] A plasma state estimation module, used for obtaining the plasma velocity between plasma images by estimating the centroid position of each denoised image;

[0081] The fluid velocity inversion module is used to generate a feature matrix based on all denoised images, and use a multi-head attention mechanism to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity; according to the plasma velocity and the weight of the plasma velocity, the velocity of the plasma in the flow field is obtained.

[0082] The exemplary system is a system embodiment corresponding to the exemplary method described above. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0083] Based on the same concept, the present invention also provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements any of the above-mentioned deep learning-based laser induced plasma image fluid velocity inversion methods.

[0084] Based on the same concept, the present invention also provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, any of the above-mentioned deep learning-based laser induced plasma image fluid velocity inversion methods is implemented.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inverting fluid velocity from laser-induced plasma images based on deep learning, characterized in that: The method comprises: For n consecutive plasma images, removing pump light noise and global background interference in each plasma image; wherein the n consecutive plasma images are generated by photographing the movement of plasma in a flow field using a high-speed camera; The plasma velocity between plasma images is obtained by estimating the centroid position of each denoised image; Generate a feature matrix based on all denoised images, and use a multi-head attention mechanism to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity; The moving speed of the plasma in the flow field is obtained according to the plasma speed and the weight of the plasma speed.

2. The method according to claim 1, characterized in that The removing of local pump light noise and global background interference in each plasma image comprises: Among the consecutive n plasma images, the last m plasma images in the time dimension are obtained, m <n; extracting pump noise features from the last m plasma images based on the image segmentation model, and suppressing pump light noise of all plasma images based on the pump noise features; A multi-head filter integrating Gaussian and OTSU filters was used to suppress global noise in all plasma images.

3. The method according to claim 1, characterized in that The method of estimating the centroid position of each denoised image to obtain the plasma velocity between plasma images comprises: Generate a feature vector for each denoised image and normalize the feature vector; The normalized feature vector is used as the input of the neural network to obtain the centroid position x of the plasma image. i ; Wherein, the neural network is based on a plurality of modules consisting of a fully connected layer, a batch normalization layer and a ReLU activation function; Calculate the actual distance corresponding to each pixel in the plasma image and obtaining a time interval Δt between consecutive frames corresponding to the high-speed camera; Based on the centroid position x i , the actual distance and the time interval Δt, obtaining the plasma velocity between every two plasma images.

4. The method according to claim 2, characterized in that: Generate a feature vector for each denoised image, including: Calculate the sum of pixel values ​​S in each column of the denoised image j ; The pixel value sum S j The denoised images are combined into a one-dimensional feature vector with a dimension of 1×W, wherein W is the width of the denoised image.

5. The method according to claim 1, characterized in that The multi-head attention mechanism is used to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity, including: The feature matrix is ​​distributed to n-1 attention heads for multi-head attention, and a multi-head attention calculation result of n×H dimensions is obtained, where H represents the height of the denoised image; Using a fully connected layer to reduce the dimension of the multi-head attention calculation result to obtain an n×(n-1)-dimensional fully connected layer output; Performing a transpose-sum operation on the n×(n-1)-dimensional fully connected layer output to obtain a (n-1)×n-dimensional transposed result; The (n-1)×n dimensional transposed result is processed using a batch normalization layer to obtain a (n-1)×1 dimensional weight matrix.

6. A laser-induced plasma image fluid velocity inversion system based on deep learning, characterized in that: The system comprises: A preprocessing module, for removing pump light noise and global background interference in each of n consecutive plasma images; wherein the n consecutive plasma images are generated by photographing the movement of plasma in a flow field using a high-speed camera; A plasma state estimation module, used for obtaining the plasma velocity between plasma images by estimating the centroid position of each denoised image; The fluid velocity inversion module is used to generate a feature matrix based on all denoised images, and use a multi-head attention mechanism to distribute the feature matrix to n-1 attention heads to obtain the weight of each plasma velocity; according to the plasma velocity and the weight of the plasma velocity, the velocity of the plasma in the flow field is obtained.

7. An electronic device, characterized in that: The electronic device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the deep learning-based laser induced plasma image fluid velocity inversion method as described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by the processor, the deep learning-based laser-induced plasma image fluid velocity inversion method as described in any one of claims 1 to 5 is implemented.

Images