Pulverized coal concentration detection method based on MobileViT

Abstract

The invention relates to a pulverized coal concentration detection method based on MobileViT. The method comprises the following steps: acquiring a microwave signal attenuation value corresponding to a pulverized coal gas-solid mixture in a detected space through a microwave sensor; the microwave sensor comprises N electrodes which are annularly and uniformly distributed; preprocessing the original pulverized coal concentration data set; the preprocessed data set is input into an improved MobileViT model for training; and acquiring a microwave signal attenuation value of the pulverized coal conveying pipeline in real time, preprocessing the microwave signal attenuation value, inputting the preprocessed microwave signal attenuation value into the trained MobileViT model, and outputting a pulverized coal concentration detection result. The method has the beneficial effects that spatial information of numerical values is completely reserved and utilized in the calculation process, two-dimensional imaging processing is performed on one-dimensional data, features are scientifically extracted, and concentration detection is effectively performed, so that the pulverized coal concentration is detected by using MobileViT.

Description

Technical Field

[0001] The present invention relates to the technical field of pulverized coal concentration detection, and more precisely, it relates to a method for detecting pulverized coal concentration based on MobileViT. Background Art

[0002] In a coal combustion system, accurately measuring the pulverized coal concentration is a complex but crucial step. As the main fuel for boilers, the concentration of pulverized coal directly affects the stability and efficiency of the combustion process. Effective monitoring of the pulverized coal concentration will not only help control the emissions of harmful substances during the production process but also optimize the operating parameters of each production link and improve energy utilization efficiency. Excessive or insufficient pulverized coal may also cause blockages or even deflagrations in the pipeline, resulting in inevitable losses.

[0003] The difficulty in detecting the pulverized coal concentration lies in its complex flow characteristics and measurement environment limitations. During the transportation of pulverized coal, the flow pattern is prone to change over time, and the size and position distribution of pulverized coal particles are uneven, making the dynamic characteristics of the pulverized coal airflow extremely complex. At the same time, the long-term high-speed erosion of pulverized coal in the pipeline is likely to cause pollution and damage to the detection equipment, which also poses challenges to the detection of pulverized coal concentration.

[0004] In common microwave attenuation techniques, numerical calculations are usually focused on analyzing data. However, numerical calculation analysis has the following problems: First, it cannot solve the problem of uneven distribution during the transportation of pulverized coal, which is accidental; second, it does not start from the physical meaning of the sensor itself and ignores the spatial information contained in the numerical values, which is one-sided; third, pure numerical calculations are difficult to handle complex non-linear relationships and may not be able to provide effective solutions; fourth, data augmentation usually relies on transforming and expanding images, videos, or other forms of spatial data, and simply using data cannot perform spatial transformations such as rotation, scaling, cropping, or flipping to increase the diversity and richness of data; fifth, it cannot be visualized, and the result presentation is not intuitive enough. Summary of the Invention

[0005] The purpose of the present invention is to address the deficiencies of the prior art and propose a method for detecting pulverized coal concentration based on MobileViT.

[0006] In the first aspect, a method for detecting pulverized coal concentration based on MobileViT is provided, including:

[0007] S1. Collect the microwave signal attenuation values corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed, and multiple groups of microwave signal attenuation values are collected by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels, forming an original pulverized coal concentration data set;

[0008] S2. Preprocess the original pulverized coal concentration dataset;

[0009] S3. Input the preprocessed dataset into the improved MobileViT model for training. The improved MobileViT model includes a convolutional layer for feature extraction, a global average pooling layer, and a classification layer;

[0010] S4. Real-time collect the microwave signal attenuation value of the pulverized coal conveying pipeline, preprocess it, and then input it into the trained MobileViT model to output the pulverized coal concentration detection result.

[0011] Preferably, in S1, multiple signal paths are constructed through a host computer, a microwave signal conditioning system, and a microwave sensor; the original pulverized coal concentration dataset is obtained according to the multiple signal paths.

[0012] Preferably, S2 includes:

[0013] S201. Subtract the microwave signal attenuation value collected under each concentration label from the microwave signal attenuation value under the reference concentration label to obtain the relative change in microwave signal attenuation;

[0014] S202. Perform data cleaning. Divide each signal path into different dimensions, and then use the Z-score index to truncate the singular values in each dimension;

[0015] S203. Encode the microwave signal attenuation value into a matrix form according to the transmitting electrode number T and the receiving electrode number R. After filling in the missing values, use the maximum-minimum normalization to unify the dimension, and then store all the matrices in the form of a bitmap;

[0016] S204. Perform data augmentation according to the rule of cyclic shift.

[0017] Preferably, in S3, during the training process, select the loss in the execution of the cross-entropy loss calculation task, and use the stochastic gradient descent as the optimizer to minimize the loss function by updating the model parameters in the direction of the negative gradient of the loss function.

[0018] Preferably, in S204, the rule of cyclic shift includes: the pixel points in the original sample are longitudinally translated row by row to obtain a new sample.

[0019] In the second aspect, a pulverized coal concentration detection system based on MobileViT is provided for executing any method in the first aspect, including:

[0020] A collection module, configured to collect the microwave signal attenuation value corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed, and collects multiple groups of microwave signal attenuation values by sequentially combining electrode pairs, and the multiple groups of microwave signal attenuation values have different concentration labels, forming an original pulverized coal concentration data set;

[0021] A preprocessing module, configured to preprocess the original pulverized coal concentration data set;

[0022] A training module, configured to input the preprocessed data set into an improved MobileViT model for training, and the improved MobileViT model includes a convolutional layer, a global average pooling layer and a classification layer for feature extraction;

[0023] An output module, configured to collect the microwave signal attenuation value of the pulverized coal conveying pipeline in real time, input it into the trained MobileViT model after preprocessing, and output the pulverized coal concentration detection result.

[0024] In a third aspect, a pulverized coal concentration detection device based on MobileViT is provided, which is used to execute any of the methods in the first aspect, and includes: a host computer, a microwave signal conditioning system and a microwave sensor;

[0025] Wherein, the host computer and the microwave signal conditioning system are communicatively connected, and the microwave signal conditioning system and the microwave sensor are communicatively connected.

[0026] In a fourth aspect, a computer storage medium is provided, and a computer program is stored in the computer storage medium; when the computer program runs on the computer, the computer is enabled to execute any of the methods in the first aspect.

[0027] In a fifth aspect, an electronic device is provided, including:

[0028] A memory, configured to store a computer program;

[0029] A processor, configured to execute the computer program to implement any of the methods in the first aspect.

[0030] The beneficial effects of the present invention are:

[0031] 1. Based on the physical meaning of the sensor itself, according to the arrangement positions of the electrodes on the sensor, the microwave attenuation values obtained for each electrode pair are placed on the corresponding pixel points to obtain the bitmap to be subsequently input into the model. Then, the MobileViT network, which is characterized by being lightweight and having low latency, is used for image classification to adapt to the offline and high-speed real industrial scenarios. During this process, the present invention completely retains and utilizes the spatial information of the numerical values during the calculation process, performs two-dimensional imaging processing on the one-dimensional data, scientifically extracts features, and effectively conducts concentration detection to achieve the detection of pulverized coal concentration using MobileViT.

[0032] 2. Due to the random distribution of pulverized coal and the complex flow process, it is difficult to grasp the distribution of the entire cross-section by relying solely on a single pair of electrodes for linear measurement. The present invention uses multiple electrodes surrounding the surface of the sensor for full cross-section measurement, which can make up for this deficiency and increase the reliability.

[0033] 3. Different from general data augmentation methods for images such as scaling, flipping, rotating, random cropping, and randomly adding masks, the cyclic shift method in the present invention completes data augmentation from a physical perspective. It is equivalent to rotating the sensor by an angle of k / N * 360° each time (where k is the number of shifts and N is the total number of electrodes). Although the numbers of each electrode change, it does not affect their spatial relationships with each other. Therefore, it does not affect the signal paths represented by each electrode pair in each group, and their responses to different concentrations will not change due to the change in numbers. Furthermore, the present invention adopts the rule of cyclic shift for data augmentation to increase the number of samples.

[0034] 4. The present invention is designed for a specific hardware system, but can be applied to different parameters of such hardware systems, such as the size and material of the sensor, without affecting the effectiveness of the method, and thus has high generality. Description of the Drawings

[0035] Figure 1 Schematic diagram of the multi-frequency pulverized coal concentration measurement system provided by this application;

[0036] Figure 2 Schematic diagram of the structure of the pulverized coal conveying process provided by this application;

[0037] Figure 3 Schematic diagram of the matrix form representation of the sample data provided by this application;

[0038] Figure 4 Bitmap of the sample data provided by this application;

[0039] Figure 5 MobileViT network structure diagram provided by this application;

[0040] Figure 6 Flow chart of the pulverized coal concentration measurement method based on multiple frequencies and full cross-section provided by this application;

[0041] Figure 7 Loss / accuracy curve provided by this application;

[0042] Figure 8 Schematic diagram of the confusion matrix provided by this application;

[0043] Explanation of reference numerals: host computer 1, microwave signal conditioning system 2, microwave sensor 3, boiler 4, pulverized coal conveying pipe 5, coal mill 6, object to be measured 7, microwave sensing system 8. Specific embodiments

[0044] The present invention will be further described below in conjunction with embodiments. The description of the following embodiments is only for helping to understand the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several modifications can still be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0045] Embodiment 1:

[0046] When microwave signals collide with the gas-solid mixture in the measured space, attenuation will occur. This attenuation value is often related to the dielectric constant of the gas-solid mixture, and the dielectric constant of the gas-solid mixture is in turn related to the concentrations of the solid and gas phases therein. In the application scenario of this application, it is manifested that by measuring the attenuation value of microwave signals, the pulverized coal concentration value in the pulverized coal airflow can be indirectly obtained.

[0047] To solve the problems of the prior art, Embodiment 1 of this application provides a pulverized coal concentration detection method based on MobileViT. Since there is currently no consensus solution for the current pulverized coal concentration measurement, this method, in cooperation with a microwave sensing system that can freely adjust the working frequency and signal path, can realize real-time measurement of pulverized coal concentration. Specifically, the method includes:

[0048] S1. Collect the attenuation values of microwave signals corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed, and multiple groups of microwave signal attenuation values are collected by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels, forming an original pulverized coal concentration data set.

[0049] In S1, multiple signal paths are constructed through the host computer 1, the microwave signal conditioning system 2, and the microwave sensor 3; the original pulverized coal concentration data set is obtained according to the multiple signal paths.

[0050] Specifically, during actual measurement, the electrodes on the microwave sensor with N electrodes are sequentially combined with the remaining electrodes to form electrode pairs, and microwave signals are passed between the electrode pairs, obtaining sets of microwave signal attenuation values. Set multiple concentration labels { }, and collect multiple sets of microwave signal attenuation values under each concentration label, then the original pulverized coal concentration data set can be obtained.

[0051] S2. Preprocess the original pulverized coal concentration data set.

[0052] S2 includes:

[0053] S201. Subtract the microwave signal attenuation value collected under each concentration label from the microwave signal attenuation value under the reference concentration label to obtain the relative change in microwave signal attenuation;

[0054] S202. Perform data cleaning. Divide each signal path into different dimensions, and then use the Z-score index to truncate the singular values in each dimension.

[0055] Specifically, according to different electrode pairs, use the Z-score index to truncate the singular values of the data set to achieve the effect of data cleaning.

[0056] S203. Encode the microwave signal attenuation values into matrix form according to the transmitting electrode number T and the receiving electrode number R. After filling in the missing values, use the maximum-minimum normalization to unify the dimension, and then store all the matrices in the form of bitmaps.

[0057] Specifically, represent each sample in matrix form, where the row number is the transmitting electrode number T, the column number is the receiving electrode number R, and use the mean filling method to process the missing values of T = R. After using the maximum-minimum normalization to unify the dimension, store all the matrices in the form of bitmaps. For example, Figure 3 is the matrix form representation of the sample data (taking 16 electrodes as an example). The row number is the transmitting electrode number T, the column number is the receiving electrode number R, and the number in each cell is the combination of the transmitting electrode number and the receiving electrode number (T, R).

[0058] S204. Perform data augmentation according to the rule of cyclic shift.

[0059] To increase the data volume and diversity of the samples, perform data augmentation on the preprocessed data set. Since the pulverized coal concentration data set is generated based on a fixed sensor structure and electrode numbers, each pixel point in the sample corresponds to a preset transmitting electrode and receiving electrode. At the same time, the electrodes on the microwave sensor have the geometric feature of being evenly distributed in a ring, and the electrodes are numbered continuously in a clockwise manner. Therefore, adopt the rule of cyclic shift to globally adjust the electrode numbering system.

[0060] In S204, the rules of the cyclic shift include: the pixel points in the original sample are longitudinally translated row by row to obtain a new sample.

[0061] S3. Input the preprocessed data set into the improved MobileViT model for training. The improved MobileViT model includes a convolutional layer for feature extraction, a global average pooling layer, and a classification layer.

[0062] Send the coal powder concentration data set after data augmentation into the lightweight and low-latency MobileViT model for training and prediction. In this network structure, MobileViT serves as the backbone network. Apply an n×n standard convolutional layer to the input tensor, and then apply a pointwise convolutional layer for feature extraction. Subsequently, use global average pooling as the neck and linear classification as the head to further process the features extracted by the backbone network and generate the final output of the model.

[0063] S4. Real-time collect the microwave signal attenuation value of the coal powder conveying pipeline, and input it into the trained MobileViT model after preprocessing to output the coal powder concentration detection result.

[0064] During actual measurement, use the preprocessed bitmap as the model input, and the prediction result of the coal powder concentration label can be obtained, realizing the effect of real-time detection of coal powder concentration. Figure 4 This is a bitmap example of sample data (taking 16 electrodes and 3 concentrations as an example). Each pixel point in the bitmap corresponds to Figure 3 each cell in. The pixel value is obtained after normalizing the dimensions of all values in the cell using the maximum and minimum values and then grayscaling.

[0065] Embodiment 2:

[0066] Based on Embodiment 1, Embodiment 2 of the present application provides a more specific method for detecting coal powder concentration based on MobileViT, as Figure 6 shown, including:

[0067] S1. Collect the microwave signal attenuation value corresponding to the coal powder gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed, and collect multiple groups of microwave signal attenuation values by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels to form an original coal powder concentration data set.

[0068] Specifically, in Figure 1In the structural schematic diagram, the host computer 1 controls the microwave signal conditioning system 2 to send microwave signals of a specified frequency. After processing such as amplification, attenuation, and filtering, it is connected to any two electrodes in the microwave sensor 3 to form a signal path, and then the signal is sent back to the microwave signal conditioning system 2 for detection processing. Finally, the microwave signal reaches the host computer 1 in the form of a digital signal. Each measurement can adjust the signal path of the microwave sensor 3 according to the host computer 1, that is, the N electrodes on the sensor are sequentially combined with the remaining electrodes to form electrode pairs, and the microwave signal is passed between the electrode pairs, and sets of microwave signal attenuation values can be obtained. Set multiple concentration labels { }, where represents that the pulverized coal concentration is 0. Under each concentration label, multiple sets of microwave signal attenuation values are collected, and the original pulverized coal concentration data set can be obtained, and its size is .

[0069] It should be noted that the size of the microwave pulverized coal concentration data set used in this method depends on the number of electrodes of the microwave sensor 3. When the number of electrodes is N, a total of sets of microwave signal attenuation values can be obtained. Therefore, the size of the microwave pulverized coal concentration data set is not fixed, and this method is applicable to all sizes. In addition, the number of concentration labels in this method is not fixed and can be subdivided according to the usage scenario and requirements.

[0070] S2. Preprocess the original pulverized coal concentration data set.

[0071] Specifically, preprocess the data set. Among them, the first step is to subtract the microwave signal attenuation value collected under each concentration from the microwave signal attenuation value to obtain the relative change in microwave signal attenuation. This relative value is the basis for subsequent judgment of pulverized coal concentration. The second step is data cleaning. Each signal path is divided into different dimensions, and then the Z-score index is used to truncate the singular values in each dimension to eliminate the random error introduced by the unevenness of pulverized coal flow during the measurement process. The formula is as follows:

[0072]

[0073] In the formula, is the original data, , is the average value of the original data set, is the standard deviation of the original data set, is 's Z-score index, is the signal path number.

[0074] The measured sample values are represented in matrix form, where the row number is the transmitting electrode number T, the column number is the receiving electrode number R, and the missing values of T = R are filled with the mean value of the T-th row, as Figure 3 shown. In addition, there is no fixed matrix encoding method in this method, which is manifested in that only the corresponding electrode pairs in each grid of each matrix need to be fixed. At the same time, the processing method for missing values can also be replaced by other data processing methods, such as deletion, filling, or interpolation, etc. The method of unifying the dimension can also be replaced by other methods, such as normalization, centering, and initialization, etc.

[0075] After using the maximum-minimum normalization to unify the dimension, all matrices are stored in the form of bitmaps, that is, the preprocessed pulverized coal concentration data set is obtained. The obtained bitmap is as Figure 4 shown.

[0076] Next, it is necessary to perform data augmentation on the preprocessed pulverized coal concentration data set to increase the data volume and diversity of the samples. The overall adjustment of the electrode number system is carried out according to the rule of cyclic shift. In this process, the pixel points in the original sample are longitudinally translated row by row to obtain new samples. The formula for cyclic shift is as follows:

[0077]

[0078] where j is the new electrode number, i is the original electrode number, k is the number of bits shifted to the right in a cycle, and N is the total number of electrodes.

[0079] S3. Input the preprocessed data set into the improved MobileViT model for training. The improved MobileViT model includes a convolutional layer for feature extraction, a global average pooling layer, and a classification layer. The improved MobileViT network structure is as Figure 5 shown in (b) in

[0080] Specifically, a lightweight and low-latency MobileViT network structure is used to train the data set. Figure 5 (a) in Figure 5 is the traditional MobileViT network structure. As the backbone network, MobileViT applies an n×n standard convolutional layer to the input tensor, and then applies a pointwise convolutional layer for feature extraction. Subsequently, global average pooling is used as the neck and linear classification is used as the head to further process the features extracted by the backbone network and generate the final output of the model. At the same time, if the bitmap size generated in the preprocessing process of this method is small and the number of categories in the image classification task is small, a lighter network based on the MobileViT network structure can be used. Figure 5In (b) is a lighter MobileViT network structure. To retain more information and resolution in the image, downsampling is abandoned during the processing, and 1×1 convolutional kernels are used instead of 3×3 convolutional kernels, thereby reducing the network's parameters and number of channels, decreasing the computational amount and memory usage, and improving the training speed. In the figure, MV2 is MobileNetV2, is 2-fold downsampling, h and w are the height and width of each patch, and L is the number of Transformer layers. During the training process, cross-entropy loss is selected to calculate the loss during task execution, and stochastic gradient descent is used as the optimizer. The model parameters are updated in the direction of the negative gradient of the loss function to minimize the loss function. The batch size is set to 32, and the initial learning rate is 0.001, which will be updated according to a linear decay strategy. The number of iterations is determined by the training effect. In addition, the structure of the MobileViT network used in this method can be appropriately reduced according to the bitmap size and the number of categories of the classification task, such as abandoning downsampling or using smaller convolutional kernels, etc.

[0081] Finally, after comprehensively evaluating the cross-entropy loss amount and accuracy, a trained model and parameters can be obtained.

[0082] S4. Real-time collect the microwave signal attenuation value of the pulverized coal conveying pipeline, and after preprocessing, input it into the trained MobileViT model to output the pulverized coal concentration detection result.

[0083] Specifically, during each measurement, only the microwave signal attenuation value of the entire cross-section needs to be measured, and after converting it into a bitmap according to a preset rule and sending it into the model, the predicted pulverized coal concentration can be obtained. On the premise of non-invasion, it is possible to achieve rapid and accurate classification of the pulverized coal concentration during the pneumatic conveying of pulverized coal.

[0084] In addition, this method is not only applicable to the pulverized coal concentration measurement scenario, but can also be used in similar gas-solid two-phase flow scenarios, such as in application fields like minerals, food, and building materials.

[0085] This application uses artificial intelligence methods to deal with the complex fluid mechanics during the flow of pulverized coal. Through the analysis and prediction of algorithms, it effectively avoids the limitations and high computational costs brought by simply relying on pure numerical calculations. During the model construction process, a total of 67,500 groups of samples were collected and divided into a training set, a validation set, and a test set according to a ratio of 6:2:2. The model training results show that as Figure 7 shown, as the number of iterations increases, the loss function value steadily decreases, the model accuracy significantly improves, and it tends to converge after about 100 iterations. As Figure 8As shown, the analysis through the confusion matrix indicates that the degree of coincidence between the predicted concentration and the actual concentration reaches over 93%, fully verifying the effectiveness and reliability of the artificial intelligence method in coal powder concentration prediction and successfully achieving the accurate learning and prediction of the characteristics of complex datasets.

[0086] In addition, this application extends the measured area from a one-dimensional straight line to a two-dimensional cross-section to avoid the problem of insufficient detection resolution caused by uneven coal powder distribution. To verify this effect, we conducted a comparative experiment: as shown in Table 1, using 2 electrodes and 16 electrodes as the input of the microwave sensor respectively, after constructing a prediction model, the recall rate was used as the evaluation index. The experimental results show that the recall rate of two-dimensional detection in each concentration range is significantly better than that of one-dimensional detection, with an average improvement of 31.7%, fully demonstrating the optimization effect of the two-dimensional detection scheme on the detection process.

[0087] Table 1 Comparative experiment of one-dimensional and two-dimensional detection

[0088]

[0089] It can be seen that this application can represent all one-dimensional microwave attenuation values on the cross-section using a two-dimensional image, increasing the interpretability of the results, and can, starting from the physical meaning of the sensor, mine the spatial information in the values for data enhancement, effectively increasing the amount of data and overcoming the limitation that pure numerical calculations cannot make full use of this spatial information for data enhancement.

[0090] It should be noted that the same or similar parts in this embodiment and Embodiment 1 can be referred to each other and will not be elaborated in this application.

[0091] Embodiment 3:

[0092] Based on Embodiment 1, Embodiment 3 of this application provides a coal powder concentration measurement device, as Figure 1 shown, including a host computer 1, a microwave signal conditioning system 2, and a microwave sensor 3. The arrow direction is the direction of the main signal or information, which is divided into a DC signal and a microwave signal. The components of the multi-frequency coal powder concentration measurement system communicate bidirectionally to complete data acquisition and instruction issuance.

[0093] Figure 2 It is a structural schematic diagram of the coal powder conveying process, including a boiler 4, a coal powder conveying pipe 5, and a coal mill 6. The coal powder flow direction is from bottom to top. After being split from the coal mill 6 into multiple coal powder conveying pipes 5, it is fed into the boiler 4 for combustion from different angles respectively.

[0094] It should be noted that the device provided in this embodiment is the device corresponding to the method provided in Embodiment 1. Therefore, the same or similar parts in this embodiment and Embodiment 1 can be referred to each other and will not be elaborated in this application.

[0095] Example 4:

[0096] Based on Example 1, Example 4 of the present application provides a pulverized coal concentration detection system based on MobileViT, including:

[0097] A collection module for collecting the microwave signal attenuation value corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed, and multiple groups of microwave signal attenuation values are collected by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels to form an original pulverized coal concentration data set;

[0098] A preprocessing module for preprocessing the original pulverized coal concentration data set;

[0099] A training module for inputting the preprocessed data set into an improved MobileViT model for training. The improved MobileViT model includes a convolutional layer, a global average pooling layer, and a classification layer for feature extraction;

[0100] An output module for collecting the microwave signal attenuation value of the pulverized coal conveying pipeline in real time, inputting it into the trained MobileViT model after preprocessing, and outputting the pulverized coal concentration detection result.

[0101] It should be noted that the system provided in this embodiment is the system corresponding to the method provided in Example 1. Therefore, for the parts that are the same or similar in this embodiment and Example 1, they can be referred to each other and will not be elaborated in this application.

Claims

1. A method for detecting pulverized coal concentration based on MobileViT, characterized in that, Including: S1. Collect the microwave signal attenuation value corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; The microwave sensor includes N annular electrodes evenly distributed. Multiple groups of microwave signal attenuation values are collected by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels, forming an original pulverized coal concentration data set; S2. Preprocess the original pulverized coal concentration data set; S3. Input the preprocessed data set into an improved MobileViT model for training. The improved MobileViT model includes a convolutional layer, a global average pooling layer, and a classification layer for feature extraction; S4. Real-time collect the microwave signal attenuation value of the pulverized coal conveying pipeline, input it into the trained MobileViT model after preprocessing, and output the pulverized coal concentration detection result.

2. The coal powder concentration detection method based on MobileViT according to claim 1, characterized in that In S1, multiple signal paths are constructed through a host computer, a microwave signal conditioning system, and a microwave sensor; the original pulverized coal concentration data set is obtained according to the multiple signal paths.

3. The method for detecting pulverized coal concentration based on MobileViT according to claim 2, wherein S2 Including: S201. Subtract the microwave signal attenuation value collected under each concentration label from the microwave signal attenuation value under the reference concentration label to obtain the relative change in microwave signal attenuation; S202. Perform data cleaning. Divide each signal path into different dimensions, and then use the Z-score index to truncate the singular values in each dimension; S203. Encode the microwave signal attenuation value into a matrix form according to the transmitting electrode number T and the receiving electrode number R. After filling in the missing values, use the maximum-minimum normalization to unify the dimension, and then store all the matrices in the form of a bitmap; S204. Perform data augmentation according to the rule of cyclic shift.

4. The method for detecting pulverized coal concentration based on MobileViT according to claim 3, wherein In S3, during the training process, select the loss in the execution of the cross-entropy loss calculation task, and use stochastic gradient descent as the optimizer to minimize the loss function by updating the model parameters in the direction of the negative gradient of the loss function.

5. The method for detecting pulverized coal concentration based on MobileViT according to claim 4, characterized in that In S204, the rule of cyclic shift includes: the pixel points in the original sample are longitudinally translated row by row to obtain a new sample.

6. A pulverized coal concentration detection system based on MobileViT, characterized in that, For implementing the method according to any one of claims 1 to 5, including: A collection module for collecting the microwave signal attenuation value corresponding to the pulverized coal gas-solid mixture in the measured space through a microwave sensor; the microwave sensor includes N annular electrodes evenly distributed. Multiple groups of microwave signal attenuation values are collected by sequentially combining electrode pairs. The multiple groups of microwave signal attenuation values have different concentration labels, forming an original pulverized coal concentration data set; A preprocessing module for preprocessing the original pulverized coal concentration data set; A training module for inputting the preprocessed data set into an improved MobileViT model for training. The improved MobileViT model includes a convolutional layer, a global average pooling layer, and a classification layer for feature extraction; An output module for real-time collecting the microwave signal attenuation value of the pulverized coal conveying pipeline, inputting it into the trained MobileViT model after preprocessing, and outputting the pulverized coal concentration detection result.

7. A pulverized coal concentration detection device based on MobileViT, characterized in that, For implementing the method according to any one of claims 1 to 5, including: a host computer, a microwave signal conditioning system, and a microwave sensor; Among them, the host computer is communicatively connected to the microwave signal conditioning system, and the microwave signal conditioning system is communicatively connected to the microwave sensor.

8. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program runs on a computer, it causes the computer to execute the method according to any one of claims 1 to 5.

9. An electronic device, characterized in that, Including: A memory for storing the computer program; A processor for executing the computer program to implement the method according to any one of claims 1 to 5.

Images