An atmospheric particulate matter concentration detection method, device, medium and product

Abstract

The application discloses an atmospheric particulate matter concentration detection method and device, medium and product, and relates to the field of atmospheric particulate matter detection. The method comprises the following steps: adsorbing an atmospheric particulate matter sample according to a film, determining a corresponding visible near-infrared reflectance spectrum and a particulate matter mass concentration true value; performing spectrum conversion by using a ring mapping method according to the visible near-infrared reflectance spectrum; constructing a feature extractor by using a ShuffleNetV2 network improved based on an attention mechanism according to a ring mapping image; performing feature fusion on spectral features and corresponding image features; constructing a film-adsorbed particulate matter concentration rapid quantitative detection model by using a machine learning method according to the fused features and the corresponding particulate matter mass concentration true value; and detecting the atmospheric particulate matter concentration by using the film-adsorbed particulate matter concentration rapid quantitative detection model. The application can improve the timeliness and accuracy of film-adsorbed particulate matter mass concentration detection, and greatly reduce the labor and equipment costs.

Description

Technical Field

[0001] This application relates to the field of atmospheric particulate matter detection, and in particular to a method, equipment, medium, and product for detecting atmospheric particulate matter concentration. Background Technology

[0002] With rapid economic development and population growth worldwide, increasing amounts of pollutants are being released into the environment, significantly raising atmospheric particulate matter levels and severely impacting air quality, climate conditions, the ecological environment, and human health. In 2013, the International Agency for Research on Cancer (IARC), a branch of the WHO, officially classified atmospheric particulate matter as a Group 1 carcinogen, recognizing it as a major cause of environmental health problems and a widely distributed environmental carcinogen. Particulate matter can remain suspended in the atmosphere for extended periods and travel long distances; its mass concentration, chemical composition, and particle size are the main factors determining its degree of harm. The higher the concentration of particulate matter, the greater the harm to human health.

[0003] In atmospheric particulate matter hazard risk analysis, particulate matter in the air is typically collected continuously using a suspended particulate matter sampler and adsorbed onto a sampling membrane. The mass concentration and physicochemical properties of the particles on the membrane are then determined. Currently, the national standard for membrane particulate matter determination methods includes the gravimetric method and the Beta ray method. The gravimetric method calculates particulate matter concentration by weighing the difference in mass of the filter membrane before and after sampling. It is the benchmark method for particulate matter mass concentration determination, with a simple principle and reliable data. However, it is easily affected by environmental humidity, and the filter membrane needs to be equilibrated in a constant temperature and humidity environment for a considerable period before and after mass determination, which is time-consuming, labor-intensive, and has poor timeliness. The Beta ray method calculates particulate matter concentration using the linear relationship between the amount of Beta ray absorbed by particulate matter and its mass. It is less affected by interference, but Beta ray detection equipment is expensive, with high operating costs and maintenance requirements, and it requires pre-calibration according to standard reference methods. Therefore, developing a relatively low-cost, efficient, and stable membrane particulate matter mass concentration detection technology is of great significance for improving the timeliness of particulate matter analysis.

[0004] Spectroscopic techniques refer to optical analysis methods based on the changes in wavelength and intensity of light, such as absorption, emission, and scattering, caused by the interaction of light (electromagnetic waves) with matter, resulting in transitions between quantized energy levels within atoms and molecules. Visible and near-infrared spectroscopy is based on the electromagnetic spectrum of molecular electronic transitions. These transitions cause the molecule to absorb light of a specific wavelength that matches its transition energy level, resulting in changes in the intensity of the reflected spectrum. Visible and near-infrared spectroscopy offers advantages such as simple sample preparation, non-destructive testing, high detection efficiency, low cost, portability, and ease of operation. It enables in-situ detection and online analysis and has been widely applied in online analysis across multiple fields, including chemical engineering, agriculture, and medicine.

[0005] The mass of particulate matter on a membrane affects its absorption of visible and near-infrared light. Therefore, the mass concentration of membrane-adsorbed particulate matter can be quantitatively detected by combining visible and near-infrared reflectance spectroscopy with machine learning models. However, traditional machine learning models based on visible and near-infrared spectroscopy suffer from poor robustness and low accuracy, and are easily affected by spectral fluctuations, making it difficult to meet the requirements for stable and accurate analysis of particulate matter mass concentration.

[0006] Therefore, in order to improve the timeliness of membrane adsorption particulate matter concentration detection and reduce the cost of manual equipment, there is an urgent need to provide a new method for detecting atmospheric particulate matter concentration. Summary of the Invention

[0007] The purpose of this application is to provide a method, device, medium, and product for detecting atmospheric particulate matter concentration, which can improve the timeliness and accuracy of membrane adsorption particulate matter mass concentration detection, and greatly reduce labor and equipment costs.

[0008] To achieve the above objectives, this application provides the following solution:

[0009] In a first aspect, this application provides a method for detecting atmospheric particulate matter concentration, the method comprising:

[0010] Based on the membrane adsorbed atmospheric particulate matter sample, determine the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration;

[0011] Based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter samples, the ring mapping method was used to perform spectrum conversion to obtain the ring mapping image.

[0012] Based on the ring map image, a feature extractor is constructed using the ShuffleNetV2 network, which is an improvement based on the attention mechanism. The feature extractor is used to extract image features, and a one-dimensional feature extraction algorithm is used for dimensionality reduction.

[0013] Feature fusion is performed on the spectral features corresponding to the visible and near-infrared reflectance spectra and the corresponding image features.

[0014] Based on the fusion characteristics and the corresponding true values ​​of particulate matter mass concentration, a rapid quantitative detection model for membrane-adsorbed particulate matter concentration is constructed using machine learning methods; and the atmospheric particulate matter concentration is detected using the rapid quantitative detection model for membrane-adsorbed particulate matter concentration.

[0015] Optionally, the step of determining the corresponding visible and near-infrared reflectance spectra and the true value of particulate matter mass concentration based on the membrane-adsorbed atmospheric particulate matter sample further includes:

[0016] Atmospheric particulate matter samples adsorbed by membranes were obtained using an atmospheric particulate matter sampler.

[0017] Optionally, the step of converting the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample into a ring-mapped image using a ring mapping method specifically includes:

[0018] The reflectance of the visible and near-infrared reflectance spectra is preprocessed; the preprocessing includes: multivariate scattering correction and smoothing filtering;

[0019] The spectrum was determined based on the pre-processed reflectance.

[0020] The spectral image is converted using the ring mapping method based on the spectral image to obtain the ring-mapped image.

[0021] Optionally, the ShuffleNetV2 network improved based on the attention mechanism is to replace the last fully connected layer in the ShuffleNetV2 network with an attention module and a global average pooling layer connected in sequence.

[0022] Optionally, the attention module includes two 1×1 convolutional layers.

[0023] Optionally, feature fusion is performed on the spectral features corresponding to the visible and near-infrared reflectance spectra and the corresponding image features, which also includes:

[0024] The average spectrum is determined based on the visible and near-infrared reflectance spectra.

[0025] The average spectrum is normalized by area.

[0026] The spectral features of the spectrum after area normalization are extracted using a feature extraction algorithm.

[0027] Secondly, this application provides an atmospheric particulate matter concentration detection device, the atmospheric particulate matter concentration detection device comprising:

[0028] The data determination module is used to determine the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration based on the membrane-adsorbed atmospheric particulate matter sample.

[0029] The ring mapping image determination module is used to perform spectrum conversion based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample to obtain a ring mapping image.

[0030] The feature extractor building module is used to construct a feature extractor based on the ring map image using an improved ShuffleNetV2 network with an attention mechanism. The feature extractor is used to extract image features and performs dimensionality reduction using a one-dimensional feature extraction algorithm.

[0031] The feature fusion module is used to fuse the spectral features corresponding to the visible and near-infrared reflectance spectra with the corresponding image features.

[0032] The concentration detection module is used to construct a rapid quantitative detection model for membrane-adsorbed particulate matter concentration based on fusion features and the corresponding true values ​​of particulate matter mass concentration using machine learning methods; and to use the rapid quantitative detection model for membrane-adsorbed particulate matter concentration to detect atmospheric particulate matter concentration.

[0033] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the atmospheric particulate matter concentration detection method.

[0034] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the atmospheric particulate matter concentration detection method described above.

[0035] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the atmospheric particulate matter concentration detection method.

[0036] According to the specific embodiments provided in this application, this application has the following technical effects:

[0037] This application provides a method, device, medium, and product for detecting atmospheric particulate matter concentration. It utilizes a ring mapping method for spectral conversion and constructs a feature extractor using an improved ShuffleNetV2 network based on an attention mechanism. This eliminates the shortcomings of traditional spectral-based machine learning methods, such as interference from spectral fluctuations and the inability to capture the overall trend of spectral lines, thus improving detection stability and accuracy. The feature extractor extracts image features from the ring-mapped image and fuses them with spectral features. A rapid quantitative detection model for membrane-adsorbed particulate matter concentration is then established using machine learning methods. This application extracts effective information, eliminates redundant information interference, and further improves detection accuracy and timeliness of particulate matter mass concentration detection on membranes through spectral feature fusion and simple machine learning methods, while significantly reducing labor and equipment costs. This application improves the robustness and accuracy of the rapid quantitative detection model for membrane-adsorbed particulate matter concentration by combining spectral conversion mapping with a lightweight deep learning network. It effectively achieves rapid and accurate detection of membrane-adsorbed atmospheric particulate matter concentration based on visible and near-infrared spectroscopy, reducing the interference of environmental factors on concentration analysis and providing a reference for efficient and accurate detection of atmospheric particulate matter. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of a method for detecting atmospheric particulate matter concentration in one embodiment of this application;

[0040] Figure 2 This is a schematic diagram of the overall process of an atmospheric particulate matter concentration detection method in one embodiment of this application;

[0041] Figure 3 A flowchart for determining the ring-shaped mapping image;

[0042] Figure 4 This is a scatter plot of the predicted particulate matter concentration results for the test set samples. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0044] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for detecting atmospheric particulate matter concentration is provided, which includes the following steps S101 to S105. Wherein:

[0046] S101. Based on the membrane adsorbed atmospheric particulate matter sample, determine the corresponding visible and near-infrared reflectance spectra and the true value of particulate matter mass concentration.

[0047] S101 also includes:

[0048] Atmospheric particulate matter samples adsorbed by membranes were obtained using an atmospheric particulate matter sampler.

[0049] Visible and near-infrared reflectance spectra of atmospheric particulate matter samples adsorbed by membranes were collected using a visible and near-infrared spectrometer. Thirty-two spectra were continuously collected for each membrane, and the reflectance was calculated. The mass method was used, and the mass of atmospheric particulate matter was weighed using a balance with a mass of 0.0001 g / L. The true mass concentration of atmospheric particulate matter was calculated by combining the mass of the sampled air with the mass of the atmospheric particulate matter.

[0050] S102. Based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample, the ring mapping method is used to perform spectrum conversion to obtain the ring mapping image.

[0051] S102 specifically includes:

[0052] The reflectance of the visible and near-infrared reflectance spectra is preprocessed; the preprocessing includes: multiplicative scattering correction (MSC) and Savitzky-Golay (SG) smoothing filtering;

[0053] The spectrum was determined based on the pre-processed reflectance.

[0054] The spectral image is converted using the ring mapping method based on the spectral image to obtain the ring-mapped image.

[0055] like Figure 3 As shown, with a fixed range of horizontal and vertical axes, a white background, and black spectral lines, a preprocessed reflectance spectrum is plotted. The spectral region is cropped and resized to 600×600 pixels. Each sample yields 32 spectra, all converted to grayscale, resulting in 32 two-dimensional matrices of 600×600 pixels, denoted as I1, I2, I3, ..., I... 32 .

[0056] Sixteen random integer sequences [1, 32] are generated, and 32 spectral images of the same sample are sorted according to these 16 random sequences. For the k-th sort, the matrices of the 32 reordered images are denoted as follows: Will Rotate 180°, and the resulting image is denoted as and Form pairs (n is a positive integer, n≤16), and insert a white image W with a size of 600×200 pixels between each pair of images. 600×200 That is, to splice them together in sequence W 600×200 , Create a new image with a size of 600×1400 pixels. (k, n are positive integers, k, n ≤ 16). [The rest of the text is missing.] Rotate by 11.25n°, fill the edges with white, expand to a new image of size 1524×1524 pixels, and invert the colors. Let the transformed image be denoted as . The new image U is obtained using the following formula. k :

[0057]

[0058] U k =(255) 1524×1524 -B k ;

[0059] S103. Based on the ring mapping image, a feature extractor is constructed using the ShuffleNetV2 network, which is an improvement based on the attention mechanism. The feature extractor is used to extract image features, and a one-dimensional feature extraction algorithm is used for dimensionality reduction.

[0060] The samples were randomly divided into training and testing sets at a 3:1 ratio for model training and testing, respectively. An improved ShuffleNetV2 model was constructed, using a pre-trained ShuffleNetV2 model (ShuffleNetV2_x0.5) as the feature extractor. The input channels of the initial convolutional layers were changed from 3 channels in RGB images to 1 channel to adapt to grayscale images. The last fully connected layer of the ShuffleNetV2 model was removed, and a custom attention module was introduced. The attention module contained two 1×1 convolutional layers. The first convolutional layer was used to reduce the number of channels, and the second convolutional layer was used to restore the number of channels. Channel-level attention weights were generated using the sigmoid function. These weights were then multiplied element-wise with the feature map to highlight the feature regions that the network considers more important. After applying the attention mechanism, the feature map was processed by a global average pooling layer, which compressed the spatial features of each channel into a single value, resulting in an output feature vector with a shape of 1024×1.

[0061] During training, the mean squared error loss function (MSELoss) was used as the loss function for the regression task. The Adam optimizer, combined with a learning rate scheduler (StepLR), was used to dynamically adjust the learning rate, ensuring that the model converged to a better local minimum in the later stages of training. To fully utilize the dataset and reduce the risk of overfitting, K-Fold Cross-Validation was employed to train the model, and the model with the smallest validation error was ultimately selected as the optimal model.

[0062] S104, feature fusion is performed on the spectral features corresponding to the visible and near-infrared reflectance spectra and the corresponding image features;

[0063] S104 also includes:

[0064] The average spectrum is determined based on the visible and near-infrared reflectance spectra.

[0065] The average spectrum is normalized by area.

[0066] The spectral features of the spectrum after area normalization are extracted using a feature extraction algorithm.

[0067] The annular mapping image of the sample is input into the improved ShuffleNetV2 model to extract image feature vectors. Each detected sample has 32 reflectance spectra, generating 16 spectral transformation images. The one-dimensional feature vectors extracted from these 16 images after adjustment by the ShuffleNetV2 network and attention mechanism are averaged to form a single feature vector, denoted as IF. The 32 spectra of the sample are averaged to form a single feature vector, denoted as SAN, after area normalization preprocessing. Spectral features and dimensionality-reduced image features are extracted using a one-dimensional feature extraction algorithm, and feature fusion is performed. Based on the fused features, a rapid quantitative detection model for membrane-adsorbed particulate matter concentration is established using machine learning methods.

[0068] S105. Based on the fusion characteristics and the corresponding true values ​​of particulate matter mass concentration, a rapid quantitative detection model for membrane-adsorbed particulate matter concentration is constructed using machine learning methods; and the rapid quantitative detection model for membrane-adsorbed particulate matter concentration is used to detect atmospheric particulate matter concentration.

[0069] This application converts visible and near-infrared reflectance spectral lines into images using a ring mapping method. Based on the spectral ring mapping image, a membrane adsorption particulate matter concentration detection model is established using a lightweight deep learning method. This eliminates the shortcomings of traditional spectral-based machine learning models, such as poor stability and difficulty in capturing overall trend information. The feature layers of the improved ShuffleNetV2 model are extracted and dimensionality reduced, and then fused with spectral features to establish a fast and accurate membrane adsorption particulate matter concentration detection model based on spectral feature fusion. Instead of directly outputting results through fully connected layers, the model fuses image features and spectral features, combined with machine learning algorithms, to improve the accuracy of particulate matter detection.

[0070] The following is an illustration through specific examples:

[0071] The sampling instrument used to collect atmospheric particulate matter samples adsorbed by the membrane was a Zhongrui Environmental Air Particulate Matter Comprehensive Sampler (ZR-2933, Qingdao, China), with a sampling flow rate of 100 L / min. The sample was obtained through PM2.5 sampling. 10 Cutter and PM 2.5The cutter cut particles with a diameter of 2.5 μm, and the sampling time for a single session was 3 hours. The filter membrane used was an 81 mm diameter quartz membrane, which was wrapped in aluminum foil before sampling and fired at 450℃ for 10 hours to remove initial organic matter. After natural cooling, it was equilibrated in a constant temperature and humidity chamber (HWS-250, China) for 24 hours, with the temperature set at 25 ± 1℃ and the humidity at 50 ± 5%. Twenty samples were taken at different locations, resulting in 20 images containing PM2.5. 2.5 The filter membrane is used for PM during sampling at each sampling point. 2.5 Mass concentration.

[0072] Will carry PM 2.5 The filter membranes were equilibrated in a constant temperature and humidity chamber for 24 hours (temperature set at 25±1℃, humidity set at 50±5%). Before and after sampling, the filter membranes were weighed using a 0.0001 g balance (Mettler Toledo XS105DU, Switzerland) and the weight was recorded. Each filter membrane was weighed three times and the average weight was taken. 2.5 The mass concentration was calculated by dividing the difference in filter membrane mass before and after sampling by the volume of air sampled. After weighing, the filter membrane was placed in a polystyrene filter membrane storage box and stored in a -4°C freezer.

[0073] Five 8mm diameter discs were cut from each sampling membrane to form parallel samples, resulting in a total of 100 membrane samples. The reflectance spectra of the samples were collected using a portable visible-near-infrared spectrometer (EE2063, Wuling Optics, Shanghai, China). The spectrometer's wavelength range was 180–1100 nm, with a resolution of 1 nm. A halogen lamp was used as the light source, connected to the spectrometer via a two-in-one fiber optic cable. The fiber optic acquisition end emitted the light source and collected the reflectance spectra. The membrane samples were placed in a self-made dark box, with the fiber optic acquisition end positioned 1 cm directly above the sample. Thirty-two reflectance spectra were continuously collected from each sample. The reflectance spectrum of the sample was calculated by comparing the dark spectrum collected under the same conditions with the reflectance spectrum from a white board.

[0074] Reflectance spectral lines in the band range of 400–1023 nm were extracted to eliminate the influence of the fluctuating band. The spectral lines were then preprocessed by multiplicative scattering correction (MSC) and Savitzky-Golay (SG) smoothing filtering. The parameter window size of the SG smoothing filter was set to 7 and the polynomial order was set to 2.

[0075] Using wavelength as the x-axis and reflectance as the y-axis, with a fixed x-axis range of 400–1023 nm and a y-axis range of 40–120%, and using white as the background and black as the spectral color, a reflectance spectrum without coordinate axes was plotted. After cropping the white edges, the image size was adjusted to 600×600 pixels, resulting in 32 spectra for each sample, thus generating 32 spectra.

[0076] The system generates 16 integer sequences [1, 32] and sorts 32 spectral images of the same sample based on these 16 random sequences. For the k-th sort, the matrices of the 32 reordered images are denoted as follows: Will Rotate 180°, and the resulting image is denoted as and Form pairs (n is a positive integer, n≤16), and insert a white image W with a size of 600×200 pixels between each pair of images. 600×200 That is, to splice them together in sequence W 600×200 , Create a new image with a size of 600×1400 pixels. (k, n are positive integers, k, n ≤ 16). [The rest of the text is missing.] Rotate counterclockwise by 11.25n°, fill the surrounding area with white, expand it into a new image with a size of 1524×1524 pixels, and invert the colors. Let the transformed new image be denoted as .

[0077] Each sample generates 16 spectral ring map images, for a total of 1600 images. The ring map image generation process is as follows: Figure 3 As shown.

[0078] The samples were randomly divided into a training set and a test set in a 3:1 ratio (parallel samples from the same membrane were assigned to the same dataset). The training set contained 75 samples and 1200 spectral ring mapping images, while the test set contained 25 samples and 400 spectral ring mapping images.

[0079] An improved ShuffleNetV2 model is constructed, using a pre-trained ShuffleNetV2 model (ShuffleNetV2_x0.5) as the feature extractor. The input channels of the initial convolutional layers are modified from 3 channels in RGB images to 1 channel to adapt to grayscale images. The last fully connected layer of the ShuffleNetV2 model is removed, retaining only the preceding convolutional layers. A custom attention module is introduced, containing two 1×1 convolutional layers. The first convolutional layer reduces the number of channels, and the second convolutional layer restores the number of channels. Channel-level attention weights are generated using the sigmoid function. These weights are then multiplied element-wise with the feature map to highlight the feature regions that the network considers more important. After applying the attention mechanism, the feature map is processed by a global average pooling layer, which compresses the spatial features of each channel into a single value, resulting in an output feature vector with a shape of 1024×1.

[0080] During training, the mean squared error loss function (MSELoss) was used as the loss function for the regression task. The Adam optimizer, combined with a learning rate scheduler (StepLR), was used to dynamically adjust the learning rate to ensure the model converges to a better local minimum in the later stages of training. To fully utilize the dataset and reduce the risk of overfitting, K-Fold Cross-Validation was employed for model training. In each fold, the model weights were reset, and training and validation were performed using different data splits. The model with the smallest validation error was selected as the optimal model. The training process consisted of 25 epochs, with 5 folds used for cross-validation. After training, the model was evaluated using a test set, and the coefficient of determination (R²) was used to measure the performance. 2 The regression performance of the model is measured by the root mean square error (RMSE) and mean absolute percentage error (MAPE). After hyperparameter tuning and optimization, the initial learning rate was 0.001, the batch size was 16, the step size of StepLR was 7, and the gamma was 0.5. The optimal model test set results were: R 2 =0.9890, RMSE=8.66, MAPE=8.8%. Since 16 ring mapping images were generated for each sample, the mean of the output results of the 16 images was used as the final detection result for that sample, and the test set result was calculated as R. 2 =0.9935, RMSE=6.66, MAPE=7.1%.

[0081] The circular mapping image of the sample is input into the ShuffleNetV2 model with an attention mechanism module. High-dimensional feature maps obtained after multiple convolutions and activation functions are extracted. After weight adjustment by the attention mechanism and adaptive average pooling, these are flattened into a one-dimensional feature vector with 1024 variables, denoted as IF = [imf1, imf2, imf3, ..., imf...]. 1023 ,imf 1024 The one-dimensional feature vectors of 16 circular mapping images generated from the same sample are averaged into one feature vector. The dimensionality of the circular mapping image feature vectors in the training set is reduced using the successive projections algorithm (SPA), ultimately extracting six features, which are the 2nd, 5th, 6th, 7th, 8th, and 12th variables, respectively. This yields the dimensionality-reduced image feature vector IFSPA = [imf2, imf5, imf6, imf7, imf8, imf...]. 12 ].

[0082] The 32 original spectra of each sample are averaged into one, and area normalization is performed as a preprocessing step, denoted as SAN = [ref] 400nm ,ref 401nm ,ref 402nm ,……,ref 1022nm ,ref 1023nm Using the same diversity partitioning method as the mapped image, the spectral dataset was divided into a training set and a test set. Feature extraction was performed on the spectra of the training set samples using SANSPA, ultimately yielding five features corresponding to the wavelengths 400, 405, 413, 942, and 1020 nm. This resulted in the spectral feature SANSPA = [ref]. 400nm ,ref 405nm ,ref 413nm ,ref 942nm ,ref 1020nm ].

[0083] The extracted ring map image features and spectral features are fused to obtain 11 fused features, denoted as FSPA = [imf2, imf5, imf6, imf7, imf8, imf...]. 12 ,ref 400nm ,ref 405nm ,ref 413nm ,ref 942nm ,ref 1020nm Based on the training set, a quantitative regression model for particulate matter mass concentration was established using partial least squares method. The model was trained using five-fold cross-validation, and the optimal model was selected based on the RMSE of the cross-validation set. The overall data analysis process is as follows: Figure 3 As shown. The final atmospheric particulate matter concentration regression model is as follows:

[0084]

[0085] In the formula, C represents the predicted atmospheric particulate matter concentration, in μg·m³. -3 .

[0086] The sample to be tested is the test set sample. The concentration of the sample in the optimal detection test set is used, and the result is: R 2 =0.9947, RMSE=6.02, MAPE=4.2%, the scatter plot of the prediction results for each sample is as follows. Figure 4 As shown.

[0087] In this embodiment, the improved ShuffleNetV2 network based on visible and near-infrared spectral ring mapping images performed excellently on the test set. After feature extraction and fusion with spectral features, the detection accuracy was further improved, with an average relative deviation of less than 5%, which can effectively achieve efficient and accurate detection of the concentration of membrane-adsorbed particulate matter in the air.

[0088] This application achieves spectral conversion through a ring mapping method and establishes a quantitative regression model for particulate matter concentration on the membrane using an improved lightweight deep network, ShuffleNetV2. This eliminates the shortcomings of traditional spectral-based machine learning methods, such as interference from spectral fluctuations and the inability to capture the overall trend of spectral lines, thus improving detection stability and accuracy. The improved ShuffleNetV2 network extracts the feature layer of the spectral ring mapping image, performs dimensionality reduction using a feature extraction algorithm, and then fuses it with spectral features. A regression model is then established using machine learning methods. Compared with the ShuffleNetV2 model that directly outputs prediction results through fully connected layers, this method further extracts effective information, eliminates redundant information interference, and further improves detection accuracy and timeliness of particulate matter concentration detection on the membrane through spectral feature fusion and simple machine learning methods, while significantly reducing labor and equipment costs.

[0089] Based on the same inventive concept, this application also provides an atmospheric particulate matter concentration detection device for implementing the atmospheric particulate matter concentration detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more atmospheric particulate matter concentration detection device embodiments provided below can be found in the limitations of the atmospheric particulate matter concentration detection method described above, and will not be repeated here.

[0090] In one exemplary embodiment, an atmospheric particulate matter concentration detection device is provided, comprising:

[0091] The data determination module is used to determine the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration based on the membrane-adsorbed atmospheric particulate matter sample.

[0092] The ring mapping image determination module is used to perform spectrum conversion based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample to obtain a ring mapping image.

[0093] The feature extractor building module is used to construct a feature extractor based on the ring map image using an improved ShuffleNetV2 network with an attention mechanism. The feature extractor is used to extract image features and performs dimensionality reduction using a one-dimensional feature extraction algorithm.

[0094] The feature fusion module is used to fuse the spectral features corresponding to the visible and near-infrared reflectance spectra with the corresponding image features.

[0095] The concentration detection module is used to construct a rapid quantitative detection model for membrane-adsorbed particulate matter concentration based on fusion features and the corresponding true values ​​of particulate matter mass concentration using machine learning methods; and to use the rapid quantitative detection model for membrane-adsorbed particulate matter concentration to detect atmospheric particulate matter concentration.

[0096] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting atmospheric particulate matter concentration.

[0097] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0098] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0099] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0100] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0101] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0102] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting atmospheric particulate matter concentration, characterized in that, The method for detecting atmospheric particulate matter concentration includes: Based on the membrane adsorbed atmospheric particulate matter sample, determine the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration; Based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter samples, the ring mapping method was used to perform spectrum conversion to obtain the ring mapping image. Based on the ring map image, a feature extractor is constructed using the ShuffleNetV2 network, which is an improvement based on the attention mechanism. The feature extractor is used to extract image features, and a one-dimensional feature extraction algorithm is used for dimensionality reduction. Feature fusion is performed on the spectral features corresponding to the visible and near-infrared reflectance spectra and the corresponding image features. Based on the fusion characteristics and the corresponding true values ​​of particulate matter mass concentration, a rapid quantitative detection model for membrane-adsorbed particulate matter concentration is constructed using machine learning methods; and the atmospheric particulate matter concentration is detected using the rapid quantitative detection model for membrane-adsorbed particulate matter concentration.

2. The method for detecting atmospheric particulate matter concentration according to claim 1, characterized in that, The process of determining the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration based on the membrane-adsorbed atmospheric particulate matter sample also includes: Atmospheric particulate matter samples adsorbed by membranes were obtained using an atmospheric particulate matter sampler.

3. The method for detecting atmospheric particulate matter concentration according to claim 1, characterized in that, The step involves converting the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample using a ring mapping method to obtain a ring-mapped image. Specifically, this includes: The reflectance of the visible and near-infrared reflectance spectra is preprocessed; the preprocessing includes: multivariate scattering correction and smoothing filtering; The spectrum was determined based on the pre-processed reflectance. The spectral image is converted using the ring mapping method based on the spectral image to obtain the ring-mapped image.

4. The method for detecting atmospheric particulate matter concentration according to claim 1, characterized in that, The improved ShuffleNetV2 network based on the attention mechanism replaces the last fully connected layer in the ShuffleNetV2 network with sequentially connected attention modules and a global average pooling layer.

5. The method for detecting atmospheric particulate matter concentration according to claim 4, characterized in that, The attention module consists of two 1×1 convolutional layers.

6. The method for detecting atmospheric particulate matter concentration according to claim 1, characterized in that, Feature fusion is performed on the spectral features corresponding to the visible and near-infrared reflectance spectra and the corresponding image features. This previously included: The average spectrum is determined based on the visible and near-infrared reflectance spectra. The average spectrum is normalized by area. The spectral features of the spectrum after area normalization are extracted using a feature extraction algorithm.

7. An atmospheric particulate matter concentration detection device, characterized in that, The atmospheric particulate matter concentration detection equipment includes: The data determination module is used to determine the corresponding visible and near-infrared reflectance spectra and the true values ​​of particulate matter mass concentration based on the membrane-adsorbed atmospheric particulate matter sample. The ring mapping image determination module is used to perform spectrum conversion based on the visible and near-infrared reflectance spectra of the membrane-adsorbed atmospheric particulate matter sample to obtain a ring mapping image. The feature extractor building module is used to construct a feature extractor based on the ring map image using an improved ShuffleNetV2 network with an attention mechanism. The feature extractor is used to extract image features and performs dimensionality reduction using a one-dimensional feature extraction algorithm. The feature fusion module is used to fuse the spectral features corresponding to the visible and near-infrared reflectance spectra with the corresponding image features. The concentration detection module is used to construct a rapid quantitative detection model for membrane-adsorbed particulate matter concentration based on fusion features and the corresponding true values ​​of particulate matter mass concentration using machine learning methods; and to use the rapid quantitative detection model for membrane-adsorbed particulate matter concentration to detect atmospheric particulate matter concentration.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the atmospheric particulate matter concentration detection method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the atmospheric particulate matter concentration detection method according to any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the atmospheric particulate matter concentration detection method according to any one of claims 1-6.

Images