Method and system for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry

Abstract

The present invention provides a method and system for analyzing the source of atmospheric particles based on single-particle aerosol mass spectrometry, which belongs to the field of particle identification. The method includes: collecting mass spectrum data of atmospheric particles through a single-particle aerosol mass spectrometer; analyzing the mass spectrum data based on a pre-trained deep learning model to determine the source of atmospheric particles; the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron connected in sequence. The present invention performs end-to-end particle source analysis through a deep learning model, achieving a high degree of automation from input to output, reducing the influence of subjective factors, and the deep learning model has better feature extraction and generalization capabilities, and can better capture the weak linear and nonlinear relationship of particles, thereby improving the accuracy of atmospheric particle source analysis.

Description

Technical Field

[0001] The present invention relates to the field of particle identification, and in particular to a method and system for analyzing the source of atmospheric particles based on single-particle aerosol mass spectrometry. Background Art

[0002] When atmospheric particulate matter enters the human body through the respiratory tract, it will harm the heart and lung function and cause health effects on humans. The chemical composition of particulate matter can reflect its source and formation process. Therefore, studying the chemical composition of particulate matter can achieve source analysis and prompt humans to take corresponding measures to prevent and control air pollution to protect their own health.

[0003] In the study of atmospheric particulate matter, online source analysis of particulate matter is often achieved using a single particle aerosol mass spectrometer (SPAMS). SPAMS is a mass spectrometer that can directly monitor the chemical composition of atmospheric particulate matter online without pretreatment. It can provide high-time resolution particle mass spectrum information and is widely used in the source analysis of atmospheric particulate matter.

[0004] Currently, the main methods for analyzing the sources of atmospheric particulate matter are positive definite matrix factorization, ART-2a algorithm and similarity method. Positive definite matrix factorization is to decompose the observed atmospheric data into a linear combination of multiple potential factors and calculate the contribution of each factor to the observed data. Positive definite matrix factorization requires artificial selection of the input data type and the source of the judgment result factor, which will introduce certain subjective factors. The selected features are difficult to cover the laws of pollution sources, and the physical meaning of the factors is not clear enough. In addition, it cannot handle weak linear and nonlinear relationships well. The ART-2a algorithm is based on an adaptive neural network. It clusters particles and identifies the mass spectrum characteristics of the clusters to determine the pollution source to which the clusters belong. In use, it is often necessary to manually merge redundant clusters, which is subjective and time-consuming. The similarity method determines the attribution by comparing the cosine similarity of the mass spectra of two particles. Usually, the mass spectrum data of particles from known pollution sources are collected, and then the cosine similarity is calculated with the particles in the environment. The similarity method has limitations in handling weak linear and nonlinear relationships and cannot capture the difference between two highly similar particles from different sources. In addition, the threshold of cosine similarity needs to be manually determined in the process of judging the attribution of particles, which is subjective.

[0005] With the in-depth development of the work on the source analysis of atmospheric particulate matter, the demand for fine analysis has gradually increased. Fine analysis refers to the identification of two highly similar particles whose sources are similar, such as particulate matter emitted by gasoline and diesel vehicles, or particulate matter emitted by agricultural machinery and construction machinery with diesel engines. Identifying highly similar particles often requires the use of weak linear or even nonlinear relationships. In addition, although particulate matter emitted by some pollution sources has significant tracer characteristics, it accounts for a low proportion of the emitted particulate matter, resulting in incomplete coverage when using tracer characteristics to determine the source. Therefore, the above commonly used methods have certain limitations for the fine analysis of highly similar particles.

[0006] In summary, the current fine analysis technology for atmospheric particulate matter is not yet mature, and cannot capture the weak linear and nonlinear relationships of particulate matter well, and thus cannot identify more complex patterns and relationships in particulate matter. It has problems such as high subjectivity, a time-consuming analysis process, and an inability to distinguish similar particulate matter from different sources. Summary of the invention

[0007] The purpose of the present invention is to provide a method and system for analyzing the sources of atmospheric particulate matter based on single-particle aerosol mass spectrometry, which can improve the accuracy of atmospheric particulate matter source analysis.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry, comprising:

[0010] Collect mass spectrometric data of atmospheric particulate matter using a single particle aerosol mass spectrometer;

[0011] Based on a pre-trained deep learning model, the mass spectrometry data is analyzed to determine the source of atmospheric particulate matter; the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron connected in sequence.

[0012] Optionally, the mass spectrometry data is parsed based on a pre-trained deep learning model to determine the source of the atmospheric particulate matter, specifically including:

[0013] Normalizing the positive spectrum and the negative spectrum of the mass spectrum data respectively to obtain one-dimensional mass spectrum data;

[0014] Using a pre-trained one-dimensional convolutional neural network to extract multiple local features of the mass spectrum one-dimensional data;

[0015] The pre-trained long short-term memory network is used to extract the dependencies between multiple local features to obtain sequence data;

[0016] Based on the sequence data, a pre-trained multi-layer perceptron is used to determine the source of the atmospheric particulate matter.

[0017] Optionally, the positive spectrum and the negative spectrum of the mass spectrum data are normalized respectively using the L2 norm.

[0018] Optionally, the one-dimensional convolutional neural network extracts local features of the one-dimensional mass spectrum data using the following formula:

[0019] Y1=σ1(W1·X1+b1);

[0020] Among them, Y1 is the local feature, w1 is the weight matrix of the one-dimensional convolutional neural network, X1 is the one-dimensional mass spectrometry data, b1 is the bias vector of the one-dimensional convolutional neural network, and σ1 is the activation function.

[0021] Optionally, the training process of the deep learning model includes:

[0022] Acquire a pollution spectrum library; the pollution spectrum library includes a plurality of single particle mass spectrum data and a known source corresponding to each single particle mass spectrum data; the single particle mass spectrum data is one-dimensional data after normalization;

[0023] Determining a loss function based on known sources of each single particle mass spectrum data;

[0024] Based on the loss function, a back propagation algorithm is used to iteratively train the deep learning model until the loss function converges or reaches a maximum number of iterations, thereby obtaining a trained deep learning model.

[0025] Optionally, the loss function is classification cross entropy loss.

[0026] To achieve the above object, the present invention also provides the following solution:

[0027] An atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry, comprising:

[0028] Single particle aerosol mass spectrometer, used to collect mass spectrometric data of atmospheric particulate matter;

[0029] A processor is connected to the single-particle aerosol mass spectrometer and is used to analyze the mass spectrum data based on a pre-trained deep learning model to determine the source of atmospheric particulate matter; the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron connected in sequence.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the present invention uses a single-particle aerosol mass spectrometer to collect mass spectrum data of atmospheric particulate matter, and can obtain the chemical composition information of each particle. A deep learning model composed of a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron is used to perform end-to-end particulate matter source analysis, achieving a high degree of automation from input to output, reducing the influence of subjective factors, and the deep learning model has better feature extraction and generalization capabilities, and can better capture the weak linear and nonlinear relationships of particulate matter, thereby improving the accuracy of atmospheric particulate matter source analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0032] Figure 1 A flow chart of the atmospheric particulate matter source analysis method based on single-particle aerosol mass spectrometry provided by the present invention;

[0033] Figure 2 A schematic diagram of a deep learning model established for the present invention;

[0034] Figure 3 This is a schematic diagram of the atmospheric particulate matter source analysis process;

[0035] Figure 4 A schematic diagram of the atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry provided by the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] The purpose of the present invention is to provide a method and system for analyzing the sources of atmospheric particulate matter based on single-particle aerosol mass spectrometry, which uses a single-particle aerosol mass spectrometer to obtain the chemical composition information of each particle, and performs end-to-end particle source analysis. A one-dimensional convolutional neural network (1D-CNN), a long short-term memory network (LSTM) and a multi-layer perceptron (MLP) are used to form a deep learning model, which achieves a high degree of automation from input to output, reduces the influence of subjective factors, and shortens the model running time. The deep learning model has better feature extraction and generalization capabilities, better captures the weak linear and nonlinear relationships of particulate matter, and realizes high-resolution, high-accuracy real-time fine analysis of the sources of atmospheric particulate matter.

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] Embodiment 1

[0040] like Figure 1 As shown, this embodiment provides a method for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry, including:

[0041] Step 100: collecting mass spectrum data of atmospheric particulate matter by a single particle aerosol mass spectrometer;

[0042] Step 200: Analyze the mass spectrum data based on a pre-trained deep learning model to determine the source of atmospheric particulate matter. The deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a multi-layer perceptron connected in sequence.

[0043] Specifically, step 200 includes:

[0044] Step 201: normalize the positive spectrum and the negative spectrum of the mass spectrum data to obtain one-dimensional mass spectrum data. In this embodiment, the positive spectrum and the negative spectrum of the mass spectrum data are normalized by using the L2 norm to convert the positive spectrum and the negative spectrum into a one-dimensional array with a value between 0 and 1.

[0045] Step 202: Use a pre-trained one-dimensional convolutional neural network to extract multiple local features of the mass spectrum one-dimensional data:

[0046] Y1=σ1(W1·X1+b1);

[0047] Among them, Y1 is a local feature, W1 is a weight matrix of a one-dimensional convolutional neural network, X1 is a one-dimensional mass spectrum data, b1 is a bias vector of a one-dimensional convolutional neural network, and σ1 is an activation function. In this embodiment, the activation function here adopts a Relu activation function, and the expression of the Relu activation function is max(0,x), which is a nonlinear function, and x is a parameter in the activation function.

[0048] Step 203: Use a pre-trained long short-term memory network to extract the dependency relationship between multiple local features to obtain sequence data.

[0049] For the study of the chemical composition of atmospheric particulate matter, it is necessary to consider not only the local features but also the complex dependencies between the features. Therefore, the local features extracted by the one-dimensional convolutional neural network are input into the long short-term memory network to further extract the dependencies between the features and identify more complex patterns or relationships. The long short-term memory network includes an input gate, a forget gate, a first memory unit, a second memory unit, an output gate, and a hidden state. Its structure is as follows:

[0050] Input gate: i t =σ(W i [h t-1 ,x t ]+b i );

[0051] Forget gate: f t =σ[W f [h t-1 ,x t ]+b f );

[0052] First memory unit:

[0053] Second memory unit:

[0054] Output gate: o t =σ(W o [h t-1 ,x t ]+b o );

[0055] Hidden state:

[0056] Among them, x t is the input at the current moment, h t-1 is the output of the previous moment, i t is the added degree value of the current state, f t To retain the degree value, is the intermediate state, C t is the state at the previous moment, Ct is the current state, o t is the output degree value, h t is the output at the current moment, W i is the weight matrix of the input gate, W f is the weight matrix of the forget gate, W c is the weight matrix of the first memory unit, W o is the weight matrix of the output gate, b i is the bias vector of the input gate, b f is the bias vector of the forget gate, b c is the bias vector of the first memory unit, b o is the bias vector of the output gate, tanh is the hyperbolic tangent function, is the dot multiplication operation.

[0057] The principle of long short-term memory network is as follows:

[0058] ① Input gate: LSTM receives an input sequence and uses a sigmoid activation function to determine which information needs to be updated. The input gate controls the importance of the input. The sigmoid activation function expression is It is a non-linear function.

[0059] ② Forget gate: LSTM uses another sigmoid activation function to decide which information needs to be forgotten. The forget gate controls whether the memory of the previous moment needs to be forgotten.

[0060] ③Memory unit: LSTM uses a memory unit to store and transmit information. It updates the memory through the input gate, forgets the memory through the forget gate, and outputs the memory through the output gate.

[0061] ④Output gate: LSTM uses a sigmoid activation function to determine which information needs to be output. The output gate controls the importance of the output.

[0062] ⑤Hidden state: LSTM calculates the hidden state based on the results of the input gate, forget gate, and output gate. The hidden state is the output of LSTM and the input of the next time step.

[0063] Step 204: According to the sequence data, a pre-trained multi-layer perceptron is used to determine the source of the atmospheric particulate matter.

[0064] The multilayer perceptron is used to receive sequence data processed by the one-dimensional convolutional neural network and the long short-term memory network. The multilayer perceptron includes an input layer, a hidden layer, and an output layer connected in sequence, and outputs the final particle classification result, that is, the source of the particles. The expression of the multilayer perceptron is Y3 = σ3 (W3 · X3 + b3), and its activation function σ3 is the sotfmax activation function, and the expression is It is a nonlinear function. Y3 is the probability of output classification, X3 is the input sequence data, and x i is the data in X3, W3 is the weight matrix of the multilayer perceptron, and b3 is the bias vector of the multilayer perceptron.

[0065] like Figure 2 Schematic diagram of the deep learning model shown.

[0066] Furthermore, the training process of the deep learning model includes:

[0067] Step 300: Obtain a pollution spectrum library. The pollution spectrum library includes multiple single-particle mass spectrum data and known sources corresponding to each single-particle mass spectrum data. The single-particle mass spectrum data is one-dimensional data after normalization. First, pollutants of known sources in the atmospheric environment are collected by vacuum bottles, and then analyzed by a single-particle aerosol mass spectrometer to obtain single-particle mass spectrum data.

[0068] Step 400: Determine a loss function based on the known sources of each single particle mass spectrometry data. In this embodiment, the loss function is a classification cross entropy loss: Among them, L is the loss function value, the value of outputsize depends on the number of categories, y j is the value of the true label of a sample in the jth category, using one-hot encoding. is the predicted probability of belonging to the jth category.

[0069] Step 500: Based on the loss function, the back propagation algorithm is used to iteratively train the deep learning model until the loss function converges or the maximum number of iterations is reached, thereby obtaining a trained deep learning model. The core idea of ​​the back propagation algorithm is to use the chain rule to calculate the gradient of the model parameters. Assume that there is an N-layer neural network, where the weight matrix of the Nth layer is W N , the bias vector is b N , the activation function is σ N , the loss function is L, and the linear output is Z N . Then the process of calculating the gradient is as follows:

[0070] ① Calculate the output Y of the deep learning model through forward propagation:

[0071] Y=σ N (W N·X N-1 +b N ).

[0072] ②For the Dth layer, the gradient is calculated using the following formula:

[0073] δ D =(W D ·X D-1 +b D )Tδ D+1 ⊙Y' D ;

[0074] Among them, δ D is the gradient of the Dth layer, ⊙ represents element-by-element multiplication, Y' D Represents the derivative of the activation function of the Dth layer.

[0075] ③ Use the following formula to calculate the weight gradient from the Dth layer to the D+1th layer and bias gradient

[0076] During the training process, the Adam algorithm is used to update the weight matrix and bias vector. First, the first-order moment estimate and the second-order moment estimate of the gradient are calculated, and then the deviation is corrected. Finally, the first-order moment estimate and the second-order moment estimate after the deviation correction are used to update the weight matrix and bias vector:

[0077]

[0078] Where k is the number of iterations, m is k is the first-order moment estimate of the gradient at the kth iteration, v k is the second-order moment estimate of the gradient at the kth iteration, is the first-order moment estimate of the gradient after deviation correction at the kth iteration, is the second-order moment estimate of the gradient after deviation correction at the kth iteration, β1 and β2 are hyperparameters, is the gradient at the kth iteration, η is the learning rate, ε is the numerical stability constant, and θ k is the weight matrix and bias vector at the kth iteration.

[0079] In summary, the specific training process of the deep learning model is as follows: first, the mass spectrometry data of the input particles is passed to the deep learning model through forward propagation, and then the loss function is calculated. The deep learning model training process aims to minimize the loss function, calculates the gradient through back propagation, and then uses the Adam algorithm to update the weight matrix and bias vector of the deep learning model. This process is iterated repeatedly until the maximum number of iterations is reached or the loss function converges.

[0080] like Figure 3As shown, the present invention mainly consists of four modules: pollution source collection, pollution spectrum library construction, model building and actual environment source analysis. The deep learning model is pre-trained using the data of the pollution spectrum library, and the trained deep learning model is saved. The mass spectrum data of a single particle is collected in the actual atmospheric environment through a single particle aerosol mass spectrometer, and the mass spectrum data is imported into the trained deep learning model to achieve a detailed analysis of the source of particulate matter.

[0081] The present invention does not require feature selection and realizes an end-to-end training process, which effectively reduces subjectivity, shortens the analysis time of particle sources, can better capture the weak linear and nonlinear relationships of particulate matter, recognize complex patterns and relationships, improves the ability to distinguish similar particles, and realizes fine analysis of the sources of atmospheric particulate matter.

[0082] Embodiment 2

[0083] In order to execute the method corresponding to the above-mentioned embodiment 1 and realize the corresponding functions and technical effects, a system for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry is provided below.

[0084] like Figure 4 As shown, the atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry provided in this embodiment includes: a single-particle aerosol mass spectrometer 1 and a processor 2.

[0085] The single-particle aerosol mass spectrometer 1 is used to collect mass spectrum data of atmospheric particulate matter.

[0086] The processor 2 is connected to the single-particle aerosol mass spectrometer 1, and is used to analyze the mass spectrum data based on a pre-trained deep learning model to determine the source of atmospheric particulate matter. The deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network, and a multi-layer perceptron connected in sequence.

[0087] Compared with the prior art, the atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry provided in this embodiment has the same beneficial effects as the atmospheric particulate matter source analysis method based on single-particle aerosol mass spectrometry provided in Example 1, which will not be repeated here.

[0088] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0089] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry, characterized in that: The method for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry includes: Collect mass spectrometric data of atmospheric particulate matter using a single particle aerosol mass spectrometer; Based on a pre-trained deep learning model, the mass spectrometry data is analyzed to determine the source of atmospheric particulate matter; the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron connected in sequence; the training process of the deep learning model includes: Acquire a pollution spectrum library; the pollution spectrum library includes a plurality of single particle mass spectrum data and a known source corresponding to each single particle mass spectrum data; the single particle mass spectrum data is one-dimensional data after normalization; According to the known sources of each single particle mass spectrometry data, a loss function is determined; the loss function is a classification cross entropy loss: Among them, L is the loss function value, the value of outputsize depends on the number of categories, y j is the value of the true label of a sample in the jth category, using one-hot encoding. To predict the probability of belonging to the jth category; Based on the loss function, a back propagation algorithm is used to iteratively train the deep learning model until the loss function converges or reaches a maximum number of iterations, thereby obtaining a trained deep learning model; the back propagation algorithm uses the chain rule to calculate the gradient of the model parameters; Based on the pre-trained deep learning model, the mass spectrometry data is analyzed to determine the source of atmospheric particulate matter, including: Normalizing the positive spectrum and the negative spectrum of the mass spectrum data respectively to obtain one-dimensional mass spectrum data; Using a pre-trained one-dimensional convolutional neural network to extract multiple local features of the mass spectrum one-dimensional data; The pre-trained long short-term memory network is used to extract the dependencies between multiple local features to obtain sequence data; Based on the sequence data, a pre-trained multi-layer perceptron is used to determine the source of the atmospheric particulate matter.

2. The method for analyzing the source of atmospheric particulate matter based on single-particle aerosol mass spectrometry according to claim 1, characterized in that: The L2 norm is used to normalize the positive spectrum and the negative spectrum of the mass spectrum data respectively.

3. The method for source analysis of atmospheric particulate matter based on single-particle aerosol mass spectrometry according to claim 1, characterized in that: The one-dimensional convolutional neural network extracts the local features of the one-dimensional mass spectrum data using the following formula: Y1=σ1(W1·X1+b1); Among them, Y1 is the local feature, W1 is the weight matrix of the one-dimensional convolutional neural network, X1 is the one-dimensional mass spectrum data, b1 is the bias vector of the one-dimensional convolutional neural network, and σ1 is the activation function.

4. An atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry, characterized in that: The atmospheric particulate matter source analysis system based on single-particle aerosol mass spectrometry includes: Single particle aerosol mass spectrometer, used to collect mass spectrometric data of atmospheric particulate matter; A processor is connected to the single-particle aerosol mass spectrometer and is used to analyze the mass spectrum data based on a pre-trained deep learning model to determine the source of atmospheric particulate matter; the deep learning model includes a one-dimensional convolutional neural network, a long short-term memory network and a multi-layer perceptron connected in sequence; the training process of the deep learning model includes: Acquire a pollution spectrum library; the pollution spectrum library includes a plurality of single particle mass spectrum data and a known source corresponding to each single particle mass spectrum data; the single particle mass spectrum data is one-dimensional data after normalization; According to the known sources of each single particle mass spectrometry data, a loss function is determined; the loss function is a classification cross entropy loss: Among them, L is the loss function value, the value of outputsize depends on the number of categories, y j is the value of the true label of a sample in the jth category, using one-hot encoding. To predict the probability of belonging to the jth category; Based on the loss function, a back propagation algorithm is used to iteratively train the deep learning model until the loss function converges or reaches a maximum number of iterations, thereby obtaining a trained deep learning model; the back propagation algorithm uses the chain rule to calculate the gradient of the model parameters; Based on the pre-trained deep learning model, the mass spectrometry data is analyzed to determine the source of atmospheric particulate matter, including: Normalizing the positive spectrum and the negative spectrum of the mass spectrum data respectively to obtain one-dimensional mass spectrum data; Using a pre-trained one-dimensional convolutional neural network to extract multiple local features of the mass spectrum one-dimensional data; The pre-trained long short-term memory network is used to extract the dependencies between multiple local features to obtain sequence data; Based on the sequence data, a pre-trained multi-layer perceptron is used to determine the source of the atmospheric particulate matter.

Images