A method and system for detecting and identifying trace substances using Raman spectroscopy

Abstract

The present invention relates to a Raman spectrum detection and identification method and system for trace substances. The identification method uses a portable Raman spectrometer to collect the spectrum signal of the detected object on the electric field coupling enhanced SERS substrate, and then transmits the spectrum signal to the Raman spectrum identification algorithm integrated in the industrial computer to judge the type of the substance. By combining the SERS substrate of the gold nano-array structure with the noble metal nano-sol, the noble metal nano-particles excite the local surface plasmon resonance (LSP), resulting in the enhancement of the local electromagnetic field on the surface of the noble metal nano-particles. At the same time, the SERS substrate of the gold nano-array structure excites the surface propagation plasmon (SPP), resulting in the enhancement of the electromagnetic field on the array surface. The SPP near field and the LSP near field are coupled to each other to form a composite SERS substrate with enhanced electric field coupling, thereby significantly improving the Raman spectrum detection accuracy. The present invention can realize the accurate identification of weak Raman signals through the Raman spectrum identification algorithm in the industrial computer.

Description

Technical Field

[0001] The invention relates to detection and identification of trace substances, in particular to a Raman spectrum detection and identification method and system for trace substances. Background Art

[0002] Surface enhanced Raman scattering (SERS) technology is a non-destructive, label-free detection technology that has a unique photon fingerprint for specific substances. It is an important technical means for the current research on trace substance detection. Raman spectroscopy detection technology has the advantages of short measurement cycle and high sensitivity, and can be used for online detection. At present, the use of SERS to detect trace substances mainly faces two difficulties: first, the detection accuracy is not high enough, which puts forward higher requirements on the enhanced performance of the SERS substrate; second, the spectral recognition accuracy is not high enough, which poses a challenge to the rapid recognition and judgment of Raman spectroscopy.

[0003] Gold nanoparticles (AuNPs) and silver nanoparticles (AgNPs) are one of the most commonly used SERS substrates due to their low production cost and ease of synthesis. AgNPs are easily oxidized and unstable, while AuNPs are relatively stable and have greater potential in practical applications. Two-dimensional arrays are an ordered, controllable SERS substrate that can be prepared over a large area. They are polarization-independent and extremely sensitive. The preparation process is relatively simple, which is conducive to the formation of uniform and stable "hot spots". At present, most of the research on improving the detection performance of SERS substrates is focused on substrate structure design and new material research, while there are relatively few methods that use the principle of electric field coupling enhancement to improve the detection performance of SERS substrates.

[0004] Raman spectrum recognition is mostly achieved by fitting the characteristic peaks of the Raman spectrum of the detected substance, and then comparing them with the peak shape, peak position and other characteristics of the material spectrum in the database. The quality of these methods depends on the standard Raman database and is greatly affected by the signal intensity, especially for weak signal spectrum recognition. The signal intensity of the currently collected trace substance Raman signals is sometimes weak, the characteristic peaks are not obvious or even difficult to distinguish, and there is still little research on the recognition of such weak Raman signals. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for detecting and identifying trace substances using Raman spectroscopy, so as to solve the problems of low precision of existing SERS detection and inaccurate identification of weak signals.

[0006] The Raman spectrum detection and identification method of trace substances of the present invention utilizes a Raman spectrum detector to collect spectrum signals of a detection object placed on an electric field coupling enhanced SERS substrate, and the collected spectrum signals are processed by a Raman spectrum identification algorithm to obtain a judgment result on the material type of the detection object; the Raman spectrum identification algorithm includes: Raman spectrum preprocessing, construction of an identification model that integrates a one-dimensional dilated convolutional network and a two-dimensional dilated convolutional network, and model training and spectrum identification.

[0007] Further, the Raman spectrum preprocessing includes the following steps:

[0008] Set the Raman data format of the original sample to:

[0009] W raw =[W1,W2,...,W n ]、X raw =[X1,X2,...,X n ] and y raw =[y1,y2,...,y n ];

[0010] Among them, W raw represents the Raman frequency shift space of all samples, W i represents the Raman frequency shift vector of the i-th sample, n represents the sample data volume, X raw represents the signal intensity space of the sample, X i represents the signal strength vector of the i-th sample, y raw represents the sample category space, y i Indicates the type of the i-th sample;

[0011] S1.1, interpolation operation: through piecewise linear interpolation operation, a function is created based on the original Raman data, and the frequency shift range of the unified sample is: [201,2200cm -1 ], interval: 1cm -1 , a total of 2048 Raman data were generated for each sample;

[0012] S1.2, baseline correction: Adaptive iterative weighted penalty least squares method is used to perform baseline correction on the data obtained by S1.1 interpolation operation to eliminate the fluorescence effect of Raman spectrum;

[0013] S1.3, smooth denoising: use Savitzky-Golay filter to smooth the data generated by S1.2 baseline correction;

[0014] S1.4, Normalization: The data set obtained by smooth denoising in S1.3 is normalized according to the following formula to unify the Raman signal intensity of each sample to the interval [0,1]:

[0015]

[0016] Among them, x i,j is the signal strength of the jth frequency shift position of the ith sample, min(x i, ) is the minimum signal strength of the ith sample, max(x i, ) is the maximum signal strength of the i-th sample;

[0017] S1.5, continuous wavelet transform: The data set obtained by normalization in S1.4 is transformed using the Mexican hat wavelet as the mother wavelet function, and the generated wavelet coefficient map is used as the input of the two-dimensional dilated convolutional network.

[0018] Furthermore, the recognition model of the fusion of the one-dimensional dilated convolutional network and the two-dimensional dilated convolutional network includes a feature input layer, a feature extraction layer, a feature classification prediction layer and an output layer;

[0019] S2.1, the feature input layer includes: using the data obtained by normalization in S1.4 as the input of the one-dimensional dilated convolutional network, and using the wavelet coefficient map obtained by continuous wavelet transform in S1.5 as the input of the two-dimensional dilated convolutional network;

[0020] S2.2, the feature extraction layer includes two parts: one-dimensional dilated convolutional network feature extraction layer and two-dimensional dilated convolutional network feature extraction layer:

[0021] The one-dimensional dilated convolutional network feature extraction layer uses four one-dimensional blocks and a flattening layer. Each block includes a dilated convolution module and a one-dimensional maximum pooling module.

[0022] The 2D dilated convolutional network feature extraction layer uses four 2D blocks and a flattening layer. Each block includes a 2D dilated convolution module and a 2D maximum pooling module.

[0023] The data after each dilated convolution operation is activated by the ReLU function, and the data after the dilated convolution changes to:

[0024] x conv_out =floor((x con_in +2×padding-dilation×(kernel_size-1)-1) / stride+1)

[0025] Among them, x con_in is the input data of the dilated convolutional layer, x conv_out is the output data of the expanded convolution layer, kernel_size is the convolution kernel size, dilation is the expansion coefficient, padding is the padding size, and stride is the convolution kernel movement step size;

[0026] S2.3, the classification prediction layer includes a splicing layer and a fully connected layer. The splicing layer is used to splice the flattening layer of the one-dimensional dilated convolutional network and the two-dimensional dilated convolutional network. The fully connected layer uses the activation function tanh to implement nonlinearity and improves the generalization ability by adopting the dropout strategy;

[0027] S2.4, the output layer uses the activation function softmax, and the output of the activation function softmax is the prediction result of the entire recognition algorithm.

[0028] Furthermore, model training and spectrum recognition include the following steps:

[0029] S3.1, model training: The input spectral sample data is randomly selected by layer and divided into a training set and a validation set according to a ratio of 8:2 to ensure that the training set contains samples of each substance; the Adam optimizer is used in model training, and the cross entropy loss function (Cross Entropy Loss) is used as the training indicator for iterative training to obtain a neural network model, and the model is saved;

[0030] S3.1, model training: The input spectral sample data is randomly selected by layer and divided into training set and validation set according to 8:2 to ensure that the training set contains samples of each substance; the Adam optimizer is used in model training, and the cross entropy loss function (Cross Entropy Loss) is used as the training indicator for iterative training to obtain the neural network model, and the model is saved.

[0031] The Raman spectrum detection and identification method of trace substances of the present invention utilizes a portable Raman spectrometer to collect the spectrum signal of the detected object on the electric field coupling enhanced SERS substrate, and then transmits the spectrum signal to the Raman spectrum identification algorithm to judge the type of substance respectively. By combining the SERS substrate of the gold nanoarray structure with the noble metal nanosol, the noble metal nanoparticles excite the local surface plasmon resonance (LSP), resulting in the enhancement of the local electromagnetic field on the particle surface, and at the same time, the gold nanoarray structure excites the surface propagation plasmon (SPP), resulting in the enhancement of the electromagnetic field on the array surface, and the SPP near field and the LSP near field are coupled to each other to form a composite SERS substrate with enhanced electric field coupling, which greatly improves the accuracy of Raman spectrum detection, thereby improving the detection sensitivity.

[0032] The present invention constructs a recognition model based on the fusion of a one-dimensional dilated convolutional network and a two-dimensional dilated convolutional network, and realizes the automatic recognition of Raman spectra. A one-dimensional Raman spectrum sequence is actually a function image of the spectrum intensity relative to the Raman frequency, and has local correlation. The one-dimensional dilated convolutional network can be used to mine information such as characteristic peaks in the Raman spectrum curve; similarly, the two-dimensional wavelet coefficient map generated after continuous wavelet transform contains the spatial information of the spectral characteristic peaks. The fusion of the two features improves the recognition accuracy. Compared with the traditional Raman spectrum matching method, this recognition method has a stronger generalization ability, especially for Raman spectra with weak signals or unclear characteristic peaks, the recognition accuracy is significantly improved.

[0033] The recognition algorithm in the present invention combines a one-dimensional dilated convolutional network and a two-dimensional dilated convolutional network to construct a recognition model. By fusing the local feature information of the spectral curve extracted by the one-dimensional dilated convolutional network and the spectral wavelet coefficient map features extracted by the two-dimensional dilated convolutional network, the one-dimensional and two-dimensional features of Raman spectral data are automatically extracted, thereby improving the recognition accuracy.

[0034] The Raman spectrum detection and identification system of trace substances of the present invention comprises a Raman spectrometer, a Raman probe, an industrial computer, a 785nm laser, a CCD, and an electric field coupling enhanced SERS substrate.

[0035] The electric field coupling enhanced SERS substrate is prepared by combining a SERS substrate with a gold nanoarray structure and a noble metal nanosol, wherein the noble metal nanoparticles excite local surface plasmon resonance (LSP), thereby enhancing the local electromagnetic field on the surface of the noble metal nanoparticles, and at the same time, the SERS substrate with a gold nanoarray structure excites surface propagation plasmons (SPP), thereby enhancing the electromagnetic field on the array surface, and the near field of the surface propagation plasmons couples with the near field of the array surface electromagnetic field to form a composite SERS substrate with enhanced electric field coupling.

[0036] The industrial computer executes the Raman spectrum identification algorithm in the Raman spectrum detection and identification method of trace substances of the present invention, processes the Raman spectrum of the trace substances, and identifies the trace substances.

[0037] The Raman spectrum detection and identification system of the present invention adopts a portable Raman spectrometer, which is easy to carry and easy to use. The identification algorithm can be deployed on a cloud server, thereby improving the sensitivity of the operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a system structure block diagram of the detection and identification system of the present invention.

[0039] In the figure: 1. Raman tester and host; 2. Sample collection device; 3. 785nm laser; 4. Raman spectrometer; 5. CCD; 6. Industrial computer; 7. Raman probe; 8. Electric field coupling enhanced SERS substrate.

[0040] Figure 2 This is a Raman spectrum diagram of the R6G reagent tested using the coupling enhanced SERS substrate of the present invention.

[0041] Figure 1. Using coupling enhanced SERS substrate to detect 10 -6 Mol / L R6G Raman spectrum; 2. Using coupling enhanced SERS substrate to detect 10 -8 Mol / L R6G Raman spectrum; 3. Using coupling enhanced SERS substrate to detect 10 -9 Mol / L R6G Raman spectrum; 4. Using nanoarray structure SERS substrate, detection of 10 -6 Mol / L R6G Raman spectrum; 5. Using nanoarray structure SERS substrate to detect 10 -7 Raman spectrum of R6G at mol / L.

[0042] Figure 3 This is a flow chart of the Raman spectrum recognition algorithm based on the fusion of one-dimensional dilated convolutional network and two-dimensional dilated convolutional network.

[0043] Figure 4 This is a structural diagram of the recognition model based on the fusion of one-dimensional dilated convolutional network and two-dimensional dilated convolutional network.

[0044] In the figure: 1. Wavelet coefficient graph; 2. Normalized Raman spectrum curve; 3. Two-dimensional convolution and two-dimensional pooling modules; 4. One-dimensional dilated convolution and maximum pooling modules; 5. Flattening layer of two-dimensional dilated convolution network; 6. Flattening layer of one-dimensional dilated convolution network; 7. Splicing layer; 8. Fully connected layer; 9. Output layer.

[0045] Figure 5 It is a comparison chart of the test accuracy of the present invention and the commonly used Raman spectrum recognition method on a partial data set A of public Raman data.

[0046] Figure 6 This is a comparison chart of the test accuracy of the present invention and the commonly used Raman spectrum recognition method on a partial data set B of public Raman data.

[0047] Figure 7 It is a comparison chart of the test accuracy, precision, recall rate and F1-score indicators of the present invention and the commonly used Raman spectrum recognition method on the public Raman data part data set B.

[0048] In the figure: 1. Accuracy; 2. Precision; 3. Recall; 4. F1-score. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0050] like Figure 1 As shown, the Raman spectrum detection and identification system of trace substances of the present invention includes two major parts: a Raman tester and a host 1 and a sample collection device, specifically including: a Raman spectrometer 4, a Raman probe 7, a CCD 5, an industrial computer 6, a 785nm laser 3 and an electric field coupling enhanced SERS substrate 8.

[0051] Among them, the electric field coupling enhanced SERS substrate 8 is achieved by combining a SERS substrate with a gold nanoarray structure and a noble metal nanosol. The noble metal nanoparticles excite local surface plasmon resonance (LSP), resulting in an enhancement of the local electromagnetic field on the surface of the noble metal nanoparticles. At the same time, the SERS substrate with a gold nanoarray structure excites surface propagation plasmons (SPP), resulting in an enhancement of the electromagnetic field on the array surface. The near-field of the surface propagation plasmons is coupled with the near-field of the array surface electromagnetic field to form a composite SERS substrate with enhanced electric field coupling.

[0052] The industrial computer 6 executes the Raman spectrum identification algorithm in the Raman spectrum detection and identification method of trace substances of the present invention, processes the Raman spectrum of the trace substances, and identifies the trace substances.

[0053] The Raman spectrum detection and identification method of trace substances of the present invention refers to Figure 1 , Figure 3 and Figure 4 As shown, 2 μl of gold particles with a radius of 40 to 50 nm were taken and dropped onto the enhanced chip of the gold nanoarray, and dried in a 60°C constant temperature oven for 2 minutes to prepare a composite SERS substrate enhanced by electric field coupling. Then, 4 substances to be tested were prepared: A was R6G, B was ethanol, C was acetonitrile, and D was pure water; 10 -6 Mol / L, 10 -7 Mol / L, 10 -8 Mol / L, 10 -9 Mol / L R6G solution was prepared with ethanol solution and acetonitrile solution in volume ratio of 1:100 and 1:1000. The prepared solutions were dropped on the composite SERS substrate and detected by a portable Raman spectrometer with 785nm laser. For comparison, the Raman spectrum signal of R6G with only nanoarray enhanced chip was detected at the same time. The detection results of the two enhanced substrates are shown in Figure 2. Figure 2 shown.

[0054] The detection point and detection time were changed, and each solution was detected multiple times, and the Raman data were saved. Among them, ethanol and acetonitrile are volatile, and their Raman signal intensity weakens with the increase of detection time. The tested Raman data are used as sample data and are sorted, as shown in Table 1.

[0055] Table 1: The detection results of the present invention on different concentrations of R6G, ethanol and acetonitrile.

[0056]

[0057] The collected spectral signals are processed by Raman spectrum recognition algorithm, which specifically includes three steps: Raman spectrum preprocessing, recognition model construction based on the fusion of one-dimensional dilated convolutional network and two-dimensional dilated convolutional network, and model training and spectrum recognition.

[0058] S1, Raman spectrum preprocessing:

[0059] Set the Raman data format of the original sample to:

[0060] W raw =[W1,W2,...,W n ]、X raw =[X1,X2,...,X n ] and y raw =[y1,y2,...,y n ]

[0061] Among them, W raw represents the Raman shift space of all samples, W i represents the Raman shift of the i-th sample, n represents the number of sample data (n=1042 in this embodiment), X raw represents the signal intensity space of the sample, X i represents the signal strength of the i-th sample, y raw represents the sample category space, y i Indicates the type of the i-th sample.

[0062] S1.1, interpolation operation: Through piecewise linear interpolation operation, a function is created based on the original Raman data, and the frequency shift range of the unified sample is: [201,2248cm -1 ], interval: 1cm -1 , a total of 2048 Raman data are generated for each sample, that is, m=2048.

[0063] S1.2, baseline correction: Adaptive iterative weighted penalty least squares method is used to perform baseline correction on the data obtained by the interpolation operation in S1.1 to eliminate the fluorescence effect of Raman spectroscopy.

[0064] S1.3, smoothing and denoising: Use the Savitzky-Golay filter to smooth the data generated by the S1.2 baseline correction.

[0065] S1.4, Normalization: The data set obtained by smooth denoising in S1.3 is normalized according to the following formula to unify the Raman signal intensity of each sample to the interval [0,1]:

[0066]

[0067] Among them, x i,j is the signal strength of the jth frequency shift position of the ith sample, min(x i, ) is the minimum signal strength of the ith sample, max(x i, ) is the maximum signal intensity of the i-th sample.

[0068] S1.5 Continuous wavelet transform: The data set normalized in S1.4 is transformed using the Mexican hat wavelet as the mother wavelet function. The scale range of the wavelet function is set to (1,31), and the generated wavelet coefficient map is used as the input of the two-dimensional dilated convolutional network:

[0069]

[0070] Here, ν is the Raman shift and δ is the scale.

[0071] S2, build a recognition model that combines a one-dimensional dilated convolutional network and a two-dimensional dilated convolutional network. The recognition model includes an input layer, a feature extraction layer, a feature classification prediction layer, and an output layer. The specific recognition model is as follows: Figure 3 shown.

[0072] S2.1, the data normalized by S1.4 is used as the input of the one-dimensional dilated convolutional network, and the wavelet coefficient map obtained by the continuous wavelet transform in S1.5 is used as the input of the two-dimensional dilated convolutional network.

[0073] 2.2, the feature extraction layer includes two parts: one-dimensional dilated convolutional network feature extraction layer and two-dimensional dilated convolutional network feature extraction layer:

[0074] The one-dimensional dilated convolutional network feature extraction layer consists of four one-dimensional blocks and a flattening layer. Each block includes a dilated convolution module and a one-dimensional maximum pooling module. The parameters of the four dilated convolution modules (in_channels, filters, kernel_size, dilation, stride) are (1, 6, 3, 2, 1), (6, 16, 3, 2, 1), (16, 32, 3, 2, 1) and (32, 64, 3, 2, 1), respectively, and all paddings are 1;

[0075] The 2D dilated convolutional network feature extraction layer consists of four 2D blocks and a flattening layer. Each block includes a 2D dilated convolution module and a 2D maximum pooling module. The parameters (in_channels, filters, kernel_size, dilation, stride) of the four 2D dilated convolution modules are (3, 8, 3, 2, 1), (8, 16, 3, 2, 1), (16, 32, 3, 2, 1) and (32, 64, 3, 2, 1), respectively. All paddings are 1. The data after each dilated convolution operation is activated by the ReLU function:

[0076] The data after the dilated convolution operation is activated using the ReLU function;

[0077] x conv_out =floor((x con_in +2×padding-dilation×(kernel_size-1)-1) / stride+1)

[0078] Among them, x con_in is the input data of the dilated convolutional layer, x conv_out is the output data of the expanded convolution layer, kernel_size is the convolution kernel size, dilation is the expansion coefficient, padding is the padding size, and stride is the convolution kernel movement step size.

[0079] The parameters (pool_size, stride) of the maximum pooling module are both (2, 2), and the output after pooling is:

[0080] x pooling_out =floor((x conv_out -pool_size+2×padding) / stride+1)

[0081] Among them, x conv_out is the input data of the pooling module, x pooling_out is the output data of the pooling module, pool_size is the pooling module size, padding is the padding size, and stride is the step size.

[0082] The S2.3 classification prediction layer includes a splicing layer and a fully connected layer. The splicing layer is composed of a one-dimensional dilated convolutional network and a flattened layer of a two-dimensional dilated convolutional network. The number of neurons in the fully connected layer is 1024. The activation function tanh is used for nonlinearization, and the dropout = 0.5 strategy is set to improve the generalization ability.

[0083] S3, model training and spectrum recognition, the specific steps are as follows:

[0084] S3.1, model training: The input spectral data is randomly selected by layer and divided into training set and validation set according to 8:2 to ensure that the training set contains samples of each substance. The Adam optimizer is used in model training, the learning rate (Learning_rate) is 0.0001, the cross entropy loss function (Cross Entropy Loss) is used as the training indicator, and the number of iterations (Epoch) is 50; after iterative training, the recognition model is stable and the model is saved.

[0085] S3.2, spectral identification: For the substance to be tested, the Raman spectral signal of the substance is detected using the coupled enhanced SERS substrate, and then after spectral preprocessing, the identification model obtained by model training in S3.1 is input to make predictions and judgments, and the model is evaluated.

[0086] The model was evaluated using the test set, and the model recognition accuracy is shown in Table 1. In order to compare the impact of spectral preprocessing, the model was further used to train and predict the original Raman data, and the results are shown in Table 1.

[0087] pass Figure 2 It can be seen that the detection accuracy of the Raman spectrum detection and identification method of the present invention is improved by 2 orders of magnitude compared with the SERS enhanced chip. From the conclusion of Table 1, it can be seen that the Raman spectrum detection and identification method of the present invention can accurately identify even weaker Raman spectrum signals, and thus has a higher recognition accuracy.

[0088] In order to further verify the performance of the Raman spectroscopy detection and recognition method (DiCNN) of the present invention, a public Raman spectroscopy dataset (Rruff) was used for testing and compared with other commonly used Raman spectroscopy recognition methods, such as support vector machine (SVM), decision tree (DT), random forest (RF), partial least squares (PLS), K nearest neighbor algorithm (KNN), linear discriminant analysis (LDA), logistic regression (LG), Xgboost, etc.

[0089] Rruff data, due to its imbalanced data volume, some samples have more than 20 data volumes, while some samples have only 1 data volume. In this embodiment, two data sets are selected, as follows:

[0090] Dataset A, selected data with sample size > 10 in the Rruff data, collected 1045 samples of 66 categories of substances;

[0091] Dataset B, selected data with sample size > 3 in the Rruff data, collected a total of 578 types of substances and 3394 samples;

[0092] According to the steps of the Raman spectrum detection and identification method of the present invention, the Raman spectrum of each substance in the two data sets (A, B) is identified.

[0093] The same data division was used to identify Raman spectra on two data sets (A, B) using other methods.

[0094] The recognition results are as follows: Comparison of the accuracy of different methods in data set A Figure 5 As shown; the accuracy comparison of different methods in data set B is Figure 6 The accuracy, precision, recall and F1-score of different methods on dataset B are shown in Figure 7 The accuracy of different methods on dataset B for 30 times is shown in Table 2.

[0095] Table 2: Multiple recognition results of different recognition algorithms on the Rruff dataset

[0096]

[0097]

[0098] By comparison, it is found that the detection method of the present invention has better accuracy and other indicators than commonly used recognition methods such as LR, RF and Xgboost, and has strong generalization ability, whether it is for data set A or data set B. Since the number of samples of many substances in data set B is small, the generalization ability of recognition methods such as RF and Xgboost is significantly reduced, although the training accuracy is high.

[0099] The detection method of the present invention has the advantages of high detection accuracy, accurate discrimination, etc., and is particularly suitable for portable identification of the type and variety of trace substances and their "Raman fingerprints".

Claims

1. A Raman spectroscopy detection and identification method for trace substances, characterized in that: The Raman spectrometer is used to collect the spectral signal of the detection object placed on the electric field coupling enhanced SERS substrate. The collected spectral signal is processed by the Raman spectrum recognition algorithm to obtain the judgment result of the material type of the detection object. The Raman spectrum recognition algorithm includes: Raman spectrum preprocessing, construction of a recognition model integrating a one-dimensional dilated convolutional network and a two-dimensional dilated convolutional network, and model training and spectrum recognition.

2. The Raman spectrum detection and identification method of trace substances according to claim 1 is characterized in that: Raman spectrum preprocessing includes the following steps: Set the Raman data format of the original sample to: W raw = [W1, W2,..., W n , X raw = [X1, X2,..., X n , and y raw = [y1, y2,..., y n ; Among them, W raw represents the Raman frequency shift space of all samples, W i represents the Raman frequency shift vector of the i-th sample, n represents the sample data volume, X raw represents the signal intensity space of the sample, X i represents the signal strength vector of the i-th sample, y raw represents the sample category space, y i Indicates the type of the i-th sample; S1.1, interpolation operation: through piecewise linear interpolation operation, a function is created based on the original Raman data, and the frequency shift range of the unified sample is: [201,2200cm -1 ], interval: 1cm -1 , a total of 2048 Raman data were generated for each sample; S1.2, baseline correction: Adaptive iterative weighted penalty least squares method is used to perform baseline correction on the data obtained by S1.1 interpolation operation to eliminate the fluorescence effect of Raman spectrum; S1.3, smooth denoising: use Savitzky-Golay filter to smooth the data generated by S1.2 baseline correction; S1.4, Normalization: The data set obtained by smooth denoising in S1.3 is normalized according to the following formula to unify the Raman signal intensity of each sample to the interval [0,1]: Among them, x i,j is the signal strength of the jth frequency shift position of the ith sample, min(x i, ) is the minimum signal strength of the ith sample, max(x i, ) is the maximum signal strength of the i-th sample; S1.5, continuous wavelet transform: The data set obtained by normalization in S1.4 is transformed using the Mexican hat wavelet as the mother wavelet function, and the generated wavelet coefficient map is used as the input of the two-dimensional dilated convolutional network.

3. The Raman spectrum detection and identification method of trace substances according to claim 2 is characterized in that: The recognition model of the fusion of the dilated convolutional network and the two-dimensional dilated convolutional network includes a feature input layer, a feature extraction layer, a feature classification prediction layer and an output layer; S2.1, the feature input layer includes: using the data obtained by normalization in S1.4 as the input of the one-dimensional dilated convolutional network, and using the wavelet coefficient map obtained by continuous wavelet transform in S1.5 as the input of the two-dimensional dilated convolutional network; S2.2, the feature extraction layer includes two parts: one-dimensional dilated convolutional network feature extraction layer and two-dimensional dilated convolutional network feature extraction layer: The one-dimensional dilated convolutional network feature extraction layer uses four one-dimensional blocks and a flattening layer. Each block includes a dilated convolution module and a one-dimensional maximum pooling module. The 2D dilated convolutional network feature extraction layer uses four 2D blocks and a flattening layer. Each block includes a 2D dilated convolution module and a 2D maximum pooling module. The data after each dilated convolution operation is activated by the ReLU function, and the data after the dilated convolution changes to: x conv_out =floor((x con_in +2×padding-dilation×(kernel_size-1)-1) / stride+1) Among them, x con_in is the input data of the dilated convolutional layer, x conv_out is the output data of the expanded convolution layer, kernel_size is the convolution kernel size, dilation is the expansion coefficient, padding is the padding size, and stride is the convolution kernel movement step size; S2.3, the classification prediction layer includes a splicing layer and a fully connected layer. The splicing layer is used to splice the flattening layer of the one-dimensional dilated convolutional network and the two-dimensional dilated convolutional network. The fully connected layer uses the activation function tanh to implement nonlinearity and improves the generalization ability by adopting the dropout strategy; S2.4, the output layer uses the activation function softmax, and the output of the activation function softmax is the prediction result of the entire recognition algorithm.

4. The Raman spectrum detection and identification method of trace substances according to claim 3 is characterized in that: Model training and spectrum recognition include the following steps: S3.1, model training: The input spectral sample data is randomly selected by layer and divided into a training set and a validation set according to a ratio of 8:2 to ensure that the training set contains samples of each substance; the Adam optimizer is used in model training, and the cross entropy loss function (Cross Entropy Loss) is used as the training indicator for iterative training to obtain a neural network model, and the model is saved; S3.2, spectral identification: For the substance to be tested, the coupled enhanced SERS substrate is used to detect the Raman spectral signal of the substance, and then after spectral preprocessing, it is input into the neural network model obtained by S3.1 model training to make predictions and judgments, and the model is evaluated.

5. A Raman spectrum detection and identification system for trace substances, characterized in that: Including Raman spectrometer, Raman probe, industrial computer, 785nm laser, CCD, electric field coupling enhanced SERS substrate; The electric field coupling enhanced SERS substrate is prepared by combining a SERS substrate with a gold nano-array structure with a noble metal nano-sol, wherein the noble metal nano-particles excite local surface plasmon resonance (LSP), thereby enhancing the local electromagnetic field on the surface of the noble metal nano-particles, and at the same time, the SERS substrate with a gold nano-array structure excites surface propagation plasmons (SPP), thereby enhancing the electromagnetic field on the array surface, and the near field of the surface propagation plasmons is coupled with the near field of the electromagnetic field on the array surface, thereby forming a composite SERS substrate with enhanced electric field coupling; The industrial computer executes the Raman spectrum identification algorithm in the Raman spectrum detection and identification method of trace substances according to any one of claims 1 to 4, and processes the Raman spectrum of the trace substances to identify the trace substances.

Images