Multi-channel drug fuzzy identification method and system based on surface-enhanced raman spectroscopy

Abstract

The application discloses a multi-channel drug fuzzy identification method and system based on surface-enhanced Raman spectroscopy. The method comprises the following steps: acquiring surface-enhanced Raman spectroscopy sequences of N sampling points of a sample to be measured; inputting the N surface-enhanced Raman spectroscopy sequences into N independent channel input ends of a trained CNN-GRU deep learning model respectively, and inputting the N surface-enhanced Raman spectroscopy sequences into a centralized channel input end of the trained CNN-GRU deep learning model; and the CNN-GRU deep learning model outputs a drug category or a drug category and a drug concentration grade. The CNN-GRU deep learning model introduces CNN and GRU into the analysis of Raman spectroscopy data features, so as to completely extract the detailed features and spatial correlation in the spectrum and be used for judging the drug type, can realize the rapid, simultaneous and high-accuracy identification of multiple drug types, and can also accurately quantitatively identify the concentration of the drug.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug detection, in particular to a multi-channel drug fuzzy recognition method and system based on surface-enhanced Raman spectroscopy. BACKGROUND

[0002] Common drug detection methods include chromatography, capillary electrophoresis, chemical coloration and colloidal gold method. Chromatography is favored by the drug detection industry due to its high efficiency, high selectivity, small sample size and other advantages. However, these methods have complex and tedious pretreatment processes, long detection times, require professional personnel to operate, and have great detection limitations. In recent years, substance analysis methods based on surface-enhanced Raman spectroscopy (SERS) have attracted widespread attention in drug detection due to their rapidity, sensitivity and high selectivity. However, with the practical application of SERS technology in complex matrices, it is difficult to separate the signal of trace substances from the matrix background, and the signal of the target substance cannot be observed. Therefore, the use of Raman spectroscopy for substance recognition in practical applications requires extremely high data analysis of the spectrum.

[0003] The advent of artificial intelligence (AI) technology has provided a new approach to scientific research. Deep learning (DL), as a subfield of artificial intelligence, has the ability to uncover non-linear relationships within complex systems. Deep learning models can learn important raw data patterns without using advanced preprocessing and feature extraction techniques. This technology is suitable for discovering complex relationships in complex systems and has far superior capabilities to traditional programs. In recent years, deep learning has attracted widespread attention in biology, chemistry, industry, medicine and many other fields. Fuzzy recognition refers to identifying useful parts in fuzzy data, i.e., making relevant connections between received information and past memories and experiences, and eliminating irrelevant information. For high-throughput identification of target substances in complex systems, fuzzy recognition is one of the few effective ways. In the field of artificial intelligence, fuzzy recognition strategies are widely used in face recognition, vehicle recognition, and text recognition.

[0004] In 2019, Ju L et al. constructed a three-dimensional fluorescence spectrum based on the interaction of silver nanocluster probes with different drugs, and used a deep learning model to capture the fingerprint information of drug molecules in the three-dimensional fluorescence difference spectrum, achieving fuzzy recognition of five different drugs in urine. However, this method has low sensitivity for simultaneous detection of multiple drugs, and is prone to fluorescence errors leading to false negatives and false positives. In addition, the acquisition of three-dimensional fluorescence difference spectrum is complex and not suitable for on-site rapid detection. In 2021, Ciloglu F U et al. proposed a deep neural network to identify antibiotic-resistant bacteria using surface-enhanced Raman spectroscopy. Although this method shows the great potential of deep learning in using SERS spectral data to characterize and detect substances, it cannot achieve fuzzy recognition of multiple drugs. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art, realize fast and high-accuracy identification of multiple drugs at the same time, and simultaneously display qualitative classification and quantitative identification, and particularly innovatively proposes a multi-channel drug fuzzy identification method and system based on surface-enhanced Raman spectrum.

[0006] In order to achieve the above-mentioned purpose of the present application, the present application provides a multi-channel drug fuzzy identification method based on surface-enhanced Raman spectrum, comprising: acquiring surface-enhanced Raman spectrum sequences of N sampling points of a to-be-tested sample, wherein N is a positive integer greater than 1; inputting the N surface-enhanced Raman spectrum sequences into N independent channel input ends of a trained CNN-GRU deep learning model respectively, and inputting the N surface-enhanced Raman spectrum sequences into a centralized channel input end of the trained CNN-GRU deep learning model; the CNN-GRU deep learning model outputs a drug category, or the CNN-GRU deep learning model outputs a drug category and a drug concentration level.

[0007] In order to achieve the above-mentioned purpose of the present application, based on the same inventive concept, the second aspect of the present application provides an electronic device comprising an input unit and a processor; the input unit is used to receive surface-enhanced Raman spectrum sequence data of N sampling points of a to-be-tested sample and transmit to the processor; the processor acquires a drug category in the to-be-tested sample according to the multi-channel drug fuzzy identification method based on surface-enhanced Raman spectrum described in the first aspect of the application, or acquires a drug category and a drug concentration level of the to-be-tested sample.

[0008] In order to achieve the above-mentioned purpose of the present application, based on the same inventive concept, the third aspect of the present application provides a multi-channel drug fuzzy identification system based on surface-enhanced Raman spectrum, comprising: a to-be-tested sample carrier for carrying a to-be-tested sample, wherein gold nanoparticles are arranged in the to-be-tested sample carrier as an enhanced substrate; a laser for outputting laser light towards the to-be-tested sample; a Raman spectrometer for measuring surface-enhanced Raman spectrum of the to-be-tested sample; and the electronic device described in the second aspect of the present application, wherein the electronic device is connected with the Raman spectrometer.

[0009] To sum up, by adopting the technical scheme, the application has the beneficial effects that the CNN-GRU deep learning model introduces the CNN and the GRU into the analysis of the Raman spectrum data features, so as to completely extract the detail features and the spatial correlation in the spectrum and be used for judging the drug type, can realize the fast, simultaneous and high-accuracy identification of multiple drug types, can also accurately quantitatively identify the concentration of the drug, can collect the surface enhanced Raman spectrum sequence of the sample at multiple sampling points, so that the spatial information extraction process of different sampling points has independence, the multi-channel spectral spatial feature extraction process makes the CNN-GRU deep learning model have stronger robustness to data anomalies, can limit the influence of abnormal data, the application establishes a rapid detection technology suitable for the seized drugs and the drug of the human body examination material, provides a reference, is beneficial to the maximum efficiency of the drug detection link in the drug suppression work, and promotes the development of the drug suppression work. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of a multi-channel drug fuzzy identification method based on surface enhanced Raman spectrum of the embodiment 1 of the application;

[0011] Figure 2 is a structure block diagram of the CNN-GRU deep learning model in the embodiment 1 of the application;

[0012] Figure 3 is a specific structure schematic diagram of the CNN-GRU deep learning model in the embodiment 1 of the application;

[0013] Figure 4 is a structure schematic diagram of the GRU in the embodiment 1 of the application;

[0014] Figure 5 is a training curve in one application scenario of the embodiment 1 of the application;

[0015] Figure 6 is a test machine confusion matrix schematic diagram in one application scenario of the embodiment 1 of the application;

[0016] Figure 7 is the t-SNE dimension reduction visualization result of qualitative classification and quantitative classification in one application scenario of the embodiment 1 of the application;

[0017] Figure 8 is the test machine confusion matrix and the t-SNE dimension reduction visualization result of the CNN-GRU deep learning model and the CNN model when the sampling point data is abnormal in one application scenario of the embodiment 1 of the application;

[0018] Figure 9 is the training curve of the model when multiple types of drugs are mixed in one application scenario of the embodiment 1 of the application;

[0019] Figure 10 is a confusion matrix of a test machine of a classification model in a mixed type of drugs in an application scenario of embodiment 1 of the present application;

[0020] Figure 11 is a t-SNE dimensionality reduction visualization result of a classification model in a mixed type of drugs in an application scenario of embodiment 1 of the present application.

[0021] Reference signs:

[0022] 1 single-channel CNN feature extraction module; 2 multi-channel feature splicing module; 3 sequence correlation feature extraction module; 4 centralized channel CNN feature extraction module; 5 drug qualitative classification module; 6 drug quantitative classification module. DETAILED DESCRIPTION

[0023] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary, only for explaining the present application, and cannot be understood as a limitation of the present application.

[0024] In the description of the present application, it should be understood that the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0025] In the description of the present application, unless otherwise specified and limited, it should be noted that the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be mechanical connection or electrical connection, or the communication between two elements, or direct connection, or indirect connection through intermediate medium, and the specific meaning of the above terms can be understood by those skilled in the art according to the specific circumstances.

[0026] Embodiment 1

[0027] The present embodiment discloses a multi-channel drug fuzzy identification method based on surface enhanced Raman spectrum, as shown in Figure 1 , comprising:

[0028] Step S1, obtaining N surface-enhanced Raman spectrum sequences of N sampling points of a sample to be measured, N being a positive integer greater than 1. The N sampling points can be N different regions or position points of the sample to be measured. Preferably, N is 5 to 15, and more preferably, N is 10, considering the amount of calculation and accuracy. The sample to be measured is preferably but not limited to heroin or ketamine or 3,4-methylenedioxymethamphetamine, or a mixture of at least two of the three. A single surface-enhanced Raman spectrum sequence is shown in Figure 3 , in which the abscissa is the wavelength number and the ordinate is the excitation intensity curve.

[0029] Further preferably, to improve the measurement accuracy, the sample to be measured not only contains the drug to be measured, but also includes a gold nanoparticle (AuNP) enhanced substrate. Gold nanoparticles (AuNPs) are used as a surface-enhanced Raman spectrum SERS enhanced substrate. Due to the signal amplification effect of AuNPs on the target, they are applied to the signal enhancement of heroin, ketamine, and 3,4-methylenedioxymethamphetamine in solution.

[0030] Step S2, inputting the N surface-enhanced Raman spectrum sequences into the N independent channel input ends of the trained CNN-GRU deep learning model, and inputting the N surface-enhanced Raman spectrum sequences into the centralized channel input end of the trained CNN-GRU deep learning model.

[0031] The CNN-GRU deep learning model has N independent input channels, each of which is used to extract the spatial detail features of the sequence in the channel. The mode of inputting the N surface-enhanced Raman spectrum sequences collectively is preferably but not limited to combining the N surface-enhanced Raman spectrum sequences into a two-dimensional matrix for input. The centralized channel input end of the CNN-GRU deep learning model is used to globally extract the spatial detail features of the enhanced Raman spectrum of the sample to be measured.

[0032] Step S3, the CNN-GRU deep learning model outputs the drug category, or the CNN-GRU deep learning model outputs the drug category and the drug concentration level. The CNN-GRU deep learning model can output P category labels, P being a positive integer, representing whether P types of drugs exist. The CNN-GRU deep learning model can output Q concentration level labels, Q being a positive integer.

[0033] In this embodiment, preferably, the structure diagram of the CNN-GRU deep learning model is shown in Figure 2 Figure 3 , and the network layers and network parameters in the model are shown in Table 1.

[0034] Table 1 CNN-GRU deep learning model structure table

[0035] ​

[0036]

[0037] Specifically, as shown in Figure 2 and Figure 3 The CNN-GRU deep learning model comprises:

[0038] N single-channel CNN feature extraction modules 1 are respectively used to extract independent spectral features of the N surface-enhanced Raman spectrum sequences, and the independent spectral features are detailed features of the surface-enhanced Raman spectrum sequences, which can be spatial features. The surface-enhanced Raman spectrum sequence is one-dimensional data. Preferably, in order to better extract the detailed features of each surface-enhanced Raman spectrum sequence, the single-channel CNN feature extraction module 1 comprises more than one cascaded first convolutional pooling unit, and the first convolutional pooling unit comprises a first one-dimensional convolutional layer and a first pooling layer connected in series. Further preferably, the single-channel CNN feature extraction module 1 comprises two cascaded first convolutional pooling units, as shown in Table 1.

[0039] A multi-channel feature splicing module 2 splices the independent spectral features output by the N single-channel CNN feature extraction modules 1 to obtain multi-channel spliced features.

[0040] A sequence-related feature extraction module 3 extracts sequence-related features based on the multi-channel spliced features, and the sequence-related features are spatial correlation features of each sequence. Preferably, the sequence-related feature extraction module 3 comprises more than one second convolutional pooling unit and more than one first GRU layer connected in sequence, and the second convolutional pooling unit comprises a second one-dimensional convolutional layer, a second pooling layer and a first batch normalization layer connected in sequence. The sequence-related feature extraction module 3 further extracts spatial detailed features of the multi-channel spliced features by convolution, and then extracts spatial correlation features between sequences by a gated neural unit GRU. Preferably, as shown in Table 1, the second convolutional pooling unit comprises two cascaded second convolutional pooling units and two cascaded GRU layers.

[0041] A centralized channel CNN feature extraction module 4 extracts centralized spectral features of the N surface-enhanced Raman spectrum sequences input in a centralized manner; and the centralized channel CNN feature extraction module 4 realizes extraction of global spatial detailed features of the N surface-enhanced Raman spectrum sequences. Preferably, in order to better extract the global spatial detailed features, the data is standardized, and the centralized channel CNN feature extraction module 4 comprises more than one third convolutional pooling unit and a first two-dimensional convolutional layer connected in cascade, and the third convolutional pooling unit comprises two first two-dimensional convolutional layers, a third pooling layer and a second batch normalization layer connected in sequence. As shown in Table 1, the centralized channel CNN feature extraction module 4 comprises three cascaded third convolutional pooling units.

[0042] The drug qualitative classification module 5 obtains a qualitative classification feature by splicing the sequence correlation feature and the centralized spectrum feature, and outputs a drug category contained in the to-be-tested sample based on the qualitative classification feature. The qualitative classification feature combines the spatial feature and the correlation feature between the spatial features in the sequence, and the classification is performed after the two features are fused, which can improve the accuracy of the qualitative classification. Preferably, as shown in Table 1, the drug qualitative classification module 5 includes a second GRU layer, a second two-dimensional convolution layer, a first global pooling layer, a first feature splicing layer, a first full connection layer, a third batch normalization layer, a third full connection layer, a fourth full connection layer and a drug category output layer connected in sequence.

[0043] The drug quantitative classification module 6 obtains a quantitative classification feature by splicing the sequence correlation feature and the centralized spectrum feature, and outputs a concentration level of the drug in the to-be-tested sample based on the quantitative classification feature. The quantitative classification feature combines the spatial feature and the correlation feature between the spatial features in the sequence, and the classification is performed after the two features are fused, which can improve the accuracy of the quantitative classification. Preferably, as shown in Table 1, the drug quantitative classification module 6 includes a third GRU layer, a third two-dimensional convolution layer, a second global pooling layer, a second feature splicing layer, a fifth full connection layer, a fourth batch normalization layer, a sixth full connection layer, a seventh full connection layer and a drug concentration level output layer connected in sequence.

[0044] In this embodiment, the CNN-GRU deep learning model extracts the spatial feature and the sequence feature in the input spectrum information in sequence, which can effectively establish the mapping relationship between the spectrum information and the material attribute, and the CNN-GRU deep learning model structure further strengthens the information processing ability of the model for the spatial and sequence features, and has good drug attribute analysis ability. As shown in Table 1, the CNN-GRU deep learning model has 23 hidden layers, and the 16-Output layer is a multi-task output module, which is used to apply the key features extracted by the 1-15 layers to two different tasks, i.e., qualitative analysis and quantitative analysis. Through the model, the type and concentration of the drug can be directly detected by the model at the same time. For the spectrum data input by the sampling point set, the wavelength number and the excitation intensity are used as the horizontal and vertical coordinates respectively, and a convolution kernel with (3, 2) and stride=(1, 2) is used to process this type of data. Through the convolution kernel, the model can effectively use the correlation information of the wavelength number and the excitation intensity in the spectrum. For the spectrum data input by the sampling point independently, a one-dimensional convolution kernel is used to scan the overall spectrum data of the sampling point to obtain detailed spatial feature information.

[0045] In this embodiment, in order to realize flexible use and obtain a drug quantitative detection result, a switch module is further included, which is used to start the drug quantitative classification module 6 to output the concentration level of the drug in the to-be-tested sample when the number of drug categories output by the drug qualitative classification module 5 is 1, and to close the drug quantitative classification module 6 when the number of drug categories output by the drug qualitative classification module 5 is greater than 1.

[0046] In the present embodiment, the CNN-GRU deep learning model introduces a convolutional neural network (CNN) and a gated recurrent unit (GRU) into the extraction of Raman spectral features, so as to completely extract the detailed features and spatial correlation in the spectrum. The feature extraction process of the CNN ignores the sequence correlation of the data, which treats all features equally and extracts spatial features by a convolution kernel. The recurrent neural network structure of the GRU is better at analyzing sequence data, that is, the network can analyze the sequential correlation from the sequence information of the spectral features. The structure of the GRU is shown in Figure 4 .

[0047] As shown in Figure 4 , the recurrent neural network structure is composed of sequence input features x t , sequence output features y t , and hidden output h t . Wherein t ∈ [1, 2, …, T] represents a certain point in the sequence, and T is the sequence length. r is the gate threshold for controlling reset, and z is the gate threshold for controlling update. These two thresholds are the key structures for the recurrent structure to remember sequence information. r = σ(W r [x t || h t-1 ]), z = σ(W z [x t || h t-1 ]). In the formula, σ is a sigmoid function, and W r and W z are learnable gate unit parameters. The hidden state of the previous sequence can affect the calculation process of the current feature. h d = tanh(W[x t || h t-1 ]), where h d mainly contains the current input x t data. Adding h d to the current hidden state is equivalent to remembering the state of the current sequence position. Thus, the hidden state of the sequence position is: h t = y t = (1-z)⊙h t-1 +z⊙h d . As can be seen from the above formula, the output of the recurrent neural network at a certain sequence position is determined by the previous sequence features and the sequence features at the current position. This structure can effectively extract sequence information. Meanwhile, compared with a long short-term memory unit (LSTM), the learning process of the GRU is faster.

[0048] In this embodiment, the spectral information is extracted by a Multichannel-CNN-GRU (MCG) deep model structure. Meanwhile, each sampling point is analyzed by an independent CNN model. The data information of N sampling points is spliced after the feature extraction by the CNN model, and the feature extraction is further completed by the GRU model. In this process, the spatial correlation and sequence correlation of each sampling point data are extracted respectively. Thus, the detailed features in the Raman spectrum are extracted and used to determine the type of drugs. The fuzzy recognition process of surface-enhanced Raman spectrum of drugs based on multichannel CNN-GRU is as shown in Figure 3 .

[0049] In this embodiment, as shown in Figure 2 and Figure 3 , the CNN-GRU deep learning model has two parallel outputs, which are qualitative function and quantitative function respectively. The output layer of the multi-task classifier uses the soft max function as the activation function, as shown in the following formula:

[0050]

[0051] wherein, soft max(Z i ) represents the probability output of the category, Z i is the output of the neuron, and C is the number of categories in the output layer. Therefore, the output threshold of the neural network can be adjusted to adapt to different tasks. Thanks to the effective feature extraction and information recognition of the CNN-GRU structure, the model can simultaneously complete the qualitative and quantitative analysis of the substance. At the same time, the multi-channel model structure makes the spatial information extraction process of different sampling points independent. The multi-channel spectral spatial feature extraction process makes the model have stronger robustness to data anomalies. When the data of some sampling points is wrong, the features of this sampling point will not spread to the spatial information extraction process of other sampling points. Therefore, the model has the ability to limit the influence of abnormal data on the analysis results of the substance.

[0052] In one application scenario of the embodiment, three drugs, such as heroin, ketamine and 3,4-methylenedioxymethamphetamine, need to be tested simultaneously, and each drug has 9 concentration levels, so P is 3 and Q is 9. The receptive field and kernel number of each neural network layer in the CNN-GRU deep learning model are shown in Table 1. Among them, the single-channel CNN feature extraction module 1 includes the 2nd to 6th layers on the independent input side (left side) of the sampling points in Table 1, the multi-channel feature splicing module 2 includes the 7th layer on the independent input side (left side) of the sampling points in Table 1, the sequence-related feature extraction module 3 includes the 7th to 15th layers on the independent input side (left side) of the sampling points in Table 1, the centralized channel CNN feature extraction module 4 includes the 2nd to 15th layers on the centralized input side (right side) of the sampling points in Table 1, the drug qualitative classification module 5 includes the 16th to Output layers on the independent input side (left side) of the sampling points in Table 1, and the drug quantitative classification module 6 includes the 16th to Output layers on the centralized input side (right side) of the sampling points in Table 1.

[0053] In this application scenario, preferably, the training process of the CNN-GRU deep learning model includes:

[0054] Step one, gold nanometer substrate preparation.

[0055] Since AuNPs have the effect of amplifying the signal of the target, gold nanoparticles (AuNPs) are used as the surface-enhanced Raman spectroscopy (SERS) enhancement substrate. The AuNPs solution preparation process is as follows: 10 mg of chloroauric acid (HAuCl4·4H2O) is dissolved in 20 mL of distilled water. The solution is heated to boiling. Then, 1 mL of 1% sodium citrate is added dropwise to the solution, the heating is stopped, and the AuNPs solution is stirred for 5 min and turns wine red.

[0056] Step two, sample preparation.

[0057] A uniform solution of ketamine (1.0 mg / mL) was prepared by dissolving 5 mg of ketamine in 5 mL of ultrapure water and sonicating for 5 min. This solution was used to prepare a series of standard samples with different concentrations of ketamine: 1000 ppm, 500 ppm, 100 ppm, 50 ppm, 10 ppm, 5 ppm, 1 ppm, 0.5 ppm, 0.1 ppm. Similarly, a solution of methamphetamine (1.0 mg / mL) and a solution of heroin (1.0 mg / mL) were prepared by dissolving 5 mg of each drug in 5 mL of ethanol. A series of standard samples with different concentrations of each drug was prepared by diluting the solutions with ethanol: 500 ppm, 100 ppm, 50 ppm, 10 ppm, 5 ppm, 1 ppm, 0.5 ppm, 0.1 ppm. Five uL of each drug was added to 100 uL of the AuNPs solution and mixed well to obtain the respective mixtures. On this basis, 20 uL of each mixture was taken and mixed well to obtain the mixture of the three drugs. Five uL of each mixture was immediately dropped on a CaF2 slide and dried at room temperature for about 30 min for SERS measurements.

[0058] Step three, raw data set acquisition.

[0059] SERS measurements were performed with a Raman spectrometer using a 50x (0.75 NA) microscope objective at 3 mW laser power with 785 nm excitation for 1 s exposure. The laser spot size was 1.3 pm (1.22 x l / NA). Spectra were collected with a 5 pm step to prevent overlap. For reproducibility, two data sets were acquired on different days. In the first data set, spectra were collected from 4 different areas of the sample prepared for each isolate. From each mixture, 60 spectra were collected at the given measurement parameters. The samples were also collected from the second data set in the same way. In total, 120 spectra were collected from each isolate in the two data sets on different days for the solutions without and with different concentrations of drugs. Thus, the total data set consisted of 3840 spectra, using 1200 lines / mm -1 grating, providing a spectral range from 100 to 2500 cm -1 .

[0060] Step four, information-enhanced training data set construction.

[0061] The original data set required for constructing the deep learning model is obtained by step three. It is assumed that the original spectral image contains heroin, ketamine, and 3,4-methylenedioxymethamphetamine, and therefore has three categories for qualitative identification. It is assumed that heroin, ketamine, and 3,4-methylenedioxymethamphetamine have nine concentration gradients, respectively. Therefore, the quantitative experiment has nine categories. Heroin, ketamine, and 3,4-methylenedioxymethamphetamine have 60 sampling information at different positions at each concentration gradient. The sample space of the deep model is constructed by random combination of the spectral data of the sampling points. In actual situations, the time and economic cost of testing 60 sampling points is relatively high. For the original 60 sampling points, 10 points of sampling data are randomly selected to form a sample and repeated 100 times in this paper. Through this downsampling method, each sample only contains 10 sampling points of data, and the number of samples is expanded by 100 times. This downsampling method is more suitable for actual situations, increases the number of samples, and also puts higher requirements on the deep learning model. The deep learning model needs to have stronger feature analysis capability to process less input information. Through the downsampling method, the number of spectral information samples is expanded to 21*100=2100. The spectral data set corresponds to three types for qualitative identification and nine types for semi-quantitative identification. 70% of the two data sets are used as the training data set, and the remaining 30% are used as the test data set.

[0062] Step five, according to Figure 2 、 Figure 3 and Table 1, the network structure of the CNN-GRU deep learning model is constructed, the network structure is trained and the network parameters are adjusted using the training set, and the trained network structure is tested and verified using the test data set, and the CNN-GRU deep learning model is obtained.

[0063] In this application scenario, the convergence verification of the CNN-GRU deep learning model training process is as shown in Figure 5 , Figure 5 Fig. a in the middle indicates the qualitative classification training curve for drug category identification, Figure 5 Fig. b in the middle is the training curve of the quantitative classification result for drug quantitative identification. As can be seen from Figure 5 , the model has strong fitting ability and can effectively learn the feature correlation between the drug spectral data and the drug type and concentration. The qualitative and quantitative accuracy rates of the model on the test set are 99.9% and 99.6%, respectively. This shows that through the spectral data, the model can give accurate material analysis results.

[0064] In this application scenario, the numbers of the drug classification data and the drug quantitative data input are shown in Table 2, and the test machine confusion matrix is shown in Figure 6 , Figure 6 Fig. a in the middle is the type confusion matrix of the drug qualitative classification module 5,Figure 6 Fig. b is the concentration confusion matrix of the drug quantitative classification module 6. From the confusion matrix, it can be seen that the model has accurate drug type and concentration analysis capability. In order to further analyze the performance of the model, the output of the second-to-last layer of the model is taken out and dimensionally reduced by the t-SNE algorithm. The output layer of the deep model is generally directly related to the number of task categories, while the second-to-last layer output is generally a vector feature closely related to the task. By t-SNE, the model vector output is reduced to two dimensions, and the output of the model for different samples can be intuitively displayed on the two-dimensional coordinate plane. The t-SNE dimension reduction pictures of the picture input and the spectrum input and the two tasks (qualitative classification task and quantitative classification task) are as shown in Fig. a and Fig. b of Figure 7 Figure 7 Fig. a of the drawings shows the type feature visualization result, Figure 7 Fig. b of the drawings shows the concentration feature visualization result. From Figure 7 It can be seen that the spectrum data is completely mapped into different feature vectors by the output category after passing through the model. The model proposed in this application can effectively process spectrum data and perform drug type and concentration analysis tasks.

[0065] Table 2 Output category number

[0066]

[0067] In this application scenario, the robustness of the CNN-GRU deep learning model under data anomaly is verified. The CNN-GRU deep learning model proposed in this application can effectively utilize drug spectrum features for qualitative and quantitative detection. In actual application, data anomaly of sampling points occurs from time to time. For example, the sampling point data of a certain drug type is mistakenly mixed with other unknown category data. That is, the 10 sampling point data is not completely the data of this drug type at this concentration. At this time, for the sample affected by data confusion, the deep learning model should have the ability to resist such data anomaly and complete the correct analysis of the drug through the effective features of most sampling points. In this case, the core of the deep learning model is to establish a mapping relationship from spectrum data input to drug type output. The more effective the mapping relationship established by the model, the stronger the robustness of the model against data anomaly. Experiments prove that the CNN-GRU deep learning model proposed in this application has stronger ability to resist data anomaly. It still has higher analysis performance when the sampling point is mixed with unknown substances. The analysis performance of the CNN-GRU deep learning model (referred to as MCG) proposed in this application and the ordinary CNN under different degrees of data anomaly is as shown in Table 3.

[0068] Table 3 Abnormal situation analysis

[0069] ​

[0070] As shown in Table 3, the CNN-GRU deep learning model network proposed in the present application has stronger robustness to data anomalies, and has stronger feature extraction ability and information mapping ability than CNN. When the data of one sampling point is randomly wrong, the accuracy of drug qualitative and quantitative of MCG is 97.3% and 92.1%, which is decreased by 2.6% and 7.5% compared with the normal situation. The accuracy of CNN is decreased to 94.8% and 85.9%, which is decreased by 4.1% and 12.1% compared with the normal situation. The qualitative and quantitative accuracy attenuation of the CNN-GRU deep learning model proposed in the present application when one sampling point data is abnormal is reduced by 57.7% and 61.3% compared with CNN. In the case of more serious data anomaly, the CNN-GRU deep learning model proposed in the present application has better qualitative and quantitative analysis accuracy. In order to further analyze the influence of data anomaly on the performance of the model, the output of the CNN-GRU deep learning model proposed in the present application under different data anomaly conditions is analyzed by using t-SNE algorithm and confusion matrix. Through the visualization display of the output vector of the CNN-GRU deep learning model proposed in the present application, the mapping effect of the CNN-GRU deep learning model proposed in the present application for different sample data can be observed on the two-dimensional plane. When the data of one sampling point is randomly wrong, the sample feature mapping of MCG and CNN is as shown in Figure 8 , wherein, Figure 8 (a1) represents CNN missing 2 type confusion, (a2) CNN represents missing 2 type feature map, (a3) represents MCG missing 2 type confusion; (a4) represents MCG missing 2 type feature map.

[0071] As can be seen from Figure 8 , the CNN-GRU deep learning model proposed in the present application has higher performance in output accuracy and feature mapping. When the data of two sampling points is abnormal, the quantitative analysis ability of CNN is seriously disturbed, and it cannot guarantee the analysis accuracy of class 3. The CNN-GRU deep learning model proposed in the present application still has higher accuracy and performance advantage, and can complete the analysis task of all concentration types. From the feature analysis result of t-SNE, when the data of two sampling points is abnormal, the features of class 3 of CNN are seriously confused with the features of other classes. The samples of this class are similar to the features of other classes after being processed by CNN, so the analysis performance of CNN for this class is seriously affected. The class 3 of the CNN-GRU deep learning model proposed in the present application is still significantly different from the features of other classes in the feature map, which makes the MCG model still have good analysis ability for this class.

[0072] In the application scenario, the simultaneous fuzzy recognition of multiple drugs is verified. In actual situations, drugs generally appear in mixed form. Therefore, it is necessary to detect different drug mixing types. Experiments show that, thanks to the effective spectral feature extraction capability of the CNN-GRU deep learning model structure, the CNN-GRU deep learning model proposed in the application can simultaneously analyze different drug types and their mixing states. In the construction of the deep learning label, the mixing types of heroin, ketamine and 3,4-methylenedioxymethamphetamine drugs are encoded. The combination of the three drugs is used to form a mixed drug data set. Therefore, the output of the classification and recognition model is 3-dimensional, and each dimension refers to a drug type. If the output is 1, it means that the corresponding drug exists. It should be noted that when all three types of drugs exist, an interfering substance is added to detect the performance of the model. The training process of the model is as shown in Figure 9 Fig. 1. It can be known from Figure 9 that the model can effectively learn the feature correlation between the drug spectral data and the drug type. The qualitative accuracy of the model on the test set is 99.9%. This shows that through the spectral data, the model can give accurate mixed drug qualitative analysis results. Table 4 shows the output category number of the qualitative classification module when different drugs are mixed.

[0073] Table 4 Output category number

[0074]

[0075] Figure 10 The mixed drug type confusion matrix is shown in Table 5. The t-SNE algorithm is used to obtain the multi-drug type feature visualization result as shown in Figure 11 Fig. 2. It can be known from Figure 11 that the CNN-GRU deep learning model proposed in the application can effectively complete the task of mixed multi-substance qualitative analysis, and has strong practical significance. HER and Heroin represent heroin, KET and Ketamine represent ketamine, and MDMA represents methylenedioxymethamphetamine.

[0076] In this embodiment, gold nanoparticles AuNPs are prepared by using sodium citrate and chloroauric acid as a SERS enhancement substrate, which is applied to signal enhancement of heroin, ketamine, 3,4-methylenedioxymethamphetamine and other drugs in solution, and SERS spectra of the three drugs are collected. By downsampling method, the number of spectral information samples is expanded. The spectral data set corresponds to the qualitative 3 types and the semi-quantitative 9 types. 70% of the data set of the two types is used as the training data set, and the remaining 30% is used as the test data set. CNN and GRU are introduced into the extraction of Raman spectral features to completely extract the detailed features and spatial correlation in the spectrum. Based on the multi-channel CNN-GRU model, spatial and sequence high-dimensional distribution features can be extracted from the original data, and classified with a high accuracy of 99.9%, and a semi-quantitative accuracy of 99.6%. This method can realize the fuzzy recognition of the three drugs, has stronger resistance to data anomalies, and has stronger application in the actual crime scene detection of drugs. It provides a reference for establishing a rapid detection technology suitable for complex systems of seized drugs and human body drug detection, which is beneficial to the efficiency maximization of the drug detection link in the drug suppression work and promotes the development of drug suppression work.

[0077] Embodiment 2

[0078] The embodiment discloses an electronic device, comprising an input unit and a processor; the input unit is used for receiving surface-enhanced Raman spectrum sequence data of N sampling points of a to-be-tested sample and transmitting to the processor; the processor obtains a drug category in the to-be-tested sample according to the multi-channel drug fuzzy recognition method based on surface-enhanced Raman spectrum provided in embodiment 1, or obtains the drug category and the drug concentration grade of the to-be-tested sample.

[0079] Embodiment 3

[0080] The embodiment discloses a multi-channel drug fuzzy recognition system based on surface-enhanced Raman spectrum, comprising: a to-be-tested sample carrier for carrying a to-be-tested sample, wherein the to-be-tested sample carrier is provided with gold nanoparticles as an enhancement substrate; a laser for outputting laser to the to-be-tested sample; a Raman spectrometer for measuring surface-enhanced Raman spectrum of the to-be-tested sample; and the electronic device provided in embodiment 2, wherein the electronic device is connected with the Raman spectrometer and obtains Raman enhancement spectrum data of different sampling points of the to-be-tested sample from the Raman spectrometer.

[0081] In the description of the specification, reference to "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that a particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the application. The appearances of the above expressions in various places in the specification are not necessarily referring to the same embodiment or example. Moreover, the particular features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0082] Although embodiments of the application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

Claims

1. A multi-channel drug fuzzy identification method based on surface-enhanced Raman spectroscopy, characterized in that: include: Obtaining a surface-enhanced Raman spectrum sequence of N sampling points of the sample to be tested, where N is a positive integer greater than 1; N sampling points are N different areas or locations of the sample to be tested; Inputting N surface-enhanced Raman spectroscopy sequences into N independent channel inputs of a trained CNN-GRU deep learning model respectively, and inputting N surface-enhanced Raman spectroscopy sequences into a centralized channel input of a trained CNN-GRU deep learning model; wherein, the N surface-enhanced Raman spectroscopy sequences are combined into a two-dimensional matrix and input into the centralized channel input of the trained CNN-GRU deep learning model; The CNN-GRU deep learning model outputs the drug category, or the CNN-GRU deep learning model outputs the drug category and drug concentration level; The CNN-GRU deep learning model includes: N single-channel CNN feature extraction modules, each used to extract independent spectral features of N surface-enhanced Raman spectroscopy sequences; Multi-channel feature splicing module, splicing the independent spectral features output by N single-channel CNN feature extraction modules to obtain multi-channel splicing features; Sequence-related feature extraction module, which extracts sequence-related features based on multi-channel splicing features; The centralized channel CNN feature extraction module extracts the centralized spectral features of the N surface-enhanced Raman spectral sequences input; The drug qualitative classification module combines sequence-related features and concentrated spectral features to obtain qualitative classification features, and outputs the drug category contained in the test sample based on the qualitative classification features; The drug quantitative classification module combines sequence-related features and concentrated spectral features to obtain quantitative classification features, and outputs the concentration level of drugs in the test sample based on the quantitative classification features; The sequence-related feature extraction module includes one or more second convolutional pooling units and one or more first GRU layers connected in sequence, and the second convolutional pooling unit includes a second one-dimensional convolutional layer, a second pooling layer and a first batch of normalization layers connected in sequence.

2. The multi-channel fuzzy drug identification method based on surface enhanced Raman spectroscopy according to claim 1, characterized in that: It also includes a switch module, which is used to start the drug quantitative classification module to output the concentration level of drugs in the sample to be tested when the number of drug categories output by the drug qualitative classification module is 1, and to turn off the drug quantitative classification module when the number of drug categories output by the drug qualitative classification module is greater than 1.

3. The multi-channel fuzzy drug identification method based on surface enhanced Raman spectroscopy according to claim 1, characterized in that: The single-channel CNN feature extraction module includes one or more cascaded first convolution pooling units, and the first convolution pooling unit includes a first one-dimensional convolution layer and a first pooling layer connected in series.

4. The multi-channel fuzzy drug identification method based on surface enhanced Raman spectroscopy according to claim 1, 2 or 3, characterized in that: The centralized channel CNN feature extraction module includes one or more cascaded third convolution pooling units and a first two-dimensional convolution layer, and the third convolution pooling unit includes two first two-dimensional convolution layers, a third pooling layer and a second batch normalization layer connected in sequence.

5. The multi-channel fuzzy drug identification method based on surface enhanced Raman spectroscopy according to claim 4, characterized in that: The drug qualitative classification module includes the second GRU layer, the second two-dimensional convolutional layer, the first global pooling layer, the first feature splicing layer, the first fully connected layer, the third batch normalization layer, the third fully connected layer, the fourth fully connected layer and the drug category output layer, which are connected in sequence.

6. The multi-channel fuzzy drug identification method based on surface enhanced Raman spectroscopy according to claim 1, 2, 3 or 5, characterized in that: The drug quantitative classification module includes the third GRU layer, the third two-dimensional convolution layer, the second global pooling layer, the second feature splicing layer, the fifth fully connected layer, the fourth batch normalization layer, the sixth fully connected layer, the seventh fully connected layer and the drug concentration level output layer, which are connected in sequence.

7. An electronic device, characterized in that: including an input unit and a processor; The input unit is used to receive surface enhanced Raman spectroscopy sequence data of N sampling points of the sample to be tested and transmit it to the processor; The processor obtains the drug category in the sample to be tested according to the multi-channel drug fuzzy identification method based on surface enhanced Raman spectroscopy as described in one of claims 1-6, or obtains the drug category and drug concentration level of the sample to be tested.

8. A multi-channel drug fuzzy identification system based on surface-enhanced Raman spectroscopy, characterized in that: include: A sample carrier for testing, used for carrying the sample for testing, wherein gold nanoparticles are provided in the sample carrier for testing as a reinforcing substrate; A laser, which outputs laser light toward the sample to be tested; Raman spectrometer, used to measure the surface enhanced Raman spectrum of the sample to be tested; The electronic device according to claim 7, wherein the electronic device is connected to a Raman spectrometer.

Images