An efficient SERS sensor signal preprocessing method

Abstract

The application discloses a kind of efficient SERS sensor signal preprocessing methods, it includes: constructing the deep learning network model for preprocessing SERS spectrum: STFT-SRNet, the STFT-SRNet of construction is trained, and the predicted data and label data are used in the training process with the Loss calculation of ISTFT transformation;SERS spectrum is preprocessed using the STFT-SRNet of training qualification.And the SERS sensor signal preprocessing method of the application is verified by simulation experiment, can realize the baseline shift correction and noise removal of SERS spectrum, and in the preprocessing effect of SERS spectrum compared with existing CNN, Unet, ResNet and BRN network, STFT-SRNet model in the application is all superior in each evaluation index, prove the effectiveness and advancement of the application.

Description

Technical Field

[0001] This invention relates to the field of sensor technology, and more specifically to a signal processing method for SERS sensors. Background Technology

[0002] Surface-enhanced Raman scattering (SERS) is an analytical technique developed from Raman spectroscopy. SERS technology is widely used in trace substance detection fields such as artifact identification, food / pharmaceutical testing, biological / medical analysis, chemical detection, and medical testing due to its advantages of high sensitivity, short detection time, small sample size, and non-destructive testing. Currently, the spectral data generated by SERS sensors can only be processed by professionals, leading to the application of deep learning in the qualitative and quantitative analysis of SERS spectra. However, due to factors such as experimental environment, Raman spectrometer stability, and fluorescence interference, the spectral data acquired by SERS sensors contains noise and exhibits baseline drift, severely affecting subsequent qualitative and quantitative signal analysis. Therefore, preprocessing of the spectra is essential before analysis.

[0003] Currently, research on SERS spectral data preprocessing can be divided into baseline correction and spectral denoising. Regarding baseline correction, previous studies have proposed a series of correction methods, such as wavelet transform, morphological methods, polynomial fitting, and penalized least squares methods. For spectral denoising, researchers have proposed methods such as Savitzky-Golay filtering, wavelet transform variants, sparse representation, and low-rank estimation. While these traditional methods can achieve both baseline correction and denoising, the results are not ideal or require manual parameter setting, which inevitably introduces subjective factors into the spectral preprocessing. To eliminate the influence of subjective factors, subsequent research has focused on developing automated spectral preprocessing methods. For example, Zhang et al. proposed a Fast Raman Baseline Correction Algorithm (FRBCA) based on automatic linear fitting, which can automatically complete the baseline correction of the spectrum. In addition, Wang et al. also proposed a spectral baseline correction algorithm based on median filtering and non-uniform B-splines, which can effectively perform baseline fitting. However, while automatic algorithms based on traditional methods can automatically complete the baseline correction task of the spectrum, they are limited to simple spectra and their preprocessing effect is not obvious for complex spectra with noise.

[0004] The rapid development of deep learning has brought new solutions to the processing and analysis of SERS spectra, and existing research has begun to use deep learning for qualitative and quantitative studies of SERS spectra. Although some studies have already used deep learning models to perform qualitative analysis of unprocessed SERS spectra, quantitative analysis of SERS spectra still requires preprocessing. Therefore, in order to achieve fully automated analysis of SERS spectra, related research has begun to apply deep learning to SERS preprocessing tasks. Currently, deep learning networks used in the field of SERS spectral preprocessing include CNN, UNet, ResNet, BRN, and various variant hybrid networks. Most of the research using deep learning tends to achieve baseline correction and does not involve spectral denoising; such research has not achieved a complete spectral preprocessing task. Research that has achieved complete spectral preprocessing is still rare. For example, Gebrekidan et al. proposed a spectral preprocessing model based on UNet that can automatically recover Raman spectra from raw spectra. However, the effect on the recovery of complex Raman spectra needs further improvement. Gao et al. proposed a multi-processing Raman spectral preprocessing method that can achieve baseline correction and denoising tasks. However, its core idea is to use two models for correction and denoising separately, which undoubtedly complicates model training. In summary, current work on spectral preprocessing using deep learning still faces challenges such as incomplete spectral preprocessing or unsatisfactory results for complex spectra.

[0005] To further improve the performance of SERS spectral preprocessing, this paper proposes an end-to-end deep learning method, STFT-SRNet, for signal preprocessing of SERS sensors. Specifically, regarding the dataset, a simulated SERS spectral dataset based on real SERS spectra and Gaussian functions was designed and established. For the preprocessing model, the designed STFT-SRNet network is a complex network constructed from Short Time Fourier Transform (STFT), Residual Neural Network (ResNet), and SAMim attention mechanism, enabling feature extraction and preprocessing of SERS spectra in the frequency domain. For model training, both frequency and time domain losses were designed to train the constructed STFT-SRNet. Extensive experiments demonstrate the effectiveness of the proposed STFT-SRNet. Experimental results show that the proposed method can effectively preprocess spectral signals from SERS sensors, exhibiting good performance on both simulated and real data, providing a new approach for promoting the further practical application of intelligent SERS preprocessing algorithms. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by proposing an efficient SERS sensor signal preprocessing method to solve the technical problem of preprocessing the spectral signals of SERS sensors and improving the effect of SERS spectral preprocessing.

[0007] The SERS sensor signal preprocessing method of this invention includes the following steps:

[0008] 1) Construct a deep learning network model for preprocessing SERS spectra: STFT-SRNet. The STFT-SRNet includes an STFT module, a four-layer complex residual network module, a fully connected layer module and an ISTFT module connected in sequence. A SimAM attention module is connected between adjacent complex residual network modules.

[0009] The STFT module is used to transform the input spectral data from the time domain to the frequency domain, that is, one-dimensional spectral data is transformed into a two-dimensional complex tensor after passing through the STFT.

[0010] The four-layer complex residual network module is used to extract features from the data output by the STFT module;

[0011] The SimAM attention module is used to assign different weights to the input of the complex residual network module;

[0012] The fully connected layer module is used to enlarge the size of the output data of the last complex residual network module to be consistent with the input data of the first complex residual network module;

[0013] The ISTFT module is used to transform the data output by the fully connected layer module into SERS spectral data with the same length as the spectral data input to the STFT module.

[0014] 2) Train the constructed STFT-SRNet, and calculate the loss using the predicted data and label data after ISTFT transformation during the training process;

[0015] 3) Preprocess the SERS spectra using a trained STFT-SRNet.

[0016] Furthermore, the efficient SERS sensor signal preprocessing method also includes constructing a dataset for training STFT-SRNet:

[0017] 1) Constructing a standard spectrum to serve as a label, including:

[0018] Gaussian peaks are randomly constructed using Gaussian functions, and a standard spectrum is built from a random number of Gaussian peaks.

[0019] Characteristic Raman peaks are extracted from real SERS spectral data and then combined to form a standard spectrum;

[0020] 2) By randomly generating coordinate points and interpolating the generated coordinate points with cubic spline curves to form a background baseline, the background baseline is added to the constructed standard spectrum, and then randomly generated noise is added to construct a simulated Raman spectrum.

[0021] Furthermore, in step 2), the root mean square error is used as the loss function for calculating the loss:

[0022]

[0023] Where: Y True,i This represents the actual value, i.e., the tag data input to the STFT module; Y Pre,i This represents the predicted value, i.e., the predicted data output by the ISTFT module.

[0024] The beneficial effects of this invention are:

[0025] Simulation experiments verified that the efficient SERS sensor signal preprocessing method of this invention can remove baseline drift and noise from the SERS spectrum. Furthermore, in comparison with existing CNN, Unet, ResNet, and BRN networks, the STFT-SRNet model in this invention outperformed the model in all evaluation metrics, demonstrating the effectiveness and advancement of this invention. Attached Figure Description

[0026] Figure 1 This is a diagram of the STFT-SRNet network model structure.

[0027] Figure 2 A diagram illustrating the process of generating the SERS spectral dataset.

[0028] Figure 3 This paper presents two loss calculation methods during the training of the STFT-SRNet network model. Method A directly calculates the loss between the label and prediction data in the frequency domain; Method B transforms the calculation to the time domain. The evaluation metrics for both methods are calculated in the time domain.

[0029] Figure 4 This is a comparison chart of the preprocessing effects of two loss calculation methods.

[0030] Figure 5 Part (a) shows the training and loss curves of the STFT-SRNet model trained using method B, which gradually converges over 200 training rounds. Part (b) shows the evaluation index change curves of the STFT-SRNet model using method B during training.

[0031] Figure 6This is a diagram showing the preprocessing effect of the STFT-SRNet model on simulation data.

[0032] Figure 7 The figure shows the results of the experimental data preprocessing. In the figure: (a) is the SERS spectrum obtained by measuring R6G, (b) is the SERS spectrum of R6G after preprocessing by the STFT-SRNet model, (c) is the SERS spectrum obtained by measuring sodium nitrate, and (d) is the SERS spectrum of sodium nitrate after preprocessing by the STFT-SRNet model. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] The SERS sensor signal preprocessing method in this embodiment includes the following steps:

[0035] 1) Construct a deep learning network model for preprocessing SERS spectra: STFT-SRNet, such as... Figure 1 As shown, the STFT-SRNet includes an STFT module, a four-layer complex residual network module, a fully connected layer module, and an ISTFT module connected in sequence, with a SimAM attention module connecting adjacent complex residual network modules.

[0036] The STFT module is used to transform the input spectral data from the time domain to the frequency domain; that is, one-dimensional spectral data is transformed into a two-dimensional complex tensor after STFT. The Short-Time Fourier Transform (STFT) is a relatively mature analysis method in signal processing, suitable for spectral analysis of slowly varying time-varying signals. Its transformation process can be summarized as follows: by setting a very narrow time window, a windowed signal is extracted from the signal to be processed, and the Fourier spectrum is calculated. The time-frequency representation of the signal is obtained by sliding the window along the time axis. The STFT representation of a continuous time-domain signal x(t) is as follows:

[0037]

[0038] Where ω represents the angular frequency, ω(t) represents the window function, τ represents the translation position of the window function on the time axis, t represents the change of the continuous signal within each time window, and j represents the imaginary unit.

[0039] In practical applications, continuous signals need to be discretized. Let the discrete signal of the continuous time-domain signal x(t) be x(n), and the window function ω(m) shifts along the time axis with a length of N. Then the discrete form of the STFT is:

[0040]

[0041] Where x(n+m)ω(m) is a short-time sequence.

[0042] In this embodiment, the window function, window width, window overlap length, and FFT point count of the STFT module are Hanning window, 64, 32, and 64, respectively. The input to the STFT in this embodiment is SERS spectral data of length 512. This 512-length spectral data is transformed into a 33*17 complex tensor after passing through the STFT.

[0043] The four-layer complex residual network module is used to extract features from the data output by the STFT module. Since the commonly used residual network (ResNet) is suitable for real tensors, this embodiment employs a complex residual network module to handle complex tensors. The complex network separates the real and imaginary parts of the complex tensor and then merges them back into a complex tensor.

[0044] The SimAM attention module is used to assign different weights to the input of the complex residual network module. Since the preprocessing task using deep learning networks requires spectral reconstruction and output, the entire process is relatively difficult. To achieve high-quality reconstruction of spectral data after feature extraction and to better suit the designed model structure for SERS spectral preprocessing tasks, a SimAM attention module is inserted between the four ResNet layers, specifically as follows: Figure 1 As shown. The attention mechanism guides the model's feature learning by assigning different weights to the network's input, distinguishing the importance of relevant information to ensure the model prioritizes important information and suppresses irrelevant information. Traditional attention mechanisms require introducing additional parameters into the network, increasing network complexity while optimizing performance. SimAM's attention module is an efficient and lightweight 3D attention module. Its advantage lies in inferring 3D attention weights through the feature layer, considering both spatial and channel dimensional correlations simultaneously. Throughout this process, no new parameters need to be introduced into the base model, significantly controlling model complexity. The SimAM model structure is as follows. Figure 1 As shown. The SimAM attention module uses three-dimensional weights and an energy function to accelerate weight calculation. The energy function for each neuron is:

[0045]

[0046] Where k is the target neuron, x i It refers to other neurons in a single channel of the input features, y k y, y o The output value is M, which is the number of neurons in the channel. and It is k and x i The linear transformation will result in the following energy function:

[0047]

[0048] Where, ω k and b k The calculation method is as follows:

[0049]

[0050] Where μ k and To calculate the mean and variance of all neurons except k, λ is a weighting constant. The minimum energy is calculated as follows:

[0051]

[0052] in It represents the minimum energy.

[0053] The fully connected layer module is used to enlarge the size of the output data of the last complex residual network module to match the input data of the first complex residual network module. Since the input and output shapes of the preprocessing task need to be consistent, the overall ResNet network is configured to first gradually reduce the feature size, and then enlarge it to the original input size through the fully connected layer module.

[0054] The ISTFT module transforms the data output from the fully connected layer module into SERS spectral data with the same length as the spectral data input to the STFT module. The output complex tensor is finally transformed by ISTFT into spectral data of length 512, with the parameter settings consistent with the STFT parameters.

[0055] 2) Train the constructed STFT-SRNet. During training, the predicted data and label data after ISTFT transformation are used to calculate the loss. In this embodiment, the root mean square error is used as the loss function to calculate the loss.

[0056]

[0057] Where: Y True,i This represents the actual value, i.e., the tag data input to the STFT module; Y Pre,i This represents the predicted value, i.e., the predicted data output by the ISTFT module.

[0058] 3) Preprocess the SERS spectra using a trained STFT-SRNet.

[0059] The quality and quantity of the dataset used to train a deep learning model are key factors determining its performance. In the field of SERS spectral preprocessing, obtaining a large enough dataset for deep learning is extremely difficult because spectral data requires actual detection. Furthermore, model training requires calculating loss and evaluation metrics using standard SERS spectra, but the process of providing standard data through traditional methods and processing software inevitably involves subjective influences. This embodiment uses a dataset composed of actual SERS spectra and simulated spectra, specifically as follows... Figure 2 As shown, the method for constructing the dataset for training STFT-SRNet is as follows:

[0060] 1) Constructing a standard spectrum to serve as a label, including:

[0061] Gaussian peaks are randomly constructed using Gaussian functions, and a standard spectrum is built from a random number of Gaussian peaks.

[0062] Characteristic Raman peaks are extracted from real SERS spectral data and then combined to form a standard spectrum;

[0063] 2) By randomly generating coordinate points and interpolating the generated coordinate points with cubic spline curves to form a background baseline, the background baseline is added to the constructed standard spectrum, and then randomly generated noise is added to construct a simulated Raman spectrum.

[0064] In this embodiment, the Gaussian peak width, peak height, peak center position, and number of peaks used in the entire construction process are set to 4 to 15, 500 to 2000, 400 to 911, and 1 to 7, respectively. The number of coordinate points, the horizontal coordinate, and the vertical coordinate are set to 2 to 5, 400 to 911, and 1 to 2000, respectively. The signal-to-noise ratio (SNR) is set to 25 to 32.

[0065] The following simulation experiment verifies the performance of the SERS sensor signal preprocessing method in the SERS spectral preprocessing task in this embodiment.

[0066] Model training plays a crucial role in performance. To determine the optimal training method, two training strategies were designed to train STFT-SRNet, and comparative experiments were conducted using simulation data. Figure 3As shown, the comparative experiments involved STFT-SRNet models trained using methods A and B, respectively. Both models were trained using the established SERS spectral dataset, containing 260,000 spectra. The SERS spectral dataset was divided into training, validation, and test sets in a 6:2:2 ratio. The preprocessed model was trained using the training set data, while the validation set was used to monitor for overfitting and correct it. The trained model was then evaluated using the test set data. By comparing evaluation metrics, the optimal network model was obtained through multiple hyperparameter adjustments and training, validation, and testing processes. The optimization method used during training was the Adaptive Moment Estimation (Adam) algorithm. The entire training and testing process utilized the PyTorch 1.9.0 deep learning framework based on Python 3.7, and the computation was performed using an AMD Ryzen 5 3600 6-Core Processor CPU and an NVIDIA GeForce GTX 1650 GPU. Method A calculates the loss in the frequency domain using prediction data without ISTFT transformation and tag data after STFT transformation, while Method B calculates the loss in the time domain using prediction data after ISTFT transformation and tag data. Both methods use the root mean square error (RMSE) as the loss function, and its calculation formula is as follows:

[0067]

[0068] Where Y True,i and Y Pre,i These represent the actual value and the predicted value, respectively. This represents the sample mean. Method B calculates the loss in the time domain and can directly use the formula. Method A, however, calculates the loss in the frequency domain and cannot directly use the formula. Since the data in the frequency domain is complex, the real and imaginary parts are separated during calculation, and the RMSE is calculated separately before merging. The specific formula is as follows:

[0069]

[0070] In this context, the subscripts real and imaginary represent the real and imaginary parts of the complex number, respectively.

[0071] During testing, two models trained with the same parameters were used: epoch = 200, batch size = 64, and learning rate = 1 × 10⁻⁶. -3 To characterize the effect of the model on spectral preprocessing, Ri, a commonly used regression method, was selected. 2The RMSE and MAE indicators are used for evaluation, and the mathematical formulas for these indicators are as follows:

[0072]

[0073] Where Y True,i and Y Pre,i These represent the actual value and the predicted value, respectively. This represents the sample mean.

[0074] The comparative experimental results of the two training strategies were calculated using the selected evaluation metrics, as shown below. Figure 4 As shown, the STFT-SRNet trained using method B outperforms method A in all metrics, especially in R... 2 The most obvious difference is reflected in the metrics. Since the test results were obtained under the same parameters, this experiment shows that calculating the loss in the time domain is more suitable for training the preprocessed model. Through analysis, since the final output of the preprocessed model is a time-domain spectrum and the metric evaluation is calculated in the time domain, Method B can more accurately reflect the difference between the predicted spectrum and the label spectrum during training. Furthermore, compared to Method A's method of calculating the loss in the time domain, Method B can better guide the model convergence and achieve better preprocessing results. Therefore, Method B was ultimately used to train the model, and the changes in the loss value and evaluation metrics were monitored during training. The results are as follows: Figure 5 As shown. According to Figure 5 As shown in (a) of the training and loss curves, the STFT-SRNet model gradually converges within 200 training rounds, and its performance is relatively stable in the later stages of convergence. Figure 5 As shown in (b) of the evaluation index value change curve, the accuracy of the model increases continuously with the increase of the number of training rounds, and the performance improvement is most obvious in the first 40 rounds, and then slowly improves until it stabilizes in subsequent training. The rapid convergence of the model is due to the Adam optimization algorithm used in the model training process, which designs independent adaptive learning rates for different parameters by calculating the first moment estimate and second moment estimate of the gradient during training.

[0075] The simulation data was preprocessed using the trained STFT-SRNet model, and the preprocessing effect is as follows: Figure 6 As shown. Figure 6 As shown in (a), the SERS spectrum containing seven characteristic peaks was preprocessed to remove baseline drift and noise, and all Raman characteristic peaks were accurately recovered. Figure 6 As shown in (b), in the case of overlapping Raman characteristic peaks, the STFT-SRNet model accurately recovers both the overlapping characteristic peaks and the SERS spectrum while performing preprocessing. Figure 6As shown in (c)-(f), the STFT-SRNet model achieves good preprocessing results even with severe baseline drift, noise, and characteristic peak distortion in the SERS spectrum. The results demonstrate that the proposed STFT-SRNet model can perform preprocessing tasks on simulated data with significant preprocessing effects. By learning more important spectral features in the frequency domain, it can effectively recover true characteristic peaks while performing baseline correction and denoising. To further evaluate the performance of the STFT-SRNet model in SERS spectrum preprocessing, CNN, Unet, ResNet, BRN, and STFT-SRNet were used for preprocessing tests on the same dataset. The specific evaluation metrics for each model are shown in Table 1. STFT-SRNet achieved outstanding performance in preprocessing among different network models, achieving high performance in R... 2 Among the metrics, it ranked first with the highest value, while in RMSE and MAE, it had the lowest error and was lower than other comparison networks. The evaluation metrics of STFT-SRNet in the SERS spectral preprocessing task were R... 2 =0.9992, RMSE=0.0022 and MAE=0.0016. Where R... 2 The numerical range of λ is [0,1], and the calculated value of 0.9992 indicates that the STFT-SRNet model has excellent spectral preprocessing performance. The numerical ranges of RMSE and MAE are [0,+∞], and the calculated values ​​of 0.0022 and 0.0016 respectively indicate that the STFT-SRNet model has good accuracy. The results show that the proposed STFT-SRNet network model can be used in SERS spectral preprocessing tasks with excellent performance.

[0076] Table 1. Evaluation metrics for each network model trained on the same dataset

[0077]

[0078] Real data experiments

[0079] To further verify the effectiveness of the STFT-SRNet model in real-world detection, the actual SERS spectra collected by the SESR sensor-based detection system were also tested. The range was selected from Raman frequency shifts of 500-2000 cm⁻¹. -1 The experiments were conducted using R6G and sodium nitrate, which exhibit abundant Raman characteristic peaks within the range. SERS detection was performed using a 50X telephoto lens with a wavelength of 532 nm, a power of 5 mW, an integration time of 10 s, and Raman spectra acquired during the measurement process. The measurement was performed using a HORIBA / Jobin Yvon Raman Spectrometer (Model: LabRAM HR Evolution).

[0080] R6G and sodium nitrate were detected using the SERS sensor and testing system described above, and the resulting SERS spectra are as follows: Figure 7 (a) and Figure 7 As shown in (c), the SERS spectra of the two substances reveal complex baseline drift and noise, necessitating preprocessing of the SERS spectra before analysis. To apply these SERS spectra to a deep learning preprocessing model, their length was set to 512. The experimental spectra of length 512 were preprocessed using the STFT-SRNet model, and the results are shown below. Figure 7 (b) and Figure 7 As shown in (d), the results demonstrate that the STFT-SRNet model effectively achieves baseline shifting and noise reduction in the experimentally measured SERS spectra, and efficiently and accurately recovers key parameters such as the position and shape of characteristic peaks. Experimental results indicate that the proposed STFT-SRNet model can be fully utilized in spectral preprocessing tasks for SERS sensors with excellent performance.

Claims

1. A highly efficient SERS sensor signal preprocessing method, characterized in that: include: 1) Construct a deep learning network model for preprocessing SERS spectra: STFT-SRNet. The STFT-SRNet includes an STFT module, a four-layer complex residual network module, a fully connected layer module and an ISTFT module connected in sequence. A SimAM attention module is connected between adjacent complex residual network modules. The STFT module is used to transform the input spectral data from the time domain to the frequency domain, that is, one-dimensional spectral data is transformed into a two-dimensional complex tensor after passing through the STFT. The four-layer complex residual network module is used to extract features from the data output by the STFT module; The SimAM attention module is used to assign different weights to the input of the complex residual network module; The fully connected layer module is used to enlarge the size of the output data of the last complex residual network module to be consistent with the input data of the first complex residual network module; The ISTFT module is used to transform the data output by the fully connected layer module into SERS spectral data with the same length as the spectral data input to the STFT module. 2) Train the constructed STFT-SRNet, and calculate the loss using the predicted data and label data after ISTFT transformation during the training process; 3) Preprocess the SERS spectra using a trained STFT-SRNet.

2. The efficient SERS sensor signal preprocessing method according to claim 1, characterized in that: It also includes building the dataset for training STFT-SRNet: 1) Constructing a standard spectrum to serve as a label, including: Gaussian peaks are randomly constructed using Gaussian functions, and a standard spectrum is built from a random number of Gaussian peaks. Characteristic Raman peaks are extracted from real SERS spectral data and then combined to form a standard spectrum; 2) By randomly generating coordinate points and interpolating the generated coordinate points with cubic spline curves to form a background baseline, the background baseline is added to the constructed standard spectrum, and then randomly generated noise is added to construct a simulated Raman spectrum.

3. The efficient SERS sensor signal preprocessing method according to claim 1, characterized in that: In step 2), the root mean square error is used as the loss function to calculate the loss: Where: Y True,i This represents the actual value, i.e., the tag data input to the STFT module; Y Pre,i This represents the predicted value, i.e., the predicted data output by the ISTFT module.

Images