Raman spectrum noise reduction method, device and equipment and storage medium

Abstract

The invention provides a Raman spectrum noise reduction method and device, computer equipment and a computer readable storage medium, and the method comprises the steps: obtaining Raman data which comprises a surface-enhanced Raman spectrum set; preprocessing the Raman data: removing the background of the Raman data, and performing spectrum baseline correction and standardization based on an asymmetric least square method to obtain preprocessed Raman data; and performing down-sampling and up-sampling on the preprocessed Raman data, performing feature extraction on the Raman data based on one-dimensional convolution in the down-sampling process to obtain a feature spectrum, and reconstructing the feature spectrum in the up-sampling process to obtain a noise-reduced spectrum.

Description

Technical Field

[0001] This invention relates to the field of optical detection technology, and in particular to a Raman spectroscopy noise reduction method, apparatus, computer equipment, and computer-readable storage medium. Background Technology

[0002] Raman spectroscopy (RS) is a non-destructive, label-free optical detection method based on the principle of inelastic scattering. It obtains information about a sample by analyzing the Raman scattering spectrum generated by the interaction of light and matter. However, only one out of every 10⁸ photons may undergo Raman scattering, meaning the Raman scattering signal is inherently weak, resulting in a weak original signal that is easily overwhelmed by sensor noise, environmental interference, and background fluorescence. Surface-enhanced Raman spectroscopy (SERS) enhances signal intensity through plasmon nanostructures, theoretically enabling single-molecule detection and real-time intraoperative imaging. It has wide applications in food testing, drug screening, and medical diagnosis.

[0003] Surface-enhanced Raman spectroscopy molecular omics (SERSome), a significant breakthrough in this field (pioneered by the team led by Jian Ye at Shanghai Jiao Tong University), integrates nanomaterials, artificial intelligence, and multi-omics methods to achieve high-throughput analysis of metabolic molecules in complex biological samples. Its core principle is to construct a statistically significant dataset by collecting hundreds of spectra from samples, leveraging signal fluctuations caused by the randomness of molecular distribution to improve reliability. However, this technology places higher demands on the efficiency and quality of real-time noise reduction for massive spectra: traditional noise reduction methods struggle to meet its high-throughput processing requirements, and noise can severely interfere with the training accuracy of subsequent artificial intelligence models.

[0004] Currently, mainstream noise reduction algorithms fall into two categories: Firstly, there are moving window methods based on polynomial fitting (such as Savitzky-Golay filters and moving average methods): these methods smooth noise through dynamic fitting within a local window, but require manual adjustment of the window size and polynomial order, making them highly dependent on experience. In scenarios with extremely weak signals (such as single-molecule detection), improper parameter selection can easily lead to the loss of effective signals or residual noise.

[0005] Secondly, wavelet transform methods based on power spectrum estimation (such as WTD) can retain peak characteristics by using wavelet basis functions to decompose the spectrum and threshold filtering noise. However, the selection of wavelet basis, threshold setting and the number of decomposition layers still require manual intervention, and the ability to suppress random noise is limited.

[0006] Although traditional mainstream noise reduction algorithms are fast, their parameter sensitivity and lack of universality make them difficult to cope with complex and varied noise types (such as shot noise, readout noise, and cosmic ray interference), and they cannot meet the real-time processing needs of high-throughput data.

[0007] While deep learning-based methods improve upon traditional mainstream denoising algorithms by automatically learning noise features through end-to-end nonlinear mapping, avoiding the bottleneck of manual parameter tuning, two key drawbacks remain: 1. Supervised learning relies on "clean-noise" data pairs: Most models require training with simulated noise data, but the distribution of virtual noise differs from real noise (such as CCD dark current and ambient light interference), leading to insufficient generalization ability in real-world scenarios. 2. High computational resource consumption: Typical convolutional neural networks (CNNs) have over 128MB of parameters, requiring GPU acceleration for training and inference, making them difficult to deploy on edge detection devices (CPU-only environments, such as microcomputers and handheld devices).

[0008] Therefore, it is necessary to provide a Raman spectroscopy noise reduction method, a Raman spectroscopy noise reduction device, a computer device, and a computer-readable storage medium to effectively solve the above problems. Summary of the Invention

[0009] This invention provides a Raman spectroscopy noise reduction method, a Raman spectroscopy noise reduction device, a computer device, and a computer-readable storage medium.

[0010] This invention provides a Raman spectroscopy noise reduction method, comprising: Acquire Raman data, including a surface-enhanced Raman spectrum set; The Raman data preprocessing includes: removing the Raman data background, performing spectral baseline correction based on asymmetric least squares method, and standardization to obtain preprocessed Raman data; The preprocessed Raman data is downsampled and upsampled. During the downsampling process, feature extraction is performed on the Raman data based on one-dimensional convolution to obtain feature spectra. During the upsampling process, the feature spectra are reconstructed to obtain the denoised spectra.

[0011] Preferably, the noise reduction process for downsampling and upsampling the preprocessed Raman data is based on a self-supervised autoencoder convolutional neural network, which includes an encoder and a decoder. The encoder includes a first module and a second module. The first module and the second module each include a one-dimensional convolutional layer, a GELU activation layer, and a max pooling layer. The one-dimensional convolutional layer, the GELU activation layer, and the max pooling layer perform feature extraction. The features extracted by the first module are passed to the second module. The decoder includes a third module and a fourth module. The third module and the fourth module each include a 3×1 deconvolution processing layer, a GELU activation function, and a 2x upsampling layer. The fourth module also includes a 1x1 convolutional layer with 1 restored channel. The third module and the fourth module are used to reconstruct the spectrum based on the features output by the second module.

[0012] Preferably, the downsampling and upsampling of the preprocessed Raman data during the model training process is based on the Huber loss function for iterative optimization.

[0013] Preferably, the downsampling and upsampling of the preprocessed Raman data are optimized using a stochastic gradient descent algorithm and trained using a learning rate cosine annealing strategy.

[0014] Preferably, during the training process based on the learning rate cosine annealing strategy, convergence is determined by the Huber loss function of two iterations, and training stops when the loss difference reaches a preset value.

[0015] Preferably, the method further includes analyzing the spectrum after noise reduction based on preset evaluation indicators.

[0016] Preferably, the analysis of the spectrum after noise reduction based on preset evaluation indicators includes signal-to-noise ratio analysis.

[0017] A Raman spectroscopy noise reduction device, comprising: Acquisition module: used to acquire Raman data, the Raman data including surface-enhanced Raman spectra; Preprocessing module: used to preprocess the Raman data, including: removing the background of the Raman data, performing spectral baseline correction based on asymmetric least squares method, and standardization to obtain preprocessed Raman data; The noise reduction module is used to downsample and upsample the preprocessed Raman data. During the downsampling process, feature extraction is performed on the Raman data based on one-dimensional convolution to obtain the feature spectrum. During the upsampling process, the feature spectrum is reconstructed to obtain the noise-reduced spectrum.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of any of the methods described above.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0020] Compared with the prior art, the technical solution of the embodiments of the present invention has the following beneficial effects: The Raman spectral noise reduction method, device, computer equipment, and computer-readable storage medium provided in this invention are based on a self-supervised, autoencoder convolutional neural network model. They reduce noise in Raman spectra through downsampling, eliminating the need for extensive training data and manual parameter settings. By using a single Raman spectrum as input, our convolutional neural network model can quickly and effectively eliminate noise in the spectrum. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention, but not all embodiments. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a Raman spectroscopy noise reduction method provided in one embodiment of the present invention; Figure 2 A schematic diagram of a Raman spectroscopy noise reduction model structure provided for one embodiment of the present invention; Figure 3 A schematic diagram of the training steps of a Raman spectroscopy noise reduction method provided in an embodiment of the present invention; Figure 4 A schematic diagram illustrating the effect of the Raman spectroscopy noise reduction method provided in an embodiment of the present invention on the Raman spectrum of SERSome; Figure 5 A schematic diagram of the original spectrum and the spectrum after denoising by a CNN network is provided for one embodiment of the present invention; Figure 6 A schematic diagram of the original spectrum and the spectrum after noise reduction by the U-Net network is provided for one embodiment of the present invention; Figure 7 A schematic diagram of the original spectrum and the spectrum after noise reduction using the WTD algorithm is provided for one embodiment of the present invention; Figure 8 A schematic diagram of the original spectrum and the spectrum after denoising using the SG algorithm is provided for one embodiment of the present invention; Figure 9 This is a schematic diagram of a Raman spectroscopy device provided in one embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0025] Figure 1 A flowchart illustrating a Raman spectroscopy noise reduction method provided in one embodiment of the present invention. Figure 1 As shown, a Raman spectroscopy noise reduction method includes: Step S11: Acquire Raman data, which includes surface-enhanced Raman spectra.

[0026] Specifically, the acquired Raman data is a spectral set based on surface-enhanced Raman spectroscopy molecular omics (SERSome) technology, thus fully demonstrating the applicability of the method of the present invention.

[0027] Step S12: Preprocess the Raman data, including: removing the Raman data background, performing spectral baseline correction based on the asymmetric least squares method, and standardizing to obtain the preprocessed Raman data.

[0028] Specifically, Raman data preprocessing is performed. The acquired surface-enhanced Raman spectroscopy molecular omics data typically contains background noise, such as instrument electronic noise causing spectral peaks to appear lower or higher, and baseline drift due to sample autofluorescence, leading to overall spectral tilt or fluctuations. Preprocessing first involves background removal of the Raman spectral data by subtracting the Raman spectral data without the sample from the sample's Raman spectral data. Secondly, asymmetric least squares (AsLs) is used for baseline correction, and the baseline-removed spectra are standardized to obtain the preprocessed spectral data. Asymmetric least squares achieves its objective by solving a penalized least squares problem. Its objective function is in the form of: Formula (1), in, It is the original signal value. It is the baseline of the fit. λ represents the weight of each data point, which is key to achieving "asymmetry," while λ is a smoothness parameter used to control the smoothness of the fitted baseline. The larger λ is, the smoother the baseline.

[0029] Step S13: Perform downsampling and upsampling on the preprocessed Raman data. During downsampling, feature extraction is performed on the Raman data based on one-dimensional convolution to obtain the feature spectrum. During upsampling, the feature spectrum is reconstructed to obtain the denoised spectrum.

[0030] Specifically, in step S13, data input is performed first. The preprocessed spectral data is directly input into the denoising algorithm, and the data itself is used for effective supervision. No other processing and parameter selection steps are required to denoise a single spectrum.

[0031] Then, noise reduction is performed based on a self-supervised autoencoder convolutional neural network model: the preprocessed Raman spectral data is subjected to downsampling feature extraction and upsampling spectral recovery in the convolutional neural network to obtain the denoised spectrum.

[0032] The self-supervised autoencoder convolutional neural network model uses one-dimensional convolutional and pooling layers to construct network modules, and employs a stride of 2 to avoid loss of spectral information, making it suitable for high-information-density spectral processing. The self-supervised strategy uses only the input spectrum itself to construct supervision, enabling single-spectral noise reduction.

[0033] The Raman spectroscopy noise reduction method provided in this invention is based on a self-supervised autoencoder convolutional neural network model. It reduces noise in Raman spectra through downsampling, without requiring a large amount of training data or manual parameter settings. By using a single Raman spectrum as input, our convolutional neural network model can quickly and effectively eliminate noise in the spectrum.

[0034] Figure 2 This is a schematic diagram of a Raman spectroscopy noise reduction model structure provided for one embodiment of the present invention. Figure 2 As shown, step S13 is based on noise reduction using a self-supervised autoencoder convolutional neural network, which includes an encoder and a decoder.

[0035] The encoder includes a first module and a second module. Both the first and second modules include a one-dimensional convolutional layer, a Gaussian Error Linear Unit (GELU) activation layer, and a max pooling layer. The one-dimensional convolutional layer, GELU activation layer, and max pooling layer are used for feature extraction. The features extracted by the first module are passed to the second module.

[0036] The decoder includes a third module and a fourth module. Each of the third and fourth modules includes a 3×1 deconvolution processing layer, a GELU activation function, and a 2x upsampling layer. The fourth module also includes a 1x1 convolutional layer with one restored channel. The third and fourth modules are used to reconstruct the spectrum based on the features output by the second module.

[0037] Specifically, the downsampling consists of a first module and a second module based on a self-supervised autoencoder convolutional neural network, containing convolutional layers and pooling layers. The first module's convolutional layers have 1×5 kernels and use padding to compress the spectrum to half its original length, followed by a 1×2 max-pooling layer, further compressing the spectrum to 1 / 4 of its original length. The second module's convolutional layers have 1×3 kernels and use padding to compress the spectrum to 1 / 8 of its original length, followed by a 1×2 max-pooling layer, further compressing the spectrum to 1 / 16 of its original length. The GELU activation function is used.

[0038] Upsampling consists of the third and fourth modules of a self-supervised autoencoder-based convolutional neural network model. The input data is compressed spectra, including transposed convolutional layers (deconvolutional processing layers) and upsampling layers. The transposed convolutional layer in the third module has a 1×3 kernel size and uses padding elements to restore the compressed spectrum to 1 / 8 of its original length. This is then passed through a 2x upsampling layer, restoring the compressed spectrum to 1 / 4 of its original length. The fourth module's convolutional layer (deconvolutional processing layer) has a 1×5 kernel size and uses padding elements to restore the compressed spectrum to 1 / 2 of its original length. This is followed by a 2x upsampling layer, restoring the compressed spectrum to its original length. The GELU activation function is used. Finally, a one-dimensional convolutional layer restores the number of channels to 1.

[0039] In convolutional neural networks, the activation function used is the GELU function, whose expression is: Formula (2), Where x represents the training data, and an approximate formula is used in the code implementation: Formula (3), In this embodiment, the autoencoder framework employs a 4-layer encoder-decoder structure. The encoder progressively compresses the feature dimensions using stride=2 convolutions (e.g., input spectrum length 1600 → compressed to spectrum length 100), and the decoder restores the original resolution using transposed convolutions. This autoencoder framework uses only 4 layers, resulting in a simple structure, few model parameters, and suitability for the high-efficiency processing requirements of high-throughput data.

[0040] Furthermore, the autoencoder structure is more lightweight: it uses one-dimensional convolutional kernels, and all convolutional layers use narrow kernels (kernel_size=3 or 5), reducing the number of parameters. The network has only 4 layers (traditional U-Net usually has more than 10 layers), and the maximum number of channels is 128 (traditional CNNs often have 512). The entire model has approximately 60,000 parameters and a model size of approximately 300KB, which can be easily embedded into the microcomputer of portable Raman devices (such as handheld detectors), thereby reducing parameter dependency.

[0041] Figure 3 This is a schematic diagram illustrating the training steps of a Raman spectroscopy noise reduction method provided in an embodiment of the present invention. Figure 3 As shown, in step S13, the model is iteratively optimized based on the Huber loss function during the training process.

[0042] Specifically, during training, the model iteratively optimizes using the Huber Loss function. Huber loss is a regression task loss function that cleverly combines the advantages of mean squared error (MSE) and mean absolute error (MAE). It uses a squared penalty of mean squared error (MSE) when the error is small, and switches to a linear penalty of mean absolute error (MAE) when the error is large. This enhances the model's robustness to outliers while ensuring optimization stability.

[0043] Suppose there are n training data Each training data The actual output is y, and the model... The predicted value is The loss function uses a threshold parameter δ (delta) to divide the "small error" and "large error" regions. Its mathematical expression is a piecewise function, which is defined as follows: Formula (4), The Raman spectroscopy noise reduction method adaptively trains and denoises spectra of different quality. The training process uses the difference in the loss function between two iterations as a convergence criterion. When the loss difference reaches a certain threshold... The model stops training immediately, requiring no pre-training or dataset partitioning, achieving "out-of-the-box" noise reduction for single-spectral input. This embodiment achieves adaptive convergence, eliminating the need for a large number of "clean-noise" spectral pairs and demonstrating strong training generalization ability.

[0044] In some embodiments, step S13 employs a stochastic gradient descent algorithm for optimization and a learning rate-based cosine annealing strategy for training.

[0045] In some embodiments, during the training process based on the learning rate cosine annealing strategy in step S13, convergence is determined according to the Huber loss function of two iterations, and training stops when the loss difference reaches a preset value.

[0046] Specifically, the cosine annealing learning rate strategy dynamically adjusts the learning rate by simulating the periodic changes of a cosine function, helping the model converge more efficiently. This strategy simulates the "annealing" process in metal processing, first heating the metal to a high temperature (high learning rate) and then slowly cooling it (gradually decreasing the learning rate) to make its structure more stable. In deep learning, it adjusts the learning rate through the smooth decreasing characteristics of the cosine function. The learning rate change in cosine annealing is typically described by the following formula: Formula (5), In this embodiment, the initial learning rate is set to 0.05, the minimum learning rate is set to 0.005, and a stochastic gradient descent (SGD) optimizer is used, which periodically decays the learning rate according to a cosine function. Corresponding to the formula above, , . This indicates how many rounds of training have been completed so far. This represents the total number of training rounds. Utilizing the periodic learning rate variation of the cosine annealing strategy helps the model escape local optima, making it easier to approach the global optimum for denoising performance and improving convergence stability. Simultaneously, compared to a fixed learning rate, training time is reduced by 55%, adapting to the needs of rapid optimization in a single spectrum. In this embodiment, the cosine annealing strategy improves both denoising performance and efficiency.

[0047] By comparing different learning strategies, it can be found that the cosine annealing strategy has the highest signal-to-noise ratio improvement among the three, and also the best noise reduction effect. Compared with cosine annealing and fixed learning rate strategies, the noise reduction time is improved by nearly 220% for the same noise reduction effect.

[0048] As shown in Table 1, experiments have demonstrated that the cosine annealing strategy plays a significant role in spectral noise reduction, not only greatly shortening the noise reduction time but also achieving excellent noise reduction results.

[0049] Table 1

[0050] In some embodiments, a general one-dimensional spectral data denoising loss function is used for the design of the loss function. Indicating that the input Raman spectrum is an n-dimensional real vector, we first define the regression loss between the original spectrum and the denoised spectrum. Mean squared error (MSE, ||∙||2) is used to ensure important peak values: Formula (6), We use SGD as the gradient optimization method for the denoising network and employ cosine annealing. The initial learning rate is 0.05, and after 350 rounds of cosine decay, the learning rate decreases to 0.005, remaining unchanged in subsequent rounds. To adapt to spectra of varying quality, when the loss function changes by less than 5 × 10⁻⁶, [the following parameters are used]. -5 At that point, model training stops.

[0051] In some implementations, the Raman spectroscopy noise reduction method further includes analyzing the denoised spectrum based on preset evaluation indicators. Preferably, this includes signal-to-noise ratio analysis.

[0052] Specifically, multiple evaluation indicators are used to conduct qualitative and quantitative analysis of the noise reduction method results.

[0053] Firstly, the signal-to-noise ratio (SNR): For a characteristic peak in the reference spectrum, five wavenumbers near that peak are selected as the range for calculating the average peak intensity. The formula for calculating the average peak intensity is as follows: Formula (7), in, It is the signal intensity corresponding to the i-th Raman shift within the signal range, and N=5 is the number of Raman shifts contained within the signal range.

[0054] The frequency band without characteristic peaks in the spectrum is selected as the range of the noise standard deviation. The formula for calculating the noise standard deviation is: Formula (8), in, This represents the intensity corresponding to the j-th Raman shift within the noise range. M represents the average intensity within the noise range, and M=50 represents the number of Raman shifts contained within the noise range.

[0055] The formula for calculating the signal-to-noise ratio of the characteristic peak is as follows: Formula (9), If the signal-to-noise ratio of the obtained characteristic peak is greater than 3, then a confidence level of 99% indicates that there is a characteristic peak at that location that is different from random noise.

[0056] Secondly, noise reduction speed: Considering that the computer configuration of edge instruments such as detection systems is not high, this speed measurement will use a CPU i5-12400. The running speed, model size and signal-to-noise ratio (SNR) of different algorithms under the same conditions will be calculated multiple times on this device. This invention utilizes a self-supervised autoencoder-based convolutional neural network model to reduce noise in Raman spectra through downsampling, without requiring extensive training data or manual parameter settings. By using a single Raman spectrum as input, the CNN denoising network can quickly and effectively eliminate noise from the spectrum.

[0057] The performance, generalization ability, and robustness of the proposed general denoising algorithm were verified using SERSome spectroscopy in the 1650–1750 cm⁻¹ range. -1 The Raman shift at 1700 cm⁻¹ indicates the signal of a certain molecule. The spectral noise is significantly compressed after noise reduction, at 1700 cm⁻¹. -1 The characteristic peaks are more clearly identified, the noise around the characteristic peaks becomes smoother, and the signal-to-noise ratio is improved from 9.83 to 23.35, an improvement of about 237%.

[0058] Next, we recorded the model parameter count and performance of the self-supervised convolutional autoencoder (CNN), U-Net, SG, and WTD algorithms. To ensure accurate comparisons, we ran the denoising algorithm five times on each of the five spectra and calculated the average time. We also recorded the SNR improvement for each spectrum and calculated the average SNR improvement. In this experiment, we controlled the experimental environment for each algorithm. Before running the SG and WTD denoising algorithms, the data was standardized to ensure consistency with the CNN and U-Net algorithms. The learning rate for CNN and U-Net was also kept consistent at 0.05. Figure 5-8 The diagram shows the original spectrum and the spectral effects of four denoising methods. Table 2 shows the evaluation metrics of the four denoising methods on the SERSome Raman spectrum.

[0059] Table 2

[0060] Experiments show that, under the conditions of minimizing time and maximizing SNR, the self-supervised convolutional autoencoder denoising network of this invention is the optimal model. In the reality of an average detection time of 1 second, it can achieve the effect of simultaneous detection and denoising without waiting.

[0061] like Figure 9 As shown, the present invention also provides a Raman spectroscopy noise reduction device 10, comprising: an acquisition module 11 for acquiring Raman data, wherein the Raman data includes a surface-enhanced Raman spectrum set; a preprocessing module 12 for preprocessing the Raman data, including: removing the Raman data background, performing spectral baseline correction based on asymmetric least squares method, and standardization to obtain preprocessed Raman data; and a noise reduction module 13 for downsampling and upsampling the preprocessed Raman data, wherein during downsampling, feature extraction of the Raman data is performed based on one-dimensional convolution to obtain a feature spectrum, and during upsampling, the feature spectrum is reconstructed to obtain a noise-reduced spectrum.

[0062] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any of the above embodiments.

[0063] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the methods of any of the above embodiments.

[0064] The Raman spectroscopy noise reduction device 10, computer equipment, and computer-readable storage medium provided by the present invention can implement the steps in any of the above embodiments, and will not be repeated here.

[0065] The Raman spectroscopy noise reduction method, Raman spectroscopy noise reduction device, computer equipment, and computer-readable storage medium provided by this invention can achieve the following effects: Adaptive denoising eliminates the need for manual parameter tuning: Traditional methods (such as SG filtering and wavelet transform) require manual adjustment of window size, polynomial order, or threshold parameters, which can easily lead to loss of effective signal or noise residue in ultra-weak signal scenarios (such as single-molecule detection). This invention achieves end-to-end automatic learning of noise features through a self-supervised convolutional autoencoder network (CAENN), completely avoiding manual intervention. Experiments show that in processing the 1700 cm⁻¹ feature peak of the SERSome spectrum, the signal-to-noise ratio (SNR) is improved from 9.83 to 23.35 (an improvement of approximately 237%), significantly outperforming SG filtering (186.68% SNR improvement) and wavelet threshold denoising (175% SNR improvement). It also addresses the dependence of supervised learning on simulation data: Existing deep learning models require a large amount of "clean-noise" data pairs for training, but the difference in the distribution of simulated noise and real noise (such as CCD dark current and ambient light interference) leads to poor generalization. This invention employs a self-supervised mechanism, requiring only a single input spectrum to construct a supervisory signal. The noise distribution is directly learned through an encoder-decoder framework, eliminating the need for paired label data. For example, in the SERSome spectral set of complex biological samples, the model successfully handled instrument electronic noise and fluorescence baseline drift, and adapted to input spectra of varying quality.

[0066] Lightweight Design and Edge Deployment Capabilities: Extremely Low Parameter and Computational Resource Requirements: The network employs a 4-layer one-dimensional convolutional structure (2 encoder layers + 2 decoder layers), with narrow convolutional kernel size (1×3 or 1×5), a maximum of only 128 channels, and a model parameter count of only 60,033 (approximately 300 KB), a reduction of 87.2% compared to the typical U-Net (469,157 parameters). In a CPU-only environment (Intel i5-12400), the average processing time for a single spectrum is only 0.357 seconds, meeting the high-throughput requirement of 1 frame per second for SERSome technology, and can be directly embedded into handheld Raman devices. Dynamic Convergence Strategy Improves Efficiency: By monitoring the difference in the loss function (HuberLoss) over continuous iterations, when the difference is <5×10⁻ 5 Training is automatically terminated when the time is right, avoiding redundant calculations in fixed rounds. Combined with a cosine annealing learning rate strategy (initial 0.05 → minimum 0.005), training time is reduced by nearly 55%, achieving real-time optimization of the single spectrum "out of the box".

[0067] Feature Preservation and Robustness Advantages: High-fidelity recovery of key peak signals: The encoder compresses the spectral length (1600→100) through stride=2 convolution, preserving low-frequency feature peak information; the decoder restores the original resolution through transposed convolution and suppresses negative noise by combining the GELU activation function. In the 1650-1750 cm⁻¹ Raman shift range, the model effectively compresses background noise, and the feature peak shape is smooth with no peak shift. Strong Adaptability to Complex Noise Scenes: The self-supervised framework handles mixed noise types through multi-scale feature fusion (such as local noise and global baseline drift). On the SERSome dataset containing shot noise and cosmic ray interference, the average SNR is improved by 237%, outperforming other algorithms.

[0068] This invention addresses three core issues in Raman spectroscopy denoising: parameter dependence, poor generalization, and computational bottlenecks, through a lightweight autoencoder architecture, a self-supervised training mechanism, and a dynamic convergence strategy. Experiments demonstrate its significant advantages in SNR improvement (average 237%), processing speed (0.357 seconds / spectrum), and edge deployment (300 KB model), providing a reliable solution for high-throughput resolution of SERSome technology and real-time processing of portable Raman devices.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A Raman spectroscopy noise reduction method, characterized in that, include: Acquire Raman data, including a surface-enhanced Raman spectrum set; The Raman data preprocessing includes: removing the Raman data background, performing spectral baseline correction based on asymmetric least squares method, and standardization to obtain preprocessed Raman data; The preprocessed Raman data is downsampled and upsampled. During the downsampling process, feature extraction is performed on the Raman data based on one-dimensional convolution to obtain feature spectra. During the upsampling process, the feature spectra are reconstructed to obtain the denoised spectra.

2. The Raman spectroscopy noise reduction method according to claim 1, characterized in that, The noise reduction process for downsampling and upsampling the preprocessed Raman data is based on a self-supervised autoencoder convolutional neural network, which includes an encoder and a decoder. The encoder includes a first module and a second module. The first module and the second module each include a one-dimensional convolutional layer, a GELU activation layer, and a max pooling layer. The one-dimensional convolutional layer, the GELU activation layer, and the max pooling layer perform feature extraction. The features extracted by the first module are passed to the second module. The decoder includes a third module and a fourth module. The third module and the fourth module each include a 3×1 deconvolution processing layer, a GELU activation function, and a 2x upsampling layer. The fourth module also includes a 1x1 convolutional layer with 1 restored channel. The third module and the fourth module are used to reconstruct the spectrum based on the features output by the second module.

3. The Raman spectroscopy noise reduction method according to claim 1, characterized in that, The downsampling and upsampling of the preprocessed Raman data are performed during the model training process, which is then iteratively optimized based on the Huber loss function.

4. The Raman spectroscopy noise reduction method according to claim 3, characterized in that, The downsampling and upsampling of the preprocessed Raman data are optimized using the stochastic gradient descent algorithm and trained using a learning rate cosine annealing strategy.

5. The Raman spectroscopy noise reduction method according to claim 4, characterized in that, During the training process based on the learning rate cosine annealing strategy, convergence is determined by the Huber loss function of two iterations, and training stops when the difference in loss reaches a preset value.

6. The Raman spectroscopy noise reduction method according to claim 1, characterized in that, It also includes analyzing the spectrum after noise reduction based on preset evaluation indicators.

7. The Raman spectroscopy noise reduction method according to claim 6, characterized in that, The analysis of the spectrum after noise reduction based on preset evaluation indicators includes signal-to-noise ratio analysis.

8. A Raman spectroscopy noise reduction device, characterized in that, include: Acquisition module: used to acquire Raman data, the Raman data including surface-enhanced Raman spectra; Preprocessing module: used to preprocess the Raman data, including: removing the background of the Raman data, performing spectral baseline correction based on asymmetric least squares method, and standardization to obtain preprocessed Raman data; The noise reduction module is used to downsample and upsample the preprocessed Raman data. During the downsampling process, feature extraction is performed on the Raman data based on one-dimensional convolution to obtain the feature spectrum. During the upsampling process, the feature spectrum is reconstructed to obtain the noise-reduced spectrum.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images