A method for automatic adjustment and reconstruction of Raman spectra based on neural network

Abstract

The present invention discloses a method for automatic adjustment and reconstruction of Raman spectrum based on neural network, which is applied to Raman spectrum system. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network includes: collecting and obtaining a first Raman spectrum signal according to a preset first temperature, a preset first integration time, a preset first gain and a preset first bias. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network automatically adjusts the integration time, bias and gain through an automatic adjustment model without manual operation, effectively improves the dynamic range and signal-to-noise ratio of the Raman spectrum system, makes weak Raman spectrum signals easier to identify, enhances detection sensitivity, and also improves the dynamic range of detection; and by weighting the width module and the depth module in the automatic adjustment model, the accuracy and flexibility of the output of the automatic adjustment model are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of Raman spectroscopy, and in particular relates to a Raman spectroscopy automatic adjustment and reconstruction method based on a neural network. Background Art

[0002] Raman spectroscopy is an analytical method that obtains molecular structure information by measuring the scattering spectrum generated by the interaction between matter and laser, and is widely used in materials science, chemical analysis, biomedicine and other fields. However, since Raman scattering signals are inherently weak and susceptible to noise interference, the dynamic range and signal-to-noise ratio of the system become key factors affecting the quality of Raman spectroscopy.

[0003] In the Raman spectroscopy system, CCD is used as a detector to convert optical signals into electrical signals, and outputs digital signals through analog-to-digital conversion chips (such as AD9826). However, the strength of the Raman spectroscopy signal depends largely on the characteristics of the sample, and the signal intensity will fluctuate greatly under different conditions (such as changes in integration time). In the case of long-term integration, strong Raman signals may cause CCD overexposure, exceeding the upper limit of the detector range (such as 65535), thereby reducing the spectral quality and data accuracy. When the integration time is short, although overexposure of strong signals can be avoided, weak Raman signals may not be detected, resulting in a lack of necessary information in the spectrum.

[0004] The existing technology usually optimizes the signal-to-noise ratio and dynamic range by manually adjusting the gain, offset and integration time, but this manual adjustment method is not only time-consuming, but also difficult to meet the requirements under real-time detection and dynamic changes. Therefore, there is a demand to achieve dynamic adjustment of gain, offset and integration time through an automated control system to adapt to signal changes under different detection conditions, thereby improving system performance. Summary of the invention

[0005] The purpose of the present invention is to solve the problems raised in the background technology and to propose a method for automatic adjustment and reconstruction of Raman spectrum based on neural network.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] The present invention proposes a method for automatic adjustment and reconstruction of Raman spectrum based on a neural network, which is applied to a Raman spectrum system. The method for automatic adjustment and reconstruction of Raman spectrum based on a neural network includes:

[0008] Acquire a first Raman spectrum signal according to a preset first temperature, a preset first integration time, a preset first gain, and a preset first bias;

[0009] The first integration time, the first temperature and the first Raman spectrum signal are used as input features and input into a trained automatic adjustment model, and the automatic adjustment model outputs a second integration time, a second bias and a second gain that reduce the background noise to a preset position;

[0010] According to the second integration time, the second bias and the second gain, a corresponding second Raman spectrum signal is collected;

[0011] Determine whether there is peak overexposure in the second Raman spectrum signal, and whether the second integration time is less than a preset first threshold;

[0012] When there is an overexposed peak in the second Raman spectrum signal and the second integration time is less than a preset first threshold, the second integration time is reduced so that the overexposed peak is not overexposed, and the non-overexposed peak is extracted;

[0013] The extracted signal is replaced with a signal within a preset range around an overexposed peak value in the second Raman spectrum signal before the second integration time is reduced to form a new Raman spectrum signal.

[0014] Preferably, the first temperature is the operating temperature of a spectrometer in the Raman spectroscopy system, and the first temperature is a preset second threshold value or is obtained by rounding up the current operating temperature of the spectrometer.

[0015] Preferably, after the automatic adjustment model outputs the second integration time, the second bias and the second gain, the spectrometer in the Raman spectroscopy system is configured to correspond to the second integration time, the second bias and the second gain, and then the second Raman spectroscopy signal is acquired.

[0016] Preferably, the automatic adjustment model includes an adaptive weight module, a width module, a depth module and a linear layer;

[0017] The input features of the automatic adjustment model are passed through the adaptive weight module to obtain weighted input features. The weighted input features are passed through the width module and the depth module in parallel to obtain their respective outputs. The output of the width module and the output of the depth module are fused, and then passed through the linear layer output to obtain the second integration time, second bias and second gain that reduce the background noise to a preset position.

[0018] Preferably, the input features of the automatic adjustment model It is expressed as:

[0019] ;

[0020] in, represents the first integration time, represents the first temperature, represents the first Raman spectrum signal, Indicates transpose.

[0021] Preferably, the adaptive weight module comprises a fully connected layer and an activation layer sequentially connected from data input to output, the input feature passes through the fully connected layer to obtain the original weight of the input feature, the original weight of the input feature passes through the activation layer to obtain the dynamic weight, and the input feature is weighted according to the dynamic weight to obtain the weighted input feature;

[0022] The calculation formula of the activation layer is as follows:

[0023] ;

[0024] in, represents the dynamic weight, represents the original weight of the input feature, Indicates the number of each vector in the input feature;

[0025] Weighted input features It is expressed as:

[0026] ;

[0027] in, represents the dynamic weight of the first integration time, represents the dynamic weight of the first temperature, represents the dynamic weight of the first Raman spectrum signal.

[0028] Preferably, the width module performs linear mapping on the weighted input features, and the calculation formula is as follows:

[0029] ;

[0030] in, represents the output of the width module, represents the weight matrix of the width module, represents the bias vector of the width module, Represents the weighted input features.

[0031] Preferably, the depth module includes two parallel embedding layers and a hidden layer, the weighted input features pass through the two embedding layers in parallel, the outputs of the two embedding layers pass through the hidden layer to obtain the output of the depth module, the hidden layer includes a preset number of activation layers arranged in sequence, and the calculation formula of each layer in the hidden layer is as follows:

[0032] ;

[0033] in, represents the output of each layer in the hidden layer, represents the input of each layer in the hidden layer, ReLU activation function calculation, represents the bias vector of each layer in the hidden layer, Represents the weight matrix of each layer in the hidden layer.

[0034] Preferably, the output of the width module and the output of the depth module are fused, wherein the calculation formula for the fusion is as follows:

[0035] ;

[0036] in,

[0037] ;

[0038] ;

[0039] in, It represents the features after the output of the width module and the output of the depth module are fused. represents the output of the width module, represents the output of the depth module, Represents the weight matrix of the fusion process for the width module, Represents the weight matrix of the fusion process for the depth module, Represents the Sigmoid activation function calculation, represents the weighted input features, Indicates transpose.

[0040] Preferably, the output of the width module and the output of the depth module are fused to obtain the features, and then output through a linear layer to obtain a second integration time, a second bias, and a second gain that reduce the background noise to a preset position, wherein the calculation formula of the linear layer is as follows:

[0041] ;

[0042] in,

[0043] Y output = [ x 1 ', Bias , Gain ] T ;

[0044] in, represents the output of the linear layer, represents the weight matrix of the linear layer, The linear layer represents the bias vector, represents the second integration time, represents the second bias, represents the second gain, Indicates transpose.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This neural network-based Raman spectrum automatic adjustment and reconstruction method automatically adjusts the integration time, bias and gain through the automatic adjustment model without manual operation, effectively improving the dynamic range and signal-to-noise ratio of the Raman spectrum system, making weak Raman spectrum signals easier to identify, enhancing detection sensitivity, and improving the dynamic range of detection; and by weighting the width module and the depth module in the automatic adjustment model, the accuracy and flexibility of the automatic adjustment model output are improved;

[0047] By controlling the first temperature, the influence of temperature fluctuation on the Raman spectrum signal is significantly reduced, thereby improving the stability of the data and ensuring that the Raman spectrum can obtain accurate results under different environmental conditions;

[0048] This method achieves optimized detection of strong and weak signals by merging and reconstructing the strong signal peak data and the weak signal data, which not only avoids the overexposure problem of strong signals, but also retains the complete information of weak signals, thereby improving the overall quality of Raman spectroscopy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the method for automatic adjustment and reconstruction of Raman spectrum based on neural network of the present invention;

[0050] Figure 2 A schematic diagram of the Raman spectrum of 95% ethanol before adjustment by the automatic adjustment model of the present invention;

[0051] Figure 3 A schematic diagram of the Raman spectrum of 95% ethanol after adjustment by the automatic adjustment model of the present invention;

[0052] Figure 4 It is a structural schematic diagram of the automatic adjustment model of the present invention;

[0053] Figure 5 It is a structural schematic diagram of the depth module of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] like Figure 1-Figure 5 As shown, a method for automatic adjustment and reconstruction of Raman spectrum based on neural network is provided, which is applied to Raman spectrum system. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network includes:

[0056] Step 1: acquiring a first Raman spectrum signal according to a preset first temperature, a preset first integration time, a preset first gain, and a preset first bias;

[0057] The first temperature is the operating temperature of a spectrometer in the Raman spectroscopy system, and the first temperature is a preset second threshold or is obtained by rounding up the current operating temperature of the spectrometer.

[0058] It should be noted that the Raman spectroscopy system includes a laser light source, a Raman probe and a spectrometer. The temperatures mentioned in this application are all the working temperatures of the spectrometer, and specifically the working temperatures of the CCD image sensor in the spectrometer. The Raman spectroscopy system is started according to the preset second threshold temperature (such as the second threshold temperature is 15°C). If it is found that the temperature cannot be stabilized within the preset range within the preset number of seconds, the current working temperature of the spectrometer is rounded up as the first temperature. If the temperature can be stabilized within the preset number of seconds, the second threshold temperature is used as the first temperature. For example, if it cannot be stabilized to 15±0.3°C within 5 seconds, the first temperature is set to the current temperature rounded up, such as 16.8°C is set to 17°C. If it can be stabilized to 15±0.3°C within 5 seconds, the first temperature is 15°C.

[0059] Wherein, if the first integration time is 1s, the first gain is 2×, and the first bias is -209mv, the first Raman spectrum signal (i.e., the ordinate is the light intensity of Raman scattering, and the abscissa is the Raman shift or the Raman spectrum of the wave number) is acquired according to these preset parameters. The specific process of acquiring the Raman spectrum signal belongs to the prior art well known in the art and is not described in this application.

[0060] Step 2: Take the first integration time, the first temperature and the first Raman spectrum signal as input features, input them into the trained automatic adjustment model, and automatically adjust the model output to reduce the background noise to the preset position (not limited, set according to actual needs) of the second integration time, the second bias and the second gain. Figure 2 and Figure 3 All are Raman spectra of 95% ethanol (the horizontal axis is Raman shift, the unit is , the vertical axis is the light intensity, the unit is ), Figure 2 After the Raman spectrum of the model is automatically adjusted, we can get Figure 3 Raman spectrum of .

[0061] It should be noted that background noise refers to interference components other than Raman spectral signals. These interference components include noise introduced by system hardware characteristics, environmental factors or sample background signals, which will directly affect the quality and accuracy of Raman spectral signals. The method of obtaining background noise is to select an area without obvious Raman spectral signals in the Raman spectrum (such as Figure 2The light intensity in these areas is used as an approximation of the background noise.

[0062] Among them, Figure 4 As shown, the automatic adjustment model includes an adaptive weight module, a width module, a depth module, and a linear layer;

[0063] The input features of the automatic adjustment model are passed through the adaptive weight module to obtain weighted input features. The weighted input features are passed through the width module and the depth module in parallel to obtain their respective outputs. The output of the width module and the output of the depth module are fused, and then passed through the linear layer output to obtain the second integration time, second bias and second gain that reduce the background noise to a preset position.

[0064] Among them, the input features of the automatic adjustment model It is expressed as:

[0065] X= [ x 1 , x 2 , x 3 ] T ;

[0066] in, represents the first integration time, represents the first temperature, represents the first Raman spectrum signal, Indicates transpose.

[0067] Among them, the adaptive weight module includes a fully connected layer and an activation layer (using the Softmax activation function) connected in sequence from data input to output. After the input feature passes through the fully connected layer, the original weight of the input feature is obtained. After the original weight of the input feature passes through the activation layer, the dynamic weight is obtained. The input feature is weighted according to the dynamic weight to obtain the weighted input feature;

[0068] The calculation formula of the activation layer is as follows:

[0069] ;

[0070] in, represents the dynamic weight, represents the original weight of the input feature (the original weight calculated by the fully connected layer), Indicates the number of each vector in the input feature;

[0071] Weighted input features It is expressed as:

[0072] X ' =[ ω 1 ⋅ x 1 , ω 2 ⋅ x 2 , ω 3 ⋅ x 3 ] T ;

[0073] in, represents the dynamic weight of the first integration time, represents the dynamic weight of the first temperature, Represents the dynamic weight of the first Raman spectrum signal; the weight of each vector in each input feature is adjusted through the dynamic weight to ensure that important features can participate in the calculation more effectively under different detection environments; for example, when the temperature fluctuates greatly, the importance of temperature increases; when the integration time is extremely long or extremely short, the importance of integration time increases.

[0074] Among them, the width module performs linear mapping on the weighted input features, and the calculation formula is as follows:

[0075] ;

[0076] in, represents the output of the width module, represents the weight matrix of the width module, represents the bias vector of the width module, Represents the weighted input features. The function of the width module (simple mapping) is to learn the simple mapping relationship between the input features (first integration time, first temperature, first Raman spectrum signal) and the target output (second integration time, second bias and second gain).

[0077] Among them, Figure 5 As shown, the depth module includes two parallel embedding layers and a hidden layer. The weighted input features pass through the two embedding layers in parallel (the main function of the embedding layer is to map the weighted input features (such as category features or discrete features) to a dense vector space, so that the relationship between the features can be better modeled by the subsequent network layer). The outputs of the two embedding layers pass through the hidden layer to obtain the output of the depth module. The hidden layer includes a preset number of activation layers that are sequentially set (three activation layers are used in this embodiment, and all are ReLU activation functions. The input of the first activation layer from the data input to the output direction is 64 dimensions, and the inputs of the second activation layer and the third activation layer are both 32 dimensions), and the calculation formulas of each layer in the hidden layer are as follows:

[0078] ;

[0079] in, represents the output of each layer in the hidden layer, represents the input of each layer in the hidden layer, ReLU activation function calculation, represents the bias vector of each layer in the hidden layer, Represents the weight matrix of each layer in the hidden layer.

[0080] The output of the width module and the output of the depth module are fused, and the calculation formula for the fusion is as follows:

[0081] ;

[0082] in,

[0083] ;

[0084] ;

[0085] in, It represents the features after the output of the width module and the output of the depth module are fused. represents the output of the width module, represents the output of the depth module, Represents the weight matrix of the fusion process for the width module, Represents the weight matrix of the fusion process for the depth module, Represents the Sigmoid activation function calculation, represents the weighted input features, Denotes transposition. The function of the deep module is to extract the nonlinear relationship and interaction pattern between weighted input features through a multi-layer neural network. It can directly process the high-dimensional data of Raman spectra and learn the complex characteristic relationship between peaks and backgrounds in the spectrum. At the same time, it accepts the input of original numerical features such as integration time and temperature to further enrich the feature extraction capability of input data.

[0086] Among them, the output of the width module and the output of the depth module are fused, and then the linear layer output is used to obtain the second integration time, the second bias and the second gain that reduce the background noise to a preset position. The calculation formula of the linear layer is as follows:

[0087] ;

[0088] in,

[0089] Y output = [ x 1 ', Bias , Gain ] T ;

[0090] in, represents the output of the linear layer, represents the weight matrix of the linear layer, The linear layer represents the bias vector, represents the second integration time, represents the second bias, represents the second gain, Indicates transpose.

[0091] Step 3: acquiring a corresponding second Raman spectrum signal according to a second integration time, a second bias and a second gain;

[0092] After the automatic adjustment model outputs the second integration time, the second bias and the second gain, the spectrometer in the Raman spectroscopy system is configured to correspond to the second integration time, the second bias and the second gain (the specific configuration process belongs to the prior art well known in the art and is not described in detail in this solution), and then the second Raman spectroscopy signal is collected.

[0093] Step 4: determine whether there is peak overexposure in the second Raman spectrum signal (in Raman spectrum intensity measurement, the intensity of one or some peaks exceeds the range or saturation threshold of the detector, resulting in signal distortion or inability to accurately measure), and whether the second integration time is less than a preset first threshold (based on experience, the first threshold is 300ms);

[0094] Step 5: When there is an overexposed peak in the second Raman spectrum signal and the second integration time is less than the preset first threshold, the second integration time is reduced so that the overexposed peak is not overexposed, and a signal within a preset range around the non-overexposed peak (i.e., the corresponding peak after the second integration time is reduced so that the overexposed peak is not overexposed) is extracted;

[0095] In this embodiment, the signal of 50 wave numbers around the peak value that is not overexposed is extracted.

[0096] Step 6, replacing the extracted signal with the signal within a preset range of the overexposed peak in the second Raman spectrum signal before reducing the second integration time to form a new Raman spectrum signal; that is, extracting a signal of 50 wavenumbers around the peak that is not overexposed, and replacing it with the signal of 50 wavenumbers around the overexposed peak in the second Raman spectrum signal before reducing the second integration time to form a new Raman spectrum signal, thereby realizing reconstruction of the Raman spectrum signal. By merging and reconstructing the strong signal peak data under a short integration time with the weak signal data under a long integration time, optimized detection of strong and weak signals is realized, which not only avoids the overexposure problem of strong signals, but also retains the complete information of weak signals, thereby improving the overall quality of the Raman spectrum.

[0097] The neural network-based Raman spectrum automatic adjustment and reconstruction method automatically adjusts the integration time, bias and gain through the automatic adjustment model without manual operation, effectively improving the dynamic range and signal-to-noise ratio of the Raman spectrum system, making weak Raman spectrum signals easier to identify, enhancing the detection sensitivity, and improving the dynamic range of detection; and by weighting the width module and the depth module in the automatic adjustment model, the accuracy and flexibility of the automatic adjustment model output are improved; by controlling the first temperature, the influence of temperature fluctuations on the Raman spectrum signal is significantly reduced, thereby improving the stability of the data and ensuring that the Raman spectrum can obtain accurate results under different environmental conditions; this method realizes the optimized detection of strong and weak signals by merging and reconstructing the strong signal peak data and the weak signal data, which not only avoids the overexposure problem of the strong signal, but also retains the complete information of the weak signal, thereby improving the overall quality of the Raman spectrum.

[0098] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0099] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for automatic adjustment and reconstruction of Raman spectrum based on neural network, characterized in that: The neural network-based Raman spectrum automatic adjustment and reconstruction method is applied to a Raman spectrum system, and the neural network-based Raman spectrum automatic adjustment and reconstruction method includes: Acquire a first Raman spectrum signal according to a preset first temperature, a preset first integration time, a preset first gain, and a preset first bias; The first integration time, the first temperature and the first Raman spectrum signal are used as input features and input into a trained automatic adjustment model, and the automatic adjustment model outputs a second integration time, a second bias and a second gain that reduce the background noise to a preset position; According to the second integration time, the second bias and the second gain, a corresponding second Raman spectrum signal is collected; Determine whether there is peak overexposure in the second Raman spectrum signal, and whether the second integration time is less than a preset first threshold; When there is an overexposed peak in the second Raman spectrum signal and the second integration time is less than a preset first threshold, the second integration time is reduced so that the overexposed peak is not overexposed, and the non-overexposed peak is extracted; replacing the extracted signal with a signal within a preset range around the overexposed peak value in the second Raman spectrum signal before the second integration time is reduced, to form a new Raman spectrum signal; Wherein, the automatic adjustment model includes an adaptive weight module, a width module, a depth module and a linear layer; The input features of the automatic adjustment model are passed through the adaptive weight module to obtain weighted input features, and the weighted input features are passed through the width module and the depth module in parallel to obtain their respective outputs, and the output of the width module and the output of the depth module are fused, and then passed through the linear layer output to obtain the second integration time, the second bias and the second gain that reduce the background noise to a preset position; Input features of the automatic tuning model It is expressed as: ； in, represents the first integration time, represents the first temperature, represents the first Raman spectrum signal, represents transpose; The width module performs linear mapping on the weighted input features, and the calculation formula is as follows: ； in, represents the output of the width module, represents the weight matrix of the width module, represents the bias vector of the width module, Represents the weighted input features; The depth module includes two parallel embedding layers and a hidden layer. The weighted input features pass through the two embedding layers in parallel. The outputs of the two embedding layers pass through the hidden layer to obtain the output of the depth module. The hidden layer includes a preset number of activation layers set in sequence, and the calculation formula of each layer in the hidden layer is as follows: ； in, represents the output of each layer in the hidden layer, represents the input of each layer in the hidden layer, ReLU activation function calculation, represents the bias vector of each layer in the hidden layer, Represents the weight matrix of each layer in the hidden layer; The output of the width module and the output of the depth module are fused, wherein the calculation formula for the fusion is as follows: ； in, ； ； in, It represents the features after the output of the width module and the output of the depth module are fused. represents the output of the depth module, Represents the weight matrix of the fusion process for the width module, Represents the weight matrix of the fusion process for the depth module, Indicates the calculation of Sigmoid activation function; The output of the width module and the output of the depth module are fused to obtain the features, and then the linear layer output is used to obtain the second integration time, the second bias and the second gain that reduce the background noise to a preset position, wherein the calculation formula of the linear layer is as follows: ； in, ； in, represents the output of the linear layer, represents the weight matrix of the linear layer, The linear layer represents the bias vector, represents the second integration time, represents the second bias, Indicates the second gain.

2. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network according to claim 1, characterized in that: The first temperature is the operating temperature of a spectrometer in the Raman spectroscopy system, and the first temperature is a preset second threshold or is obtained by rounding up the current operating temperature of the spectrometer.

3. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network according to claim 1, characterized in that: After the automatic adjustment model outputs the second integration time, the second bias and the second gain, the spectrometer in the Raman spectroscopy system is configured to correspond to the second integration time, the second bias and the second gain, and then a second Raman spectroscopy signal is acquired.

4. The method for automatic adjustment and reconstruction of Raman spectrum based on neural network according to claim 1, characterized in that: The adaptive weight module includes a fully connected layer and an activation layer connected in sequence from data input to output, the input feature passes through the fully connected layer to obtain the original weight of the input feature, the original weight of the input feature passes through the activation layer to obtain the dynamic weight, and the input feature is weighted according to the dynamic weight to obtain the weighted input feature; The calculation formula of the activation layer is as follows: ； in, represents the dynamic weight, represents the original weight of the input feature, Indicates the number of each vector in the input feature; Weighted input features It is expressed as: ； in, represents the dynamic weight of the first integration time, represents the dynamic weight of the first temperature, represents the dynamic weight of the first Raman spectrum signal.

Images