A single-channel raman spectral unmixing method

Abstract

The present application relates to the technical field of spectral unmixing, and particularly relates to a single-channel Raman spectral unmixing method. The method comprises the following steps: obtaining a mixed Raman spectrum, and performing data coding on the mixed Raman spectrum to obtain high-dimensional features; the high-dimensional features are cut into a plurality of three-dimensional tensors that overlap with each other; the three-dimensional tensors are input into an RSSNet separation module to obtain separated feature tensors; the separated feature tensors are overlapped and added to obtain a high-dimensional feature sequence; the high-dimensional feature sequence is input into a mask network to obtain a mask; the high-dimensional features are input into a feature separation module to obtain the mask; and the high-dimensional features and the mask are input into a decoding and spectrum reconstruction module to obtain pure spectra. The present application breaks through the dependence on multi-spectral input by using a specially designed coding-separation-decoding deep neural network architecture, realizes single-channel unmixing, has high robustness and noise resistance, and has practical value in real-world applications.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spectral unmixing technology, and particularly relates to a single-channel Raman spectral unmixing method. BACKGROUND

[0002] Raman spectroscopy is a scattering spectroscopy technique based on molecular vibration and rotation, which can provide detailed chemical structure information of a substance, like a "molecular fingerprint" of the substance. Due to its advantages of rapidity, non-destructiveness, non-contact, etc., Raman spectroscopy has been widely applied in many fields such as material science, geological exploration, life science, industrial process control, etc.

[0003] In many practical application scenarios, the sample to be measured is often not a single pure substance, but a mixture composed of multiple chemical components. At this time, the collected Raman spectrum is the result of linear or nonlinear superposition of the pure spectrum signals of each component and background noise, resulting in the characteristic peaks of different substances overlapping and masking each other, which is difficult to directly identify. In order to identify the original components and their contents from such a complex mixed spectrum, a technique called "spectral unmixing" must be used. The goal of spectral unmixing is to decompose a mixed spectrum into pure spectra and corresponding abundances.

[0004] At present, for Raman spectral unmixing, the existing technologies can be mainly divided into the following categories, but all have significant limitations:

[0005] 1. Methods based on geometry and statistics:

[0006] This kind of method is the mainstream of traditional methods, including geometric methods such as VCA (Vertex Component Analysis); statistical methods such as ICA (Independent Component Analysis), NMF (Non-negative Matrix Factorization) or some methods based on Bayesian estimation. The design basis of these methods is an overdetermined system, that is, it is assumed to have multiple mixed spectrum data collected from different spatial points or under different conditions, for example, a Raman imaging image, each pixel of which corresponds to a spectrum. By analyzing the overall characteristics of this set of spectral data, the pure spectrum of the substance and its distribution are inversely calculated.

[0007] The disadvantage is that this kind of method is heavily dependent on "multi-channel" or "multi-spectral" input. However, in many key applications such as on-site rapid detection, portable device analysis or non-cooperative target detection, only a single mixed spectrum can usually be obtained. In these "single-channel" scenarios, the above methods are not applicable due to not meeting their basic data input requirements, greatly limiting the application range of Raman spectroscopy technology.

[0008] 2. Deep learning methods based on autoencoders:

[0009] In recent years, with the development of deep learning techniques, autoencoder networks have also been applied to hyperspectral image unmixing, and some research has explored their application in Raman imaging unmixing. This type of method compresses the input spectral data into a low-dimensional latent space through an encoder, and then reconstructs it through a decoder. During the training process, the weights of the decoder are optimized to fit the characteristics of the pure spectrum, thereby achieving unmixing.

[0010] However, this type of autoencoder-based method has similar fundamental flaws as traditional methods:

[0011] First, it also relies on multispectral input. This type of method is essentially still designed for image data. Like traditional geometric or statistical methods, they require a Raman image containing hundreds or even thousands of pixel points of spectra as input. Therefore, when faced with a "single-channel" detection task with only a single mixed spectrum, this type of method is also not applicable. Second, it is difficult to deal with large-scale spectral libraries. Existing autoencoder unmixing models can usually only handle a limited number of pure components. In scenarios where unmixing needs to be performed from a large-scale spectral library containing hundreds or even thousands of substances, the performance of these models will be severely degraded, or even completely ineffective, failing to meet the complex demands of practical applications.

[0012] 3. Methods based on sparse regression:

[0013] There are also unmixing algorithms based on sparse regression in the prior art, such as SUnSAL (Sparse Unmixing by Variable Splitting and Augmented Lagrange). This method assumes that the mixed spectrum can be sparsely linearly combined from a few pure spectra in a large pre-defined spectral library, and the unmixing process is to find this optimal sparse subset. In theory, this method can handle a single mixed spectrum.

[0014] The fatal flaw is that such algorithms are very sensitive to noise. The actual collected Raman spectrum will inevitably be disturbed by noise from multiple sources such as sample fluorescence, detector dark current, stray light, etc. In the presence of noise, the performance of the sparse regression algorithm will decrease dramatically, often leading to unmixing failure or producing seriously erroneous results, making it lack practicality and reliability in real and complex application environments.

[0015] Therefore, there is an urgent need in the art for a completely new technical solution that can achieve accurate and robust unmixing based only on a single, noisy mixed Raman spectrum, especially when faced with a large-scale spectral library containing hundreds or even thousands of substances. The prior art has failed to effectively solve this difficult problem. SUMMARY

[0016] The present application aims to at least solve one of the technical problems existing in the related art. To this end, the present application provides a single-channel Raman spectrum unmixing method, which breaks through the dependence on multi-spectral input, realizes single-channel unmixing, has high robustness and noise resistance, and has practical value in real environment applications.

[0017] The present application provides a single-channel Raman spectrum unmixing method, comprising:

[0018] S1: obtaining a mixed Raman spectrum, and performing data encoding on the mixed Raman spectrum to obtain a high-dimensional feature;

[0019] S2: inputting the high-dimensional feature into a feature separation module to obtain a mask, comprising:

[0020] S21: cutting the high-dimensional feature into a plurality of three-dimensional tensors that overlap with each other;

[0021] S22: inputting the three-dimensional tensors into an RSSNet separation module to obtain a separation feature tensor;

[0022] S23: overlapping and adding the separation feature tensors to obtain a high-dimensional feature sequence;

[0023] S24: inputting the high-dimensional feature sequence into a mask network to obtain a mask;

[0024] S3: inputting the high-dimensional feature and the mask into a decoding and spectrum reconstruction module to obtain a pure spectrum.

[0025] According to the single-channel Raman spectrum unmixing method provided by the present application, step S1 comprises:

[0026] S11: performing one-dimensional convolution on the mixed Raman spectrum to obtain a local feature;

[0027] S12: performing global layer normalization on the local feature to obtain a normalized local feature;

[0028] S13: performing parameterized linear correction on the normalized local feature to obtain a high-dimensional feature.

[0029] According to the single-channel Raman spectrum unmixing method provided by the present application, the RSSNet separation module is based on a U-Net module and is composed of a global attention mechanism and a local attention mechanism.

[0030] According to the single-channel Raman spectrum unmixing method provided by the present application, the process of the RSSNet separation module is:

[0031] S221: inputting the three-dimensional tensors into an intra-block processing module to obtain an intra-block processing result;

[0032] The flow of the intra-block processing module is:

[0033] S2211: performing deep separable convolution on the three-dimensional tensor to obtain an intra-block convolution result;

[0034] S2212: transposing the three-dimensional tensor to obtain a first transposed three-dimensional tensor;

[0035] S2213: inputting the first transposed three-dimensional tensor into a TDA module to obtain an intra-block TDA result;

[0036] S2214: element-wise adding the intra-block convolution result and the intra-block TDA result to obtain an intra-block processing result;

[0037] S222: inputting the intra-block processing result into an inter-block processing module to obtain an inter-block processing result, and the flow of the inter-block processing module is:

[0038] S2221: performing deep separable convolution on the inter-block processing result to obtain an inter-block convolution result;

[0039] S2222: transposing the inter-block processing result to obtain a second transposed three-dimensional tensor;

[0040] S2223: inputting the second transposed three-dimensional tensor into a TDA module to obtain an inter-block TDA result;

[0041] S2224: element-wise adding the inter-block convolution result and the inter-block TDA result to obtain the inter-block processing result,

[0042] S223: element-wise adding the inter-block processing result and the three-dimensional tensor to obtain a separated feature tensor.

[0043] According to the single-channel Raman spectrum unmixing method provided in the application, the transposition directions in the intra-block processing module and the inter-block processing module are different, the dimension of the first transposed three-dimensional tensor is , and the dimension of the second transposed three-dimensional tensor is , wherein is the number of split blocks, is the data block length, is the feature dimension.

[0044] According to the single-channel Raman spectrum unmixing method provided in the application, the separated feature tensor is subjected to deep separable convolution to obtain a cyclic input; the cyclic input is taken as the input of an RSSNet separation module, and one or more cycles are used to improve the separation effect of the RSSNet separation module.

[0045] According to the single-channel Raman spectrum unmixing method provided in the application, the flow of the TDA module is:

[0046] S100: Convolve the input of the TDA module to obtain the TDA convolution result;

[0047] S200: The TDA convolution result is downsampled multiple times to obtain multiple downsampled results;

[0048] S300: Perform average pooling on each downsampling result to obtain multiple corresponding average pooling results, and combine the average pooling results with the last downsampling result. The overall information and global features are obtained by adding elements together. :

[0049]

[0050] in, For the first The second downsampling result For the number of samples, , The total number of samples, The result of TDA convolution. For average pooling function, This is element-wise addition;

[0051] S400: Input the overall information global features into the global attention module to obtain the top-level attention signal. ;

[0052] S500: Upsample the top-level attention signal to obtain an upsampling result, and add the upsampling result and the downsampling result element-wise to obtain the LA input value.

[0053]

[0054] in, For the first Second upsampling result, For the first Next LA input value;

[0055] S600: Input the LA input value to the LA module to obtain the LA output result:

[0056]

[0057] in, For the first The next LA output result, Functions of the LA module;

[0058] S700: The last LA output result is the output result of the TDA module.

[0059] According to the single-channel Raman spectrum demixing method provided by the application, the flow of the global attention module is:

[0060] S110: input the high-dimensional feature into the global attention module to obtain a global attention result; input the global attention result into the multi-head self-attention layer to obtain a multi-head self-attention result;

[0061] S120: add the multi-head self-attention result and the high-dimensional feature element by element to obtain a feedforward network layer input value; S130: input the feedforward network layer input value into the feedforward network layer to obtain a feedforward network layer output;

[0062] S140: add the feedforward network layer input value and the feedforward network layer output element by element to obtain a high-dimensional feature output value;

[0063] .

[0064] According to the single-channel Raman spectrum demixing method provided by the application, the flow of the LA module is:

[0065] S210: input the high-dimensional feature into the LA module to obtain a LA first convolution result; S230: add the LA first convolution result and the high-dimensional feature element by element to obtain a LA second convolution result;

[0066] S240: add the LA second convolution result and the high-dimensional feature element by element to obtain a LA third intermediate value;

[0067] S250: multiply the LA first intermediate value and the LA third intermediate value element by element to obtain a LA fourth intermediate value; S260: add the LA fourth intermediate value and the LA second intermediate value element by element to obtain a high-dimensional feature output value;

[0068] .

[0069] According to the single-channel Raman spectrum demixing method provided by the application, step S3 comprises:

[0070] S31: multiply the high-dimensional feature and the mask element by element to obtain a shielded feature;

[0071] S32: perform one-dimensional transpose convolution on the shielded feature to obtain a pure spectrum.

[0072]

[0073] ​​​​According to the single-channel Raman spectrum demixing method provided by the application, when the component of the mixed Raman spectrum is greater than 2, the mixed Raman spectrum is demixed, and two demixed spectra are output, and the demixed spectra are continuously demixed until a pure spectrum is output.

[0074] According to the single-channel Raman spectrum demixing method provided by the application, when the component of the mixed Raman spectrum is greater than 2, the mixed Raman spectrum is demixed, and two demixed spectra are output, and the demixed spectra are continuously demixed until a pure spectrum is output.

[0075] The one or more technical solutions in the embodiments of the application have at least one of the following technical effects:

[0076] The single-channel Raman spectrum demixing method provided by the application overcomes many defects in the prior art and has the following remarkable beneficial effects:

[0077] 1. The dependence on multi-spectrum input is broken, and single-channel demixing is realized: the existing mainstream geometric, statistical and auto-encoder demixing methods all need to rely on “multi-channel” or “multi-spectrum” data input, and cannot be applied to the fundamental problem of more general single-channel detection scenarios. The application proposes a new end-to-end demixing paradigm. The application only needs a single mixed Raman spectrum obtained in a single measurement as input, and can complete the high-precision demixing task. This greatly expands the application scenarios of Raman spectrum demixing.

[0078] 2. It has high robustness and anti-noise ability, and has practical value for application in real environment: the existing sparse regression method is the only method that can be used for single-channel, but it is very sensitive to noise, which leads to a fatal defect that the performance sharply decreases and the result is unreliable in actual application. The application uses a specially designed encoding-separating-decoding deep neural network architecture, especially its unique double-path separation mechanism and attention module, to deeply learn the internal structure and dependency of spectral features. This enables the application to effectively suppress noise interference and obtain accurate and reliable demixing results even under low signal-to-noise ratio conditions. The application fundamentally improves the stability and practicality of single-channel Raman demixing technology in complex and real environments.

[0079] 3. Support open-domain unmixing of large-scale spectral library, significantly improve the scope of application of the method: for the existing part of the deep learning method (such as autoencoder) and the method based on geometry can only handle a few known components, it is difficult to deal with the limitation of complex mixture system, the network structure designed by the invention can efficiently unmix from a large-scale pre-defined spectral library containing thousands of candidate substances. This makes the invention capable of performing material identification tasks in open-domain, non-cooperative scenarios, such as quickly identifying the components of unknown mixtures from a large chemical library in security checks, which has important practical application significance.

[0080] 4. A training method based on synthetic data is proposed, which realizes low-cost model training: the invention trains the model by constructing a large-scale synthetic mixed spectrum dataset, avoiding the huge cost and time required for collecting and annotating a large number of diverse mixture spectrum samples in the real world. Experiments show that the model trained only by synthetic data can successfully generalize to real measurement data and achieve excellent results, which provides an economical and efficient technical path for the wide application of this technology.

[0081] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0082] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0083] Figure 1 is the flowchart of the single-channel Raman spectrum unmixing method provided by the present application.

[0084] Figure 2 is the flowchart of the single-channel Raman spectrum unmixing method provided by the present application.

[0085] Figure 3 is the flowchart of step S2 of the present application.

[0086] Figure 4 is the flowchart of the TDA module of the present application.

[0087] Figure 5 is the flowchart of the LA module of the present application.

[0088] Figure 6 is the decoding result comparison diagram of embodiment 1 and other methods of the present application.

[0089] Figure 7 is a decoding result graph of embodiment 2 of the present application.

[0090] Figure 8 is a decoding result graph of three components in embodiment 3 of the present application.

[0091] Figure 9 is a decoding result graph of four components in embodiment 3 of the present application.

[0092] Figure 10 is a decoding result graph of five components in embodiment 3 of the present application. DETAILED DESCRIPTION

[0093] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be clearly and completely described below. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application. The following embodiments are used to illustrate the present application, but cannot be used to limit the scope of the present application.

[0094] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in combination with the embodiment or example are contained in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms is not necessarily for the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0095] The present application will be described below in combination with Figures 1 to 10 The present application is described.

[0096] Embodiment 1

[0097] As shown in Figure 1 and Figure 2 The present application provides a single-channel Raman spectrum unmixing method, comprising:

[0098] S1: obtaining a mixed Raman spectrum, and performing data encoding on the mixed Raman spectrum to obtain a high-dimensional feature;

[0099] S2: inputting the high-dimensional feature into a feature separation module to obtain a mask;

[0100] S3: inputting the high-dimensional feature and the mask into a decoding and spectrum reconstruction module to obtain a pure spectrum.

[0101] Specifically, step S1 includes:

[0102] S11: performing one-dimensional convolution on the mixed Raman spectrum to obtain a local feature.

[0103] Specifically, the embodiment of the present application performs one-dimensional convolution operation on the input spectrum to extract the local feature. The convolution layer has a specific convolution kernel size and step.

[0104] S12: performing global layer normalization on the local feature to obtain a normalized local feature.

[0105] Specifically, the embodiment of the present application performs normalization processing on the convolution feature to stabilize the training process.

[0106] S13: performing parameterized linear correction on the normalized local feature to obtain a high-dimensional feature.

[0107] Specifically, the embodiment of the present application uses the PReLU function as the activation function for the normalized local feature, and performs nonlinear transformation on the normalized feature. After processing by the encoding module, a high-dimensional feature is obtained, where is a tensor with a dimension of , wherein is the feature dimension, and is the length of the transformed sequence.

[0108] In summary, the formula of step S1 is:

[0109]

[0110] wherein, is a parameterized linear correction unit activation function, is a global layer normalization operation, is one-dimensional convolution.

[0111] Specifically, the RSSNet separation module is based on a U-Net module, and is composed of a global attention mechanism and a local attention mechanism. As shown in Figure 3 , step S2 includes:

[0112] S2: inputting the high-dimensional feature into a feature separation module to obtain a mask, including:

[0113] S21: cutting the high-dimensional feature into a plurality of three-dimensional tensors that overlap with each other.

[0114] Specifically, in order to model the local and global dependencies of the spectrum at the same time, first cut into multiple blocks that overlap each other. For example, a sequence of length is divided into data blocks of length with 50% overlap between adjacent blocks. After stacking all the blocks, a three-dimensional tensor is formed, with dimensions .

[0115] S22: input the three-dimensional tensor into the RSSNet separation module to obtain a separation feature tensor.

[0116] Specifically, the process of the RSSNet separation module is as follows:

[0117] S221: input the three-dimensional tensor into the intra-block processing module to obtain an intra-block processing result, wherein the process of the intra-block processing module is as follows:

[0118] S2211: perform deep separable convolution (DWConv) on the three-dimensional tensor to obtain an intra-block convolution result.

[0119] S2212: transpose the three-dimensional tensor to obtain a first transposed three-dimensional tensor; wherein the dimensions of the first transposed three-dimensional tensor after transposition are .

[0120] S2213: input the first transposed three-dimensional tensor into the TDA module to obtain an intra-block TDA result.

[0121] S2214: add the intra-block convolution result and the intra-block TDA result element by element to obtain the intra-block processing result.

[0122] S222: input the intra-block processing result into the inter-block processing module to obtain an inter-block processing result, wherein the process of the inter-block processing module is as follows:

[0123] S2221: perform deep separable convolution on the intra-block processing result to obtain an inter-block convolution result;

[0124] S2222: transpose the intra-block processing result to obtain a second transposed three-dimensional tensor; wherein the dimensions of the second transposed three-dimensional tensor after transposition are .

[0125] S2223: input the second transposed three-dimensional tensor into the TDA module to obtain an inter-block TDA result;

[0126] ​​S2224: add the inter-block convolution result and the inter-block TDA result element by element to obtain an inter-block processing result.

[0127] S223: add the inter-block processing result and the three-dimensional tensor element by element to obtain a separated feature tensor .

[0128] S23: add the separated feature tensors element by element to obtain a high-dimensional feature sequence. Specifically, the separated feature tensors are seamlessly spliced back into a continuous high-dimensional feature sequence (dimension ) This operation is the inverse process of step S21.

[0129] S24: input the high-dimensional feature sequence into a mask network to obtain a mask. Specifically, the high-dimensional feature sequence is input into a mask network, which is usually composed of a PReLU activation layer and a one-dimensional convolution layer with a kernel size of 1. The output of the mask network is a mask , wherein, is the mask sequence number, is the preset number of pure spectra to be separated, and in the embodiment . The dimension of each mask is the same as that of the high-dimensional feature .

[0130] Specifically, as shown in Figure 4 , the TDA module (Top-Down Attention Module) is a processing unit for deep feature extraction and modulation, and the internal architecture is an enhanced structure based on a U-Net network. The process of the TDA module described in the present application is as follows:

[0131] S100: convolve the input of the TDA module to obtain a TDA convolution result;

[0132] S200: downsample the TDA convolution result to obtain a downsampled result:

[0133]

[0134] wherein, is the downsampled result, is the downsampled function, is the sampling number, , is the total number of samples, is the TDA convolution result; in the embodiment of the present application .

[0135] ​S300: average pooling is performed on each down-sampling result to obtain a plurality of corresponding average pooling results, and the average pooling results and the last down-sampling result are element-wise added to obtain a whole information global feature. The whole information global feature is obtained by element-wise addition.

[0136]

[0137] wherein, is an average pooling function, is element-wise addition.

[0138] S400: the whole information global feature is input into a global attention module to obtain a top-level attention signal. ;

[0139] S500: the top-level attention signal is up-sampled to obtain an up-sampling result, and the up-sampling result and the down-sampling result are element-wise added to obtain an LA input value:

[0140]

[0141]

[0142] wherein, is the nth up-sampling result, is an up-sampling operation, is the nth LA input value, in particular, ; S600: the LA input value is input into an LA module to obtain an LA output result:

[0143]

[0144]

[0145] wherein, is the nth LA output result, is the nth LA output result, is a function of the LA module, in particular ,

[0146] S700: the last LA output result is the output result of the TDA module.

[0147] wherein, the process of the global attention module is:

[0148] S110: the is input into a multi-head self-attention layer (MHA) to obtain a multi-head self-attention result;

[0149] ​​​​​​S120: element-wise add the multi-head self-attention result and element-wise add to obtain a feedforward network layer input value;

[0150] S130: input the feedforward network layer input value into a feedforward network layer (FFN) to obtain a feedforward network layer output;

[0151] S140: element-wise add the feedforward network layer input value and the feedforward network layer output to obtain .

[0152] As shown in Figure 5 , the application introduces a LA module (Up-sampling and Local Attention Module), which dynamically generates parameters such as gain and bias for fine adjustment of the current scale feature according to the information of adjacent and coarser scale feature maps. This local and adaptive modulation process can gradually and finely reconstruct a feature representation containing rich details.

[0153] The process of the LA module is as follows:

[0154] S210: perform convolution operation on to obtain a LA first convolution result, perform Sigmoid activation on the LA first convolution result, and then perform interpolation to obtain a LA first intermediate value;

[0155] S230: perform convolution operation on to obtain a LA second convolution result, and perform interpolation on the LA second convolution result to obtain a LA second intermediate value;

[0156] S240: perform convolution operation on to obtain a LA third intermediate value;

[0157] S250: element-wise multiply the LA first intermediate value and the LA third intermediate value to obtain a LA fourth intermediate value;

[0158] S260: element-wise add the LA fourth intermediate value and the LA second intermediate value to obtain .

[0159] Specifically, step S3 comprises:

[0160] S31: element-wise multiply the high-dimensional feature and the mask to obtain a screened feature.

[0161] Specifically, the high-dimensional feature is element-wise multiplied with each mask to obtain C screened features . This operation is equivalent to filtering out the exclusive feature information of each pure component for each pure component.

[0162] S32: One-dimensional transposed convolution is performed on the shielded features to obtain a pure spectrum.

[0163] Each shielded feature is subjected to one-dimensional transposed convolution operation. This operation is the inverse operation of one-dimensional convolution in the encoding module, responsible for mapping high-dimensional features back to the original spectral dimension, thereby obtaining the final estimated pure spectrum .

[0164] Specifically, the formula of step S3 is:

[0165]

[0166] wherein, is a one-dimensional transposed convolution function, is an element-wise multiplication.

[0167] In particular, the system can work accurately and needs to be trained. The present application also includes a unique training method, the process of which is as follows:

[0168] 1. Construct a synthetic training data set:

[0169] First, obtain an open-source large-scale pure substance Raman spectrum library, such as the RRUFF database, which contains thousands of pure spectra of known substances.

[0170] Randomly select D different pure spectra from the library.

[0171] According to the linear mixing model, assign random abundance coefficients to the D pure spectra, the sum of which is 1, and linearly add them.

[0172] Add random red noise and white noise superimposed noise in a specific signal-to-noise ratio (SNR) range to the synthesized mixed spectrum, and finally generate a single mixed spectrum with noise simulating real conditions as a training data. Generate tens of thousands of mixed spectra as a training data set.

[0173] This process can be represented by the following formula:

[0174]

[0175] wherein, is a mixed spectrum, is the th pure spectrum, is the superimposed noise.

[0176] 2. Define the loss function:

[0177] The negative value of the scale-invariant signal-to-noise ratio (SI-SNR) is used as the loss function. The SI-SNR measures the difference between the estimated pure spectrum output by the system and the true pure spectrum. A higher SI-SNR value indicates that the estimation result is closer to the true value.

[0178] 3. Model optimization:

[0179] The synthesized mixed spectrum is input into the system to obtain the estimated spectrum. The SI-SNR loss between the estimated spectrum and the pure spectrum is calculated. Using optimization algorithms such as gradient descent and the Adam optimizer, the weight parameters of all neural network modules in the system are continuously adjusted through backpropagation to maximize the SI-SNR value, i.e., minimize the loss. This process is repeated until the model converges.

[0180] like Figure 6 As shown, Figure 6 This is a comparison chart of the decoding results of Embodiment 1 of the present invention and other methods. Figure 6 In the middle (a), the Raman spectrum of a 1:1 mixture of orpiment and amazonite is shown. The spectrum separated by the method (RSSNet prediction) used in this invention is basically the same as the pure spectrum, while the spectra separated by other methods (SUnSAL prediction and NNOMP prediction) are completely different from the pure spectrum.

[0181] Figure 6 In the middle (b), the Raman spectrum of a 1:1 mixture of realgar and amazonite is shown. The spectrum separated by the method (RSSNet prediction) used in this invention is basically the same as the pure spectrum, while the spectra separated by other methods (SUnSAL prediction and NNOMP prediction) are completely different from the pure spectrum.

[0182] Figure 6 In the middle (c), the Raman spectrum formed by a 1:2 mixture of phlogopite and ochre is shown. The spectrum separated by the method (RSSNet prediction) used in this invention is basically the same as the pure spectrum, while the spectra separated by other methods (SUnSAL prediction and NNOMP prediction) are completely different from the pure spectrum.

[0183] Figure 6 In the middle (d), the Raman spectrum is formed by a 1:2 mixture of orpiment and calcite. The spectrum separated by the method used in this invention (RSSNet prediction) is basically the same as the pure spectrum, while the spectra separated by other methods (SUnSAL prediction and NNOMP prediction) are completely different from the pure spectrum.

[0184] Example 2

[0185] When the Raman spectrum has more than 2 components, the Raman spectrum is demixed. By modifying the number of outputs of the mask in the mask network, several pure spectra corresponding to the number of components can be directly output.

[0186] As Figure 7 shown, Figure 7 is the decoding result graph of embodiment 2 of the present application.

[0187] Among them, Figure 7 (a) is the Raman spectrum of the spectrum mixed by three components of calcio-roselite, eakelite and partheite, and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0188] Figure 7 (b) is the Raman spectrum of the spectrum mixed by three components of chernovite, xonotlite and rathite-(Ce), and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0189] Figure 7 (c) is the Raman spectrum of the spectrum mixed by three components of vanoxite, plumbosulphate and pollucite, and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0190] Embodiment 3

[0191] When the component of the Raman spectrum is greater than 2, the Raman spectrum can be first de-mixed, and two de-mixed spectra are outputted, and the de-mixed spectra are continuously de-mixed until the pure spectrum is outputted. For example, the mixed spectrum of three components

[0192] As Figure 8 shown, Figure 8 is the decoding result graph of three components in embodiment 3 of the present application.

[0193] Among them, Figure 8 (a) is the Raman spectrum of the spectrum mixed by three components of bassetite, stibiocassiterite and dypingite, and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0194] Figure 8 (b) is the Raman spectrum of the spectrum mixed by three components of sodalite, montebrasite and langbeinite, and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0195] Figure 8 (c) is the Raman spectrum of the spectrum mixed by three components of vikite-(Y), dypingite and barite, and the spectrum graph separated by the method (RSSNet prediction) adopted by the present application is basically the same as the pure spectrum.

[0196] As Figure 9 shown, Figure 9 is the decoding result graph of four components in embodiment 3 of the present application.

[0197] wherein, Figure 9 In (a), the Raman spectrum of the spectrum mixed by the four components of dolomite, manganhumite, chenevixite and zabuyelite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0198] Figure 9 In (b), the Raman spectrum of the spectrum mixed by the four components of chengite, bassetite, clinochlore and barite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0199] Figure 9 In (c), the Raman spectrum of the spectrum mixed by the four components of boehmite, vanadinite, lithiophorite and kurchatovite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0200] As shown in Figure 10 , Figure 10 is the decoding result graph of five components in embodiment 3 of the present application.

[0201] wherein, Figure 10 In (a), the Raman spectrum of the spectrum mixed by the five components of wickenburgite, barylite, red lead chromate, aluminocerine and pollucite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0202] Figure 10 In (b), the Raman spectrum of the spectrum mixed by the five components of xanthiosite, chudobaite, strontioaragonite, pollucite and thenardite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0203] Figure 10 In (c), the Raman spectrum of the spectrum mixed by the five components of dolomite, natrolite, eakerite, hemimorphite and barite is basically the same as the pure spectrum separated by the method (RSSNet prediction) adopted in the present application.

[0204] In summary, the present application proposes a new end-to-end unmixing paradigm to solve the fundamental problem that the existing mainstream geometric, statistical and autoencoder unmixing methods all need to rely on “multi-channel” or “multi-spectral” data input and cannot be applied to more general single-channel detection scenarios. The present application only needs a single mixed Raman spectrum obtained in a single measurement as input to complete the high-precision unmixing task. This greatly expands the application scenarios of Raman spectrum unmixing. At the same time, the method proposed in the present application can also be used for the unmixing of noisy Raman spectra.

[0205] Finally, it should be noted that the above embodiments are merely used to illustrate the technical solutions of the present application, instead of limiting the present application; even though the present application has been described in detail with reference to the foregoing embodiments, those ordinarily skilled in the art should understand: the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0206] It should be noted that embodiments of the present disclosure can be realized by hardware, software, or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control codes, for example, such codes are provided on a programmable memory or a data carrier such as an optical or electronic signal carrier.

[0207] In addition, although the operations of the methods of the present disclosure are described in a particular order in the drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the illustrated operations must be performed to achieve the desired result. On the contrary, the steps depicted in the flowcharts can change the order of execution. Additionally or alternatively, some steps can be omitted, combined into one step, and / or divided into multiple steps. It should also be noted that the features and functions of two or more devices according to the present disclosure can be embodied in one device. Conversely, the features and functions of one device described above can be further divided into devices embodied by multiple devices.

[0208] Although the present disclosure has been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments. The present disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A single channel Raman spectral unmixing method, characterized in that, The method comprises the following steps: S1: obtaining a mixed Raman spectrum, and performing data coding on the mixed Raman spectrum to obtain a high-dimensional feature; S2: inputting the high-dimensional feature into a feature separation module to obtain a mask, comprising: S21: cutting the high-dimensional feature into a plurality of three-dimensional tensors that overlap with each other; S22: inputting the three-dimensional tensors into an RSSNet separation module to obtain a separation feature tensor; S23: overlapping and adding the separation feature tensors to obtain a high-dimensional feature sequence; S24: inputting the high-dimensional feature sequence into a mask network to obtain a mask; S3: inputting the high-dimensional feature and the mask into a decoding and spectrum reconstruction module to obtain a pure spectrum.

2. The single channel Raman spectral unmixing method of claim 1, wherein, Step S1 comprises: S11: performing one-dimensional convolution on the mixed Raman spectrum to obtain a local feature; S12: performing global layer normalization on the local feature to obtain a normalized local feature; S13: performing parameterized linear correction on the normalized local feature to obtain a high-dimensional feature.

3. The single channel Raman spectral unmixing method of claim 1, wherein, The RSSNet separation module is based on a double-path structure and comprises a global feature processing module and a local feature processing module.

4. The single channel Raman spectral unmixing method of claim 1, wherein, The process of the RSSNet separation module is as follows: S221: inputting the three-dimensional tensor into an intra-block processing module to obtain an intra-block processing result; The process of the intra-block processing module is as follows: S2211: performing depth separable convolution on the three-dimensional tensor to obtain an intra-block convolution result; S2212: transposing the three-dimensional tensor to obtain a first transposed three-dimensional tensor; S2213: inputting the first transposed three-dimensional tensor into a TDA module to obtain an intra-block TDA result; S2214: elementarily adding the intra-block convolution result and the intra-block TDA result to obtain the intra-block processing result; S222: inputting the intra-block processing result into an inter-block processing module to obtain an inter-block processing result, and the process of the inter-block processing module is as follows: S2221: performing depth separable convolution on the inter-block processing result to obtain an inter-block convolution result; S2222: transposing the inter-block processing result to obtain a second transposed three-dimensional tensor; S2223: inputting the second transposed three-dimensional tensor into a TDA module to obtain an inter-block TDA result; S2224: elementarily adding the inter-block convolution result and the inter-block TDA result to obtain the inter-block processing result, S223: elementarily adding the inter-block processing result and the three-dimensional tensor to obtain a separation feature tensor.

5. The single channel Raman spectral unmixing method of claim 4, wherein, The transposition directions in the intra-block processing module and the inter-block processing module are different, and the dimensions of the first transposed three-dimensional tensor are ; The dimensions of the second transposed three-dimensional tensor are wherein, is the number of cut blocks, is the data block length, is the feature dimension.

6. The single channel Raman spectral unmixing method of claim 3, wherein, The separation feature tensor is subjected to depth separable convolution to obtain a cyclic input; the cyclic input is taken as the input of the RSSNet separation module and is subjected to one or more cycles to improve the separation effect of the RSSNet separation module.

7. The single channel Raman spectral unmixing method of claim 4, wherein, The process of the TDA module is as follows: S100: performing convolution on the input of the TDA module to obtain a TDA convolution result; S200: performing multiple times of down-sampling on the TDA convolution result to obtain a plurality of down-sampling results; S300: average pooling is respectively performed on each down-sampling result to obtain a plurality of corresponding average pooling results, and the average pooling results and the last down-sampling result are added element by element to obtain the global feature of the whole image The global feature of the whole image is obtained by adding elements : wherein, is the first down-sampling result, is the number of sampling, , is the total number of sampling, is the TDA convolution result, is the average pooling function, is the element-wise addition; S400: input the whole information global feature into a global attention module to obtain a top-level attention signal ; S500: performing up-sampling on the top-layer attention signal to obtain an up-sampling result, and elementarily adding the up-sampling result and the down-sampling result to obtain an LA input value: wherein, is the first upsampled result, is the first LA input value; S600: inputting the LA input value into an LA module to obtain an LA output result: wherein, is the LA output result, is a function of the LA module; S700: The last LA output result is the output result of the TDA module.

8. The single channel Raman spectral unmixing method of claim 7, wherein, The flow of the global attention module is as follows: S110: input to the multi-head self-attention layer to obtain a multi-head self-attention result; input to the multi-head self-attention layer to obtain a multi-head self-attention result; S120: add the multi-head self-attention result and element-wise addition to obtain the feedforward network layer input value; S130: Input the feedforward network layer input value into the feedforward network layer to obtain a feedforward network layer output; S140: element-wise adding the feedforward network layer input values and the feedforward network layer output values to obtain .

9. The single channel Raman spectral unmixing method of claim 7, wherein, The flow of the LA module is as follows: S210: obtaining performing a convolution operation to obtain a LA first convolution result, performing a Sigmoid activation on the LA first convolution result, and performing interpolation to obtain a LA first intermediate value; S230: obtaining a second convolution result of LA by performing a convolution operation, and obtaining a second intermediate value of LA by performing interpolation on the second convolution result of LA. performing a convolution operation to obtain a second convolution result of LA, and performing interpolation on the second convolution result of LA to obtain a second intermediate value of LA. S240: obtaining the third intermediate value LA by performing a convolution operation on the first intermediate value LB and the second intermediate value LC. obtaining the third intermediate value LA by performing a convolution S250: Multiply the LA first intermediate value and the LA third intermediate value element by element to obtain an LA fourth intermediate value; S260: element-wise adding the LA fourth intermediate value and the LA second intermediate value to obtain .

10. The single channel Raman spectral unmixing method of claim 1, wherein, Step S3 comprises: S31: Multiply the high-dimensional feature and the mask element by element to obtain a screened feature; S32: Perform one-dimensional transpose convolution on the screened feature to obtain a pure spectrum.

11. The single channel Raman spectral unmixing method of claim 1, wherein, When the number of components of the mixed Raman spectrum is greater than 2, the mixed Raman spectrum is unblended, and two unblended spectra are output. The unblended spectra are continuously unblended until a pure spectrum is output.

12. The single channel Raman spectral unmixing method of claim 1, wherein, When the number of components of the mixed Raman spectrum is greater than 2, the mixed Raman spectrum is unblended, and a plurality of pure spectra corresponding to the number of components are directly output.

Images