Multi-dimensional Raman spectrum in-situ analysis system and method based on high-speed line scanning

Abstract

The invention relates to the technical field of material in-situ analysis, in particular to a multi-dimensional Raman spectrum in-situ analysis system and method based on high-speed line scanning, and the system comprises a laser focus stabilization unit, a material signal registration unit, a correlation analysis unit, a background correction unit and a material characteristic analysis unit. According to the method, the position drift vector is calculated by collecting the position coordinates of the nanometer displacement table, the predicted position drift vector is obtained in combination with Kalman filtering, and dynamic focus compensation is executed; obtaining a material analysis signal, and carrying out preprocessing and registration to obtain a registered material analysis signal; constructing a two-way analysis network to process and register the material analysis signal to obtain a spectrum-structure relation model; determining a background contribution weight and a background interference signal according to the registered Raman scattering spectrum to obtain a clean Raman scattering spectrum; the method solves the problems that the signal-to-noise ratio is reduced and the data error is large, and is beneficial for obtaining an accurate material structure and a grain boundary analysis result.

Description

Technical Field

[0001] The present invention relates to the technical field of in-situ analysis of materials, and specifically to a multi-dimensional Raman spectroscopy in-situ analysis system and method based on high-speed line scanning. Background Art

[0002] Raman spectroscopy is a non-destructive optical analysis technique that obtains information on molecular vibration and rotation energy levels by detecting the inelastic scattering of incident light by a sample, thereby achieving precise characterization of the composition of substances, molecular structure, crystal phase, etc. The high-speed line scanning Raman spectroscopy technique significantly improves the detection throughput and efficiency of nanomaterials by focusing the laser into a line to collect continuous spectral information, and has shown unique advantages especially in the large-area characterization of two-dimensional materials and heterostructures.

[0003] However, in practical applications, high-speed line scanning faces the following technical bottlenecks: First, high-speed scanning leads to a decrease in the signal-to-noise ratio of the original spectrum, affecting the accuracy and reliability of spectral analysis; second, there is a large data space registration error in the multi-dimensional signals required for material analysis, and it is impossible to accurately establish the correlation between structural parameters and non-linear optical responses; third, the focus drift caused by the vibration of the nano-displacement stage will cause the excitation spot to defocus on the sample, affecting the accuracy of material analysis.

[0004] Therefore, a multi-dimensional Raman spectroscopy in-situ analysis system and method based on high-speed line scanning are proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-dimensional Raman spectroscopy in-situ analysis system and method based on high-speed line scanning, to solve the problems of decreased signal-to-noise ratio and large data errors, and to be able to obtain accurate material structure and grain boundary analysis results. The present invention includes a laser focus stabilization unit, a material signal registration unit, a correlation analysis unit, a background correction unit, and a material property analysis unit. By collecting the position coordinates of the nano-displacement stage, the position drift vector is calculated, and combined with the Kalman filter to obtain the predicted position drift vector, and dynamic focus compensation is performed; the material analysis signal is obtained and preprocessed and registered to obtain the registered material analysis signal; a dual-channel analysis network is constructed to process the registered material analysis signal to obtain a spectrum-structure relationship model; the background contribution weight and background interference signal are determined according to the registered Raman scattering spectrum to obtain a clean Raman scattering spectrum; based on the spectrum-structure relationship model, the registered material analysis signal, and the clean Raman scattering spectrum, the material structure analysis result and the grain boundary analysis result are obtained.

[0006] To achieve the above object, the present invention provides the following technical solutions: A multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning, including.

[0007] The laser focus stabilization unit collects the position coordinates of the nano-displacement stage to calculate the position drift vector, combines the Kalman filter to obtain the predicted position drift vector, and performs dynamic focus compensation based on the predicted position drift vector; The material signal registration unit acquires the material analysis signal, performs preprocessing and registration to obtain the registered material analysis signal; the material analysis signal includes Raman scattering spectrum, second harmonic signal, and laser power; The correlation analysis unit constructs a dual-channel analysis network to process the registered material analysis signal and obtains a spectrum-structure relationship model; The background correction unit applies wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal, determines the background contribution weight in combination with the KL divergence, and obtains the clean Raman scattering spectrum based on the background contribution weight and the background interference signal; The material property analysis unit obtains the material structure analysis result and the grain boundary analysis result based on the spectrum-structure relationship model, the registered material analysis signal, and the clean Raman scattering spectrum.

[0008] Preferably, combining the Kalman filter to obtain the predicted position drift vector and performing dynamic focus compensation based on the predicted position drift vector includes: using the Kalman filter algorithm to process the position drift vector, constructing a state equation and an observation equation; according to the state equation and the observation equation, recursively calculating the predicted position drift vector through the prediction step and the update step; calculating the beam path compensation amount according to the predicted position drift vector; generating a compensation control signal according to the beam path compensation amount to adjust the laser beam path to achieve dynamic focus compensation.

[0009] Preferably, the material signal registration unit specifically includes: Adopting a time-division multiplexing strategy to sequentially collect the Raman scattering spectrum, second harmonic signal, and laser power at each scanning point to obtain the material analysis signal, and recording a time stamp for each collection; Performing preprocessing on the material analysis signal, and using the time stamp and the position coordinates, mapping the preprocessed material analysis signal to a unified spatial coordinate system through an interpolation algorithm to obtain the registered material analysis signal.

[0010] Preferably, constructing a dual-channel analysis network to process the registered material analysis signal and obtaining a spectrum-structure relationship model specifically includes: Organizing the registered material analysis signals at the same scanning point into a third-order tensor, and inputting the third-order tensor into the dual-channel analysis network; the dual-channel analysis network includes a frequency domain processing path, a spatial domain processing path, a feature fusion layer, and an output layer; The frequency domain processing path uses a one-dimensional convolutional layer to process the third-order tensor along the spectral wave number to extract spectral features and obtains a frequency domain feature vector; The spatial domain processing path uses a two-dimensional convolutional layer to process the third-order tensor along the spatial dimension to extract material structure features, obtaining a spatial domain feature vector; The feature fusion layer uses an attention mechanism to merge the frequency domain feature vector and the spatial domain feature vector, obtaining a fused feature vector; The output layer uses a fully connected layer to map the fused feature vector to the predicted value of the material microstructure parameters, constituting the spectral-structure relationship model.

[0011] Preferably, applying wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify background interference signals includes: Performing j-layer wavelet packet decomposition on the registered Raman scattering spectrum to obtain wavelet coefficients of 2 j frequency bands; Calculating the similarity between the wavelet coefficients of each frequency band and the typical background wavelet coefficients, and identifying the frequency bands with similarity greater than the similarity threshold to obtain a background frequency band set; Using the wavelet coefficients corresponding to the background frequency band set to perform inverse wavelet packet transform to reconstruct the background interference signal.

[0012] Preferably, determining the background contribution weight in combination with the KL divergence, and obtaining a clean Raman scattering spectrum based on the background contribution weight and the background interference signal includes: Normalizing the registered Raman scattering spectrum and the background interference signal into a first probability distribution and a second probability distribution respectively; Calculating the KL divergence value between the first probability distribution and the second probability distribution, and calculating the initial background contribution weight based on the KL divergence value; Establishing an optimization process with the goal of maximizing the signal-to-noise ratio of the target characteristic peak, iteratively adjusting the initial background contribution weight, and obtaining the background contribution weight when the iteration stop condition is reached; Calculating the clean Raman scattering spectrum according to the registered Raman scattering spectrum, the background contribution weight, and the background interference signal.

[0013] A multi-dimensional Raman spectrum in-situ analysis method based on high-speed line scanning includes: Collecting the position coordinates of the nano-displacement stage to calculate the position drift vector, combining with Kalman filtering to obtain the predicted position drift vector, and performing dynamic focus compensation based on the predicted position drift vector; Obtaining the material analysis signal and performing preprocessing and registration to obtain the registered material analysis signal; the material analysis signal includes Raman scattering spectrum, second harmonic signal, and laser power; Constructing a dual-channel analysis network to process the registered material analysis signal to obtain a spectral-structure relationship model; Wavelet packet decomposition is applied to the registered Raman scattering spectrum in the analysis signal of the registration material to identify background interference signals. The KL divergence is combined to determine the background contribution weight, and a clean Raman scattering spectrum is obtained based on the background contribution weight and the background interference signal. Based on the spectral-structure relationship model, the analysis signal of the registration material, and the clean Raman scattering spectrum, the material structure analysis result and the grain boundary analysis result are obtained.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By collecting real-time position coordinates, calculating the position drift vector, and then combining the Kalman filtering algorithm to construct the state equation and the observation equation, the predicted position drift vector is recursively calculated through the prediction step and the update step. Based on the predicted position drift vector, the system can calculate the beam path compensation amount and generate a compensation control signal to automatically adjust the laser beam path and achieve dynamic focus compensation. This predictive compensation mechanism can anticipate upcoming drifts in advance and reduce the system response delay. By correcting the laser focus position in real time, the spectral acquisition quality under high-speed scanning conditions is ensured. Especially in the microstructural change regions such as material grain boundaries, the problem of missed reports in grain boundary detection caused by focus drift is effectively solved, providing a stable and reliable raw data basis.

[0015] 2. Raman scattering spectra, second harmonic signals, and laser power are sequentially collected at each scanning point, and an accurate timestamp is recorded for each acquisition to establish an accurate correspondence between the data and the acquisition time and spatial position. By using the timestamp and position coordinate information and adopting the interpolation algorithm, the preprocessed material analysis signal is mapped to a unified spatial coordinate system, realizing the accurate spatial registration of multi-dimensional data. Accurate data registration eliminates the spatial inconsistency caused by the difference in data acquisition time, provides a high-quality registered data set for subsequent relationship analysis, and enables the system to accurately characterize the correlation between material structure parameters and nonlinear responses.

[0016] 3. The dual-channel analysis network extracts the frequency-domain feature vector and the spatial-domain feature vector through the frequency-domain processing path and the spatial-domain processing path respectively, and then combines these two feature vectors through the feature fusion layer using the attention mechanism to obtain the fused feature vector. In terms of interference separation, the system applies wavelet packet decomposition to the registered Raman scattering spectrum to identify background interference signals, determines the background contribution weight by calculating the KL divergence, and based on the optimization strategy of maximizing the signal-to-noise ratio of the target characteristic peak, effectively separates the background interference signals from the original spectrum to obtain a clean Raman scattering spectrum. Combining the dual-channel analysis network and the wavelet packet decomposition technology improves the analysis ability of the system in a high-noise environment, solves the problem of the decrease in the signal-to-noise ratio of the raw data caused by the high scanning rate, and provides high-quality data support for the accurate analysis of material structures and grain boundaries. Description of the Drawings

[0017] Figure 1 Schematic structural diagram of the multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning of the present invention; Figure 2 Schematic structural diagram of the dual-channel analysis network of the present invention; Figure 3 Schematic flow diagram of the multi-dimensional Raman spectroscopy in-situ analysis method based on high-speed line scanning of the present invention. Specific embodiments

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0019] Please refer to Figures 1 to 3 , the present invention provides a multi-dimensional Raman spectroscopy in-situ analysis system and method based on high-speed line scanning, and the technical solutions are as follows: Embodiment 1: As Figure 1 shown, the multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning includes: A laser focus stabilization unit that collects the position coordinates of the nano-displacement stage to calculate the position drift vector, combines Kalman filtering to obtain the predicted position drift vector, and performs dynamic focus compensation based on the predicted position drift vector; A material signal registration unit that acquires the material analysis signal and performs preprocessing and registration to obtain the registered material analysis signal; the material analysis signal includes Raman scattering spectrum, second harmonic signal, and laser power; An association analysis unit that constructs a dual-channel analysis network to process the registered material analysis signal and obtains a spectrum-structure relationship model; A background correction unit that applies wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal, combines the KL divergence to determine the background contribution weight, and obtains the clean Raman scattering spectrum based on the background contribution weight and the background interference signal; A material property analysis unit that obtains the material structure analysis result and the grain boundary analysis result based on the spectrum-structure relationship model, the registered material analysis signal, and the clean Raman scattering spectrum.

[0020] Furthermore, the material property analysis unit specifically includes: obtaining the distribution of material structure parameters through a spectrum-structure relationship model; performing peak fitting on the clean Raman scattering spectrum to extract the peak parameter distribution, where the peak parameters include peak position, integrated intensity, and full width at half maximum; combining the material structure parameter distribution and the peak parameter distribution to obtain the material structure analysis result; applying an edge detection algorithm to process the gradient of the peak parameter distribution in space, and at the same time using the registered second harmonic signal in the registered material analysis signal as an aid to obtain the grain boundary analysis result.

[0021] Furthermore, combining the Kalman filter to obtain the predicted position drift vector, and performing dynamic focus compensation based on the predicted position drift vector includes: using the Kalman filter algorithm to process the position drift vector to construct a state equation and an observation equation; according to the state equation and the observation equation, recursively calculating the predicted position drift vector through a prediction step and an update step; calculating the beam path compensation amount according to the predicted position drift vector, and generating a compensation control signal according to the beam path compensation amount to adjust the laser beam path to achieve dynamic focus compensation.

[0022] Specifically, the piezoelectric ceramic position sensor data integrated on the nano-displacement stage is used to monitor the position coordinates in real time, and the difference between the real-time monitored position coordinates and the preset ideal scanning trajectory coordinates is calculated to obtain the position drift vector in three-dimensional space; The state equation describes the evolution of the state vector over time. The state vector includes the position deviation vector and velocity. The observation equation describes the relationship between the observation vector and the state vector. The observation vector is obtained by multiplying the observation matrix by the state vector and adding the observation noise. The observation matrix is used to map the state vector to the measurement space; At the current moment of the beam path compensation amount where, is the optical system magnification matrix, which converts the position drift of the nano-displacement stage into the required beam path adjustment, depends on the design parameters of the optical system and is determined through experimental calibration; is the predicted position drift vector output by the Kalman filter; is the predicted time step of the Kalman filter; through the response characteristic curve of the optical deflector (this curve describes the relationship between the input control signal and the generated physical deflection amount), the beam path compensation amount is converted into the corresponding compensation control signal, and a high-speed digital-to-analog converter is used to convert the digitized compensation control signal into an analog signal to drive the optical deflector to deflect the laser beam accordingly, so as to adjust the laser focus position in real time and offset the drift of the nano-displacement stage.

[0023] By constructing the state equation and the observation equation, the predicted position drift vector is calculated recursively, and a compensation control signal is generated accordingly to adjust the laser beam path. This predictive compensation mechanism can track and correct the focus drift in real time, has a lower system response delay than traditional feedback compensation, can effectively eliminate the laser focus drift caused by the vibration of the nano-displacement stage, and ensures the spectral acquisition stability under high-speed scanning conditions.

[0024] Further, the material signal registration unit specifically includes: The Raman scattering spectrum, the second harmonic signal, and the laser power are sequentially collected at each scanning point by using a time-division multiplexing strategy to obtain the material analysis signal, and a timestamp is recorded for each collection; The material analysis signal is preprocessed, and using the timestamp and the position coordinates, the preprocessed material analysis signal is mapped to a unified spatial coordinate system through an interpolation algorithm to obtain the registered material analysis signal.

[0025] Specifically, by using the known timestamp and the accurate position coordinates after dynamic focus compensation, the material analysis signal values in different dimensions are mapped to the unified spatial coordinate system, that is, the predefined regular grid points, through mathematical interpolation, so as to achieve spatial registration.

[0026] By collecting multi-dimensional data through a time-division multiplexing strategy and recording timestamps, and combining the interpolation algorithm to map the data to a unified spatial coordinate system, high-precision multi-dimensional data spatial registration is achieved. By fully utilizing the time information and the real-time position coordinates after performing dynamic focus compensation to assist in registration, the registration accuracy is significantly improved, laying a foundation for accurately characterizing the correlation between material structure parameters and non-linear responses.

[0027] Further, constructing a dual-path analysis network to process the registered material analysis signal to obtain the spectral-structure relationship model specifically includes: The registered material analysis signals at the same scanning point are organized into a third-order tensor, and the third-order tensor is input into the dual-path analysis network; as Figure 2 shown, the dual-path analysis network includes a frequency domain processing path, a spatial domain processing path, a feature fusion layer, and an output layer; The frequency domain processing path uses a one-dimensional convolutional layer to process the third-order tensor along the spectral wave number to extract spectral features and obtain a frequency domain feature vector; The spatial domain processing path uses a two-dimensional convolutional layer to process the third-order tensor along the spatial dimension to extract material structure features and obtain a spatial domain feature vector; The feature fusion layer uses an attention mechanism to merge the frequency domain feature vector and the spatial domain feature vector to obtain a fused feature vector; The output layer uses a fully connected layer to map the fused feature vector to the predicted value of the material microstructure parameters to form the spectral-structure relationship model.

[0028] The microstructural parameters of the material mainly include: the number of layers (characterizing the thickness of the two-dimensional material, such as monolayer, bilayer or multilayer state), the stacking angle (describing the relative rotation angle between adjacent layers in the multilayer material, which affects the electronic and optical properties of the material), the lattice constant (characterizing the basic parameter of the lattice structure and reflecting the atomic spacing), the strain distribution (quantifying the degree of local tensile or compressive deformation of the material, usually expressed as a percentage), and the grain boundary orientation (describing the structural characteristics at the junction of adjacent grains), etc. These microstructural parameters jointly determine the electronic, optical, thermal and mechanical properties of the two-dimensional material. By establishing a quantitative relationship between them and the spectral characteristics, non-destructive and rapid characterization of the material microstructure can be achieved.

[0029] The dual-path analysis network extracts features through two processing paths in the frequency domain and the spatial domain respectively, and uses the attention mechanism to fuse the features. This dual-path parallel processing structure can capture frequency domain and spatial domain information simultaneously, more comprehensively mine data features, improve the robustness of the model to noise, and can still accurately establish the relationship model between the microstructural parameters of the material and the spectral characteristics under the condition of low signal-to-noise ratio caused by high-speed scanning.

[0030] Furthermore, applying wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify background interference signals includes: Performing j-layer wavelet packet decomposition on the registered Raman scattering spectrum to obtain wavelet coefficients of 2 j frequency bands; Calculating the similarity between the wavelet coefficients of each frequency band and the typical background wavelet coefficients, and identifying the frequency bands with similarity greater than the similarity threshold to obtain the background frequency band set; Using the wavelet coefficients corresponding to the background frequency band set to perform inverse wavelet packet transform to reconstruct the background interference signal.

[0031] Specifically, in the substrate area without sample coverage, collecting the Raman spectrum to obtain the typical background, performing j-layer wavelet packet decomposition on the typical background to obtain the typical background wavelet coefficients of 2 j frequency bands, and calculating the similarity between the wavelet coefficients of each frequency band and the typical background wavelet coefficients of the corresponding frequency band.

[0032] By performing multi-layer wavelet packet decomposition on the registered Raman scattering spectrum, calculating the similarity between each frequency band and the typical background, identifying the background frequency band and reconstructing the background interference signal. This method can accurately separate complex background noise, is not affected by the change of background shape and intensity, significantly improves the anti-interference ability of the system in a complex sample environment, and provides technical guarantee for obtaining high-quality clean Raman scattering spectrum.

[0033] Further, determining the background contribution weight in combination with the KL divergence, and obtaining the clean Raman scattering spectrum based on the background contribution weight and the background interference signal includes: Normalize the registered Raman scattering spectrum and the background interference signal into a first probability distribution and a second probability distribution, respectively; Calculate the KL divergence value between the first probability distribution and the second probability distribution , and calculate the initial background contribution weight based on the KL divergence value ; where is the initial background contribution weight of the scanning point ; is the natural exponential function; is an adjustable smoothing factor; Establish an optimization process aiming to maximize the signal-to-noise ratio of the target characteristic peak, iteratively adjust the initial background contribution weight, and obtain the background contribution weight when the iteration stop condition is reached; Calculate the clean Raman scattering spectrum according to the registered Raman scattering spectrum, the background contribution weight, and the background interference signal ; where is the registered Raman scattering spectrum, and its specific meaning is the Raman scattering spectrum signal intensity varying with the wavenumber at the scanning point ; is the background contribution weight of the scanning point ; is the background interference signal.

[0034] Specifically, the gradient ascent method is used for optimization, and the iteration stop conditions include: the increment of the signal-to-noise ratio of the target characteristic peak between two iterations is less than a preset increment threshold, or the number of iterations reaches a preset maximum number of iterations.

[0035] The KL (Kullback-Leibler) divergence is an asymmetric measure of the difference between two probability distributions. By converting the spectrum into a probability distribution to calculate the KL divergence and then optimizing based on maximizing the signal-to-noise ratio of the target characteristic peak, the background contribution weight is accurately determined, thereby obtaining the clean Raman scattering spectrum. This method avoids the overcorrection or undercorrection problems in traditional background removal, can retain the characteristic peaks of the original signal while effectively removing background interference, improves the accuracy of spectral analysis, and is applicable to the processing of low signal-to-noise ratio data under high-speed scanning conditions.

[0036] Through the collaborative work of five functional units, the present invention constructs an efficient and reliable in-situ analysis system for multi-dimensional Raman scattering spectroscopy. The laser focus stabilization unit uses the Kalman filtering algorithm to achieve dynamic focus compensation, effectively solving the problem of the excitation light focus drift caused by the vibration of the nano-displacement stage; the material signal registration unit adopts a time-division multiplexing strategy and an interpolation algorithm based on timestamps to achieve high-precision spatial registration of multi-dimensional data; the correlation analysis unit and the background correction unit respectively extract and fuse features from two dimensions of the frequency domain and the spatial domain through a dual-channel convolutional neural network and wavelet packet decomposition technology, and at the same time effectively separate the background interference signals, jointly solving the problem of the decrease in the signal-to-noise ratio of the original data caused by the high scanning rate. The present invention realizes the precise analysis of the material structure and grain boundaries under high-speed scanning conditions, effectively overcoming the problems of missed reports in grain boundary detection, data registration errors, and signal-to-noise ratio reduction.

[0037] Embodiment 2: As Figure 3 shown, the present invention also provides a multi-dimensional Raman spectroscopy in-situ analysis method based on high-speed line scanning, including: Collect the position coordinates of the nano-displacement stage to calculate the position drift vector, combine with the Kalman filter to obtain the predicted position drift vector, and perform dynamic focus compensation based on the predicted position drift vector; Obtain the material analysis signal and perform preprocessing and registration to obtain the registered material analysis signal; the material analysis signal includes Raman scattering spectrum, second harmonic signal, and laser power; Construct a dual-channel analysis network to process the registered material analysis signal to obtain a spectrum-structure relationship model; Apply wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal, determine the background contribution weight in combination with the KL divergence, and obtain the clean Raman scattering spectrum based on the background contribution weight and the background interference signal; Based on the spectrum-structure relationship model, the registered material analysis signal, and the clean Raman scattering spectrum, obtain the material structure analysis result and the grain boundary analysis result.

[0038] Further, combining with the Kalman filter to obtain the predicted position drift vector and performing dynamic focus compensation based on the predicted position drift vector includes: using the Kalman filtering algorithm to process the position drift vector, constructing a state equation and an observation equation; according to the state equation and the observation equation, recursively calculate the predicted position drift vector through the prediction step and the update step; calculate the beam path compensation amount according to the predicted position drift vector, generate a compensation control signal according to the beam path compensation amount to adjust the laser beam path, and realize dynamic focus compensation.

[0039] Further, obtaining the material analysis signal and performing preprocessing and registration to obtain the registered material analysis signal specifically includes: Adopt a time-division multiplexing strategy to sequentially collect Raman scattering spectra, second harmonic signals, and laser power at each scanning point to obtain material analysis signals, and record a timestamp for each collection. Preprocess the material analysis signals, and use the timestamp and the position coordinates to map the preprocessed material analysis signals to a unified spatial coordinate system through an interpolation algorithm to obtain the registered material analysis signals.

[0040] Furthermore, construct a dual-path analysis network to process the registered material analysis signals to obtain a spectral-structure relationship model, which specifically includes: Organize the registered material analysis signals at the same scanning point into a third-order tensor, and input the third-order tensor into the dual-path analysis network; the dual-path analysis network includes a frequency-domain processing path, a spatial-domain processing path, a feature fusion layer, and an output layer; The frequency-domain processing path uses a one-dimensional convolutional layer to process the third-order tensor along the spectral wave number to extract spectral features and obtain a frequency-domain feature vector; The spatial-domain processing path uses a two-dimensional convolutional layer to process the third-order tensor along the spatial dimension to extract material structure features and obtain a spatial-domain feature vector; The feature fusion layer uses an attention mechanism to combine the frequency-domain feature vector and the spatial-domain feature vector to obtain a fused feature vector; The output layer uses a fully connected layer to map the fused feature vector to the predicted value of the material microstructure parameters to form the spectral-structure relationship model.

[0041] This embodiment mainly aims at the high-speed line-scanning Raman analysis application scenario of two-dimensional transition metal chalcogenide materials. Table 1 describes the performance of the dual-path analysis network of the present invention. Compared with a single-path network that only uses frequency-domain information, the dual-path network of the present invention has significantly improved performance in various performance indicators. Two-dimensional materials can exist in monolayer or multilayer forms, and the number of layers has a great influence on the material properties. The 98.7% layer prediction accuracy means that the system can have high accuracy in judging the number of layers of the material. The stacking angle is the rotation angle between adjacent layers in multilayer two-dimensional materials, and the average deviation between the stacking angle predicted by the system and the actual value is about 0.86 degrees.

[0042] Table 2 shows the performance comparison between the background correction method of the present invention and other methods. Among them, polynomial fitting is a traditional background correction method, which fits the background signal with a polynomial function and then subtracts the fitted background from the original signal. The sliding window minimum value finds the minimum point in each window by sliding a window on the spectrum and then connecting these points to form a background estimate; the results show that the method of the present invention is significantly superior to the traditional method in terms of signal-to-noise ratio improvement and peak position accuracy. The peak position deviation value can reflect the precision of the system in measuring the Raman peak position, in units of wave number (cm -1） , with ±0.12 cm -1It means that the deviation between the peak position measured by the system and the true value is within 0.12 wavenumbers, which is crucial for identifying specific molecular vibration modes.

[0043] Table 1 Performance Evaluation of Dual-Path Analysis Network Performance metrics The dual-channel network of the present invention Single-channel network (only frequency-domain processing path) Accuracy of layer number prediction 98.7% 92.3% Stacking angle prediction error 0.86° 2.57° Table 2 Performance Comparison of Background Correction Methods Performance metrics Signal-to-noise ratio improvement (dB) <![CDATA[Peak position deviation value (cm -1 )]]> The method of the present invention 18.3 ±0.12 Polynomial fitting 9.7 ±0.43 Minimum value of sliding window 10.5 ±0.44 Furthermore, applying wavelet packet decomposition to the registered Raman scattering spectrum in the analysis signal of the registration material to identify background interference signals includes: Performing j-layer wavelet packet decomposition on the registered Raman scattering spectrum to obtain wavelet coefficients of 2 j frequency bands; Calculating the similarity between the wavelet coefficients of each frequency band and the typical background wavelet coefficients, and identifying the frequency bands with similarity greater than the similarity threshold to obtain a set of background frequency bands; Using the wavelet coefficients corresponding to the set of background frequency bands for inverse wavelet packet transform to reconstruct the background interference signal.

[0044] Furthermore, determining the background contribution weight in combination with KL divergence, and obtaining a clean Raman scattering spectrum based on the background contribution weight and the background interference signal includes: Normalizing the registered Raman scattering spectrum and the background interference signal into a first probability distribution and a second probability distribution respectively; Calculating the KL divergence value between the first probability distribution and the second probability distribution, and calculating the initial background contribution weight based on the KL divergence value; Establishing an optimization process with the goal of maximizing the signal-to-noise ratio of the target characteristic peak, iteratively adjusting the initial background contribution weight, and obtaining the background contribution weight when the iteration stop condition is reached; Calculating the clean Raman scattering spectrum according to the registered Raman scattering spectrum, the background contribution weight, and the background interference signal.

[0045] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning, characterized in that: include: The laser focus stabilization unit collects the position coordinates of the nano-displacement stage to calculate the position drift vector, obtains the predicted position drift vector by combining with the Kalman filter, and performs dynamic focus compensation based on the predicted position drift vector; A material signal registration unit, which acquires material analysis signals and performs preprocessing and registration to obtain registered material analysis signals; the material analysis signals include Raman scattering spectra, second harmonic signals and laser power; The correlation analysis unit constructs a dual-path analysis network to process the registration material analysis signals and obtain a spectrum-structure relationship model; A background correction unit, which applies wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal, determines the background contribution weight in combination with the KL divergence, and obtains the clean Raman scattering spectrum based on the background contribution weight and the background interference signal; The material property analysis unit obtains the material structure analysis results and grain boundary analysis results based on the spectrum-structure relationship model, the registered material analysis signal and the clean Raman scattering spectrum.

2. The multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning according to claim 1 is characterized in that: The predicted position drift vector is obtained in combination with Kalman filtering, and dynamic focus compensation is performed based on the predicted position drift vector, including: using a Kalman filtering algorithm to process the position drift vector, and constructing a state equation and an observation equation; according to the state equation and the observation equation, the predicted position drift vector is obtained by recursive calculation through a prediction step and an update step; according to the predicted position drift vector, a beam path compensation amount is calculated, and a compensation control signal is generated according to the beam path compensation amount to adjust the laser beam path, so as to realize dynamic focus compensation.

3. The multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning according to claim 1 is characterized in that: The material signal registration unit specifically includes: A time-division multiplexing strategy is used to sequentially collect Raman scattering spectra, second harmonic signals, and laser power at each scanning point to obtain material analysis signals, and a timestamp is recorded for each acquisition; The material analysis signal is preprocessed, and the preprocessed material analysis signal is mapped to a unified spatial coordinate system by using the timestamp and the position coordinates through an interpolation algorithm to obtain the registered material analysis signal.

4. The multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning according to claim 1 is characterized in that: Construct a dual-path analysis network to process and register material analysis signals, and obtain a spectrum-structure relationship model, which specifically includes: Organizing the registered material analysis signals of the same scanning point into a third-order tensor, and inputting the third-order tensor into a two-way analysis network; the two-way analysis network includes a frequency domain processing path, a spatial domain processing path, a feature fusion layer and an output layer; The frequency domain processing path uses a one-dimensional convolution layer to process the third-order tensor along the spectral wavenumber to extract spectral features and obtain a frequency domain feature vector; The spatial domain processing path uses a two-dimensional convolutional layer to process the third-order tensor along the spatial dimension to extract material structure features and obtain a spatial domain feature vector; The feature fusion layer uses an attention mechanism to merge the frequency domain feature vector and the spatial domain feature vector to obtain a fused feature vector; The output layer uses a fully connected layer to map the fused feature vector to the predicted value of the material microstructure parameter to form the spectrum-structure relationship model.

5. The multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning according to claim 1 is characterized in that: Applying wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal includes: The registered Raman scattering spectrum is decomposed by j layers of wavelet packets to obtain 2 j The wavelet coefficients of frequency bands; Calculate the similarity between the wavelet coefficient of each frequency band and the typical background wavelet coefficient, identify the frequency bands whose similarity is greater than the similarity threshold to obtain the background frequency band set; The background interference signal is reconstructed by performing inverse wavelet packet transform using the wavelet coefficients corresponding to the background frequency band set.

6. The multi-dimensional Raman spectroscopy in-situ analysis system based on high-speed line scanning according to claim 1 is characterized in that: Combining KL divergence to determine the background contribution weight, obtaining a clean Raman scattering spectrum based on the background contribution weight and the background interference signal includes: Normalizing the registered Raman scattering spectrum and the background interference signal to a first probability distribution and a second probability distribution, respectively; Calculate the KL divergence value between the first probability distribution and the second probability distribution, and calculate the initial background contribution weight based on the KL divergence value; Establishing an optimization process with the goal of maximizing the signal-to-noise ratio of the target characteristic peak, iteratively adjusting the initial background contribution weight, and obtaining the background contribution weight when the iteration stop condition is reached; The clean Raman scattering spectrum is calculated according to the registered Raman scattering spectrum, the background contribution weight and the background interference signal.

7. A multi-dimensional Raman spectroscopy in-situ analysis method based on high-speed line scanning, characterized in that: include: Collect the position coordinates of the nano-displacement stage to calculate the position drift vector, combine with Kalman filtering to obtain the predicted position drift vector, and perform dynamic focus compensation based on the predicted position drift vector; Acquire material analysis signals and perform preprocessing and registration to obtain registered material analysis signals; the material analysis signals include Raman scattering spectrum, second harmonic signal and laser power; A dual-path analysis network was constructed to process the analysis signals of the registration materials and obtain a spectrum-structure relationship model; Apply wavelet packet decomposition to the registered Raman scattering spectrum in the registered material analysis signal to identify the background interference signal, determine the background contribution weight in combination with KL divergence, and obtain the clean Raman scattering spectrum based on the background contribution weight and the background interference signal; The material structure analysis results and grain boundary analysis results are obtained based on the spectrum-structure relationship model, the registered material analysis signal and the clean Raman scattering spectrum.

Images