Optical material defect depth detection method based on spectrum

Abstract

The invention discloses a spectrum-based optical material defect depth detection method, which belongs to the technical field of defect detection, and comprises the following steps: obtaining hyperspectral image characteristic parameters of a to-be-detected optical material, and determining whether a defect exists or not based on the hyperspectral image characteristic parameters; if no defect exists, ending the detection; if the defect exists, determining a preliminary defect area and a defect type; obtaining Raman spectrum characteristic parameters of the preliminary defect area, and determining a final defect area based on the Raman spectrum characteristic parameters; performing terahertz spectrum depth scanning in the final defect area to obtain terahertz spectrum characteristic parameters, and constructing a defect depth information atlas; and performing defect depth inversion based on the hyperspectral image characteristic parameters, the Raman spectrum characteristic parameters and the terahertz spectrum characteristic parameters to obtain a defect depth prediction result. The method solves the problem that the traditional method is difficult to accurately detect the depth of the internal defect of the optical material, and has high precision, comprehensiveness and adaptability.

Description

Technical Field

[0001] The present invention belongs to the technical field of defect detection, and specifically relates to a method for detecting the depth of optical material defects based on spectroscopy. Background Art

[0002] Optical materials are a type of materials that can regulate the propagation direction, intensity, polarization state, spectral distribution, etc. of light to meet the specific functional requirements of optical systems. For example, CN114136979A discloses a device and method for detecting surface micro-defects of volume scattering optical materials. By adjusting the polarizer, the illumination light source is in a specific polarization mode, and the incident angle of the illumination light source can be adjusted through a bracket to reduce the transmittance of the illumination light source, thereby reducing the influence of the volume scattering light of the sample on the detection of surface micro-defects of the material, so that the microscope provided directly above the sample holder can clearly observe the defects on the surface of the volume scattering optical material sample.

[0003] However, the scattering centers inside the volume scattering optical materials in the existing optical material defect detection methods will cause the incident light to be scattered multiple times, resulting in the reflection / scattering signals of surface micro-defects being easily submerged by internal scattered light. Moreover, the coupling of defect depth and light penetration depth will cause signal superposition. At the same time, the anisotropy and inhomogeneity of optical properties also make the signals show non-linear characteristics, making it difficult to accurately distinguish defect signals and reverse the depth through the ratio of reflectivity to transmittance at a single incident angle. Therefore, there are problems in accurately detecting the defects and depths of optical materials. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a method for detecting the depth of optical material defects based on spectroscopy. By integrating hyperspectral imaging, Raman spectroscopy, and terahertz spectroscopy technologies and combining with deep neural network analysis, it can accurately identify and locate the defects and their depths of optical materials, solve the problem that traditional methods are difficult to accurately detect the depths of internal defects of optical materials, and has high precision, comprehensiveness, and adaptability.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for detecting the depth of optical material defects based on spectroscopy, comprising the following steps: Obtain the hyperspectral image feature parameters of the optical material to be detected, and determine whether there are defects based on the hyperspectral image feature parameters; If there are no defects, the detection ends; If there are defects, determine the preliminary defect area and defect type; Obtain the Raman spectral feature parameters of the preliminary defect area, and determine the final defect area based on the Raman spectral feature parameters; Perform terahertz spectral depth scanning in the final defect area to obtain terahertz spectral feature parameters, and construct a defect depth information map; Based on the hyperspectral image feature parameters, Raman spectrum feature parameters, and terahertz spectrum feature parameters, defect depth inversion is performed to obtain the defect depth prediction result.

[0006] Preferably, the hyperspectral image feature parameters include reflectivity, peak intensity, full width at half maximum, and inter-band correlation coefficient; the Raman spectrum feature parameters include Raman spectrum peak position, Raman spectrum peak intensity, and Raman spectrum full width at half maximum; the terahertz spectrum feature parameters include terahertz reflection spectrum reflection peak position, terahertz reflection spectrum intensity, and terahertz reflection spectrum phase change.

[0007] Preferably, determining whether there are defects based on the hyperspectral image feature parameters includes the following steps: The optical material to be detected moves on the motion control platform of the hyperspectral image data device according to a preset scanning path, and the hyperspectral image data of each coordinate point on the optical material to be detected is obtained based on the hyperspectral image data device; After preprocessing the hyperspectral image data, principal component analysis is performed for dimensionality reduction to obtain the hyperspectral image feature parameters of each coordinate point; The hyperspectral image feature parameters of each coordinate point are averaged to obtain the homogenized hyperspectral image feature parameters; The homogenized hyperspectral image feature parameters are subjected to similarity analysis with the standard hyperspectral image feature parameters stored in the database to obtain the feature cosine similarity value; If the feature cosine similarity value is greater than the set similarity threshold, there are no defects; If the feature cosine similarity value is not greater than the set similarity threshold, there are defects.

[0008] Preferably, when there are defects, the process of determining the preliminary defect area and defect type is as follows: The hyperspectral image feature parameters of each coordinate point are subjected to similarity analysis with the standard hyperspectral image feature parameters stored in the database to obtain the coordinate point feature cosine similarity value; If the coordinate point feature cosine similarity value of a certain coordinate point is greater than the set coordinate point feature similarity threshold, then this coordinate point is not recorded as a defect coordinate point; If the coordinate point feature cosine similarity value of a certain coordinate point is not greater than the set coordinate point feature similarity threshold, then this coordinate point is recorded as a defect coordinate point; All the defect coordinate points are connected to determine the preliminary defect area; The hyperspectral image feature parameters of each defect coordinate point in the preliminary defect area are averaged to obtain the homogenized hyperspectral image feature parameters of the defect area; The homogenized hyperspectral image feature parameters of the defect area are used as the input of the trained support vector machine model to obtain the defect type.

[0009] Preferably, determining the final defect area based on the Raman spectral characteristic parameters includes the following steps: Performing Raman spectral scanning on the preliminary defect area by a Raman spectrometer to obtain Raman spectral data corresponding to each defect coordinate point; Performing feature processing on the Raman spectral data corresponding to each defect coordinate point to obtain Raman spectral characteristic parameters corresponding to each defect coordinate point; Determining a machine learning model; Taking the Raman spectral characteristic parameters corresponding to each defect coordinate point as the input of the machine learning model, outputting the probability value of each defect coordinate point belonging to the defect type, and recording the defect coordinate points with probability values greater than the probability threshold as the final defect area coordinate points; Connecting all the final defect area coordinate points to determine the final defect area.

[0010] Preferably, the method for determining the machine learning model is as follows: Determining the molecular structure composition of the preliminary defect area based on the Raman spectral characteristic parameters corresponding to each defect coordinate point; Determining the difference in Raman peak positions before and after stress action for each defect coordinate point, denoted as the Raman shift, and determining the stress distribution of the preliminary defect area based on the stress-Raman shift correlation formula; Determining the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area.

[0011] Preferably, determining the molecular structure composition of the preliminary defect area includes the following steps: Comparing the Raman spectral characteristic parameters corresponding to each defect coordinate point with a known molecular vibration mode database one by one to determine the molecular structure composition of the defect area at each coordinate point; The stress-Raman shift correlation formula is , is the Raman shift, is the stress coefficient, is the stress.

[0012] Preferably, determining the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area includes the following steps: Obtaining the historical molecular structure composition and stress distribution combinations corresponding to the defect types, and performing clustering analysis to obtain multiple clustering points, each clustering point corresponding to a historical molecular structure composition and stress distribution combination; Presetting multiple groups of machine learning model parameters for each clustering point; Under the combination of historical molecular structure composition and stress distribution, the historical Raman spectral characteristic parameters corresponding to the defect type are used as the input of the machine learning models corresponding to each group of machine learning model parameters, and the boundary coordinates of the corresponding defect region are output. The accuracy of the machine learning models corresponding to each group of machine learning model parameters is calculated, and the machine learning model corresponding to the group of machine learning model parameters with the highest accuracy is selected; Based on the machine learning model corresponding to this group of machine learning model parameters, training is carried out to obtain machine learning models corresponding to different combinations of historical molecular structure composition and stress distribution for different defect types; Perform similarity analysis on the molecular structure composition and stress distribution of the preliminary defect region and each group of historical molecular structure composition and stress distribution combinations under the corresponding defect type to determine the most similar historical molecular structure composition and stress distribution combination, and obtain the corresponding machine learning model.

[0013] Preferably, obtaining terahertz spectral characteristic parameters and constructing a defect depth information map includes the following steps: Start the depth scanning program of the terahertz time-domain spectrometer. The motion control platform of the terahertz time-domain spectrometer controls the optical material to be detected to move in the Z-axis direction at a preset scanning interval. After moving a set distance each time, it pauses and collects the terahertz reflection spectral data at this depth point; Extract the characteristics of the terahertz reflection spectral data at each depth point to obtain the terahertz spectral characteristic parameters at each depth point; Based on three-dimensional visualization software, with the depth as the Z-axis and the X and Y coordinates of the final defect region as the plane axes, perform visual mapping on the terahertz spectral characteristic parameters at each depth point to generate a defect depth information map of the defect region.

[0014] Preferably, obtaining the defect depth prediction result includes the following steps: Merge the hyperspectral image characteristic parameters, Raman spectral characteristic parameters, and terahertz spectral characteristic parameters according to the same spatial coordinates to form multimodal spectral characteristic data at each spatial position point; Perform standardization processing on the multimodal spectral characteristic data to obtain the standardized multimodal spectral characteristics; Use the standardized multimodal spectral characteristics as the input of the trained deep neural network, output the defect depth prediction values at each spatial position point, and summarize them to form a defect depth prediction result.

[0015] The present invention has the following beneficial effects: The present invention determines the presence, type, and preliminary region of defects through hyperspectral images, uses Raman spectroscopy to accurately delimit the final defect region, combines terahertz spectroscopy for depth scanning to construct a defect depth information map, and then fuses multi-spectral characteristic parameters to invert the defect depth through a deep neural network, realizing a full-process detection from defect identification to depth positioning, improving the accuracy and comprehensiveness of detection, and solving the problem in the prior art that it is difficult to accurately detect the defects and depths of optical materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flowchart of the method of the present invention; Figure 2 is a flowchart for determining the final defect region of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] Example 1: As Figure 1 shown, a method for detecting the depth of defects in optical materials based on spectroscopy includes the following steps: Obtain the hyperspectral image characteristic parameters of the optical material to be detected, and determine whether there are defects based on the hyperspectral image characteristic parameters; If there are no defects, the detection ends; If there are defects, determine the preliminary defect region and the type of defect; Obtain the Raman spectroscopic characteristic parameters of the preliminary defect region, and determine the final defect region based on the Raman spectroscopic characteristic parameters; Perform terahertz spectroscopic depth scanning on the final defect region to obtain terahertz spectroscopic characteristic parameters, and construct a defect depth information map; Perform defect depth inversion based on the hyperspectral image characteristic parameters, Raman spectroscopic characteristic parameters, and terahertz spectroscopic characteristic parameters to obtain the defect depth prediction result.

[0018] The hyperspectral image characteristic parameters include reflectivity, peak intensity, full width at half maximum, and inter-band correlation coefficient; the Raman spectroscopic characteristic parameters include Raman spectral peak position, Raman spectral peak intensity, and Raman spectral full width at half maximum; the terahertz spectroscopic characteristic parameters include terahertz reflection spectral peak position, terahertz reflection spectral intensity, and terahertz reflection spectral phase change.

[0019] Determining whether there are defects based on the hyperspectral image characteristic parameters includes the following steps: The optical material to be detected moves along a preset scanning path (such as grid scanning) on the motion control platform of the hyperspectral image data device, and hyperspectral image data of each coordinate point on the optical material to be detected is obtained based on the hyperspectral image data device; the sample is driven by the motion control platform to move along the preset path to ensure full surface coverage of the sample. Each coordinate point corresponds to a spatial resolution of 50μm / pixel, and a three-dimensional data cube containing spatial and spectral information is obtained, that is, the hyperspectral image data of each coordinate point. The full coordinate point data provides an accurate spatial index for subsequent defect localization, ensuring traceability of each suspicious point.

[0020] After preprocessing the hyperspectral image data (eliminating uneven illumination through radiometric correction, removing noise through median filtering, and improving the signal-to-noise ratio of spectral data), principal component analysis is performed for dimensionality reduction to obtain the hyperspectral image feature parameters of each coordinate point. The preprocessing solves the problem of "feature distortion caused by noise interference in the original data" and ensures the reliability of subsequent analysis; PCA dimensionality reduction eliminates redundant information, highlights defect-sensitive features (such as principal component 1 reflecting the overall reflectance trend), improves computational efficiency, and avoids the impact of the "curse of dimensionality" on model performance.

[0021] The hyperspectral image feature parameters of each coordinate point are averaged to obtain the homogenized hyperspectral image feature parameters; the homogenized hyperspectral image feature parameters are analyzed for similarity with the standard hyperspectral image feature parameters stored in the database to obtain the feature cosine similarity value; if the feature cosine similarity value is greater than the set similarity threshold, there are no defects; if the feature cosine similarity value is not greater than the set similarity threshold, there are defects.

[0022] The averaging process filters out local outliers, avoids misjudging the overall state due to individual noise points, and solves the problem of "false alarms caused by accidental interference"; the similarity judgment is quantified (such as setting the threshold to 0.9), converting defect recognition into a computable numerical comparison, standardizing the detection process, and avoiding the subjectivity of manual interpretation.

[0023] Through the process of "full coordinate acquisition → noise reduction and dimensionality reduction → overall comparison", a "pre-screening mechanism" for optical material defects is constructed. By excluding defect-free samples and focusing on suspicious areas, it saves computing power and time for subsequent Raman spectroscopy fine analysis and terahertz depth scanning, and is the basic link to achieve "accurate detection".

[0024] When there are defects, the process of determining the preliminary defect area and defect type is as follows: The hyperspectral image feature parameters of each coordinate point are analyzed for similarity with the standard hyperspectral image feature parameters stored in the database to obtain the cosine similarity value of the coordinate point features. "Pixel-level positioning" of defects is achieved, and micron-scale tiny defects (such as scratches and dot-like impurities) can be detected, solving the problem of "difficulty in identifying tiny defects by traditional methods", quantifying single-point anomalies, and avoiding missed judgments caused by local feature fluctuations (such as slight contamination on the material surface).

[0025] If the cosine similarity value of the coordinate point features of a certain coordinate point is greater than the set coordinate point feature similarity threshold (such as 0.8), then this coordinate point is not recorded as a defective coordinate point; if the cosine similarity value of the coordinate point features of a certain coordinate point is not greater than the set coordinate point feature similarity threshold, then this coordinate point is recorded as a defective coordinate point; all defective coordinate points are connected to determine the preliminary defective area; a connected component analysis algorithm (such as 8-neighborhood connection) is used to merge the discrete defective coordinate points into a continuous area.

[0026] The discrete abnormal points are aggregated into physically interpretable defective areas, which conform to the characteristics of actual defects "distributed in patches", solving the problem of "misjudging scattered points as multiple defects". The preliminarily delimited preliminary defective area narrows the range for subsequent Raman spectroscopy scanning, reduces the data acquisition volume (such as only scanning the defective area instead of the entire surface), and improves the detection efficiency.

[0027] The hyperspectral image feature parameters of each defective coordinate point in the preliminary defective area are averaged to obtain the averaged hyperspectral image feature parameters of the defective area; the averaged hyperspectral image feature parameters of the defective area are used as the input of the trained support vector machine model to obtain the defect type.

[0028] The feature parameters of all coordinate points in the defective area are averaged to weaken local noise and highlight overall features; the hyperspectral data with known defect types are used for training in advance, the kernel function is selected as the radial basis function (RBF), and the parameters are optimized (such as C = 10, γ = 0.1) to improve the classification accuracy. Automatic classification based on machine learning avoids the subjectivity of manual interpretation, can distinguish multiple defect types simultaneously (such as distinguishing air bubbles and metal impurities in glass), and solves the problems of "low efficiency and high error rate in manual identification of defect types".

[0029] As Figure 2 shown, determining the final defective area based on Raman spectroscopy feature parameters includes the following steps: Perform Raman spectroscopy scanning on the preliminary defect area using a Raman spectrometer to obtain Raman spectroscopy data corresponding to each defect coordinate point; perform feature processing on the Raman spectroscopy data corresponding to each defect coordinate point to obtain Raman spectroscopy characteristic parameters corresponding to each defect coordinate point; the scanning range is limited to the determined preliminary defect area, and the scanning step is set to 0.5 μm to achieve sub-micron spatial resolution, which can capture subtle structural changes at the defect edge (such as lattice distortion at the crack tip), providing high-fidelity data for accurate boundary delineation. Feature processing includes using Savitzky-Golay filtering to remove high-frequency noise and correcting baseline drift through polynomial fitting to ensure the accuracy of spectral features (such as peak positions).

[0030] Determine the machine learning model; use the Raman spectroscopy characteristic parameters corresponding to each defect coordinate point as the input of the machine learning model, output the probability value of each defect coordinate point belonging to the defect type, and record the defect coordinate points with probability values greater than the probability threshold as the final defect area coordinate points; connect all the final defect area coordinate points to determine the final defect area.

[0031] In the training stage of the machine learning model, historical Raman data with labeled defect boundaries (such as manually marked crack edge coordinates) is used, with peak position, stress distribution, etc. as the input and defect probability as the output; the optimal threshold (such as 0.6) is determined through cross-validation to balance the recall rate (to avoid missed detections) and the precision rate (to avoid misjudgments). The machine learning model can automatically capture the non-linear relationship between spectral features and defect boundaries (such as the correlation between the stress concentration area and the crack propagation path), solving the problem that "traditional threshold methods cannot adapt to complex defect morphologies". Through the process of "high-resolution spectral acquisition → feature enhancement → intelligent boundary recognition", the upgrade of the defect area from "preliminary positioning" to "accurate delineation" is realized.

[0032] Determining the machine learning model includes the following steps: determining the molecular structure composition of the preliminary defect area based on the Raman spectroscopy characteristic parameters corresponding to each defect coordinate point; determining the difference in Raman peak positions before and after stress action for each defect coordinate point, denoted as Raman shift, and determining the stress distribution in the preliminary defect area based on the stress-Raman shift correlation formula; determining the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area.

[0033] Determining the molecular structure composition of the preliminary defect area includes the following steps: comparing the Raman spectroscopy characteristic parameters corresponding to each defect coordinate point with a known molecular vibration mode database (such as the RamanShift database) one by one to determine the molecular structure composition of the defect area at each coordinate point, and matching the chemical components (such as SiO2, SiC) or impurities (such as metal ions) in the defect area by comparing the peak positions and intensities of the measured spectrum with the standard spectrum in the database.

[0034] Analyze the essence of defects at the chemical level, distinguish different causes of similar defects (such as "oxide cracks" and "stress cracks"), and solve the problem of "poor model generalization ability due to unclear defect causes". The molecular composition information can be used as a key feature of the machine learning model to improve the model's discrimination ability for defect types (for example, areas containing oxides are more likely to be judged as corrosion defects).

[0035] The stress-Raman shift correlation formula is: , is the Raman shift, is the stress coefficient such as the 520 cm standard peak of silicon, -1 ), is the stress. Quantify the mechanical state of the defect area, identify stress concentration points (such as σ>100 MPa at the crack tip), and provide a basis for defect propagation prediction.

[0036] Determine the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area, including the following steps: Obtain the historical molecular structure composition and stress distribution combinations corresponding to the defect types, and perform clustering analysis to obtain multiple clustering points, each clustering point corresponding to a historical molecular structure composition and stress distribution combination; preset multiple groups of machine learning model parameters for each clustering point (such as the kernel function type and penalty factor C of SVM, the number of convolutional layers of CNN); preset machine learning models with different parameters for each clustering category (such as the kernel function type of SVM, the number of convolutional layers of CNN), train through historical data and calculate the accuracy rate, avoid "one-size-fits-all" modeling, dynamically call the optimal model according to different defect characteristics, and improve the boundary recognition accuracy rate.

[0037] Under the historical molecular structure composition and stress distribution combination, use the historical Raman spectral characteristic parameters corresponding to the defect type as the input of the machine learning model corresponding to each group of machine learning model parameters, output the corresponding defect area boundary coordinates, calculate the accuracy rate of the machine learning model corresponding to each group of machine learning model parameters, and select the machine learning model corresponding to the group of machine learning model parameters with the highest accuracy rate; Train based on the machine learning model corresponding to this group of machine learning model parameters to obtain machine learning models corresponding to different historical molecular structure compositions and stress distribution combinations of different defect types; perform similarity analysis on the molecular structure composition and stress distribution of the preliminary defect area and each group of historical molecular structure compositions and stress distribution combinations under the corresponding defect type to determine the most similar historical molecular structure composition and stress distribution combination, and obtain the corresponding machine learning model.

[0038] An adaptive selection system for the defect boundary recognition model is constructed through the mechanism of "data clustering - directional modeling - dynamic matching". By introducing data-driven clustering analysis and model optimization strategies, the chemical - mechanical characteristics of Raman spectroscopy are deeply integrated with machine learning algorithms, providing an efficient and interpretable technical solution for the accurate recognition of defect boundaries, which is the key technical support for "deepening and refining defect characteristics" in this embodiment.

[0039] Perform a terahertz spectroscopy depth scan on the final defect area to obtain terahertz spectroscopy characteristic parameters and construct a defect depth information map; the terahertz spectroscopy characteristic parameters include the position of the reflection peak of the terahertz reflection spectrum, the intensity of the terahertz reflection spectrum, and the phase change of the terahertz reflection spectrum.

[0040] Start the depth scan program of the terahertz time-domain spectrometer. The motion control platform of the terahertz time-domain spectrometer controls the optical material to be detected to move in the Z-axis direction according to a preset scan interval, pauses after moving a set distance each time, and collects the terahertz reflection spectrum data at this depth point.

[0041] From the material surface (Z = 0) to the preset maximum depth (such as Z = 2 mm), the movement interval is matched with the terahertz wave wavelength (such as a 50-μm interval is suitable for a 0.1-THz wavelength), ensuring that the depth resolution is better than λ / 2 (about 1.5 mm), achieving sub-millimeter-level depth positioning. By collecting data point by point in depth, the problem of "confusion between defect layers due to insufficient depth resolution in traditional terahertz imaging" (such as distinguishing surface cracks from deep delamination) is solved.

[0042] Extract the characteristics of the terahertz reflection spectrum data at each depth point to obtain the terahertz spectroscopy characteristic parameters at each depth point, including the position of the reflection peak of the terahertz reflection spectrum, the intensity of the terahertz reflection spectrum, and the phase change of the terahertz reflection spectrum; based on three-dimensional visualization software (such as ParaView), with depth as the Z-axis and the X and Y coordinates of the final defect area as the plane axes, visually map the terahertz spectroscopy characteristic parameters at each depth point to generate a defect depth information map of the defect area. This realizes the core link of "defect depth detection". Through the penetrability and depth resolution ability of terahertz spectroscopy, the internal defects of the optical material are changed from "invisible" to "quantifiable, locatable, and analyzable".

[0043] Based on the hyperspectral image characteristic parameters, Raman spectroscopy characteristic parameters, and terahertz spectroscopy characteristic parameters, perform defect depth inversion to obtain the defect depth prediction result.

[0044] Merge the hyperspectral image feature parameters, Raman spectral feature parameters, and terahertz spectral feature parameters according to the same spatial coordinates to form multimodal spectral feature data at each spatial position point; perform standardization processing on the multimodal spectral feature data to eliminate the dimensional differences of different spectral parameters (such as reflectivity being dimensionless and stress unit being Pa), and obtain the standardized multimodal spectral features, solving the problem of "depth inversion deviation caused by one-sided features in a single spectral technique" (such as misjudging surface contamination as a deep defect when only using terahertz).

[0045] Use the standardized multimodal spectral features as the input of the trained deep neural network, and output the defect depth prediction value at each spatial position point, and summarize to form the defect depth prediction result. The input is the original multimodal features, and the output is directly the depth value, without the need to manually design feature fusion rules, reducing the subjective deviation caused by manual intervention.

[0046] Example 2: For the deep neural network in Example 1, adopt a transfer learning strategy for optimization, including the following steps: Select a pre-trained model as the initial model. The pre-trained model is a convolutional neural network model trained on a large-scale general spectral dataset; Replace the last few fully connected layers of the pre-trained model with customized fully connected layers adapted to the current optical material defect detection task, and adjust the number of neurons and activation functions of the customized fully connected layers to make it more suitable for processing the spectral feature data of the current task; Use a small amount of labeled optical material defect sample data to fine-tune and train the adjusted model. By setting a small learning rate, the model can learn the specific features of the current task while retaining the general spectral feature representations learned in the pre-trained model; Use the validation set to validate and evaluate the fine-tuned model, and further adjust the hyperparameters of the model, such as the learning rate, regularization parameter, etc., according to the evaluation results to improve the generalization ability and prediction accuracy of the model, so as to realize the optimization of the deep neural network and make the defect depth prediction result based on multimodal spectral feature data more accurate and reliable.

[0047] Example 3: An electronic device, including: a processor; and a memory, in which computer program instructions are stored, and when the computer program instructions are run by the processor, the processor executes the detection method in Example 1 or Example 2.

Claims

1. A method for detecting the depth of optical material defects based on spectroscopy, characterized in that, It includes the following steps: Obtain the hyperspectral image feature parameters of the optical material to be detected, and determine whether there are defects based on the hyperspectral image feature parameters; If there are no defects, the detection ends; If there are defects, determine the preliminary defect area and defect type; Obtain the Raman spectral feature parameters of the preliminary defect area, and determine the final defect area based on the Raman spectral feature parameters; Perform terahertz spectral depth scanning on the final defect area to obtain terahertz spectral feature parameters, and construct a defect depth information map; Perform defect depth inversion based on the hyperspectral image feature parameters, Raman spectral feature parameters, and terahertz spectral feature parameters to obtain the defect depth prediction result.

2. The method for detecting the depth of optical material defects based on spectroscopy according to claim 1, wherein The hyperspectral image feature parameters include reflectivity, peak intensity, full width at half maximum, and inter-band correlation coefficient; The Raman spectral feature parameters include Raman spectral peak position, Raman spectral peak intensity, and Raman spectral full width at half maximum; the terahertz spectral feature parameters include terahertz reflection spectral peak position, terahertz reflection spectral intensity, and terahertz reflection spectral phase change.

3. The method for detecting the depth of optical material defects based on spectroscopy according to claim 2, wherein Determining whether there are defects based on the hyperspectral image feature parameters includes the following steps: The optical material to be detected moves on the motion control platform of the hyperspectral image data device according to a preset scanning path, and the hyperspectral image data of each coordinate point on the optical material to be detected is obtained based on the hyperspectral image data device; After preprocessing the hyperspectral image data, perform principal component analysis for dimensionality reduction to obtain the hyperspectral image feature parameters of each coordinate point; Perform mean processing on the hyperspectral image feature parameters of each coordinate point to obtain the homogenized hyperspectral image feature parameters; Perform similarity analysis on the homogenized hyperspectral image feature parameters and the standard hyperspectral image feature parameters stored in the database to obtain the characteristic cosine similarity value; If the characteristic cosine similarity value is greater than the set similarity threshold, there are no defects; If the characteristic cosine similarity value is not greater than the set similarity threshold, there are defects.

4. The method for detecting the depth of optical material defects based on spectroscopy according to claim 3, wherein When there are defects, the process of determining the preliminary defect area and defect type is as follows: Perform similarity analysis on the hyperspectral image feature parameters of each coordinate point and the standard hyperspectral image feature parameters stored in the database to obtain the coordinate point characteristic cosine similarity value; If the coordinate point characteristic cosine similarity value of a certain coordinate point is greater than the set coordinate point characteristic similarity threshold, then this coordinate point is not recorded as a defect coordinate point; If the coordinate point characteristic cosine similarity value of a certain coordinate point is not greater than the set coordinate point characteristic similarity threshold, then this coordinate point is recorded as a defect coordinate point; Connect all the defect coordinate points to determine the preliminary defect area; Perform homogenization processing on the hyperspectral image feature parameters of each defect coordinate point in the preliminary defect area to obtain the homogenized hyperspectral image feature parameters of the defect area; Use the homogenized hyperspectral image feature parameters of the defect area as the input of the trained support vector machine model to obtain the defect type.

5. The method for detecting the depth of optical material defects based on spectroscopy according to claim 2, wherein Determining the final defect area based on the Raman spectral feature parameters includes the following steps: Perform Raman spectral scanning on the preliminary defect area based on a Raman spectrometer to obtain the Raman spectral data corresponding to each defect coordinate point; Perform feature processing on the Raman spectral data corresponding to each defect coordinate point to obtain the Raman spectral characteristic parameters corresponding to each defect coordinate point; Determine a machine learning model; Use the Raman spectral characteristic parameters corresponding to each defect coordinate point as the input of the machine learning model, output the probability values of each defect coordinate point belonging to the defect type, and record the defect coordinate points with probability values greater than the probability threshold as the final defect area coordinate points; Connect all the final defect area coordinate points to determine the final defect area.

6. The method for detecting the depth of optical material defects based on spectroscopy according to claim 5, characterized in that The method for determining the machine learning model is as follows: Determine the molecular structure composition of the preliminary defect area based on the Raman spectral characteristic parameters corresponding to each defect coordinate point; Determine the difference in Raman peak positions before and after stress action for each defect coordinate point, denoted as the Raman shift, and determine the stress distribution of the preliminary defect area based on the stress-Raman shift correlation formula; Determine the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area.

7. The method for detecting the depth of optical material defects based on spectroscopy according to claim 6, characterized in that, Determining the molecular structure composition of the preliminary defect area includes the following steps: Compare the Raman spectral characteristic parameters corresponding to each defect coordinate point with the known molecular vibration mode database one by one to determine the molecular structure composition of the defect area at each coordinate point; The stress-Raman shift correlation formula is , is the Raman shift, is the stress coefficient, is the stress.

8. The method for detecting the depth of optical material defects based on spectroscopy according to claim 6, characterized in that Determining the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area includes the following steps: Obtain the historical molecular structure composition and stress distribution combinations corresponding to the defect type, and perform clustering analysis to obtain multiple clustering points, each clustering point corresponding to a historical molecular structure composition and stress distribution combination; Preset multiple sets of machine learning model parameters for each clustering point; Under the historical molecular structure composition and stress distribution combination, use the historical Raman spectral characteristic parameters corresponding to the defect type as the input of the machine learning model corresponding to each set of machine learning model parameters, output the corresponding defect area boundary coordinates, calculate the accuracy of the machine learning model corresponding to each set of machine learning model parameters, and select the machine learning model corresponding to the set of machine learning model parameters with the highest accuracy; Train the machine learning model corresponding to this set of machine learning model parameters to obtain the machine learning models corresponding to different historical molecular structure compositions and stress distribution combinations of different defect types; Perform similarity analysis on the molecular structure composition and stress distribution of the preliminary defect area and the corresponding groups of historical molecular structure compositions and stress distribution combinations under the corresponding defect type to determine the most similar historical molecular structure composition and stress distribution combination, and obtain the corresponding machine learning model.

9. The method for detecting the depth of optical material defects based on spectroscopy according to claim 2, wherein Obtain terahertz spectral characteristic parameters and construct a defect depth information map, including the following steps: Start the depth scanning program of the terahertz time-domain spectrometer. The motion control platform of the terahertz time-domain spectrometer controls the optical material to be detected to move in the Z-axis direction at a preset scanning interval, pause after each movement of a set distance, and collect the terahertz reflection spectral data at this depth point; Perform feature extraction on the terahertz reflection spectral data at each depth point to obtain the terahertz spectral characteristic parameters at each depth point; Based on 3D visualization software, with depth as the Z-axis and the X and Y coordinates of the final defect area as the plane axes, the terahertz spectral characteristic parameters of each depth point are visually mapped to generate a defect depth information map of the defect area.

10. The method for detecting the depth of optical material defects based on spectroscopy according to claim 2, wherein The defect depth prediction results are obtained through the following steps: Merge the hyperspectral image characteristic parameters, Raman spectral characteristic parameters, and terahertz spectral characteristic parameters according to the same spatial coordinates to form multi-modal spectral characteristic data for each spatial position point; Perform standardization processing on the multi-modal spectral characteristic data to obtain the standardized multi-modal spectral characteristics; Use the standardized multi-modal spectral characteristics as the input of the trained deep neural network, output the defect depth prediction values for each spatial position point, and summarize them to form the defect depth prediction results.

Images