Depth detection method for optical material defects based on spectroscopy

Abstract

The present invention discloses a spectrum-based optical material defect depth detection method, which belongs to the field of defect detection technology and includes the following steps: obtaining hyperspectral image characteristic parameters of the optical material to be detected, and determining whether a defect exists based on the hyperspectral image characteristic parameters; if no defect exists, the detection ends; if a defect exists, determining a preliminary defect area and defect type; obtaining Raman spectrum characteristic parameters of the preliminary defect area, and determining the final defect area based on the Raman spectrum characteristic parameters; performing a terahertz spectrum depth scan on the final defect area to obtain terahertz spectrum characteristic parameters and construct a defect depth information map; performing defect depth inversion based on the hyperspectral image characteristic parameters, Raman spectrum characteristic parameters, and terahertz spectrum characteristic parameters to obtain a defect depth prediction result. The present invention solves the problem that traditional methods have difficulty in accurately detecting the depth of defects inside optical materials, and has high precision, comprehensiveness, and adaptability.

Description

Technical Field

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

[0002] Optical materials are a type of material that can regulate properties such as the propagation direction, intensity, polarization state, and spectral distribution of light to meet the specific functional requirements of optical systems. For example, CN114136979A discloses a device and method for detecting micro-defects on the surface of volume-scattering optical materials. The device adjusts the polarizer to place the illumination light source in a specific polarization mode, and uses a bracket to adjust the incident angle of the illumination light source to reduce the transmittance of the illumination light source, thereby reducing the impact of the sample's volume-scattered light on the detection of micro-defects on the material surface. This allows a microscope located directly above the sample holder to clearly observe defects on the surface of the volume-scattering optical material sample.

[0003] However, in existing optical material defect detection methods, the scattering centers inside the volume scattering optical material will cause the incident light to be scattered multiple times, resulting in the reflection / scattering signals of surface micro-defects being easily submerged by the internal scattered light, and the coupling of the defect depth and the light penetration depth will cause signal superposition. At the same time, the anisotropy and inhomogeneity of the optical properties also make the signal exhibit nonlinear characteristics. It is difficult to accurately distinguish the defect signal and infer the depth through the ratio of reflectivity to transmittance under a single incident angle. Therefore, there is a problem of difficulty in accurately detecting optical material defects and depth. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a spectral-based method for detecting the depth of optical material defects. By integrating hyperspectral imaging, Raman spectroscopy and terahertz spectroscopy technology, and combining it with deep neural network analysis, it can accurately identify and locate defects in optical materials and their depth, solving the problem that traditional methods are difficult to accurately detect the depth of internal defects in optical materials. It has high precision, comprehensiveness and adaptability.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] The optical material defect depth detection method based on spectroscopy includes the following steps:

[0007] Acquire characteristic parameters of a hyperspectral image of the optical material to be inspected, and determine whether there is a defect based on the characteristic parameters of the hyperspectral image;

[0008] If there are no defects, the inspection ends;

[0009] If defects exist, determine the preliminary defect area and defect type;

[0010] Obtaining Raman spectrum characteristic parameters of a preliminary defect area, and determining a final defect area based on the Raman spectrum characteristic parameters;

[0011] Perform a terahertz spectrum depth scan on the final defect area to obtain the terahertz spectrum characteristic parameters and construct a defect depth information map;

[0012] The defect depth is inverted based on the characteristic parameters of hyperspectral images, Raman spectrum and terahertz spectrum to obtain the defect depth prediction results.

[0013] Preferably, the hyperspectral image characteristic parameters include reflectivity, peak intensity, half-peak width and inter-band correlation coefficient; the Raman spectrum characteristic parameters include Raman spectrum peak position, Raman spectrum peak intensity and Raman spectrum half-peak width; the terahertz spectrum characteristic parameters include terahertz reflection spectrum reflection peak position, terahertz reflection spectrum intensity and terahertz reflection spectrum phase change.

[0014] Preferably, determining whether a defect exists based on the hyperspectral image characteristic parameters includes the following steps:

[0015] The optical material to be inspected moves along a preset scanning path on a motion control platform of a hyperspectral image data device, and the hyperspectral image data of each coordinate point on the optical material to be inspected is obtained based on the hyperspectral image data device;

[0016] After preprocessing the hyperspectral image data, principal component analysis and dimensionality reduction are performed to obtain the hyperspectral image characteristic parameters of each coordinate point;

[0017] Performing mean processing on the hyperspectral image characteristic parameters of each coordinate point to obtain the averaged hyperspectral image characteristic parameters;

[0018] Perform similarity analysis on the characteristic parameters of the averaged hyperspectral image and the characteristic parameters of the standard hyperspectral image stored in the database to obtain the characteristic cosine similarity value;

[0019] If the feature cosine similarity value is greater than the set similarity threshold, there is no defect;

[0020] If the feature cosine similarity value is not greater than the set similarity threshold, there is a defect.

[0021] Preferably, when a defect exists, the process of determining the preliminary defect area and defect type is as follows:

[0022] 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 feature cosine similarity value of the coordinate point;

[0023] If the cosine similarity value of the coordinate point feature of a certain coordinate point is greater than the set coordinate point feature similarity threshold, the coordinate point will not be recorded as a defect coordinate point;

[0024] If the cosine similarity value of the coordinate point feature of a certain coordinate point is not greater than the set coordinate point feature similarity threshold, the coordinate point is recorded as a defective coordinate point;

[0025] Connect all defect coordinate points to determine the preliminary defect area;

[0026] Performing average processing on the hyperspectral image characteristic parameters of each defect coordinate point in the preliminary defect area to obtain averaged hyperspectral image characteristic parameters of the defect area;

[0027] The characteristic parameters of the averaged hyperspectral image of the defect area are used as the input of the trained support vector machine model to obtain the defect type.

[0028] Preferably, determining the final defect area based on Raman spectral characteristic parameters includes the following steps:

[0029] Perform Raman spectrum scanning on the preliminary defect area based on a Raman spectrometer to obtain Raman spectrum data corresponding to each defect coordinate point;

[0030] Perform feature processing on the Raman spectrum data corresponding to each defect coordinate point to obtain the Raman spectrum feature parameters corresponding to each defect coordinate point;

[0031] Identify the machine learning model;

[0032] The Raman spectrum characteristic parameters corresponding to each defect coordinate point are used as the input of the machine learning model, and the probability value of each defect coordinate point belonging to the defect type is output. The defect coordinate point with a probability value greater than the probability threshold is recorded as the final defect area coordinate point;

[0033] Connect all the final defect area coordinate points to determine the final defect area.

[0034] Preferably, the method for determining the machine learning model is:

[0035] Determine the molecular structure composition of the preliminary defect area based on the Raman spectrum characteristic parameters corresponding to each defect coordinate point;

[0036] Determine the difference in Raman peak position before and after stress at each defect coordinate point, record it as Raman shift, and determine the stress distribution of the preliminary defect area based on the stress-Raman shift correlation formula;

[0037] The machine learning model is determined based on the molecular structure composition and stress distribution of the preliminary defect area.

[0038] Preferably, determining the molecular structure composition of the preliminary defect region comprises the following steps:

[0039] 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;

[0040] The stress-Raman shift correlation formula is: , is the Raman shift, is the stress coefficient, For stress.

[0041] Preferably, determining a machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area comprises the following steps:

[0042] Obtain the historical molecular structure composition and stress distribution combination corresponding to the defect type, and perform cluster analysis to obtain multiple cluster points, each of which corresponds to a historical molecular structure composition and stress distribution combination;

[0043] Preset multiple sets of machine learning model parameters for each cluster point;

[0044] 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 model corresponding to each group of machine learning model parameters, the corresponding defect area boundary coordinates are output, the accuracy of the machine learning model 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;

[0045] Training the machine learning model corresponding to the set 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;

[0046] The molecular structure composition and stress distribution of the preliminary defect area are analyzed for similarity with 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.

[0047] Preferably, obtaining the characteristic parameters of the terahertz spectrum and constructing a defect depth information map comprises the following steps:

[0048] Start the depth scanning program of the terahertz time-domain spectrometer. The motion control platform of the terahertz time-domain spectrometer controls the movement of the optical material to be tested in the Z-axis direction according to the preset scanning interval. After each movement of the set distance, it pauses and collects the terahertz reflection spectrum data of the depth point.

[0049] Extract the characteristics of the terahertz reflection spectrum data at each depth point to obtain the characteristic parameters of the terahertz spectrum at each depth point;

[0050] 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 spectrum characteristic parameters at each depth point are visualized and mapped to generate a defect depth information map of the defect area.

[0051] Preferably, obtaining the defect depth prediction result includes the following steps:

[0052] The hyperspectral image characteristic parameters, Raman spectrum characteristic parameters and terahertz spectrum characteristic parameters are merged according to the same spatial coordinates to form multimodal spectrum characteristic data of each spatial position point;

[0053] Performing standardization processing on the multimodal spectral feature data to obtain standardized multimodal spectral features;

[0054] The standardized multimodal spectral features are used as the input of the trained deep neural network, and the defect depth prediction value of each spatial position point is output, which is summarized to form the defect depth prediction result.

[0055] The present invention has the following beneficial effects:

[0056] The present invention uses hyperspectral images to determine the presence, type and preliminary area of ​​defects, uses Raman spectroscopy to accurately delineate the final defect area, combines terahertz spectroscopy for deep scanning to construct a defect depth information map, and then integrates multi-spectral feature parameters to invert the defect depth through a deep neural network, realizing full-process detection from defect identification to depth positioning, improving the accuracy and comprehensiveness of detection, and solving the problem of difficulty in accurately detecting defects and depth of optical materials in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of the method of the present invention;

[0058] Figure 2 Flowchart for determining the final defect area for the present invention. DETAILED DESCRIPTION

[0059] Example 1: Figure 1 As shown, the optical material defect depth detection method based on spectroscopy includes the following steps:

[0060] Acquire characteristic parameters of a hyperspectral image of the optical material to be inspected, and determine whether there is a defect based on the characteristic parameters of the hyperspectral image;

[0061] If there are no defects, the inspection ends;

[0062] If defects exist, determine the preliminary defect area and defect type;

[0063] Obtaining Raman spectrum characteristic parameters of a preliminary defect area, and determining a final defect area based on the Raman spectrum characteristic parameters;

[0064] Perform a terahertz spectrum depth scan on the final defect area to obtain the terahertz spectrum characteristic parameters and construct a defect depth information map;

[0065] The defect depth is inverted based on the characteristic parameters of hyperspectral images, Raman spectrum and terahertz spectrum to obtain the defect depth prediction results.

[0066] The characteristic parameters of hyperspectral images include reflectivity, peak intensity, half-maximum width and inter-band correlation coefficient; the characteristic parameters of Raman spectrum include Raman spectrum peak position, Raman spectrum peak intensity and Raman spectrum half-maximum width; the characteristic parameters of terahertz spectrum include terahertz reflection spectrum reflection peak position, terahertz reflection spectrum intensity and terahertz reflection spectrum phase change.

[0067] Determining whether there is a defect based on the characteristic parameters of the hyperspectral image includes the following steps:

[0068] The optical material to be inspected moves along a preset scanning path (e.g., grid scanning) on ​​the hyperspectral imaging data equipment's motion control platform. The hyperspectral imaging data equipment then acquires hyperspectral image data for each coordinate point on the optical material to be inspected. The motion control platform drives the sample along the preset path, ensuring full surface coverage. Each coordinate point corresponds to a spatial resolution of 50μm / pixel, generating a three-dimensional data cube containing spatial and spectral information—the hyperspectral image data for each coordinate point. This full-point data provides a precise spatial index for subsequent defect location, ensuring traceability of each suspicious point.

[0069] After preprocessing the hyperspectral image data (using radiometric correction to eliminate uneven illumination and median filtering to remove noise and improve the signal-to-noise ratio of the spectral data), principal component analysis (PCA) is used for dimensionality reduction to obtain the hyperspectral image characteristic parameters for each coordinate point. Preprocessing addresses the problem of feature distortion caused by noise interference in the raw data, ensuring the reliability of subsequent analysis. PCA dimensionality reduction eliminates redundant information and highlights defect-sensitive features (for example, principal component 1 reflects the overall reflectance trend), improving computational efficiency and preventing the "curse of dimensionality" from affecting model performance.

[0070] The hyperspectral image characteristic parameters of each coordinate point are averaged to obtain the averaged hyperspectral image characteristic parameters; the averaged hyperspectral image characteristic parameters are analyzed for similarity with the standard hyperspectral image characteristic 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 is no defect; if the characteristic cosine similarity value is not greater than the set similarity threshold, there is a defect.

[0071] Averaging processing filters out local anomalies, preventing individual noise points from misjudging the overall status and resolving the problem of "false alarms caused by accidental interference." Quantifying similarity judgment (e.g., setting the threshold to 0.9) converts defect identification into a calculable numerical comparison, standardizing the inspection process and avoiding the subjectivity of manual interpretation.

[0072] Through the process of "full coordinate acquisition → noise reduction and dimensionality reduction → overall comparison", a "primary screening mechanism" for optical material defects was constructed. By excluding defect-free samples and focusing on suspicious areas, computing power and time are saved for subsequent Raman spectroscopy fine analysis and terahertz deep scanning, which is the basic link for achieving "precise detection".

[0073] When defects exist, the process of determining the preliminary defect area and defect type is as follows:

[0074] The hyperspectral image feature parameters at each coordinate point are analyzed for similarity with those of standard hyperspectral images stored in a database to obtain the cosine similarity value of the coordinate point features. This enables "pixel-level positioning" of defects, detecting micron-level defects (such as scratches and point-like impurities), resolving the difficulty of identifying tiny defects with traditional methods. It also quantifies single-point anomalies and avoids missed detections due to local feature fluctuations (such as minor contamination on the material surface).

[0075] If the cosine similarity value of the coordinate point feature of a certain coordinate point is greater than the set coordinate point feature similarity threshold (such as 0.8), then the coordinate point is not recorded as a defective coordinate point; if the cosine similarity value of the coordinate point feature of a certain coordinate point is not greater than the set coordinate point feature similarity threshold, then the coordinate point is recorded as a defective coordinate point; all defective coordinate points are connected to determine the preliminary defect area; and a connected domain analysis algorithm (such as 8-neighborhood connectivity) is used to merge discrete defective coordinate points into a continuous area.

[0076] Aggregating discrete outliers into physically explainable defect regions aligns with the "patchy" distribution of actual defects, resolving the problem of scattered points being misidentified as multiple defects. This initial delineation of the defect region narrows the scope for subsequent Raman spectroscopy scans, reducing data collection (e.g., scanning only the defect region rather than the entire surface), and improving detection efficiency.

[0077] The hyperspectral image feature parameters of each defect coordinate point in the preliminary defect area are averaged to obtain the averaged hyperspectral image feature parameters of the defect area; the averaged 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.

[0078] The characteristic parameters of all coordinate points within the defect area are averaged to reduce local noise and highlight overall features. Hyperspectral data of known defect types is pre-trained, and the radial basis function (RBF) kernel is selected. Parameters are optimized (e.g., C=10, γ=0.1) to improve classification accuracy. Automatic classification based on machine learning avoids the subjectivity of manual interpretation and can simultaneously distinguish multiple defect types (e.g., bubbles in glass versus metal impurities), addressing the inefficient and error-prone manual identification of defect types.

[0079] like Figure 2 As shown, determining the final defect area based on Raman spectrum characteristic parameters includes the following steps:

[0080] A Raman spectrometer scans the initial defect area to obtain Raman spectral data corresponding to each defect coordinate point. Feature processing is then performed on the Raman spectral data corresponding to each defect coordinate point to obtain the corresponding Raman spectral characteristic parameters. The scanning range is limited to the determined initial defect area, and the scanning step size is set to 0.5μm, achieving submicron spatial resolution. This allows the capture of subtle structural changes at the defect edge (such as lattice distortion at the crack tip), providing high-fidelity data for precise boundary delineation. Feature processing includes Savitzky-Golay filtering to remove high-frequency noise and polynomial fitting to correct baseline drift, ensuring the accuracy of spectral features (such as peak positions).

[0081] Determine the 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 value of each defect coordinate point belonging to the defect type, and record the defect coordinate point with a probability value greater than the probability threshold as the final defect area coordinate point; connect all the final defect area coordinate points to determine the final defect area.

[0082] The machine learning model training phase uses historical Raman data with labeled defect boundaries (e.g., manually marked crack edge coordinates), with peak position and stress distribution as input and defect probability as output. Cross-validation is used to determine the optimal threshold (e.g., 0.6), balancing recall (to avoid missed detections) with precision (to avoid false positives). The machine learning model automatically captures the nonlinear relationship between spectral features and defect boundaries (e.g., the correlation between stress concentration areas and crack propagation paths), addressing the inability of traditional threshold methods to adapt to complex defect morphologies. Through the "high-resolution spectral acquisition → feature enhancement → intelligent boundary recognition" process, the system achieves an upgrade from "initial location" to "precise delineation" of defect areas.

[0083] Determining a machine learning model includes the following steps: determining the molecular structure composition of a preliminary defect area based on the Raman spectral characteristic parameters corresponding to each defect coordinate point; determining the difference in Raman peak position before and after stress action at each defect coordinate point, recorded as Raman shift, and determining the stress distribution of the preliminary defect area based on a stress-Raman shift correlation formula; and determining a machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area.

[0084] 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 (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 composition (such as SiO2, SiC) or impurities (such as metal ions) of the defect area by comparing the peak position and intensity of the measured spectrum with the standard spectrum in the database.

[0085] Analyze the nature of defects at the chemical level, distinguish different causes of similar defects (such as "oxide cracks" and "stress cracks"), and solve the problem of "fuzzy defect causes leading to poor model generalization ability". Molecular composition information can be used as a key feature of the machine learning model to improve the model's ability to distinguish defect types (for example, areas containing oxides are more likely to be judged as corrosion defects).

[0086] The stress-Raman shift correlation formula is: , is the Raman shift, For stress coefficients such as 520cm for silicon -1 Standard peak), Quantify the mechanical state of the defect area and identify stress concentration points (such as crack tip σ>100MPa), providing a basis for defect growth prediction.

[0087] Determine the machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area, including the following steps:

[0088] Obtain the historical molecular structure composition and stress distribution combination corresponding to the defect type, and perform cluster analysis to obtain multiple cluster points, each cluster point corresponding to a historical molecular structure composition and stress distribution combination; preset multiple sets of machine learning model parameters for each cluster point (such as the kernel function type of SVM, the penalty factor C, and the number of convolution layers of CNN); preset machine learning models with different parameters for each cluster category (such as the kernel function type of SVM, the number of convolution layers of CNN), train and calculate the accuracy through historical data, avoid "one-size-fits-all" modeling, dynamically call the optimal model for different defect characteristics, and improve the accuracy of boundary recognition.

[0089] 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 model corresponding to each group of machine learning model parameters, the corresponding defect area boundary coordinates are output, the accuracy of the machine learning model 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;

[0090] The machine learning model corresponding to the set of machine learning model parameters is trained to obtain machine learning models corresponding to different historical molecular structure compositions and stress distribution combinations of different defect types. The molecular structure composition and stress distribution of the preliminary defect area are subjected to similarity analysis with the historical molecular structure composition and stress distribution combinations of each group 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.

[0091] Through the mechanism of "data clustering-directional modeling-dynamic matching", an adaptive selection system for defect boundary identification models was constructed. By introducing data-driven clustering analysis and model optimization strategies, the chemical-mechanical characteristics of Raman spectroscopy were deeply integrated with machine learning algorithms, providing an efficient and explainable technical solution for the accurate identification of defect boundaries. This is the key technical support for the "deepening and refinement of defect features" in this embodiment.

[0092] A terahertz spectrum depth scan is performed on the final defect area to obtain terahertz spectrum characteristic parameters and construct a defect depth information map; the terahertz spectrum characteristic parameters include the terahertz reflection spectrum reflection peak position, terahertz reflection spectrum intensity and terahertz reflection spectrum phase change.

[0093] Start the depth scanning program of the terahertz time-domain spectrometer. The motion control platform of the terahertz time-domain spectrometer controls the movement of the optical material to be detected in the Z-axis direction according to the preset scanning interval. It pauses after each movement of the set distance to collect the terahertz reflection spectrum data of the depth point.

[0094] From the material surface (Z=0) to the preset maximum depth (e.g., Z=2mm), the moving interval matches the terahertz wave wavelength (e.g., a 50μm interval is suitable for a 0.1THz wavelength), ensuring a depth resolution better than λ / 2 (approximately 1.5mm), achieving submillimeter depth positioning, and solving the problem of "insufficient depth resolution leading to confusion between defect layers" in traditional terahertz imaging by collecting depth points one by one (e.g., distinguishing surface cracks from deep stratification).

[0095] Feature extraction is performed on the terahertz reflection spectrum data at each depth point to obtain the characteristic parameters of the terahertz spectrum at each depth point, including the position of the terahertz reflection spectrum reflection peak, the terahertz reflection spectrum intensity, and the terahertz reflection spectrum phase change. Using 3D 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, the terahertz spectrum characteristic parameters at each depth point are visualized and mapped to generate a defect depth information map of the defect area. This realizes the core link of "defect depth detection" and, through the penetration and depth resolution capabilities of terahertz spectroscopy, transforms internal defects of optical materials from "invisible" to "quantifiable, locatable, and analyzable."

[0096] The defect depth is inverted based on the characteristic parameters of hyperspectral images, Raman spectrum and terahertz spectrum to obtain the defect depth prediction results.

[0097] The hyperspectral image characteristic parameters, Raman spectrum characteristic parameters and terahertz spectrum characteristic parameters are merged according to the same spatial coordinates to form multimodal spectral characteristic data for each spatial position point; the multimodal spectral characteristic data are standardized to eliminate the dimensional differences of different spectral parameters (such as reflectivity is dimensionless and stress is in Pa), and the standardized multimodal spectral characteristics are obtained to solve the problem of "one-sided characteristics leading to depth inversion deviation" of a single spectral technology (for example, using only terahertz may misjudge surface contamination as deep defects).

[0098] The normalized multimodal spectral features are used as input to a trained deep neural network. The predicted defect depth values ​​for each spatial location are output and aggregated to form the defect depth prediction results. The input is the original multimodal features, and the output is directly the depth value, eliminating the need for manually designed feature fusion rules and reducing subjective bias caused by human intervention.

[0099] Example 2: The deep neural network in Example 1 is optimized using a transfer learning strategy, including the following steps:

[0100] A pre-trained model is selected as the initial model. The pre-trained model is a convolutional neural network model trained on a large-scale general spectral dataset.

[0101] 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 function of the customized fully connected layers to make them more suitable for processing the spectral feature data of the current task;

[0102] The adjusted model is fine-tuned using a small amount of labeled optical material defect sample data. By setting a small learning rate, the model can learn the specific features of the current task while retaining the general spectral feature representation learned in the pre-trained model.

[0103] The fine-tuned model is verified and evaluated using the validation set. Based on the evaluation results, the model's hyperparameters, such as the learning rate and regularization parameters, are further adjusted to improve the model's generalization ability and prediction accuracy, thereby optimizing the deep neural network and making the results of defect depth prediction based on multimodal spectral feature data more accurate and reliable.

[0104] Embodiment 3: An electronic device comprises: a processor; and a memory, wherein computer program instructions are stored in the memory, and when the computer program instructions are executed by the processor, the processor executes the detection method in embodiment 1 or embodiment 2.

Claims

1. A method for detecting the depth of optical material defects based on spectroscopy, characterized in that: The following steps are involved: Acquire characteristic parameters of a hyperspectral image of the optical material to be inspected, and determine whether there is a defect based on the characteristic parameters of the hyperspectral image; If there are no defects, the inspection ends; If defects exist, determine the preliminary defect area and defect type; Obtaining Raman spectrum characteristic parameters of a preliminary defect area and determining a final defect area based on the Raman spectrum characteristic parameters includes the following steps: Perform Raman spectrum scanning on the preliminary defect area based on a Raman spectrometer to obtain Raman spectrum data corresponding to each defect coordinate point; Perform feature processing on the Raman spectrum data corresponding to each defect coordinate point to obtain the Raman spectrum feature parameters corresponding to each defect coordinate point; The molecular structure composition of the preliminary defect area is determined based on the Raman spectral characteristic parameters corresponding to each defect coordinate point: the Raman spectral characteristic parameters corresponding to each defect coordinate point are compared one by one with the known molecular vibration mode database 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 stress; Determine the difference in Raman peak position before and after stress at each defect coordinate point, record it as Raman shift, and determine the stress distribution of the preliminary defect area based on the stress-Raman shift correlation formula; Determine a machine learning model based on the molecular structure composition and stress distribution of the preliminary defect area; The Raman spectrum characteristic parameters corresponding to each defect coordinate point are used as the input of the machine learning model, and the probability value of each defect coordinate point belonging to the defect type is output. The defect coordinate point with a probability value greater than the probability threshold is recorded as the final defect area coordinate point; Connect all the final defect area coordinate points to determine the final defect area; Perform a terahertz spectrum depth scan on the final defect area to obtain the terahertz spectrum characteristic parameters and construct a defect depth information map; The defect depth is inverted based on the characteristic parameters of hyperspectral images, Raman spectrum and terahertz spectrum to obtain the defect depth prediction results.

2. The optical material defect depth detection method based on spectroscopy according to claim 1, characterized in that: Hyperspectral image characteristic parameters include reflectance, peak intensity, half-peak width and inter-band correlation coefficient; The characteristic parameters of Raman spectrum include Raman spectrum peak position, Raman spectrum peak intensity and Raman spectrum half-maximum width; the characteristic parameters of THz spectrum include THz reflection spectrum reflection peak position, THz reflection spectrum intensity and THz reflection spectrum phase change.

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

4. The optical material defect depth detection method based on spectroscopy according to claim 3, characterized in that: When defects exist, 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 feature cosine similarity value of the coordinate point; If the cosine similarity value of the coordinate point feature of a certain coordinate point is greater than the set coordinate point feature similarity threshold, the coordinate point will not be recorded as a defect coordinate point; If the cosine similarity value of the coordinate point feature of a certain coordinate point is not greater than the set coordinate point feature similarity threshold, the coordinate point is recorded as a defective coordinate point; Connect all defect coordinate points to determine the preliminary defect area; Performing average processing on the hyperspectral image characteristic parameters of each defect coordinate point in the preliminary defect area to obtain averaged hyperspectral image characteristic parameters of the defect area; The characteristic parameters of the averaged hyperspectral image of the defect area are used as the input of the trained support vector machine model to obtain the defect type.

5. The optical material defect depth detection method based on spectroscopy according to claim 1, characterized in that: 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 combination corresponding to the defect type, and perform cluster analysis to obtain multiple cluster points, each of which corresponds to a historical molecular structure composition and stress distribution combination; Preset multiple sets of machine learning model parameters for each cluster 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 model corresponding to each group of machine learning model parameters, the corresponding defect area boundary coordinates are output, the accuracy of the machine learning model 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; Training the machine learning model corresponding to the set 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; The molecular structure composition and stress distribution of the preliminary defect area are analyzed for similarity with 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.

6. The optical material defect depth detection method based on spectroscopy according to claim 2, characterized in that: Obtaining the characteristic parameters of the terahertz spectrum 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 movement of the optical material to be tested in the Z-axis direction according to the preset scanning interval. After each movement of the set distance, it pauses and collects the terahertz reflection spectrum data of the depth point. Extract the characteristics of the terahertz reflection spectrum data at each depth point to obtain the characteristic parameters of the terahertz spectrum 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 spectrum characteristic parameters at each depth point are visualized and mapped to generate a defect depth information map of the defect area.

7. The optical material defect depth detection method based on spectroscopy according to claim 2, characterized in that: Obtaining the defect depth prediction result includes the following steps: The hyperspectral image characteristic parameters, Raman spectrum characteristic parameters and terahertz spectrum characteristic parameters are merged according to the same spatial coordinates to form multimodal spectrum characteristic data of each spatial position point; Performing standardization processing on the multimodal spectral feature data to obtain standardized multimodal spectral features; The standardized multimodal spectral features are used as the input of the trained deep neural network, and the defect depth prediction value of each spatial position point is output, which is summarized to form the defect depth prediction result.

Images