Metal aging evaluation method and system based on LIBS (Laser-induced Breakdown Spectroscopy) multi-dimensional feature fusion

Abstract

The invention discloses a metal aging evaluation method and system based on LIBS multi-dimensional feature fusion, and belongs to the technical field of metal material life evaluation, and the method comprises the steps: collecting LIBS spectral data of a to-be-detected material; performing quality control and standardization treatment on the LIBS spectral data; extracting features from the preprocessed spectral data; performing adaptive weighted fusion on the multi-dimensional features based on a deep learning attention mechanism; selecting a corresponding neural network model according to the downstream task, and outputting performance indexes of the material; training and optimizing the neural network model by using the sample data of the known label; and performing performance evaluation on the trained model, and deploying the model to an actual detection system after the evaluation is passed. According to the method, data multi-dimensional features are extracted, and the limitation of single-index evaluation is broken through; through multi-dimensional feature collaborative modeling, the prediction accuracy is improved; multi-index complementary fusion effectively reduces the influence that a single index is susceptible to interference, and the stability of the model under different measurement conditions is enhanced.

Description

Technical Field

[0001] This invention relates to the field of metal material life assessment technology, and in particular to a metal aging assessment method and system based on LIBS multidimensional feature fusion. Background Technology

[0002] Metallic materials undergo structural degradation during long-term high-temperature service, leading to a decline in mechanical properties, which is known as material aging. Traditional evaluation methods, such as metallographic analysis and hardness testing, require destructive testing of samples and cannot achieve online monitoring. Although laser-induced breakdown spectroscopy (LIBS) technology has advantages such as non-destructive, rapid, and in-situ detection, existing LIBS-based material evaluation methods have the following shortcomings: (1) Single information utilization: Existing studies mostly rely on a single LIBS index, such as the intensity of a certain characteristic spectral line or a simple intensity ratio, to judge material properties, without fully exploring the multidimensional information value of LIBS data; (2) Insufficient accuracy: Single index evaluation is easily affected by matrix effects, self-absorption effects, surface conditions, and other factors, resulting in large prediction errors and making it difficult to meet the needs of industrial applications; (3) Simple feature fusion: Although a few studies have attempted to combine multiple indices, they mostly use simple linear combinations or traditional machine learning methods, such as support vector machines, which fail to effectively capture the nonlinear correlation and synergistic effect between different LIBS indices; (4) Insufficient utilization of physical parameters: Parameters such as plasma temperature and electron density, which contain rich physical information, are often ignored or used only as independent indices without deep integration with spectral features. Summary of the Invention

[0003] The purpose of this invention is to provide a metal aging assessment method and system based on LIBS multidimensional feature fusion. By constructing a LIBS multidimensional feature fusion system, spectral features, plasma physical parameters and time evolution features are integrated. The attention mechanism in deep learning is used to achieve adaptive weighted fusion of multidimensional features, establish a high-precision material performance assessment model and apply it to the aging assessment of metal materials.

[0004] To achieve the above objectives, the present invention provides a metal aging assessment system based on LIBS multidimensional feature fusion, including a data acquisition module, a data preprocessing module, a multidimensional feature extraction module, a feature fusion module, and a performance prediction module; The data acquisition module is used to acquire LIBS spectral data of the surface of the material to be tested, including setting laser parameters and spectral acquisition parameters. The laser parameters include energy, pulse width, and focusing conditions, and the spectral acquisition parameters include wavelength range, resolution, and delay time. The data preprocessing module is used to perform quality control on the raw spectrum, including noise reduction, baseline correction, intensity normalization, and outlier spectrum removal. The multidimensional feature extraction module is used to extract three levels of features from the preprocessed spectral data, including a spectral feature layer, a physical parameter layer, and a time evolution feature layer. Spectral feature layer: includes the intensity of several characteristic spectral lines, the intensity ratio of ionic / atomic spectral lines, the intensity ratio between different elements, the peak width and peak position of spectral lines; Physical parameter layer: Calculates plasma temperature, electron density, and plasma lifetime based on physical models; Time evolution feature layer: Extracts the rate of change and decay characteristics of plasma parameters over time; The feature fusion module is based on a deep learning attention mechanism, which adaptively weights and fuses multi-dimensional features, and automatically learns the importance and interrelationship of different features. The performance prediction module selects a neural network model based on the fused feature vector and outputs several performance indicators, including the material's aging level, hardness value, and grain size.

[0005] Based on the above system, the present invention also provides a metal aging assessment method based on LIBS multidimensional feature fusion, comprising the following steps: S1. Collect LIBS spectral data of the material to be tested; S2. Data preprocessing: Quality control and standardization of LIBS spectral data, including noise reduction, baseline correction, normalization and outlier removal. S3. Multidimensional feature extraction: Extract three levels of features from the preprocessed spectral data, namely the spectral feature layer, the physical parameter layer, and the time evolution feature layer; S4. Feature Fusion: Based on the deep learning attention mechanism, the multi-dimensional features in S3 are adaptively weighted and fused to automatically learn the importance and interrelationship of different features; S5. Material performance prediction: Based on the fused features, select the corresponding neural network model according to the downstream task, and output the material performance indicators. S6. Model Training and Optimization: Train and optimize the neural network model using sample data with known labels; S7. Model Evaluation and Application: The trained model is evaluated on an independent test set. After passing the evaluation, it is deployed to the actual detection system.

[0006] Preferably, in S1, the following steps are taken: the surface of the material to be tested is sampled using a LIBS system, and laser parameters and spectral acquisition parameters are set; a pulsed laser is focused on the surface of the material, ablation generates high-temperature and high-pressure plasma, and the plasma emission spectrum is collected by a fiber-coupled spectrometer; the spectrum is collected several times at each measurement point according to the actual situation and averaged to reduce random errors.

[0007] Preferably, S2 specifically includes the following steps: S21. Wavelet transform denoising: Daubechies wavelet is used to decompose and reconstruct the spectral signal to remove high-frequency noise; S22. Baseline Correction: Spectral baseline drift is removed using one of iterative polynomial fitting or asymptotic least squares method. S23. Normalization: Using one of Z-score normalization and maximum value normalization, intensity differences under different measurement conditions are eliminated. ; in I For original strength, μ σ and σ' are the mean and standard deviation, respectively; S24. Outlier Removal: Outlier detection based on the 3σ criterion and principal component analysis, removing abnormal spectra.

[0008] Preferably, step S3 specifically includes the following steps: S31. Extract features from the spectral feature layer: Characteristic spectral line intensity: Identify elemental characteristic spectral lines related to material aging and extract their peak intensity. I i ; Intensity ratio characteristics: Calculate the intensity ratio of ionic / atomic spectral lines, as well as the intensity ratio between different elements, to reflect changes in the chemical state and microstructure of the material; Spectral morphology characteristics: Peak width, peak position, and peak area were extracted by fitting spectral lines using Lorentzian and Voigt functions to reflect elemental distribution and lattice distortion; S32. Extract features from the physical parameter layer: plasma temperature T e The Boltzmann plot method is used for calculation, selecting spectral lines of the same element, based on the Boltzmann distribution equation: ; in, For spectral line intensity, For wavelength, For statistical weighting, For the probability of jump, E k For higher energy levels, k This is the Boltzmann constant; obtained by linear fitting of the slope. T e ; electron density n eCalculations based on the Stark broadening effect; isolated atomic / ion spectral lines were selected, and their full width at half maximum (FWHM) was measured. ,according to: ; in w The electron collision broadening parameter is used to calculate... n e ; Plasma lifetime: In time-resolved LIBS mode, the decay curve of plasma emission intensity over time is recorded, and the decay time constant is obtained by fitting. S33. Extracting features from the time evolution feature layer: Spectra were acquired at different delay times, and the rate of temperature change over time was extracted. dT e / dt Electron density decay rate dn e / dt And the time evolution curves of characteristic spectral line intensities; these characteristics reflect plasma dynamics processes and are closely related to the surface state and microstructure of the material. S34. Combine the features from the above three levels into a high-dimensional feature vector. ,in N The total number of features.

[0009] Preferably, in S4, the deep learning attention mechanism is the multi-head self-attention mechanism in the Transformer architecture, and the specific fusion process is as follows: S41. Feature Embedding: Mapping the feature vector F to a linear transformation. d Dimensional space: ; in, For embedding weight matrix; S42. Multi-head self-attention calculation: The embedded features are linearly projected into query, key, and value matrices respectively. ; Calculate the attention score and weighted output: ; in d k The dimension is the key vector; multi-head attention mechanism for parallel computation. h Pay attention to each point, then piece them together: ; S43, Feedforward Network: Two fully connected layers are added after the attention layer, introducing nonlinearity. ; S44, Residual Connections and Layer Normalization: Apply residual connections and layer normalization after each sublayer to stabilize training. ; S45. By stacking several Transformer blocks, the neural network model learns the complex correlations between features; the attention weights automatically reflect the contribution of different features to the target task, achieving adaptive weighted fusion.

[0010] Preferably, in S5, the downstream tasks include classification and regression tasks; the classification task determines the aging level, using the softmax function to output the probability distribution of each category and determine the predicted category; the regression task predicts parameters including hardness and grain size, outputting specific values ​​through a fully connected layer; specifically: S51, The fused feature representation H (L) , L For the final layer, input is fed into the task-specific prediction head; S52. For classification tasks: ,in For each category's probability distribution, M The number of categories; the predicted categories are: ; S53. For regression tasks: .

[0011] Preferably, in S6, the specific steps are as follows: S61. Constructing the loss function: The classification task uses cross-entropy loss: ; The regression task uses mean squared error loss: ; in B Batch size; S62. Update model parameters using the AdamW optimizer: ; Among them, η Let m be the learning rate. t and v t λ represents the first and second moments of the gradient, and λ is the weight decay coefficient. S63. Develop training strategies, specifically including: Dataset partitioning: Divide the dataset into training, validation, and test sets according to a certain ratio; Data augmentation: Applying slight perturbations to spectral data enhances robustness; Early stopping strategy: Monitor the performance on the validation set and stop training when there is no improvement after several consecutive epochs; Learning rate scheduling: The learning rate is dynamically adjusted using one of the cosine annealing and step descent strategies.

[0012] Preferably, in S7, the evaluation metric used is classification accuracy.

[0013] Therefore, this invention adopts the aforementioned metal aging assessment method and system based on LIBS multidimensional feature fusion. By constructing a LIBS multidimensional feature fusion system, it integrates spectral features, plasma physical parameters, and time evolution features. It utilizes the attention mechanism in deep learning to achieve adaptive weighted fusion of multidimensional features, establishing a high-precision material performance assessment model. Information is fully utilized by systematically extracting multidimensional features from LIBS data, maximizing information utilization and overcoming the limitations of single-index assessment. Multidimensional feature collaborative modeling improves prediction accuracy. Complementary fusion of multiple indices effectively reduces the susceptibility of single indices to interference, enhancing the model's stability under different measurement conditions. LIBS physical mechanisms, such as Boltzmann distribution and Stark broadening theory, are embedded in the feature extraction process, improving the model's physical interpretability and generalization ability.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This is a LIBS multidimensional feature fusion and inference network diagram according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 3 The metallographic structures of (a) aging level 0 (original sample) and (b) aging level 3 samples according to embodiments of the present invention are shown. Figure 4 This is a hardness measurement diagram of a sample with aging level 3 according to an embodiment of the present invention; Figure 5 The variation of spectral intensity of matrix elements (Fe I 373.486) and alloying elements (VI 321.243) of P92 steel with the number of laser ablation cycles is shown in the aging grade 3 sample of this invention. Figure 6 These are the LIBS spectral data corresponding to aging levels 1, 2 and 3 of the present invention, respectively, as shown in Figures (a), (b) and (c). Figure 7 The average spectral intensity of the matrix element Fe I (373.486 nm) and the alloy element VI (321.243 nm) of the sample with aging level 3 in this embodiment of the invention varies with the duration of aging. Figure 8The results are the correlation analysis results of the intensity of the matrix element (Fe I 373.486nm) and alloy element spectral line (VI 321.243) and the aging state of P92 steel in the embodiments of the present invention. Figure 9 This is the correlation analysis result between the intensity ratio of the matrix element ionic and atomic spectral lines of the aging grade 3 sample in this embodiment of the invention and the aging state; Figure 10 This is the correlation analysis result of the plasma temperature and aging state of the matrix element (Fe II 263.104) in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] Example 1 This invention provides a metal aging assessment system based on LIBS multidimensional feature fusion, wherein the LIBS multidimensional feature fusion and inference network are as follows: Figure 1 As shown, it includes a data acquisition module, a data preprocessing module, a multi-dimensional feature extraction module, a feature fusion module, and a performance prediction module.

[0019] The data acquisition module is used to acquire LIBS spectral data of the surface of the material under test, including setting laser parameters and spectral acquisition parameters. Laser parameters include energy, pulse width, and focusing conditions, while spectral acquisition parameters include wavelength range, resolution, and delay time.

[0020] The data preprocessing module is used for quality control of the raw spectra, including noise reduction, baseline correction, intensity normalization, and removal of outlier spectra.

[0021] The multidimensional feature extraction module is used to extract three levels of features from the preprocessed spectral data, including the spectral feature layer, the physical parameter layer, and the time evolution feature layer. The spectral feature layer includes the intensity of several characteristic spectral lines, the intensity ratio of ionic / atomic spectral lines, the intensity ratio between different elements, the peak width, and the peak position. The physical parameter layer calculates plasma temperature, electron density, and plasma lifetime based on a physical model. The time evolution feature layer extracts the rate of change and decay characteristics of plasma parameters over time.

[0022] The feature fusion module is based on a deep learning attention mechanism. It adaptively weights and fuses multi-dimensional features, automatically learning the importance and interrelationships of different features.

[0023] The performance prediction module selects a neural network model based on the fused feature vectors, and outputs several performance indicators, including the material's aging level, hardness value, and grain size.

[0024] The workflow of the metal aging assessment method based on LIBS multidimensional feature fusion is as follows: Figure 2 As shown, it includes the following steps: S1. Acquire LIBS spectral data of the material to be tested; specifically: acquire the spectrum of the surface of the material to be tested through the LIBS system, set the laser parameters and spectral acquisition parameters; focus the pulsed laser on the surface of the material, ablate to generate high temperature and high pressure plasma, and acquire the plasma emission spectrum through a fiber-coupled spectrometer; acquire the spectrum of each measurement point several times according to the actual situation and take the average to reduce random error.

[0025] This embodiment uses P92, a high-temperature pressure-bearing pipeline material in a supercritical thermal power unit, as the object of detection and prediction. P92 samples with different aging levels were artificially prepared. Specifically, the original material samples were treated in a muffle furnace at 625℃ for 100h, 1000h, and 3000h, respectively labeled as aging levels 1, 2, and 3; for comparison, the original material was labeled as aging level 0. Metallographic samples of materials with different aging levels were prepared, and their metallographic structures were obtained using an optical microscope; as an example, Figure 3 The metallographic structures of samples with aging level 0 (original sample) and aging level 3 are shown. The hardness values ​​of samples with different aging levels were measured using a hardness tester. Figure 4This section demonstrates an example of measuring the hardness of a sample at aging level 3. Materials at different aging levels were tested using a LIBS instrument to acquire LIBS spectral data. Before testing, the samples were cleaned with organic solvents such as anhydrous ethanol or acetone to reduce interference from external contaminants on the LIBS test data. The laser energy was adjusted to 80 mJ, the laser frequency to 2 Hz, and the laser focus point was set approximately 2 mm below the sample surface. The diameter of the laser spot on the sample surface was approximately 120 μm. Ten points were measured on each sample, and each point was ablated 60 times with the laser. The average value was calculated to reduce random errors. The wavelength range for LIBS detection was selected as 200nm~500nm, which includes the spectral lines of the matrix elements and main alloying elements of P92 steel. To study the evolution characteristics of laser spectral lines over time, a delay trigger was used to set 10 different delay durations, including 100ns, 200ns, 400ns, 600ns, 800ns, 1000ns, 1500ns, 2000ns, 2500ns, and 3000ns. LIBS spectral data were collected at different delay times. By processing the spectral data, the evolution of the spectral line intensity of the matrix elements and alloying elements over time could be obtained. Considering that the spectral information obtained from approximately the first 15 laser ablations contained interference information caused by the sample surface quality, this embodiment discarded it; the spectral information images obtained from laser ablations 16-60 were averaged to obtain a stable spectral image. In this embodiment, a dataset of 4×10×10=400 (aging level, number of measurement points for each sample, and 10 different delay durations) LIBS spectra was obtained. Figure 5 The variation of Fe I and VI atomic spectral line intensities with the number of laser pulses was shown; Figure 6 LIBS spectral data for samples with aging levels 1, 2 and 3; Figure 7 The variation of the average spectral intensity of the matrix element Fe I (373.486 nm) and the alloy element VI (321.243 nm) with the aging time of the sample of aging level 3 is shown.

[0026] S2. Data Preprocessing: Quality control and standardization of LIBS spectral data are performed, including denoising, baseline correction, normalization, and outlier removal; specifically, the following steps are included: S21. Wavelet transform denoising: Daubechies wavelet is used to decompose and reconstruct the spectral signal to remove high-frequency noise; S22. Baseline Correction: Spectral baseline drift is removed using one of iterative polynomial fitting or asymptotic least squares method. S23. Normalization: Using one of Z-score normalization and maximum value normalization, intensity differences under different measurement conditions are eliminated. ; in I For original strength, μ σ and σ' are the mean and standard deviation, respectively; S24. Outlier Removal: Outlier detection based on the 3σ criterion and principal component analysis, removing abnormal spectra.

[0027] S3. Multidimensional Feature Extraction: Extracting three levels of features from the preprocessed spectral data: spectral feature layer, physical parameter layer, and time evolution feature layer; specifically including the following steps: S31. Extract features from the spectral feature layer: Characteristic spectral line intensities: Identify elemental characteristic spectral lines associated with material aging, such as Fe, Cr, and Ni, and extract their peak intensities. I i ; This embodiment confirms that the matrix element Fe and alloying elements such as V and Cr are all highly correlated with the aging level of the material. Figure 8 The results of the correlation analysis between the spectral line intensities of matrix elements and alloying elements and the aging state of P92 steel are presented.

[0028] Intensity ratio characteristics: Calculate the intensity ratio of ionic / atomic spectral lines, as well as the intensity ratio between different elements, such as CrI / FeI, to reflect the changes in the chemical state and microstructure of the material; This embodiment calculates the intensity ratio of the ionic and atomic spectral lines of the matrix element and the alloy element, as well as the intensity ratio of the atomic spectral lines of the matrix element and the ionic spectral lines of the alloy element, based on the spectral line intensity information of characteristic elements under different samples, different measuring points, and different delay times. This value can reflect the matrix characteristics of the material, that is, the aging state. Figure 9 The results of the correlation analysis between the intensity ratio of the matrix element ionic state (Fe II 263.104) and atomic state spectral lines (Fe I 249.064) and the aging state are presented.

[0029] Spectral morphology characteristics: Peak width, peak position, and peak area were extracted by fitting spectral lines using Lorentzian and Voigt functions to reflect elemental distribution and lattice distortion; S32. Extract features from the physical parameter layer: plasma temperature T e The Boltzmann plot method is used for calculation, selecting spectral lines of the same element, based on the Boltzmann distribution equation: ; in, For spectral line intensity, For wavelength, For statistical weighting, For the probability of jump, E k For higher energy levels, k This is the Boltzmann constant; obtained by linear fitting of the slope. T e ; electron density n e Calculations based on the Stark broadening effect; isolated atomic / ion spectral lines were selected, and their full width at half maximum (FWHM) was measured. ,according to: ; in w The electron collision broadening parameters are obtained by looking up a table, and are then calculated. n e ; Plasma lifetime: In time-resolved LIBS mode, the decay curve of plasma emission intensity over time is recorded, and the decay time constant is obtained by fitting.

[0030] In this embodiment, the Boltzmann plot method was used to calculate the plasma temperature of the matrix element ion states in samples under four aging conditions, and the electron density was calculated using the Stark broadening effect. Based on the decay curve of plasma emission intensity over time, the decay time constant was determined by fitting, and the plasma lifetime was further determined. Figure 10 The results of the correlation analysis between plasma temperature and aging state are presented.

[0031] S33. Extracting features from the time evolution feature layer: Spectra were acquired at different delay times, and the rate of temperature change over time was extracted. dT e / dt Electron density decay rate dn e / dt And the time evolution curves of characteristic spectral line intensities; these characteristics reflect plasma dynamics processes and are closely related to the surface state and microstructure of the material. S34. Combine the features from the above three levels into a high-dimensional feature vector. ,in N The total number of features.

[0032] S4. Feature Fusion: Based on a deep learning attention mechanism, the multi-dimensional features in S3 are adaptively weighted and fused to automatically learn the importance and interrelationships of different features. The deep learning attention mechanism is the multi-head self-attention mechanism in the Transformer architecture. The specific fusion process is as follows: S41. Feature Embedding: Mapping the feature vector F to a linear transformation. d 3D space, in this embodiment d= 1024: ; in, For embedding weight matrix; S42. Multi-head self-attention calculation: The embedded features are linearly projected into query, key, and value matrices respectively. ; Calculate the attention score and weighted output: ; in d k =128 is the dimension of the key vector; multi-head attention mechanism for parallel computation. h =8 attention points, then spliced ​​together: ; S43, Feedforward Network: Two fully connected layers are added after the attention layer, introducing nonlinearity. ; S44, Residual Connections and Layer Normalization: Apply residual connections and layer normalization after each sublayer to stabilize training. ; S45. By stacking several Transformer blocks, the neural network model learns the complex correlations between features; the attention weights automatically reflect the contribution of different features to the target task, achieving adaptive weighted fusion.

[0033] S5. Material Performance Prediction: Based on the fused features, a corresponding neural network model is selected according to the downstream task to output the material's performance indicators. Downstream tasks include classification and regression tasks. The classification task determines the aging level, using a softmax function to output the probability distribution of each category and determine the predicted category. The regression task predicts parameters including hardness and grain size, outputting specific values ​​through a fully connected layer. Specifically: S51, The fused feature representation H (L) , L For the final layer, input is fed into the task-specific prediction head; S52. For classification tasks: ,in For each category's probability distribution, M The number of categories; the predicted categories are: ; S53. For regression tasks: .

[0034] S6. Model Training and Optimization: Train and optimize the neural network model using sample data with known labels; the specific steps are as follows: S61. Constructing the loss function: The classification task uses cross-entropy loss: ; The regression task uses mean squared error loss: ; in B =32 is the batch size; S62. Update model parameters using the AdamW optimizer: ; Among them, η Let m be the learning rate. t and v t The first and second moments of the gradient are estimated, and λ is the weight decay coefficient; in this embodiment... η =1e -4 , λ=0.01.

[0035] S63. Develop training strategies, specifically including: Dataset partitioning: The dataset is divided into training, validation, and test sets according to a certain ratio; in this embodiment, the specific ratio is 7:2:1, and the specific quantities are 280, 80, and 40 respectively.

[0036] Data augmentation: Applying slight perturbations to spectral data enhances robustness; Early stopping strategy: Monitor validation set performance and stop training when there is no improvement after 10 consecutive epochs; Learning rate scheduling: The learning rate is dynamically adjusted using cosine annealing or step-down strategies.

[0037] S7. Model Evaluation and Application: The trained model is evaluated on an independent test set. After passing the evaluation, it is deployed to the actual detection system. The evaluation metrics used include classification accuracy.

[0038] In this embodiment, based on the aging level assessment dataset, the training / validation / test sets are divided in a 7:2:1 ratio, and this method is compared with existing methods.

[0039] This method achieves an accuracy of 90% on the test set, which is a significant improvement in prediction performance compared to traditional single-index methods and multivariate prediction methods, and outperforms existing technologies.

[0040] Therefore, this invention adopts the aforementioned metal aging assessment method and system based on LIBS multidimensional feature fusion. By constructing a LIBS multidimensional feature fusion system, it integrates spectral features, plasma physical parameters, and time evolution features. It utilizes the attention mechanism in deep learning to achieve adaptive weighted fusion of multidimensional features, establishing a high-precision material performance assessment model. Information is fully utilized by systematically extracting multidimensional features from LIBS data, maximizing information utilization and overcoming the limitations of single-index assessment. Multidimensional feature collaborative modeling improves prediction accuracy. Complementary fusion of multiple indices effectively reduces the susceptibility of single indices to interference, enhancing the model's stability under different measurement conditions. LIBS physical mechanisms, such as Boltzmann distribution and Stark broadening theory, are embedded in the feature extraction process, improving the model's physical interpretability and generalization ability.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A metal aging assessment system based on LIBS multidimensional feature fusion, characterized in that, It includes a data acquisition module, a data preprocessing module, a multi-dimensional feature extraction module, a feature fusion module, and a performance prediction module; The data acquisition module is used to acquire LIBS spectral data of the surface of the material to be tested, including setting laser parameters and spectral acquisition parameters. The laser parameters include energy, pulse width, and focusing conditions, and the spectral acquisition parameters include wavelength range, resolution, and delay time. The data preprocessing module is used to perform quality control on the raw spectrum, including noise reduction, baseline correction, intensity normalization, and outlier spectrum removal. The multidimensional feature extraction module is used to extract three levels of features from the preprocessed spectral data, including a spectral feature layer, a physical parameter layer, and a time evolution feature layer. Spectral feature layer: includes the intensity of several characteristic spectral lines, the intensity ratio of ionic / atomic spectral lines, the intensity ratio between different elements, the peak width and peak position of spectral lines; Physical parameter layer: Calculates plasma temperature, electron density, and plasma lifetime based on physical models; Time evolution feature layer: Extracts the rate of change and decay characteristics of plasma parameters over time; The feature fusion module is based on a deep learning attention mechanism, which adaptively weights and fuses multi-dimensional features, and automatically learns the importance and interrelationship of different features. The performance prediction module selects a neural network model based on the fused feature vector and outputs several performance indicators, including the material's aging level, hardness value, and grain size.

2. A metal aging assessment method based on LIBS multidimensional feature fusion, characterized in that, Includes the following steps: S1. Collect LIBS spectral data of the material to be tested; S2. Data preprocessing: Quality control and standardization of LIBS spectral data, including noise reduction, baseline correction, normalization and outlier removal. S3. Multidimensional feature extraction: Extract three levels of features from the preprocessed spectral data, namely the spectral feature layer, the physical parameter layer, and the time evolution feature layer; S4. Feature Fusion: Based on the deep learning attention mechanism, the multi-dimensional features in S3 are adaptively weighted and fused to automatically learn the importance and interrelationship of different features; S5. Material performance prediction: Based on the fused features, select the corresponding neural network model according to the downstream task, and output the material performance indicators. S6. Model Training and Optimization: Train and optimize the neural network model using sample data with known labels; S7. Model Evaluation and Application: The trained model is evaluated on an independent test set. After passing the evaluation, it is deployed to the actual detection system.

3. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 2, characterized in that, In S1, specifically: the LIBS system is used to collect the spectrum of the material surface under test, and the laser parameters and spectral acquisition parameters are set; the pulsed laser is focused on the material surface, and ablation generates high-temperature and high-pressure plasma, which is then collected by a fiber-coupled spectrometer; the spectrum of each measurement point is collected several times according to the actual situation and averaged to reduce random errors.

4. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 3, characterized in that, S2 specifically includes the following steps: S21. Wavelet transform denoising: Daubechies wavelet is used to decompose and reconstruct the spectral signal to remove high-frequency noise; S22. Baseline Correction: Spectral baseline drift is removed using one of iterative polynomial fitting or asymptotic least squares method. S23. Normalization: Using one of Z-score normalization and maximum value normalization, intensity differences under different measurement conditions are eliminated. ； in I For original strength, μ σ and σ' are the mean and standard deviation, respectively; S24. Outlier Removal: Outlier detection based on the 3σ criterion and principal component analysis, removing abnormal spectra.

5. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 4, characterized in that, S3 specifically includes the following steps: S31. Extract features from the spectral feature layer: Characteristic spectral line intensity: Identify elemental characteristic spectral lines related to material aging and extract their peak intensity. I i ; Intensity ratio characteristics: Calculate the intensity ratio of ionic / atomic spectral lines, as well as the intensity ratio between different elements, to reflect changes in the chemical state and microstructure of the material; Spectral morphology characteristics: Peak width, peak position, and peak area were extracted by fitting spectral lines using Lorentzian and Voigt functions to reflect elemental distribution and lattice distortion; S32. Extract features from the physical parameter layer: plasma temperature T e The Boltzmann plot method is used for calculation, selecting spectral lines of the same element, based on the Boltzmann distribution equation: ； in, For spectral line intensity, For wavelength, For statistical weighting, For the probability of jump, E k For higher energy levels, k This is the Boltzmann constant; obtained by linear fitting of the slope. T e ; electron density n e Calculations based on the Stark broadening effect; isolated atomic / ion spectral lines were selected, and their full width at half maximum (FWHM) was measured. ,according to: ； in w The electron collision broadening parameter is used to calculate... n e ; Plasma lifetime: In time-resolved LIBS mode, the decay curve of plasma emission intensity over time is recorded, and the decay time constant is obtained by fitting. S33. Extracting features from the time evolution feature layer: Spectra were acquired at different delay times, and the rate of temperature change over time was extracted. dT e / dt Electron density decay rate dn e / dt And the time evolution curves of characteristic spectral line intensities; these characteristics reflect plasma dynamics processes and are closely related to the surface state and microstructure of the material. S34. Combine the features from the above three levels into a high-dimensional feature vector. ,in N The total number of features.

6. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 5, characterized in that, In S4, the deep learning attention mechanism is the multi-head self-attention mechanism in the Transformer architecture, and the specific fusion process is as follows: S41. Feature Embedding: Mapping the feature vector F to a linear transformation. d Dimensional space: ; in, For embedding weight matrix; S42. Multi-head self-attention calculation: The embedded features are linearly projected into query, key, and value matrices respectively. ; Calculate the attention score and weighted output: ； in d k The dimension is the key vector; multi-head attention mechanism for parallel computation. h Pay attention to each point, then piece them together: ； S43, Feedforward Network: Two fully connected layers are added after the attention layer, introducing nonlinearity. ; S44, Residual Connections and Layer Normalization: Apply residual connections and layer normalization after each sublayer to stabilize training. ; S45. By stacking several Transformer blocks, the neural network model learns the complex correlations between features; the attention weights automatically reflect the contribution of different features to the target task, achieving adaptive weighted fusion.

7. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 6, characterized in that, In S5, downstream tasks include classification and regression tasks. The classification task determines the aging level, using a softmax function to output the probability distribution of each category and determine the predicted category. The regression task predicts parameters including hardness and grain size, outputting specific values ​​through a fully connected layer. Specifically: S51, The fused feature representation H (L) , L For the final layer, input is fed into the task-specific prediction head; S52. For classification tasks: ,in For each category's probability distribution, M The number of categories; the predicted categories are: ; S53. For regression tasks: .

8. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 7, characterized in that, In S6, the specific steps are as follows: S61. Constructing the loss function: The classification task uses cross-entropy loss: ; The regression task uses mean squared error loss: ; in B Batch size; S62. Update model parameters using the AdamW optimizer: ； Among them, η Let m be the learning rate. t and v t λ represents the first and second moments of the gradient, and λ is the weight decay coefficient. S63. Develop training strategies, specifically including: Dataset partitioning: Divide the dataset into training, validation, and test sets according to a certain ratio; Data augmentation: Applying slight perturbations to spectral data enhances robustness; Early stopping strategy: Monitor the performance on the validation set and stop training when there is no improvement after several consecutive epochs; Learning rate scheduling: The learning rate is dynamically adjusted using one of the cosine annealing and step descent strategies.

9. The metal aging assessment method based on LIBS multidimensional feature fusion according to claim 2, characterized in that, In S7, the evaluation metric used is classification accuracy.

Images