Alloy identification system and method based on deep recurrent neural network in combination with fs-LA-SIBS

Abstract

The invention discloses an alloy identification system and method based on a deep recurrent neural network in combination with fs-LA-SIBS, and belongs to the technical field of alloy material classification and identification. The system comprises an excitation module, a spectrum detection module and a data processing module, the excitation module is used for emitting a laser pulse, and the laser pulse acts on an alloy to be detected; the spectrum detection module is connected with the excitation module and is used for collecting spectrum data of the to-be-detected alloy subjected to the laser pulse effect; and the data processing module is connected with the spectrum detection module and is used for processing the spectrum data to obtain an alloy classification result. According to the method, the spectrum classification precision is ensured, and an extensible technical basis is provided for algorithm rapid iteration and multi-scene adaptation in an industrial scene.

Description

Technical Field

[0001] This invention belongs to the field of alloy material classification and identification technology, specifically involving an alloy identification system and method based on deep recurrent neural network combined with fs-LA-SIBS. Background Technology

[0002] Precise alloy classification is fundamental to ensuring stable material properties, quality and safety, and efficient resource utilization in industries such as aerospace and automotive manufacturing. The main existing alloy classification technologies include:

[0003] 1. Traditional alloy testing methods:

[0004] (1) X-ray fluorescence spectroscopy (XRF) and spark emission spectroscopy (OES) are commonly used alloy detection techniques in industry, but they have problems such as poor real-time performance, high cost, and insufficient sensitivity to trace elements.

[0005] (2) Laser-induced breakdown spectroscopy (LIBS) has gradually become the mainstream technology due to its rapid multi-element detection capability, lack of complex sample preparation and non-contact characteristics. However, traditional LIBS is easily affected by plasma fluctuations, resulting in insufficient spectral stability; and the spectral signal-to-noise ratio is low, making it difficult to capture the weak elemental characteristics of complex alloys.

[0006] 2. Combining Machine Learning with LIBS:

[0007] Existing methods such as Random Forest (RF), Support Vector Machine (SVM), and Partial Least Squares Discriminant Analysis (PLS-DA) rely on manual feature extraction, which makes it difficult to model the dynamic features of spectral time series, thus limiting classification accuracy.

[0008] Therefore, existing technologies lack a deep fusion of global statistical features and local temporal features of the spectrum, resulting in insufficient model robustness. Summary of the Invention

[0009] This invention aims to address the shortcomings of existing technologies by proposing an alloy identification method based on deep recurrent neural networks combined with fs-LA-SIBS, which improves signal sensitivity and anti-interference capabilities. Furthermore, by incorporating a deep recurrent neural network (DeepRNN) framework and fusing spectral temporal features with statistical features, a breakthrough in alloy classification accuracy is achieved.

[0010] To achieve the above objectives, the present invention provides the following solution: an alloy identification system based on a deep recurrent neural network combined with fs-LA-SIBS, comprising: an excitation module, a spectral detection module, and a data processing module;

[0011] The excitation module is used to emit laser pulses, which act on the alloy to be tested;

[0012] The spectral detection module is connected to the excitation module and is used to collect the spectral data of the alloy to be tested after being subjected to laser pulses.

[0013] The data processing module is connected to the spectral detection module and is used to process the spectral data to obtain alloy classification results.

[0014] More preferably, the excitation module includes a femtosecond laser unit and a spark discharge re-excitation unit;

[0015] The femtosecond laser unit and the spark discharge re-excitation unit work together to suppress thermal interference using ultrashort pulses and to re-excite the plasma using discharge energy.

[0016] More preferably, the data processing module includes: a preprocessing unit and a feature extraction unit;

[0017] The preprocessing unit is used to preprocess the spectral data to obtain processed spectral data;

[0018] The feature extraction unit is connected to the preprocessing unit and is used to extract features from the processed spectral data to obtain the classification result.

[0019] More preferably, the preprocessing unit includes: a normalization subunit, a first extraction subunit, a statistics subunit, and a combination subunit;

[0020] The normalization subunit is used to normalize the spectral data to obtain normalized data;

[0021] The method for obtaining the normalized data by the normalization subunit includes: normalizing the wavelength and intensity of the spectral data to the [0,1] interval to obtain the normalized data.

[0022] The first extraction subunit is connected to the normalization subunit and is used to extract peak features from the normalized data to obtain extracted data; the extracted data includes: mean peak intensity.

[0023] The statistical subunit is connected to the normalization subunit and is used to obtain statistical data based on the normalized data;

[0024] The combination subunit is connected to the normalization subunit, the first extraction subunit, and the statistics subunit, and is used to combine the normalized data, the extracted data, and the statistical data to obtain processed spectral data.

[0025] More preferably, the method by which the first extraction subunit obtains the extracted data includes:

[0026] Calculate the peak threshold based on the normalized data, and obtain the number of peaks based on the peak threshold;

[0027] The mean peak intensity is obtained based on the number of peaks and the normalized data:

[0028]

[0029] In the formula, μ peak N represents the mean peak intensity. peak Indicates the number of peak values; I norm This represents the intensity normalization result; This represents the normalized intensity value of the peak value.

[0030] More preferably, the method by which the statistical subunit obtains the statistical data includes:

[0031] Calculate the global intensity mean based on the normalized data:

[0032]

[0033] In the formula, μ global L represents the global intensity mean; L represents the total number of normalized intensity data points; I represents the global intensity mean. norm This represents the intensity normalization result;

[0034] Calculate the global intensity standard deviation based on the global intensity mean and the normalized data:

[0035]

[0036] More preferably, the feature extraction unit includes: an input layer, a recurrent layer, a feature fusion subunit, a fully connected layer, and an output layer;

[0037] The input layer is used to process the spectral data to obtain splicing features;

[0038] The recurrent layer is used to perform temporal analysis on the spliced ​​features to obtain temporal features;

[0039] The feature fusion subunit is used to concatenate the spliced ​​features and the temporal features to obtain a joint feature vector;

[0040] The fully connected layer is used to perform multi-level linear transformations on the joint feature vector to obtain dimensionality-reduced features;

[0041] The output layer is used to generate the classification result based on the dimensionality-reduced features normalized by Softmax.

[0042] This invention also provides an alloy identification method based on a deep recurrent neural network combined with fs-LA-SIBS, comprising the following steps:

[0043] S1. Emit laser pulses to the alloy under test;

[0044] S2. Detect the spectral data of the alloy to be tested after being subjected to laser pulses;

[0045] S3. Process the spectral data to obtain alloy classification results.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention utilizes the combined effect of femtosecond laser ablation (35fs pulse) and spark discharge re-excitation to suppress thermal interference with ultrashort pulses and to re-excite plasma with discharge energy, thereby significantly improving the spectral signal-to-noise ratio and stability, overcoming the shortcomings of traditional LIBS signals such as large fluctuations and insufficient sensitivity for trace element detection.

[0048] This invention designs a bidirectional recurrent neural network architecture, combines time-series modeling to analyze the dynamic evolution of spectra (such as the time correlation of element emission peaks), and integrates statistical features such as the number of peaks and the global mean to construct a noise-resistant multimodal feature space, breaking through the classification accuracy bottleneck of traditional models that rely on static features.

[0049] This invention achieves decoupling of data loading, model building, training, and evaluation processes through modular design. It adopts standardized interfaces and parameterized configurations (supporting flexible switching of key parameters such as RNN type, hidden layer dimension, and learning rate), combined with a multi-dataset loop experiment framework and automated evaluation scripts in the main function, significantly improving model tuning efficiency. While ensuring the accuracy of spectral classification, it provides a scalable technical foundation for rapid algorithm iteration and multi-scenario adaptation in industrial applications. Attached Figure Description

[0050] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the alloy identification method based on deep recurrent neural network combined with fs-LA-SIBS according to an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1:

[0055] This embodiment provides an alloy identification system based on a deep recurrent neural network combined with fs-LA-SIBS, including: an excitation module, a spectral detection module, and a data processing module. The excitation module emits laser pulses, which are applied to the alloy to be detected. The spectral detection module is connected to the excitation module and is used to acquire the spectral data of the alloy after being applied by the laser pulse. In this embodiment, the spectral detection module can use a three-channel fiber optic spectrometer (200-500nm), a non-gated signal detection mode, and supports high-throughput acquisition of 30 spectra / second. The data processing module is connected to the spectral detection module and is used to process the spectral data to obtain the alloy classification results.

[0056] A further implementation involves the excitation module comprising a femtosecond laser unit and a spark discharge re-excitation unit; the femtosecond laser unit and the spark discharge re-excitation unit work together to suppress thermal interference using ultrashort pulses and to re-excite the plasma through discharge energy. Preferably, the laser emission module can employ a Ti:sapphire ultrafast laser (800nm, 35fs, 1kHz).

[0057] A further implementation involves a data processing module comprising a preprocessing unit and a feature extraction unit. The preprocessing unit is used to preprocess the spectral data to obtain processed spectral data. The feature extraction unit is connected to the preprocessing unit and is used to extract features from the processed spectral data to obtain classification results.

[0058] Specifically, the preprocessing unit includes: a normalization subunit, a first extraction subunit, a statistical subunit, and a combination subunit. The normalization subunit is used to normalize the spectral data to obtain normalized data. The method by which the normalization subunit obtains normalized data includes: normalizing the wavelength λ and intensity I channels of the spectral data to the [0,1] interval to obtain normalized data. Wavelength and intensity are normalized independently to eliminate dimensional differences and provide standardized input for subsequent peak detection and statistical feature extraction.

[0059] For the original wavelength sequence λ, calculate its minimum value min(λ) and maximum value max(λ), and normalize it according to the following formula:

[0060]

[0061] For the original intensity sequence I, calculate its minimum value min(I) and maximum value max(I), and normalize it according to the following formula:

[0062]

[0063] The first extraction subunit is connected to the normalization subunit and is used to extract peak features from the normalized data to obtain extracted data. The extracted data includes the mean peak intensity. The method is as follows: calculate the peak threshold based on the normalized data, and obtain the number of peaks based on the peak threshold.

[0064] Specifically, based on the normalized intensity sequence I norm A threshold method is used to detect the number of peaks and the mean peak intensity. Dynamic thresholding adapts to the intensity distribution of different samples, improving feature robustness.

[0065] In this embodiment, the peak intensity threshold is set to 80% of the normalized maximum intensity (i.e., the threshold is 0.8), and the peak positions of all intensity values ​​exceeding this threshold are detected.

[0066] Threshold=0.8×max(I norm ).

[0067] The number of peak values ​​N detected peak And calculate the average of all peak intensities; if there are no peaks, the average peak intensity is set to 0:

[0068]

[0069] In the formula, μ peak N represents the mean peak intensity. peak Indicates the number of peak values; I norm This represents the intensity normalization result; This represents the normalized intensity value of the peak value.

[0070] The statistical subunit is connected to the normalization subunit, used to obtain statistical data based on the normalized data. Based on the normalized intensity sequence I... norm The global intensity mean and standard deviation are calculated. Statistical characteristics describe the intensity distribution from a global perspective, complementing local peak characteristics.

[0071] First, calculate the global intensity mean to characterize the center of the overall intensity distribution.

[0072]

[0073] In the formula, μ global The value represents the global intensity mean; L represents the total number of normalized intensity data points.

[0074] Next, the global intensity standard deviation is calculated to characterize the degree of intensity fluctuation.

[0075]

[0076] The combination subunit is connected to the normalization subunit, the first extraction subunit, and the statistical subunit, and is used to combine and process the normalized data, extracted data, and statistical data to obtain processed spectral data.

[0077] Specifically, the processed data and features are integrated into the model input, with normalized data as the original input and statistical features as auxiliary features. The two are combined to enhance the model's ability to represent the spectrum at multiple scales.

[0078] (1) Spectral data: The normalized wavelength λ norm and intensity I norm The two columns of data are merged into a two-dimensional array, with each row corresponding to a wavelength-intensity pair.

[0079] (2) The number of peaks N peak Peak intensity mean μ peak Global intensity mean μ global and standard deviation σ global The four values ​​are combined into a one-dimensional feature vector.

[0080] The data is then processed by a feature extraction unit to obtain the classification results. In this embodiment, the feature extraction unit adopts a DeepRNN model architecture, including an input layer, a recurrent layer, a feature fusion subunit, a fully connected layer, and an output layer. Its core design achieves efficient classification of spectral data through multimodal feature fusion and bidirectional temporal modeling. The input layer first performs dual-channel normalization on the original spectral data, mapping wavelength and intensity to the [0,1] interval through minimum-maximum normalization to eliminate dimensional differences. At the same time, a peak detection algorithm is used to extract local features of the spectral sequence (number of peaks N). peak Peak intensity mean μ peak ) and global statistical characteristics (global mean μ) global Standard deviation σ globalThe normalized dual-channel spectral sequence is concatenated with a four-dimensional statistical feature vector to form an initial feature matrix. The recurrent layer employs bidirectional LSTM, GRU, or RNN units, capturing local dynamic patterns and global temporal dependencies of the spectral sequence through forward and backward propagation, respectively. The bidirectional structure can simultaneously model the positive trend of spectral intensity increasing with wavelength and potential negative correlations. Finally, the feature vector of the last time step is extracted from the temporal feature tensor output by the recurrent layer as a global temporal representation. The feature fusion subunit concatenates the temporal feature vector (e.g., 512-dimensional) output by the recurrent layer with the four-dimensional statistical feature vector of the input layer to generate a joint feature vector. This deeply integrates temporal dynamic information with global distribution characteristics, enhancing the model's ability to jointly analyze short-term spectral fluctuations and long-term trends. The fully connected layer consists of two linear transformation modules with ReLU activation functions. The first layer maps the fused high-dimensional features (e.g., 516-dimensional) to 256-dimensionality, and the second layer further reduces the dimensionality to 128-dimensionality. During this process, Dropout with a probability of 0.2 is used to suppress overfitting, progressively extracting discriminative high-level semantic features. The output layer maps the dimensionality-reduced features to the class space through linear transformation and uses the Softmax function to generate a normalized probability distribution. Finally, the model parameters are optimized using the cross-entropy loss function and backpropagation. This architecture, through the synergy of statistical features and recurrent neural networks, introduces a global distribution prior on the basis of bidirectional temporal modeling, effectively balancing the model's robustness to spectral sequence noise and sensitivity to complex patterns, providing a highly efficient and accurate solution for industrial alloy classification.

[0081] Model training and evaluation methods include:

[0082] (1) Training method: The model is trained end-to-end using the Adam optimizer and cross-entropy loss function, and the computation is accelerated by dynamic device adaptation (GPU / CPU). In each training epoch, the data is loaded in batches of 512. After forward propagation, the loss is calculated and the parameters are updated in reverse. At the same time, the average loss is accumulated to monitor the convergence status. After each training round, the loss, accuracy and macro F1 score are evaluated on the validation set to ensure that the model optimization direction is correct.

[0083] (2) Evaluation method: During evaluation, switch to gradient-free mode, disable Dropout and BatchNorm to ensure stability, traverse all batches of the test set and collect prediction results and true labels, and finally jointly calculate the accuracy and macro F1 score. For multiple independent test sets, output classification metrics to verify the model's generalization ability and robustness in complex scenarios.

[0084] Example 2:

[0085] This embodiment provides a detailed description of the solution of the present invention with reference to examples.

[0086] 1. Experimental materials and equipment:

[0087] Nine certified steel alloy standard samples (numbered 0-8) were used, containing trace elements such as Al, Cr, Cu, Mn, Mo, Ni, and V, with mass concentrations ranging from 0.002% to 4.21%. The experimental equipment consisted of a femtosecond laser ablation-spark-induced breakdown spectroscopy (fs-LA-SIBS) system (laser parameters: 800 nm, 35 fs, 1 kHz, single pulse energy 0.5 mJ) and a three-channel fiber optic spectrometer (wavelength range 200-500 nm, resolution <0.1 nm).

[0088] 2. Dataset Partitioning and Preprocessing

[0089] Spectral data from 100 different regions were collected for each sample, with 500 pulses accumulated per region (integration time 10 ms), resulting in a total of 14,804 spectral sequences. The data was divided into a 5:1:4 ratio for training (50%), validation (10%), and test (40%). The test set was further divided into four subsets to verify the model's robustness. After dual-channel normalization, the number of peaks, mean peak intensity, global mean, and standard deviation were extracted as features.

[0090] 3. Model Training and Testing Results

[0091] A unified hyperparameter configuration (512-dimensional hidden layers, 300 training epochs, 512 batch size, 0.0005 learning rate, and 0.2 dropout rate) was used to train GRU, LSTM, and RNN models within the DeepRNN framework. The test set results are shown in Table 1.

[0092] Table 1

[0093]

[0094] Example 3:

[0095] This embodiment illustrates the performance of the proposed solution by comparing it with traditional machine learning methods.

[0096] 1. Experimental Design

[0097] Using the same dataset (14,804 spectral data points) and features (number of peaks, global mean, etc.), we compared Random Forest (RF), Support Vector Machine (SVM), and Partial Least Squares Discriminant Analysis (PLS-DA). The data partitioning was consistent with Example 1 to ensure fairness in the comparison.

[0098] 2. Results Analysis

[0099] As shown in Table 2:

[0100] Table 2

[0101]

[0102] As can be seen, the DeepRNN proposed in this invention significantly outperforms traditional methods (by 7-15%), validating the synergistic advantages of temporal modeling and feature fusion. Traditional methods, relying on static features and unable to capture the dynamic evolution of the spectrum, suffer from limited classification performance.

[0103] Example 4:

[0104] like Figure 1 As shown, this embodiment provides an alloy identification method based on a deep recurrent neural network combined with fs-LA-SIBS, including the following steps:

[0105] S1. Emit laser pulses to the alloy under test;

[0106] S2. Detect the spectral data of the alloy to be tested after being subjected to laser pulses;

[0107] S3. Process the spectral data to obtain the alloy classification results.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An alloy identification system based on deep recurrent neural networks combined with fs-LA-SIBS, characterized in that, include: The module consists of an excitation module, a spectral detection module, and a data processing module. The excitation module is used to emit laser pulses, which act on the alloy to be tested; The spectral detection module is connected to the excitation module and is used to collect the spectral data of the alloy to be tested after being subjected to laser pulses. The data processing module is connected to the spectral detection module and is used to process the spectral data to obtain alloy classification results.

2. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 1, characterized in that, The excitation module includes a femtosecond laser unit and a spark discharge re-excitation unit; The femtosecond laser unit and the spark discharge re-excitation unit work together to suppress thermal interference using ultrashort pulses and to re-excite the plasma using discharge energy.

3. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 1, characterized in that, The data processing module includes: a preprocessing unit and a feature extraction unit; The preprocessing unit is used to preprocess the spectral data to obtain processed spectral data; The feature extraction unit is connected to the preprocessing unit and is used to extract features from the processed spectral data to obtain the classification result.

4. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 3, characterized in that, The preprocessing unit includes: a normalization subunit, a first extraction subunit, a statistical subunit, and a combination subunit; The normalization subunit is used to normalize the spectral data to obtain normalized data; The method for obtaining the normalized data by the normalization subunit includes: normalizing the wavelength and intensity of the spectral data to the [0,1] interval to obtain the normalized data. The first extraction subunit is connected to the normalization subunit and is used to extract peak features from the normalized data to obtain extracted data; the extracted data includes: mean peak intensity. The statistical subunit is connected to the normalization subunit and is used to obtain statistical data based on the normalized data; The combination subunit is connected to the normalization subunit, the first extraction subunit, and the statistics subunit, and is used to combine the normalized data, the extracted data, and the statistical data to obtain processed spectral data.

5. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 4, characterized in that, The method by which the first extraction subunit obtains the extracted data includes: Calculate the peak threshold based on the normalized data, and obtain the number of peaks based on the peak threshold; The mean peak intensity is obtained based on the number of peaks and the normalized data: In the formula, μ peak N represents the mean peak intensity. peak Indicates the number of peak values; I norm This represents the intensity normalization result; This represents the normalized intensity value of the peak value.

6. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 4, characterized in that, The method by which the statistical subunit obtains the statistical data includes: Calculate the global intensity mean based on the normalized data: In the formula, μ global L represents the global intensity mean; L represents the total number of normalized intensity data points; I represents the global intensity mean. norm This represents the intensity normalization result; Calculate the global intensity standard deviation based on the global intensity mean and the normalized data:

7. The alloy identification system based on deep recurrent neural network combined with fs-LA-SIBS according to claim 4, characterized in that, The feature extraction unit includes: an input layer, a recurrent layer, a feature fusion subunit, a fully connected layer, and an output layer; The input layer is used to process the spectral data to obtain splicing features; The recurrent layer is used to perform temporal analysis on the spliced ​​features to obtain temporal features; The feature fusion subunit is used to concatenate the spliced ​​features and the temporal features to obtain a joint feature vector; The fully connected layer is used to perform multi-level linear transformations on the joint feature vector to obtain dimensionality-reduced features; The output layer is used to generate the classification result based on the dimensionality-reduced features normalized by Softmax.

8. An alloy identification method based on a deep recurrent neural network combined with fs-LA-SIBS, wherein the method is applied to the system described in any one of claims 1-7, characterized in that, Includes the following steps: S1. Emit laser pulses to the alloy under test; S2. Detect the spectral data of the alloy to be tested after being subjected to laser pulses; S3. Process the spectral data to obtain alloy classification results.

Images