Neural network model training method, material detection method, system and equipment

Abstract

The invention is suitable for the technical field of spectrum detection, and provides a neural network model training method, a material detection method, a system and equipment, and the neural network model training method comprises the steps: obtaining a sample training set; wherein each piece of sample data in the sample training set comprises deviation between ideal spectral intensity and actual spectral intensity and corresponding sample fusion feature data; the sample fusion feature data of each sample is obtained based on a neural network model of an attention mechanism; the actual spectrum intensity of each sample is obtained according to the preprocessed plasma radiation spectrum data of the corresponding sample; and taking the sample fusion feature data as an input feature, taking the deviation between the ideal spectral intensity and the actual spectral intensity as an output feature, and training to obtain a first neural network model. According to the method, the analysis accuracy of the LIBS technology can be improved.

Description

Technical Field

[0001] This application belongs to the field of spectral detection technology, and in particular relates to a neural network model training method, a material detection method, a system and equipment. Background Technology

[0002] In recent years, laser-induced breakdown spectroscopy (LIBS) has attracted widespread attention from researchers due to its many advantages, such as minimally invasive detection, low detection limit, rapid detection, high sensitivity, and real-time analysis. It has been applied in many fields, including industry, agriculture, biology, archaeology, and aerospace.

[0003] In practical applications, LIBS test signals are unstable, primarily due to material properties, environmental factors, and system parameters. Physical properties of materials, such as roughness, hardness, moisture content, and particle size, as well as chemical properties like elemental valence states and abundance, can lead to matrix effects, causing fluctuations in the LIBS test signal. Furthermore, environmental factors such as temperature, humidity, air pressure, light intensity, and wind speed can all affect signal strength. Simultaneously, in complex environments, LIBS system parameters, such as laser energy, can fluctuate significantly due to scattering and particle absorption. These factors severely reduce the accuracy of LIBS analysis, limiting its application scope. Summary of the Invention

[0004] This application provides a model training method, a material testing method, a system, and an apparatus to solve or improve the technical problem in the related art where unstable test signals of LIBS lead to a decrease in the accuracy of LIBS technical analysis.

[0005] In a first aspect, embodiments of this application provide a neural network model training method, including:

[0006] Obtain a sample training set; wherein, each sample data in the sample training set includes the deviation between the ideal spectral intensity and the actual spectral intensity and the corresponding sample fusion feature data;

[0007] The sample fusion feature data for each sample is obtained based on a neural network model with an attention mechanism. The neural network model takes plasma radiation spectral feature data and plasma image feature data of the corresponding sample as input and the sample fusion feature data as output. The actual spectral intensity of each sample is obtained based on the preprocessed plasma radiation spectral data of the corresponding sample.

[0008] The sample fusion feature data is used as input features, and the deviation between the ideal spectral intensity and the actual spectral intensity is used as output features to train the first neural network model.

[0009] The above technical solution achieves multimodal deep fusion of plasma radiation spectral features and image features by introducing an attention mechanism, and constructs a neural network model with dynamic noise compensation capability. This model can accurately predict the deviation between the ideal spectral intensity and the actual spectral intensity, thereby achieving signal correction and effectively overcoming the signal instability problem caused by environmental interference, equipment fluctuations or sample inhomogeneity in traditional LIBS technology.

[0010] Furthermore, the neural network model training method also includes:

[0011] Acquire plasma event stream data for each sample;

[0012] Reconstruct the plasma image based on the plasma event stream data;

[0013] After applying Gaussian filtering to the plasma image, a Gaussian-filtered plasma image is obtained.

[0014] Based on the Gaussian filtered plasma image, plasma image feature data of each sample is extracted.

[0015] By acquiring plasma event stream data, we can capture the temporal information of plasma dynamic changes, providing a richer raw data foundation for subsequent feature extraction. Gaussian filtering of plasma images effectively smooths image noise, suppresses interference, and highlights key plasma morphology and brightness distribution characteristics, improving the accuracy of image feature extraction. Extracting feature data based on the processed Gaussian-filtered image yields cleaner and more representative plasma image features. These processing and extraction steps ensure that the image features integrated into the sample fusion feature data more accurately reflect the plasma state, synergistically enhancing the explanatory power of deviations with spectral features. This lays a high-quality feature foundation for the first neural network model to more accurately predict spectral intensity deviations, thereby contributing to improved accuracy of LIBS technology analysis.

[0016] Furthermore, the ideal spectral intensity of each sample is calculated using the plasma temperature of the corresponding sample as the correction formula, the electron density of the corresponding sample as the correction formula, and the radiation path length of the corresponding sample as the correction formula.

[0017] Furthermore, the plasma radiation spectral feature data is obtained by extracting plasma radiation spectral data using a sequence feature extraction neural network model; the plasma image feature data is obtained by extracting plasma image data using an image feature extraction neural network model.

[0018] And / or, the first neural network model is a regression neural network model.

[0019] Furthermore, the sequence feature extraction neural network model includes one of the following: a one-dimensional convolutional neural network, a long short-term memory network, a recurrent neural network, a bidirectional long short-term memory network, or a neural network model based on a self-attention mechanism; the image feature extraction neural network model includes one of the following: a two-dimensional convolutional neural network, a spiking neural network, a residual neural network, or a visual neural network model based on a self-attention mechanism.

[0020] Furthermore, the preprocessed plasma radiation spectrum data is obtained by performing background removal, baseline correction, and noise reduction post-processing on the original plasma radiation spectrum;

[0021] And / or, the attention mechanism includes one of self-attention, cross-attention, multi-head attention, sparse attention, channel / spatial attention, or linearized attention.

[0022] Secondly, embodiments of this application provide a material testing method, including:

[0023] Acquire detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested.

[0024] The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs.

[0025] The fused feature data is used as input features, and the first neural network model obtained by the neural network model training method is input to the model, and the predicted value of the deviation is output.

[0026] The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

[0027] Furthermore, the material detection method further includes acquiring plasma image feature data of the material to be detected; the acquisition of plasma image feature data of the material to be detected includes:

[0028] Acquire plasma event flow data of the material to be tested;

[0029] After applying Gaussian filtering to the plasma image, a Gaussian-filtered plasma image is obtained.

[0030] Based on the Gaussian filtered plasma image, plasma image feature data of the material to be detected is extracted.

[0031] Thirdly, embodiments of this application provide a laser-induced breakdown spectroscopy system, comprising:

[0032] Optical path system;

[0033] Laser, connected to the optical path system;

[0034] Delay controller, connected to the laser;

[0035] The spectrometer is connected to the computer, optical system, and delay controller, respectively.

[0036] Dynamic vision sensors are used to acquire plasma event flow data of materials or samples;

[0037] And a computer, connected to the spectrometer and the dynamic vision sensor respectively, for:

[0038] Acquire detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested.

[0039] The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs.

[0040] The fused feature data is used as input features, and the first neural network model obtained by the neural network model training method is input to the model, and the predicted value of the deviation is output.

[0041] The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

[0042] Fourthly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the model training method and the material detection method.

[0043] Fifthly, embodiments of this application provide a computer-readable storage medium, comprising: the computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the model training method and the material detection method.

[0044] In a sixth aspect, embodiments of this application provide a computer program product that, when run on a terminal device, causes the terminal device to execute the model training method and material testing method described in any of the first aspects above.

[0045] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0046] The beneficial effects of the embodiments in this application compared with the prior art are:

[0047] The neural network model training method, material detection method, system, and device of this application embodiment obtain sample fusion feature data for each sample based on an attention mechanism neural network model; obtain the actual spectral intensity of each sample based on the preprocessed plasma radiation spectrum data of the corresponding sample; use the sample fusion feature data as input features and the deviation between the ideal spectral intensity and the actual spectral intensity as output features to train a first neural network model. By using this first neural network model for material detection, the technical problem of unstable LIBS test signals leading to decreased accuracy of LIBS technical analysis in related technologies can be improved, thereby enhancing the accuracy of LIBS technical analysis. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram of a laser-induced breakdown spectroscopy system provided in an embodiment of this application.

[0050] Figure 2 This is a flowchart illustrating a neural network model training method according to an embodiment of this application.

[0051] Figure 3 This is a schematic diagram of the framework of a multimodal neural network model.

[0052] Figure 4 This is a plasma image reconstructed based on plasma event stream data according to an embodiment of this application; wherein, (a) is the original plasma grayscale image reconstructed based on plasma event stream data; and (b) is the Gaussian filtered plasma image obtained after Gaussian filtering the image in (a).

[0053] Figure 5 This is a flowchart illustrating the material testing method.

[0054] Figure 6These are schematic diagrams illustrating the analytical results of CI (493.202 nm) and Mn (403.076 nm) in different types of carbon steel. (a) shows the relationship between the original intensity obtained using preprocessed plasma radiation spectral data and the elemental content of carbon (C); (b) shows the relationship between the normalized intensity obtained after normalizing the preprocessed plasma radiation spectral data and the elemental content of carbon (C); (c) shows the relationship between the corrected intensity obtained after processing using the material detection method of this application and the elemental content of carbon (C); (d) shows the relationship between the original intensity obtained using preprocessed plasma radiation spectral data and the elemental content of manganese (Mn); (e) shows the relationship between the normalized intensity obtained after normalizing the preprocessed plasma radiation spectral data and the elemental content of manganese (Mn); and (f) shows the relationship between the corrected intensity obtained after processing using the material detection method of this application and the elemental content of manganese (Mn).

[0055] Figure 7 This is a schematic diagram of the structure of a neural network model training system.

[0056] Figure 8 This is a schematic diagram of the material testing system. Detailed Implementation

[0057] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0058] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0059] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0060] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0061] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0062] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0063] Before introducing the embodiments of this application, the relevant names in the embodiments of this application will be explained:

[0064] 1. Laser-induced breakdown spectroscopy (LIBS): The principle is to ablate the sample surface using a high-energy laser pulse, forming a high-temperature, high-electron-density plasma. During the cooling process, some of the plasma's energy is radiated in the form of a spectrum. By collecting the spectral signals with a spectrometer, the wavelength and intensity of the spectral lines can be obtained. The wavelength of the spectral lines can determine their elemental composition, while the intensity of the spectral lines can reflect the elemental content to some extent.

[0065] LIBS is widely used in industrial testing. By analyzing the changes in spectral characteristics caused by differences in the types and contents of elements in different materials, LIBS technology can distinguish between them. However, due to the complexity and variability of LIBS spectral signals, it is affected not only by the types and contents of elements in the sample but also by various external conditions such as laser energy, plasma state, sample surface condition, and environmental factors. These interfering factors lead to lower accuracy in material classification using LIBS technology.

[0066] 2. Plasma Event Stream Data: Plasma event stream data refers to the continuously changing data stream of plasma acquired in real time during the ablation of a sample surface by a high-energy laser pulse, forming a high-temperature, high-electron-density plasma. It can be acquired using a high-speed camera or a dynamic vision sensor.

[0067] 3. Plasma radiation spectral data: The sample surface is ablated by high-energy laser pulses to form high-temperature, high-electron-density plasma. During the cooling process, some of the plasma energy is radiated in the form of a spectrum, which is collected by a spectrometer to form the spectral signal.

[0068] 4. Plasma Image: Generated based on plasma event stream data reconstruction. For example, a dynamic vision sensor can be used to acquire plasma event stream data. It is understood that a dynamic vision sensor differs from a traditional camera or high-speed camera; its output is not a frame image, but rather contains pixel coordinates (x, y), a timestamp t, and an event polarity p. When the light intensity changes beyond a threshold, the pixel responds: an ON event (p=1) is generated when the brightness increases, and an OFF event (p=0) is generated when the brightness decreases. The dynamic vision sensor outputs these events and data as an event stream.

[0069] 5. Plasma characteristic data: refers to a set of parameters and information that can reflect various properties and states of plasma. Plasma characteristic data can be the number of events obtained based on plasma event stream data, the plasma image reconstructed based on plasma event stream data, and the plasma image area, shape features, and texture features obtained based on the plasma image.

[0070] 6. Event Count: Generated based on plasma event stream data. Plasma event stream data is obtained by accumulating event data over a certain period of time from the plasma generated by the sample under laser ablation using a dynamic vision sensor. When the sample generates plasma under laser ablation, each time the brightness change of the plasma exceeds the threshold of the dynamic vision sensor, it is recorded as an event, and the pixel data of each event is recorded, including the pixel number, timestamp, and event polarity. The accumulated events over a certain period of time generate plasma event stream data, and the number of events generated within that period is counted to generate the event count.

[0071] 7.1D-CNN (One-Dimensional Convolutional Neural Network): This is a combination of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM), specifically designed for processing one-dimensional sequence data. 1D-CNN is responsible for extracting local features from the sequence, while LSTM is responsible for capturing the long-term dependencies between these features.

[0072] 8. RNN (Recurrent Neural Network): RNNs are a type of neural network specifically designed for processing sequential data. They pass information through hidden states and have memory capabilities. Working principle: Hidden state recursion: The hidden state at each time step is determined by the current input and the hidden state of the previous time step. Output calculation: The hidden states generate the output (such as classification probability or regression value) through fully connected layers. Parameter sharing: All time steps share the same set of weight parameters. Core advantages: Strong sequence modeling capability: Naturally suitable for handling variable-length sequences. Simple and efficient: Simple structure, easy to implement and train.

[0073] 9. LSTM (Long Short-Term Memory Network): LSTM is an improved version of RNN. It solves the problem of long sequence dependencies through a gating mechanism, enabling it to remember and forget information. Working principle: Gating mechanism: Input gate: Controls the degree to which new information flows into the cell state. Forget gate: Determines whether to retain or discard historical information in the cell state. Output gate: Controls the influence of the cell state on the current output. Cell state: Acts as a "highway" for information transmission, storing key information long-term. Time step recursion: The output of each time step depends on the current input and the hidden state of the previous time step.

[0074] 10. Bi-LSTM (Bidirectional Long Short-Term Memory Network): Bi-LSTM is a variant of LSTM that enhances the model's contextual modeling ability by simultaneously capturing the forward and backward dependencies of a sequence. How it works: Forward LSTM layer: Processes the input sequentially, capturing historical information. Backward LSTM layer: Processes the input in reverse order, capturing future contextual information. Output fusion: The hidden states from both directions are concatenated at each time step to form the final bidirectional representation.

[0075] 11. Transformers (self-attention based neural network models): These are deep learning models based on a self-attention mechanism, originally designed for natural language processing tasks. Unlike RNNs and LSTMs, Transformers efficiently process sequential data through parallel computation and global attention mechanisms.

[0076] 12. 2D-CNN (Two-Dimensional Convolutional Neural Network): 2D-CNN is a deep learning model specifically designed for processing two-dimensional data (such as images). It extracts spatial features from the input data through convolution operations. Core components: Convolutional layers: Multiple convolutional kernels (filters) slide across the input image, calculating the dot product between the kernel and a local region of the image to generate a feature map. Each convolutional kernel captures a specific pattern or structure in the image. Activation functions: Such as ReLU, which increase the non-linear expressive power of the network, enabling it to fit more complex functional relationships. Pooling layers: Such as max pooling or average pooling, which downsample the output of the convolutional layers, reducing data dimensionality while preserving important features and improving the robustness of the model. Fully connected layers: Mapping the extracted features to the output category, acting as a classification or regression layer. Output layer: Depending on the task, the output layer may be a softmax layer (for classification tasks) or a linear regression layer (for regression tasks).

[0077] 13. SNN (Spiking Neural Network): SNN is a computational model inspired by biological neural systems. It simulates the spiking behavior of neurons in the brain, transmitting information through discrete spiking signals and combining this with a time dimension for computation. Core components: Neuron model: such as the Leaky Integrate-and-Fire (LIF) model, which simulates the membrane potential changes and spiking process of neurons. Synapse model: simulates the connections and signal transmission processes between neurons. Network structure: composed of multiple neurons and synapses, forming a complex network topology.

[0078] 14. ResNet (Residual Neural Network): ResNet is a deep convolutional neural network architecture that addresses the degradation problem in deep network training by introducing "residual learning." It creates residual blocks by directly linking the activation values ​​of layers to subsequent layers, bypassing certain intermediate layers. These residual blocks are stacked to build the ResNet. Core components: Residual Block: Contains multiple convolutional layers and connects the input directly to the output through shortcut connections, forming a residual map. Bottleneck Structure: In deeper ResNet models, a Bottleneck structure is used to reduce model complexity and computational requirements. It contains three convolutional layers: a 1x1 convolutional layer for dimensionality reduction, a 3x3 convolutional layer for handling complex features, and another 1x1 convolutional layer for dimensionality recovery. Global Average Pooling Layer: After all the residual blocks, a global average pooling layer replaces the fully connected layers, reducing the number of model parameters and thus mitigating overfitting.

[0079] 14. Vision Transformer (a visual neural network model based on self-attention): This model applies the Transformer architecture to computer vision tasks. It treats images as a sequence of smaller patches, thus capturing global relationships and long-range dependencies in visual data. Core components: Image patch creation: The input image is segmented into a grid of fixed-size patches, and each patch is flattened into a one-dimensional vector. Linear embedding of patches: Each patch vector is passed through a linear layer, projecting it into a higher-dimensional space to create an embedding sequence. Adding positional encoding: Positional encoding is added to each image patch embedding, injecting information about the spatial location of each image patch. Transformer encoder: Composed of multiple layers of multi-head self-attention and feedforward neural networks, it learns the relationships between different image patches through a self-attention mechanism.

[0080] Laser-induced breakdown spectroscopy (LIBS) has attracted widespread attention from researchers due to its many advantages, such as minimally invasive detection, low detection limit, rapid detection, high sensitivity, and real-time analysis. It has been applied in many fields, including industry, agriculture, biology, archaeology, and aerospace.

[0081] In practical applications, LIBS test signals are unstable, primarily due to material properties, environmental factors, and system parameters. Physical properties of materials, such as roughness, hardness, moisture content, and particle size, as well as chemical properties like elemental valence states and abundance, can lead to matrix effects, causing fluctuations in the LIBS test signal. Furthermore, environmental factors such as temperature, humidity, air pressure, light intensity, and wind speed can all affect signal strength. Simultaneously, in complex environments, LIBS system parameters, such as laser energy, can fluctuate significantly due to scattering and particle absorption. These factors severely reduce the accuracy of LIBS analysis, limiting its application scope.

[0082] To address or improve the technical problem of decreased accuracy in LIBS analysis due to unstable test signals in related technologies, and to enhance the accuracy of LIBS analysis, the present application proposes the following inventive concept.

[0083] First, in this application embodiment, a dynamic visual sensor is added to the laser-induced breakdown spectroscopy system of the related technology to form a new laser-induced breakdown spectroscopy system.

[0084] The laser-induced breakdown spectroscopy system of related technologies can ablate the sample surface with high-energy laser pulses to form high-temperature, high-electron-density plasma. During the cooling process, some of the energy of the plasma is radiated in the form of a spectrum. The spectral signal is collected by a spectrometer to generate plasma radiation spectrum data of the sample or the material to be tested.

[0085] To eliminate interference from certain factors, the plasma radiation spectrum data of the sample or the material to be tested is the preprocessed plasma radiation spectrum data of the corresponding sample or the material to be tested obtained after preprocessing such as background removal, baseline correction and / or noise reduction, and the corresponding plasma radiation spectrum feature data is extracted from the preprocessed plasma radiation spectrum data.

[0086] High-energy laser pulses ablate the sample surface, forming a high-temperature, high-electron-density plasma. During the cooling process, plasma radiation spectrum data and plasma event flow data can be collected. For example, a dynamic vision sensor can simultaneously acquire plasma event flow data of the sample while collecting plasma radiation spectrum data, and plasma feature data can be extracted based on the plasma event flow data.

[0087] By combining the plasma radiation spectral characteristic data and plasma characteristic data of the sample, that is, fusing the plasma radiation spectral characteristic data and plasma characteristic data into fused characteristic data, and training a neural network model, a first neural network model that can improve the accuracy of LIBS technology analysis can be obtained.

[0088] It is understood that the technical solutions of this application embodiment can be applied to various laser-induced breakdown spectroscopy systems. For example, the technical solutions of this application embodiment can be applied to systems that incorporate dynamic visual sensors into various traditional laser-induced breakdown spectroscopy systems. The dynamic visual sensor is used to capture plasma event flow data during the period from when the sample is ablated to generate plasma until the plasma cools down. This plasma event flow data can be used to reconstruct a plasma image. It is understood that a high-speed camera can be used instead of a dynamic visual sensor; however, a dynamic visual sensor generates less data than a high-speed camera, which is more advantageous for extracting plasma feature data from the plasma event flow data to reconstruct a plasma image. Therefore, in this application embodiment, a dynamic visual sensor is used to capture plasma event flow data during the period from when the sample is ablated to generate plasma until the plasma cools down.

[0089] For example, Figure 1 This is a laser-induced breakdown spectroscopy system applicable to embodiments of this application. Figure 1 The replacement of the dynamic vision sensor with a high-speed camera is also applicable to the technical solutions in the embodiments of this application.

[0090] See Figure 1 As shown, the laser-induced breakdown spectroscopy system includes a laser 101, a delay controller 102, a spectrometer 103, a computer 104, a dynamic vision sensor 105, an optical path system 106, a sample, and a sample excitation stage 107.

[0091] Laser 101 is connected to optical path system 106. Laser 101, delay controller 102, spectrometer 103, computer 104, and dynamic vision sensor 105 are connected in sequence. The optical path of the optical path system points towards the sample and sample excitation stage 107. Spectrometer 103 receives the plasma spectrum generated by the sample and the sample excitation stage 107. Dynamic vision sensor 105 is used to capture event data generated by the brightness change of the plasma during sample ablation. By selecting appropriate laser energy, light collection angle, and spectrometer delay time, high spectral signals with high signal-to-noise ratio and signal-to-background ratio can be obtained. In addition, a dynamic vision sensor is added to capture the plasma signal.

[0092] To improve the accuracy of LIBS technology analysis, this application provides a neural network model training method. This method can run on the computer of the aforementioned laser-induced breakdown spectroscopy system, or on the aforementioned laser-induced breakdown spectroscopy system after replacing the computer with other devices. These other devices can be mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, ultra-mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), and other terminal devices. This application does not limit the specific type of terminal device.

[0093] For example, the terminal device may be a station (STAION, ST) in a WLAN, a cellular phone, a cordless phone, a Session Initiation Protocol (SIP) phone, a Wireless Local Loop (WLL) station, a Personal Digital Assistant (PDA) device, a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, an in-vehicle device, a vehicle networking terminal, a computer, a laptop computer, a handheld communication device, a handheld computing device, a satellite wireless device, a wireless modem card, a set-top box (STB), customer premises equipment (CPE), and / or other devices used for communication over a wireless system, as well as next-generation communication systems, such as mobile terminals in 5G networks or mobile terminals in future evolved Public Land Mobile Network (PLMN) networks.

[0094] As an example and not a limitation, when the terminal device is a wearable device, the term "wearable device" can also refer to any device that utilizes wearable technology to intelligently design and develop everyday wearables, such as glasses, gloves, watches, clothing, and shoes. Wearable devices are portable devices worn directly on the body or integrated into a user's clothing or accessories. Wearable devices are not merely hardware devices; they achieve powerful functions through software support, data interaction, and cloud interaction. Broadly defined, wearable smart devices include those with comprehensive functions, large sizes, and the ability to perform complete or partial functions without relying on a smartphone, such as smartwatches or smart glasses, as well as those focused on a specific application function that require interaction with other devices such as smartphones, such as various smart bracelets and smart jewelry for vital sign monitoring.

[0095] Figure 2 This is a schematic flowchart illustrating a neural network model training method according to an embodiment of this application. (See attached diagram.) Figure 2 As shown, in one embodiment of this application, the neural network model training method can be applied to the aforementioned laser-induced breakdown spectroscopy system. (See reference...) Figure 2 As shown, it includes the following steps:

[0096] A1. Obtain a sample training set; wherein, each sample data in the sample training set includes the deviation between the ideal spectral intensity and the actual spectral intensity and the corresponding sample fusion feature data;

[0097] The sample fusion feature data for each sample is obtained based on a neural network model with an attention mechanism. The neural network model takes plasma radiation spectrum feature data and plasma image feature data of the corresponding sample as input and the sample fusion feature data as output.

[0098] It is understandable that plasma radiation spectral characteristic data is key information extracted from the radiation spectrum of plasma (such as peak wavelength, peak intensity, full width at half maximum, radiation intensity at a specific wavelength, etc.), reflecting information such as the chemical composition and excited state distribution of the plasma. Plasma image characteristic data is key information extracted from images of plasma (such as plasma morphology images taken by high-speed cameras) (such as plasma shape, boundary contour, brightness distribution, turbulent regions, etc.), reflecting information such as the spatial structure and stability of the plasma.

[0099] By integrating the plasma radiation spectral feature data and plasma image feature data of the corresponding samples, a neural network with an attention mechanism can be used.

[0100] Single-modal features (spectral or image-only) often have information limitations (e.g., spectra cannot reflect spatial distribution, and images cannot reflect detailed chemical information). The core logic of fusing the two through an "attention-based neural network" is that the attention mechanism allows the model to automatically "focus on key features": for example, when turbulence appears in a region of a plasma image, the model will focus on the spectral features corresponding to that region (because turbulence may cause spectral distortion); conversely, when the intensity of a certain wavelength in the spectrum is abnormal, the model will prioritize associating it with the morphological features of the corresponding location in the image. The final output "sample fusion feature data" is a "comprehensive feature vector" that integrates the chemical information of the spectrum and the spatial information of the image. It is more comprehensive than single-modal features and can provide a more reliable basis for subsequent bias prediction.

[0101] It is understandable that the ideal spectral intensity can be calculated using the spectral intensity formula.

[0102] It is understandable that plasma radiation spectral feature data can be obtained by using a sequence feature extraction neural network model to extract preprocessed plasma radiation spectral data; preprocessed plasma radiation spectral data is obtained by preprocessing the original plasma radiation spectral data, such as background removal, baseline correction and / or noise reduction; the original plasma radiation spectral data can be plasma radiation spectral data directly obtained by a spectrometer.

[0103] For example, the sequence feature extraction neural network model can be one of 1D-CNN, LSTM, RNN, BiLSTM, or Transformers. Optionally, the preprocessed plasma radiation spectrum data can be extracted using 1D-CNN.

[0104] It is understood that the plasma image feature data is obtained by extracting plasma image data using an image feature extraction neural network model; for example, the image feature extraction neural network model includes one of 2D-CNN, SNN, ResNet, or Vision Transformer. Optionally, the plasma image feature data can be extracted using 2D-CNN.

[0105] It is understood that the actual spectral intensity of each sample can be obtained from the preprocessed plasma radiation spectral data of the corresponding sample; for example, the preprocessed plasma radiation spectral data is obtained by preprocessing the original plasma radiation spectral data, such as background removal, baseline correction, and / or noise reduction. The actual spectral intensity can be obtained from the processed plasma radiation spectral data.

[0106] Understandably, the attention mechanism in a neural network model employing an attention mechanism can include one of the following: self-attention, cross-attention, multi-head attention, sparse attention, channel / spatial attention, or linearized attention. Optionally, the neural network model employing an attention mechanism can be a neural network model employing a multi-head attention mechanism.

[0107] For example, the first neural network model is a regression neural network model.

[0108] A2. Using the sample fusion feature data as input features and the deviation between the ideal spectral intensity and the actual spectral intensity as output features, a first neural network model is trained.

[0109] For example, the deviation between the ideal spectral intensity and the actual spectral intensity can be expressed as I. error I error The calculation formula is shown in (1):

[0110] I error =I ideal -I real (1)

[0111] Among them, I ideal Indicates ideal spectral intensity; I real This represents the actual spectral intensity.

[0112] Alternatively, the neural network model of this method can also be a multimodal neural network model. Figure 3 This is a schematic diagram of the framework of a multimodal neural network model. It may include:

[0113] Spectral data feature extraction: Plasma radiation spectral data 201 passes through the first layer (one-dimensional convolutional layer + ReLU activation layer) 203, the first max pooling layer 205, and the second layer (global average pooling layer + fully connected layer) 206 to extract spectral data feature data.

[0114] Specifically, for a one-dimensional spectral sequence of length 10542, a three-layer 1D-CNN is used, with each layer containing convolution, batch normalization, ReLU activation, and max pooling operations. The convolutional kernel widths are 64, 32, and 16 respectively, and the number of filters increases from 64 to 256. With each downsampling layer, the sequence length continuously decreases. Finally, global average pooling and fully connected layers are used to reduce the dimensionality of the feature vector to 128 dimensions.

[0115] Plasma image feature extraction: Plasma subbody image 202 passes through the third layer (two-dimensional convolutional layer + ReLU activation layer) 204, the second max pooling layer 212, and the fourth layer (global average pooling layer + fully connected layer) 213 to extract plasma image feature data.

[0116] Specifically, for a 100×100 two-dimensional grayscale image, the input is a three-layer 2D-CNN structure, each layer consisting of convolution, batch normalization, ReLU activation, and 2×2 max pooling. The convolution kernel size is 3×3, and the number of filters increases from 64 to 256. Subsequently, global average pooling is used to obtain a 128-dimensional image feature vector.

[0117] Spectral data features and plasma image features are combined to form CNN-extracted feature data 207. This extracted feature data is then fused through a network layer 208 with an attention mechanism to obtain fused feature data 209. After processing by the fifth layer (fully connected layer + regression layer) 210, the fused feature data 209 yields the prediction result 211. Finally, a multimodal neural network model is obtained, using the sample fused feature data as input features and the deviation between the ideal spectral intensity and the actual spectral intensity as output features.

[0118] Attention Mechanism: To achieve effective information fusion of spectral and image features, a multi-head attention mechanism is introduced to weightedly integrate the two types of features. First, the 128-dimensional feature vectors extracted by 1D-CNN and 2D-CNN are horizontally concatenated. The multi-head attention mechanism can model the correlations between features in parallel across different subspaces, learning the intra- and cross-modal correlations of spectral and image features through multiple attention heads, and assigning higher weights to useful information. After fusion by the multi-head attention mechanism, a comprehensive feature vector is output, improving the model's ability to express key features and its noise suppression effect.

[0119] Regression analysis network: Incorporating ideal spectrum I ideal Compared with actual spectrum Ireal Deviation I error As the target variable (output feature), the feature fused by the multi-head attention mechanism is used as the input feature. A fully connected network and regression layer are used for regression analysis, and finally the first neural network model is trained.

[0120] Therefore, in this embodiment, the sample fusion feature data of each sample is obtained based on the neural network model of the attention mechanism; the actual spectral intensity of each sample is obtained according to the preprocessed plasma radiation spectrum data of the corresponding sample; the sample fusion feature data is used as input features, and the deviation between the ideal spectral intensity and the actual spectral intensity is used as output features to train a first neural network model. By using the first neural network model to perform material detection, the technical problem of unstable LIBS test signals leading to a decrease in the accuracy of LIBS technology analysis in related technologies can be improved, thereby increasing the accuracy of LIBS technology analysis.

[0121] Furthermore, the neural network model training method also includes:

[0122] A01. Obtain plasma event stream data for each sample;

[0123] A02. Reconstruct the plasma image based on the plasma event stream data;

[0124] A03. After performing Gaussian filtering on the plasma image, a Gaussian-filtered plasma image is obtained;

[0125] A04. Extract plasma image feature data for each sample based on the Gaussian filtered plasma image.

[0126] Optionally, A02. Reconstructing a plasma image based on the plasma event stream data; including:

[0127] A021. Based on the plasma event stream data, a plasma image is reconstructed using deep neural networks, spiking neural networks, cluster analysis, or event accumulation.

[0128] The plasma image is either a first plasma image generated based on the plasma event stream data or a Gaussian-filtered plasma image obtained by Gaussian filtering the first plasma image. The first plasma image can be the original plasma image directly reconstructed from the plasma event stream data.

[0129] Exemplary reference Figure 4 As shown, the event frame image of plasma can be reconstructed by event accumulation. Plasma event stream data is obtained by accumulating event data within a certain time range. The plasma image can then be obtained from the plasma event stream data. Figure 4 As shown in the figure, the cumulative time is 200ms, but the cumulative time can also be selected based on the actual plasma event data. Figure 4 Image (a) is a plasma image generated based on the plasma event stream data. Figure 4 Image (b) is the Gaussian-filtered plasma image obtained after Gaussian filtering in image (a). Subsequent steps can employ... Figure 4 The area enclosed by the plasma shape contour, the near-circularity of the plasma image, or other plasma feature data such as the shape features and texture features of the plasma image are extracted from the image in (a) or (b).

[0130] Understandably, identifying plasma boundaries in plasma images typically involves using image processing algorithms, such as the Canny operator and the Sobel operator, to detect edges in the image. Key plasma parameters, such as temperature and density, are calculated based on spectral or image data. Image generation then requires reconstructing the plasma image from the plasma event stream data using image processing software or algorithms. These images can be two-dimensional grayscale or color images, or three-dimensional stereo images.

[0131] Furthermore, the ideal spectral intensity of each sample is calculated using the plasma temperature of the corresponding sample as the correction formula, the electron density of the corresponding sample as the correction formula, and the radiation path length of the corresponding sample as the correction formula.

[0132] Specifically, the modified formula (2) is as follows:

[0133]

[0134] Among them, I ideal The value of F0 represents the intensity of atomic or ion spectral lines, and F0 represents the instrument and environmental gain factor under standard conditions, the value of which is chosen to make I... ideal with I real (This can be obtained from preprocessed plasma radiation spectrum data) The closest approximation is used, where C0 represents elemental content, l0 represents radiation path length, and T0 represents plasma temperature. Us(T0) is the partition function, E... i A represents the excitation energy of energy level i. ij Spontaneous emission coefficient, g i k represents the degeneracy of energy level i. B is the Boltzmann constant.

[0135] The average plasma temperature of the corresponding sample is used as the plasma temperature T0 in the correction formula, and the electron density of the corresponding sample is used as the electron density n in the correction formula. eThe radiation path length of the corresponding sample is used as the radiation path length l0 of the correction formula.

[0136] R is the number density ratio of atoms to ions, which can be determined by the Saha equation based on the electron density n. e The plasma temperature T0 was calculated. Details are as follows:

[0137] According to the Saha equation, R can be expressed as:

[0138]

[0139] Where, n I and n II Us represents the number of atoms and the number of ions produced by a single ionization of an element, respectively. I (T0) and Us II (T0) represent the partition functions of atoms and ions, respectively, m e E represents the electron mass. ion The ionization energy of an atom, ΔE ion This is a correction term for atomic ionization energy.

[0140] Alternatively, to simplify calculations, the plasma temperature T0 can be determined using the Boltzmann diagram method, and the electron density n e The radiation path length can be obtained through the Stark broadening method, while the average value of the plasma image in the four directions of horizontal, vertical and diagonal is obtained.

[0141] Specifically, the commonly used method for determining plasma temperature is the Boltzmann plot method. This method uses spectral line data of the same element in the same ionization state for calculation, and the specific calculation formula is as follows:

[0142]

[0143] Among them, I ij For the characteristic spectral line intensity, A ij G represents the transition probability. j E represents the energy level degeneracy of the upper energy level j. j Let kB represent the energy of the upper energy level j, kB represent the Boltzmann constant, T be the plasma temperature, F be a constant related to the system parameters, and N be the plasma temperature. s U represents the total particle number density. s (T) represents the partition function. In solving for the plasma temperature, the above equation can be transformed to obtain...

[0144] y = kx + b (5)

[0145]

[0146] x = E j (7)

[0147]

[0148] By selecting multiple characteristic spectral lines, a set of data points (x, y) can be obtained. The slope of the curve can be obtained through linear fitting, and then the plasma temperature T can be calculated.

[0149] In laser-induced plasma, the main mechanism for atomic emission line broadening is Stark broadening caused by the field effect of surrounding charged particles on the luminescent atoms, with electron number density n e Calculations can be made using Stark widening:

[0150] Δλ 1 / 2 =2ω(n) e / 10 16 (10)

[0151] In the formula, ω is the electron collision parameter, and Δλ 1 / 2 The full width at half maximum (FWHM) of the spectral line.

[0152] Based on the neural network model trained above, embodiments of this application provide a material detection method, see below. Figure 5 As shown, it includes:

[0153] S1. Obtain detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested;

[0154] The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs.

[0155] S2. Using the fused feature data as input features, inputting the first neural network model obtained by the neural network model training method, and outputting the predicted value of the deviation;

[0156] S3. The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

[0157] Optionally, the material detection method further includes S0. acquiring plasma image feature data of the material to be detected; S0. acquiring plasma image feature data of the material to be detected includes:

[0158] S01. Acquire plasma event flow data of the material to be tested;

[0159] S02. After performing Gaussian filtering on the plasma image, a Gaussian filtered plasma image is obtained;

[0160] S03. Extract plasma image feature data of the material to be detected based on the Gaussian filtered plasma image.

[0161] Optionally, the spectral intensity I of the material corr It can be represented as:

[0162]

[0163] in, This indicates the model's prediction results.

[0164] For example, LIBS testing was performed on different grades of carbon steel to verify the accuracy of the material testing method. Different grades of carbon steel were selected for LIBS testing, and the sample content information is shown in Table 1.

[0165] Table 1. Elemental content information of 9 carbon steel samples

[0166]

[0167] Analysis was performed using CI (493.202 nm) and MnI (403.076 nm), and the results are as follows: Figure 6 As shown. Figure 6 As shown, when calibrating using the original strength, the fitting determination coefficient R of element C is... 2 The mean squared error (MSD) was 0.232, the mean squared deviation (MSD) was 0.129, and the mean relative standard deviation (MRSD) was 0.317. Figure 6 a) R of Mn element 2 The value was 0.425, the MSTD was 0.138, and the MRSD was 0.351. Figure 6 d). After normalization, the R of element C 2 It decreased to 0.054, MSD was 0.064, and MRSD was 0.118. Figure 6 b), while the R of the Mn element 2 The value was 0.388, the MSD was 0.073, and the MRSD was 0.117. Figure 6 e) The fitting effect is still poor. After testing using the detection method described above for the materials, the R-value of element C is... 2 The value increased significantly to 0.992, the MSD decreased to 0.042, and the MRSD was 0.087. Figure 6 c) R of Mn element 2 The value was 0.985, the MSD was 0.052, and the MRSD was 0.073. Figure 6 f).

[0168] Therefore, the detection method described above significantly improves the linear relationship and analytical accuracy between the corrected signal and element content, far superior to traditional or simple normalization methods, thereby improving the accuracy of LIBS technology analysis.

[0169] This application also provides a neural network model training system 300, see below. Figure 7 As shown, it includes:

[0170] The first acquisition unit 301 is used to acquire a sample training set; wherein, each sample data in the sample training set includes the deviation between the ideal spectral intensity and the actual spectral intensity and the corresponding sample fusion feature data;

[0171] The sample fusion feature data for each sample is obtained based on a neural network model with an attention mechanism. The neural network model takes plasma radiation spectral feature data and plasma image feature data of the corresponding sample as input and the sample fusion feature data as output. The actual spectral intensity of each sample is obtained based on the preprocessed plasma radiation spectral data of the corresponding sample.

[0172] The second acquisition unit 302 is used to acquire plasma image feature data of the corresponding sample; the second acquisition unit 302 includes:

[0173] The third acquisition unit 3021 acquires plasma event stream data for each sample;

[0174] Processing unit 3022 is used to perform Gaussian filtering on the plasma image to obtain a Gaussian filtered plasma image;

[0175] Extraction unit 3033 is used to extract plasma image feature data of each sample based on the Gaussian filtered plasma image.

[0176] And a training unit 303, used to take the sample fusion feature data as input features and the deviation between the ideal spectral intensity and the actual spectral intensity as output features to train a first neural network model.

[0177] This application also provides a material testing system 400, see below. Figure 8 As shown, it includes:

[0178] The material data acquisition unit 401 acquires the detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion feature data of the material to be tested; wherein, the actual spectral intensity of the material to be tested is obtained based on the preprocessed plasma radiation spectrum data of the material to be tested; the fusion feature data is obtained based on a neural network model with an attention mechanism, the neural network model taking the plasma radiation spectrum feature data and the plasma image feature data of the corresponding sample as input, and the sample fusion feature data as output;

[0179] The generation unit 402 is used to take the fused feature data as input features, input the first neural network model obtained by the neural network model training method, and output the predicted value of the deviation.

[0180] And a calculation unit 403, used to calculate the spectral intensity of the material based on the predicted value and the actual spectral intensity.

[0181] This application also provides a laser-induced breakdown spectroscopy system, see below. Figure 1 As shown, it includes:

[0182] Optical path system 106;

[0183] Laser 101 is connected to the optical path system;

[0184] Delay controller 102 is connected to the laser;

[0185] The spectrometer 103 is connected to the computer, the optical path system, and the delay controller, respectively.

[0186] Dynamic vision sensor 105 is used to acquire plasma event flow data of materials or samples;

[0187] And computer 104, connected to the spectrometer and dynamic vision sensor respectively, for:

[0188] Acquire detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested.

[0189] The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs.

[0190] The fused feature data is used as input features, and the first neural network model obtained by the neural network model training method is input to the model, and the predicted value of the deviation is output.

[0191] The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

[0192] This application also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the model training method and the material detection method.

[0193] This application also provides a computer-readable storage medium, including: the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the model training method and the material detection method.

[0194] This application also provides a computer program product that, when run on a terminal device, causes the terminal device to execute the model training method and material testing method described in any of the first aspects above.

[0195] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0196] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0197] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0198] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0199] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0200] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the technical solution in this embodiment, depending on actual needs.

[0201] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for training a neural network model, characterized in that, include: A sample training set is obtained; wherein, each sample data in the sample training set includes the deviation between the ideal spectral intensity and the actual spectral intensity and the corresponding sample fusion feature data; the sample fusion feature data of each sample is obtained based on a neural network model with an attention mechanism, wherein the neural network model takes plasma radiation spectral feature data and plasma image feature data of the corresponding sample as input and takes sample fusion feature data as output; the actual spectral intensity of each sample is obtained based on the preprocessed plasma radiation spectral data of the corresponding sample; The sample fusion feature data is used as input features, and the deviation between the ideal spectral intensity and the actual spectral intensity is used as output features to train the first neural network model.

2. The neural network model training method according to claim 1, characterized in that, Also includes: Acquire plasma event stream data for each sample; After applying Gaussian filtering to the plasma image, a Gaussian-filtered plasma image is obtained. Based on the Gaussian filtered plasma image, plasma image feature data of each sample is extracted.

3. The neural network model training method according to claim 1, characterized in that, The ideal spectral intensity of each sample is calculated using the plasma temperature of the corresponding sample as the correction formula, the electron density of the corresponding sample as the correction formula, and the radiation path length of the corresponding sample as the correction formula.

4. The neural network model training method according to any one of claims 1-3, characterized in that, The plasma radiation spectral feature data is obtained by extracting plasma radiation spectral data using a sequence feature extraction neural network model; the plasma image feature data is extracted from plasma image data using an image feature extraction neural network model. And / or, the first neural network model is a regression neural network model.

5. The neural network model training method according to claim 4, characterized in that, The sequence feature extraction neural network model includes one of the following: a one-dimensional convolutional neural network, a long short-term memory network, a recurrent neural network, a bidirectional long short-term memory network, or a neural network model based on a self-attention mechanism; the image feature extraction neural network model includes one of the following: a two-dimensional convolutional neural network, a spiking neural network, a residual neural network, or a visual neural network model based on a self-attention mechanism.

6. The neural network model training method according to claim 4, characterized in that, The preprocessed plasma radiation spectrum data is obtained by performing background removal, baseline correction, and noise reduction on the original plasma radiation spectrum. And / or, the attention mechanism includes one of self-attention, cross-attention, multi-head attention, sparse attention, channel / spatial attention, or linearized attention.

7. A material testing method, characterized in that, include: Acquire detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested. The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs. The fused feature data is used as input features, and the first neural network model obtained by the neural network model training method according to any one of claims 1-5 is used as input. The predicted value of the deviation is then output. The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

8. The material testing method according to claim 7, characterized in that, It also includes acquiring plasma image feature data of the material to be tested; the acquisition of plasma image feature data of the material to be tested includes: Acquire plasma event flow data of the material to be tested; Reconstruct the plasma image based on the plasma event stream data; After applying Gaussian filtering to the plasma image, a Gaussian-filtered plasma image is obtained. Based on the Gaussian filtered plasma image, plasma image feature data of the material to be detected is extracted.

9. A laser-induced breakdown spectroscopy system, characterized in that, include: Optical path system; Laser, connected to the optical path system; Delay controller, connected to the laser; The spectrometer is connected to the computer, optical system, and delay controller, respectively. Dynamic vision sensors are used to acquire plasma event flow data of materials or samples; And a computer, connected to the spectrometer and the dynamic vision sensor respectively, for: Acquire detection data of the material to be tested; the detection data includes the actual spectral intensity of the material to be tested and the fusion characteristic data of the material to be tested. The actual spectral intensity of the material to be tested is obtained from the preprocessed plasma radiation spectrum data of the material to be tested; the fused feature data is obtained from a neural network model based on an attention mechanism, wherein the neural network model takes the plasma radiation spectrum feature data and the plasma image feature data of the corresponding samples as inputs and the sample fused feature data as outputs. The fused feature data is used as input features, and the first neural network model obtained by the neural network model training method according to any one of claims 1-6 is used as input, and the predicted value of the deviation is output. The spectral intensity of the material is calculated based on the predicted value and the actual spectral intensity.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6 or any one of claims 7 to 8.

Images