Libs spectral time sequence feature mining prediction method and system based on space-time mapping mechanism and storage medium

Abstract

This invention discloses a method, system, and storage medium for mining and predicting the temporal features of Libs spectra based on a spatiotemporal mapping mechanism. The method includes: acquiring spectral data collected by a Libs device to form a spectral matrix, where rows correspond to laser spotting times, columns correspond to spectral segments, and cell values ​​are spectral intensities; performing three-dimensional spatiotemporal mapping on the spectral matrix to generate a single-frame color image, which simultaneously retains temporal, spatial, and spectral intensity dimensions; obtaining multiple frames of color images by sampling multiple times and performing three-dimensional spatiotemporal mapping on each frame, arranging them in chronological order of sampling time to construct a temporal image sequence; inputting the temporal image sequence into a prediction model to output the prediction results of the target parameters. This invention transforms Libs spectral data into a visual image sequence through three-dimensional spatiotemporal mapping, and combines this with a prediction model to mine temporal dependencies, achieving accurate temporal prediction of material composition content and lowering the threshold for spectral data analysis.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis and time series prediction technology, specifically to a method, system, and storage medium for mining and predicting Libs spectral time series features based on spatiotemporal mapping mechanism. Background Technology

[0002] Laser-induced breakdown spectroscopy (Libs) is an atomic emission spectrometry technique used for qualitative and quantitative analysis of multiple chemical elements. Libs technology requires no sample preparation, is virtually non-destructive, rapid, and safe, and is particularly suitable for the detection of light elements such as carbon, lithium, and silicon.

[0003] The spectral data collected by the Libs device is usually output in the form of standardized spectral data files. Its data structure presents clear temporal and spatial characteristics: each row corresponds to the reaction intensity of the full spectrum at a laser dotting moment, and each column corresponds to different points on the material surface hit by the Libs device at different times, forming a three-dimensional data structure of time, space, and spectral intensity. Directly analyzing this type of ultra-large matrix data makes it difficult to intuitively discover the distribution law and temporal change characteristics of the material composition.

[0004] Among existing Libs spectral analysis methods, Chinese patent CN115839942A discloses a quantitative Libs analysis method combining wavelet transform and lightweight convolutional neural networks. It transforms a one-dimensional spectrum into a two-dimensional image using wavelet transform and achieves quantitative analysis through lightweight CNN modeling. However, its core focus is on spectral preprocessing, two-dimensional image conversion, and static modeling, without addressing the precise mapping of three-dimensional spectral information based on spatiotemporal mapping mechanisms or conducting temporal feature mining. It only models single-sample spectral data and cannot capture temporal changes from multiple samplings. Similarly, Chinese patent CN119416670A discloses a method for establishing a laser-induced breakdown spectral quantitative detection model. This method optimizes the quantitative detection model based on sample uncertainty and system uncertainty, focusing on model error correction and static quantitative detection. It does not address spatiotemporal mapping mechanisms or conduct temporal feature mining and prediction, and only optimizes the model for a single detection, failing to capture temporal patterns from multiple samplings.

[0005] In summary, existing technologies have not designed a dedicated spatiotemporal mapping mechanism for the three-dimensional ultra-large matrix data collected by rotating testing equipment, nor do they have suitable temporal feature mining algorithms. They cannot simultaneously meet the requirements of spatiotemporal information preservation and temporal prediction. There is an urgent need for a Libs spectral temporal feature mining and prediction method, system, and storage medium based on the spatiotemporal mapping mechanism to solve the problems existing in the current technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a method for mining and predicting Libs spectral temporal features based on a spatiotemporal mapping mechanism, aiming to solve the technical problems of existing Libs spectral analysis being unable to effectively process three-dimensional spatiotemporal data and lacking temporal prediction capabilities. The specific technical solution is as follows: A method for mining and predicting Libs spectral temporal features based on spatiotemporal mapping mechanism includes: The spectral data collected by the Libs device is used to form a spectral matrix; wherein, the rows of the spectral matrix correspond to the laser spotting time, the columns correspond to the spectral segments, and the values ​​of the matrix cells are spectral intensities; A single-frame color image is generated by performing a three-dimensional spatiotemporal mapping on the spectral matrix; wherein, the three-dimensional spatiotemporal mapping refers to the operation of converting the spectral matrix into a color image while simultaneously retaining the time dimension information, spatial dimension information and spectral intensity dimension information; Multiple samples were taken to obtain the spectral matrix, and three-dimensional spatiotemporal mapping was performed on each sample to obtain multiple frames of color images. The multiple frames of color images were then arranged in order of sampling time to construct a time-series image sequence. The time-series image sequence is input into the prediction model, and the prediction result of the target parameters is output.

[0007] Preferably, the three-dimensional spatiotemporal mapping specifically includes: Map the row indices of the spectral matrix to time steps; The column indices of the spectral matrix are fused with the spectral intensities of the corresponding cells to generate a spatiotemporal feature vector; The spectral intensity values ​​in each spatiotemporal feature vector are linearly normalized and mapped to the pixel value range; The normalized pixel matrix is ​​mapped into a color image according to a preset color scheme.

[0008] Preferably, the spectral intensity values ​​are linearly normalized and mapped to the pixel value range, using the following formula: in, These are the original spectral intensity values; This is the minimum value of the spectral matrix; This represents the maximum value of the spectral matrix; Original spectral intensity values After normalization, the pixel values ​​are mapped to the range [0, 255]. Rounding is required.

[0009] Preferably, after acquiring the spectral data collected by the Libs device and forming a spectral matrix, a data filtering step is also included: Determine whether the number of rows in the spectral matrix has reached a preset standard number of rows; if not, delete the spectral matrix.

[0010] Preferably, the prediction model is a spatiotemporal correlation Transformer model, which specifically includes the following modules: An embedding layer is used to map each spatiotemporal feature vector in a time-series image sequence to a high-dimensional feature vector through a linear transformation, and to add a learnable spatiotemporal location code; wherein, the learnable spatiotemporal location code includes a spatial location code and a temporal location code, the spatial location code is used to identify the spatial spectral location corresponding to the high-dimensional feature vector, and the temporal location code is used to identify the time step within the time-series image sequence; The spatiotemporal correlation Transformer encoder consists of multiple stacked spatiotemporal attention modules, each of which simultaneously performs spatial attention computation and temporal attention computation. The spatial attention is used to capture the spatial spectral correlation features of all high-dimensional feature vectors within the same time step; the temporal attention is used to capture the temporal dependency between high-dimensional feature vectors at the same spatial spectral position at different time steps. The aggregation layer uses an average pooling strategy to aggregate the global spatiotemporal feature vectors output by the spatiotemporal correlation Transformer encoder into global features of fixed dimensions. The regression network consists of at least two fully connected layers, with the intermediate layers using the GELU activation function, and the output layer outputting continuous predicted values ​​of the target parameters.

[0011] Preferably, the spatiotemporal correlation Transformer encoder consists of 6 stacked spatiotemporal attention modules; When training the prediction model, mean squared error is used as the loss function, and a time regularization term is added to penalize abrupt changes in features at adjacent time steps. At the same time, a cosine annealing learning rate strategy is used for optimization.

[0012] This invention also provides a Libs spectral time-series feature mining and prediction system based on a spatiotemporal mapping mechanism. This system employs the aforementioned Libs spectral time-series feature mining and prediction method based on a spatiotemporal mapping mechanism, and includes: The data acquisition module is used to acquire spectral data collected by the Libs device and form a spectral matrix; The spatiotemporal mapping module is used to perform three-dimensional spatiotemporal mapping on the spectral matrix to generate a single-frame color image; The time-series construction module performs three-dimensional spatiotemporal mapping on the spectral matrices obtained from multiple samplings to obtain multiple frames of color images, and arranges the multiple frames of color images in order according to the sampling time to construct a time-series image sequence. The prediction module has a built-in prediction model, which is used to input the time-series image sequence into the prediction model and output the prediction result of the target parameters.

[0013] The present invention also provides another Libs spectral time series feature mining and prediction system based on spatiotemporal mapping mechanism, including a memory and a processor. The memory stores a computer program, and the processor executes the Libs spectral time series feature mining and prediction method based on spatiotemporal mapping mechanism when running the computer program.

[0014] The present invention also provides a storage medium storing a computer program, which is executed by a processor to perform the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism.

[0015] The application of the technical solution of the present invention has the following beneficial effects: This invention transforms ultra-large matrix spectral data into intuitive color images through three-dimensional mapping of time, space-spectral fusion, and intensity normalization dimensions. These images visually reflect the spectral intensity distribution of materials, facilitating rapid assessment of component homogeneity and locating anomalous regions, significantly lowering the barrier to spectral data analysis.

[0016] This invention constructs a time-series image sequence by repeatedly sampling and performing three-dimensional spatiotemporal mapping, arranging multiple frames of color images in chronological order of sampling time. This sequence fully preserves the temporal variation characteristics of spectral intensity, providing a visual basis for material quality control and expanding the application scenarios of Libs spectroscopy in process monitoring.

[0017] This invention designs a spatiotemporal correlation Transformer model and simultaneously implements the superposition of spatial attention and temporal attention in the encoder. Spatial attention captures the correlation features of different spectral points in the same sample, while temporal attention captures the evolution of the same spectral point across different samples. The fusion of the two attention mechanisms enables joint modeling of single-frame spatial dependencies and inter-frame temporal dependencies, improving prediction accuracy and generalization ability.

[0018] This invention introduces learnable spatial and temporal location codes in the embedding layer to accurately distinguish between "the same spatial spectral location at different time steps" and "different spatial spectral locations at the same time step," avoiding confusion of location information and improving the model's ability to identify complex spatiotemporal structures.

[0019] This invention adds a temporal regularization term to the loss function to penalize abrupt changes in features at adjacent time steps, thus avoiding overfitting of the model to temporal noise; at the same time, it adopts a cosine annealing learning rate strategy to improve training efficiency and model stability.

[0020] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism in Example 1; Figure 2 This is a schematic diagram of the collaborative operation of the rotating testing equipment and the Libs equipment in Example 1. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description is provided below, along with preferred embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0024] Example 1: See Figure 1 This embodiment provides a method for mining and predicting Libs spectral temporal features based on a spatiotemporal mapping mechanism, including: The spectral data collected by the Libs device is used to form a spectral matrix; wherein, the rows of the spectral matrix correspond to the laser spotting time, the columns correspond to the spectral segments, and the values ​​of the matrix cells are spectral intensities; A single-frame color image is generated by performing a three-dimensional spatiotemporal mapping on the spectral matrix; wherein, the three-dimensional spatiotemporal mapping refers to the operation of converting the spectral matrix into a color image while simultaneously retaining the time dimension information, spatial dimension information and spectral intensity dimension information; Multiple samples were taken to obtain the spectral matrix, and three-dimensional spatiotemporal mapping was performed on each sample to obtain multiple frames of color images. The multiple frames of color images were then arranged in order of sampling time to construct a time-series image sequence. The time-series image sequence is input into the prediction model, and the prediction result of the target parameters is output.

[0025] The mining prediction method of this embodiment will be described in detail below: See Figure 2First, complete the synchronous debugging of the rotary testing equipment and the Libs equipment. The material to be tested is evenly spread in the groove of the rotary testing equipment, the equipment is set to rotate at a constant speed, and the Libs equipment is synchronously triggered to continuously laser-point, collect the reaction intensity data of the full spectrum of the material surface, and output a standardized spectral data file.

[0026] Specifically, the spectral data file output in this embodiment is a very large matrix (i.e., a spectral matrix) of 600 rows × 4097 columns. Each row corresponds to a laser acupuncture point moment, each column corresponds to the reaction intensity of a spectral band, and the value of each matrix cell is the spectral intensity of the corresponding detection point. In this embodiment, the 600 rows × 4097 columns spectral matrix obtained from a single sampling corresponds to the spectral intensity data of 600 different spatial detection points on the surface of a tested material.

[0027] Preferably, to ensure the data quality of subsequent visualization transformation, the integrity and accuracy of the data should be guaranteed to avoid image distortion caused by dirty data, thereby ensuring the rationality of spatiotemporal mapping. In this embodiment, the preset standard number of acquisition rows is 600 rows (those skilled in the art can flexibly adjust the standard number of acquisition rows), meaning that each complete sampling should obtain spectral data corresponding to 600 laser point times. If the number of acquired spectral matrix rows is less than 600 due to equipment malfunction or uneven material distribution, the sampling data is deemed incomplete, and the spectral matrix is ​​directly deleted and not included in the subsequent 3D spatiotemporal mapping and time-series image sequence construction process. Through this data filtering operation, it is ensured that each frame of color image participating in modeling has a complete temporal length, avoiding the impact of missing data on the rationality of spatiotemporal mapping and the training stability of the prediction model.

[0028] Preferably, to achieve accurate and lossless dimensionality reduction of three-dimensional spectral data in terms of time, space, and spectral intensity, this embodiment constructs a proprietary three-dimensional spatiotemporal mapping mechanism. This mapping mechanism differs from existing simple dimensionality transformation methods; its core lies in completely preserving three-dimensional information and orderly converting it into a two-dimensional color image. Specifically, the three-dimensional spatiotemporal mapping in this embodiment is as follows: The row index of the preprocessed (i.e., after data filtering) spectral matrix is ​​directly mapped to the time step index of the spatiotemporal feature sequence. In this embodiment, the row index corresponds to 600 dot times, and the row order strictly follows the time sequence of laser dotting. Through this mapping, the temporal characteristics of material detection are completely preserved, and accurate mapping of time information is achieved.

[0029] The column index of the spectral matrix is ​​fused with the spectral intensity of the corresponding cell to generate a spatiotemporal feature vector. In this embodiment, the column index corresponds to the wavelength of the spectral segment, which ranges from 159.449 nm to 561.869 nm. Each spatiotemporal feature vector contains both spectral wavelength information and intensity information of the corresponding spatial point, thereby achieving joint preservation of the spatial dimension and spectral dimension and solving the problem of spatiotemporal information separation in existing methods.

[0030] The spectral intensity values ​​in each spatiotemporal feature vector are linearly normalized, mapping them to the interval [0, 255]. This step eliminates the interference of differences in intensity value ranges across different spectral segments on subsequent modeling, improving the stability of model training. Specifically, the normalization formula in this embodiment is: in, These are the original spectral intensity values; This is the minimum value of the spectral matrix; This represents the maximum value of the spectral matrix; Original spectral intensity values After normalization, the pixel values ​​are mapped to the range [0, 255], and the result needs to be rounded up or down.

[0031] After completing the three-dimensional spatiotemporal mapping, all standardized spatiotemporal feature vectors are sequentially concatenated according to the row order of laser dotting times to generate a single batch of Libs spectral spatiotemporal feature sequences. This sequence fully preserves the core temporal, spatial, and spectral intensity information of the original three-dimensional spectral data, laying the foundation for subsequent temporal feature mining.

[0032] Furthermore, after obtaining the normalized pixel matrix, it can be mapped to a color image according to a preset color scheme. In this embodiment, the Viridis color scheme is selected, strictly following the spatiotemporal mapping mechanism to map the normalized pixel matrix to a color image. Of course, in addition to Viridis, the color scheme can also use Plasma, Inferno, Magma, or other sequential color schemes, as long as they can map the normalized pixel values ​​to a visually distinguishable color image.

[0033] Preferably, this embodiment constructs a complete time-series image sequence by repeatedly sampling and performing the above detection and visualization conversion process, thereby realizing the time-series extension of Libs spectral data visualization, fully capturing the time-series changes in the spectral intensity of materials, and providing visualized time-series data for subsequent in-depth prediction.

[0034] Specifically, the process involves repeatedly executing the entire workflow of material loading, rotation detection, spectral data acquisition, data preprocessing, 3D spatiotemporal mapping, and color image generation. Each sampling ensures that the Libs device completes comprehensive material detection and uses the same screening conditions (600 rows or more) and the same 3D spatiotemporal mapping mechanism to generate a single color image. This color image corresponds to the spectral intensity distribution of the sampled material, thus ensuring consistent visualization standards between frames.

[0035] Multiple frames of color images generated from multiple samplings are arranged in chronological order of sampling time to construct a time-series image sequence. The sequence length T (i.e., the number of time steps) can be set according to actual needs. For example, T=10 means 10 consecutive samplings to generate 10 frames of images. Each frame corresponds to the spectral distribution of the material in one sampling, and the order between frames strictly corresponds to the sampling time, thus completely preserving the temporal variation characteristics of spectral intensity and realizing the temporal extension of visualized data.

[0036] Preferably, for the constructed temporal image sequence, this embodiment designs an adaptive prediction model, which is a spatiotemporal correlation Transformer model. This model can simultaneously model the spatial dependency within a single frame and the temporal dependency between frames, thereby realizing temporal feature mining and target parameter prediction.

[0037] Specifically, the spatiotemporal correlation Transformer model includes the following modules: An embedding layer is used to map each spatiotemporal feature vector in the input time-series image sequence to a high-dimensional feature vector and add learnable spatiotemporal location encoding. The learnable spatiotemporal location encoding includes spatial location encoding and temporal location encoding. The spatial location encoding is used to identify the spatial spectral location corresponding to the high-dimensional feature vector, and the temporal location encoding is used to identify the time step within the time-series image sequence.

[0038] The spatiotemporal correlation Transformer encoder consists of multiple stacked spatiotemporal attention modules, each of which simultaneously performs spatial attention computation and temporal attention computation.

[0039] Preferably, in this embodiment, the spatiotemporal correlation Transformer encoder consists of 6 stacked spatiotemporal attention modules. Spatial attention is used to calculate the attention weights between high-dimensional feature vectors of all spatial spectral locations within the same time step, thereby capturing the spatial spectral correlation features of a single batch of Libs spectral data. Temporal attention is used to calculate the attention weights between high-dimensional feature vectors of the same or similar spatial spectral locations at different time steps, thereby capturing the inter-frame temporal dependencies of multiple batches of sampling.

[0040] The aggregation layer employs an average pooling strategy.

[0041] The regression network consists of at least two fully connected layers, with the GELU activation function used in the intermediate layers (i.e., hidden layers).

[0042] Furthermore, the data processing flow of the spatiotemporal correlation Transformer model is as follows: First, the temporal image sequence is input into the embedding layer. The embedding layer first maps each spatiotemporal feature vector into a high-dimensional feature vector through a linear transformation, and then adds a learnable spatiotemporal location code. Through the above processing, the embedding layer outputs the encoded feature sequence, denoted as the encoded feature sequence. Each element in this encoded feature sequence corresponds to a high-dimensional encoded representation of a specific spatial spectral location at a time step, which can accurately distinguish between "the same spatial spectral location at different time steps" and "different spatial spectral locations at the same time step," providing a foundation for subsequent spatiotemporal correlation modeling.

[0043] Then, the encoded feature sequence is input into the spatiotemporal correlation Transformer encoder. Through the stacking of spatial attention and temporal attention, the spatiotemporal correlation Transformer encoder performs layer-by-layer feature extraction and fusion on the input encoded feature sequence, and finally outputs a global spatiotemporal feature vector. This global spatiotemporal feature vector integrates the feature representations of all time steps and all spatial spectral positions in the deepest layer of the encoder.

[0044] Next, the global spatiotemporal feature vector is input into the aggregation layer. The aggregation layer employs an average pooling strategy, averaging the vectors along both the time step and spatial spectral location dimensions, thus aggregating the global spatiotemporal feature vector into a fixed-dimensional global feature. This global feature is a compact vector representation of the entire temporal image sequence, preserving the main components of all input information.

[0045] Finally, the global features output from the aggregation layer are input into a regression network, which is a fully connected network with at least two layers. The intermediate layers use the GELU activation function to improve the model's nonlinear fitting ability, and the output layer uses a linear activation function to output continuous predicted values ​​of the target parameters. For example, it can output the iron content in the material at the next time step.

[0046] Preferably, to improve prediction accuracy and training stability, this embodiment uses mean squared error (MSE) as the loss function and adds a temporal regularization term to penalize abrupt changes in features between adjacent time steps, avoiding overfitting of the model to temporal noise. Furthermore, a cosine annealing learning rate strategy is employed to optimize the training process, adaptively adjusting the learning rate to accelerate model convergence. Through these optimization strategies, a successfully trained temporal feature mining and prediction model is finally obtained.

[0047] It should be noted that the number of stacked layers in the spatiotemporal correlation Transformer encoder is not limited to 6 layers; 4, 8, or other layers can be selected depending on the data scale and computing resources. The number of layers in the regression network can be adjusted from 2 to 4 layers or other layers, and the activation function can be ReLU, SiLU, etc., instead of GELU. The temporal regularization term in the loss function can use the L1 norm or L2 norm to penalize the feature difference between adjacent time steps. The average pooling can also be replaced by max pooling or attention pooling. That is, those skilled in the art can flexibly adjust the composition of the spatiotemporal correlation Transformer model in this embodiment according to the actual situation.

[0048] In practical applications, the Libs spectral data (i.e., spectral matrix) of the material under test in a continuous time series are sequentially subjected to three-dimensional spatiotemporal mapping to generate a time series image sequence to be predicted. This sequence is then input into a trained prediction model, which outputs the time series prediction result of the target parameters corresponding to the material after Libs spectral detection at the next moment. The materials that can be detected by the method in this embodiment include, but are not limited to, ores, sintered ore, raw coal, blended coal, cement, and coke. The predictable elements include iron, calcium, magnesium, aluminum, potassium, sodium, copper, chromium, manganese, and sulfur, and their content values ​​are the time series prediction results output by the algorithm in this embodiment.

[0049] Example 2: This embodiment provides a Libs spectral time-series feature mining and prediction system based on the spatiotemporal mapping mechanism. The system includes: The data acquisition module is used to acquire spectral data collected by the Libs device and form a spectral matrix; wherein, the rows of the spectral matrix correspond to the laser spotting time, the columns correspond to the spectral segments, and the values ​​of the matrix cells are spectral intensities; The spatiotemporal mapping module is used to perform three-dimensional spatiotemporal mapping on the spectral matrix to generate a single-frame color image; the three-dimensional spatiotemporal mapping refers to the operation of converting the spectral matrix into a color image while simultaneously retaining time dimension information, spatial dimension information and spectral intensity dimension information. The time-series construction module performs three-dimensional spatiotemporal mapping on the spectral matrices obtained from multiple samplings to obtain multiple frames of color images, and arranges the multiple frames of color images in order according to the sampling time to construct a time-series image sequence. The prediction module has a built-in prediction model, which is used to input the time-series image sequence into the prediction model and output the prediction result of the target parameters. Preferably, the prediction model is a spatiotemporal correlation Transformer model, the specific structure of which is described in Embodiment 1.

[0050] The system in this embodiment can execute the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism described in Embodiment 1. It acquires spectral data through a data acquisition module, generates a single-frame color image through a spatiotemporal mapping module, constructs a time-series image sequence through a time-series sequence construction module, and outputs the target parameter prediction result through a prediction module. This system can be used to achieve time-series prediction of material composition content, lowering the threshold for spectral data analysis.

[0051] Example 3: This embodiment provides a Libs spectral time-series feature mining and prediction system based on spatiotemporal mapping mechanism, including a memory and a processor. The memory stores a computer program, and when the processor runs the computer program, it executes the Libs spectral time-series feature mining and prediction method based on spatiotemporal mapping mechanism in Embodiment 1.

[0052] Example 4: This embodiment provides a storage medium storing a computer program, which is executed by a processor to perform the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism in Embodiment 1.

[0053] Preferably, the storage medium can be any medium capable of storing data, such as read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, etc.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for mining and predicting Libs spectral temporal features based on spatiotemporal mapping mechanism, characterized in that, include: The spectral data collected by the Libs device is used to form a spectral matrix; wherein, the rows of the spectral matrix correspond to the laser spotting time, the columns correspond to the spectral segments, and the values ​​of the matrix cells are spectral intensities; A single-frame color image is generated by performing a three-dimensional spatiotemporal mapping on the spectral matrix; wherein, the three-dimensional spatiotemporal mapping refers to the operation of converting the spectral matrix into a color image while simultaneously retaining the time dimension information, spatial dimension information and spectral intensity dimension information; Multiple samples were taken to obtain the spectral matrix, and three-dimensional spatiotemporal mapping was performed on each sample to obtain multiple frames of color images. The multiple frames of color images were then arranged in order of sampling time to construct a time-series image sequence. The time-series image sequence is input into the prediction model, and the prediction result of the target parameters is output.

2. The Libs spectral temporal feature mining and prediction method based on spatiotemporal mapping mechanism according to claim 1, characterized in that, The three-dimensional spatiotemporal mapping specifically includes: Map the row indices of the spectral matrix to time steps; The column indices of the spectral matrix are fused with the spectral intensities of the corresponding cells to generate a spatiotemporal feature vector; The spectral intensity values ​​in each spatiotemporal feature vector are linearly normalized and mapped to the pixel value range; The normalized pixel matrix is ​​mapped into a color image according to a preset color scheme.

3. The Libs spectral temporal feature mining and prediction method based on spatiotemporal mapping mechanism according to claim 2, characterized in that, The spectral intensity values ​​are linearly normalized and mapped to pixel value ranges using the following formula: in, These are the original spectral intensity values; This is the minimum value of the spectral matrix; This represents the maximum value of the spectral matrix; Original spectral intensity values After normalization, the pixel values ​​are mapped to the range [0, 255]. Rounding is required.

4. The Libs spectral temporal feature mining and prediction method based on spatiotemporal mapping mechanism according to claim 1, characterized in that, After acquiring the spectral data collected by the Libs device and forming a spectral matrix, the process also includes a data filtering step: Determine whether the number of rows in the spectral matrix has reached a preset standard number of rows; if not, delete the spectral matrix.

5. The Libs spectral temporal feature mining and prediction method based on spatiotemporal mapping mechanism according to claim 1, characterized in that, The prediction model is a spatiotemporal correlation Transformer model, which specifically includes the following modules: An embedding layer is used to map each spatiotemporal feature vector in a time-series image sequence to a high-dimensional feature vector through a linear transformation, and to add a learnable spatiotemporal location code; wherein, the learnable spatiotemporal location code includes a spatial location code and a temporal location code, the spatial location code is used to identify the spatial spectral location corresponding to the high-dimensional feature vector, and the temporal location code is used to identify the time step within the time-series image sequence; The spatiotemporal correlation Transformer encoder consists of multiple stacked spatiotemporal attention modules, each of which simultaneously performs spatial attention computation and temporal attention computation. The spatial attention is used to capture the spatial spectral correlation features of all high-dimensional feature vectors within the same time step; the temporal attention is used to capture the temporal dependency between high-dimensional feature vectors at the same spatial spectral position at different time steps. The aggregation layer uses an average pooling strategy to aggregate the global spatiotemporal feature vectors output by the spatiotemporal correlation Transformer encoder into global features of fixed dimensions. The regression network consists of at least two fully connected layers, with the intermediate layers using the GELU activation function, and the output layer outputting continuous predicted values ​​of the target parameters.

6. The Libs spectral temporal feature mining and prediction method based on spatiotemporal mapping mechanism according to claim 5, characterized in that, The spatiotemporal correlated Transformer encoder consists of 6 stacked spatiotemporal attention modules; When training the prediction model, mean squared error is used as the loss function, and a time regularization term is added to penalize abrupt changes in features at adjacent time steps. At the same time, a cosine annealing learning rate strategy is used for optimization.

7. A Libs spectral temporal feature mining and prediction system based on spatiotemporal mapping mechanism, characterized in that, The system employs the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism as described in any one of claims 1-6, including: The data acquisition module is used to acquire spectral data collected by the Libs device and form a spectral matrix; The spatiotemporal mapping module is used to perform three-dimensional spatiotemporal mapping on the spectral matrix to generate a single-frame color image; The time-series construction module performs three-dimensional spatiotemporal mapping on the spectral matrices obtained from multiple samplings to obtain multiple frames of color images, and arranges the multiple frames of color images in order according to the sampling time to construct a time-series image sequence. The prediction module has a built-in prediction model, which is used to input the time-series image sequence into the prediction model and output the prediction result of the target parameters.

8. A Libs spectral temporal feature mining and prediction system based on spatiotemporal mapping mechanism, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism as described in any one of claims 1-6 when running the computer program.

9. A storage medium, characterized in that, The storage medium stores a computer program, which is executed by a processor to perform the Libs spectral temporal feature mining and prediction method based on the spatiotemporal mapping mechanism as described in any one of claims 1-6.

Images