A method for determining high calorific value of biomass based on laser-induced breakdown spectroscopy

Abstract

The application discloses a kind of biomass high calorific value determination method based on laser-induced breakdown spectroscopy, it is related to spectral determination technical field, for solving the problem that the mapping relationship between spectral characteristics and heat value is not clear, resulting in the problem of reduced prediction accuracy, by collecting spectral characteristics to biomass sample and constructing spectral data, combined with calibration heat value executes different data enhancement processing to form synthetic training set, the spectral characteristics are compressed using attention mechanism and the feature weight is calculated, the features are screened and marked by gradient response, based on random forest, the importance of the labeled spectral characteristics is evaluated and the dimension is suppressed to obtain the target feature, a regression model is constructed to accurately predict the heat value of biomass samples, improve the utilization efficiency of spectral characteristics, reduce the interference of redundant features, enhance the recognition ability of the model to key features, and improve the stability and consistency of the heat value prediction result.

Description

Technical Field

[0001] This invention relates to the field of spectroscopic measurement technology, and more specifically, to a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy. Background Technology

[0002] As the application of biomass energy in alternative energy systems continues to expand, the demand for rapid determination of the calorific value of biomass materials is gradually increasing. Laser-induced breakdown spectroscopy (LAS) technology, due to its advantages such as no need for complex chemical processing, fast response speed, and ability to acquire multi-element information simultaneously, is gradually being applied to the field of biomass calorific value detection and has good application prospects in online detection and batch analysis scenarios.

[0003] The existing technology has the following shortcomings: Currently, existing technologies for predicting the calorific value of biomass based on spectral data struggle to construct stable feature representation systems under conditions of uneven sample distribution and complex feature expressions. They also lack data augmentation mechanisms that differentiate between samples with varying fitting difficulties, easily leading to over-reliance on mainstream sample features during model training. This results in a decreased responsiveness of prediction results to marginal samples. Furthermore, they lack hierarchical screening and effective compression methods for the large amount of redundant information in high-dimensional spectral features, making it impossible to establish a screening mechanism driven by feature contribution. This leads to reduced calorific value prediction accuracy and increased fluctuations in prediction results. Therefore, this paper proposes a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy. This method constructs evenly distributed training data by employing a differential sample enhancement mechanism, combines an attention mechanism to achieve hierarchical compression and weight allocation of spectral features, establishes a feature contribution-driven screening path based on gradient response, and further integrates random forest feature importance assessment and feature subset optimization strategies to construct a regression model centered on the target feature, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy, comprising the following steps: Step S1: After slicing the biomass sample to be tested, a sheet sample to be tested is generated. The calibrated calorific value of the sheet sample to be tested is collected. The spectral characteristics of the sheet sample to be tested are obtained using a laser-induced device and spectral data is constructed. Step S2: Perform prediction processing on the sheet-like sample to be tested, obtain the predicted calorific value of the sheet-like sample to be tested, calculate the fitting score by combining the calibrated calorific value, and output the synthetic training set after interpolating and merging the spectral data of the sheet-like sample to be tested according to the fitting score. Step S3: After performing layer-by-layer compression of the spectral features of the synthetic training set using the attention mechanism, propagation calculation is performed on the synthetic training set to generate gradient response values ​​of the spectral features. The spectral features are then filtered and labeled based on the gradient response values. Step S4: Dimensionality suppression of the labeled spectral features is performed using random forest and the target features are output. The training set and target features are combined to construct a regression model and map the calorific value of the biomass sample to be tested.

[0007] In a preferred embodiment, in step S1, the biomass sample to be tested is sliced. The biomass sample to be tested refers to the original sample derived from crop straw, wood or other organic biomass materials. Slicing process refers to placing the biomass sample to be tested in a dry environment and treating it at a constant temperature. Then, the dried sample is crushed into granular powder by mechanical grinding equipment. The powder particle size is then uniformly screened by a sieving device. The screened powder is placed in a pressing mold and pressed into shape under a set pressure condition to form several slices of the sample to be tested. The calibrated calorific value of the sheet sample to be tested is collected. The calibrated calorific value refers to the heat released by the complete combustion of a unit mass, as measured by a standard calorific value measuring device.

[0008] In a preferred embodiment, in step S1, a laser-induced device is used to acquire spectral features of the sheet-like sample to be tested. The laser-induced device is a detection system that can emit high-energy pulsed lasers and excite the sheet-like sample to be tested to generate plasma luminescence, including a laser emitting unit, an optical focusing unit, a sample positioning unit, and a spectral acquisition unit. The laser pulse from the laser emitting unit irradiates the surface of the sheet sample under test, causing the local area of ​​the sheet sample to heat up instantaneously and form a high-temperature plasma. The plasma releases emitted light signals of different wavelengths during the cooling process. The emitted light signal is decomposed into different wavelength components by the spectral acquisition unit, and the light intensity value corresponding to each wavelength is recorded, thereby forming spectral features and constructing spectral data; Spectral characteristics refer to the light intensity values ​​measured at different wavelength positions, while spectral data refers to the set of values ​​formed by arranging the spectral characteristics at each wavelength position in a fixed order.

[0009] In a preferred embodiment, in step S2, the spectral data is input into the prediction model for prediction processing. The prediction model refers to a regression calculation model used to establish the mapping relationship between spectral features and calibrated calorific values. Predictive processing refers to using an established predictive model to calculate the input spectral data and output the corresponding predicted calorific value. The predicted calorific value is the estimated calorific value calculated by the predictive model based on spectral features. The absolute difference between the predicted calorific value and the calibrated calorific value is calculated to obtain the fitting residual of the sheet sample to be tested. The Max-Min normalization method is used to normalize the fitting residual to obtain the fitting score. Based on the fitting score, all the sheet-like samples to be tested are sorted, and the sheet-like samples to be tested are divided into a set of difficult-to-fit samples and a set of easy-to-fit samples according to the preset scoring threshold.

[0010] In a preferred embodiment, in step S2, a differential interpolation merging process is performed on the hard-to-fit sample set and the easy-to-fit sample set. The interpolation merging process includes interpolation processing and merging processing. Interpolation refers to generating new feature samples and their corresponding calibrated calorific values ​​by constructing a linear combination in the spectral feature space of a known sample, based on the continuous relationship between neighboring samples. Merging refers to the process of incorporating the original test samples and new samples into the same dataset and rearranging the order of the samples to form a synthetic training set containing both the original information and the augmented information.

[0011] In a preferred embodiment, in step S3, the spectral features in the synthetic training set are standardized to obtain standardized spectral features. The standardized spectral features are input into the encoder layer of the attention neural network, and the features are transformed layer by layer to output the encoded feature matrix. Based on the encoded output feature matrix, the intermediate mapping value corresponding to each feature dimension of each sample is calculated, and normalization processing is performed on all feature dimensions of the same sample to obtain the attention weights corresponding to each feature dimension. The attention weights are matched dimension-by-dimensionally with the feature values ​​at the corresponding positions in the encoded output feature matrix to obtain the weighted feature result.

[0012] In a preferred embodiment, in step S3, regression prediction is performed based on the weighted feature results to obtain the predicted heat value; with the predicted heat value as the output, backpropagation processing is performed on the encoded output feature matrix to obtain the gradient response value of each sample in each feature dimension. The gradient response values ​​of the same feature dimension across all samples are processed by absolute value and averaged to obtain the gradient score of the corresponding feature dimension. All gradient scores are normalized and sorted. The screening boundary is determined according to the preset percentile position. The feature dimensions that meet the screening conditions are marked as labeled spectral features, and their feature index positions in the encoded output feature matrix are recorded. The labeled spectral feature set is then output.

[0013] In a preferred embodiment, in step S4, the feature index position in the marked spectral feature set is read, and the value of the corresponding feature dimension under each sample is extracted from the encoded output feature matrix to form an initial feature set; Training data is constructed based on the initial feature set, and the calibrated heat value corresponding to each sample is used as the output label to train the random forest regression model. During model training, the contribution of each feature to the reduction of prediction error when participating in node partitioning is statistically analyzed to obtain the importance score of each feature; The features in the initial feature set are filtered based on the importance scores of each feature, and features with importance scores below the filtering boundary are removed to obtain the target feature set.

[0014] In a preferred embodiment, in step S4, corresponding feature data are extracted from the synthetic training set based on the target feature set, a target training feature matrix is ​​constructed, and the calibration heat value corresponding to each sample is used as the training label to train the random forest regression model and obtain the final regression model. The spectral characteristics of the biomass sample to be tested are input into the final regression model, and the calorific value of the biomass sample to be tested is output.

[0015] The technical effects and advantages of this invention are as follows: This invention compensates for the distribution of training data by constructing a differentiated sample enhancement mechanism, combines an attention mechanism to achieve hierarchical compression and weight allocation of spectral features, establishes a feature contribution-driven screening path based on gradient response, and integrates random forest feature importance assessment and feature subset optimization to perform multi-level screening and dimensionality suppression of spectral features. This improves the expressive power of key features, reduces redundant feature interference, improves the accuracy of heat value prediction, and enhances the stability and consistency of prediction results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy, as described in this invention.

[0017] Figure 2 This is a schematic diagram illustrating the steps of a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention compensates for the distribution of training data by constructing a differentiated sample enhancement mechanism, combines an attention mechanism to achieve hierarchical compression and weight allocation of spectral features, establishes a feature contribution-driven screening path based on gradient response, and integrates random forest feature importance assessment and feature subset optimization to perform multi-level screening and dimensionality suppression of spectral features, thereby achieving accurate mapping and prediction of the calorific value of the biomass sample to be tested.

[0020] Example 1, as Figures 1 to 2 As shown, a method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy includes the following steps: Step S1: After slicing the biomass sample to be tested, a sheet sample to be tested is generated. The calibrated calorific value of the sheet sample to be tested is collected. The spectral characteristics of the sheet sample to be tested are obtained using a laser-induced device and spectral data is constructed. Step S2: Perform prediction processing on the sheet-like sample to be tested, obtain the predicted calorific value of the sheet-like sample to be tested, calculate the fitting score by combining the calibrated calorific value, and output the synthetic training set after interpolating and merging the spectral data of the sheet-like sample to be tested according to the fitting score. Step S3: After performing layer-by-layer compression of the spectral features of the synthetic training set using the attention mechanism, propagation calculation is performed on the synthetic training set to generate gradient response values ​​of the spectral features. The spectral features are then filtered and labeled based on the gradient response values. Step S4: Dimensionality suppression of the labeled spectral features is performed using random forest and the target features are output. The training set and target features are combined to construct a regression model and map the calorific value of the biomass sample to be tested.

[0021] The specific implementation is as follows: In step S1, the biomass sample to be tested is sliced, its calorific value is calibrated, and its spectral data is acquired and processed to form the basic data for subsequent analysis.

[0022] The biomass sample to be tested refers to the original sample derived from crop straw, wood or other organic biomass materials. Its components are combustible organic matter such as cellulose, hemicellulose and lignin, which are the objects of subsequent calorific value determination and spectral analysis.

[0023] First, the biomass sample to be tested is sliced ​​to generate sheet-like samples. Specifically, the biomass sample is placed in a dry environment and subjected to constant temperature treatment to reduce the moisture content to a preset level, so as to avoid the interference of moisture with subsequent spectral signals and calorific value measurements. Then, the dried sample is pulverized into granular powder using a mechanical grinding device, and the powder particle size is uniformly screened by a sieving device to ensure that the sample particle size is consistent and to ensure the uniformity of the laser interaction area. On this basis, the screened powder is placed in a pressing mold and pressed into shape under a set pressure condition to form a sheet-like structure with a fixed shape and a certain density, thereby obtaining several sheet-like samples to be tested.

[0024] The sheet-like sample to be tested refers to a regular solid sample formed after being pressed. Its surface is flat and its structure is uniform, which is used to ensure the stability and repeatability of the laser excitation process.

[0025] It should be noted that mechanical grinding equipment refers to a processing device used to apply mechanical force to the dried biomass sample to reduce the particle size. It breaks the biomass sample into powder particles through shearing, impact, or extrusion. Sieving device refers to a device used to classify and screen the powder after grinding. It separates the particles through a sieve with a preset aperture to obtain a sample set with the required particle size. Pressing mold refers to a molding device used to press the powder sample into a sheet-like structure under external pressure. It achieves sample densification by limiting the geometry and volume of the sample.

[0026] After obtaining the sheet-like sample to be tested, its calibrated calorific value is collected. The calibrated calorific value refers to the heat released by the complete combustion of a unit mass, measured using a standard calorific value measuring device. It is a physical quantity reflecting the energy content of biomass and serves as a reference label for subsequent model training. Specifically, the sheet-like sample to be tested is placed in a closed combustion environment, and the temperature change before and after combustion is recorded through a complete combustion process under sufficient oxygen conditions. The calibrated calorific value of the sample is then calculated using the heat capacity parameters of the standard calorific value measuring device. The calibrated calorific value characterizes the true energy level of the sheet-like sample; a higher value indicates higher combustible energy.

[0027] It should be noted that a standard calorific value measuring device refers to a testing device used to measure the heat released by the complete combustion of a unit mass of sample, including a sealed combustion container, a temperature measuring unit, and a heat capacity calibration unit.

[0028] After obtaining the calibrated calorific value, a laser-induced device is used to acquire the spectral characteristics of the sheet-like sample under test. The laser-induced device is a detection system capable of emitting high-energy pulsed laser light and exciting the sheet-like sample to produce plasma luminescence. It includes a laser emitting unit, an optical focusing unit, a sample positioning unit, and a spectral acquisition unit. The laser emitting unit outputs high-energy laser pulses, the optical focusing unit focuses the laser light onto a specific area on the surface of the sheet-like sample, the sample positioning unit adjusts the position of the sample to achieve multi-point sampling, and the spectral acquisition unit records the luminescence signal.

[0029] During the specific data acquisition process, laser pulses from the laser emission unit irradiate the surface of the sheet-like sample under test, causing a localized area of ​​the sample to heat up instantaneously and form high-temperature plasma. As the plasma cools, it releases emitted light signals of different wavelengths. These emitted light signals are decomposed into different wavelength components by the spectral acquisition unit, and the intensity values ​​corresponding to each wavelength are recorded, thus forming spectral characteristics. Spectral characteristics refer to the light intensity values ​​measured at different wavelengths, reflecting the elements and their energy level transition characteristics in the sheet-like sample. Different elements correspond to different characteristic wavelengths, and the magnitude of their intensity values ​​reflects the content and excitation degree of the corresponding element. Spectral data refers to the numerical set formed by arranging the spectral characteristics at each wavelength position in a fixed order, used to characterize the overall spectral response of the sheet-like sample under test.

[0030] To improve data stability, multiple different locations were selected on the surface of each sheet sample to be tested for repeated excitation and sampling. The spectral features obtained from multiple acquisitions were averaged to reduce the impact of local inhomogeneity and random noise.

[0031] In step S2, prediction evaluation and differential data augmentation processing are performed on the spectral data corresponding to the sheet-like sample obtained in step S1 to construct a balanced synthetic training set.

[0032] Specifically, the spectral data is first input into the prediction model for prediction processing. Prediction processing refers to using the established prediction model to calculate the input spectral data and output the corresponding predicted calorific value. The predicted calorific value is the estimated calorific value calculated by the prediction model based on spectral features, which is used to characterize the model's ability to fit the current sample.

[0033] After obtaining the predicted calorific value, it is compared with the calibrated calorific value to construct a fit score. Specifically, the absolute difference between the predicted calorific value and the calibrated calorific value is calculated to obtain the fitting residual of the sheet-like sample to be tested. The larger the fitting residual value, the greater the fitting error of the prediction model to the sheet-like sample to be tested, and the more difficult it is for the prediction model to accurately learn the sheet-like sample to be tested.

[0034] To eliminate the dimensional differences in the fitting residuals among different sheet-like samples, the Max-Min normalization method was used to normalize the fitting residuals and obtain a fitting score. The fitting score is an evaluation index used to quantify the degree of deviation between the prediction result and the true value. Its value reflects the difficulty of the prediction model fitting the sheet-like sample. The closer the fitting score is to 1, the higher the difficulty of fitting the sheet-like sample. The closer the value is to 0, the higher the degree of fitting of the sheet-like sample.

[0035] It should be noted that the prediction model refers to a regression calculation model used to establish the mapping relationship between spectral features and calibrated calorific value. It is pre-built based on historical training data, and its parameters are obtained by minimizing the objective function of prediction error. The input is the spectral features of the sheet sample to be tested, and the output is the predicted calorific value.

[0036] Based on the fitting score, all the test sheet samples are sorted and divided into a difficult-to-fit sample set and a easy-to-fit sample set according to a preset scoring threshold. Specifically, test sheet samples with a fitting score greater than or equal to the preset scoring threshold are assigned to the difficult-to-fit sample set, and test sheet samples with a fitting score less than the preset scoring threshold are assigned to the easy-to-fit sample set.

[0037] It should be noted that the preset scoring threshold is a discrimination parameter used to distinguish between the set of difficult-to-fit samples and the set of easy-to-fit samples. It is derived from the statistical analysis of the distribution of all fitting scores and is obtained by calculating the quantiles of all fitting scores.

[0038] After the sample division is completed, differential interpolation merging is performed on the hard-to-fit sample set and the easy-to-fit sample set. The interpolation merging process includes interpolation processing and merging processing.

[0039] Interpolation refers to the process of generating new samples and their corresponding calibrated calorific values ​​by constructing linear combinations in the spectral feature space of known samples, based on the continuous relationship between neighboring samples. Essentially, it involves continuously expanding the sample space while maintaining the original data distribution structure.

[0040] For a set of hard-to-fit samples, a hard-to-fit expansion ratio is set. Neighboring samples whose Euclidean distance to the spectral features of the current sheet-like sample to be tested is less than a preset distance threshold are selected from the corresponding spectral features. Sample pairs are constructed based on the neighborhood relationship, and interpolation is performed between the sample pairs based on the hard-to-fit expansion ratio to generate new spectral features, thereby generating new samples. The interpolation operation refers to generating intermediate features between two spectral features according to a random ratio, so that the generated new samples are within the data distribution range of the original samples, thereby ensuring the physical rationality of the data.

[0041] At the same time, the corresponding calibration calorific value is calculated synchronously according to the same interpolation ratio to generate the calibration calorific value of the new sample, so that the new sample maintains a consistent mapping relationship between spectral characteristics and calibration calorific value.

[0042] For the set of easily fittable samples, an expansion ratio is set. The expansion ratio for difficult-to-fit samples is greater than that for easily fit samples, reflecting the emphasis on enhancing the difficult-to-fit samples. Interpolation is performed on the easily fittable sample set while maintaining data diversity to avoid introducing too many redundant samples. If there is insufficient neighboring samples during sample generation, new samples are generated by expanding the neighborhood range to ensure the continuity of the generation process.

[0043] After interpolation, all generated spectral features are merged with the original spectral features. Merging involves incorporating the original sheet-like samples and the new samples into the same dataset and rearranging the sample order to form a synthetic training set containing both original and enhanced information.

[0044] The final output synthetic training set refers to the training data set obtained after differential augmentation guided by fitting score. Its data distribution is more balanced and can cover the areas that are difficult to fit in the original data, thus providing a higher quality data foundation for subsequent feature selection and model construction.

[0045] In step S3, the spectral features in the synthetic training set are subjected to normalization preprocessing, which specifically includes the following: Select spectral features from the synthetic training set and perform standardization on each spectral feature dimension: ; in, Let i be the normalized spectral features of the i-th synthetic training sample in the j-th dimension. Let i be the spectral feature of the i-th synthetic training sample in the j-th dimension. Let be the mean of the j-th spectral feature. Let be the standard deviation of the j-th spectral characteristic; The standardized spectral features are compressed layer by layer using an attention neural network, which includes an encoder layer, an attention layer, and a regression layer. The encoder layer is used to perform hierarchical compression of spectral features. For the k-th encoder layer, its output is calculated according to the following formula: ; in, This represents the output of the k-th layer encoder. This represents the output of the (k-1)th layer, where when k=1, A standardized spectral feature matrix is ​​formed by arranging all standardized spectral features in order of sample number and feature dimension. This represents the weight matrix corresponding to the k-th layer encoder. This represents the bias vector corresponding to the k-th layer encoder. This represents the linear rectification activation function, whose calculation rule is: when the input value is greater than 0, the output is the original value; when the input value is less than or equal to 0, the output is 0. " indicates matrix multiplication; in, and All of these are trainable parameters of the k-th layer encoder, derived from the initial parameter set of the attention neural network, and obtained through backpropagation of the loss function and parameter updates during network training using the synthetic training set and corresponding calibration heat values.

[0046] During the layer-by-layer compression process, each layer calculates the compressed interlayer feature vector by applying the encoder layer calculation formula to the output of the previous layer. The interlayer feature vector refers to the feature representation obtained by transforming the input spectral features in the current encoder layer, which is used as the input data for the next encoder layer. The compression process is considered complete when the preset number of compression layers is reached. It should be explained that the preset number of compression layers can be set according to the initial feature dimension of the standardized spectral features and the convergence of the prediction error during network training.

[0047] After compression, the encoded output feature matrix is ​​obtained, denoted as... The encoded output feature matrix is ​​the spectral feature representation result after multi-layer compression.

[0048] Attention weights for each spectral feature are calculated using an attention layer. Specifically, a nonlinear mapping is first performed on the encoded output feature matrix, and then softmax normalization is used to obtain the weight values ​​for each dimension. The calculation process is as follows: ; ; in: This represents the intermediate mapping value of the i-th sample on the j-th encoded output feature dimension. This represents the encoded output feature vector corresponding to the i-th sample. This represents the weight vector corresponding to the j-th encoded output feature dimension. This represents the vector dot product operation. This represents the bias parameter corresponding to the j-th encoded output feature dimension. This represents the attention weight of the i-th sample on the j-th encoded output feature dimension. The expression represents the exponential operation, t represents the index of the spectral feature dimension participating in the weight normalization calculation, and m represents the total number of spectral feature dimensions.

[0049] The intermediate mapping value is the transitional calculation result obtained by the attention layer after performing mapping calculation on the j-th spectral feature in the encoded output feature matrix, and is used to participate in the subsequent attention weight normalization calculation. and These are all trainable parameters of the attention layer, derived from the initial parameter set of the attention neural network, and obtained through backpropagation of the loss function and parameter updates during network training using the synthetic training set and corresponding calibration heat values.

[0050] The attention weights are combined to obtain the attention weight sequence, denoted as . , where m represents the number of dimensions of the compressed spectral features.

[0051] Each weight value in the attention weight sequence is obtained through normalization calculation, and the sum of all weights is 1, which is used to represent the relative contribution ratio of each spectral feature in the current prediction task. Subsequently, the weighted spectral features are obtained by combining the attention weight sequence and the encoded output features, and the calculation formula is as follows: ; in, This represents the weighted result of the i-th sample on the j-th spectral feature dimension; , This represents the value of the i-th sample in the j-th encoded output feature dimension.

[0052] The weighted spectral features are combined to obtain a weighted spectral feature matrix, denoted as H, which is used as input to the regression layer to perform calorific value prediction.

[0053] In the regression layer, the predicted calorific value is generated using the weighted spectral feature matrix H. For the i-th sample, its predicted calorific value is denoted as y^(i), and the corresponding calculation relationship can be expressed as: ; in, This represents the predicted calorific value of the i-th sample. This represents the weighted spectral feature vector corresponding to the i-th sample. This represents the output mapping function established by the regression layer.

[0054] The output mapping function is a nonlinear mapping relationship determined by the regression layer parameters in the current training state of the attention neural network, which is used to convert the weighted spectral feature vector of the i-th sample into the corresponding predicted heat value.

[0055] After obtaining the predicted calorific value, backpropagation calculation is performed on the regression layer output relative to the encoded output feature matrix, and each feature dimension in the encoded output feature matrix is ​​used as the gradient response calculation object to measure the influence of each encoded feature on the predicted calorific value, so as to obtain the gradient response value of each spectral feature. The gradient response value for the j-th spectral feature is calculated as follows: ; in, This represents the gradient response value of the j-th spectral feature in the i-th sample to the predicted calorific value. This represents the partial derivative of the predicted calorific value with respect to the corresponding weighted spectral characteristics.

[0056] Gradient response values ​​are used to predict the magnitude of the response to calorific value when spectral features change slightly. To eliminate individual fluctuations between different samples, the absolute values ​​of the gradient response values ​​of all samples on the same spectral feature dimension are taken and averaged to obtain the gradient score for that spectral feature. The calculation formula is as follows: ; in: Represents the gradient score of the j-th spectral feature. Indicates the number of samples involved in the calculation. This represents the gradient response value of the i-th sample in the j-th spectral feature dimension; The gradient scores of all spectral features are normalized to obtain the normalized gradient scores, calculated using the following formula: in, Represents the normalized gradient score of the j-th spectral feature. This represents the original gradient score of the j-th spectral feature. This represents the minimum value among all spectral feature gradient scores. This represents the maximum value among all spectral feature gradient scores.

[0057] After obtaining the normalized gradient score, all spectral features are sorted according to the numerical value of the normalized gradient score, and the screening boundary is determined according to the preset percentile threshold.

[0058] Specifically, a percentile threshold of p is set, all normalized gradient scores are sorted according to their numerical values, and the score value corresponding to the p-th percentile is taken as the selection threshold. ; If a certain spectral feature satisfies the following formula: ; Then the spectral feature is preserved and labeled; If a certain spectral feature satisfies the following formula: ; Then remove that spectral feature from the current feature set.

[0059] Record the feature index position of each marker spectral feature in the encoded output feature matrix, and output the marker spectral feature set.

[0060] This represents the filtering boundary value calculated based on the preset percentile threshold, where p represents the preset percentile parameter, used to control the range of the number of features to be retained.

[0061] It should be explained that the preset percentile threshold can be set according to the distribution of normalized gradient scores corresponding to each spectral feature in the synthetic training set and the trend of model prediction error.

[0062] This step involves constructing an attention neural network and combining it with gradient response calculation to quantify the contribution of spectral features in the synthetic training set and perform percentile screening. This process can eliminate redundant spectral features while maintaining the correlation information for calorific value prediction, thereby providing more compact and targeted input data for subsequent feature optimization and regression model construction.

[0063] In step S4, the feature index positions in the marked spectral feature set output in step S3 are read, and the values ​​of the corresponding feature dimensions are extracted from the encoded output feature matrix according to the feature index positions to form an initial feature set; It should be noted that, in this embodiment, the feature importance score obtained by the random forest model is only used to construct the initial search space of the feature selection optimization algorithm. Its role is to provide probabilistic guidance for the initial feature set in order to improve the convergence efficiency and search effectiveness of the subsequent feature subset search process, rather than serving as a direct basis for judging the final feature selection result. The encoded output feature matrix is ​​derived from the output of the encoder processing of the synthetic training set in step S3.

[0064] A random forest regression model is trained using an initial feature set, and the importance score corresponding to each spectral feature is calculated. Specifically, the spectral features of each sample in the synthetic training set in the corresponding dimension of the initial feature set are selected as input data, and the calibrated calorific value of each sample is selected as output data to train the random forest regression model. During the training of the random forest, the error changes corresponding to each split node of each decision tree are statistically analyzed. For the k-th split node in the (t)-th decision tree, the error reduction before and after the split is calculated as follows: ; in, This represents the error reduction at the k-th split node in the (t)-th decision tree; This represents the squared error of the split node before the split; This represents the sum of the squared errors of the left and right child nodes after the split node splits.

[0065] The squared error before splitting can be obtained by squaring and summing the differences between the calibrated calorific value and the predicted calorific value of all samples in the splitting node. The sum of the squared errors after splitting can be obtained by squaring and summing the differences between the calibrated calorific value and the predicted calorific value of the corresponding child node for all samples in the left and right child nodes respectively, and then summing the squared errors of the left and right child nodes.

[0066] For the (j)th spectral feature, the sum of the error reduction corresponding to the splitting nodes in all decision trees is calculated to obtain the importance score of the spectral feature: ; in, represents the importance score of the (j)th spectral feature; (T) represents the number of decision trees in the random forest; Let represent the set of nodes in the (t)th decision tree that are split using the (j)th spectral feature; The importance scores corresponding to each spectral feature are arranged according to the spectral feature dimension number to form a feature importance vector. , where (d) represents the number of spectral feature dimensions in the initial feature set.

[0067] Feature Importance Score It is a quantitative result obtained by summing the error reduction of each split node, used to characterize the cumulative effect of the (j)th spectral feature on the reduction of the random forest prediction error. The larger the value, the greater the contribution of the spectral feature to the error reduction on the current training data.

[0068] After obtaining the feature importance vector, the importance scores of each spectral feature are normalized, and the normalization results are used to construct the initial ISABO population.

[0069] After obtaining the feature importance score and completing the normalization process, the initial population distribution of the ISABO algorithm is constructed based on the score result. Then, the ISABO algorithm is used to perform iterative optimization in the continuous feature selection space. With prediction error and feature dimension constraint as the joint optimization objective, the feature subset is globally searched and dynamically updated, and finally the target feature set that meets the requirements of accuracy and sparsity is output. Specifically, select each score value from the feature importance vector. The feature importance score (I'j) is obtained by mapping it to the interval ([0,1]) according to the normalization formula. Then, the initial selection probability of the (j)th spectral feature is calculated based on the normalized feature importance score: ; in, This represents the probability of the (j)th spectral feature being selected during the initial population construction; Represents the basic probability coefficient, used to control the overall range of probabilities of all spectral features being selected; ) represents the normalized feature importance score of the (j)th spectral feature.

[0070] To prevent the initial selection probability of certain spectral features from being too high, thus leading to homogenization of the initial population, [the following measures are taken]. Setting an upper limit constraint can be based on the requirements for the degree of feature overlap of individuals in the initial population and the requirements for maintaining population diversity, so that a single spectral feature will not appear repeatedly in the initial population with an excessively high probability; for example, the upper limit constraint can be set to not exceed a preset probability upper bound.

[0071] Based on the initial selection probability of each spectral feature, an initial population matrix containing (N) individuals is constructed. and the corresponding individual optimal position matrix ,in, This represents the continuous position value of the (i)th individual in the (j)th spectral feature dimension, which is used to characterize the selection tendency of the (i)th individual for the (j)th spectral feature; This represents the historical best position value of the (i)th individual in the (j)th spectral feature dimension; During initialization, set During the initial population generation process, the following constraint is simultaneously applied: each individual corresponds to at least two selected spectral features to avoid generating empty feature subsets or single feature subsets; The number of selected spectral features for each individual does not exceed the preset upper limit of the total number of spectral features, in order to ensure the dimensionality suppression effect; The preset upper limit of the ratio can be set according to the dimensionality of the initial feature set and the compression requirements of the target feature set, so that the initial individual retains a sufficient number of candidate spectral features.

[0072] If an individual generated based on the initial selection probability does not meet the condition of having at least two spectral features, then supplementary spectral features with higher importance scores are selected until the constraint is met.

[0073] After the initial population is constructed, the fitness of each individual is calculated. Before fitness evaluation, the continuous position vector of the individual is converted into a binary feature selection mask. For the position value of the (i)th individual in the (j)th spectral feature dimension... Binarization is performed according to the following rules: ; in, This represents the binary feature selection mask for the (i)th individual in the (j)th spectral feature dimension; when When, it means that the (i)th individual selects the (j)th spectral feature; when When, it means that the (i)th individual has its (j)th spectral feature removed; This represents the feature selection threshold.

[0074] The feature selection threshold can be set according to the distribution of continuous position values ​​in the 0 to 1 interval and the sparsity requirements of the target feature subset, and is used to control the retention strength when continuous position values ​​are converted into binary selection results.

[0075] Based on the binary feature selection mask, the corresponding spectral feature dimension is extracted from the initial feature set to form the feature subset corresponding to the current individual.

[0076] If the feature subset corresponding to the current individual is empty, or contains only one spectral feature, then the individual is determined to be an invalid solution, and its fitness value is set to infinity to prevent it from entering the subsequent optimal position update process. For valid individuals, a random forest regression model is trained using their corresponding feature subset, and the out-of-bag root mean square error is calculated. Prediction is performed using out-of-bag samples from the random forest that did not participate in the training of the corresponding decision tree. The root mean square error between the prediction result and the calibrated heat value is calculated, and then a fitness function that integrates prediction accuracy and feature size is constructed. ; Where (Fit) represents the fitness value of the current individual; Indicates the prediction accuracy weight; Indicates the feature size weight; This represents the out-of-bag root mean square error obtained after training a random forest using the feature subset corresponding to the current individual. (d) represents the number of spectral features selected by the current individual; (d) represents the total number of spectral features in the initial feature set.

[0077] The prediction accuracy weight and feature size weight can be set according to the emphasis relationship between the prediction accuracy constraint requirements and the dimensionality compression requirements; when it is necessary to prioritize the accuracy of the predicted calorific value, the prediction accuracy weight is increased; when it is necessary to further compress the number of spectral features, the feature size weight is increased.

[0078] The fitness value is composed of the error term and the dimension ratio term. The smaller the value, the less spectral features the current individual uses while maintaining a small prediction error.

[0079] After obtaining the fitness values ​​of each individual, the iterative optimization phase of ISABO is entered. ISABO is an improved subtractive average optimizer used to search for the optimal subset of features in a continuous feature selection space.

[0080] It should be explained that the improved subtractive averaging optimizer refers to a feature subset search method that introduces adaptive learning parameters and diversity perturbation terms on the basis of the subtractive averaging optimization mechanism. It is used to iteratively update and optimally select candidate feature combinations in a continuous feature selection space. In particular, by introducing a differential update term based on the global optimal position, the global worst position, and the population mean position, the position of the individual is guided and adjusted. Combined with the diversity perturbation term, the discreteness of the population distribution is maintained, thereby improving the stability and effectiveness of the feature subset optimization process while ensuring the search convergence ability.

[0081] For the position value of the i-th individual in the j-th spectral feature dimension in the t-th iteration, it is updated according to the following formula: ; in, This represents the position value of the (i)th individual in the (j)th spectral feature dimension at the (t)th iteration; This indicates the updated value at the corresponding position. This represents the position value of the currently globally optimal individual in the (j)th spectral feature dimension; This represents the position value of the current worst-performing individual in the (j)th spectral feature dimension; Let represent the mean position of the population in the j-th spectral feature dimension at the (t)-th iteration; , , Indicates the learning parameters; Let represent the diversity perturbation term in the (j)th spectral feature dimension at the (t)th iteration.

[0082] It needs to be explained that, A random scalar with values ​​in the range [0,1] is used to scale the population mean vector.

[0083] In learning parameters and Adaptively adjust according to the iteration progress. For example, let the current iteration progress be... ,(t) represents the current iteration number, Indicates the maximum number of iterations, then , , The variable is a uniformly random number within the interval ([0,1]). The diversity perturbation term is calculated using the following formula: ; in, Let represent the standard deviation of the population's position in the j-th spectral feature dimension at the (t)-th iteration; This represents a random number that follows a standard normal distribution.

[0084] The diversity perturbation term is used to maintain the diversity of the population across spectral feature dimensions in the early stages of iteration and gradually decays as the iteration progresses. After an individual's position is updated, if the position value exceeds the range ([0,1]), boundary pruning is performed to bring it back into the range ([0,1]).

[0085] After each iteration, the individual's best position and the global best position are updated. Specifically, the fitness value corresponding to the current individual position is compared with the fitness value corresponding to its historical best position. If the current fitness value is less than the historical best fitness value, then the current individual position is updated to the corresponding individual best position, and the historical best fitness value of that individual is updated simultaneously.

[0086] The historical best fitness value refers to the minimum fitness value obtained by the i-th individual in each iteration during the ISABO iterative optimization process. The position of the individual corresponding to this minimum value is taken as the historical best position of that individual.

[0087] Select the optimal position of the individual with the smallest fitness value as the new global optimal position, and record the corresponding global optimal fitness value.

[0088] When any of the termination conditions are met, the ISABO iterative optimization ends and the optimal feature subset is output. The termination conditions include: 1. The current iteration number reaches the preset maximum iteration number; 2. The change in the global optimal fitness value in 10 consecutive iterations is lower than the preset convergence change threshold; 3. The population diversity is lower than the preset threshold.

[0089] Population diversity is calculated by averaging the standard deviations of all spectral feature dimensions. A preset convergence change threshold can be set based on the fluctuation range of the global optimal fitness value over several generations, used to determine whether the current iteration has entered a stable convergence state; a preset population diversity threshold can be set based on the dispersion of the population's location distribution, used to determine whether the population has shown significant homogenization.

[0090] After termination, the global optimal position vector is converted into a target feature selection mask according to the aforementioned binarization rules, and the corresponding spectral feature dimensions are extracted from the initial feature set to obtain the target feature set. The target feature set is the final feature result, used for subsequent construction of the random forest regression model.

[0091] Therefore, in this embodiment, the target feature set is determined by the iterative optimization process of the ISABO algorithm, and the random forest feature importance only participates in the feature subset search process as initialization guidance information, and does not directly participate in the final feature selection decision. After obtaining the target feature set, a random forest regression model is constructed based on the synthetic training set and the target feature set; Specifically, the spectral features of each sample in the synthetic training set in the corresponding dimension of the target feature set are selected to form the target training feature matrix, and the corresponding calibrated heat value is selected as the training label to train the random forest regression model; The Bayesian optimization method was used to search for the key parameters of the random forest regression model. The key parameters included: number of decision trees, maximum depth, minimum number of leaf node samples, and maximum number of features.

[0092] In one embodiment, the parameter search range can be set as follows: based on the sample size of the synthetic training set, the number of dimensions of the target feature set, and the training complexity requirements of the random forest regression model, a corresponding search interval is set for each key parameter. The Bayesian optimization process uses cross-validation determination coefficient as the evaluation index, and sets a preset number of random initial search points and a preset number of parameter iteration search processes to select the parameter combination with the best score as the final parameter configuration of the random forest regression model.

[0093] The number of random initial search points and the number of parameter iterations can be set according to the size of the parameter space and the required accuracy of parameter optimization.

[0094] After the random forest regression model is trained, calorific value mapping is performed on the biomass samples to be tested.

[0095] Specifically, the spectral features of the biomass sample to be tested are selected, and the corresponding spectral features are extracted in the order of the dimensions corresponding to the target feature set to form the feature vector to be tested; the feature vector to be tested is input into the trained random forest regression model to output the calorific value of the biomass sample to be tested.

[0096] Through the above processing, the effectiveness of the feature search starting point is improved by using the importance-guided initialization strategy of random forest. Then, ISABO is used to iteratively optimize the subset of spectral features. Based on the target feature set, a random forest regression model is constructed to realize the quantitative mapping of the calorific value of the biomass sample to be tested, thereby completing the complete closed-loop processing from labeled spectral features to target features and then to calorific value results.

[0097] Finally, it should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0098] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0099] In this document, the singular forms “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that terms such as “comprising / including” or “having” specify the presence of the stated features, integrals, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, integrals, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0100] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0101] The above description of the disclosed embodiments will enable those skilled in the art to make or use various modifications to these embodiments. It will be readily apparent to those skilled in the art that the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy, characterized in that: Includes the following steps: Step S1: After slicing the biomass sample to be tested, a sheet sample to be tested is generated. The calibrated calorific value of the sheet sample to be tested is collected. The spectral characteristics of the sheet sample to be tested are obtained using a laser-induced device and spectral data is constructed. Step S2: Perform prediction processing on the sheet-like sample to be tested, obtain the predicted calorific value of the sheet-like sample to be tested, calculate the fitting score by combining the calibrated calorific value, and output the synthetic training set after interpolating and merging the spectral data of the sheet-like sample to be tested according to the fitting score. Step S3: After performing layer-by-layer compression of the spectral features of the synthetic training set using the attention mechanism, propagation calculation is performed on the synthetic training set to generate gradient response values ​​of the spectral features. The spectral features are then filtered and labeled based on the gradient response values. Step S4: Dimensionality suppression of the labeled spectral features is performed using random forest and the target features are output. The training set and target features are combined to construct a regression model and map the calorific value of the biomass sample to be tested.

2. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 1, characterized in that: In step S1, the biomass sample to be tested is sliced. The biomass sample to be tested refers to the original sample derived from crop straw, wood or other organic biomass materials. Slicing process refers to placing the biomass sample to be tested in a dry environment and treating it at a constant temperature. Then, the dried sample is crushed into granular powder by mechanical grinding equipment. The powder particle size is then uniformly screened by a sieving device. The screened powder is placed in a pressing mold and pressed into shape under a set pressure condition to form several slices of the sample to be tested. The calibrated calorific value of the sheet sample to be tested is collected. The calibrated calorific value refers to the heat released by the complete combustion of a unit mass, as measured by a standard calorific value measuring device.

3. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 1, characterized in that: In step S1, the spectral characteristics of the sheet-like sample to be tested are acquired using a laser induction device. The laser induction device is a detection system that can emit high-energy pulsed laser and excite the sheet-like sample to be tested to generate plasma luminescence, including a laser emitting unit, an optical focusing unit, a sample positioning unit, and a spectral acquisition unit. The laser pulse from the laser emitting unit irradiates the surface of the sheet sample under test, causing the local area of ​​the sheet sample to heat up instantaneously and form a high-temperature plasma. The plasma releases emitted light signals of different wavelengths during the cooling process. The emitted light signal is decomposed into different wavelength components by the spectral acquisition unit, and the light intensity value corresponding to each wavelength is recorded, thereby forming spectral features and constructing spectral data; Spectral characteristics refer to the light intensity values ​​measured at different wavelength positions, while spectral data refers to the set of values ​​formed by arranging the spectral characteristics at each wavelength position in a fixed order.

4. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 3, characterized in that: In step S2, the spectral data is input into the prediction model for prediction processing. The prediction model is a regression calculation model used to establish the mapping relationship between spectral features and calibrated calorific values. Predictive processing refers to using an established predictive model to calculate the input spectral data and output the corresponding predicted calorific value. The predicted calorific value is the estimated calorific value calculated by the predictive model based on spectral features. The absolute difference between the predicted calorific value and the calibrated calorific value is calculated to obtain the fitting residual of the sheet sample to be tested. The Max-Min normalization method is used to normalize the fitting residual to obtain the fitting score. Based on the fitting score, all the sheet-like samples to be tested are sorted, and the sheet-like samples to be tested are divided into a set of difficult-to-fit samples and a set of easy-to-fit samples according to the preset scoring threshold.

5. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 4, characterized in that: In step S2, differential interpolation merging is performed on the hard-to-fit sample set and the easy-to-fit sample set. The interpolation merging process includes interpolation processing and merging processing. Interpolation refers to generating new feature samples and their corresponding calibrated calorific values ​​by constructing a linear combination in the spectral feature space of a known sample, based on the continuous relationship between neighboring samples. Merging refers to the process of incorporating the original test samples and new samples into the same dataset and rearranging the order of the samples to form a synthetic training set containing both the original information and the augmented information.

6. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 5, characterized in that: In step S3, the spectral features in the synthetic training set are standardized to obtain standardized spectral features; The standardized spectral features are input into the encoder layer of the attention neural network, and the features are transformed layer by layer to output the encoded feature matrix. Based on the encoded output feature matrix, the intermediate mapping value corresponding to each feature dimension of each sample is calculated, and normalization processing is performed on all feature dimensions of the same sample to obtain the attention weights corresponding to each feature dimension. The attention weights are matched dimension-by-dimensionally with the feature values ​​at the corresponding positions in the encoded output feature matrix to obtain the weighted feature result.

7. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 6, characterized in that: In step S3, regression prediction is performed based on the weighted feature results to obtain the predicted heat value; with the predicted heat value as the output, backpropagation processing is performed on the encoded output feature matrix to obtain the gradient response value of each sample in each feature dimension; The gradient response values ​​of the same feature dimension across all samples are processed by absolute value and averaged to obtain the gradient score of the corresponding feature dimension. All gradient scores are normalized and sorted. The screening boundary is determined according to the preset percentile position. The feature dimensions that meet the screening conditions are marked as labeled spectral features, and their feature index positions in the encoded output feature matrix are recorded. The labeled spectral feature set is then output.

8. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 7, characterized in that: In step S4, the feature index positions in the marked spectral feature set are read, and the values ​​of the corresponding feature dimensions under each sample are extracted from the encoded output feature matrix to form an initial feature set; Training data is constructed based on the initial feature set, and the calibrated heat value corresponding to each sample is used as the output label to train the random forest regression model. During model training, the contribution of each feature to the reduction of prediction error when participating in node partitioning is statistically analyzed to obtain the importance score of each feature; The features in the initial feature set are filtered based on the importance scores of each feature, and features with importance scores below the filtering boundary are removed to obtain the target feature set.

9. The method for determining the high calorific value of biomass based on laser-induced breakdown spectroscopy according to claim 8, characterized in that: In step S4, corresponding feature data are extracted from the synthetic training set based on the target feature set, a target training feature matrix is ​​constructed, and the calibration heat value corresponding to each sample is used as the training label to train the random forest regression model and obtain the final regression model. The spectral characteristics of the biomass sample to be tested are input into the final regression model, and the calorific value of the biomass sample to be tested is output.

Images