CN-121983176-A - Heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method and system

CN121983176ACN 121983176 ACN121983176 ACN 121983176ACN-121983176-A

Abstract

The invention provides a heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method and system, which comprise the steps of establishing a parameter-wave band mapping relation, fusing and outputting mixed coding sequence characteristics through heterogeneous characteristic attention, inputting the mixed coding sequence characteristics and coal quality parameter coding characteristics into a spectrum-coal quality joint learning module, outputting high-quality potential representation of spectrum characteristics and coal quality parameters, using the coal quality parameters as inversion prediction results when inversion is carried out on coal quality to be inverted, fusing the high-quality potential representation of the spectrum characteristics of the existing coal quality and the coded real coal quality parameters as history spectrum-coal quality potential characteristics when the new coal quality appears, inputting the history spectrum-coal quality potential characteristics and the mixed coding sequence characteristics of the new coal quality into an incremental characteristic learning module together, and generating coal quality parameter inversion prediction results of history knowledge and new characteristics through experience transfer and gate-controlled feedback cross-coal quality learning. The invention can invert the coal quality parameters.

Inventors

YUAN LI
ZHOU HEBIN
LI JIANGYUN
WANG HONG
GUO ANTING
ZHANG TIANXIANG
Zhuang Peixian

Assignees

北京科技大学

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (10)

1. The heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method is characterized by comprising the following steps of: S1, acquiring two complementary multi-source data of laser-induced breakdown spectroscopy LIBS and near infrared spectroscopy NIR of the existing coal, forming a data set by the multi-source data, and dividing a training set and a verification set; S2, respectively inputting multi-source data in the training set into a spectral domain feature mapping module, identifying sensitive wave band intervals of each coal quality parameter in LIBS and NIR, and establishing a parameter-wave band mapping relation to obtain LIBS spectral domain features and NIR spectral domain features; S3, mapping the LIBS spectral domain features into parameter specific sub-bands and intensity values thereof, inputting the parameter specific sub-bands and the intensity values thereof into a heterogeneous encoder cooperative feature extraction module, mapping the NIR spectral domain features into parameter specific sub-bands and absorption values thereof, inputting the NIR spectral domain features into the heterogeneous encoder cooperative feature extraction module, respectively generating LIBS sequence features and NIR sequence features, inputting the LIBS sequence features and the NIR sequence features into a heterogeneous feature attention fusion module, and outputting mixed coding sequence features fused with double-source spectral information; S4, inputting the mixed coding sequence characteristics and the coal quality parameter coding characteristics into a spectrum-coal quality joint learning module, and outputting high-quality potential representations subjected to alignment constraint and reconstruction verification through a bidirectional coding-decoding mechanism and spectrum-coal quality bidirectional association constraint, wherein the high-quality potential representations comprise spectrum characteristic high-quality potential representations and coal quality parameters; S5, when inversion is carried out on the coal quality to be inverted, if no new coal quality appears, the coal quality parameters output by the spectrum-coal quality joint learning module are used as inversion prediction results, when the new coal quality appears, the spectrum characteristics high-quality potential representation of the existing coal quality output by the spectrum-coal quality joint learning module and the real coal quality parameters of the existing coal quality after MLP characteristic coding are fused, the result is used as a history spectrum-coal quality potential characteristic, and the history code sequence characteristics output by the heterogeneous characteristic attention fusion module of the new coal quality are input into the incremental characteristic learning module together, and coal quality parameter inversion prediction results considering both history knowledge and new characteristics are generated through experience transmission and door control feedback of cross-coal quality learning.
2. The method according to claim 1, wherein the processing of the spectral domain feature mapping module specifically comprises: Full-band data of LIBS and NIR are input into a one-dimensional convolutional neural network, and multi-scale features of spectrums and correlations among wavelengths are captured by utilizing convolution kernels of different scales through layer-by-layer feature extraction of Conv1, conv2 and Conv 3; Then, the features output by the one-dimensional convolutional neural network are used as the input of a self-attention mechanism module, the self-attention mechanism module maps the features output by the one-dimensional convolutional neural network into a matrix Q, K, V, and the attention weighted enhancement features are calculated to represent K 'and V' by calculating the dependency relationship between different wavelengths in the attention weight learning spectrum sequence; Then, inputting coal parameters as priori knowledge into a multi-layer perceptron MLP to perform semantic coding, and converting discrete parameter labels into continuous high-dimensional vector representations Q'; And then, the cross attention module generates an attention weight matrix by calculating a dot product similarity matrix between Q 'and K', wherein each element in the attention weight matrix quantitatively reflects the sensitivity and importance score of the corresponding wavelength to the specific coal parameters, identifies the specific sensitive wave band of each coal parameter through the attention weight matrix, establishes a parameter-wave band mapping relation, obtains LIBS spectral domain characteristics and NIR spectral domain characteristics, and each sub-wave band of the parameter-wave band mapping relation is used for extracting the characteristics of the corresponding coal parameters, so that the mixed interference of the spectral information of different physical mechanisms is avoided, and the accurate characteristic extraction of parameter guidance is realized.
3. The method of claim 1, wherein the LIBS spectral domain features include a LIBS with atomic emission lines, a full moisture corresponding to H, O atomic emission lines, an ash corresponding to Si, al atomic emission lines, a volatile corresponding to C, H atomic emission lines, a full sulfur corresponding to S atomic emission lines, and a calorific value corresponding to C atomic feature emission lines; The NIRS spectrum domain features that through the absorption characteristic of the molecular groups, the NIRS leads the total moisture to correspond to the absorption spectrum section of the hydroxyl groups, the ash to correspond to the mineral oxide, the volatile to correspond to the hydrocarbon, the total sulfur to correspond to the sulfur-hydrogen wave band and the calorific value to correspond to the carbon-hydrogen wave band.
4. The method according to claim 1, wherein the heterogeneous encoder cooperates with a processing procedure of the feature extraction module, specifically comprising: The parameter specificity sub-band of LIBS spectrum and intensity value thereof, and the parameter specificity sub-band of NIR spectrum and absorption rate thereof are respectively used as input sequence data, and the two input sequence data are firstly processed by a word embedding layer and then are respectively input into two parallel coding paths: The converter encoder path is composed of 6 layers CSwin modules, each CSwin module comprises a multi-head self-attention mechanism, a feedforward neural network and a residual error connection structure, global association weights among different band characteristics are adaptively learned through the multi-head self-attention mechanism, spatial position information of a spectrum sequence is kept by utilizing position codes, and long-distance dependency relation and global sequence characteristics of the spectrum data are captured; The CNN encoder path consists of 6 layers MobileNet modules, so that the computational complexity is greatly reduced while the feature extraction effect is ensured, and local spectrum texture features with different frequencies are captured through multi-scale convolution kernel combination, so that local modes and detail information in spectrum data are effectively extracted; Then the fusion module passes 、、 The three branches process features from the transducer and CNN paths respectively, And Derived from CNN encoder features, computing the correlation weights between different path features using a cross-attention mechanism yields a cross-attention output that is matched to the non-LN transformed multi-scale features provided by CNN and the transform encoder The weighted fusion realizes the effective fusion of the global sequence features and the local texture features, the features after each layer of fusion are used as the multi-scale feature input of the fusion module of the next layer to form a progressive fusion mechanism of the hierarchy, and the fusion module of the last layer is used for the multi-scale feature input of the fusion module of the next layer 、、 The final sequence features were obtained from the fransformer encoder sixth layer CSwin and the CNN encoder sixth layer MobileNet by 1 x 1 convolution.
5. The method according to claim 1, wherein the processing of the heterologous feature attention fusion module specifically comprises: The LIBS sequence features and the NIR sequence features are input into a full-connection layer at the same time by adopting a gating fusion mechanism, a gating weight G between 0 and 1 is output through a Sigmoid activation function, the G represents the contribution duty ratio of the LIBS sequence features in a fusion result, and 1-G is the contribution duty ratio of the NIR sequence features; Then, weighting LIBS sequence features by G, weighting NIRS sequence features by 1-G, adding the two to obtain mixed coding sequence features fused with double-source spectrum information, and through the gating mechanism, the model can adaptively adjust the fusion proportion of LIBS and NIRS features according to the spectrum characteristics of different samples, the gating weight G automatically tends to be 1 for samples with more obvious LIBS spectrum information, and the dynamic weighting strategy effectively integrates complementary information of double-source heterogeneous spectrum to form the mixed coding sequence features fused with double-source heterogeneous spectrum information, wherein the mixed coding sequence features not only maintain the global dependency relation extracted by a transform encoder, but also fuse the local feature modes captured by a CNN encoder, and simultaneously realize the effective fusion of LIBS and NIRS heterogeneous spectrum, thereby providing high-quality feature input for the subsequent spectrum-coal parameter joint learning.
6. The method according to claim 1, wherein the processing procedure of the spectrum-coal quality joint learning module specifically comprises: Hybrid coding sequence features Coal quality parameter coding features Mapping two heterogeneous data into a unified representation space by a latent spatial mapping module comprising a spectral encoder and a parametric encoder, the method comprising the steps of Generating a spectral representation by the spectral encoder mapping The said Generating a coal representation by the parameter encoder mapping The two representations realize unified modeling in potential space with the same dimension; Then the said And The input spectrum-coal quality alignment constraint module implements consistency constraint in potential space through a double-reference comparison learning strategy, and ensures that the double-reference comparison learning loss function designed by the formula (1) is utilized to ensure that the same coal sample comes from And Maintaining high similarity in potential space while maximizing representation differences between different coal samples, achieving accurate alignment of matched sample pairs and efficient discrimination of non-matched sample pairs for the first A sample of coal whose spectrum represents And coal quality representation To form positive sample pairs Coal quality representation with other coal samples , The negative sample pair is formed, and the double reference pair learning loss function is shown as a formula (1): (1) Wherein the method comprises the steps of Representing the similarity score for a positive sample pair, Represent the first Cosine similarity between the spectrum representation of each coal sample and the coal quality representation; represent the first The spectrum of each coal sample represents the sum of similarity scores with all coal sample coal quality representations, including 1 positive sample pair and N-1 negative sample pair, through maximizing the duty ratio of the positive sample pair similarity in all sample pairs, make the spectrum-coal quality representation of the same coal sample the nearest in potential space, and the representing distance among different coal samples is pushed away, when the positive sample pair similarity is far higher than the negative sample pair, the loss value tends to 0; finally, the spectrum decoder outputs a potential representation from the spectrum-coal alignment constraint module Mid-reconstruction hybrid coded spectral sequence features, potential representations output by a coal decoder from a spectral-coal alignment constraint module The reconstruction constraint comprises a spectrum reconstruction loss of a formula (2) and a reconstruction loss of a coal parameter of a formula (3), the reconstruction constraint requires that a model keeps complete original information in a potential space, the coal parameter can be accurately restored, the rationality of an inversion result can be verified, and the spectrum reconstruction loss is shown as the formula (2): (2) Spectral reconstruction loss guarantees that a spectral decoder can be derived from the Accurately recovering the original mixed coding spectrum characteristics; the coal quality parameter characteristic coding reconstruction loss is shown in the formula (3): (3) the design of the coal quality parameter reconstruction penalty ensures that the coal quality decoder can reconstruct the data from the data Accurately reconstructing the original coal quality parameter value; Total reconstruction loss By minimizing the reconstruction errors, the encoder is forced to learn the potential representation that includes the complete original information.
7. The method according to claim 1, wherein the incremental feature learning module processes, in particular, comprise: the balance between historical knowledge maintenance and new feature learning is realized through parallel processing of an experience transfer network and an adaptive learning network, wherein the experience transfer network carries out depth coding on the potential features of the historical spectrum and coal through a four-layer depth structure, namely a hidden layer I, a hidden layer II, an attention layer, an LN+Softmax, the adaptive learning network adopts a residual connection architecture, namely a residual block 1, an self-attention mechanism, a residual block 2, an LN+Softmax, and the mixed coding sequence features of new coal are processed; And then dynamically adjusting the fusion weight according to the similarity of the new coal data and the historical data, generating a coal quality parameter inversion prediction result taking both historical knowledge and new characteristics through a characteristic regression head, introducing gating feedback control, automatically triggering model parameter updating of the self-adaptive learning network when the error exceeds a preset threshold value by monitoring the error of inversion predicted coal quality parameters and real values in real time, intelligently adjusting the updating strength according to the error size and the confidence coefficient, feeding back an updating signal to the self-adaptive learning network for parameter updating to form a closed-loop self-adaptive optimization mechanism of 'characteristic regression head, inversion prediction, error evaluation, gating judgment, parameter updating and network adjustment', and gating off when the error is within an acceptable range, avoiding model overadjustment, ensuring stability and realizing continuous learning and self-adaptive optimization of the model.
8. A heterogeneous dual-source collaborative and empirical self-evolving coal quality parameter inversion system, the system comprising: the acquisition and division module is used for acquiring two complementary multi-source data of laser-induced breakdown spectroscopy LIBS and near infrared spectroscopy NIR of the existing coal, forming a data set by the multi-source data, and dividing a training set and a verification set; The identification building module is used for respectively inputting the multi-source data in the training set into the spectral domain feature mapping module, identifying sensitive wave band intervals of each coal quality parameter in LIBS and NIR, and building a parameter-wave band mapping relation to obtain LIBS spectral domain features and NIR spectral domain features; The extraction fusion module is used for mapping the LIBS spectral domain characteristics into parameter specific sub-bands and intensity values thereof, inputting the parameter specific sub-bands and the absorption rate thereof into the heterogeneous encoder cooperative characteristic extraction module, inputting the NIR spectral domain characteristics into the heterogeneous encoder cooperative characteristic extraction module, respectively generating LIBS sequence characteristics and NIR sequence characteristics, inputting the LIBS sequence characteristics and the NIR sequence characteristics into the heterogeneous characteristic attention fusion module, and outputting mixed coding sequence characteristics fused with double-source spectral information; the combined learning module is used for inputting the mixed coding sequence characteristics and the coal quality parameter coding characteristics into the spectrum-coal quality combined learning module, and outputting high-quality potential representation which is subjected to alignment constraint and reconstruction verification through a bidirectional coding-decoding mechanism and a spectrum-coal quality bidirectional association constraint, wherein the high-quality potential representation comprises spectrum characteristic high-quality potential representation and coal quality parameters; And when the new coal type appears, the spectrum characteristic high-quality potential representation of the existing coal type output by the spectrum-coal type joint learning module and the real coal type parameter of the existing coal type after MLP characteristic coding are fused to be used as a history spectrum-coal type potential characteristic, and the result and the mixed coding sequence characteristic output by the heterogeneous characteristic attention fusion module of the new coal type are input into an incremental characteristic learning module together, and the coal type parameter inversion prediction result considering both the history knowledge and the new characteristic is generated through the experience transfer and the cross-coal type learning fed back by the door control.
9. An electronic device comprising a processor and a memory having at least one instruction stored therein, wherein the at least one instruction is loaded and executed by the processor to implement the heterogeneous dual source collaborative and empirical self-evolving coal quality parameter inversion method of any of claims 1-7.
10. A computer readable storage medium having stored therein at least one instruction, wherein the at least one instruction is loaded and executed by a processor to implement the heterogeneous dual source collaborative and empirical self-evolving coal quality parameter inversion method of any of claims 1-7.

Description

Heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method and system Technical Field The invention belongs to the technical field of coal quality parameter inversion, and particularly relates to a heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method and system. Background The coal is used as an important fossil energy source, and the accurate detection of parameters of the coal has important significance for optimizing combustion efficiency, controlling environmental pollution and managing industrial production. The traditional coal quality detection mainly depends on a chemical analysis method, has the problems of long detection period, complex sample treatment, poor real-time performance and the like, and is difficult to meet the requirements of modern industry on quick, accurate and online detection. Disclosure of Invention In order to solve the technical problems in the prior art, the invention provides a heterogeneous dual-source cooperation and experience self-evolution coal quality parameter inversion method and system, wherein the technical scheme is as follows: in one aspect, a heterogeneous dual-source collaborative and experience self-evolving coal quality parameter inversion method is provided, and the method comprises the following steps: S1, acquiring two complementary multi-source data of laser-induced breakdown spectroscopy LIBS and near infrared spectroscopy NIR of the existing coal, forming a data set by the multi-source data, and dividing a training set and a verification set; S2, respectively inputting multi-source data in the training set into a spectral domain feature mapping module, identifying sensitive wave band intervals of each coal quality parameter in LIBS and NIR, and establishing a parameter-wave band mapping relation to obtain LIBS spectral domain features and NIR spectral domain features; S3, mapping the LIBS spectral domain features into parameter specific sub-bands and intensity values thereof, inputting the parameter specific sub-bands and the intensity values thereof into a heterogeneous encoder cooperative feature extraction module, mapping the NIR spectral domain features into parameter specific sub-bands and absorption values thereof, inputting the NIR spectral domain features into the heterogeneous encoder cooperative feature extraction module, respectively generating LIBS sequence features and NIR sequence features, inputting the LIBS sequence features and the NIR sequence features into a heterogeneous feature attention fusion module, and outputting mixed coding sequence features fused with double-source spectral information; S4, inputting the mixed coding sequence characteristics and the coal quality parameter coding characteristics into a spectrum-coal quality joint learning module, and outputting high-quality potential representations subjected to alignment constraint and reconstruction verification through a bidirectional coding-decoding mechanism and spectrum-coal quality bidirectional association constraint, wherein the high-quality potential representations comprise spectrum characteristic high-quality potential representations and coal quality parameters; S5, when inversion is carried out on the coal quality to be inverted, if no new coal quality appears, the coal quality parameters output by the spectrum-coal quality joint learning module are used as inversion prediction results, when the new coal quality appears, the spectrum characteristics high-quality potential representation of the existing coal quality output by the spectrum-coal quality joint learning module and the real coal quality parameters of the existing coal quality after MLP characteristic coding are fused, the result is used as a history spectrum-coal quality potential characteristic, and the history code sequence characteristics output by the heterogeneous characteristic attention fusion module of the new coal quality are input into the incremental characteristic learning module together, and coal quality parameter inversion prediction results considering both history knowledge and new characteristics are generated through experience transmission and door control feedback of cross-coal quality learning. In another aspect, a heterogeneous dual-source collaborative and empirical self-evolving coal quality parameter inversion system is provided, the system comprising: the acquisition and division module is used for acquiring two complementary multi-source data of laser-induced breakdown spectroscopy LIBS and near infrared spectroscopy NIR of the existing coal, forming a data set by the multi-source data, and dividing a training set and a verification set; The identification building module is used for respectively inputting the multi-source data in the training set into the spectral domain feature mapping module, identifying sensitive wave band intervals of each coal quality parameter in LI