Heterogeneous double-source cooperation and empirical self-evolution coal quality parameter inversion method and system

Abstract

The invention provides a coal quality parameter inversion method and system based on heterogeneous double-source collaboration and empirical self-evolution, and the method comprises the steps: building a parameter-waveband mapping relation, and outputting a hybrid coding sequence feature through heterogeneous feature attention fusion; inputting the mixed coding sequence features and the coal quality parameter coding features into a spectrum-coal quality combined learning module, and outputting spectral feature high-quality potential representation and coal quality parameters; when the to-be-inverted coal quality is inverted, if no new coal type appears, using the coal quality parameter as an inversion prediction result; when a new coal type appears, spectral feature high-quality potential representation of existing coal quality and encoded real coal quality parameters are fused to serve as historical spectrum-coal quality potential features, the historical spectrum-coal quality potential features and mixed encoding sequence features of the new coal type are input into an increment feature learning module together, and through cross-coal-type learning of experience transmission and gating feedback, the coal quality of the new coal type is obtained. And generating a coal quality parameter inversion prediction result considering historical knowledge and new features. The method can be used for coal quality parameter inversion.

Description

Technical Field

[0001] This invention belongs to the field of coal quality parameter inversion technology, and specifically refers to a heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion method and system. Background Technology

[0002] As an important fossil energy source, accurate detection of coal parameters is of great significance for optimizing combustion efficiency, controlling environmental pollution, and managing industrial production.

[0003] Traditional coal quality testing mainly relies on chemical analysis methods, which have problems such as long testing cycles, complex sample processing, and poor real-time performance, making it difficult to meet the needs of modern industry for rapid, accurate, and online testing. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a heterogeneous dual-source collaborative and empirical self-evolutionary coal quality parameter inversion method and system, the technical solution of which is as follows: On the one hand, a method for inverting coal quality parameters based on heterogeneous dual-source synergy and empirical self-evolution is provided, which includes: S1. Collect two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The multi-source data constitutes a dataset, which is then divided into a training set and a validation set. S2. Input the multi-source data in the training set into the spectral domain feature mapping module respectively, identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. S3. Map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. Map the NIR spectral domain features to parameter-specific sub-bands and their absorption rates, and input them into the heterogeneous encoder collaborative feature extraction module to generate LIBS sequence features and NIR sequence features respectively. Then, input the LIBS sequence features and NIR sequence features into the heterogeneous feature attention fusion module to output the hybrid coding sequence features that fuse dual-source spectral information. S4. Input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module. Through the bidirectional encoding-decoding mechanism and the bidirectional correlation constraint between the spectral and coal quality, output a high-quality latent representation that has been verified by alignment constraint and reconstruction, including a high-quality latent representation of spectral features and coal quality parameters. S5. When inverting the coal quality to be inverted, if no new coal type appears, the coal quality parameters output by the spectral-coal quality joint learning module are used as the inversion prediction result. When a new coal type appears, the high-quality potential representation of the spectral features output by the spectral-coal quality joint learning module and the real coal quality parameters after MLP feature encoding of the existing coal quality are fused together as historical spectral-coal quality potential features. These features are then input together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

[0005] On the other hand, a heterogeneous dual-source collaborative and empirically self-evolving coal quality parameter inversion system is provided, the system comprising: The data acquisition and partitioning module is used to acquire two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The dataset is composed of the multi-source data and is divided into a training set and a validation set. The identification and establishment module is used to input the multi-source data in the training set into the spectral domain feature mapping module to identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. The extraction and fusion module is used to map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. The NIR spectral domain features are also mapped to parameter-specific sub-bands and their absorption rates, and input into the heterogeneous encoder collaborative feature extraction module to generate LIBS sequence features and NIR sequence features respectively. Then, the LIBS sequence features and NIR sequence features are input into the heterogeneous feature attention fusion module to output a hybrid coded sequence feature that fuses dual-source spectral information. The joint learning module is used to input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module. Through a bidirectional encoding-decoding mechanism and bidirectional correlation constraints between the spectral and coal quality, it outputs a high-quality latent representation that has been verified by alignment constraints and reconstruction, including a high-quality latent representation of spectral features and coal quality parameters. The inversion prediction module is used to use the coal quality parameters output by the spectral-coal quality joint learning module as the inversion prediction result when no new coal type appears during the inversion of coal quality to be inverted. When a new coal type appears, the high-quality potential representation of the spectral features output by the spectral-coal quality joint learning module and the real coal quality parameters after MLP feature encoding of the existing coal quality are fused together as historical spectral-coal quality potential features. These features are then input together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

[0006] On the other hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to realize the above-mentioned heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion method.

[0007] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to realize the above-mentioned heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion method.

[0008] The beneficial effects of the technical solution provided by this invention include at least the following: This invention proposes a heterogeneous dual-source collaborative and empirically self-evolving coal quality spectral inversion method and system, which solves the problems of low feature extraction efficiency, insufficient physical meaning, and poor cross-coal adaptability in traditional coal quality detection methods. First, by combining parameter-band mapping of spectral domain feature mapping, a heterogeneous encoder collaborative feature extraction module, and a heterogeneous attention fusion module to generate hybrid coded sequence features, the problem of different sensitivity of coal quality parameters to spectral bands is solved, achieving efficient coal quality inversion. Second, through joint representation learning of spectra and coal quality parameters, the problem of unidirectional mapping between spectral features and coal quality parameters is solved, ensuring bidirectional correlation between them and improving the physical rationality of the inversion process and the generalization ability of the model. Finally, by adopting an empirical transfer and gated feedback cross-coal type learning mechanism, the problem of static models being unable to adapt to changes in the spectral characteristics of new coal types is solved. Through knowledge distillation and incremental update algorithms, continuous optimization and cross-coal type adaptability of the model are ensured. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a heterogeneous dual-source synergy and empirical self-evolution coal quality parameter inversion method provided by an embodiment of the present invention; Figure 2 This is a general block diagram of a heterogeneous dual-source synergistic and empirical self-evolution coal quality parameter inversion method provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the processing procedure of the spectral domain feature mapping module provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of hybrid coded sequence feature generation provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the processing procedure of the heterogeneous encoder collaborative feature extraction module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the heterogeneous feature attention module processing procedure provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the processing procedure of the spectrum-coal quality joint learning module provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of cross-coal type learning based on experience transfer and gating feedback provided in an embodiment of the present invention; Figure 9 This is a block diagram of a heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion system provided by an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0012] This invention proposes a heterogeneous dual-source collaborative and empirical self-evolutionary coal quality spectral inversion method, which constructs an end-to-end coal quality inversion algorithm to realize the inversion prediction of coal quality parameters from multi-source spectral data. It aims to solve the key problems in existing coal quality detection methods, such as low feature extraction efficiency, lack of physical meaning, and poor cross-coal type adaptability. It includes three main processes: feature extraction, heterogeneous encoder collaborative feature extraction and heterogeneous feature attention fusion, spectral-coal quality joint learning, and cross-coal type learning based on empirical transfer and gating feedback.

[0013] This invention provides a method for inverting coal quality parameters using heterogeneous dual-source synergy and empirical self-evolution. This method can be implemented by an electronic device, which can be a terminal or a server. Figure 1 The flowchart of this method is shown below. Figure 2 The diagram shown is an overall block diagram of the method. The processing flow may include the following steps:

[0014] S1. Collect two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The multi-source data constitutes a dataset, which is then divided into a training set and a validation set. Laser-induced breakdown spectroscopy (LIBS) excites coal samples with high-power laser pulses to generate plasma, detecting the characteristic spectral lines emitted by the plasma to provide compositional information of key elements such as carbon, hydrogen, nitrogen, and sulfur in the coal sample. It exhibits high sensitivity and rapid detection capability for elemental content. Near-infrared spectroscopy (NIR), on the other hand, detects the absorption characteristics of near-infrared light in coal samples to obtain information on molecular vibrations and functional groups, and is particularly sensitive to parameters such as moisture content, organic matter content, and chemical bond structure. The two spectral data are significantly complementary in terms of physical mechanisms and information dimensions, providing a fundamental guarantee for comprehensive and accurate extraction of coal quality characteristics.

[0015] S2. Input the multi-source data in the training set into the spectral domain feature mapping module respectively, identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. To address the issues of traditional full-band spectral processing methods failing to consider the differences in sensitivity of various coal quality parameters across different wavelength ranges and resulting in low feature extraction efficiency due to the mixing of spectral information with completely different physical mechanisms, this invention performs spectral domain feature mapping on five key coal quality parameters: total moisture, ash, volatile matter, total sulfur, and calorific value. The spectral domain feature mapping module identifies the sensitive band ranges of each coal quality parameter in laser-induced breakdown spectroscopy and near-infrared spectroscopy, establishes parameter-band mapping relationships, and divides the full-band spectral data into multiple parameter-specific sub-bands.

[0016] Optionally, such as Figure 3 As shown, the processing procedure of the spectral domain feature mapping module specifically includes: Full-band data from LIBS and NIR are input into a one-dimensional convolutional neural network. Through layer-by-layer feature extraction using three convolutional layers (Conv1, Conv2, and Conv3), multi-scale features of the spectrum and correlations between wavelengths are captured using convolutional kernels of different scales. Then, the features output by the one-dimensional convolutional neural network are used as input to the self-attention mechanism module. The self-attention mechanism module maps the features output by the one-dimensional convolutional neural network into matrices Q, K, and V. By calculating the attention weights, it learns the dependency between different wavelengths within the spectral sequence and calculates the attention-weighted enhanced feature representations K' and V'. Then, coal quality parameters (total moisture, ash, volatile matter, total sulfur, and calorific value) are used as prior knowledge and input into a multilayer perceptron (MLP) for semantic encoding, converting discrete parameter labels into continuous high-dimensional vector representations Q'. Then, the cross-attention module generates an attention weight matrix by calculating the dot product similarity matrix between Q' and K'. Each element in the attention weight matrix quantitatively reflects the sensitivity and importance score of the corresponding wavelength to a specific coal quality parameter. The attention weight matrix identifies the specific sensitive bands of each coal quality parameter, establishes a parameter-band mapping relationship, and obtains LIBS spectral domain features and NIR spectral domain features. Each sub-band of the parameter-band mapping relationship is used for feature extraction of the corresponding coal quality parameter, avoiding the mixed interference of spectral information from different physical mechanisms, and achieving accurate feature extraction guided by parameters.

[0017] Optionally, the LIBS spectral domain features include: LIBS uses atomic emission lines to make total water content correspond to H and O atomic emission lines, ash content correspond to Si and Al atomic emission lines, volatile matter correspond to C and H atomic emission lines, total sulfur content correspond to S atomic emission lines, and calorific value correspond to C atomic characteristic emission lines. The NIR spectral characteristics include: NIRS uses the absorption characteristics of molecular groups to make total water correspond to the absorption band of hydroxyl groups, ash correspond to mineral oxides, volatile matter correspond to hydrocarbons, total sulfur correspond to the sulfur-hydrogen band, and calorific value correspond to the carbon-hydrogen band.

[0018] S3. Map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. Map the NIR spectral domain features to parameter-specific sub-bands and their absorption rates, and input them into the heterogeneous encoder collaborative feature extraction module as well, generating LIBS sequence features and NIR sequence features respectively. Then, input the LIBS sequence features and NIR sequence features into the heterogeneous feature attention fusion module, and output the hybrid coded sequence features that fuse the dual-source spectral information, such as... Figure 4 As shown; Optionally, such as Figure 5 As shown, the processing procedure of the heterogeneous encoder collaborative feature extraction module specifically includes: The parameter-specific sub-bands and their intensity values ​​of LIBS spectroscopy and the parameter-specific sub-bands and their absorbance of NIR spectroscopy are used as input sequence data, respectively. These two types of input sequence data are first processed through a word embedding layer, and then input into two parallel encoding paths: The Transformer encoder path consists of 6 CSwin modules. Each CSwin module includes a multi-head self-attention mechanism, a feedforward neural network, and a residual connection structure. It adaptively learns the global correlation weights between features of different bands through the multi-head self-attention mechanism, uses position encoding to maintain the spatial position information of the spectral sequence, and captures the long-distance dependencies and global sequence features of the spectral data. The CNN encoder path consists of 6 MobileNet modules, which greatly reduces computational complexity while ensuring feature extraction performance. It captures local spectral texture features of different frequencies through multi-scale convolutional kernel combinations, effectively extracting local patterns and details from spectral data. Then the fusion module passes , , The three branches process features from the Transformer and CNN paths respectively. and Derived from CNN encoder features, the cross-attention mechanism is used to calculate the correlation weights between features from different paths, resulting in a cross-attention output. This output is combined with the multi-scale features provided by the CNN without LN transformation and those provided by the Transformer encoder. Weighted fusion achieves effective fusion of global sequence features and local texture features. The features obtained from each fusion layer serve as multi-scale feature inputs for the next fusion module, forming a hierarchical fusion mechanism. The final fusion module... , , The final sequence features are derived from the sixth layer CSwin of the Transformer encoder and the sixth layer MobileNet of the CNN encoder, and then obtained through 1×1 convolution.

[0019] Optionally, such as Figure 6 As shown, the processing procedure of the heterogeneous feature attention fusion module specifically includes: A gated fusion mechanism is adopted, in which the LIBS sequence features and NIR sequence features are simultaneously input into a fully connected layer, and a gate weight G between 0 and 1 is output by the Sigmoid activation function. This G represents the contribution ratio of the LIBS sequence features in the fusion result, and 1-G is the contribution ratio of the NIR sequence features. Then, the LIBS sequence features are weighted by G, and the NIRS sequence features are weighted by 1-G. The two are then added together to obtain a hybrid coded sequence feature that integrates dual-source spectral information. Through this gating mechanism, the model can adaptively adjust the fusion ratio of LIBS and NIRS features according to the spectral characteristics of different samples. For samples with more significant LIBS spectral information, the gating weight G automatically tends to 1; while for samples with more significant NIRS spectral information, G tends to 0. This dynamic weighting strategy effectively integrates the complementary information of dual-source heterogeneous spectra to form a hybrid coded sequence feature that integrates dual-source spectral information. The hybrid coded sequence feature retains the global dependencies extracted by the Transformer encoder and integrates the local feature patterns captured by the CNN encoder. At the same time, it realizes the effective fusion of LIBS and NIRS heterogeneous spectra, providing high-quality feature input for subsequent joint learning of spectral-coal quality parameters.

[0020] S4. Input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module. Through the bidirectional encoding-decoding mechanism and the bidirectional correlation constraint between the spectral and coal quality, output a high-quality latent representation that has been verified by alignment constraint and reconstruction, including a high-quality latent representation of spectral features and coal quality parameters. To address the problems of traditional coal quality testing methods that employ a one-way mapping approach, only invert coal quality parameters from spectral features, and ignore the bidirectional correlation and inherent consistency constraints between spectral data and coal quality parameters, this invention constructs a joint learning module for spectral and coal quality parameters. Conventional methods typically employ a one-way mapping process of "spectral feature extraction → regression prediction → coal quality parameters," which lacks constraint verification of the mapping relationship, resulting in insufficient model generalization ability. This invention solves this problem through a bidirectional encoding-decoding architecture and multiple constraint mechanisms.

[0021] Optionally, such as Figure 7 As shown, the processing procedure of the spectrum-coal quality joint learning module specifically includes: Hybrid coded sequence features and coal quality parameter coding characteristics (That is, Q' obtained earlier), the two heterogeneous data are mapped to a unified representation space through a latent spatial mapping module, which includes a spectral encoder and a parameter encoder. The spectral representation is generated by mapping through the spectral encoder. The Coal quality representation is generated through the parameter encoder mapping. The two representations achieve unified modeling in a latent space of the same dimension; Then the and The input spectrum-coal quality alignment constraint module implements consistency constraints in the latent space through a two-parameter contrastive learning strategy. The two-parameter contrastive learning loss function designed using equation (1) ensures consistency from the same coal sample. and Maintaining high similarity in the latent space while maximizing representational differences between different coal samples, achieving accurate alignment of matched sample pairs and effective distinction of non-matched sample pairs, for the ... A coal sample, its spectral representation Coal quality representation Forming positive sample pairs, while Coal quality representation compared to other coal samples , The negative sample pairs are constructed, and the two-parameter contrastive learning loss function is shown in equation (1): (1) in This represents the similarity score between positive sample pairs. Indicates the first The cosine similarity between the spectral representation and the coal quality representation of a coal sample; Indicates the first The sum of similarity scores between the spectral representation of a coal sample and the coal quality representation of all coal samples, including one positive sample pair and N-1 negative sample pairs, maximizes the proportion of positive sample pair similarity among all sample pairs, so that the spectral-coal quality representation of the same coal sample is closest in the latent space, while the representation distance between different coal samples is pushed far apart. When the similarity of positive sample pairs is much higher than that of negative sample pairs, the loss value tends to 0; when the similarity of positive and negative sample pairs is close, the loss value increases, enabling the model to learn more discriminative feature representations. Finally, the latent representation output by the spectral decoder from the spectral-coal quality alignment constraint module. The latent representation output from the spectrum-coal quality alignment constraint module of the coal quality decoder is reconstructed from the hybrid encoded spectral sequence features. The coal quality parameter encoding features are reconstructed in the model. The reconstruction constraints consist of the spectral reconstruction loss in equation (2) and the coal quality parameter reconstruction loss in equation (3). The reconstruction constraints require the model to retain complete original information in the latent space, so as to accurately restore the coal quality parameters and verify the rationality of the inversion results. The spectral reconstruction loss is shown in equation (2). (2) Spectral reconstruction loss ensures that the spectral decoder can obtain the spectral reconstruction loss from the spectral reconstruction loss. Accurately recover the original mixed-coded spectral features; The coal quality parameter feature encoding reconstruction loss is shown in equation (3): (3) The design of the coal quality parameter reconstruction loss ensures that the coal quality decoder can obtain the parameters from the specified parameters. Accurately reconstruct the original coal quality parameter values; Total Reconstruction Loss By minimizing the reconstruction error, the encoder is forced to learn a potential representation that includes the complete original information.

[0022] The bidirectional encoding-decoding mechanism of this invention involves encoders generating latent representations for both spectral and coal quality parameters, followed by decoders reconstructing the original data to form an encoding-decoding closed loop. The bidirectional correlation constraint between spectral and coal quality parameters ensures the physical correlation between spectral features and coal quality parameters by establishing bidirectional alignment and consistency verification in the latent space. This stage outputs a high-quality latent representation verified by alignment constraints and reconstruction, including a high-quality latent representation of spectral features and collaborative information on coal quality parameters, providing features for subsequent incremental learning and final coal quality parameter inversion.

[0023] S5. When inverting the coal quality to be inverted, if no new coal type appears, the coal quality parameters output by the spectral-coal quality joint learning module are used as the inversion prediction result. When a new coal type appears (such as expanding from lignite to anthracite, or from Northeast coal type to Southwest coal type), the high-quality latent representation of the spectral features of the existing coal quality output by the spectral-coal quality joint learning module and the real coal quality parameters of the existing coal quality after MLP feature encoding (that is, Q' obtained above) are fused together as historical spectral-coal quality latent features. These features are then input together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

[0024] To address the challenges of traditional coal quality testing models in adapting to the differences in characteristics of coal samples from different coal types and origins, as well as the performance degradation caused by the continuous emergence of new sample data in practical industrial applications, this invention employs an incremental feature learning module based on an experience transfer and gated feedback cross-coal type inversion algorithm to solve the problems of knowledge transfer and adaptation to new samples in coal quality testing.

[0025] Optionally, such as Figure 8 As shown, the processing procedure of the incremental feature learning module specifically includes: The parallel processing of the experience transfer network and the adaptive learning network achieves a balance between preserving historical knowledge and learning new features. The experience transfer network uses a four-layer deep structure: hidden layer 1 → hidden layer 2 → attention layer → LN+Softmax to deeply encode the historical spectral-coal quality potential features. The adaptive learning network adopts a residual connection architecture: residual block 1 → self-attention mechanism → residual block 2 → LN+Softmax to process the mixed encoded sequence features of new coal types. Then, the fusion weights are dynamically adjusted based on the similarity between the new coal type data and historical data. A coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated through the feature regression head. Gated feedback control is introduced. By monitoring the error between the predicted coal quality parameters and the true values ​​(which can be obtained by chemical detection methods after sampling some new coal types) in real time, the model parameters of the adaptive learning network are automatically updated when the error exceeds a preset threshold. The update intensity is intelligently adjusted according to the error magnitude and confidence level, and the update signal is fed back to the adaptive learning network for parameter update, forming a closed-loop adaptive optimization mechanism of "feature regression head → inversion prediction → error evaluation → gating judgment → parameter update → network adjustment". When the error is within an acceptable range, the gating is closed to avoid over-adjustment of the model, ensure stability, and realize the continuous learning and adaptive optimization of the model.

[0026] like Figure 9 As shown, this embodiment of the invention also provides a heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion system, the system comprising: The data acquisition and partitioning module 910 is used to acquire two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The multi-source data constitutes a dataset, which is then divided into a training set and a validation set. The identification and establishment module 920 is used to input the multi-source data in the training set into the spectral domain feature mapping module to identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. The extraction and fusion module 930 is used to map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. The NIR spectral domain features are also mapped to parameter-specific sub-bands and their absorption rates, and input into the heterogeneous encoder collaborative feature extraction module to generate LIBS sequence features and NIR sequence features respectively. Then, the LIBS sequence features and NIR sequence features are input into the heterogeneous feature attention fusion module to output a hybrid coded sequence feature that fuses dual-source spectral information. The joint learning module 940 is used to input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module, and output a high-quality latent representation verified by alignment constraints and reconstruction through a bidirectional encoding-decoding mechanism and spectral-coal quality bidirectional correlation constraints, including a high-quality latent representation of spectral features and coal quality parameters. The inversion prediction module 950 is used to, when inverting the coal quality to be inverted, if no new coal type appears, use the coal quality parameters output by the spectral-coal quality joint learning module as the inversion prediction result; when a new coal type appears, the fusion result of the high-quality latent representation of the spectral features output by the spectral-coal quality joint learning module and the real coal quality parameters of the existing coal quality after MLP feature encoding is used as the historical spectral-coal quality latent features, and together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type, is input into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

[0027] The coal quality parameter inversion system provided in this embodiment of the invention has a functional structure that corresponds to the coal quality parameter inversion method provided in this embodiment of the invention, and will not be described again here.

[0028] Figure 10 This is a schematic diagram of the structure of an electronic device 1000 provided in an embodiment of the present invention. The electronic device 1000 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 1001 and one or more memories 1002. The memory 1002 stores at least one instruction, which is loaded and executed by the processor 1001 to implement the steps of the above-mentioned heterogeneous dual-source collaborative and empirical self-evolution coal quality parameter inversion method.

[0029] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the aforementioned heterogeneous dual-source collaborative and empirically self-evolving coal quality parameter inversion method. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0030] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0031] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inverting coal quality parameters using heterogeneous dual-source synergy and empirical self-evolution, characterized in that, The method includes: S1. Collect two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The multi-source data constitutes a dataset, which is then divided into a training set and a validation set. S2. Input the multi-source data in the training set into the spectral domain feature mapping module respectively, identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. S3. Map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. Map the NIR spectral domain features to parameter-specific sub-bands and their absorption rates, and input them into the heterogeneous encoder collaborative feature extraction module to generate LIBS sequence features and NIR sequence features respectively. Then, input the LIBS sequence features and NIR sequence features into the heterogeneous feature attention fusion module to output the hybrid coding sequence features that fuse dual-source spectral information. S4. Input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module. Through the bidirectional encoding-decoding mechanism and the bidirectional correlation constraint between the spectral and coal quality, output a high-quality latent representation that has been verified by alignment constraint and reconstruction, including a high-quality latent representation of spectral features and coal quality parameters. S5. When inverting the coal quality to be inverted, if no new coal type appears, the coal quality parameters output by the spectral-coal quality joint learning module are used as the inversion prediction result. When a new coal type appears, the high-quality potential representation of the spectral features output by the spectral-coal quality joint learning module and the real coal quality parameters after MLP feature encoding of the existing coal quality are fused together as historical spectral-coal quality potential features. These features are then input together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

2. The method according to claim 1, characterized in that, The processing procedure of the spectral domain feature mapping module specifically includes: Full-band data from LIBS and NIR are input into a one-dimensional convolutional neural network. Through layer-by-layer feature extraction using three convolutional layers (Conv1, Conv2, and Conv3), multi-scale features of the spectrum and correlations between wavelengths are captured using convolutional kernels of different scales. Then, the features output by the one-dimensional convolutional neural network are used as input to the self-attention mechanism module. The self-attention mechanism module maps the features output by the one-dimensional convolutional neural network into matrices Q, K, and V. By calculating the attention weights, it learns the dependency between different wavelengths within the spectral sequence and calculates the attention-weighted enhanced feature representations K' and V'. Then, the coal quality parameters are used as prior knowledge to input the multilayer perceptron (MLP) for semantic encoding, converting the discrete parameter labels into a continuous high-dimensional vector representation Q'. Then, the cross-attention module generates an attention weight matrix by calculating the dot product similarity matrix between Q' and K'. Each element in the attention weight matrix quantitatively reflects the sensitivity and importance score of the corresponding wavelength to a specific coal quality parameter. The attention weight matrix identifies the specific sensitive bands of each coal quality parameter, establishes a parameter-band mapping relationship, and obtains LIBS spectral domain features and NIR spectral domain features. Each sub-band of the parameter-band mapping relationship is used for feature extraction of the corresponding coal quality parameter, avoiding the mixed interference of spectral information from different physical mechanisms, and achieving accurate feature extraction guided by parameters.

3. The method according to claim 1, characterized in that, The LIBS spectral domain features include: LIBS uses atomic emission lines to make total water content correspond to H and O atomic emission lines, ash content correspond to Si and Al atomic emission lines, volatile matter correspond to C and H atomic emission lines, total sulfur content correspond to S atomic emission lines, and calorific value correspond to C atomic characteristic emission lines. The NIR spectral characteristics include: NIRS uses the absorption characteristics of molecular groups to make total water correspond to the absorption band of hydroxyl groups, ash correspond to mineral oxides, volatile matter correspond to hydrocarbons, total sulfur correspond to the sulfur-hydrogen band, and calorific value correspond to the carbon-hydrogen band.

4. The method according to claim 1, characterized in that, The processing procedure of the heterogeneous encoder collaborative feature extraction module specifically includes: The parameter-specific sub-bands and their intensity values ​​of LIBS spectroscopy and the parameter-specific sub-bands and their absorbance of NIR spectroscopy are used as input sequence data, respectively. These two types of input sequence data are first processed through a word embedding layer, and then input into two parallel encoding paths: The Transformer encoder path consists of 6 CSwin modules. Each CSwin module includes a multi-head self-attention mechanism, a feedforward neural network, and a residual connection structure. It adaptively learns the global correlation weights between features of different bands through the multi-head self-attention mechanism, uses position encoding to maintain the spatial position information of the spectral sequence, and captures the long-distance dependencies and global sequence features of the spectral data. The CNN encoder path consists of 6 MobileNet modules, which greatly reduces computational complexity while ensuring feature extraction performance. It captures local spectral texture features of different frequencies through multi-scale convolutional kernel combinations, effectively extracting local patterns and details from spectral data. Then the fusion module passes , , The three branches process features from the Transformer and CNN paths respectively. and Derived from CNN encoder features, the cross-attention mechanism is used to calculate the correlation weights between features from different paths, resulting in a cross-attention output. This output is combined with the multi-scale features provided by the CNN without LN transformation and those provided by the Transformer encoder. Weighted fusion achieves effective fusion of global sequence features and local texture features. The features obtained from each fusion layer serve as multi-scale feature inputs for the next fusion module, forming a hierarchical fusion mechanism. The final fusion module... , , The final sequence features are derived from the sixth layer CSwin of the Transformer encoder and the sixth layer MobileNet of the CNN encoder, and then obtained through 1×1 convolution.

5. The method according to claim 1, characterized in that, The processing procedure of the heterogeneous feature attention fusion module specifically includes: A gated fusion mechanism is adopted, in which the LIBS sequence features and NIR sequence features are simultaneously input into a fully connected layer, and a gate weight G between 0 and 1 is output by the Sigmoid activation function. This G represents the contribution ratio of the LIBS sequence features in the fusion result, and 1-G is the contribution ratio of the NIR sequence features. Then, the LIBS sequence features are weighted by G, and the NIRS sequence features are weighted by 1-G. The two are then added together to obtain a hybrid coded sequence feature that integrates dual-source spectral information. Through this gating mechanism, the model can adaptively adjust the fusion ratio of LIBS and NIRS features according to the spectral characteristics of different samples. For samples with more significant LIBS spectral information, the gating weight G automatically tends to 1; while for samples with more significant NIRS spectral information, G tends to 0. This dynamic weighting strategy effectively integrates the complementary information of dual-source heterogeneous spectra to form a hybrid coded sequence feature that integrates dual-source spectral information. The hybrid coded sequence feature retains the global dependencies extracted by the Transformer encoder and integrates the local feature patterns captured by the CNN encoder. At the same time, it realizes the effective fusion of LIBS and NIRS heterogeneous spectra, providing high-quality feature input for subsequent joint learning of spectral-coal quality parameters.

6. The method according to claim 1, characterized in that, The processing procedure of the spectrum-coal quality joint learning module specifically includes: Hybrid coded sequence features and coal quality parameter coding characteristics The latent spatial mapping module maps two heterogeneous data types to a unified representation space. This module includes a spectral encoder and a parameter encoder. The spectral representation is generated by mapping through the spectral encoder. The Coal quality representation is generated through the parameter encoder mapping. The two representations achieve unified modeling in a latent space of the same dimension; Then the and The input spectrum-coal quality alignment constraint module implements consistency constraints in the latent space through a two-parameter contrastive learning strategy. The two-parameter contrastive learning loss function designed using equation (1) ensures consistency from the same coal sample. and Maintaining high similarity in the latent space while maximizing representational differences between different coal samples, achieving accurate alignment of matched sample pairs and effective distinction of non-matched sample pairs, for the ... A coal sample, its spectral representation Coal quality representation Forming positive sample pairs, while Coal quality representation compared to other coal samples , The negative sample pairs are constructed, and the two-parameter contrastive learning loss function is shown in equation (1): （1） in This represents the similarity score between positive sample pairs. Indicates the first The cosine similarity between the spectral representation and the coal quality representation of a coal sample; Indicates the first The sum of similarity scores between the spectral representation of a coal sample and the coal quality representation of all coal samples, including one positive sample pair and N-1 negative sample pairs, maximizes the proportion of positive sample pair similarity among all sample pairs, so that the spectral-coal quality representation of the same coal sample is closest in the latent space, while the representation distance between different coal samples is pushed far apart. When the similarity of positive sample pairs is much higher than that of negative sample pairs, the loss value tends to 0; when the similarity of positive and negative sample pairs is close, the loss value increases, enabling the model to learn more discriminative feature representations. Finally, the latent representation output by the spectral decoder from the spectral-coal quality alignment constraint module. The latent representation output from the spectrum-coal quality alignment constraint module of the coal quality decoder is reconstructed from the hybrid encoded spectral sequence features. The coal quality parameter encoding features are reconstructed in the model. The reconstruction constraints consist of the spectral reconstruction loss in equation (2) and the coal quality parameter reconstruction loss in equation (3). The reconstruction constraints require the model to retain complete original information in the latent space, so as to accurately restore the coal quality parameters and verify the rationality of the inversion results. The spectral reconstruction loss is shown in equation (2). （2） Spectral reconstruction loss ensures that the spectral decoder can obtain the spectral reconstruction loss from the spectral reconstruction loss. Accurately recover the original mixed-coded spectral features; The coal quality parameter feature encoding reconstruction loss is shown in equation (3): （3） The design of the coal quality parameter reconstruction loss ensures that the coal quality decoder can obtain the parameters from the specified parameters. Accurately reconstruct the original coal quality parameter values; Total Reconstruction Loss By minimizing the reconstruction error, the encoder is forced to learn a potential representation that includes the complete original information.

7. The method according to claim 1, characterized in that, The processing procedure of the incremental feature learning module specifically includes: The parallel processing of the experience transfer network and the adaptive learning network achieves a balance between preserving historical knowledge and learning new features. The experience transfer network uses a four-layer deep structure: hidden layer 1 → hidden layer 2 → attention layer → LN+Softmax to deeply encode the historical spectral-coal quality potential features. The adaptive learning network adopts a residual connection architecture: residual block 1 → self-attention mechanism → residual block 2 → LN+Softmax to process the mixed encoded sequence features of new coal types. Then, the fusion weights are dynamically adjusted based on the similarity between the new coal type data and historical data. A coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated through the feature regression head. Gated feedback control is introduced. By monitoring the error between the inverted predicted coal quality parameters and the actual values ​​in real time, the model parameters of the adaptive learning network are automatically updated when the error exceeds a preset threshold. The update intensity is intelligently adjusted according to the error magnitude and confidence level, and the update signal is fed back to the adaptive learning network for parameter update, forming a closed-loop adaptive optimization mechanism of "feature regression head → inversion prediction → error evaluation → gating judgment → parameter update → network adjustment". When the error is within an acceptable range, the gating is closed to avoid over-adjustment of the model, ensure stability, and realize the continuous learning and adaptive optimization of the model.

8. A heterogeneous dual-source collaborative and empirically self-evolving coal quality parameter inversion system, characterized in that, The system includes: The data acquisition and partitioning module is used to acquire two complementary multi-source data of existing coal quality: laser-induced breakdown spectrum (LIBS) and near-infrared spectrum (NIR). The dataset is composed of the multi-source data and is divided into a training set and a validation set. The identification and establishment module is used to input the multi-source data in the training set into the spectral domain feature mapping module to identify the sensitive band intervals of each coal quality parameter in LIBS and NIR, establish parameter-band mapping relationship, and obtain LIBS spectral domain features and NIR spectral domain features. The extraction and fusion module is used to map the LIBS spectral domain features to parameter-specific sub-bands and their intensity values, and input them into the heterogeneous encoder collaborative feature extraction module. The NIR spectral domain features are also mapped to parameter-specific sub-bands and their absorption rates, and input into the heterogeneous encoder collaborative feature extraction module to generate LIBS sequence features and NIR sequence features respectively. Then, the LIBS sequence features and NIR sequence features are input into the heterogeneous feature attention fusion module to output a hybrid coded sequence feature that fuses dual-source spectral information. The joint learning module is used to input the hybrid encoded sequence features and coal quality parameter encoded features into the spectral-coal quality joint learning module. Through a bidirectional encoding-decoding mechanism and bidirectional correlation constraints between the spectral and coal quality, it outputs a high-quality latent representation that has been verified by alignment constraints and reconstruction, including a high-quality latent representation of spectral features and coal quality parameters. The inversion prediction module is used to use the coal quality parameters output by the spectral-coal quality joint learning module as the inversion prediction result when no new coal type appears during the inversion of coal quality to be inverted. When a new coal type appears, the high-quality potential representation of the spectral features output by the spectral-coal quality joint learning module and the real coal quality parameters after MLP feature encoding of the existing coal quality are fused together as historical spectral-coal quality potential features. These features are then input together with the hybrid encoded sequence features output by the heterogeneous feature attention fusion module for the new coal type into the incremental feature learning module. Through cross-coal type learning with experience transfer and gating feedback, a coal quality parameter inversion prediction result that takes into account both historical knowledge and new features is generated.

9. An electronic device comprising a processor and a memory, wherein the memory stores at least one instruction, characterized in that, The processor loads and executes at least one instruction to implement the heterogeneous dual-source collaborative and empirical self-evolutionary coal quality parameter inversion method as described in any one of claims 1-7.

10. A computer-readable storage medium storing at least one instruction, characterized in that, The at least one instruction is loaded and executed by the processor to implement the heterogeneous dual-source collaborative and empirical self-evolutionary coal quality parameter inversion method as described in any one of claims 1-7.

Images