Coal quality near infrared spectrum correction compression self-correction integrated prediction method

Abstract

The present application discloses a coal quality near-infrared spectrum correction compression self-correction integrated prediction method, and belongs to the technical field of coal detection. The present application builds a coal quality near-infrared spectrum detection system, collects the near-infrared spectrum original data of the coal quality sample through an intelligent sensor, performs correction and compression processing on the spectrum data, constructs a self-correction integrated prediction model, integrates and predicts the coal quality index through the self-correction iterative optimization mechanism of the self-correction integrated prediction model, synchronously collects the spectrum original data and the environment and position correlation data, realizes the intelligentization and automation of the detection whole link, performs multi-factor correction and feature compression processing on the pretreated spectrum data, constructs an integrated prediction model containing a self-correction iteration module, completes the prediction of the coal quality index, triggers the self-correction mechanism based on the error threshold, completes the double iterative optimization of the model weight and the internal parameter, improves the model generalization ability and the prediction precision, and realizes the accurate and efficient integrated prediction of the coal quality index such as ash content.

Description

Technical Field

[0001] This invention relates to the field of coal detection technology, and in particular to a coal quality near-infrared spectral correction, compression, self-correction, integrated prediction method. Background Technology

[0002] Current near-infrared spectroscopy methods for coal quality detection still have many technical shortcomings. Traditional detection systems have scattered and poorly coordinated sensor units, capable of acquiring only single spectral data. They are easily affected by factors such as ambient temperature and humidity, and deviations in coal sample placement. Furthermore, existing spectral data processing methods rely on simplistic corrections, addressing only single error factors and failing to eliminate distortions caused by multiple factors such as wavelength drift, baseline shift, and light intensity attenuation. Moreover, they lack effective compression of high-dimensional spectral data, significantly increasing the computational complexity of subsequent prediction models and reducing operational efficiency. In addition, existing coal quality prediction models often employ single-algorithm models, resulting in weak generalization ability and low prediction accuracy. Some ensemble models also lack dynamic self-optimization capabilities, with fixed model parameters that cannot adaptively adjust to actual prediction errors. This leads to excessive prediction deviations when dealing with spectral data from different coal types and detection environments, making it difficult to meet the high-precision and high-stability detection requirements for coal quality indicators in industrial production. Summary of the Invention

[0003] The purpose of this invention is to provide a coal near-infrared spectral correction and compression self-correction integrated prediction method to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a coal near-infrared spectral correction, compression, self-correction, integrated prediction method, comprising:

[0005] A coal near-infrared spectroscopy detection system is built based on an intelligent sensing system, which integrates intelligent sensing elements and multiple types of intelligent sensors.

[0006] Raw near-infrared spectral data of coal samples are collected by intelligent sensors. The raw near-infrared spectral data are then converted and preprocessed by intelligent sensing elements. The preprocessed spectral data are then corrected and compressed sequentially.

[0007] A self-correcting integrated prediction model is constructed. The processed spectral data is input into the self-correcting integrated prediction model, and the coal quality indicators are integrated and predicted through the self-correcting iterative optimization mechanism of the self-correcting integrated prediction model.

[0008] Furthermore, the intelligent sensing system includes a spectral acquisition sensing module, a signal conditioning sensing module, a data transmission sensing module, and a main control sensing module;

[0009] The intelligent sensors are near-infrared spectral sensors, ambient temperature and humidity sensors, and coal sample position sensors; the intelligent sensing elements are photoelectric conversion elements, signal amplification elements, and filtering elements.

[0010] The smart sensor and smart sensing element are connected through a main control sensing module.

[0011] Furthermore, raw near-infrared spectral data of coal samples are collected using smart sensors, including:

[0012] The coal sample is placed in the preset testing station. The coal sample position sensor confirms that the coal sample is positioned. The ambient temperature and humidity sensor collects the temperature and humidity data of the testing environment in real time and uploads it to the main control sensor module.

[0013] The near-infrared spectral sensor performs near-infrared spectral scanning on the coal sample according to the preset scanning frequency and spectral band, and acquires raw near-infrared spectral data including background noise;

[0014] The temperature and humidity data are associated and labeled with the raw spectral data.

[0015] Furthermore, correction processing is performed on the preprocessed spectral data, specifically including:

[0016] A spectral correction sample library was constructed based on historical detection data from the intelligent sensing system, and wavelength drift, baseline shift, and light intensity attenuation were determined as correction factors.

[0017] Wavelength and baseline corrections were performed on the spectral data, and light intensity attenuation compensation corrections were performed on the spectral data in conjunction with environmental temperature and humidity correlation data to obtain the corrected effective near-infrared spectral data of coal.

[0018] Furthermore, compression processing is performed on the corrected spectral data, specifically including:

[0019] Feature extraction is performed on the corrected spectral data;

[0020] Determine the principal component eigenvectors and eigenvalues ​​of the spectral data;

[0021] Remove redundant features whose feature values ​​are below a preset threshold;

[0022] The core spectral feature data is compressed using a sparse representation algorithm to obtain a compressed spectral feature dataset.

[0023] Furthermore, a self-correcting ensemble prediction model is constructed, including:

[0024] Backpropagation neural network, random forest, and support vector machine were selected as the basic prediction models.

[0025] Each basic prediction model was trained independently based on the standard coal quality spectral dataset to determine the initial model parameters for each model.

[0026] A weighted ensemble algorithm is used to integrate the trained basic prediction models. Dynamic weight coefficients are set according to the prediction accuracy of each basic prediction model to construct an initial ensemble prediction model.

[0027] A self-correcting iterative module is embedded in the initial ensemble prediction model, and a model prediction error threshold is set as the self-correction trigger condition to complete the construction of the self-correcting ensemble prediction model.

[0028] Furthermore, the integrated prediction of coal quality indicators is achieved through the model's self-correcting iterative optimization mechanism, including:

[0029] The compressed spectral feature dataset is input into the self-correcting ensemble prediction model;

[0030] The coal quality index prediction results output by each basic prediction model are defined as the first prediction parameter. Each first prediction parameter corresponds to the single model prediction value of ash content, volatile matter, fixed carbon and calorific value, respectively. At the same time, the dynamic weight coefficient of each basic prediction model is extracted as the first weight parameter.

[0031] The initial integrated prediction results of coal quality indicators are obtained by weighted fusion calculation of each first prediction parameter based on the first weight parameter, and the initial integrated prediction results are defined as the second prediction parameters.

[0032] The measured index values ​​of coal quality samples are obtained as standard reference parameters. The relative error value between the second prediction parameter and the standard reference parameter is calculated, and the relative error value is defined as the first error parameter.

[0033] The first error parameter is compared with the preset error threshold parameter. If the first error parameter is less than or equal to the error threshold parameter, the second prediction parameter is determined as the final integrated prediction result of coal quality indicators. If the first error parameter is greater than the error threshold parameter, the self-correction iteration module is triggered to optimize the model parameters.

[0034] After the self-correcting iteration module is started, the current model parameters of each basic prediction model are extracted as the first adjustment parameters. The current model parameters include model weights, bias terms and kernel function parameters.

[0035] The prediction deviation of each basic prediction model is calculated as the second error parameter. The second error parameter is the difference between the first prediction parameter of each basic prediction model and the standard reference parameter.

[0036] The first weight parameter is adaptively redistributed based on the second error parameter to obtain the updated second weight parameter.

[0037] With minimizing the first error parameter as the optimization objective, an iterative optimization algorithm is used to update and correct the first adjustment parameters of each basic prediction model to obtain the second adjustment parameters, thus completing the parameter iteration of each basic prediction model.

[0038] Furthermore, after completing the parameter iteration of each basic prediction model, the compressed spectral feature dataset is re-input into each basic prediction model after parameter updates to obtain the updated first prediction parameters.

[0039] The first prediction parameters are weighted and fused based on the second weight parameters to obtain the updated second prediction parameters, and the corresponding first error parameters are recalculated.

[0040] Multiple rounds of self-correcting iterative optimization are performed until the calculated first error parameter is less than or equal to the error threshold parameter. The iteration is then stopped, and the second prediction parameter at this point is determined as the final integrated prediction result of coal quality indicators. At the same time, the second weight parameter and the second adjustment parameter of this iteration are saved as the optimal parameters of the model.

[0041] Furthermore, the parameter adaptive adjustment process of the self-correcting iterative module includes:

[0042] With minimizing model prediction error as the optimization objective, the pre-defined gradient solving algorithm in the self-correcting iterative module is extracted as the basic optimization algorithm.

[0043] The current parameters to be adjusted of each basic prediction model are retrieved as the first parameter set. The first parameter set includes model weight parameters, bias term parameters and kernel function parameters, and each parameter corresponds to a unique parameter identifier.

[0044] Based on the first error parameter, the loss function of each basic prediction model is constructed, and the partial derivative of the loss function with respect to each parameter in the first parameter set is solved to obtain the gradient change value corresponding to each parameter.

[0045] All gradient change values ​​are integrated into a first gradient set, and the parameter identifiers of the first gradient set correspond one-to-one with the first parameter set.

[0046] The momentum factor preset in the correction iteration module is retrieved as the first correction coefficient. Each gradient change value in the first gradient set is multiplied by the first correction coefficient to obtain the corrected second gradient set.

[0047] The learning rate is introduced as a second correction coefficient, which is an adaptive dynamic value.

[0048] The iterative update step size of each parameter is obtained by multiplying each gradient change value in the second gradient set with the second correction coefficient.

[0049] Furthermore, after obtaining the iteration update step size for each parameter, subtract the corresponding iteration update step size from each parameter in the first parameter set to obtain the updated candidate parameter set, which is defined as the second parameter set.

[0050] A preset proportion of validation set data is extracted from the spectral feature dataset as the first validation sample. The first validation sample is then replaced with each of the basic prediction models in the second parameter set to obtain the validation set prediction results, which are defined as the third prediction parameters.

[0051] The measured coal quality index value corresponding to the first verification sample was retrieved as the second standard parameter.

[0052] Calculate the verification error value between the third predicted parameter and the second standard parameter, define it as the third error parameter, and compare the third error parameter with the first error parameter before the parameter update;

[0053] If the third error parameter is less than the first error parameter, the parameter update is deemed valid. The second parameter set is determined as the optimal parameter set for each basic prediction model, and the parameter iteration update is completed. At the same time, the current second correction coefficient is retained as the initial learning rate for the next iteration.

[0054] If the third error parameter is greater than or equal to the first error parameter, the parameter update is deemed invalid, the second parameter set is abandoned and restored to the first parameter set, and the second correction coefficient is adaptively adjusted by multiplying it by the preset attenuation coefficient to obtain a new second correction coefficient.

[0055] Based on the adjusted second correction coefficient, the parameter gradient is solved and iterated again until the obtained third error parameter is less than the first error parameter and the gradient change value of each parameter tends to 0. The model parameters are then determined to have converged to the optimal value, and the parameter set at this time is determined as the final updated parameter set of each basic prediction model.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. This invention establishes a coal near-infrared spectroscopy detection system by integrating an intelligent sensing system. It is equipped with multiple types of intelligent sensors and sensing elements and achieves module collaboration. It can simultaneously collect raw spectral data and environmental and location-related data, ensuring the integrity, accuracy and standardization of spectral data. It replaces the traditional manual operation process, realizes the intelligentization and automation of the entire detection process, and greatly improves the efficiency of spectral data acquisition.

[0058] 2. This invention performs multi-factor correction and feature compression on preprocessed spectral data, constructs a correction sample library based on historical data, and achieves accurate correction for problems such as wavelength drift, while incorporating temperature and humidity compensation to effectively eliminate data distortion. Through principal component analysis and sparse representation algorithms, core features are extracted and data dimensionality is reduced. While fully preserving key information about coal quality characteristics, the data volume is significantly reduced, the complexity of subsequent model calculations is decreased, and the computational efficiency is improved. At the same time, the storage of correction parameters enables experience reuse, making the data processing process more practical and efficient.

[0059] 3. This invention constructs an integrated prediction model containing a self-correcting iterative module to predict coal quality indicators. It integrates the algorithmic advantages of multiple basic models such as BP neural networks, optimizes and integrates prediction capabilities through dynamic weights, defines quantitative parameters to make the prediction process traceable, triggers a self-correction mechanism based on an error threshold, completes dual iterative optimization of model weights and internal parameters, and ensures the effectiveness of parameter updates through a validation set until the error reaches the target to achieve high-precision prediction. This invention overcomes the limitations of single models and traditional integrated models, improves the model's generalization ability and prediction accuracy, and achieves accurate and efficient integrated prediction of coal quality indicators such as ash content. Attached Figure Description

[0060] Figure 1 This is a flowchart of the coal near-infrared spectral correction, compression, self-correction, integrated prediction method of the present invention. Detailed Implementation

[0061] 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.

[0062] Please see Figure 1 The present invention provides the following technical solutions:

[0063] A coal near-infrared spectral correction and compression self-correction integrated prediction method includes:

[0064] A coal near-infrared spectroscopy detection system was built based on an intelligent sensing system. The intelligent sensing system integrates intelligent sensing elements and multiple types of intelligent sensors.

[0065] Raw near-infrared spectral data of coal samples are collected by intelligent sensors. The raw near-infrared spectral data are then converted and preprocessed by intelligent sensing elements. The preprocessed spectral data are then corrected and compressed sequentially.

[0066] A self-correcting integrated prediction model is constructed. The processed spectral data is input into the self-correcting integrated prediction model, and the coal quality indicators are integrated and predicted through the self-correcting iterative optimization mechanism of the self-correcting integrated prediction model.

[0067] In the above embodiments, an intelligent sensing system integrating intelligent sensing elements and multiple types of intelligent sensors is built as the basic carrier for coal near-infrared spectral detection. This achieves integrated spectral data acquisition and signal processing, allowing the detection process to simultaneously acquire spectral information and environmental and location-related data. The intelligent sensing elements achieve efficient conversion of optical signals to electrical signals, ensuring the integrity and validity of spectral data from the source of detection. The construction of the detection system realizes the intelligent and automated detection of coal quality spectra, replacing the cumbersome process of manual operation and significantly improving the efficiency and accuracy of raw spectral data acquisition.

[0068] The intelligent sensing system includes a spectrum acquisition sensing module, a signal conditioning sensing module, a data transmission sensing module, and a main control sensing module;

[0069] The intelligent sensors include near-infrared spectroscopy sensors, ambient temperature and humidity sensors, and coal sample position sensors; the intelligent sensing elements include photoelectric conversion elements, signal amplification elements, and filtering elements.

[0070] The smart sensors and smart sensing elements are connected through the main control sensing module.

[0071] When collecting raw near-infrared spectral data of coal samples:

[0072] The coal sample is placed in the preset testing station. The coal sample position sensor confirms that the coal sample is positioned. The ambient temperature and humidity sensor collects the temperature and humidity data of the testing environment in real time and uploads it to the main control sensor module.

[0073] The near-infrared spectral sensor performs near-infrared spectral scanning on the coal sample according to the preset scanning frequency and spectral band, and acquires raw near-infrared spectral data including background noise;

[0074] The temperature and humidity data are associated and labeled with the raw spectral data.

[0075] In the above embodiments, the intelligent sensing system is divided into spectral acquisition, signal conditioning, data transmission, and main control sensing modules. The specific configurations of three types of intelligent sensors (near-infrared spectroscopy, temperature and humidity, and coal sample location) and three types of intelligent sensing elements (photoelectric conversion, signal amplification, and filtering) are clearly defined. Communication between these units is achieved through the main control sensing module, forming a modular and standardized sensing system architecture. This ensures precise matching between sensors and sensing elements, guarantees smooth signal transmission and processing, and allows for flexible addition or removal of sensing units according to actual detection needs. Furthermore, the centralized management of the main control module makes data interaction more efficient, effectively reducing the system's failure rate and improving the stability of coal quality spectral detection.

[0076] The preprocessed spectral data undergoes correction and compression processing, specifically including:

[0077] A spectral correction sample library was constructed based on historical detection data from the intelligent sensing system, and wavelength drift, baseline shift, and light intensity attenuation were determined as correction factors.

[0078] Wavelength and baseline corrections are performed on the spectral data, and light intensity attenuation compensation correction is performed on the spectral data in conjunction with environmental temperature and humidity correlation data to obtain the corrected effective near-infrared spectral data of coal. At the same time, the correction parameters are synchronously stored in the local database of the intelligent sensing system.

[0079] Principal component analysis algorithm is used to extract features from the corrected effective spectral data;

[0080] Determine the principal component eigenvectors and eigenvalues ​​of the spectral data;

[0081] Redundant features with eigenvalues ​​below a preset threshold are removed, and core spectral feature data that can characterize coal quality are retained;

[0082] The core spectral feature data is compressed using a sparse representation algorithm to obtain a compressed spectral feature dataset, which reduces the complexity of subsequent model calculations.

[0083] In the above embodiments, the coal sample is accurately located using a coal sample position sensor when collecting raw spectral data, avoiding spectral acquisition errors caused by coal sample placement deviations. At the same time, environmental data is collected synchronously using an environmental temperature and humidity sensor and associated with the spectral data, enabling real-time recording of environmental factors. This acquisition process integrates coal sample positioning, environmental monitoring, and spectral scanning, significantly improving the efficiency of raw spectral data acquisition. The acquired spectral data includes complete environmental and location attributes, taking into account various interference factors from the acquisition stage, effectively improving the quality of raw spectral data.

[0084] Constructing a self-correcting ensemble prediction model includes:

[0085] Backpropagation neural network, random forest, and support vector machine were selected as the basic prediction models.

[0086] Each basic prediction model was trained independently based on the standard coal quality spectral dataset to determine the initial model parameters for each model.

[0087] A weighted ensemble algorithm is used to integrate the trained basic prediction models. Dynamic weight coefficients are set according to the prediction accuracy of each basic prediction model to construct an initial ensemble prediction model.

[0088] A self-correcting iterative module is embedded in the initial ensemble prediction model, and a model prediction error threshold is set as the self-correction trigger condition to complete the construction of the self-correcting ensemble prediction model.

[0089] In the above embodiments, a multi-foundation prediction model is constructed by selecting BP neural network, random forest and support vector machine. The algorithm advantages of each model are fully utilized. The nonlinear fitting ability of BP neural network, the anti-overfitting ability of random forest and the small sample learning ability of support vector machine complement each other, which breaks through the limitations of single model prediction and improves the generalization ability of prediction model.

[0090] Each basic model was trained independently using a standard coal quality spectral dataset, ensuring the training effectiveness of each model. A weighted ensemble algorithm was then used, with dynamic weight coefficients set based on prediction accuracy, assigning higher weights to models with higher prediction accuracy. This optimized integration of the prediction capabilities of the basic models. A self-correcting iterative module was embedded in the ensemble model, and error threshold triggering conditions were set, enabling the model to self-optimize and adjust parameters in real time based on prediction errors.

[0091] The integrated prediction of coal quality indicators is achieved through a self-correcting iterative optimization mechanism of the model, including:

[0092] The compressed spectral feature dataset is input into the self-correcting ensemble prediction model;

[0093] The coal quality index prediction results output by each basic prediction model are defined as the first prediction parameter. Each first prediction parameter corresponds to the single model prediction value of ash content, volatile matter, fixed carbon and calorific value, respectively. At the same time, the dynamic weight coefficient of each basic prediction model is extracted as the first weight parameter.

[0094] The initial integrated prediction results of coal quality indicators are obtained by weighted fusion calculation of each first prediction parameter based on the first weight parameter, and the initial integrated prediction results are defined as the second prediction parameters.

[0095] The measured index values ​​of coal quality samples are obtained as standard reference parameters. The relative error value between the second prediction parameter and the standard reference parameter is calculated, and the relative error value is defined as the first error parameter.

[0096] The first error parameter is compared with the preset error threshold parameter. If the first error parameter is less than or equal to the error threshold parameter, the second prediction parameter is determined as the final integrated prediction result of coal quality indicators. If the first error parameter is greater than the error threshold parameter, the self-correction iteration module is triggered to optimize the model parameters.

[0097] After the self-correcting iteration module is started, the current model parameters of each basic prediction model are extracted as the first adjustment parameters. The current model parameters include model weights, bias terms and kernel function parameters.

[0098] The prediction deviation of each basic prediction model is calculated as the second error parameter. The second error parameter is the difference between the first prediction parameter of each basic prediction model and the standard reference parameter.

[0099] The first weight parameter is adaptively redistributed based on the second error parameter to obtain the updated second weight parameter. The allocation of the second weight parameter is inversely proportional to the second error parameter of the corresponding basic prediction model. The smaller the error, the higher the weight ratio. The sum of all the second weight parameters is 1.

[0100] With minimizing the first error parameter as the optimization objective, an iterative optimization algorithm is used to update and correct the first adjustment parameters of each basic prediction model to obtain the second adjustment parameters, thus completing the parameter iteration of each basic prediction model.

[0101] The compressed spectral feature dataset is re-input into the updated parameters of each basic prediction model to obtain the updated first prediction parameters.

[0102] The first prediction parameters are weighted and fused based on the second weight parameters to obtain the updated second prediction parameters, and the corresponding first error parameters are recalculated.

[0103] Multiple rounds of self-correcting iterative optimization are performed until the calculated first error parameter is less than or equal to the error threshold parameter. The iteration is then stopped, and the second prediction parameter at this point is determined as the final integrated prediction result of coal quality indicators. At the same time, the second weight parameter and the second adjustment parameter of this iteration are saved as the optimal parameters of the model.

[0104] In the above embodiments, by defining a series of data parameters such as the first prediction parameter, the second prediction parameter, and the first error parameter, the quantitative analysis of the coal quality index prediction process is realized. This allows for clear numerical representations of the prediction results of each basic model, the integrated prediction results, and the prediction errors, improving the standardization and traceability of the prediction process. A weighted fusion algorithm based on dynamic weight coefficients enables the scientific integration of the prediction results from each basic model. Furthermore, by comparing the prediction error with a preset threshold to determine whether self-correction is triggered, model optimization becomes more targeted, avoiding meaningless parameter adjustments.

[0105] The self-correction process involves adaptive redistribution of weight parameters and iterative updates of model parameters, achieving dual optimization of model weights and internal parameters. Multiple iterations until the error reaches the target ensure the accuracy of the prediction results. Simultaneously, the preservation of optimal parameters accumulates model parameters, allowing subsequent predictions to directly reuse high-quality parameters, further improving prediction efficiency and ultimately achieving high-precision integrated prediction of coal quality indicators such as ash content and volatile matter.

[0106] The parameter adaptive adjustment process of the self-correcting iterative module includes:

[0107] With minimizing model prediction error as the optimization objective, the pre-defined gradient solving algorithm in the self-correcting iterative module is extracted as the basic optimization algorithm.

[0108] The current parameters to be adjusted of each basic prediction model are retrieved as the first parameter set. The first parameter set includes model weight parameters, bias term parameters and kernel function parameters, and each parameter corresponds to a unique parameter identifier.

[0109] Based on the first error parameter, the loss function of each basic prediction model is constructed, and the partial derivative of the loss function with respect to each parameter in the first parameter set is solved to obtain the gradient change value corresponding to each parameter.

[0110] All gradient change values ​​are integrated into the first gradient set, and the parameter identifiers of the first gradient set correspond one-to-one with the first parameter set;

[0111] The momentum factor preset in the correction iteration module is retrieved as the first correction coefficient. The value range of the first correction coefficient is [0.8, 0.95]. Each gradient change value in the first gradient set is multiplied by the first correction coefficient to obtain the corrected second gradient set.

[0112] The learning rate is introduced as a second correction coefficient, which is an adaptive dynamic value. The initial range of the learning rate is [0.001, 0.01].

[0113] The step size of each parameter is obtained by multiplying each gradient change value in the second gradient set with the second correction coefficient.

[0114] Subtract the corresponding iteration update step size from each parameter in the first parameter set to obtain the updated candidate parameter set, which is defined as the second parameter set.

[0115] A preset proportion of validation set data is extracted from the spectral feature dataset as the first validation sample. The first validation sample is then replaced with each of the basic prediction models in the second parameter set to obtain the validation set prediction results, which are defined as the third prediction parameters.

[0116] The measured coal quality index value corresponding to the first verification sample was retrieved as the second standard parameter.

[0117] Calculate the verification error value between the third predicted parameter and the second standard parameter, define it as the third error parameter, and compare the third error parameter with the first error parameter before the parameter update;

[0118] If the third error parameter is less than the first error parameter, the parameter update is deemed valid. The second parameter set is determined as the optimal parameter set for each basic prediction model, and the parameter iteration update is completed. At the same time, the current second correction coefficient is retained as the initial learning rate for the next iteration.

[0119] If the third error parameter is greater than or equal to the first error parameter, the parameter update is deemed invalid. The second parameter set is discarded and the system is restored to the first parameter set. At the same time, the second correction coefficient is adaptively adjusted by multiplying it by a preset attenuation coefficient to obtain a new second correction coefficient. The attenuation coefficient is set to 0.5.

[0120] Based on the adjusted second correction coefficient, the parameter gradient is solved and iterated again until the obtained third error parameter is less than the first error parameter and the gradient change value of each parameter tends to 0. The model parameters are then determined to have converged to the optimal value, and the parameter set at this time is determined as the final updated parameter set of each basic prediction model.

[0121] In the above embodiments, minimizing the model prediction error is the optimization objective. A gradient descent algorithm is used to adjust the parameters, achieving precise optimization of the model parameters. The gradient change value is obtained by taking the partial derivative of the loss function, making the direction of parameter adjustment more scientific and effectively avoiding blind parameter iteration. Momentum factor and adaptive learning rate are introduced as correction coefficients. The momentum factor accelerates the parameter convergence speed, while the adaptive learning rate is dynamically adjusted according to the parameter update effect, solving the problems of slow convergence and easy getting trapped in local optima in traditional gradient descent algorithms. The validity of the updated parameters is verified by validation set data. The validity of the parameters is judged by comparing the validation error with the original error, ensuring the rationality of parameter updates and avoiding model performance degradation caused by invalid updates. Iteration is repeated until the parameters converge to the optimal value, keeping the parameters of the basic prediction model in an optimal state at all times, greatly improving the prediction accuracy of each basic model. At the same time, the setting of parameter labels makes parameter management clearer and improves the efficiency of parameter adjustment.

[0122] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A coal near-infrared spectral correction, compression, self-correcting integrated prediction method, characterized in that, include: A coal near-infrared spectroscopy detection system is built based on an intelligent sensing system, which integrates intelligent sensing elements and multiple types of intelligent sensors. Raw near-infrared spectral data of coal samples are collected by intelligent sensors. The raw near-infrared spectral data are then converted and preprocessed by intelligent sensing elements. The preprocessed spectral data are then corrected and compressed sequentially. A self-correcting integrated prediction model is constructed. The processed spectral data is input into the self-correcting integrated prediction model, and the coal quality indicators are integrated and predicted through the self-correcting iterative optimization mechanism of the self-correcting integrated prediction model.

2. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 1, characterized in that, The intelligent sensing system includes a spectral acquisition sensing module, a signal conditioning sensing module, a data transmission sensing module, and a main control sensing module. The intelligent sensors are near-infrared spectral sensors, ambient temperature and humidity sensors, and coal sample position sensors; the intelligent sensing elements are photoelectric conversion elements, signal amplification elements, and filtering elements. The smart sensor and smart sensing element are connected through a main control sensing module.

3. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 2, characterized in that, Raw near-infrared spectral data of coal samples were collected using smart sensors, including: The coal sample is placed in the preset testing station. The coal sample position sensor confirms that the coal sample is positioned. The ambient temperature and humidity sensor collects the temperature and humidity data of the testing environment in real time and uploads it to the main control sensor module. The near-infrared spectral sensor performs near-infrared spectral scanning on the coal sample according to the preset scanning frequency and spectral band, and acquires raw near-infrared spectral data including background noise; The temperature and humidity data are associated and labeled with the raw spectral data.

4. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 1, characterized in that, Correction processing is performed on the preprocessed spectral data, specifically including: A spectral correction sample library was constructed based on historical detection data from the intelligent sensing system, and wavelength drift, baseline shift, and light intensity attenuation were determined as correction factors. Wavelength and baseline corrections were performed on the spectral data, and light intensity attenuation compensation corrections were performed on the spectral data in conjunction with environmental temperature and humidity correlation data to obtain the corrected effective near-infrared spectral data of coal.

5. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 4, characterized in that, The corrected spectral data is compressed, specifically including: Feature extraction is performed on the corrected spectral data; Determine the principal component eigenvectors and eigenvalues ​​of the spectral data; Remove redundant features whose feature values ​​are below a preset threshold; The core spectral feature data is compressed using a sparse representation algorithm to obtain a compressed spectral feature dataset.

6. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 1, characterized in that, Constructing a self-correcting ensemble prediction model includes: Backpropagation neural network, random forest, and support vector machine were selected as the basic prediction models. Each basic prediction model was trained independently based on the standard coal quality spectral dataset to determine the initial model parameters for each model. A weighted ensemble algorithm is used to integrate the trained basic prediction models. Dynamic weight coefficients are set according to the prediction accuracy of each basic prediction model to construct an initial ensemble prediction model. A self-correcting iterative module is embedded in the initial ensemble prediction model, and a model prediction error threshold is set as the self-correction trigger condition to complete the construction of the self-correcting ensemble prediction model.

7. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 6, characterized in that, The integrated prediction of coal quality indicators is achieved through a self-correcting iterative optimization mechanism of the model, including: The compressed spectral feature dataset is input into the self-correcting ensemble prediction model; The coal quality index prediction results output by each basic prediction model are defined as the first prediction parameter. Each first prediction parameter corresponds to the single model prediction value of ash content, volatile matter, fixed carbon and calorific value, respectively. At the same time, the dynamic weight coefficient of each basic prediction model is extracted as the first weight parameter. The initial integrated prediction results of coal quality indicators are obtained by weighted fusion calculation of each first prediction parameter based on the first weight parameter, and the initial integrated prediction results are defined as the second prediction parameters. The measured index values ​​of coal quality samples are obtained as standard reference parameters. The relative error value between the second prediction parameter and the standard reference parameter is calculated, and the relative error value is defined as the first error parameter. The first error parameter is compared with the preset error threshold parameter. If the first error parameter is less than or equal to the error threshold parameter, the second prediction parameter is determined as the final integrated prediction result of coal quality indicators. If the first error parameter is greater than the error threshold parameter, the self-correction iteration module is triggered to optimize the model parameters. After the self-correcting iteration module is started, the current model parameters of each basic prediction model are extracted as the first adjustment parameters. The current model parameters include model weights, bias terms and kernel function parameters. The prediction deviation of each basic prediction model is calculated as the second error parameter. The second error parameter is the difference between the first prediction parameter of each basic prediction model and the standard reference parameter. The first weight parameter is adaptively redistributed based on the second error parameter to obtain the updated second weight parameter. With minimizing the first error parameter as the optimization objective, an iterative optimization algorithm is used to update and correct the first adjustment parameters of each basic prediction model to obtain the second adjustment parameters, thus completing the parameter iteration of each basic prediction model.

8. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 7, characterized in that, After completing the parameter iteration of each basic prediction model, the compressed spectral feature dataset is re-input into each basic prediction model after updating the parameters to obtain the updated first prediction parameters. The first prediction parameters are weighted and fused based on the second weight parameters to obtain the updated second prediction parameters, and the corresponding first error parameters are recalculated. Multiple rounds of self-correcting iterative optimization are performed until the calculated first error parameter is less than or equal to the error threshold parameter. The iteration is then stopped, and the second prediction parameter at this point is determined as the final integrated prediction result of coal quality indicators. At the same time, the second weight parameter and the second adjustment parameter of this iteration are saved as the optimal parameters of the model.

9. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 8, characterized in that, The parameter adaptive adjustment process of the self-correcting iterative module includes: With minimizing model prediction error as the optimization objective, the pre-defined gradient solving algorithm in the self-correcting iterative module is extracted as the basic optimization algorithm. The current parameters to be adjusted of each basic prediction model are retrieved as the first parameter set. The first parameter set includes model weight parameters, bias term parameters and kernel function parameters, and each parameter corresponds to a unique parameter identifier. Based on the first error parameter, the loss function of each basic prediction model is constructed, and the partial derivative of the loss function with respect to each parameter in the first parameter set is solved to obtain the gradient change value corresponding to each parameter. All gradient change values ​​are integrated into a first gradient set, and the parameter identifiers of the first gradient set correspond one-to-one with the first parameter set. The momentum factor preset in the correction iteration module is retrieved as the first correction coefficient. Each gradient change value in the first gradient set is multiplied by the first correction coefficient to obtain the corrected second gradient set. The learning rate is introduced as a second correction coefficient, which is an adaptive dynamic value. The iterative update step size of each parameter is obtained by multiplying each gradient change value in the second gradient set with the second correction coefficient.

10. The coal quality near-infrared spectral correction, compression, self-correcting integrated prediction method as described in claim 9, characterized in that, After obtaining the iteration update step size of each parameter, subtract the corresponding iteration update step size from each parameter in the first parameter set to obtain the updated candidate parameter set, which is defined as the second parameter set. A preset proportion of validation set data is extracted from the spectral feature dataset as the first validation sample. The first validation sample is then replaced with each of the basic prediction models in the second parameter set to obtain the validation set prediction results, which are defined as the third prediction parameters. The measured coal quality index value corresponding to the first verification sample was retrieved as the second standard parameter. Calculate the verification error value between the third predicted parameter and the second standard parameter, define it as the third error parameter, and compare the third error parameter with the first error parameter before the parameter update; If the third error parameter is less than the first error parameter, the parameter update is deemed valid. The second parameter set is determined as the optimal parameter set for each basic prediction model, and the parameter iteration update is completed. At the same time, the current second correction coefficient is retained as the initial learning rate for the next iteration. If the third error parameter is greater than or equal to the first error parameter, the parameter update is deemed invalid, the second parameter set is abandoned and restored to the first parameter set, and the second correction coefficient is adaptively adjusted by multiplying it by the preset attenuation coefficient to obtain a new second correction coefficient. Based on the adjusted second correction coefficient, the parameter gradient is solved and iterated again until the obtained third error parameter is less than the first error parameter and the gradient change value of each parameter tends to 0. The model parameters are then determined to have converged to the optimal value, and the parameter set at this time is determined as the final updated parameter set of each basic prediction model.

Images