Coal testing system and method, and storage medium

Abstract

Provided are a coal testing system and method, and a storage medium. The system comprises: a detection unit, which is used for collecting spectral data of a coal sample under test; a training unit, which is used for constructing, by means of the spectral data, neural networks across neural network dimensions on the basis of a greedy search, and obtaining a coal sample test model by means of performing training on the basis of simulated samples which have been subjected to sample augmentation; and an analysis unit, which is used for executing fusion processing on the spectral data to obtain target spectral data, and on the basis of the coal sample test model, executing target spectral data inference to obtain a coal test result, wherein the detection unit comprises a near-infrared spectroscopy collection module and an X-ray fluorescence spectroscopy collection module.

Description

Coal detection system, method and storage medium

[0001] Cross-reference to related applications

[0002] The present disclosure claims priority to Chinese Patent Application No. 202410597882.8, titled “Coal detection system, method and storage medium” and filed on May 14, 2024; Chinese Patent Application No. 202410597885.1, titled “X-ray detection system and XRF ash measurement system comprising the same” and filed on May 14, 2024; Chinese Patent Application No. 202410597883.2, titled “Coal detection method and system” and filed on May 14, 2024; Chinese Patent Application No. 202410597886.6, titled “Coal detection pre-treatment device and coal detection system” and filed on May 14, 2024; and Chinese Patent Application No. 202410597884.7, titled “Coal quality rapid detection execution system” and filed on May 14, 2024, the entire contents of which are incorporated herein by reference. TECHNICAL FIELD

[0003] The present disclosure relates to the technical field of coal detection, and in particular to a coal detection system, a coal detection method and a storage medium. BACKGROUND

[0004] Coal online detection is an important means of coal detection, which has the significant advantages of real-time and efficient, all-weather continuous operation, and assisting digital management. The main difficulty of coal online detection is how to realize high-precision detection of multiple indexes such as coal calorific value, ash content, sulfur content and moisture content at the same time under complex and variable working conditions. The existing coal composition online analysis methods include near-infrared spectroscopy signal online technology (NIRS), laser-induced breakdown spectroscopy technology, X-ray fluorescence spectroscopy technology (XRF), neutron activation technology, and machine vision image method. The above-mentioned technologies can only realize effective detection of part or single index, and in order to realize comprehensive coal detection, various tests need to be executed at a single speed, which not only leads to a large number of repeated operations, such as sample preparation, but also leads to the problem of low accuracy of test results due to the differences in test standards between devices.

[0005] A common problem with current coal monitoring solutions is that each link in the process of coal sample preparation, transportation, and testing, as well as coal sample index monitoring, is carried out independently. This decentralized execution approach leads to inconsistencies in sample status and makes it impossible to achieve unified standard control. In particular, in coal online detection, factors such as the size and surface flatness of coal particles have a significant impact on the stability of the detection system. This independent execution of the monitoring solution leads to information silos in the monitoring process. Each executor only focuses on completing their own task and lacks overall control and coordination of the entire monitoring process. Due to the lack of unified standard control, there may be execution deviations in different links, affecting the reliability and consistency of the monitoring results. In coal online detection, the size and surface flatness of coal particles are crucial to the stability of the detection system. The size of coal particles not only affects the combustion characteristics and calorific value, but also directly affects the accuracy of the detection equipment. The surface flatness affects the detection effect of optical sensors, and uneven surfaces may cause uneven light reflection, affecting the accuracy of data collection. Therefore, understanding the characteristics of coal particles and surface flatness is crucial for optimizing the monitoring system. Only by comprehensively understanding the physical properties and surface morphology of coal samples can we effectively adjust the parameters and methods of the monitoring equipment to improve the stability and accuracy of the monitoring system. At the same time, establishing a unified standard control system to ensure the coordination and information sharing of each link is the key to improving the efficiency and accuracy of coal monitoring.

[0006] In existing solutions, such as patent CN114112976A "XRF-NIRS combined coal calorific value high repeatability detection method", a XRF-NIRS combined coal calorific value high repeatability detection method is disclosed, but this solution does not mention a solution to overcome the significant impact of coal particles, surface flatness, etc. on the stability of the XRF system and NIRS system in coal online detection, and it still uses a single index (coal calorific value) testing solution, which cannot achieve comprehensive integrated detection of multiple indexes. Therefore, this isolated detection method will inevitably cause problems such as low detection accuracy and low detection efficiency. To solve this problem, a new coal detection solution is needed.

[0007] Disclosure of Invention

[0008] The purpose of the embodiments of the present disclosure is to provide a coal detection system, method and storage medium to at least solve the problems of low detection accuracy and low detection efficiency in existing coal detection solutions.

[0009] In order to achieve the above object, the first aspect of the present disclosure provides a coal detection system, comprising: a detection unit configured to collect spectral data of a detection coal sample; wherein the detection unit comprises a near-infrared spectrum collection module and an X-ray fluorescence spectrum collection module; a training unit configured to construct a neural network in each neural network dimension based on a greedy search through the spectral data, and train a coal sample detection model based on simulated samples after sample augmentation; and an analysis unit configured to perform fusion processing on the spectral data to obtain target spectral data, perform inference on the target spectral data based on the coal sample detection model, and obtain a coal detection result.

[0010] Optionally, the system further comprises a sample preparation unit configured to collect a coal sample and perform sample preparation processing on the coal sample to obtain a detection coal sample.

[0011] Optionally, the sample preparation unit comprises: a sampling module configured to randomly collect a raw coal sample during raw coal transportation or storage, or configured to receive a raw coal sample input by a user; a crushing module configured to perform crushing processing on the raw coal sample to obtain a crushed coal sample; a processing module configured to perform pretreatment on the crushed coal sample to obtain a basic coal sample; and a shaping module configured to perform shaping processing on the basic coal sample to obtain a detection coal sample.

[0012] Optionally, the pretreatment on the crushed coal sample comprises drying processing, grinding processing, and screening processing.

[0013] Optionally, the sampling module, the crushing module, the processing module, and the shaping module are connected based on a transportation conveyor belt; and the transportation conveyor belt is triggered and controlled based on a servo system.

[0014] Optionally, the servo system is configured to: start timing based on a trigger signal of a position trigger at a preset position of each module of the sample preparation unit, and control a power servo motor of the transportation conveyor belt to start until a predetermined time is reached, and then turn off the power servo motor of the transportation conveyor belt; or start timing in response to a corresponding switch trigger signal of each module of the sample preparation unit, and control a power servo motor of the transportation conveyor belt to start until a predetermined time is reached, and then turn off the power servo motor of the transportation conveyor belt.

[0015] Optionally, the shaping module comprises: a coal conveying member comprising a conveying frame and a conveying belt rotatably arranged on the conveying frame, and a detection coal sample inlet is arranged on the conveying frame; a first pretreatment member comprising a limiting frame and a scraper, the limiting frame is supported above the conveying belt by the conveying frame, and the limiting frame is located below the detection coal sample inlet, the scraper is arranged at the discharging end of the limiting frame and forms a first gap with the conveying belt; and a second pretreatment member comprising at least one press roller, the press roller is rotatably arranged on the conveying frame and located at the outlet end of the limiting frame, and a second gap in communication with the first gap is formed between the press roller and the conveying belt, and a limiting piece for limiting the width of the second gap is arranged on the press roller.

[0016] Optionally, the limiting piece comprises two limiting wheels coaxially arranged on the press roller, and the two limiting wheels are spaced apart along the extension direction of the press roller to form the second gap therebetween.

[0017] Optionally, the inner side of the limiting wheel is provided with a slope surface.

[0018] Optionally, two mounting frames are spaced apart on the conveying frame along a direction perpendicular to the conveying direction of the conveying belt, a rotating shaft is arranged through the press roller, two ends of the rotating shaft are rotatably arranged on the two mounting frames respectively, and a first driving member for driving the rotating shaft to rotate is arranged on the conveying frame.

[0019] Optionally, when the press roller is a plurality of press rollers, one end of the rotating shaft of each of the plurality of press rollers away from the first driving member is provided with a sprocket, the plurality of sprockets are connected by a chain, and the first driving member is used to drive one of the rotating shafts to rotate; and / or two blocking wheels are further arranged on the rotating shaft, and a baffle is arranged on each of the two mounting frames, and the two blocking wheels are respectively supported on the inner sides of the two baffles.

[0020] Optionally, limiting plates are arranged at the output ends of the two side edges of the limiting frame, the limiting plates extend in an arc shape, the limiting plates are arranged opposite to the press roller, and the curvature of the limiting plates matches the curvature of the press roller.

[0021] Optionally, a protective cover is arranged above the second pretreatment member.

[0022] Optionally, a feeding hopper is arranged above the material inlet.

[0023] Optionally, a mounting position for mounting a detection module is arranged on the conveying frame, and the mounting position is located on the side of the second pretreatment member away from the first pretreatment member.

[0024] Optionally, the near-infrared spectrum acquisition module comprises an X-ray tube system, a detector system and a collimator, the collimator is a straight horn structure, one side of the hypotenuse of the collimator is arranged close to the detector system, and has inhibitory effect on X-rays of large angle emitted close to the detector system; one side of the straight angle of the collimator is arranged away from the detector system, and has no inhibitory effect on X-rays of large angle emitted away from the detector system.

[0025] Optionally, the collimator is integrally arranged on the X-ray tube of the X-ray tube system.

[0026] Optionally, the collimator is connected with the X-ray tube of the X-ray tube system through an X-ray tube interface.

[0027] Optionally, the detection unit further comprises a sampling cover for arranging the near-infrared spectrum acquisition module, the X-ray fluorescence spectrum acquisition module and the visual analysis module; the sampling cover is further provided with an illumination module, an optical fiber and a photoelectric sensor inside; a diaphragm extending from the inner wall of the sampling cover is arranged between the illumination module and the optical fiber; the photoelectric sensor is arranged in the light path direction of the illumination module, and is used for monitoring the illumination intensity of the corresponding illumination module.

[0028] Optionally, the illumination module comprises a plurality of symmetrically arranged light sources; the optical fiber is arranged at the center position of the top of the sampling cover; each illumination module is provided with at least one photoelectric sensor arranged on the inner wall of the sampling cover.

[0029] Optionally, the analysis unit is further used for monitoring the state of the corresponding illumination module based on the illumination intensity collected by the photoelectric sensor, comprising: performing filtering and denoising pretreatment on the illumination data collected by the photoelectric sensor; performing interpolation or fitting calculation on the calibration curve between the voltage value or digital value of the corresponding electrical signal of the pretreated illumination data and the illumination intensity, to obtain the illumination intensity of the detection position; performing illumination intensity fitting of the corresponding illumination module based on the illumination intensity of the detection position and the positional relationship between the light point sensor and the corresponding illumination module, to obtain the detection illumination intensity of the corresponding illumination module; comparing the detection illumination intensity with a preset illumination intensity threshold value, and if the detection illumination intensity is less than the preset illumination intensity threshold value, outputting an alarm information.

[0030] Optionally, the detection unit further comprises a Mylar film replacement device, comprising a sample conveying belt, the sample conveying belt is provided with a detection coal sample, an installation rack is arranged above the detection coal sample, a driving wheel and a driven wheel are arranged on the installation rack, a Mylar film transmission channel is formed on the installation rack, the Mylar film can pass through the Mylar film transmission channel and abut against the lower edge of the outer circumferential surface of the driving wheel and the driven wheel, and move under the driving of the driving wheel.

[0031] Optionally, the system further comprises a visual analysis module, the visual analysis module comprising one or more image acquisition modules, each image acquisition module being arranged at a preset direction of each detection position of the coal sample, and being configured to acquire image information of the coal sample at a view angle; the analysis unit is further configured to evaluate the shape of the coal sample based on the image information at each view angle.

[0032] Optionally, the evaluation of the shape of the coal sample based on the image information at each view angle comprises: sequentially performing denoising, grayscale and edge detection processing on the image information at each view angle, and calibrating the coal sample region; fitting the shape size of the coal sample based on the coal sample region at each view angle; collecting the surface flatness index of the coal sample region at each view angle, and fitting the surface flatness of the coal sample based on the surface flatness index; evaluating the shape of the coal sample based on the fitted shape size and the fitted surface flatness.

[0033] Optionally, the evaluation of the shape of the coal sample based on the fitted shape size and the fitted surface flatness comprises: calculating the Euclidean distance between the fitted shape size and the preset shape size as a first deviation value; calculating the Euclidean distance between the fitted surface flatness and the preset surface flatness as a second deviation value; performing normalization processing on the first deviation value and the second deviation value, and performing arithmetic average calculation on the normalized first deviation value and the second deviation value to obtain a deviation coefficient; assigning a preset correction coefficient in the coal sample detection model based on the deviation coefficient.

[0034] Optionally, the spectral data comprises near-infrared spectrum signal and X-ray fluorescence spectrum data; the fusion processing of the spectral data to obtain target spectral data comprises: performing fusion processing on the near-infrared spectrum signal collected by the near-infrared spectrum acquisition module and the X-ray fluorescence spectrum data collected by the X-ray fluorescence spectrum acquisition module to obtain target spectral data.

[0035] Optionally, the fusion processing of the near-infrared spectrum signal collected by the near-infrared spectrum acquisition module and the X-ray fluorescence spectrum data collected by the X-ray fluorescence spectrum acquisition module to obtain target spectral data comprises: performing preprocessing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively; performing down-sampling processing on the preprocessed near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively, and scaling the down-sampled near-infrared spectrum signal and the X-ray fluorescence spectrum data to a normal distribution; performing splicing processing on the normal distribution of the near-infrared spectrum signal and the normal distribution of the X-ray fluorescence spectrum data to obtain target spectral data.

[0036] Optionally, the pre-processing is performed on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively, including: performing SG convolution smoothing processing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively to obtain SG convolution smoothing processed near-infrared spectrum signal and SG convolution smoothing processed X-ray fluorescence spectrum data; and performing area normalization processing on the SG convolution smoothing processed near-infrared spectrum signal.

[0037] Optionally, the SG convolution smoothing processing includes: based on a preset window size and a polynomial order, calculating a coefficient of an SG convolution kernel; based on the coefficient of the SG convolution kernel, performing weighted average on adjacent data points in each spectrum signal to obtain smoothed data points; performing symmetric extension or zero padding processing on a boundary of each spectrum signal; and based on the smoothed data points and the processed boundary, obtaining the SG convolution smoothing processed spectrum signal.

[0038] Optionally, the area normalization processing on the Savitzky-Golay convolution smoothing processed near-infrared spectrum signal includes: performing area normalization processing on each Savitzky-Golay convolution smoothing processed near-infrared spectrum signal to obtain area normalization processed near-infrared spectrum signal.

[0039] Optionally, a calculation rule of the area normalization processed near-infrared spectrum signal is:

[0040] wherein, represents an NIR reflectivity corresponding to an mth wavelength point of the near-infrared spectrum signal;

[0041] X NIR is the near-infrared spectrum signal before the area normalization processing, is the area normalization processed near-infrared spectrum signal.

[0042] Optionally, the pre-processing is performed on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively, including: performing SG convolution smoothing processing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively to obtain SG convolution smoothing processed near-infrared spectrum signal and SG convolution smoothing processed X-ray fluorescence spectrum data; and performing area normalization processing on the SG convolution smoothing processed near-infrared spectrum signal.

[0043] Optionally, the near-infrared spectrum signal and the X-ray fluorescence spectrum data after down-sampling are scaled to a normal distribution, including: calculating statistical characteristics of each spectrum signal respectively; wherein the statistical characteristics are variance and / or standard deviation; performing standardization processing or normalization processing on each spectrum signal based on the statistical characteristics of each spectrum signal, and scaling the numerical value of each spectrum signal to a preset numerical range to obtain a scaled numerical value of each spectrum signal; and converting each spectrum signal to a normal distribution based on a preset transformation algorithm and the scaled numerical value of each spectrum signal.

[0044] Optionally, the near-infrared spectrum signal and the X-ray fluorescence spectrum data after down-sampling are scaled to a normal distribution, including: calculating statistical characteristics of each spectrum signal respectively; wherein the statistical characteristics are variance and / or standard deviation; performing standardization processing or normalization processing on each spectrum signal based on the statistical characteristics of each spectrum signal, and scaling the numerical value of each spectrum signal to a preset numerical range to obtain a scaled numerical value of each spectrum signal; and converting each spectrum signal to a normal distribution based on a preset transformation algorithm and the scaled numerical value of each spectrum signal.

[0045] Optionally, the pre-trained coal sample detection model is: θ=argmax θ L2(f(X|θ),Y);

[0046] wherein f(·|θ) is a deep neural network with θ as a parameter; is the target spectrum data; is the coal component.

[0047] Optionally, the system further includes a training unit for performing coal sample detection model training, including: collecting historical near-infrared spectrum signals and historical X-ray fluorescence spectrum data, and constructing corresponding historical target spectrum data based on the historical near-infrared spectrum signals and the historical X-ray fluorescence spectrum data; performing PLS model parameter initialization using the historical target spectrum data as training data; generating simulation samples based on the initialized PLS model, performing model training in a pre-constructed neural network based on the simulation samples, and obtaining a coal sample detection initial model; and performing verification on the coal sample detection initial model based on the reserved historical target spectrum data, and obtaining a coal sample detection model.

[0048] Optionally, the PLS model parameters include: weights, biases, mean and variance of training set input data, and mean and variance of output data.

[0049] Optionally, the generating the simulation sample based on the initialized PLS model comprises: adaptively determining a signal type and a signal parameter matched with the target spectral data, and generating a basic signal based on the signal type and the signal parameter; adaptively selecting an augmentation scheme, and performing a data augmentation operation on the basic signal based on the selected augmentation scheme to obtain an augmented signal set; and performing bias data screening on the augmented signal set, and taking the screened augmented signal set as the simulation sample.

[0050] Optionally, the adaptively selecting the augmentation scheme comprises: randomly selecting one or more schemes from among noise addition, translation and a preset rule as preselected schemes; and wherein the preset rule is: X aug = alpha * X1 + (1 - alpha) * X2, alpha is in [0, 1] aug = f PLS (X aug )

[0051] wherein f PLS represents a PLS mapping function; X aug and y aug respectively represent the synthesized NIRS-XRF signal and the corresponding pseudo label; alpha represents an interpolation weight randomly sampled in the range of [0, 1]; X1 and X2 are signal samples randomly selected from the historical target spectral data of the coal sample; the adjustable parameters in the preset range of each preselected scheme are randomly executed to obtain the parameter-determined preselected scheme; if there is only one scheme in the parameter-determined preselected scheme, the basic signal is directly processed based on the scheme to obtain an augmented signal; if there are multiple schemes in the parameter-determined preselected scheme, each scheme is executed in turn to obtain an augmented signal.

[0052] Optionally, the performing model training based on the simulation sample to obtain the initial model for detecting the coal sample comprises: performing pseudo label annotation on the simulation sample based on the PLS model after parameter initialization, and taking the annotated simulation sample as a training sample; and performing model training in the neural network searched based on the target spectral data based on the training sample to obtain the initial model for detecting the coal sample.

[0053] Optionally, the coal sample detection model is obtained by performing verification on the initial coal sample detection model based on the reserved historical target spectrum data, including: constructing a verification set based on the reserved historical target spectrum data, inputting the verification set into the initial coal sample detection model, and obtaining a corresponding prediction result; performing performance evaluation of the initial coal sample detection model based on the prediction result and an actual result corresponding to the reserved historical target spectrum data; if the performance evaluation of the initial coal sample detection model fails, adjusting the hyperparameters, optimizing the model structure and / or adding the regularization operation to obtain an updated initial coal sample detection model; and re-performing performance evaluation of the updated initial coal sample detection model based on the reserved historical target spectrum data until an initial coal sample detection model meeting a preset performance requirement is obtained as the coal sample detection model.

[0054] Optionally, the performance evaluation of the initial coal sample detection model is performed based on the prediction result and the actual result corresponding to the reserved historical target spectrum data, including: comparing the prediction result with the actual result, and evaluating the accuracy and the recall rate of the model based on the deviation of the two; if any one of the evaluation results of the accuracy and the recall rate fails, the performance evaluation of the initial coal sample detection model fails.

[0055] Optionally, the training unit is further configured to search a neural network based on the spectrum data, including: dividing a plurality of optimization dimensions based on the structure parameters of the neural network; in each optimization dimension, adaptively adjusting the parameters in the corresponding neural network structure of the corresponding optimization dimension, and calculating the MAE index after each adjustment; comparing the MAE index corresponding to each parameter to determine the parameter combination with the highest MAE index as the optimization result in the corresponding optimization dimension; performing greedy search between the optimization dimensions to obtain the optimization result of each optimization dimension; and determining the structure parameters of the corresponding neural network structure based on the optimization result of each optimization dimension to construct a neural network corresponding to the search result.

[0056] Optionally, the greedy search is performed between the optimization dimensions to obtain the optimization result of each optimization dimension, including: determining a search order and performing the greedy search between the optimization dimensions based on the determined search order to obtain the optimization result of each optimization dimension.

[0057] Optionally, the search order is search of the basic operator, search of the input resolution, search of the network depth, and search of the network width; and the determining of the search order and the performing of the greedy search among the optimization dimensions based on the determined search order to obtain the optimization result of each optimization dimension comprises: determining a search range of the basic operator, randomly generating N groups of different configurations of the input resolution, the network depth, and the network width, determining an optimal operator for each group of configurations, outputting N optimal operators from the N groups of configurations, and taking the operator with the highest frequency of occurrence in the N optimal operators as the searched basic operator; determining a search range of the input resolution, fixing the searched basic operator as the optimal configuration, randomly generating M groups of different configurations of the network depth and the network width, determining an optimal input resolution for each group of configurations, outputting M optimal input resolutions from the M groups of configurations, and taking the input resolution with the highest frequency of occurrence in the M optimal input resolutions as the searched input resolution; determining a search range of the network depth, fixing the searched basic operator and the searched input resolution as the optimal configuration, randomly generating L groups of different configurations of the network width, determining an optimal network depth for each group of configurations, outputting L optimal network depths from the L groups of configurations, and taking the network depth with the highest frequency of occurrence in the L optimal network depths as the searched network depth; and determining a search range of the network width, fixing the searched basic operator, the searched input resolution, and the searched network depth as the optimal configuration, and searching for the optimal network width.

[0058] Optionally, the neural network dimensions comprise a basic operator dimension of the neural network layer, a resolution dimension of the input data, a network depth dimension, and a network width dimension.

[0059] Optionally, the corresponding optimization result obtaining rule in the basic operator dimension of the neural network layer comprises: traversing the convolution type of the linear layer and the activation function type of the nonlinear layer, adaptively combining the convolution type of the fully connected layer and the activation function type of the nonlinear layer, and calculating the MAE index after each combination; comparing the MAE index after each combination, selecting the combination corresponding to the maximum MAE index, and determining the convolution type of the fully connected layer and the activation function type of the nonlinear layer corresponding to the combination as the optimization result in the basic operator dimension of the neural network layer.

[0060] Optionally, the convolution type of the linear layer is a fully connected layer, a 1D convolution layer with a kernel size of 5, a 1D convolution layer with a kernel size of 10, or a 1D convolution layer with a kernel size of 15; and the activation function type of the nonlinear layer is a TanH function, an ELU function, a Sigmoid activation function, or a Softmax function.

[0061] Optionally, the calculation rule of the MAE index is:

[0062] wherein, y is a true value of a coal component corresponding to the i th training data; i m is a model prediction value corresponding to the i th training data; and m is a size of a data set.

[0063] Optionally, the system further comprises an output unit configured to visualize the coal detection result.

[0064] Optionally, the detection results of the coal sample include one or more of ash component, ash content, volatile matter, carbon and hydrogen, ash melting point, total water, total sulfur, and calorific value.

[0065] Optionally, the visualizing of the coal detection result includes: in response to a user data query instruction, determining a corresponding detection result object; selecting a preset data visualization scheme based on the corresponding detection result object, and pushing the visualized data to a user terminal.

[0066] The second aspect of the present disclosure provides a coal detection method, which is applied to the coal detection system described above, and the method comprises: collecting spectrum data of a detection coal sample; constructing a neural network in each neural network dimension based on a greedy search through the spectrum data, and training a coal sample detection model based on simulation samples after sample augmentation; performing fusion processing on the spectrum data to obtain target spectrum data, and performing inference on the target spectrum data based on the coal sample detection model to obtain a coal detection result.

[0067] Optionally, before the spectrum data of the detection coal sample is collected, the method further comprises: collecting a raw coal sample, and performing sample preparation processing on the raw coal sample to obtain a detection coal sample.

[0068] Optionally, the sample preparation processing on the raw coal sample to obtain a detection coal sample includes: randomly collecting a raw coal sample during raw coal transportation or storage, or receiving a raw coal sample stored by a user; performing crushing processing on the raw coal sample to obtain a crushed coal sample; performing pretreatment on the crushed coal sample to obtain a basic coal sample; performing shaping processing on the basic coal sample to obtain a detection coal sample.

[0069] Optionally, the pretreatment on the crushed coal sample includes drying processing, grinding processing, and screening processing.

[0070] Optionally, the method further comprises: performing corresponding lighting module state monitoring, including: performing preprocessing including filtering and denoising on the illumination data collected by the photoelectric sensor; wherein each lighting module is at least corresponding to one photoelectric sensor arranged on the inner wall of the sampling cover; performing interpolation or fitting calculation on the calibration curve between the voltage value or digital value of the corresponding electrical signal of the preprocessed illumination data and the illumination intensity to obtain the illumination intensity of the detection position; based on the illumination intensity of the detection position and the positional relationship between the light point sensor and the corresponding lighting module, performing corresponding lighting module illumination intensity fitting to obtain the detection illumination intensity of the corresponding lighting module; comparing the detection illumination intensity with the preset illumination intensity threshold, if the detection illumination intensity is less than the preset illumination intensity threshold, outputting an alarm information.

[0071] Optionally, the method further comprises: collecting image data of the detected coal sample, and performing detection of the shape of the detected coal sample based on the image data of the detected coal sample.

[0072] Optionally, the method further comprises: collecting image data of the detected coal sample, and performing detection of the shape of the detected coal sample based on the image data of the detected coal sample.

[0073] Optionally, the method further comprises: collecting image data of the detected coal sample, and performing detection of the shape of the detected coal sample based on the image data of the detected coal sample.

[0074] Optionally, the spectral data includes: near-infrared spectrum signal and X-ray fluorescence spectrum data; the method further comprises: performing fusion processing on the near-infrared spectrum signal collected by the near-infrared spectrum acquisition module and the X-ray fluorescence spectrum data collected by the X-ray fluorescence spectrum acquisition module to obtain target spectral data; performing inference on the target spectral data based on the coal sample detection model with the assigned preset correction coefficient to obtain the inference result.

[0075] Optionally, the spectral data includes near-infrared spectral data and X-ray fluorescence spectral data; the step of performing fusion processing on the spectral data to obtain target spectral data includes: performing fusion processing on the near-infrared spectral signal and the X-ray fluorescence spectral data to obtain target spectral data.

[0076] Optionally, the near-infrared spectral signal acquired by the near-infrared spectral acquisition module and the X-ray fluorescence spectral data acquired by the X-ray fluorescence spectral acquisition module are fused to obtain target spectral data. This includes: performing preprocessing on the near-infrared spectral signal and the X-ray fluorescence spectral data respectively; performing downsampling on the preprocessed near-infrared spectral signal and the X-ray fluorescence spectral data respectively, and scaling the downsampled near-infrared spectral signal and the X-ray fluorescence spectral data to a normal distribution; and performing stitching on the normally distributed near-infrared spectral signal and the normally distributed X-ray fluorescence spectral data to obtain target spectral data.

[0077] Optionally, the preprocessing of the near-infrared spectral data and the X-ray fluorescence spectral data includes: performing SG convolution smoothing on the near-infrared spectral data and the X-ray fluorescence spectral data respectively to obtain SG convolution smoothed near-infrared spectral data and SG convolution smoothed X-ray fluorescence spectral data; and performing area normalization on the SG convolution smoothed near-infrared spectral data.

[0078] Optionally, the step of performing SG convolution smoothing on near-infrared spectral data and X-ray fluorescence spectral data includes: calculating the coefficients of the SG convolution kernel based on a preset window size and polynomial order; performing a weighted average of adjacent data points in each spectral data based on the coefficients of the SG convolution kernel for both near-infrared spectral data and X-ray fluorescence spectral data to obtain smoothed data points; performing symmetric expansion or zero-padding on the boundaries of each spectral data; and obtaining the SG convolution smoothed spectral data based on the smoothed data points and the processed boundaries.

[0079] Optionally, the area normalization processing of the near-infrared spectral signals after SG convolution smoothing includes: performing area normalization processing on each near-infrared spectral signal after SG convolution smoothing, and calculating the area-normalized near-infrared spectral signal.

[0080] Optionally, the calculation rule for the area-normalized near-infrared spectral signal is as follows:

[0081] in, This represents the NIR reflectance corresponding to the m-th wavelength point of the near-infrared spectral signal;

[0082] XNIR is an area-normalized near-infrared spectrum signal before processing, is an area-normalized near-infrared spectrum signal after processing.

[0083] Optionally, the pre-processed near-infrared spectrum data and the X-ray fluorescence spectrum data are respectively subjected to down-sampling processing, including: the SG convolution smoothed X-ray fluorescence spectrum data and the area-normalized near-infrared spectrum data are respectively subjected to down-sampling processing, including: each spectrum data is respectively subjected to filtering processing, and in the spectrum data after filtering processing, every fixed interval retains one sampling point to obtain a plurality of sampling points; or, a plurality of segments are cut from the spectrum data after filtering processing, and each sampling point in each segment is subjected to average processing, and each segment obtains one sampling point to obtain a plurality of sampling points; signal reconstruction is performed based on each sampling point to obtain spectrum data after down-sampling processing.

[0084] Optionally, the down-sampled near-infrared spectrum data and the X-ray fluorescence spectrum data are scaled to a normal distribution, including: statistical characteristics of each spectrum data are respectively calculated; wherein the statistical characteristics are variance and / or standard deviation; based on the statistical characteristics of each spectrum data, each spectrum data is subjected to standardization processing or normalization processing to scale the numerical value of each spectrum data to a preset numerical range to obtain a scaled numerical value of each spectrum data; based on a preset transformation algorithm and the scaled numerical value of each spectrum data, each spectrum data is converted to a normal distribution.

[0085] Optionally, the normal distribution of the near-infrared spectrum data and the normal distribution of the X-ray fluorescence spectrum data are subjected to splicing processing to obtain target spectrum data, including: alignment operation of the normal distribution of the near-infrared spectrum data and the normal distribution of the X-ray fluorescence spectrum data is performed; after the alignment of the normal distribution of the near-infrared spectrum signal and the normal distribution of the X-ray fluorescence spectrum data is completed, the splicing of the two spectrum signals is performed based on weighted average to obtain initial target spectrum data; the initial target spectrum data is verified based on the normal distribution of the near-infrared spectrum signal and / or the normal distribution of the X-ray fluorescence spectrum data, and the initial target spectrum data that passes the verification is taken as the target spectrum data.

[0086] Optionally, the coal sample detection model is: θ=argmax θ L2(f(X|θ),Y);

[0087] Wherein, f(·|θ) is a deep neural network with θ as a parameter;

[0088] is target spectrum data;

[0089] is a coal component.

[0090] Optionally, the method further comprises: performing coal sample detection model training, including: collecting historical near-infrared spectrum signal and historical X-ray fluorescence spectrum data, and constructing corresponding historical target spectrum data based on the historical near-infrared spectrum signal and the historical X-ray fluorescence spectrum data; performing PLS model parameter initialization by taking the historical target spectrum data as training data; generating a simulation sample based on the initialized PLS model, performing model training in a pre-constructed neural network based on the simulation sample, and obtaining a coal sample detection initial model; performing verification on the coal sample detection initial model based on reserved historical target spectrum data, and obtaining a coal sample detection model.

[0091] Optionally, the PLS model parameters include: weight, bias, mean and variance of training set input data, and mean and variance of output data.

[0092] Optionally, the generating a simulation sample based on the initialized PLS model comprises: adaptively determining a signal type and signal parameters matched with the target spectrum data, generating a basic signal based on the signal type and the signal parameters; adaptively selecting an augmentation scheme, and performing a data augmentation operation on the basic signal based on the selected augmentation scheme to obtain an augmented signal set; screening out bias data in the augmented signal set, and taking the screened out augmented signal set as a simulation sample.

[0093] Optionally, the adaptively selecting an augmentation scheme comprises: randomly selecting one or more schemes in noise addition, translation and preset rules as pre-selected schemes; wherein, the preset rule is: X aug = αX1 + (1-α)X2, α ∈ [0, 1] y aug = f PLS (X aug )

[0094] Wherein, f PLS represents a PLS mapping function; X aug and y aug respectively represent the synthesized NIRS-XRF signal and the corresponding pseudo label; α represents an interpolation weight randomly sampled in the range of [0, 1]; X1 and X2 are signal samples randomly selected from the historical target spectrum data of the coal sample; the adjustable parameters in each pre-selected scheme are randomly adjusted in the preset range to obtain the parameter-determined pre-selected scheme; if there is only one scheme in the parameter-determined pre-selected scheme, the basic signal is directly processed based on the scheme to obtain an augmented signal; if there are multiple schemes in the parameter-determined pre-selected scheme, each scheme is sequentially executed to obtain an augmented signal.

[0095] Optionally, the model training based on the simulation sample is performed to obtain a coal sample detection initial model, including: performing pseudo-label annotation on the simulation sample based on the PLS model after parameter initialization, taking the simulation sample after annotation as a training sample; performing model training in a neural network based on the target spectrum data search to obtain the coal sample detection initial model.

[0096] Optionally, the coal sample detection initial model is verified based on the reserved historical target spectrum data to obtain a coal sample detection model, including: constructing a verification set based on the reserved historical target spectrum data, inputting the verification set into the coal sample detection initial model to obtain a corresponding prediction result; performing coal sample detection initial model performance evaluation based on the prediction result and the actual result of the corresponding reserved historical target spectrum data; if the coal sample detection initial model performance evaluation fails, adjusting the hyperparameters, optimizing the model structure and / or adding the regularization operation to obtain an updated coal sample detection initial model; re-performing performance evaluation on the updated coal sample detection initial model based on the reserved historical target spectrum data until a coal sample detection initial model meeting the preset performance requirement is obtained as the coal sample detection model.

[0097] Optionally, the coal sample detection initial model performance evaluation based on the prediction result and the actual result of the corresponding reserved historical target spectrum data includes: comparing the prediction result and the actual result, and evaluating the accuracy and recall rate of the model based on the deviation of the two; if any one of the accuracy and recall rate evaluation results fails, the coal sample detection initial model performance evaluation fails.

[0098] Optionally, the pre-construction rule of the neural network includes: dividing a plurality of optimization dimensions based on the structure parameters of the neural network; in each optimization dimension, the parameters in the corresponding neural network structure are adjusted adaptively, and the MAE index after each adjustment is calculated based on the target spectrum data; comparing the MAE index corresponding to each parameter, determining the parameter combination with the highest MAE index as the optimization result in the corresponding optimization dimension; performing greedy search between each optimization dimension to obtain the optimization result of each optimization dimension; determining the structure parameters corresponding to each neural network structure based on the optimization result of each optimization dimension, and constructing the neural network corresponding to the search result.

[0099] Optionally, the greedy search between each optimization dimension to obtain the optimization result of each optimization dimension includes: determining a search order and performing greedy search between each optimization dimension based on the determined search order to obtain the optimization result of each optimization dimension.

[0100] Optionally, the search order is search of the basic operator, search of the input resolution, search of the network depth, and search of the network width; and the determining of the search order and the performing of the greedy search among the optimization dimensions based on the determined search order to obtain the optimization result of each optimization dimension comprises: determining a search range of the basic operator, randomly generating N groups of different configurations of the input resolution, the network depth, and the network width, determining an optimal operator for each group of configurations, outputting N optimal operators from the N groups of configurations, and taking the operator with the highest frequency of occurrence in the N optimal operators as the searched basic operator; determining a search range of the input resolution, fixing the searched basic operator as the optimal configuration, randomly generating M groups of different configurations of the network depth and the network width, determining an optimal input resolution for each group of configurations, outputting M optimal input resolutions from the M groups of configurations, and taking the input resolution with the highest frequency of occurrence in the M optimal input resolutions as the searched input resolution; determining a search range of the network depth, fixing the searched basic operator and the searched input resolution as the optimal configuration, randomly generating L groups of different configurations of the network width, determining an optimal network depth for each group of configurations, outputting L optimal network depths from the L groups of configurations, and taking the network depth with the highest frequency of occurrence in the L optimal network depths as the searched network depth; and determining a search range of the network width, fixing the searched basic operator, the searched input resolution, and the searched network depth as the optimal configuration, and searching for the optimal network width.

[0101] Optionally, the neural network dimensions comprise a basic operator dimension of the neural network layer, a resolution dimension of the input data, a network depth dimension, and a network width dimension.

[0102] Optionally, the corresponding optimization result obtaining rule in the basic operator dimension of the neural network layer comprises: traversing the convolution type of the linear layer and the activation function type of the nonlinear layer, adaptively combining the convolution type of the fully connected layer and the activation function type of the nonlinear layer, and calculating the MAE index after each combination; comparing the MAE index after each combination, selecting the combination corresponding to the maximum MAE index, and determining the convolution type of the fully connected layer and the activation function type of the nonlinear layer corresponding to the combination as the optimization result in the basic operator dimension of the neural network layer.

[0103] Optionally, the convolution type of the linear layer is a fully connected layer, a 1D convolution layer with a kernel size of 5, a 1D convolution layer with a kernel size of 10, or a 1D convolution layer with a kernel size of 15; and the activation function type of the nonlinear layer is a TanH function, an ELU function, a Sigmoid activation function, or a Softmax function.

[0104] Optionally, the calculation rule of the MAE index is:

[0105] wherein, is the true value of the coal composition corresponding to the i-th training data;

[0106] y i is the model prediction value corresponding to the i-th training data;

[0107] m is the size of the data set.

[0108] Optionally, the system further comprises an output unit for visualizing the coal detection result.

[0109] Optionally, one or more of ash composition, ash content, volatile matter, hydrocarbon, ash melting point, total water, total sulfur, and calorific value.

[0110] Optionally, the method further comprises visualizing the coal detection result, including: in response to a user data query instruction, determining a corresponding detection result object; based on the corresponding detection result object, selecting a preset data visualization scheme, and pushing the visualized data to the user end.

[0111] In another aspect, the present disclosure provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, cause the computer to perform the coal detection method described above.

[0112] Through the above technical solutions, the present disclosure has the following beneficial effects:

[0113] 1. Improved detection accuracy: using multiple detection technologies such as near-infrared spectrum acquisition module, X-ray fluorescence spectrum acquisition module, and visual analysis module, the system can comprehensively and multi-angulary detect coal samples, improving the accuracy and accuracy of detection.

[0114] 2. Improved detection efficiency: through automated sample preparation and data acquisition, the system can quickly and efficiently complete the detection process of coal samples, saving time and labor costs, and improving detection efficiency.

[0115] 3. Optimized training model: the analysis unit performs data training based on the pre-trained coal sample detection model, which can continuously optimize the model and improve the detection capability and adaptability of the system.

[0116] 4. Result information output: the output unit determines the detection result information of the detected coal sample based on the inference result, so that users can timely obtain and analyze the detection data, providing strong support for decision-making.

[0117] Other features and advantages of the embodiments of the present disclosure will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0118] The accompanying drawings are included to provide a further understanding of embodiments of the disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and serve to explain the principles of the disclosure, but do not limit the disclosure. In the drawings:

[0119] Fig. 1 is a system structure diagram of a coal detection system according to an embodiment of the disclosure;

[0120] Fig. 2 is a system structure diagram of a coal detection system including a sample preparation module according to an embodiment of the disclosure;

[0121] Fig. 3 is a structure diagram of one specific embodiment of a shaping module according to an embodiment of the disclosure;

[0122] Fig. 4 is a structure diagram of a second pretreatment member according to an embodiment of the disclosure;

[0123] Fig. 5 is an enlarged view of A in Fig. 4;

[0124] Fig. 6 is a structure diagram of a first pretreatment member according to an embodiment of the disclosure;

[0125] Fig. 7 is a structure diagram of a mounting mode of a compression roller according to an embodiment of the disclosure;

[0126] Fig. 8 is a structure diagram of a compression roller according to an embodiment of the disclosure;

[0127] Fig. 9 is a structure diagram of the compression roller connected with a first driving member according to an embodiment of the disclosure;

[0128] Fig. 10 is a structure diagram of one specific embodiment of a detection unit according to an embodiment of the disclosure;

[0129] Fig. 11 is a near-infrared spectrum acquisition module according to an embodiment of the disclosure;

[0130] Fig. 12 is a principle diagram of an X-ray tube according to an embodiment of the disclosure;

[0131] Fig. 13 is a structure diagram of one specific embodiment of a Mylar film replacement device according to an embodiment of the disclosure;

[0132] Fig. 14 is a structure diagram of one specific embodiment of a buffer adjustment structure according to an embodiment of the disclosure;

[0133] Fig. 15 is a system structure diagram of a coal detection system including an output module according to an embodiment of the disclosure;

[0134] Fig. 16 is a step flow chart of a coal detection method according to an embodiment of the disclosure;

[0135] FIG. 17 is a step flow chart of a coal detection method including a sample preparation step according to an embodiment of the present disclosure;

[0136] FIG. 18 is a schematic diagram of a sample preparation process for detecting a coal sample according to an embodiment of the present disclosure;

[0137] FIG. 19 is a step flow chart of a spectrum signal fusion process according to an embodiment of the present disclosure;

[0138] FIG. 20 is a step flow chart of a coal detection method including an output step according to an embodiment of the present disclosure.

[0139] The following table provides a description of the reference numerals in the drawings: 1, coal conveying member; 11, conveying frame; 111, leg; 112, mounting frame; 1121, baffle; 113, second driving member; 12, conveying belt; 13, feeding hopper; 14, mounting position; 2, first pretreatment member; 21, limiting frame; 211, limiting plate; 22, scraper; 3, second pretreatment member; 31, compression roller; 311, rotating shaft; 3111, chain wheel; 3112, chain; 3113, blocking wheel; 3114, synchronization shaft; 32, limiting member; 321, limiting wheel; 33, first driving member; 34, protective cover; 35, support; 4, sampling cover; 41, illumination module; 42, optical fiber; 43, photoelectric sensor; 44, diaphragm; 5, sample conveying belt; 6, detected coal sample; 7, mounting frame; 73, Mylar film conveying channel; 71, driving wheel; 72, driven wheel; 8, Mylar film; 9, buffer adjustment structure; 901, fixed shaft; 902, longitudinal threaded shaft; 903, lower fixed sleeve; 904, upper threaded sleeve; 905, compression spring; 101, positive pressure blowing device; 102, X-ray generating device; 103, scanning device; 104, energy receiving device. DETAILED DESCRIPTION

[0140] The specific embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely intended to illustrate and explain the present disclosure, and are not intended to limit the present disclosure.

[0141] FIG. 1 is a system structure diagram of a coal detection system according to an embodiment of the present disclosure. As shown in FIG. 1, the present disclosure provides a coal detection system, which includes a detection unit configured to collect spectral data of a detected coal sample; wherein the detection unit includes a near-infrared spectrum acquisition module and an X-ray fluorescence spectrum acquisition module; a training unit configured to construct a neural network in each neural network dimension based on a greedy search through the spectral data, and train a coal sample detection model based on simulated samples after sample augmentation; and an analysis unit configured to perform fusion processing on the spectral data to obtain target spectral data, perform inference on the target spectral data based on the coal sample detection model, and obtain a coal detection result.

[0142] Preferably, as shown in FIG. 2, the system further comprises a sample preparation unit for collecting a coal sample and performing sample preparation processing on the coal sample to obtain a detection coal sample. The sample preparation unit comprises: a sampling module for randomly collecting a raw coal sample during raw coal transportation or storage, or for receiving a raw coal sample stored by a user; a crushing module for performing crushing processing on the raw coal sample to obtain a crushed coal sample; a processing module for performing pretreatment on the crushed coal sample to obtain a basic coal sample; wherein the pretreatment on the crushed coal sample comprises: drying processing, grinding processing and screening processing; and a shaping module for performing shaping processing on the basic coal sample to obtain a detection coal sample.

[0143] In one possible implementation, in the coal industry, sampling is a crucial link that directly affects the accuracy and reliability of subsequent coal quality analysis and utilization. The present disclosure proposes an intelligent sampling device that combines sensing technology and automatic control systems to achieve intelligent identification and collection of raw coal samples. By mounting high-precision sensors and image recognition technology, it can monitor the coal flow during raw coal transportation or storage in real time and randomly sample according to preset algorithms and rules. Using data analysis and simulation technology, the selection of sampling locations is optimized to ensure the representativeness and comprehensiveness of the sampling points. By analyzing the coal flow path and speed, the best sampling location is determined to avoid sampling bias and local problems and improve the reliability of sampling. A real-time monitoring and feedback mechanism is introduced to monitor key parameters such as sampling volume and sampling frequency during the sampling process and adjust the sampling strategy in a timely manner. Through data collection and analysis, real-time monitoring and quality control of the sampling process are achieved to ensure the accuracy and reliability of the sampling results. An automated sampling system is designed to achieve automatic collection and processing of raw coal samples. Machine learning and artificial intelligence technologies are combined to optimize sampling algorithms and processes, improve sampling efficiency and automation, and reduce human intervention and errors. The design of the intelligent sampling device and the optimized sampling location of the present disclosure can improve the accuracy and representativeness of sampling, reduce sampling errors and biases, and improve the reliability of sampling results. The introduction of a real-time monitoring and feedback mechanism can optimize the sampling process, improve sampling efficiency and speed, and save labor and time costs. The design of an automated sampling system can reduce human intervention, reduce operational risks and errors, and improve the controllability and stability of the sampling process.

[0144] In a possible implementation, in the coal processing process, crushing is an essential link, but at the same time, it also produces a large amount of dust pollution. In the crushing module, high-efficiency dust removal equipment such as a bag filter or an electric dust collector is integrated to capture and filter dust particles generated in the crushing process. Through equipment integration, timely removal and treatment of dust in the crushing process are realized, and the cleanliness of the production environment and the health of the employees are ensured. Wet dust removal technology is adopted, and through a spraying system or a wet scrubbing device, dust is wetted and treated during the crushing process to reduce the diffusion and flying of dust. Wet dust removal technology can effectively control dust emission and reduce the impact on air quality. A negative pressure closed structure of the crushing module is designed, and the flow of dust in the closed space is controlled through a negative pressure system to reduce dust leakage and diffusion. The negative pressure closed design can effectively prevent dust pollution and protect the surrounding environment and equipment. An online monitoring and control system is introduced to monitor the dust emission concentration and particle size distribution in real time, and the operation parameters and cleaning cycle of the dust removal equipment are adjusted according to the monitoring data. Through real-time monitoring and control, the stability and efficiency of the dust removal effect are ensured. The disclosed scheme effectively controls dust emission and reduces environmental pollution by integrating dust removal equipment and applying wet dust removal technology. The negative pressure closed design and the introduction of the online monitoring system in the disclosed scheme can protect the health of employees, reduce the harm of dust to the respiratory tract and skin, and improve the comfort of the working environment.

[0145] In a possible implementation, the disclosed scheme introduces a sample grading and screening device in the processing module to finely screen and process the sample according to the particle size and shape characteristics of the crushed coal sample. Through grading and screening, basic coal samples of different particle size ranges can be obtained to provide more accurate and representative samples for subsequent detection. The disclosed scheme designs a chemical treatment reaction tank for chemical treatment and reaction of the crushed coal sample to extract target components or remove interfering substances. The chemical treatment reaction tank can change the chemical properties of the coal sample to make it more suitable for subsequent detection and analysis, thereby improving the accuracy of the detection results. The disclosed scheme introduces a magnetic separation device to separate magnetic impurities or magnetic minerals in the crushed coal sample. The magnetic separation device can effectively remove interfering substances and purify the basic coal sample to improve the accuracy and reliability of the detection. A drying and dehumidifying system is designed to dry and dehumidify the processed basic coal sample to ensure the dryness and stability of the sample. The drying and dehumidifying system can avoid the influence of moisture on the detection results and ensure the accuracy and reliability of the detection data.

[0146] Preferably, the sampling module, the crushing module, the processing module, and the shaping module are connected based on a transfer conveyor belt; and the transfer conveyor belt is triggered and controlled based on a servo system.

[0147] Preferably, the servo system is configured to start timing based on the triggering signal of the position trigger at the preset position of each module of the sample preparation unit, and control the power servo motor of the transfer conveyor belt to start, until a predetermined time is reached, and the power servo motor of the transfer conveyor belt is turned off; or in response to the corresponding switch trigger signal of each module of the sample preparation unit, start timing, and control the power servo motor of the transfer conveyor belt to start, until a predetermined time is reached, and the power servo motor of the transfer conveyor belt is turned off.

[0148] Specifically, intelligent sensors such as infrared sensors are installed at the preset positions of each module of the sample preparation unit to detect the position of the sample and trigger corresponding operations. The sensors can monitor the sample position in real time and send signals to the servo system to start corresponding operations. According to the preset position trigger or switch trigger signal of each module of the sample preparation unit, the servo system starts timing and controls the power servo motor of the transfer conveyor belt to start. According to the predetermined time, the corresponding operation such as sampling, crushing, processing or shaping is performed, and then the power servo motor of the transfer conveyor belt is turned off. The control logic of the servo system is optimized to enable flexible adjustment of the transfer speed and dwell time of the conveyor belt according to the operation requirements of different modules, to ensure smooth and efficient completion of each operation link. The disclosed scheme realizes automatic transfer and operation execution of samples between modules, reduces manual intervention, and improves sample processing efficiency and consistency. The servo system accurately controls the start and stop of the conveyor belt according to the preset position trigger signal to ensure accurate operation and transfer of samples in each module. By optimizing the control logic, flexible adjustment of the conveyor belt speed and dwell time is realized to adapt to different operation requirements, improving the applicability and flexibility of the system.

[0149] Preferably, as shown in FIGS. 3-6, the shaping module includes a coal conveying member 1, a first pretreatment member 2, and a second pretreatment member 3.

[0150] The coal conveying member 1 is composed of a conveying frame 11 and a rotating conveying belt 12 installed thereon. The conveying frame 11 is provided with a detection coal sample inlet and has a cuboid structure. The detection coal sample is transported along the length direction of the conveying frame 11, and the bottom is provided with multiple supporting legs to ensure the supporting effect. The detection coal sample enters the conveying frame 11 through the detection coal sample inlet and directly falls into the conveying belt 12, and then is transported towards the outlet. In addition, the conveying frame 11 is also provided with a second driving member for driving the conveying belt 12 to rotate. The specific rotating mode is a conventional technology, and therefore will not be described in detail.

[0151] The first pretreatment component 2 is composed of a limiting frame 21 and a scraper 22. The limiting frame 21 is supported above the conveying belt 12 by the conveying frame 11 and is located below the detection coal sample inlet, so that the detection coal sample can smoothly fall into the limiting frame 21. The scraper 22 is located between the discharge end of the limiting frame 21 and the conveying belt 12 to form a first gap. The conveying frame 11 is provided with two mounting plates, and the two ends of the limiting frame 21 are connected to the two mounting plates to support the limiting frame 21 above the conveying belt 12. The limiting frame 21 and the conveying belt 12 are spaced apart in the vertical direction, and the gap therebetween is small, which prevents the detection coal sample from flowing out of the limiting frame 21 from the side, while not affecting the rotation of the conveying belt 12, and ensures the smooth conveying of the detection coal sample. The limiting frame 21 adopts a rectangular frame structure, including two side plates extending along the direction of the conveying belt 12. The gap between the side plates can be designed according to requirements to meet the preliminary shaping requirements of the width of the detection coal sample. The gap between the scraper 22 and the conveying belt 12 above the discharge end of the two side plates also meets the preliminary shaping requirements.

[0152] By arranging the limiting frame 21 and the conveying belt 12, the limiting frame 21 is located directly below the detection coal sample inlet of the conveying frame 11, so that the detection coal sample poured into the conveying frame 11 can fall onto the conveying belt 12, thereby limiting the coal sample within the limiting frame 21, and the rotating conveying belt 12 can convey the coal sample within the limiting frame 21. The scraper 22 is arranged on the limiting frame 21. When the detection coal sample located in the limiting frame 21 is brought into contact with the scraper 22 by the conveying belt 12, the scraper 22 can scrape the detection coal sample flat on the conveying belt 12. The thickness of the detection coal sample is equal to the gap between the scraper 22 and the conveying belt 12, and the width of the coal sample is equal to the width inside the limiting frame 21.

[0153] The second pretreatment component 3 includes at least one compression roller 31 arranged on the conveying frame 11 along a direction perpendicular to the conveying direction of the conveying belt 12 and located at the outlet end of the limiting frame 21. The compression roller 31 and the conveying belt 12 form a second gap connected to the first gap, so that the detection coal sample preliminarily shaped by the limiting frame 21 and the scraper 22 can enter between the compression roller 31 and the conveying frame 11, thereby limiting the thickness of the detection coal sample within a predetermined range. The compression roller 31 is provided with a limiting piece 32 for limiting the width of the second gap to ensure that the width of the detection coal sample is within a predetermined range.

[0154] The coal detection pretreatment device provided by the present disclosure can preliminarily shape the coal sample during the conveying process of the coal sample by the first pretreatment component 2 to limit the width and height of the coal sample within a certain range, and then secondarily shape the preliminarily shaped coal sample by the second pretreatment component 3 to roll the coal sample, so that the height and height of the rolled coal sample reach a predetermined range, and the surface of the coal sample is smooth, so that the coal sample meets the detection requirements, improves the stability of the spectral data of the coal sample, and improves the detection accuracy.

[0155] In some embodiments, referring to FIGS. 4 and 6, the limiting member 32 includes two limiting wheels 321 coaxially arranged on the compression roller 31, and the two limiting wheels 321 are arranged at intervals along the extension direction of the compression roller 31 to form a second gap between the two limiting wheels 321.

[0156] Specifically, the two limiting wheels 321 are arranged outside the two side plates respectively, so that the preliminarily shaped material flowing through the first gap can enter the second gap, and then the material is secondarily shaped by the compression roller 31. The two limiting wheels 321 can rotate synchronously with the compression roller 31, and the distance between the two limiting wheels 321 can be designed according to actual requirements to meet the requirements of material rolling and shaping.

[0157] The limiting member 32 has a simple design, and the material processed by the compression roller 31 can meet the preset requirements.

[0158] Further optimization, as shown in FIG. 7, the inner side of the limiting wheel 321 is provided with a slope surface.

[0159] Specifically, the middle part of the limiting wheel 321 is inwardly convex, so that the inner side of the limiting wheel 321 forms a circular truncated cone, and the side wall of the circular truncated cone forms a slope surface. The limiting wheel 321 in this design can guide the material entering the second gap, so as to facilitate the material entering the second gap.

[0160] Referring to FIGS. 4, 6 and 7, two mounting frames 112 are arranged at intervals on the conveying frame 11 along a direction perpendicular to the conveying direction of the conveying belt 12, a rotating shaft 311 is arranged on the compression roller 31, both ends of the rotating shaft 311 are rotatably arranged on the two mounting frames 112, and the conveying frame 11 is provided with a first driving member 33 for driving the rotating shaft 311 to rotate.

[0161] Specifically, the rotating shaft 311 can coaxially rotate the compression roller 31, both ends of the rotating shaft 311 are mounted on the two mounting frames 112 through the support 35, and the end of the rotating shaft 311 is connected with the support 35 through a bearing, so as to ensure the smoothness of the rotation of the rotating shaft 311. The support 35 includes a fixed frame arranged on the mounting frame 112 and a bearing seat arranged on the fixed frame, and the bearing is arranged on the bearing seat. The fixed frame can position the mounting position of the bearing seat, so that the mounting hole on the bearing seat can correspond to the mounting hole on the mounting frame 112, and then the external bolt can be accurately screwed into the mounting hole of the bearing seat and the mounting frame 112, thereby increasing the convenience of installation.

[0162] The first driving member 33 can be a driving motor arranged on the conveying frame 11. An output shaft of the driving motor is connected with the rotating shaft 311, so as to drive the rotating shaft 311 to rotate through the driving motor, and then drive the compression roller 31 and the limiting wheel 321 to rotate, so as to make the device run stably. The driving motor is a variable frequency motor. It can be understood that the driving motor can also be other forms of motor, which can be designed according to actual needs.

[0163] In some embodiments, as shown in FIGS. 4 to 7, when the compression roller 31 is multiple, the end of the rotating shaft 311 of each compression roller 31 away from the first driving member 33 is provided with a sprocket wheel 3111. The sprocket wheels 3111 are connected through a chain. The first driving member 33 is used to drive one rotating shaft 311 to rotate.

[0164] Specifically, the multiple compression rollers 31 are arranged at intervals along the conveying direction of the conveying belt 12. The end of the rotating shaft 311 away from the first driving member 33 extends out of the support 35, and the extending end of the rotating shaft 311 is connected with the sprocket wheel 3111. The sprocket wheel 3111 rotates synchronously with the rotating shaft 311. The sprocket wheels 3111 are connected through the chain, so that the rotation of one rotating shaft 311 can drive the other rotating shafts 311 to rotate synchronously through the cooperation of the sprocket wheels 3111 and the chain, thereby reducing the number of driving devices and the structural cost. FIG. 4 shows a design mode in which the compression roller 31 is two. The extending ends of the two rotating shafts 311 are connected with the sprocket wheels 3111. The two sprocket wheels 3111 are connected through the chain, that is, the chain is engaged with the sprocket wheels 3111.

[0165] Preferably, the compression roller 31 is two. The two compression rollers 31 are arranged at intervals along the conveying direction of the conveying belt 12. For the convenience of description, the compression roller 31 arranged close to the limiting frame 21 is referred to as the first compression roller 31, and the compression roller 31 arranged away from the limiting frame 21 is referred to as the second compression roller 31. The height of the second gap between the first compression roller 31 and the conveying belt 12 is greater than the height of the second gap between the second compression roller 31 and the conveying belt 12.

[0166] As a feasible embodiment, the height of the second gap between the first compression roller 31 and the conveying belt 12 is 3.5 cm, and the width of the second gap (the distance between the two limiting wheels 321) is 10 cm. The height of the second gap between the second compression roller 31 and the conveying belt 12 is 3 cm, and the width of the second gap (the distance between the two limiting wheels 321) is 10 cm. This design mode can compact the thickness of the material to 3 cm and limit the width of the material to 10 cm through the two compression rollers 31.

[0167] In the design, the first compression roller 31 is used to receive the preliminary shaped material delivered from the limiting frame 21. Since the amount of the delivered material is relatively large, the thickness of the first gap between the first compression roller 31 and the conveying belt 12 is relatively large, so as to compress the preliminary shaped material into a coal sample with a thickness of 3.5 cm and a width of 10 cm. Then, the second compression roller 31 is used to compress the material into a coal sample with a thickness of 3 cm and a width of 10 cm, so as to ensure the compaction effect of the material.

[0168] FIG. 8 schematically shows the compression roller 31 not connected with the first driving member 33, and FIG. 9 schematically shows the compression roller 31 connected with the first driving member 33. Specifically, the end of the rotating shaft 311 of the compression roller 31 connected with the first driving member 33 is provided with a synchronous shaft, which is used to be connected with the output shaft of the first driving member 33, so that the output shaft of the first driving member 33 can drive the rotating shaft 311 to rotate synchronously through the synchronous shaft. In order to ensure that the synchronous shaft can rotate synchronously with the output shaft of the first driving member 33, a spline can be inserted at the connection between the synchronous shaft and the output shaft of the first driving member 33.

[0169] In some embodiments, as shown in FIGS. 6 and 7, the rotating shaft 311 is further provided with two blocking wheels 3113, and the two mounting frames 112 are each provided with a baffle 1121. The two blocking wheels 3113 are supported on the inner sides of the two baffles 1121, respectively.

[0170] Specifically, as shown in FIGS. 8 and 9, the two blocking wheels 3113 on the rotating shaft 311 are arranged on the outer sides of the two limiting wheels 321, respectively, and the blocking wheels 3113 can rotate synchronously with the rotating shaft 311, or the blocking wheels 3113 are rotationally matched with the rotating shaft 311, which can be designed according to actual requirements. As shown in FIGS. 6 and 7, one end of the mounting frame 112 is connected with the conveying frame 11, so as to support the other end of the mounting frame 112 above the conveying belt 12. The mounting frame 112 is arranged in a spaced manner with the conveying belt 12, so as to limit the position of the conveying belt 12 in the vertical direction through the mounting frame 112, thereby avoiding the phenomenon of edge lifting or arching of the conveying belt 12, and ensuring the conveying effect of the conveying belt 12. The baffle 1121 is arranged on the mounting frame 112, and the inner side of the baffle 1121 is in contact with the outer side of the blocking wheel 3113, so as to limit the movement of the rotating shaft 311 along the axial direction thereof through the two baffles 1121 and the two blocking wheels 3113, thereby limiting the position of the compression roller 31 perpendicular to the conveying direction of the conveying belt 12, so that the compression roller 31 can stably receive and compress the material.

[0171] In some embodiments, as shown in FIG. 4, the output ends of the two side edges of the limiting frame 21 are each provided with a limiting plate 211. The limiting plate 211 extends in an arc shape. The limiting plate 211 is arranged opposite to the compression roller 31, and the curvature of the limiting plate 211 matches the curvature of the compression roller 31.

[0172] Specifically, the ends of the two side plates of the limiting frame 21 are respectively provided with limiting plates 211. The limiting plates 211 extend in an arc shape to the bottom of the pressure roller 31 to limit the material flowing out of the limiting frame 21, ensuring that the material can be conveyed to the bottom of the pressure roller 31. The arc of the limiting plate 211 matches the arc of the pressure roller 31, so that the gap between the arc plate and the pressure roller 31 is small, preventing the material from leaving the second gap and meeting the material conveying requirements.

[0173] As shown in Figure 3, in some embodiments, a protective cover is provided above the second pretreatment component 3. Specifically, when there are two pressure rollers 31, both the first pressure roller 31 and the second pressure roller 31 are disposed inside the protective cover, and the protective cover is located between the first drive component 33 and the chain. By providing a protective cover, external dust is less likely to contaminate the material pressed by the first pressure roller 31 and the second pressure roller 31, and the dust generated by the first pressure roller 31 and the second pressure roller 31 when pressing the material is less likely to spread to the environment around the device.

[0174] Referring to Figures 3 and 4, in some embodiments, a feed hopper 13 is provided above the material inlet, wherein the top of the feed hopper 13 extends upward beyond the conveyor frame 11, and the inlet of the feed hopper 13 can be connected to the equipment outlet in the previous process, so that the material can enter the conveyor frame 11 through the feed hopper 13. In order to increase the sealing of the connection between the feed hopper 13 and the equipment outlet in the previous process, a flange can be provided at the top of the feed hopper 13.

[0175] The conveyor frame 11 is provided with a mounting position 14 for installing a detection module. The mounting position 14 is located on the side of the second pretreatment component 3 opposite to the first pretreatment component 2. The detection module can be a rapid coal detection module, which is installed on the mounting position 14. The mounting position 14 has multiple mounting holes and detection ports. The mounting holes are used to install the rapid coal detection module, and during operation, the rapid coal detection module can detect coal samples with a width of 10cm and a height of 3cm on the conveyor belt 12 through the detection ports. The rapid coal detection module is a conventional technology in the field of coal detection; therefore, its structure and working principle are not described in detail here.

[0176] During operation, the rapid coal detection module is installed at mounting position 14. Coal samples enter the conveyor frame 11 through the feed hopper 13 and fall onto the portion of the conveyor belt 12 opposite to the limiting frame 21. The coal sample is then transported via the conveyor belt 12, and during transport, it is flattened by the scraper 22 to obtain a sample of a certain width and thickness. The flattened coal sample is then sequentially conveyed to two pressure rollers 31, allowing it to be pressed into a preset size (3cm in height and 10cm in width). The resulting coal sample has a smooth surface, meeting the detection requirements of the rapid coal detection module, improving the stability of the coal sample spectral data, and increasing detection accuracy.

[0177] As shown in FIG. 10, in some embodiments of the present application, the X-ray fluorescence spectrum acquisition module includes an X-ray tube system, a detector system, and a collimator, the collimator is a right-angle horn structure, one side of the hypotenuse of the collimator is arranged close to the detector system, and has a suppression effect on the large-angle X-ray exiting close to the detector system; the other side of the right-angle edge of the collimator is arranged away from the detector system, and has no suppression effect on the large-angle X-ray exiting away from the detector system.

[0178] Through the above implementation, the power supply supplies power to the X-ray tube and excites X-rays, the collimator adopts a right-angle horn structure, one side of the hypotenuse of which can suppress the large-angle X-ray exiting from one end of the X-ray fluorescence detector, and the other side of the right-angle edge has no suppression effect on the large-angle X-ray exiting away from one end of the X-ray fluorescence detector, which can reduce the divergence angle of the X-ray at one end of the detector, avoid the X-ray scattering to the surface of the detector, and at the same time retain the X-ray on the other side, increase the X-ray fluorescence generation area, and increase the sensing area of the detection surface.

[0179] It can be seen that the collimator with the specific structure can only constrain the large-angle X-ray close to the detector, without constraining the large-angle X-ray away from the detector, avoiding the direct incidence of the stray light of the X-ray tube into the detector while ensuring a larger sensing field on the surface of the sample, and the increased sensing field of the X-ray helps to reduce the influence of the surface topography of the to-be-measured sample on the measurement.

[0180] FIG. 11 is a schematic diagram of an X-ray tube according to an embodiment of the present disclosure. As shown in FIG. 11, the target material of the X-ray tube is selected to be a metal target material, after the cathode electron bombards the metal target material, the energy spectrum generated is emitted from the window, and the emitted X-ray has a certain divergence angle. When the divergence angle of the X-ray is large, part of the X-ray close to the detector is easy to scatter into the detector, forming a strong background in the energy spectrum, and a collimator is needed to filter out the large-angle X-ray. In general, a straight-cylinder collimator is used, which has the characteristics of filtering out all large-angle X-rays and retaining only small-angle X-rays. Although this structure avoids large-angle stray light, it also reduces the irradiation area on the coal surface, i.e., reduces the sensing field of the X-ray. The present scheme suppresses the large-angle X-ray close to the detector and retains the large-angle X-ray away from the detector, avoiding the disadvantage of small sensing field of the straight-cylinder collimator, suppressing the stray light on the side of the detector, and increasing the sensing field of the system on the coal surface, thereby improving the robustness of the X-ray fluorescence detector.

[0181] In a possible implementation, the collimator is integrally arranged on the X-ray tube of the X-ray tube system. This implementation mainly considers integrating the symmetrical collimator into the X-ray tube, which has the advantages of high integration and facilitating the miniaturization of the X-ray tube.

[0182] Optionally, the collimator is connected to the X-ray tube of the X-ray tube system through an X-ray tube interface. Another connection mode is shown in Figure 10, which connects the collimator to the X-ray tube through the X-ray tube interface. This connection mode is beneficial to the adjustment of the collimator.

[0183] Preferably, as shown in Figure 12, the detection unit further comprises a sampling cover 4 for setting the near-infrared spectrum acquisition module, the X-ray fluorescence spectrum acquisition module and the visual analysis module; the sampling cover is further provided with an illumination module 41, an optical fiber 42 and a photoelectric sensor 43 inside; the illumination module 41 is provided with a diaphragm 44 extending out of the inner wall of the sampling cover between the illumination module 41 and the optical fiber 42; the photoelectric sensor 43 is arranged in the light path direction of the illumination module 41 and is used for monitoring the illumination intensity of the corresponding illumination module 41.

[0184] In the embodiments of the present disclosure, the present scheme also tests the reflectivity of coal, i.e., sets the corresponding optical fiber to recycle the reflected light of the sample. Generally, the reflectivity of coal is tested by using a halogen lamp as a light source for illumination. The light source will be attenuated for a long time, and the reflectivity of coal is low, so the accuracy of the reflectivity test is easily affected by stray light and other factors. In order to improve the test accuracy and reduce the influence of stray light, the present scheme proposes a corresponding detection module improvement, i.e., a diaphragm is arranged between the illumination module and the optical fiber to block stray light from entering the optical fiber.

[0185] Preferably, the illumination module 41 comprises a plurality of symmetrically arranged light sources; the optical fiber 4 is arranged at the center position of the top of the sampling cover; each illumination module 41 is provided with at least one photoelectric sensor 43 arranged on the inner wall of the sampling cover.

[0186] Further, the analysis unit is further used for monitoring the state of the corresponding illumination module based on the illumination intensity collected by the photoelectric sensor, comprising: performing filtering and denoising preprocessing on the illumination data collected by the photoelectric sensor; performing interpolation or fitting calculation on the calibration curve between the voltage value or digital value of the corresponding electrical signal of the preprocessed illumination data and the illumination intensity to obtain the illumination intensity of the detection position; performing illumination intensity fitting of the corresponding illumination module based on the illumination intensity of the detection position and the positional relationship between the light point sensor and the corresponding illumination module to obtain the detection illumination intensity of the corresponding illumination module; comparing the detection illumination intensity with the preset illumination intensity threshold value, and if the detection illumination intensity is less than the preset illumination intensity threshold value, outputting an alarm information.

[0187] In the embodiments of the present disclosure, it has been explained above that the reflectivity test of general coal is usually illuminated by a halogen lamp as a light source. The light source will be attenuated after a long time of lighting. In the existing scheme, the light source life needs to be judged by the relevant personnel, and the light source needs to be replaced when it needs to be replaced. In this case, it is highly dependent on the subjective experience of personnel. Therefore, the present scheme proposes a corresponding illumination intensity monitoring scheme. The light intensity of the light source is monitored in real time by a photoelectric sensor. When the light intensity of the light source does not meet the requirements, an alarm information is output to remind the relevant personnel to replace the light source.

[0188] Preferably, as shown in FIG. 13, the detection unit further comprises a Mylar film replacement device, which comprises a sample conveying belt 5 on which a detection coal sample 6 is placed, an installation rack 7 is arranged above the detection coal sample 6, a driving wheel 71 and a driven wheel 72 are arranged on the installation rack 7, a Mylar film transmission channel 73 is formed on the installation rack 7, and the Mylar film 8 can pass through the Mylar film transmission channel 73 and abut against the lower edge of the outer circumferential surface of the driving wheel 71 and the driven wheel 72, and move under the driving of the driving wheel 71.

[0189] In the embodiments of the present disclosure, in the existing online coal component analysis process, the Mylar film of the X-ray fluorescence spectrum acquisition module (radio frequency signal light transmission module) is easily contaminated during transmission and detection. In the existing use process of the Mylar film, the degree of contamination is usually judged by manual operation, so as to determine whether the Mylar film needs to be replaced. The working efficiency is low, and the replacement cost is high. Based on this, the present scheme proposes a corresponding Mylar film replacement device.

[0190] Specifically, the two ends of the sample conveying belt 5 are respectively provided with a conveying belt driving wheel and a conveying belt driven wheel. The sample conveying belt 5 conveys the detection coal sample 6 through the conveying belt driving wheel. The conveying belt driving wheel can be connected with a motor to provide power for the conveying belt driving wheel. Of course, since the sample conveying belt 5 is relatively soft, a supporting plate can be arranged below the sample conveying belt 5 to support the sample conveying belt 5, so as to avoid that multiple detection coal samples 6 placed on the sample conveying belt 5 cause the sample conveying belt 5 to deform.

[0191] In addition, in the present disclosure, the installation rack 7 is provided with the driving wheel 71 and the driven wheel 72, and the driving wheel 71 is connected with a motor to provide power for the driving wheel 71. The installation rack 7 is provided with the Mylar film transmission channel 71, which is suitable for the Mylar film 8 to pass through. It can be conceived that the Mylar film transmission channel 71 can be provided with a protection structure to prevent the Mylar film 8 from being damaged. The Mylar film 8 abuts against the driving wheel 71 and the driven wheel 72, and the driven wheel 72 can flatten the Mylar film 8 to provide a certain tension.

[0192] In a possible implementation, as shown in FIG. 14, the driving wheel 71 is provided with a buffer adjusting structure 9, and the driven wheel 72 is also provided with a buffer adjusting structure 9. The two buffer adjusting structures 9 can be provided with the same structure, thereby reducing the production cost. The single buffer adjusting structure 9 comprises a fixed shaft 901 and a longitudinal threaded shaft 902 connected with or integrated with the fixed shaft 901. The outer circumferential surface of the longitudinal threaded shaft 902 is formed with external threads. The longitudinal threaded shaft 902 is sleeved with a lower fixed sleeve 903 and an upper threaded sleeve 904. The lower fixed sleeve 903 is spaced apart from the longitudinal threaded shaft 902. The upper threaded sleeve 904 is threadedly connected with the longitudinal threaded shaft 902. Of course, a locking nut can be further arranged above the upper threaded sleeve 904 to prevent the upper threaded sleeve 904 from loosening during use. Of course, other anti-loosening measures can also be adopted, such as using thread glue. It is conceivable that a compression spring 905 is arranged between the lower fixed sleeve 903 and the upper threaded sleeve 904, thereby providing the driving wheel 71 and the driven wheel 72 with a buffer force and a downward pressure.

[0193] In a possible implementation, bearings are arranged between the fixed shaft 901 and the driving wheel 71 and between the fixed shaft 901 and the driven wheel 72.

[0194] In one specific embodiment of the present disclosure, the two ends of the fixed shaft 901 are provided with bearings, and the outer rings of the bearings are attached to the inner circumferential surfaces of the inner holes of the driving wheel 71 or the driven wheel 72. Specifically, the two ends of the fixed shaft 901 are provided with steps, one side of the step is attached to one side of the inner ring of the bearing, and the fixed shaft 901 is further provided with an elastic shaft collar, which is attached to the outer side of the inner ring of the bearing to prevent the bearing from moving left and right. Of course, other structures can also be adopted to prevent the bearing from moving left and right, which also belongs to the protection scope of the present disclosure.

[0195] As a preferred embodiment of the present disclosure, the central region of the lower fixed sleeve 903 is formed with a through hole, and the central region of the upper threaded sleeve 904 is formed with a threaded inner hole, which is matched with the longitudinal threaded shaft 902.

[0196] In one specific embodiment of the present disclosure, the central region of the lower fixed sleeve 903 is formed with a through hole, so that the longitudinal threaded shaft 902 can move freely in the lower fixed sleeve 903. However, the longitudinal threaded shaft 902 can move freely in the lower fixed sleeve 903 on the premise that the lower fixed sleeve 903 has a fixed support. During the upward and downward movement of the longitudinal threaded shaft 902, the lower fixed sleeve 903 is always in a fixed state and will not move with the upward and downward movement of the longitudinal threaded shaft 902. Therefore, as a preferred embodiment of the present disclosure, the lower fixed sleeve 903 is fixedly connected with the mounting frame 7.

[0197] As a preferred embodiment of the present disclosure, the outer circumferential surface of the driving wheel 71 and the driven wheel 72 is provided with a silica gel protective layer.

[0198] As a specific embodiment of the present disclosure, the driving wheel 71 comprises a driving wheel shaft and a silica gel protective layer arranged on the outer circumferential surface of the driving wheel shaft, and the central region of the driving wheel shaft is provided with a mounting through hole for mounting a bearing and a fixing shaft 901.

[0199] As a preferred embodiment of the present disclosure, the Mylar film automatic replacement device further comprises a positive pressure blowing device 101, an X-ray generating device 102, a scanning device 103 and an energy receiving device 104, the positive pressure blowing device 101 is arranged above the operation window of the mounting frame 7, and the X-ray generating device 102, the scanning device 103 and the energy receiving device 104 are arranged above the positive pressure blowing device 101.

[0200] In a specific embodiment of the present disclosure, a control system is further included, and the control system is respectively connected with signals of the positive pressure blowing device 101, the X-ray generating device 102, the scanning device 103, the energy receiving device 104 and the motor of the driving wheel 71.

[0201] As a preferred embodiment of the present disclosure, the positive pressure blowing device 101, the X-ray generating device 102, the scanning device 103 and the energy receiving device 104 are fixedly connected on the mounting frame 7 through a mounting bracket.

[0202] As a preferred embodiment of the present disclosure, the surface of the sample conveying belt 5 is provided with an anti-skid layer or an anti-skid structure.

[0203] In a most preferred embodiment of the present disclosure, a sample conveying belt 5 is included, a detection coal sample 6 is placed on the sample conveying belt 5, a mounting frame 7 is arranged above the detection coal sample 6, a driving wheel 71 and a driven wheel 72 are arranged on the mounting frame 7, a motor is connected to the driving wheel 71, a Mylar film conveying channel 71 and an operation window penetrating the upper and lower thicknesses of the mounting frame 7 are formed on the mounting frame 7, the Mylar film 8 can pass through the Mylar film conveying channel 71, the operation window is arranged corresponding to the detection coal sample 6, a buffer adjusting structure 9 is arranged on the driving wheel 71 and the driven wheel 72, the buffer adjusting structure 9 comprises a fixing shaft 901 and a longitudinal threaded shaft 902 connected with the fixing shaft 901, a lower fixing sleeve 903 and an upper threaded sleeve 904 are arranged on the longitudinal threaded shaft 902, a compression spring 905 is arranged between the lower fixing sleeve 903 and the upper threaded sleeve 904, bearings are arranged between the fixing shaft 901 and the driving wheel 71 and between the fixing shaft 901 and the driven wheel 72, elastic retaining rings for shafts are further arranged at both ends of the fixing shaft 901, and silica gel protective layers are arranged on the outer circumferential surfaces of the driving wheel 71 and the driven wheel 72.

[0204] Preferably, the system further comprises a visual analysis module, the visual analysis module comprising one or more image acquisition modules, each image acquisition module being arranged in a preset direction corresponding to a detection position of the coal sample, and being configured to acquire image information of the coal sample at a view angle; the analysis unit is further configured to perform shape evaluation on the coal sample based on the image information at each view angle, including: performing denoising, grayscale and edge detection processing on the image information at each view angle in sequence, and calibrating the coal sample region; fitting the shape size of the coal sample based on the image information at each view angle; collecting the surface flatness index of the coal sample at each view angle, and fitting the surface flatness of the coal sample based on the surface flatness index; and performing shape evaluation on the coal sample based on the fitted shape size and the fitted surface flatness.

[0205] In the embodiments of the present disclosure, the shape of the coal sample may have the following effects on the result of coal component monitoring based on the spectral signal:

[0206] 1) Light path scattering effect: The irregular shape or rough surface of the coal sample may cause scattering of the light path, resulting in deflection or scattering of the light signal during transmission, affecting the collection and accuracy of the spectral signal.

[0207] 2) Light absorption difference: Different shapes of the coal sample may result in different light absorption capabilities of different parts, thereby affecting the intensity and characteristics of the spectral signal. For example, the size, shape and density differences of the particles on the surface of the coal may affect the transmission and absorption of light.

[0208] 3) Surface reflection effect: The light reflection characteristics of the surface of the coal sample may affect the collection and analysis of the spectral signal. Different shapes of the coal sample may have different reflectivities, resulting in different reflection and diffuse reflection of the light signal on the sample surface, affecting the quality and accuracy of the spectral signal.

[0209] 4) Spectral signal interference: The unevenness of the shape of the coal sample may cause interference of the spectral signal, for example, impurities, oxides or moisture on the surface of the sample may affect the purity and stability of the spectral signal, thereby affecting the accuracy of the component monitoring.

[0210] 5) Light path length difference: Different sizes and shapes of the coal sample may result in differences in the length of the light path, causing different degrees of attenuation and absorption of the light signal during transmission in the sample, affecting the intensity and clarity of the spectral signal.

[0211] In the embodiments of the present disclosure, the shape of the coal sample has a certain effect on the result of coal component monitoring based on the spectral signal, so when performing component monitoring, the shape factor needs to be considered and appropriately corrected and processed to improve the accuracy and reliability of the monitoring result.

[0212] Although the corresponding detection coal sample shape limiting scheme is proposed in the present disclosure, in order to ensure that the shape shaping conforms to the expectation, the shaping result needs to be detected. The present disclosure will also collect visual signal while collecting spectrum signal, such as image information, and based on the image information, the target detection coal sample shape recognition is carried out to judge whether it conforms to the expectation. If it does not conform to the expectation, an alarm instruction needs to be output to remind the relevant personnel to replace the sample to avoid obtaining the wrong detection result.

[0213] Preferably, the surface flatness recognition of the detection coal sample based on the visual image data comprises: sequentially performing denoising, graying and edge detection processing on the visual image to demarcate the detection coal sample area; extracting the features related to the surface flatness in the detection coal sample area; performing surface flatness index recognition on the features related to the surface flatness to determine the surface flatness of the detection coal sample based on the recognition result. The features related to the surface flatness include one or more of texture features, color features and edge features. The surface flatness index includes average gray level and / or texture uniformity.

[0214] Specifically, the following steps are included:

[0215] 1) Data collection and preprocessing: Collect the visual image data of the detection coal sample to ensure the image definition and consistency. Perform denoising processing on the visual image to eliminate the influence of noise on the subsequent processing. Convert the processed image into a gray image to simplify the processing process and highlight the surface features.

[0216] 2) Surface flatness recognition processing: Perform edge detection processing to accurately demarcate the boundary and contour of the detection coal sample. Demarcate the detection coal sample area to ensure that the subsequent processing is focused on the sample surface. Extract the features related to the surface flatness in the detection coal sample area, including texture features, color features and edge features. Feature extraction and surface flatness index recognition:

[0217] 3) Analyze and process the extracted features to determine the features related to the surface flatness. Based on one or more of the texture features, color features and edge features, calculate the surface flatness index such as average gray level and texture uniformity.

[0218] 4) Surface flatness recognition and result determination: Determine the surface flatness level of the detection coal sample according to the recognition result. Combine the surface flatness index and feature analysis result to evaluate and classify the surface flatness of the detection coal sample.

[0219] In the present disclosure, the surface flatness recognition of the detection coal sample based on the visual image data can be effectively applied to quality control and production process monitoring in the coal industry to improve production efficiency and product quality management level.

[0220] Specifically, the shape evaluation of the detected coal sample based on the fitted shape size and the fitted surface flatness comprises: calculating the Euclidean distance between the fitted shape size and the preset shape size as a first deviation value; calculating the Euclidean distance between the fitted surface flatness and the preset surface flatness as a second deviation value; performing normalization processing on the first deviation value and the second deviation value, and performing arithmetic average calculation on the normalized first deviation value and the second deviation value to obtain a deviation coefficient; and assigning a preset correction coefficient in the coal sample detection model based on the deviation coefficient.

[0221] In the embodiments of the present disclosure, the detection coal sample with certain differences in shape can also be fused in the subsequent detection model through the deviation correction coefficient, so that even if there are certain differences in the detection coal sample, the unified standard processing can be realized through the deviation correction coefficient.

[0222] Preferably, the analysis unit is configured to: perform the spectrum data reasoning based on the coal sample detection model after the assignment to obtain a reasoning result, comprising: performing fusion processing on the near-infrared spectrum signal collected by the near-infrared spectrum acquisition module and the X-ray fluorescence spectrum data collected by the X-ray fluorescence spectrum acquisition module to obtain target spectrum data; and performing target spectrum data reasoning based on the coal sample detection model after the assignment to obtain a reasoning result.

[0223] Specifically, the near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively preprocessed; the near-infrared spectrum signal and the X-ray fluorescence spectrum data after the preprocessing are respectively subjected to down-sampling processing, and the near-infrared spectrum signal and the X-ray fluorescence spectrum data after the down-sampling are scaled to a normal distribution; the near-infrared spectrum signal in the normal distribution and the X-ray fluorescence spectrum data in the normal distribution are subjected to splicing processing to obtain target spectrum data. Specifically, the method comprises the following steps:

[0224] Step 1: The near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively preprocessed.

[0225] Specifically, the near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively subjected to SG convolution smoothing processing; and the near-infrared spectrum signal after the convolution smoothing processing is subjected to area normalization processing.

[0226] Preferably, the SG convolution smoothing is respectively performed on the near-infrared spectrum signal and the X-ray fluorescence spectrum data, including: based on a preset window size and a polynomial order, coefficients of an SG convolution kernel are obtained; based on the coefficients in the convolution kernel, adjacent data points in each spectrum signal are respectively weighted and averaged, so as to obtain smoothed data points; symmetric extension or zero padding processing is performed on the boundaries of each spectrum signal; and based on the smoothed data points and the processed boundaries, the spectrum signal subjected to the SG convolution smoothing is obtained.

[0227] In the embodiments of the present disclosure, the SG convolution smoothing is a linear smoothing method, which can better smooth the data curve, eliminate the peaks and fluctuations in the data, and make the data more stable and continuous by locally fitting the data. The SG convolution smoothing can effectively smooth the noise part of the data while retaining the characteristics and trends of the signal, which helps to reduce the interference of high-frequency noise in the data on the signal and improve the readability and analysis accuracy of the data. Compared with other smoothing methods, the SG convolution smoothing can better maintain the overall shape and trend of the data while smoothing the data, without causing distortion or deviation of the data shape, and retains the original characteristics of the data. The SG convolution smoothing is a simple and efficient data smoothing method, which has a faster calculation speed and is suitable for processing large-scale data sets, can complete the data smoothing operation in a short time, and is suitable for large-scale detection of coal samples, which ensures the detection accuracy while improving the detection efficiency.

[0228] Further, the area normalization processing is performed on the near-infrared spectrum signal subjected to the convolution smoothing, including: a signal set of the near-infrared spectrum signal subjected to the convolution smoothing is determined; the signal set includes a plurality of signal samples, and the reflectivity of each signal sample under a corresponding spectrum type at a corresponding wavelength point; and a signal subjected to the area normalization processing is calculated based on the signal set of the near-infrared spectrum signal.

[0229] In the embodiments of the present disclosure, the near-infrared spectrum signal is prone to baseline drift, so the present disclosure further eliminates the baseline drift in the near-infrared spectrum signal through area normalization. Baseline drift is a signal deviation caused by instrument drift, environmental changes, or sample differences, which affects the accuracy and stability of the spectrum signal. Through area normalization processing, the signal can be shifted upward or downward as a whole, so that the baseline level is more stable. The optical path difference in the near-infrared spectrum signal can cause differences in signal intensity, affecting the comparison and analysis of the signal. The area normalization processing can normalize the overall intensity of the signal, reduce the influence of the optical path difference on the signal, and make the comparison between different samples more accurate. The area normalization processing can highlight the characteristic peaks or troughs in the near-infrared spectrum signal, making the characteristics of the signal more obvious and prominent, which helps to more accurately identify and analyze specific components or characteristics in the spectrum.

[0230] Specifically, the calculation rule of the area-normalized signal is as follows:

[0231] wherein, Rnm represents the NIR reflectance corresponding to the mth wavelength point of the nth sample; is the area-normalized signal.

[0232] Step 2: Perform downsampling processing on the preprocessed near-infrared spectrum signal and X-ray fluorescence spectrum data respectively.

[0233] Specifically, filter processing is performed on each spectrum signal respectively, and in the spectrum signal after the filter processing, every fixed interval retains one sampling point to obtain a plurality of sampling points; or a plurality of segments are intercepted from the spectrum signal after the filter processing, and average processing is performed on each sampling point in each segment to obtain one sampling point in each segment, thereby obtaining a plurality of sampling points; signal reconstruction is performed based on each sampling point to obtain a spectrum signal after downsampling processing.

[0234] In the embodiments of the present disclosure, the downsampling processing can reduce the data amount, that is, reduce the sampling rate or the number of sampling points, thereby saving storage space and computing resources. In particular, for large-scale data sets or high-frequency sampled data, downsampling can effectively reduce the volume of data, facilitating storage and processing. Downsampling can simplify the data analysis process, reduce the complexity and dimensionality of data, making the data more easily understood and processed. By reducing the resolution of the data, some detailed information can be removed, highlighting the main features of the data, simplifying the model establishment and analysis process. Downsampling processing can help remove noise and interference in the data, smooth the data curve, and improve the quality and stability of the data. By reducing the sampling rate or averaging the sampling points, the volatility of the data can be reduced, making the data clearer and more reliable. The present disclosure scheme can reduce the complexity and dimensionality of the data by retaining interval sampling points and performing multi-segment average processing, thereby reducing the data amount and making the data more easily processed and analyzed. In addition, the noise and interference in the signal are reduced, and interval sampling point retention and average processing can further smooth the signal curve, reduce the volatility of the data, and improve the signal quality and stability.

[0235] Step 3: Scale the near-infrared spectrum signal and X-ray fluorescence spectrum data after downsampling to a normal distribution.

[0236] Specifically, the statistical characteristics of each spectrum signal are calculated; wherein the statistical characteristics are variance and / or standard deviation; based on the statistical characteristics of each spectrum signal, standardization processing or normalization processing is performed on each spectrum signal, the numerical value of each spectrum signal is scaled to a preset numerical range, and the scaled numerical value of each spectrum signal is obtained; based on a preset transformation algorithm and the scaled numerical value of each spectrum signal, each spectrum signal is converted to a normal distribution.​

[0237] In the embodiments of the present disclosure, the present scheme converts the signal into a normal distribution, which can simplify the data analysis process and make the data easier to understand and process. By converting the signal into a normal distribution, the fitting effect and prediction accuracy of the model can be improved, and converting the signal into a normal distribution can better meet the requirements of hypothesis testing and ensure the effectiveness of the test results. Moreover, the normal distribution has the characteristics of standardization, i.e., the mean is 0 and the standard deviation is 1, and converting the signal into a normal distribution can standardize the data, making different signals comparable and facilitating subsequent splicing of two spectral signals.

[0238] Step 4: performing splicing processing on the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data to obtain target spectrum data.

[0239] Specifically, the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data are aligned, and after the alignment of the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data is completed, the two spectral signals are spliced based on weighted average to obtain initial target spectrum data, and the initial target spectrum data is verified based on the normal distribution near-infrared spectrum signal and / or the normal distribution X-ray fluorescence spectrum data, and the initial target spectrum data that passes the verification is taken as the target spectrum data.

[0240] Further, the alignment operation of the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data includes: selecting the same reference points in the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data, and calculating the time difference between the corresponding reference points and other data points in each spectrum signal to obtain a time difference sequence; and performing time axis adjustment of one of the spectrum signals based on the random interpolation method and the corresponding time difference sequence until the time axes of the two spectrum signals are aligned.

[0241] Further, the random interpolation method is: X aug = αX1 + (1-α)X2, α ∈ [0, 1] y aug = f PLS (X aug )

[0242] wherein, f PLS represents a PLS mapping function; X aug and y aug represent the synthesized NIRS-XRF signal and the corresponding pseudo label, respectively; α represents an interpolation weight randomly sampled in the range of [0, 1]; and X1 and X2 are signals randomly selected from the historical NIRS-XRF signal of the coal sample.

[0243] In the embodiments of the present disclosure, the present scheme performs two spectral signal splicing based on a random interpolation method, which can preserve the characteristics and information of the original data during the signal splicing process, avoiding data loss or distortion. By splicing the signals through interpolation, the integrity of the data can be better maintained. The random interpolation method can achieve smooth transition in the transition area of signal splicing, avoiding sudden changes or discontinuities, which helps to improve the continuity and smoothness of signal splicing. The random interpolation method can adjust the interpolation parameters as needed, such as the number of interpolation points, the selection of interpolation functions, etc., thereby flexibly controlling the effect of signal splicing, which makes the signal splicing process more customizable and can be customized based on user testing needs. The random interpolation method can effectively reduce the artifacts and distortion that may occur during signal splicing, improving the quality and accuracy of signal splicing. By reasonably selecting the interpolation method and parameters, the error introduced by signal splicing can be reduced, thereby improving the subsequent detection accuracy.

[0244] Preferably, the pre-trained coal sample detection model is: θ = argmax θ L2(f(X|θ),Y);

[0245] where f(·|θ) is a deep neural network with θ as a parameter; is the target spectral data; is the coal component. In one possible implementation, the present scheme uses a deep neural network to construct a prediction model for coal component prediction, and establishes the mapping between the input NIRS-XRF fusion spectrum and the corresponding true component label through an end-to-end training method. Let denote the input signal after preprocessing, denote the true value of the corresponding coal component. X n ∈R D is the splicing of the NIRS and XRF spectra after preprocessing in step S20, y n is a non-negative scalar label. N and D are the number of training set samples and the input resolution, respectively. Note that X n can be down-sampled from the original data, so D can be smaller than the resolution of the original data. f(·|θ) is a deep neural network with θ as a parameter, which can be composed of multiple linear and nonlinear layers. The nonlinear layer here is also called the activation layer. A linear layer is usually followed by a nonlinear layer, and the combination of the two becomes a module. In this paper, the network is composed of multiple identical modules stacked in sequence and ends with a fully connected layer, outputting a scalar value y. Specifically, the basic operator of the linear layer can be a fully connected layer or a 1D convolution layer with different kernel sizes, while the basic operator of the nonlinear layer can be a hyperbolic tangent function (TanH), an exponential linear unit (ELU), or a Sigmoid function, etc.

[0246] Preferably, the method further comprises: performing pre-training of the coal sample detection model, including: collecting historical near-infrared spectrum signals and historical X-ray fluorescence spectrum data, and constructing corresponding historical target spectrum data based on the historical near-infrared spectrum signals and the historical X-ray fluorescence spectrum data; performing PLS model parameter initialization by taking the historical target spectrum data as training data; generating simulation samples based on the initialized PLS model, performing model training based on the simulation samples, and obtaining a coal sample detection initial model; performing verification on the coal sample detection initial model based on reserved historical target spectrum data, and obtaining a coal sample detection model.

[0247] In the embodiments of the present disclosure, in order to obtain an accurate detection model, a large amount of historical data is actually required as training samples for model training. However, it is actually difficult to obtain comprehensive historical data. On the one hand, there is a historical data retention problem, and on the other hand, there is a data intercommunication problem. Therefore, it is actually difficult to train an accurate detection model based on existing historical data. Based on this, the present disclosure proposes a pre-training scheme based on a PLS model.

[0248] In the embodiments of the present disclosure, the historical target spectrum data is used as training data to initialize parameters of a partial least squares regression (PLS) model, thereby laying a foundation for subsequent model training. Based on the initialized PLS model, simulation sample data is generated for model training, thereby expanding a training data set and improving the generalization ability of the model. The model is trained by using the simulation sample data, thereby establishing a coal sample detection initial model by learning the patterns and features of historical data.

[0249] Preferably, the PLS model parameters include: an initial weight, a learning rate, an activation function, a regularization parameter, an initialization bias term, and an optimizer type.

[0250] Preferably, the PLS model parameters include: an initial weight, a learning rate, an activation function, a regularization parameter, an initialization bias term, and an optimizer type.

[0251] In the embodiments of the present disclosure, an adaptive algorithm and machine learning technology are used to determine the optimal signal type and parameter combination to match the characteristics and changes of different spectral signals. Based on the determined signal type and parameters, a basic signal is generated as the basis of the augmented signal set. According to the data characteristics and model requirements, an augmented scheme is adaptively selected, including data interpolation, noise reduction processing, etc., to improve the data quality and model performance. The selected augmented scheme is performed on the basic signal, such as data interpolation, noise addition, etc., to generate a diversified augmented signal set. After the data augmentation operation is completed, an augmented signal set containing diversified signals is obtained, which is used to enrich the training data and improve the generalization ability and robustness of the model. Deviation data screening is performed in the augmented signal set to identify and remove data points that may introduce errors, ensuring the accuracy and reliability of the training data. Model training and verification are performed to improve the adaptability of the model to uncertainty and noise.

[0252] In the embodiments of the present disclosure, the adaptive selection of the augmented scheme includes: randomly selecting one or more schemes from among noise addition, translation, scaling, rotation, clipping, and transformation as preselected schemes; randomly adjusting parameters within the preset adjustable parameter range of each preselected scheme to obtain parameter-determined preselected schemes; if there is only one preselected scheme, directly performing processing on the basic signal based on the scheme to obtain an augmented signal; if there are multiple preselected schemes, sequentially performing each scheme to obtain an augmented signal.

[0253] Further, the deviation data screening in the augmented signal set includes: calculating the Euclidean distance between each augmented signal in each augmented signal set and the basic signal respectively; and removing the augmented signals with a Euclidean distance greater than a preset Euclidean distance threshold to obtain the simulated samples.

[0254] In the embodiments of the present disclosure, the present scheme adaptively determines the augmented scheme, which can randomly generate a large number of simulated signals to simulate the detection signals under various coal conditions, ensuring data comprehensiveness to make up for the problem of being unable to train a precise detection model due to insufficient historical retained data.

[0255] Further, the model training based on the simulated samples to obtain a coal sample detection initial model includes: performing pseudo-label annotation on the simulated samples based on the PLS model after parameter initialization, taking the annotated simulated samples as training samples; and performing model training in the neural network based on the target spectral data search based on the training samples to obtain a coal sample detection initial model.

[0256] In the embodiments of the present disclosure, the augmented data set is obtained, and the coal component corresponding to the data set also needs to be labeled, so as to train the model based on the simulated signal corresponding to the coal component. Based on this, the coal component of the augmented signal in the augmented signal set needs to be labeled, so that each augmented signal corresponds to a simulated coal component prediction result. After the annotation is completed, the corresponding training sample is obtained, and the model training is performed based on the training sample, so that the corresponding detection model can be obtained. However, in order to ensure the accuracy of the model, the model also needs to be verified.

[0257] Further, the coal sample detection initial model is verified based on the reserved historical target spectrum data to obtain a coal sample detection model, including: constructing a verification set based on the reserved historical target spectrum data, inputting the verification set into the coal sample detection initial model, and obtaining a corresponding prediction result; based on the prediction result and the actual result of the corresponding reserved historical target spectrum data, the performance of the coal sample detection initial model is evaluated; if the performance evaluation of the coal sample detection initial model does not pass, the hyperparameters are adjusted, the model structure is optimized, and / or the regularization operation is added to obtain an updated coal sample detection initial model; the performance of the updated coal sample detection initial model is re-evaluated based on the reserved historical target spectrum data, until a coal sample detection initial model meeting the preset performance requirement is obtained as a coal sample detection model.

[0258] In the embodiments of the present disclosure, the verification set is input into the coal sample detection initial model to obtain a prediction result, which is compared with the actual result to perform performance evaluation and error analysis. According to the performance evaluation result, if the model does not meet the preset performance requirement, the hyperparameters are adjusted, the model structure is optimized, and the regularization operation is added to improve the accuracy and stability of the model. According to the adjusted model, the performance is re-evaluated until a coal sample detection initial model meeting the preset performance requirement is obtained as an updated coal sample detection model. The present disclosure realizes continuous optimization and iteration of the coal sample detection model through repeated verification, evaluation and adjustment, improves the performance and accuracy of the model. The verification set based on the reserved historical target spectrum data can more accurately evaluate the performance and generalization ability of the model, and improve the reliability of the model evaluation. According to the performance evaluation result, the hyperparameters and the model structure are adjusted in time, so that the model can better adapt to the data characteristics and task requirements, and the accuracy of the coal sample detection is improved. Through the means such as adding the regularization operation, the stability and generalization ability of the model are improved, the risk of overfitting is reduced, and the robustness of the model is enhanced.

[0259] Preferably, the initial model performance evaluation of the coal sample detection based on the prediction result and the actual result of the corresponding reserved historical target spectrum data comprises: comparing the prediction result with the actual result of the corresponding reserved historical target spectrum data, and evaluating the accuracy and recall rate of the model based on the deviation of the two; if any one of the evaluation results of the accuracy and the recall rate fails, the initial model performance evaluation of the coal sample detection fails.

[0260] Preferably, the neural network search based on the target spectrum data comprises: dividing a plurality of optimization dimensions based on the neural network structure parameters; in each optimization dimension, adaptively adjusting the parameters in the corresponding neural network structure of the corresponding optimization dimension, and calculating the MAE index after each adjustment; comparing the MAE index corresponding to each parameter to determine the parameter combination with the highest MAE index as the optimization result in the corresponding optimization dimension; performing a greedy search between each optimization dimension to obtain the optimization result of each optimization dimension; and determining the structure parameters of the corresponding neural network structure based on the optimization result of each optimization dimension to construct the neural network corresponding to the search result.

[0261] Further, the optimization dimensions comprise: a basic operator dimension of a neural network layer, a resolution dimension of input data, a network depth dimension, and a network width dimension.

[0262] In the embodiments of the present disclosure, the neural network dimensions affecting the NIRS-XRF dual-spectrum fusion include the basic operators of linear and nonlinear layers, the resolution of input data, the network depth (the number of stacked modules), and the network width (the output channel of the linear layer). The search process of the four dimensions is divided into four stages in the present scheme, each stage searches a single dimension, and a greedy search is performed between the dimensions.

[0263] In the embodiments of the present disclosure, in the basic operator dimension of the neural network layer, the corresponding optimization result obtaining rule comprises: traversing the convolution types of the linear layer, and traversing the activation function types of the nonlinear layer, adaptively performing the combination of the convolution type of the fully connected layer and the activation function type of the nonlinear layer, and calculating the MAE index after each combination; comparing the MAE index after each combination, selecting the combination corresponding to the maximum MAE index, and determining the convolution type of the fully connected layer and the activation function type of the nonlinear layer corresponding to the combination as the optimization result in the basic operator dimension of the neural network layer.

[0264] Further, the convolution type of the linear layer is: a fully connected layer, a 1D convolution layer with a kernel size of 5, a 1D convolution layer with a kernel size of 10, or a 1D convolution layer with a kernel size of 15.

[0265] Further, the activation function type of the nonlinear layer is: a TanH function, an ELU function, a Sigmoid activation function, or a Softmax function.

[0266] Further, the calculation rule of the MAE index is:

[0267] wherein, is the true value of the coal component corresponding to the i-th training data; y i is the model prediction value corresponding to the i-th training data; m is the size of the data set.

[0268] In the embodiments of the present disclosure, the present scheme repeatedly generates N (N = 10) different initial network configurations to repeat the search process multiple times, and these configurations have different input resolutions, network depths and network widths. The cross-validation MAE is calculated for each configuration with different random network initialization configurations, and the best settings of the first-stage basic operator are voted to determine the final search result using the N results.

[0269] Further, the present scheme freezes the linear and nonlinear basic operator settings in several other dimensions. Similar processes as in the first stage are also used in the search of the remaining three dimensions. Finally, a network configuration with the best basic operator, input resolution, network depth and network width settings is generated to model the NIRS-XRF dual-spectral signal for predicting coal components.

[0270] In the embodiments of the present disclosure, an automatic neural network structure search strategy is proposed for automatically searching the optimal configuration for predicting coal components from NIRS-XRF dual-spectral signals. This method avoids manual network design, so that without much knowledge of the professional features of NIRS-XRF dual-spectral signals, the optimal neural network structure can also be outputted, which can fully capture the complex features in NIRS and XRF data and effectively fuse the information from the two different sources. It is worth noting that the neural network automatic search strategy is also efficient. Assuming that there are No, Nr, Nd and Nw candidate items for the basic operator, input resolution, network depth and network width, respectively, the complexity of the enumeration exhaustive search method is O(N o N r N d N w ), and the greedy search method proposed by us can reduce the complexity to O(N o +N r +N d +N w ), thereby significantly improving the search efficiency and being beneficial to model updating and deployment in practical applications.

[0271] Preferably, the detection results of the coal sample include one or more of ash composition, ash content, volatile matter, carbon and hydrogen, ash melting point, total water, total sulfur, and calorific value.

[0272] In the embodiments of the present disclosure, the present disclosure can achieve the following coal detection goals:

[0273] 1) Ash composition analysis: Coal is a complex organic material containing various elements and compounds, and ash is the inorganic matter left over after coal combustion, including minerals, soil, metal oxides, etc. The content of ash has an important influence on the combustion characteristics and utilization value of coal.

[0274] 2) Ash content analysis: Ash content is the content of non-combustible substances in coal, which can be inferred from specific characteristic peaks in the spectral signal. Accurate determination of ash content is crucial for evaluating the purity and combustion characteristics of coal.

[0275] 3) Volatile matter analysis: The volatile matter content is the content of gases and liquids volatilized during the heating process of coal, which can be inferred from the change of spectral signal. This is important for understanding the combustion characteristics and stability of coal.

[0276] 4) Carbon and hydrogen analysis: Carbon and hydrogen analysis of coal refers to the quantitative analysis of the content of carbon (C) and hydrogen (H) elements in coal samples. Through carbon and hydrogen analysis, the proportion of carbon and hydrogen elements in coal can be understood, and the calorific value, combustion characteristics and chemical properties of coal can be inferred.

[0277] 5) Ash melting point analysis: The ash melting point of coal refers to the temperature at which the ash in coal melts during the heating process at high temperature. The ash melting point can reflect the melting property of ash in coal. Coal with low ash melting temperature is prone to form slag during combustion, which can cause adverse effects on combustion equipment and environment. Coal with high ash melting temperature is more suitable for combustion, reducing the problem of ash during combustion. Ash melting point analysis can help researchers and engineers choose the right type of coal, optimize the combustion process, reduce environmental pollution, and improve energy utilization efficiency.

[0278] 6) Total water analysis: Coal total water analysis refers to the analysis and determination of the content of all water in coal samples. The results of coal total water analysis can help researchers and engineers understand the content of water in coal, which can affect the combustion performance, combustion efficiency and temperature control during combustion. Coal with high water content will consume more heat to evaporate water during combustion, reducing combustion efficiency and increasing flue gas emissions during combustion.

[0279] 7) Total sulfur analysis: Coal total sulfur analysis refers to the process of analyzing and measuring the content of all sulfur elements in coal samples. The results of coal total sulfur analysis can help researchers and engineers understand the content of sulfur elements in coal, and thus evaluate the combustion characteristics of coal, the sulfur emissions in combustion products, and the possible environmental impact during combustion. High-sulfur coal can produce sulfur oxides and sulfuric acid mist during combustion, causing negative impacts on the environment and health.

[0280] 8) Calorific value analysis: Coal calorific value analysis refers to the process of measuring and analyzing the heat released during coal combustion. The calorific value of coal is one of the important indicators for evaluating coal combustion performance and energy utilization efficiency. High-calorific value coal usually has higher combustion efficiency, can provide more heat energy, and reduce energy consumption costs. Therefore, coal calorific value analysis is of great significance for selecting appropriate fuels, optimizing combustion processes, evaluating energy utilization efficiency, and other aspects.

[0281] Preferably, as shown in FIG. 15, the system further comprises an output unit for visualizing the coal detection results. The visualization of the coal detection results includes: in response to a user data query instruction, determining a corresponding detection result object; selecting a preset data visualization scheme based on the corresponding detection result object, and pushing the visualized data to the user end.

[0282] Further, for different detection result objects, the system can select a preset data visualization scheme to display the detection results in the form of intuitive charts, curves, or images, facilitating user understanding and analysis. After data visualization processing, the system pushes the visualized data to the user end, and the user can intuitively view each detection result of the coal sample through the interface, helping them make decisions and evaluations.

[0283] FIG. 16 is a step flowchart of a coal detection method according to an embodiment of the present disclosure. As shown in FIG. 16, the present disclosure provides a coal detection method, and the system comprises:

[0284] Step S10: Perform the detection of coal sample spectrum data collection.

[0285] Specifically, near-infrared spectroscopy (NIRS) is a valuable non-destructive analysis technique for detecting organic materials, which has been widely used and recognized in multiple research works. Since coal is mainly composed of organic matter, containing various functional groups and minerals, NIRS is very suitable for coal quality analysis. Specifically, the spectral range of NIRS includes a broad absorption band sensitive to hydrogen-containing organic functional groups, which helps to accurately and reliably analyze coal quality. On the contrary, X-ray fluorescence spectroscopy (XRF) is commonly used for elemental analysis. In this process, high-energy X-rays are used to irradiate the material, excite atoms and induce energy absorption. When the excited atoms undergo electronic transitions back to the ground state, they release energy in the form of fluorescent radiation. This fluorescent radiation can be used to detect inorganic ash-forming elements such as aluminum (Al), silicon (Si), calcium (Ca), and other similar elements. By combining NIRS and XRF techniques, we can gain a more comprehensive understanding of the organic and inorganic composition of coal, thus more accurately assessing its quality. This integrated approach will help improve the efficiency and accuracy of coal quality detection, promoting the sustainable development of the coal industry.

[0286] Recently, there have been some works on coal sample detection based on NIRS-XRF dual-spectrum fusion. These studies usually use simple network structures without precise adaptation to the coal sample detection task. Near-infrared spectroscopy (NIRS) excels in detecting molecular-level information, while X-ray fluorescence spectroscopy (XRF) is good at extracting atomic-level information. Combining NIRS and XRF sensors with different material perception capabilities poses a significant challenge to linear-based techniques (such as partial least squares regression) and simple network structures.

[0287] To solve this problem and ensure the accuracy of dual-spectrum signals in coal sample detection, the present disclosure proposes a dual-spectrum signal fusion scheme and an adaptive neural network training scheme, achieving the technical effect of accurate prediction of coal composition based on target spectral data. Based on this, the present disclosure needs to collect dual-spectrum information of the detected coal sample.

[0288] Further, the present disclosure also tests the reflectivity of coal, i.e., sets the reflected light of the corresponding optical fiber recovered sample.

[0289] In one possible implementation, as shown in FIG. 17, before step S10, the method further includes: collecting a raw coal sample and performing sample preparation on the raw coal sample to obtain a detection coal sample.

[0290] Specifically, the sample preparation of the raw coal sample includes sampling, crushing, processing, and shaping, as shown in FIG. 18, specifically including:

[0291] Step S101: Perform raw coal sampling.

[0292] Specifically, in the coal industry, sampling is a crucial step that directly affects the accuracy and reliability of subsequent coal quality analysis and utilization. The present scheme proposes an intelligent sampling device that combines sensing technology and automatic control systems to achieve intelligent recognition and collection of raw coal samples. By mounting high-precision sensors and image recognition technology, it can monitor the coal flow during the transportation or storage process in real time and perform random sampling according to preset algorithms and rules. Using data analysis and simulation technology, the selection of sampling locations is optimized to ensure the representativeness and comprehensiveness of the sampling points. Through analysis of the coal flow path and speed, the best sampling location is determined to avoid sampling bias and local problems, improving the reliability of sampling. The introduction of real-time monitoring and feedback mechanisms monitors key parameters such as sampling volume and frequency during the sampling process, allowing for timely adjustments to the sampling strategy. Through data collection and analysis, real-time monitoring and quality control of the sampling process are achieved to ensure the accuracy and reliability of the sampling results. The design of an automated sampling system enables automatic collection and processing of raw coal samples, combining machine learning and artificial intelligence technology to optimize sampling algorithms and processes, improve sampling efficiency and automation, and reduce human intervention and errors. The design of the intelligent sampling device and optimized sampling location in the present scheme can improve the accuracy and representativeness of sampling, reduce sampling errors and biases, and improve the reliability of sampling results. The introduction of real-time monitoring and feedback mechanisms can optimize the sampling process, improve sampling efficiency and speed, and save labor and time costs. The design of an automated sampling system can reduce human intervention, reduce operational risks and errors, and improve the controllability and stability of the sampling process. As shown in FIG. 18, the steps include:

[0293] Step S102: Perform raw coal crushing treatment.

[0294] Specifically, in the coal processing process, crushing is an essential step, but it also produces a large amount of dust pollution. In the crushing module, integrate efficient dust removal equipment such as bag dust collectors or electrostatic precipitators to capture and filter dust particles generated during the crushing process. Through equipment integration, timely removal and treatment of dust during the crushing process is achieved, ensuring a clean production environment and employee health. Wet dust removal technology is adopted, which can reduce the spread and flying of dust through a spraying system or wet scrubbing device. Wet dust removal technology can effectively control dust emissions and reduce the impact on air quality. The crushing module is designed with a negative pressure closed structure, which controls the flow of dust in the closed space through a negative pressure system, reducing dust leakage and dispersion. The negative pressure closed design can effectively prevent dust pollution and protect the surrounding environment and equipment. An online monitoring and control system is introduced to monitor dust emission concentration and particle size distribution in real time. Based on the monitoring data, adjust the operating parameters and cleaning cycle of the dust removal equipment. Through real-time monitoring and control, the stability and efficiency of the dust removal effect are ensured. The disclosed scheme effectively controls dust emissions and reduces environmental pollution by integrating dust removal equipment and applying wet dust removal technology. The negative pressure closed design and the introduction of the online monitoring system can protect employee health, reduce the harm of dust to the respiratory tract and skin, and improve the comfort of the working environment.

[0295] Step S103: Process the raw coal into a detection coal sample.

[0296] Specifically, the disclosed scheme introduces a sample grading and screening device in the processing module, which can finely screen and grade the sample according to the particle size and shape characteristics of the crushed coal sample. Through grading and screening, basic coal samples of different particle size ranges can be obtained, providing more accurate and representative samples for subsequent detection. The disclosed scheme designs a chemical treatment reaction tank for chemical treatment and reaction of the crushed coal sample to extract target components or remove interfering substances. The chemical treatment reaction tank can change the chemical properties of the coal sample, making it more suitable for subsequent detection and analysis, and improving the accuracy of the detection results. The disclosed scheme introduces a magnetic separation device to separate magnetic impurities or magnetic minerals from the crushed coal sample. The magnetic separation device can effectively remove interfering substances and purify the basic coal sample, improving the accuracy and reliability of the detection. A drying and dehumidifying system is designed to dry and dehumidify the processed basic coal sample, ensuring the dryness and stability of the sample. The drying and dehumidifying system can avoid the influence of moisture on the detection results, ensuring the accuracy and reliability of the detection data.

[0297] Step S104: Perform shaping treatment on the detection coal sample.

[0298] Specifically, the shape of the coal sample can have the following effects on the monitoring results of the coal composition based on the spectral signal:

[0299] 1) Light path scattering effect: The irregular shape or rough surface of the coal sample can cause scattering of the light path, resulting in deflection or scattering of the light signal during transmission, affecting the collection and accuracy of the spectral signal.

[0300] 2) Light absorption difference: Different shapes of the coal sample can result in different light absorption capabilities of different parts, thereby affecting the intensity and characteristics of the spectral signal. For example, the size, shape and density differences of the particles on the surface of the coal can affect the transmission and absorption of light.

[0301] 3) Surface reflection effect: The light reflection characteristics of the surface of the coal sample can affect the collection and analysis of the spectral signal. Different shapes of the coal sample can have different reflectivities, resulting in different reflection and diffuse reflection of the light signal on the sample surface, affecting the quality and accuracy of the spectral signal.

[0302] 4) Spectral signal interference: The unevenness of the shape of the coal sample can cause interference of the spectral signal, such as the presence of impurities, oxides or moisture on the surface of the sample, which can affect the purity and stability of the spectral signal, thereby affecting the accuracy of the composition monitoring.

[0303] 5) Light path length difference: The size and shape of the coal sample can cause differences in the length of the light path, resulting in different degrees of attenuation and absorption of the light signal during transmission in the sample, affecting the intensity and clarity of the spectral signal.

[0304] In the embodiments of the present disclosure, the shape of the coal sample has a certain effect on the monitoring results of the coal composition based on the spectral signal, so it is necessary to consider and appropriately correct and process the shape factor during composition monitoring to improve the accuracy and reliability of the monitoring results.

[0305] Further, although the present disclosure proposes corresponding detection of the shape of the coal sample, in order to ensure that the shaped shape meets the expectations, the shaped result needs to be detected. The present disclosure collects spectral signals at the same time, and also collects visual signals such as image information. Based on the image information, target detection of the shape of the coal sample is performed to determine whether it meets the expectations. If it does not meet the expectations, an alarm instruction needs to be output to remind relevant personnel to replace the sample to avoid obtaining incorrect detection results.

[0306] Preferably, the surface flatness recognition of the coal sample based on the visual image data comprises: sequentially performing denoising, grayscale and edge detection processing on the visual image, calibrating the detection coal sample area; extracting the surface flatness related features in the detection coal sample area; performing surface flatness index recognition on the surface flatness related features, and determining the surface flatness of the detection coal sample based on the recognition result. The surface flatness related features include one or more of texture features, color features and edge features. The surface flatness index includes average gray level and / or texture uniformity.

[0307] Specifically, the following steps are included:

[0308] 1) Data acquisition and preprocessing: Collect visual image data of the detection coal sample, ensure image clarity and consistency. Perform denoising processing on the visual image to eliminate the influence of noise on subsequent processing. Convert the processed image to a grayscale image to simplify the processing process and highlight the surface features.

[0309] 2) Surface flatness recognition processing: Perform edge detection processing to accurately calibrate the boundary and contour of the detection coal sample. Calibrate the detection coal sample area to ensure that subsequent processing focuses on the sample surface. Extract surface flatness related features, including texture features, color features and edge features, in the detection coal sample area. Feature extraction and surface flatness index recognition:

[0310] 3) Analyze and process the extracted features to determine the surface flatness related features. Based on one or more of the texture features, color features and edge features, calculate the surface flatness index, such as average gray level and texture uniformity.

[0311] 4) Surface flatness recognition and result determination: According to the recognition result, determine the surface flatness level of the detection coal sample. Combine the surface flatness index and feature analysis result to evaluate and classify the surface flatness of the detection coal sample.

[0312] In the embodiments of the present disclosure, the surface flatness recognition of the detection coal sample based on the visual image data can be effectively applied to quality control and production process monitoring in the coal industry, improving production efficiency and product quality management level.

[0313] Specifically, the shape evaluation of the detected coal sample based on the fitted shape size and the fitted surface flatness includes: calculating the Euclidean distance between the fitted shape size and the preset shape size as a first deviation value; calculating the Euclidean distance between the fitted surface flatness and the preset surface flatness as a second deviation value; performing normalization processing on the first deviation value and the second deviation value, and performing arithmetic average calculation on the normalized first deviation value and the second deviation value to obtain a deviation coefficient; and assigning a preset correction coefficient in the coal sample detection model based on the deviation coefficient.

[0314] In the embodiments of the present disclosure, the present scheme can also fuse the subsequent detection model through the deviation correction coefficient for the detection coal sample with certain differences in shape, so that even the detection coal sample with certain differences can be processed by the unified standard through the deviation correction coefficient.

[0315] Step S20: performing spectrum data inference based on the pre-trained coal sample detection model to obtain an inference result.

[0316] Specifically, the near-infrared spectrum signal and the X-ray fluorescence spectrum data in the detection data are subjected to fusion processing to obtain target spectrum data, and model training is performed based on the fusion signal. The near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively subjected to preprocessing; the preprocessed near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively subjected to down-sampling processing, and the down-sampled near-infrared spectrum signal and the X-ray fluorescence spectrum data are scaled to a normal distribution; the normal distribution of the near-infrared spectrum signal and the normal distribution of the X-ray fluorescence spectrum data are subjected to splicing processing to obtain target spectrum data. Specifically, as shown in FIG. 19, the following steps are included:

[0317] Step S201: respectively performing preprocessing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data.

[0318] Specifically, the near-infrared spectrum signal and the X-ray fluorescence spectrum data are respectively subjected to SG convolution smoothing processing; and the near-infrared spectrum signal subjected to the convolution smoothing processing is subjected to area normalization processing.

[0319] Preferably, the SG convolution smoothing processing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data respectively includes: calculating the coefficients of the SG convolution kernel based on a preset window size and a polynomial order; respectively weighting and averaging the adjacent data points in the near-infrared spectrum signal and the X-ray fluorescence spectrum data based on the coefficients in the convolution kernel, so as to obtain the smoothed data points; performing symmetric extension or zero padding processing on the boundaries of the spectrum signals; and obtaining the spectrum signals subjected to the SG convolution smoothing processing based on the smoothed data points and the processed boundaries.

[0320] In the embodiments of the present disclosure, the SG convolution smoothing is a linear smoothing method, which can better smooth the data curve by local fitting, eliminate the peaks and fluctuations in the data, and make the data more stable and continuous. The SG convolution smoothing processing can effectively smooth the noise part of the data while retaining the characteristics and trends of the signal, which helps to reduce the interference of high-frequency noise in the data on the signal and improve the readability and analysis accuracy of the data. Compared with other smoothing methods, the SG convolution smoothing processing can better maintain the overall shape and trend of the data while smoothing the data, without causing distortion or deviation of the data shape, and retains the original characteristics of the data. The SG convolution smoothing processing is a simple and efficient data smoothing method, which has a faster calculation speed and is suitable for processing large-scale data sets, can complete the data smoothing operation in a short time, and is suitable for large-scale detection of coal samples, ensuring the detection accuracy while improving the detection efficiency.

[0321] Further, the area normalization processing of the near-infrared spectrum signal after the convolution smoothing processing comprises: determining a signal set of the near-infrared spectrum signal after the convolution smoothing processing; the signal set comprises a plurality of signal samples, and each signal sample is a reflectivity under a corresponding spectrum type at a corresponding wavelength point; and calculating a signal after the area normalization processing based on the signal set of the near-infrared spectrum signal.

[0322] In the embodiments of the present disclosure, the near-infrared spectrum signal is prone to baseline drift, so the present disclosure also eliminates the baseline drift in the near-infrared spectrum signal through area normalization. Baseline drift is a signal offset caused by instrument drift, environmental changes, or sample differences, which affects the accuracy and stability of the spectrum signal. Through area normalization processing, the signal can be shifted up or down as a whole, so that the baseline level is more stable. The optical path difference in the near-infrared spectrum signal can cause differences in signal intensity, affecting the comparison and analysis of the signal. Area normalization processing can normalize the overall intensity of the signal, reduce the influence of optical path difference on the signal, and make the comparison between different samples more accurate. Area normalization processing can highlight the characteristic peaks or troughs in the near-infrared spectrum signal, making the characteristics of the signal more obvious and prominent, which helps to more accurately identify and analyze specific components or characteristics in the spectrum.

[0323] Specifically, the calculation rule of the signal after the area normalization processing is:

[0324] wherein, represents the NIR reflectivity of the nth sample at the mth wavelength point; is the signal after the area normalization processing.

[0325] Step S202: performing down-sampling processing on the near-infrared spectrum signal and the X-ray fluorescence spectrum data after the pre-processing is completed.

[0326] Specifically, filtering processing is performed on each spectrum signal respectively, in the spectrum signal after the filtering processing, a sampling point is reserved every fixed interval to obtain a plurality of sampling points, or a plurality of segments are intercepted from the spectrum signal after the filtering processing, and average processing is performed on each sampling point in each segment to obtain a sampling point in each segment, and a plurality of sampling points are obtained, and signal reconstruction is performed based on each sampling point to obtain a spectrum signal after the down-sampling processing.

[0327] In the embodiments of the present disclosure, the down-sampling processing can reduce the data amount, that is, reduce the sampling rate or the number of sampling points, thereby saving the storage space and the computing resources. In particular, for large-scale data sets or high-frequency sampled data, down-sampling can effectively reduce the volume of data, facilitating storage and processing. Down-sampling can simplify the data analysis process, reduce the complexity and dimensionality of data, making the data more easily understood and processed. By reducing the resolution of data, some detailed information can be removed, highlighting the main features of the data, simplifying the model establishment and analysis process. Down-sampling processing can help remove noise and interference in the data, smooth the data curve, and improve the quality and stability of the data. By reducing the sampling rate or averaging the sampling points, the volatility of the data can be reduced, making the data clearer and more reliable. The interval sampling point reservation and the multi-segment average processing in the present disclosure can reduce the complexity and dimensionality of the data, reduce the data amount, and make the data more easily processed and analyzed. In addition, the interval sampling point reservation and the average processing can further smooth the signal curve, reduce the volatility of the data, and improve the signal quality and stability by reducing noise and interference in the signal.

[0328] Step S203: scaling the near-infrared spectrum signal and the X-ray fluorescence spectrum data after the down-sampling to a normal distribution.

[0329] Specifically, the statistical characteristics of each spectrum signal are calculated respectively; wherein the statistical characteristics are variance and / or standard deviation; based on the statistical characteristics of each spectrum signal, standardization processing or normalization processing is performed on each spectrum signal, the numerical value of each spectrum signal is scaled to a preset numerical range to obtain the scaled numerical value of each spectrum signal; and based on a preset transformation algorithm and the scaled numerical value of each spectrum signal, each spectrum signal is converted to a normal distribution.

[0330] In the embodiments of the present disclosure, the present scheme converts the signal into a normal distribution, which can simplify the data analysis process and make the data easier to understand and process. By converting the signal into a normal distribution, the fitting effect and prediction accuracy of the model can be improved, and the signal converted into a normal distribution can better meet the requirements of hypothesis testing and ensure the effectiveness of the test results. Moreover, the normal distribution has the characteristics of standardization, i.e., the mean is 0 and the standard deviation is 1, and converting the signal into a normal distribution can standardize the data, making different signals comparable and facilitating subsequent splicing of two spectral signals.

[0331] Step S204: performing splicing processing on the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data to obtain target spectrum data.

[0332] Specifically, the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data are aligned, and after the alignment, the two spectral signals are spliced based on weighted average to obtain initial target spectrum data. The initial target spectrum data is verified based on the normal distribution near-infrared spectrum signal and / or the normal distribution X-ray fluorescence spectrum data, and the initial target spectrum data that passes the verification is taken as the target spectrum data.

[0333] Further, the alignment of the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data includes: selecting the same reference point in the normal distribution near-infrared spectrum signal and the normal distribution X-ray fluorescence spectrum data, and calculating the time difference between the corresponding reference point and other data points in each spectrum signal to obtain a time difference sequence; and performing time axis adjustment of one of the spectrum signals based on the random interpolation method and the corresponding time difference sequence until the time axes of the two spectrum signals are aligned.

[0334] Further, the random interpolation method is: X aug = αX1 + (1-α)X2, α ∈ [0, 1] y aug = f PLS (X aug )

[0335] wherein, f PLS represents a PLS mapping function; X aug and y aug represent the synthesized NIRS-XRF signal and the corresponding pseudo label, respectively; α represents an interpolation weight randomly sampled in the range of [0, 1]; and X1 and X2 are randomly selected signal samples from the historical NIRS-XRF signal of the coal sample.

[0336] In the embodiments of the present disclosure, the present scheme performs two spectral signal splicing based on a random interpolation method, which can preserve the characteristics and information of the original data during the signal splicing process and avoid data loss or distortion. By splicing the signals through interpolation, the integrity of the data can be better maintained. The random interpolation method can achieve smooth transition in the transition area of signal splicing and avoid sudden changes or discontinuities, which helps to improve the continuity and smoothness of signal splicing. The random interpolation method can adjust the interpolation parameters as needed, such as the number of interpolation points, the selection of interpolation functions, etc., to flexibly control the effect of signal splicing, which makes the signal splicing process more customizable and can be customized based on user testing needs. The random interpolation method can effectively reduce the artifacts and distortion phenomena that may occur during signal splicing, improving the quality and accuracy of signal splicing. By reasonably selecting the interpolation method and parameters, the error introduced by signal splicing can be reduced, thereby improving the subsequent detection accuracy.

[0337] Further, the present scheme uses a deep neural network to construct a prediction model for coal component prediction, and establishes the mapping between the input NIRS-XRF fusion spectrum and the corresponding true component label through an end-to-end training method. represents the input signal after preprocessing, represents the true value of the corresponding coal component. n ∈R D is the splicing of the NIRS and XRF spectra after preprocessing in step S20, y n is a non-negative scalar label. N and D are the number of training set samples and input resolution, respectively. Note that, since X n D can be smaller than the resolution of the original data, as the original data can be downsampled. f(·|θ) is a deep neural network with θ as a parameter, which can be composed of multiple linear and nonlinear layers. The nonlinear layer here is also called the activation layer. A linear layer is usually followed by a nonlinear layer, and the combination of the two becomes a module. In this paper, the network is composed of multiple identical modules stacked in sequence and ends with a fully connected layer, outputting a scalar value y. Specifically, the basic operator of the linear layer can be a fully connected layer or a 1D convolution layer with different kernel sizes, while the basic operator of the nonlinear layer can be a hyperbolic tangent function (TanH), an exponential linear unit (ELU), or a Sigmoid function, etc.

[0338] Preferably, the method further includes: pre-training the coal sample detection model, including: collecting historical near-infrared spectral signals and historical X-ray fluorescence spectral data, and constructing corresponding historical target spectral data based on the historical near-infrared spectral signals and historical X-ray fluorescence spectral data; using the historical target spectral data as training data to initialize PLS model parameters; generating simulated samples based on the initialized PLS model, performing model training based on the simulated samples to obtain an initial coal sample detection model; and performing verification on the initial coal sample detection model based on reserved historical target spectral data to obtain a coal sample detection model.

[0339] In this embodiment, obtaining an accurate detection model requires a large amount of historical data as training samples. However, obtaining comprehensive historical data is difficult due to issues of data retention and data interoperability. Therefore, it is challenging to train an accurate detection model based solely on existing historical data. To address this, this disclosure proposes a pre-training scheme based on the PLS model.

[0340] In this embodiment, historical target spectral data is used as training data to initialize the parameters of the partial least squares regression (PLS) model, laying the foundation for subsequent model training. Based on the initialized PLS model, simulated sample data is generated for model training, expanding the training dataset and improving the model's generalization ability. Using the simulated sample data for model training, an initial model for coal sample detection is established by learning the patterns and features of historical data.

[0341] Preferably, the PLS model parameters include: initial weights, learning rate, activation function, regularization parameter, initialization bias term, and optimizer type.

[0342] Preferably, the step of generating simulated samples based on the initialized PLS model includes: adaptively determining the signal type and signal parameters that match the target spectral data, generating a basic signal based on the signal type and the signal parameters; adaptively selecting an augmentation scheme, and performing a data augmentation operation on the basic signal based on the selected augmentation scheme to obtain an augmented signal set; filtering out bias data from the augmented signal set, and using the filtered augmented signal set as simulated samples.

[0343] In the embodiments of the present disclosure, an adaptive algorithm and machine learning technology are used to determine the optimal signal type and parameter combination to match the characteristics and changes of different spectral signals. Based on the determined signal type and parameters, a basic signal is generated as the basis of the augmented signal set. According to the data characteristics and model requirements, an augmented scheme is adaptively selected, including data interpolation, noise reduction processing, etc., to improve the data quality and model performance. The selected augmented scheme is performed on the basic signal, such as data interpolation, noise addition, etc., to generate a diversified augmented signal set. After the data augmentation operation is completed, an augmented signal set containing diversified signals is obtained, which is used to enrich the training data and improve the generalization ability and robustness of the model. Deviation data screening is performed in the augmented signal set to identify and remove data points that may introduce errors, ensuring the accuracy and reliability of the training data. Model training and verification are performed to improve the adaptability of the model to uncertainty and noise.

[0344] In the embodiments of the present disclosure, the adaptive selection of the augmented scheme includes: randomly selecting one or more schemes from the noise addition, translation, scaling, rotation, clipping and transformation as preselected schemes; randomly adjusting the parameters within the preset adjustable parameter range of each preselected scheme to obtain a parameter-determined preselected scheme; if there is only one preselected scheme, directly performing processing on the basic signal based on the scheme to obtain an augmented signal; if there are multiple preselected schemes, sequentially performing each scheme to obtain an augmented signal.

[0345] Further, the deviation data screening in the augmented signal set includes: calculating the Euclidean distance between each augmented signal in each augmented signal set and the basic signal respectively; removing the augmented signals with a Euclidean distance greater than a preset Euclidean distance threshold, and using the removed augmented signal set as the simulation sample.

[0346] In the embodiments of the present disclosure, the present scheme adaptively determines the augmented scheme, which can randomly generate a large number of simulation signals to simulate the detection signals under various coal conditions, ensuring data comprehensiveness to make up for the problem of inability to train a precise detection model due to insufficient historical retained data.

[0347] Further, the model training based on the simulation sample to obtain a coal sample detection initial model includes: performing pseudo-label annotation on the simulation sample based on the PLS model after parameter initialization, and using the annotated simulation sample as a training sample; performing model training in the neural network based on the target spectral data search based on the training sample to obtain a coal sample detection initial model.

[0348] In the embodiments of the present disclosure, the augmented data set is obtained, and the coal component corresponding to the data set also needs to be labeled, so as to train the model based on the simulated signal corresponding to the coal component. Based on this, the coal component of the augmented signal in the augmented signal set needs to be labeled, so that each augmented signal corresponds to a simulated coal component prediction result. After the annotation is completed, the corresponding training sample is obtained, and the model training is performed based on the training sample, so that the corresponding detection model can be obtained. However, in order to ensure the accuracy of the model, the model also needs to be verified.

[0349] Further, the coal sample detection initial model is verified based on the reserved historical target spectrum data to obtain a coal sample detection model, including: constructing a verification set based on the reserved historical target spectrum data, inputting the verification set into the coal sample detection initial model, and obtaining a corresponding prediction result; based on the prediction result and the actual result of the corresponding reserved historical target spectrum data, the performance of the coal sample detection initial model is evaluated; if the performance evaluation of the coal sample detection initial model does not pass, the hyperparameters are adjusted, the model structure is optimized, and / or the regularization operation is added to obtain an updated coal sample detection initial model; the performance of the updated coal sample detection initial model is re-evaluated based on the reserved historical target spectrum data, until a coal sample detection initial model meeting the preset performance requirement is obtained as a coal sample detection model.

[0350] In the embodiments of the present disclosure, the verification set is input into the coal sample detection initial model to obtain a prediction result, which is compared with the actual result to perform performance evaluation and error analysis. According to the performance evaluation result, if the model does not meet the preset performance requirement, the hyperparameters are adjusted, the model structure is optimized, and the regularization operation is added to improve the accuracy and stability of the model. According to the adjusted model, the performance is re-evaluated until a coal sample detection initial model meeting the preset performance requirement is obtained as an updated coal sample detection model. The present disclosure realizes continuous optimization and iteration of the coal sample detection model through repeated verification, evaluation and adjustment, improves the performance and accuracy of the model. The verification set based on the reserved historical target spectrum data can more accurately evaluate the performance and generalization ability of the model, and improve the reliability of the model evaluation. According to the performance evaluation result, the hyperparameters and the model structure are adjusted in time, so that the model can better adapt to the data characteristics and task requirements, and the accuracy of the coal sample detection is improved. Through the means such as adding the regularization operation, the stability and generalization ability of the model are improved, the risk of overfitting is reduced, and the robustness of the model is enhanced.

[0351] Preferably, the coal sample detection initial model performance evaluation based on the prediction result and the corresponding reserved historical target spectrum data actual result comprises: comparing the prediction result with the corresponding reserved historical target spectrum data actual result, and evaluating the accuracy and recall rate of the model based on the deviation of the two; if any one of the accuracy and recall rate evaluation results does not pass, the coal sample detection initial model performance evaluation does not pass.

[0352] Preferably, the neural network search based on the target spectrum data comprises: dividing a plurality of optimization dimensions based on the neural network structure parameters; in each optimization dimension, the parameters in the corresponding neural network structure are adaptively adjusted, and the MAE index after each adjustment is calculated; the parameter combination with the highest MAE index is determined as the optimization result in the corresponding optimization dimension by comparing the MAE indexes corresponding to each parameter; the greedy search is performed between each optimization dimension to obtain the optimization result of each optimization dimension; and the structure parameters corresponding to each neural network structure are determined based on the optimization results of each optimization dimension to construct the neural network corresponding to the search result.

[0353] Further, the optimization dimension comprises: a basic operator dimension of a neural network layer, a resolution dimension of input data, a network depth dimension, and a network width dimension.

[0354] In a possible implementation, the method specifically comprises the following steps:

[0355] 1) Determine the search order, for example, the basic operator, the input resolution, the network depth, and the network width.

[0356] 2) Search the basic operator, determine the basic operator search range (linear layer and nonlinear layer combination), randomly select 10 groups (or more) of different configurations of input resolution, network depth, and network width, determine the optimal operator for each group of configurations, and 10 groups of configurations will output 10 optimal operators (the 10 optimal operators may not all be the same operator). Finally, the operator with the highest frequency in the 10 optimal operators is selected as the searched basic operator.

[0357] 3) Search the input resolution, determine the input resolution search range, fix the basic operator to the optimal configuration since the optimal operator has been determined in step 2, randomly select 10 groups of different configurations of network depth and network width, determine the optimal input resolution for each group of configurations, and 10 groups of configurations will output 10 optimal input resolutions. Finally, the input resolution with the highest frequency in the 10 optimal input resolutions is selected as the searched input resolution.

[0358] 4) Search network depth, determine network depth search range, because the optimal operator has been determined in steps 2 and 3, and the optimal input resolution is determined, so we fix the basic operator and input resolution as the optimal configuration, randomly search 10 groups of different configurations of network width, determine the optimal network depth for each group of configurations, 10 groups of configurations will output 10 optimal network depths, and finally the network depth with the highest frequency in the 10 optimal network depths is taken as the searched network depth.

[0359] 5) Search network width, determine network width search range, because the optimal operator, the optimal input resolution, and the optimal network depth have been determined in steps 2, 3, and 4, so we fix the basic operator, the input resolution, and the network depth as the optimal configuration, search for the optimal network width, and finally determine the optimal network width.

[0360] In the embodiment of the present application, the proposed inherited greedy search is that, assuming that there are N dimensions to be searched, and the ith dimension has been searched, the 0th-i-1th dimensions are fixed as the optimal configurations that have been searched, M groups of i+1th-Nth dimension configurations are randomly searched, M groups of optimal configurations of the ith dimension are output, and the optimal configuration with the highest frequency in the M optimal configurations of the ith dimension is taken as the optimal configuration.

[0361] Based on the scheme of the present application, assuming that there are No, Nr, Nd, and Nw candidate items for the basic operator, the input resolution, the network depth, and the network width, respectively, the complexity of the enumeration exhaustive search method is O(No Nr NdNw), and the proposed greedy search method can reduce the complexity to O(No+Nr+Nd+Nw), thereby significantly improving the search efficiency and being beneficial to model updating and deployment in practical applications.

[0362] In the embodiment of the present application, the dimensions of the neural network affecting the NIRS-XRF dual-spectrum fusion include the basic operator of the linear and nonlinear layers, the resolution of the input data, the network depth (the number of stacked modules), and the network width (the output channel of the linear layer). The search process of the four dimensions is divided into four stages in the scheme of the present application, each stage searches a single dimension, and greedy search is performed between the dimensions.

[0363] In the embodiment of the present application, the corresponding optimization result obtaining rule in the basic operator dimension of the neural network layer includes: traversing the convolution type of the linear layer, and traversing the activation function type of the nonlinear layer, adaptively combining the convolution type of the fully connected layer and the activation function type of the nonlinear layer, calculating the MAE index after each combination; comparing the MAE index after each combination, selecting the combination corresponding to the maximum MAE index, and determining the convolution type of the fully connected layer and the activation function type of the nonlinear layer corresponding to the combination as the optimization result in the basic operator dimension of the neural network layer.

[0364] Further, the convolution type of the linear layer is: a fully connected layer, a 1D convolution layer with a kernel size of 5, a 1D convolution layer with a kernel size of 10, or a 1D convolution layer with a kernel size of 15.

[0365] Further, the activation function type of the nonlinear layer is: a TanH function, an ELU function, a Sigmoid activation function, or a Softmax function.

[0366] Further, the calculation rule of the MAE index is:

[0367] wherein, is the true value of the coal component corresponding to the i th training data; y i is the model prediction value corresponding to the i th training data; m is the size of the data set.

[0368] In the embodiments of the present disclosure, the present scheme repeatedly generates N (N = 10) different initial network configurations to repeat the search process multiple times, and these configurations have different input resolutions, network depths, and network widths. The cross-validation MAE is calculated for each configuration with different random network initialization configurations, and the N results are used to vote for the best settings of the first-stage basic operator to determine the final search result.

[0369] Further, the present scheme freezes the linear and nonlinear basic operator settings in other several dimensions. Similar processes as in the first stage are also adopted in the search of the remaining three dimensions. Finally, a network configuration with the best basic operator, input resolution, network depth, and network width settings is generated to model the NIRS-XRF dual-spectral signal for predicting coal components.

[0370] In the embodiments of the present disclosure, an automatic neural network structure search strategy is proposed for automatically searching for the optimal configuration for predicting coal components from NIRS-XRF dual-spectral signals. This method avoids manual network design, so that without much knowledge of the professional features of NIRS-XRF dual-spectral signals, the optimal neural network structure can also be output, which can fully capture the complex features in NIRS and XRF data and effectively fuse information from the two different sources. It is worth noting that the neural network automatic search strategy is also efficient. Assuming that there are No, Nr, Nd, and Nw candidate items for the basic operator, input resolution, network depth, and network width, respectively, the complexity of the exhaustive search method by enumeration is O(N o N r N d N w ), and the greedy search method proposed by us can reduce this complexity to O(No +N r +N d +N w ), thereby significantly improving search efficiency, which is conducive to model updating and deployment in practical applications.

[0371] In another possible implementation, as shown in FIG. 20, the method further includes visualizing and outputting the coal detection result.

[0372] Specifically, the detection results of the coal sample include one or more of ash composition, ash content, volatile matter, carbon and hydrogen, ash melting point, total water, total sulfur, and calorific value.

[0373] In the embodiments of the present disclosure, the present disclosure can achieve the following coal detection goals:

[0374] 1) Ash composition analysis: Coal is a complex organic matter containing various elements and compounds, and ash is inorganic matter left over after coal combustion, including minerals, soil, metal oxides, etc. The content of ash has an important influence on the combustion characteristics and utilization value of coal.

[0375] 2) Ash content analysis: Ash content is the content of non-combustible substances in coal, which can be inferred from specific characteristic peaks in the spectral signal. Accurate determination of ash content is crucial for evaluating the purity and combustion characteristics of coal.

[0376] 3) Volatile matter analysis: The volatile matter content is the content of gases and liquids volatilized during the heating process of coal. The volatile matter content in the coal sample can be inferred from the changes in the spectral signal, which is crucial for understanding the combustion characteristics and stability of coal.

[0377] 4) Carbon and hydrogen analysis: Carbon and hydrogen analysis of coal refers to the quantitative analysis of the content of carbon (C) and hydrogen (H) elements in coal samples. Through carbon and hydrogen analysis, the proportion of carbon and hydrogen elements in coal can be understood, and the calorific value, combustion characteristics, and chemical properties of coal can be inferred.

[0378] 5) Ash melting point analysis: The ash melting point of coal refers to the temperature at which the ash in the coal melts during the heating process at high temperatures. The ash melting point reflects the fusibility of the ash in the coal. Coal with a lower ash melting temperature is prone to form slag during combustion, which can adversely affect combustion equipment and the environment. Coal with a higher ash melting temperature is more suitable for combustion, reducing the problem of ash during combustion. Ash melting point analysis can help researchers and engineers choose the right type of coal, optimize the combustion process, reduce environmental pollution, and improve energy utilization efficiency.

[0379] 6) Total moisture analysis: Coal total moisture analysis refers to the process of analyzing and measuring the content of all water in coal samples. The results of coal total moisture analysis can help researchers and engineers understand the content of water in coal, which in turn affects the combustion performance, combustion efficiency, and temperature control during the combustion process. Coal with high water content will consume more heat to evaporate water during the combustion process, reducing combustion efficiency and increasing flue gas emissions during the combustion process.

[0380] 7) Total sulfur analysis: Coal total sulfur analysis refers to the process of analyzing and measuring the content of all sulfur elements in coal samples. The results of coal total sulfur analysis can help researchers and engineers understand the content of sulfur elements in coal, which in turn assesses the combustion characteristics of coal, the amount of sulfur emissions in combustion products, and the possible environmental impact during the combustion process. High-sulfur coal will produce sulfur oxides and sulfuric acid mist during the combustion process, causing negative impacts on the environment and health.

[0381] 8) Calorific value analysis: Coal calorific value analysis refers to the process of measuring and analyzing the heat released during coal combustion. The calorific value of coal is one of the important indicators for evaluating the combustion performance and energy utilization efficiency of coal. High-calorific-value coal usually has higher combustion efficiency and can provide more heat energy, reducing energy consumption costs. Therefore, coal calorific value analysis is of great significance for selecting appropriate fuels, optimizing combustion processes, and evaluating energy utilization efficiency.

[0382] The method further includes visualizing the coal detection results. The visualization of the coal detection results includes determining corresponding detection result objects in response to user data query instructions, selecting a preset data visualization scheme based on the corresponding detection result objects, and pushing the visualized data to the user end.

[0383] The disclosed scheme arranges and stores the coal sample detection results output by the model based on the inference results, for subsequent display and analysis. The coal sample detection results are displayed in the form of charts, images, etc. using data visualization technology, intuitively presenting the composition information of the coal sample. The result display interface is designed, and users can input coal sample information through the interface to view the corresponding coal sample detection results, realizing personalized display and query of the results. The real-time updating function of the results is realized, and when new coal samples are detected, the displayed coal sample detection results are updated in time, maintaining the timeliness and accuracy of the data. The result explanation and analysis function is provided to explain the content and significance of each component, helping users better understand the composition and characteristics of the coal sample.

[0384] In the embodiments of the present disclosure, the coal sample detection results are displayed through data visualization, making the complex data intuitive and easy to understand, and improving the user's understanding and analysis ability of the composition of the coal sample. The friendly result display interface enables users to conveniently and quickly query and view the coal sample detection results, and improves the user experience and operation efficiency. The display results are updated in real time, so that users can timely understand the latest coal sample detection situation and provide timely feedback and support for decision-making. Through the result interpretation and analysis function, users can deeply understand the content and influencing factors of the coal composition, and provide scientific basis and guidance for coal production and utilization. The present disclosure can realize the output and display of the coal sample detection results of the coal sample, provide more intelligent and convenient coal sample detection services for the coal industry, and promote the effective management and utilization of coal resources.

[0385] The present disclosure also provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, cause the computer to perform the coal detection method described above.

[0386] Those skilled in the art can understand that all or part of the steps in the method of implementing the above embodiments can be completed by programs instructing related hardware, the programs are stored in a storage medium, and include a plurality of instructions for causing a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the method described in various embodiments of the present disclosure. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0387] The above describes the optional embodiments of the present disclosure in detail in combination with the drawings, but the embodiments of the present disclosure are not limited to the specific details in the above embodiments. Within the technical concept range of the embodiments of the present disclosure, various simple modifications can be made to the technical solutions of the embodiments of the present disclosure, and these simple modifications all belong to the protection range of the embodiments of the present disclosure. In addition, it should be noted that various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present disclosure will not further describe various possible combinations.

[0388] In addition, various different embodiments of the present disclosure can also be combined in any manner, as long as they do not deviate from the idea of the embodiments of the present disclosure, and they should also be considered as disclosed by the embodiments of the present disclosure.

Claims

1. A coal detection system, the system comprising: a detection unit configured to collect spectral data of a detection coal sample; wherein, the detection unit comprises a near-infrared spectrum collection module and an X-ray fluorescence spectrum collection module; a training unit configured to perform neural network construction in each neural network dimension based on a greedy search through the spectral data, and train a coal sample detection model based on simulated samples after sample augmentation; an analysis unit configured to perform fusion processing on the spectral data to obtain target spectral data, perform target spectral data inference based on the coal sample detection model, and obtain a coal detection result.

2. The system of claim 1, wherein, The system further comprises a sample preparation unit configured to collect a coal sample and perform sample preparation processing on the coal sample to obtain a detection coal sample.

3. The system of claim 2, wherein, The sample preparation unit comprises: a sampling module configured to randomly collect raw coal samples during raw coal transportation or storage, or configured to receive raw coal samples stored by a user; a crushing module configured to perform crushing processing on the raw coal sample to obtain a crushed coal sample; a processing module configured to perform pretreatment on the crushed coal sample to obtain a base coal sample; a shaping module configured to perform shaping processing on the base coal sample to obtain a detection coal sample.

4. The system of claim 3, wherein, The pretreatment on the crushed coal sample comprises: drying processing, grinding processing, and screening processing.

5. The system of claim 3 or 4, wherein, The sampling module, the crushing module, the processing module, and the shaping module are connected based on a transportation conveyor belt; The transportation conveyor belt is triggered and controlled based on a servo system.

6. The system of claim 5, wherein, The servo system is configured to: start timing based on a trigger signal of a position trigger at a preset position of each module of the sample preparation unit, and control a power servo motor of the transportation conveyor belt to start until a predetermined time is reached, and then turn off the power servo motor of the transportation conveyor belt; or start timing in response to a corresponding switch trigger signal of each module of the sample preparation unit, and control a power servo motor of the transportation conveyor belt to start until a predetermined time is reached, and then turn off the power servo motor of the transportation conveyor belt.

7. The system of any one of claims 3 to 6, wherein, The shaping module comprises: a coal conveying member (1) comprising a conveying frame (11) and a conveying belt (12) rotatably arranged on the conveying frame (11), and a detection coal sample inlet is arranged on the conveying frame (11); a first pretreatment member (2) comprising a limiting frame (21) and a scraper (22), the limiting frame (21) is supported above the conveying belt (12) by the conveying frame (11), and the limiting frame (21) is located below the detection coal sample inlet, and the scraper (22) is arranged at the discharge end of the limiting frame (21) and forms a first gap with the conveying belt (12); a second pretreatment member (3) comprising at least one compression roller (31), the compression roller (31) is rotatably arranged on the conveying frame (11), and the compression roller (31) is located at the outlet end of the limiting frame (21), and the compression roller (31) and the conveying belt (12) form a second gap in communication with the first gap, and the compression roller (31) is provided with a limiting member (32) for limiting the width of the second gap.

8. The system of claim 7, wherein, The limiting piece (32) comprises two limiting wheels (321) coaxially arranged on the compression roller (31), and the two limiting wheels (321) are arranged at intervals along the extension direction of the compression roller (31) to form the second gap between the two limiting wheels (321).

9. The system of claim 8, wherein, The inner side of the limiting wheel (321) is provided with a slope surface.

10. The system of any one of claims 7 to 9, wherein, Two mounting racks (112) are arranged at intervals on the conveying frame (11) along a direction perpendicular to the conveying direction of the conveying belt (12), a rotating shaft (311) is arranged on the compression roller (31), and the two ends of the rotating shaft (311) are rotatably arranged on the two mounting racks (112), respectively.

11. The system of claim 10, wherein, when the compression roller (31) is a plurality of compression rollers (31), the ends of the rotating shafts (311) of the plurality of compression rollers (31) away from the first driving piece (33) are each provided with a sprocket wheel (3111), the plurality of sprocket wheels (3111) are connected by a chain (3112), and the first driving piece (33) is used to drive one of the rotating shafts (311) to rotate; and / or the rotating shaft (311) is further provided with two blocking wheels (3113), and the two mounting racks (112) are each provided with a blocking plate (1121), and the two blocking wheels (3113) are supported on the inner sides of the two blocking plates (1121), respectively.

12. The system of any one of claims 7 to 11, wherein, The output ends of the two side edges of the limiting frame (21) are each provided with a limiting plate (211), the limiting plate (211) extends in an arc shape, the limiting plate (211) is arranged opposite to the compression roller (31), and the curvature of the limiting plate (211) matches the curvature of the compression roller (31).

13. The system of any one of claims 7 to 12, wherein, The upper cover of the second pretreatment component (3) is provided with a protective cover (34).

14. The system of any one of claims 7 to 13, wherein, The upper side of the coal sample inlet is provided with a feeding hopper (13).

15. The system of any one of claims 7 to 14, wherein, The conveying frame (11) is provided with a mounting position (14) for mounting a detection module, and the mounting position (14) is located on the side of the second pretreatment component (3) away from the first pretreatment component (2).

16. The system of any one of claims 1 to 15, wherein, The near-infrared spectrum acquisition module comprises an X-ray tube system, a detector system and a collimator, the collimator is in a right-angled horn structure, one side of the hypotenuse of the collimator is arranged close to the detector system, and has a suppression effect on X-rays emitted at a large angle close to the detector system; one side of the right angle of the collimator is arranged away from the detector system, and has no suppression effect on X-rays emitted at a large angle away from the detector system.

17. The system of claim 16, wherein, The collimator is integrally arranged on the X-ray tube of the X-ray tube system.

18. The system of claim 16 or 17, wherein, The collimator is connected to the X-ray tube of the X-ray tube system through an X-ray tube interface.

19. The system of any one of claims 1 to 18, wherein, The detection unit further comprises: a sampling cover (4) for setting the near-infrared spectrum acquisition module and the X-ray fluorescence spectrum acquisition module; the inside of the sampling cover is further provided with an illumination module (41), an optical fiber (42) and a photoelectric sensor (43); A light barrier (44) is arranged between the illumination module (41) and the optical fiber (42) and extends from the inner wall of the sampling cover; The photoelectric sensor (43) is arranged in the light path direction of the illumination module and is used for monitoring the illumination intensity of the corresponding illumination module (41).

20. The system of claim 19, wherein, The illumination module (41) comprises a plurality of symmetrically arranged light sources; The optical fiber (42) is arranged at the center of the top of the sampling cover (4); Each illumination module (41) corresponds to at least one photoelectric sensor (43) arranged on the inner wall of the sampling cover.

21. The system of claim 20, wherein, The analysis unit is also used for monitoring the state of the corresponding illumination module based on the illumination intensity collected by the photoelectric sensor, comprising: The photoelectric sensor collects illumination data, which is preprocessed including filtering and denoising; According to the voltage value or digital value of the corresponding electric signal of the preprocessed illumination data, the calibration curve between the illumination intensity and the voltage value or digital value is calculated by interpolation or fitting, and the illumination intensity of the detection position is obtained; Based on the illumination intensity of the detection position and the positional relationship between the light point sensor and the corresponding illumination module, the illumination intensity of the corresponding illumination module is fitted to obtain the detection illumination intensity of the corresponding illumination module; The detection illumination intensity is compared with the preset illumination intensity threshold value, and if the detection illumination intensity is less than the preset illumination intensity threshold value, an alarm information is output.

22. The system of any one of claims 1 to 21, wherein, The detection unit further comprises: The mica membrane replacement device comprises a sample conveying belt (5), a detection coal sample (6) is placed on the sample conveying belt (5), a mounting rack (7) is arranged above the detection coal sample (6), a driving wheel (71) and a driven wheel (72) are arranged on the mounting rack (7), a mica membrane transmission channel (73) is formed on the mounting rack (7), and the mica membrane (8) can pass through the mica membrane transmission channel (73) and abut against the lower edge of the outer circumferential surface of the driving wheel (71) and the driven wheel (72), and move under the driving of the driving wheel (71).

23. The system of any one of claims 1 to 22, wherein, The system further comprises a visual analysis module, the visual analysis module comprises one or more image acquisition modules, each image acquisition module is arranged in each preset direction of the detection position of the detection coal sample, and is used for acquiring image information of the detection coal sample in the view angle; The analysis unit is also used for evaluating the shape of the detection coal sample based on the image information in each view angle.

24. The system of claim 23, wherein, The evaluation of the shape of the detection coal sample based on the image information in each view angle comprises: The image information in each view angle is sequentially subjected to denoising, grayscale and edge detection processing, and the detection coal sample region is calibrated; The shape size of the detection coal sample is fitted based on the image detection coal sample region in each view angle; The surface flatness index of the detection coal sample is fitted based on the surface flatness index collected from the image detection coal sample region in each view angle; The shape of the detection coal sample is evaluated based on the fitted shape size and the fitted surface flatness.

25. The system of claim 24, wherein, The evaluation of the shape of the detection coal sample based on the fitted shape size and the fitted surface flatness comprises: The Euclidean distance between the fitted shape size and the preset shape size is calculated as a first deviation value; Calculate the Euclidean distance between the fitted surface flatness and the preset surface flatness, and use it as the second deviation value; Normalization is performed on the first deviation value and the second deviation value, and the arithmetic mean of the normalized first deviation value and the second deviation value is calculated to obtain the deviation coefficient. The preset correction coefficient in the coal sample detection model is assigned a value based on the deviation coefficient.

26. The system of any one of claims 1 to 25, wherein, The spectral data includes near-infrared spectral data and X-ray fluorescence spectral data; The step of performing fusion processing on the spectral data to obtain the target spectral data includes: The near-infrared spectral signal and the X-ray fluorescence spectral data are fused to obtain the target spectral data.

27. The system of claim 26, wherein, The near-infrared spectral data and the X-ray fluorescence spectral data are fused to obtain target spectral data, including: Preprocessing was performed on the near-infrared spectral data and the X-ray fluorescence spectral data respectively; Downsampling was performed on the preprocessed near-infrared spectral data and X-ray fluorescence spectral data, and the downsampled near-infrared spectral data and X-ray fluorescence spectral data were scaled to a normal distribution. The target spectral data is obtained by stitching together normally distributed near-infrared spectral data and normally distributed X-ray fluorescence spectral data.

28. The system of claim 27, wherein, The preprocessing of near-infrared spectral data and X-ray fluorescence spectral data includes: SG convolution smoothing was performed on the near-infrared spectral data and X-ray fluorescence spectral data respectively to obtain the near-infrared spectral data and X-ray fluorescence spectral data after SG convolution smoothing. Area normalization was performed on the near-infrared spectral data after SG convolution smoothing.

29. The system of claim 28, wherein, The SG convolution smoothing process performed on the near-infrared spectral data and X-ray fluorescence spectral data includes: The coefficients of the SG convolution kernel are calculated based on the preset window size and polynomial order. For near-infrared spectral data and X-ray fluorescence spectral data respectively, the weighted average of adjacent data points in each spectral data is performed based on the coefficients of the SG convolution kernel to obtain smoothed data points; Perform symmetric expansion or zero-filling on the boundaries of each spectral data; The spectral data after SG convolution smoothing is obtained based on the smoothed data points and the processed boundaries.

30. The system of claim 28 or 29, wherein, The area normalization process for the near-infrared spectral signal after SG convolution smoothing includes: Area normalization is performed on each near-infrared spectral signal after SG convolution smoothing to obtain the area-normalized near-infrared spectral signal.

31. The system of claim 30, wherein, The calculation rule of the near-infrared spectrum signal after area normalization processing is: wherein This represents the NIR reflectance corresponding to the m-th wavelength point of the near-infrared spectral signal; X NIR For the near infrared spectrum signal before area normalization processing, This is the near-infrared spectral signal after area normalization.

32. The system of any one of claims 27-31, wherein, The downsampling processing performed on the preprocessed near-infrared spectral data and X-ray fluorescence spectral data includes: Downsampling was performed on the X-ray fluorescence spectral data after SG convolution smoothing and the near-infrared spectral data after area normalization, including: The filtering processing is respectively performed on each spectrum data, in the spectrum data after the filtering processing, a sampling point is reserved every fixed interval to obtain a plurality of sampling points; or, a plurality of sections are cut from the spectrum data after the filtering processing, and the average processing is performed on each sampling point in each section to obtain one sampling point in each section, and a plurality of sampling points are obtained; Signal reconstruction is performed based on each sampling point to obtain the spectrum data after the down-sampling processing.

33. The system of any one of claims 27-32, wherein, The near-infrared spectrum data and the X-ray fluorescence spectrum data after the down-sampling are scaled to a normal distribution, comprising: Statistical characteristics of each spectrum data are respectively calculated; wherein the statistical characteristics are variance and / or standard deviation; Based on the statistical characteristics of each spectrum data, the standardization processing or the normalization processing is performed on each spectrum data, and the numerical value of each spectrum data is scaled to a preset numerical range to obtain the scaled numerical value of each spectrum data; Based on the preset transformation algorithm and the scaled numerical value of each spectrum data, each spectrum data is converted to a normal distribution.

34. The system of claim 27, wherein, The normal distribution of the near-infrared spectrum data and the normal distribution of the X-ray fluorescence spectrum data are subjected to splicing processing to obtain target spectrum data, comprising: Performing alignment operation of the normal distribution of the near-infrared spectrum data and the normal distribution of the X-ray fluorescence spectrum data; After completing the alignment of the normal distribution of the near-infrared spectrum signal and the normal distribution of the X-ray fluorescence spectrum data, the splicing of the two spectrum signals is performed based on weighted average to obtain initial target spectrum data; Based on the normal distribution of the near-infrared spectrum signal and / or the normal distribution of the X-ray fluorescence spectrum data, the initial target spectrum data is verified, and the initial target spectrum data that passes the verification is taken as the target spectrum data.

35. The system of any one of claims 1 to 34, wherein, the coal sample detection model is: θ = argmaxθL2(f(X | θ), Y); where f(· | θ) is a deep neural network with θ as a parameter; the target spectrum data; the coal component.

36. The system of any one of claims 1 to 35, wherein, The training unit is specifically configured to: collect historical near-infrared spectrum signals and historical X-ray fluorescence spectrum data, and construct corresponding historical target spectrum data based on the historical near-infrared spectrum signals and the historical X-ray fluorescence spectrum data; perform PLS model parameter initialization by taking the historical target spectrum data as training data; generate simulation samples based on the initialized PLS model, perform model training based on the simulation samples, and obtain a coal sample detection initial model; perform verification on the coal sample detection initial model based on the reserved historical target spectrum data, and obtain a coal sample detection model.

37. The system of claim 36, wherein, The PLS model parameters include: weights, biases, mean and variance of input data of a training set, and mean and variance of output data.

38. The system of claim 36 or 37, wherein, The generating simulation samples based on the initialized PLS model comprises: adaptively determining a signal type and signal parameters matched with the target spectrum data, and generating a basic signal based on the signal type and the signal parameters; adaptively selecting an augmentation scheme, and performing a data augmentation operation on the basic signal based on the selected augmentation scheme to obtain an augmented signal set; deviation data is screened out in the augmented signal set, and the screened-out augmented signal set is taken as simulation samples.

39. The system of claim 38, wherein, The adaptive selection of augmentation schemes includes: One or more schemes are randomly selected from noise addition, translation, and preset rules as pre-selected schemes; wherein, the preset rules are: X aug = aX1+ (1-a)X2, a e [0, 1] y aug = f PLS (X aug ) where f PLS represents the PLS mapping function; X aug and y aug represent the synthesized NIRS-XRF signal and the corresponding pseudo label, respectively; a represents the interpolation weight randomly sampled in the range of [0, 1]; X1 and X2 are randomly selected signal samples from the historical target spectral data of the coal sample. Randomly adjust the parameters within the preset adjustable parameter range of each pre-selected scheme to obtain a pre-selected scheme with determined parameters; If there is only one pre-selected scheme with defined parameters, then the basic signal is directly processed based on that scheme to obtain an augmented signal; If there are multiple options for the pre-selected scheme with defined parameters, then each scheme is executed sequentially to obtain an augmented signal.

40. The system of any one of claims 36-39, wherein, The process of training the model based on the simulated samples to obtain an initial model for coal sample detection includes: The simulated samples are pseudo-labeled based on the PLS model after parameter initialization, and the labeled simulated samples are used as training samples. Based on the training samples, model training is performed in a neural network searched based on the target spectral data to obtain an initial model for coal sample detection.

41. The system of any one of claims 36-40, wherein, The process of validating the initial coal sample detection model based on reserved historical target spectral data to obtain the coal sample detection model includes: A validation set is constructed based on reserved historical target spectral data, and the validation set is input into the initial model for coal sample detection to obtain the corresponding prediction results. Based on the prediction results and the actual results of the corresponding reserved historical target spectral data, the initial model performance of coal sample detection is evaluated; If the initial model performance evaluation for coal sample testing fails, then perform hyperparameter adjustment, model structure optimization, and / or regularization operations to obtain an updated initial model for coal sample testing. The performance of the updated initial model for coal sample detection is re-evaluated based on the reserved historical target spectral data until an initial model for coal sample detection that meets the preset performance requirements is obtained, which is then used as the coal sample detection model.

42. The system of claim 41, wherein, The initial model performance evaluation for coal sample testing is conducted based on the predicted results and the actual results of the corresponding reserved historical target spectral data, including: By comparing the predicted results with the actual results, the accuracy and recall of the model are evaluated based on the deviation between the two. If either accuracy or recall fails the evaluation, the initial model performance evaluation for this coal sample testing will fail.

43. The system of any one of claims 36-42, wherein, The training unit is also used to perform neural network search based on the spectral data, including: Multiple optimization dimensions are defined based on the structural parameters of the neural network; Within each optimization dimension, the parameters within the corresponding neural network structure are adaptively adjusted, and the MAE index is calculated after each adjustment. Compare the MAE index corresponding to each parameter, and determine the parameter combination with the highest MAE index as the optimization result in the corresponding optimization dimension. Perform a greedy search across each optimization dimension to obtain the optimization results for each optimization dimension; Based on the optimization results of each optimization dimension, the structural parameters of each corresponding neural network structure are determined, and the neural network corresponding to the search results is constructed.

44. The system of claim 43, wherein, The step of performing a greedy search across each optimization dimension to obtain the optimization results for each optimization dimension includes: Determine the search order and perform a greedy search across each optimization dimension based on the determined search order to obtain the optimization results for each optimization dimension.

45. The system of claim 44, wherein, The search order is: search basic operator, search input resolution, search network depth, and search network width; The method comprises the following steps of: determining a search order and performing a greedy search among each optimization dimension based on the determined search order to obtain an optimization result of each optimization dimension, comprising: determining a search range of a basic operator, randomly generating N groups of different configurations of input resolution, network depth and network width, determining an optimal operator for each group for each group of configurations, outputting N optimal operators for the N groups of configurations, and taking the operator with the highest frequency of occurrence in the N optimal operators as the searched basic operator; determining a search range of input resolution, fixing the searched basic operator as the optimal configuration, randomly generating M groups of different configurations of network depth and network width, determining an optimal input resolution for each group for each group of configurations, outputting M optimal input resolutions for the M groups of configurations, and taking the input resolution with the highest frequency of occurrence in the M optimal input resolutions as the searched input resolution; determining a search range of network depth, fixing the searched basic operator and the searched input resolution as the optimal configuration, randomly generating L groups of different configurations of network width, determining an optimal network depth for each group for each group of configurations, outputting L optimal network depths for the L groups of configurations, and taking the network depth with the highest frequency of occurrence in the L optimal network depths as the searched network depth; 46. The system of any one of claims 43-45, wherein, determining a search range of network width, fixing the searched basic operator, the searched input resolution and the searched network depth as the optimal configuration, and searching for an optimal network width. The neural network dimensions comprise:

47. The system of claim 46, wherein, a basic operator dimension of a neural network layer, a resolution dimension of input data, a network depth dimension and a network width dimension. In the basic operator dimension of the neural network layer, the corresponding optimization result obtaining rule comprises: traversing the convolution type of the linear layer and the activation function type of the nonlinear layer, adaptively combining the convolution type of the fully connected layer and the activation function type of the nonlinear layer, and calculating the MAE index after each combination; 48. The system of claim 47, wherein, comparing the MAE index after each combination, selecting the combination corresponding to the maximum MAE index, and determining the convolution type of the fully connected layer and the activation function type of the nonlinear layer corresponding to the combination as the optimization result in the basic operator dimension of the neural network layer. The convolution type of the linear layer is: a fully connected layer, a 1D convolution layer with a kernel size of 5, a 1D convolution layer with a kernel size of 10 or a 1D convolution layer with a kernel size of 15; The activation function type of the nonlinear layer is:

49. The system of claim 47 or 48, wherein, The calculation rule of MAE index is: wherein a TanH function, an ELU function, a Sigmoid activation function or a Softmax function. y i the model prediction value corresponding to the i-th training data pair; a true value of a coal component corresponding to the i th training data; 50. The system of any one of claims 1-49, wherein, m is the size of the data set.

51. The system of any one of claims 1 to 50, wherein, The system further comprises an output unit for visualizing the coal detection result. The coal detection result comprises:

52. The system of claim 50 or 51, wherein, one or more of ash component, ash content, volatile matter, carbon and hydrogen, ash melting point, total water, total sulfur and calorific value. The visualization of the coal detection result comprises: in response to a user data query instruction, determining a corresponding detection result object; based on the corresponding detection result object, selecting a preset data visualization scheme, and pushing the visualized data to the user end. 53.A coal detection method applied to the coal detection system of any one of claims 1-52, the method comprising: collecting spectral data of a detection coal sample; performing neural network construction in each neural network dimension based on a greedy search through the spectral data, and training a coal sample detection model based on simulated samples after sample augmentation; performing fusion processing on the spectral data to obtain target spectral data, performing target spectral data inference based on the coal sample detection model, and obtaining a coal detection result. 54.A computer-readable storage medium having instructions stored thereon, which, when executed on a computer, cause the computer to perform the coal detection method of claim 53.

Images