Method and system for rapidly measuring coal moisture content based on convolutional neural network

Abstract

The invention relates to the technical field of coal moisture content measurement, in particular to a coal moisture content rapid measurement method and system based on a convolutional neural network. Coal flow hyperspectrums, images and pulse flows are synchronously collected, aligned segmentation is executed, and evaporation stages are divided based on spectral reflectivity time sequence fluctuation to obtain segmented codes; and constructing water film roughness distribution by combining image gray and gradient features, cascading a time-space sequence to generate a fusion feature vector, implementing time scaling and interpolation filling on an effective fragment according to a belt speed, and aggregating and outputting the coal moisture content. According to the method, hyperspectral and image pulse streams are associated to align data and eliminate missing frames, reflectivity time sequence fluctuation is analyzed to divide evaporation stages and establish indexes, gray and gradient features are fused to construct a roughness texture matrix, space-time sequences are cascaded to generate feature vectors, and sequence scaling and interpolation filling are dynamically executed according to belt speed. Speed interference is eliminated, spectrum deviation is corrected, and finally accurate analysis of the coal flow moisture under the variable running speed working condition is achieved.

Description

Technical Field

[0001] This invention relates to the field of coal moisture content measurement technology, and in particular to a rapid coal moisture content measurement method and system based on convolutional neural networks. Background Technology

[0002] The field of coal moisture content measurement technology encompasses the analysis and determination of the state, composition, and distribution characteristics of moisture in coal. Core components include the collection, preservation, and pretreatment of moisture in raw coal, washed coal, and blended coal samples; the differentiation and definition of different moisture indices such as total moisture, air-dried moisture loss, and analytical basis moisture; and the setting and execution of moisture measurement steps under different operating conditions. This technical field systematically covers the design of experimental procedures, control of measurement conditions, and correction of measurement results for rapid or conventional determination of coal sample moisture using thermogravimetric analysis, drying methods, infrared moisture measurement, near-infrared spectroscopy, and characterization methods based on images and optical features. It also includes the rational configuration of moisture detection frequency, detection location, and detection methods in mine production, coal preparation plant processing, coal storage and transportation, and coal blending processes to support the moisture data needs in application scenarios such as coal quality evaluation, combustion control, and material settlement. Based on the above, hyperspectral detection has been gradually used to measure the moisture content of coal. By acquiring hyperspectral reflectance information of the coal surface in the range of visible light to near-infrared and even wider bands, the spectral absorption characteristics, reflectance characteristics and energy distribution at different wavelengths are analyzed to characterize the correspondence between the moisture content and spectral response in coal, providing a spectral data basis for the rapid and non-destructive determination of coal moisture.

[0003] The rapid coal moisture content measurement method based on convolutional neural networks refers to a measurement method that uses a convolutional neural network model to extract features and estimate moisture content from raw image or spectral data of coal samples. The technical aspects mainly cover coal sample surface or cross-sectional image data acquisition, lighting conditions and shooting distance settings, image resolution and field of view limitations, numerical normalization and calibration of raw image or spectral data, construction of a training sample set based on coal sample moisture calibration values, multi-level convolution calculations and feature mapping of texture features, brightness distribution features, and chromaticity features in image or spectral data using a convolutional neural network structure, and a series of operational procedures and calculation steps in online or offline detection scenarios, inputting the coal sample image or spectral data to be tested into the trained convolutional neural network to output the corresponding moisture content measurement results. In hyperspectral applications, it is no longer limited to a single... Instead of collecting static spectral data at any given moment, this method acquires a sequence of hyperspectral changes on the coal surface over a specific time range, recording the dynamic spectral characteristics of the coal's moisture evaporation or adsorption process. By serializing and digitally encoding multi-band hyperspectral reflectance data at different time points, a hyperspectral time-series dataset describing the coal's moisture change process is obtained. This hyperspectral time-series data is then paired with corresponding moisture content calibration values ​​to construct a numerical mapping relationship between the coal's hyperspectral time series and its moisture content. Specifically, this involves segmenting, resampling, and normalizing the hyperspectral time-series data, extracting key band responses and their combined features that change over time, and modeling these hyperspectral time-series features using time-series modeling methods. This establishes the correspondence between the hyperspectral time-series features and the coal's moisture content, thus forming the structure and process of a rapid coal moisture content measurement method based on hyperspectral information.

[0004] Existing technologies only collect static spectra at a single moment or time series at a fixed frequency, lacking accurate correlation with the real-time motion state of the conveyor belt. In scenarios with nonlinear fluctuations in coal flow speed, fixed sampling intervals cannot maintain a strict correspondence with physical spatial locations, causing time series data to be stretched, deformed, or misaligned on the time axis. Furthermore, relying solely on spectral numerical mapping ignores the light and shadow interference caused by the coal surface accumulation morphology. When there are large height differences or non-uniform gradients on the coal pile surface, local shadows and changes in illumination angles mask the true spectral response, leading to a decrease in the generalization ability of the moisture content inversion model and drift in the measurement values. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for rapid measurement of coal moisture content based on convolutional neural networks. The technical solution is as follows: A rapid method for measuring the moisture content of coal based on convolutional neural networks includes the following steps: S1: Acquire hyperspectral frames of coal flow and record the time and wavelength order, synchronously acquire visible light and near-infrared images and mark the time, read encoder pulses and associate them with time, align them by time and remove missing frames, check the continuity of time and then segment and record the range, and generate a synchronous frame group for coal flow moisture measurement. S2: Based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, select the analysis unit, extract the moisture-sensitive band to construct the reflectance time series, identify the candidate turning point, divide the evaporation stage according to the set threshold, and generate the moisture time series segmented coding result. S3: Call the visible light and near-infrared images in the synchronous frame group of coal flow moisture measurement, analyze the visible light neighborhood and near-infrared brightness, extract the gray-level abrupt change and brightness transition to form height and gradient feature pairs, construct a two-dimensional distribution matrix according to the block statistical frequency, and generate a joint distribution map block group of water film roughness. S4: Scan the two-dimensional matrix in the joint distribution map of water film roughness, aggregate adjacent elements to extract local patterns, splice them to form map descriptors and concatenate them into a spatial description sequence, combine the moisture temporal segmentation coding results to rearrange and mark the length, and obtain the coal moisture fusion feature vector.

[0006] As a further embodiment of the present invention, the coal flow moisture measurement synchronization frame group includes a unified time reference for multi-source data, an effective coal flow segment identifier, a frame sequence continuity marker, and a spatial correspondence index. The moisture temporal segmentation coding result includes an evaporation stage type identifier, a stage boundary position code, a stage duration parameter, and a band association index. The water film roughness joint distribution map block group includes a block texture distribution matrix, roughness level combination features, local structure statistical results, and map block spatial numbering. The coal moisture fusion feature vector includes temporal moisture feature components, spatial texture feature components, spatiotemporal association identifiers, and vector structure description information.

[0007] As a further aspect of the present invention, the step of obtaining S1 is as follows: S101: Obtain the hyperspectral reflectance frame series of the coal flow area above the conveyor belt, record the acquisition time mark for each frame and record the corresponding wavelength arrangement order, and simultaneously acquire the visible light image stream and near-infrared image stream in the same coal flow area and mark the acquisition time of each image. Read the encoder pulse stream corresponding to the motion state of the conveyor belt and associate it with the time mark to generate a multi-source time mark dataset. S102: Based on the multi-source time-marked dataset, perform frame-by-frame comparison according to the acquisition time of the hyperspectral frame and the acquisition time of the visible light image frame and the near-infrared image frame, perform same-time window matching for encoder pulse records, remove records whose time deviation exceeds the alignment tolerance value, and rearrange the retained frame order according to the acquisition time order to obtain the alignment frame order index table. S103: Based on the alignment frame sequence index table, the continuity of the acquisition time difference between adjacent frames is judged, and the position where the difference exceeds the continuity threshold is taken as the segmentation point. The frame sequence is divided into segments, and the start and end times and frame sequence number range of each segment are recorded to generate a coal flow moisture measurement synchronization frame group.

[0008] As a further aspect of the present invention, the step of obtaining S2 is as follows: S201: Based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, for each coal flow segment, select a pixel point or a grid unit composed of multiple pixels in the coal flow area as an analysis unit, collect the reflectance value of the moisture-sensitive band corresponding to the analysis unit at each collection time, arrange them in the order of collection time, and record the difference of reflectance values ​​between adjacent frames to generate a reflectance time series value sequence. S202: Based on the reflectance time-series numerical sequence, the reflectance difference between adjacent frames is judged frame by frame. The rising state, falling state or stable state is marked according to the change of the sign of the difference. The position where the sign of adjacent states reverses is located. The corresponding reflectance change amplitude and the number of consecutive frames are extracted and compared with the evaporation stage discrimination threshold to obtain the stage boundary position sequence. S203: Based on the stage boundary position sequence, the reflectance time series is divided into stages according to the boundary position to form an initial evaporation stage, a stable evaporation stage, and a tail drying stage. The reflectance change trajectory is summarized according to the sliding window for each stage and a stage change contour is generated. The contours of each stage are spliced ​​according to the acquisition time sequence and merged according to the wavelength order to establish an index and generate a moisture time series segmented coding result.

[0009] As a further aspect of the present invention, the step of obtaining S3 is as follows: S301: Call the visible light image frame in the coal flow moisture measurement synchronization frame group, perform difference calculation on the gray values ​​of adjacent pixels in the analysis unit, judge according to the gray value difference and the brightness change threshold, locate the gray value change position, and map the corresponding position to the height level value in the analysis unit to generate a height level mark sequence. S302: Based on the near-infrared image frames in the coal flow moisture measurement synchronous frame group, perform difference judgment on the brightness values ​​of adjacent pixels in the analysis unit, filter continuously changing segments according to the brightness change amplitude and transition judgment threshold, and map the segments to gradient level values ​​in the analysis unit to obtain gradient level labeling sequence. S303: Based on the height level marker sequence and gradient level marker sequence, pair them together in the same analysis unit to form feature pairs, perform block processing on the image region according to the preset block size, count the frequency of feature pair combinations in each block region and arrange them into a two-dimensional numerical matrix to generate a water film roughness joint distribution map block group.

[0010] As a further aspect of the present invention, the step of obtaining S4 is as follows: S401: Based on the two-dimensional distribution matrix in the joint distribution map block group of water film roughness, read the matrix element values ​​one by one in the block region order, perform neighborhood aggregation operation on adjacent matrix elements, extract local value arrangement patterns according to the combination relationship of adjacent element values, and record them in the scanning order to form a sequence, generating a local pattern sequence set. S402: Based on the local pattern sequence set, the local pattern sequences are spliced ​​together in each block region according to the spatial arrangement order, and the splicing results are numbered and stored according to the block number to form the numerical description vector of the corresponding block region, thus obtaining the block description subset; S403: Based on the block description subset, the block description vectors are concatenated along the coal flow transport direction and the lateral arrangement order to construct a spatial arrangement numerical sequence covering the coal flow area. At the same time, the segmented coding vectors of each analysis unit in the moisture time-series segmented coding result are called and sorted according to the acquisition time to generate a whole frame time-series feature sequence. S404: Based on the acquisition time corresponding to the whole frame temporal feature sequence and the spatially arranged numerical sequence, perform time consistency judgment and complete sequence alignment, perform concatenation and order rearrangement of the aligned sequence, and add sequence length identifier to generate coal moisture fusion feature vector.

[0011] As a further aspect of the present invention, the neighbor aggregation operation performed on adjacent matrix elements is specifically defined as follows: Centered on any matrix element in the two-dimensional distribution matrix, a neighborhood window is formed by selecting a fixed number of matrix elements, including that matrix element, and combining the matrix element values ​​within the neighborhood window in a row-major scanning order. The extraction of local numerical arrangement patterns based on the numerical combination relationship of adjacent elements is specifically limited to: The matrix element values ​​within the neighborhood window are arranged according to a preset rule to form a non-repeating numerical sequence, and the numerical sequence is used as the confirmation condition for the local pattern sequence. The specific limitation of storing the splicing results by block number is as follows: Numerical description vector indices for the block regions are generated using a numbering rule consistent with the spatial numbering of the block regions in the joint distribution map of water film roughness. The additional sequence length identifier is specifically defined as follows: The effective length values ​​of the entire frame temporal feature sequence and the spatially arranged numerical sequence are recorded in the coal moisture fusion feature vector and stored as part of the vector structure description information.

[0012] As a further aspect of the present invention, the method further includes: S5: Based on the conveyor belt speed mark and pixel dwell frame number mark in the coal flow moisture measurement synchronization frame group, filter the effective coal flow time segments that meet the threshold, select the time scaling rule according to the conveyor belt speed, perform time scaling, rearrangement and interpolation filling on the moisture measurement value sequence in the effective coal flow time segment based on the moisture time sequence segmentation coding result, align with the corresponding sub-vector of the fusion feature vector and jointly aggregate to generate the coal moisture content measurement result. The specific results of the coal moisture content measurement include the moisture content value corresponding to the coal flow, the continuous time output identifier, the correspondence of valid measurement segments, and the online measurement result mark.

[0013] As a further aspect of the present invention, the step of obtaining S5 is as follows: S501: Based on the conveyor belt speed mark and pixel dwell frame number mark in the coal flow moisture measurement synchronization frame group, read the pixel dwell frame number value one by one for each coal flow time segment, compare the dwell frame number with the set dwell frame number threshold, record the index corresponding to the time segment that meets the threshold condition, and synchronously associate it with the corresponding conveyor belt speed mark to obtain the effective coal flow time segment index set. S502: Based on the effective coal flow time segment index set, select time scaling rules according to the conveyor belt speed mark corresponding to each time segment, call the moisture measurement value sequence of the corresponding time segment in the moisture time series segment coding result, scale the time axis of the sequence proportionally, and perform sequential rearrangement and numerical interpolation filling on the scaled time points to obtain the time-scaled moisture time series. S503: Based on the time-stretched moisture time series, align the corresponding feature sub-vectors in the coal moisture fusion feature vector point by point according to the collection time sequence, perform joint aggregation operation on the aligned values ​​and output a single numerical result to generate the coal moisture content measurement result.

[0014] A rapid coal moisture content measurement system based on convolutional neural networks, the system comprising: The multi-source frame synchronization module acquires hyperspectral frames of coal flow and records the time and wavelength order, synchronously acquires visible light and near-infrared images and marks the time, reads encoder pulses and associates them with time, aligns them according to time and removes missing frames, checks the continuity of time and then segments them and records the range, and generates a coal flow moisture measurement synchronization frame group. The spectral time series segmentation module, based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, selects an analysis unit, extracts moisture-sensitive bands to construct reflectance time series, identifies candidate turning points, divides the evaporation stage according to a set threshold, and generates moisture time series segmentation coding results. The texture roughness modeling module calls the visible light and near-infrared images in the synchronous frame group of coal flow moisture measurement, analyzes the brightness of the visible light neighborhood and near-infrared, extracts the height and gradient feature pairs formed by gray-level abrupt changes and brightness transitions, constructs a two-dimensional distribution matrix according to the block statistical frequency, and generates a joint distribution map block group of water film roughness. The spatiotemporal feature fusion module scans the two-dimensional matrix in the joint distribution map of water film roughness, aggregates adjacent elements to extract local patterns, splices them to form map descriptors and concatenates them into a spatial description sequence, and rearranges and marks the length of the sequence based on the moisture temporal segmentation coding results to obtain the coal moisture fusion feature vector. The timing correction estimation module filters effective coal flow time segments that meet the threshold based on the conveyor belt speed mark and pixel dwell frame number mark in the coal flow moisture measurement synchronization frame group. It selects time scaling rules according to the conveyor belt speed and performs time scaling, rearrangement and interpolation filling on the moisture measurement value sequence in the moisture time segment based on the moisture time-series segmented encoding result. It aligns with the corresponding sub-vectors of the fused feature vector and aggregates them together to generate the coal moisture content measurement result.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this invention, by associating and aligning hyperspectral frames with image streams and encoder pulses and removing missing frames, a reflectance time series is constructed for the coal flow region. The inflection points of the fluctuating state are identified to divide the initial evaporation and stable evaporation stages. The changing contours are aggregated and indexed. By combining visible light grayscale differences with near-infrared brightness transition band mapping height and gradient levels, a two-dimensional distribution matrix reflecting texture is constructed. Spatial descriptors are cascaded and time series features are connected in series. The effective segments are dynamically adjusted according to the conveyor belt speed markings. Time scaling, rearrangement, and interpolation filling are performed on the measurement sequence to eliminate motion speed fluctuation interference. The spectral response deviation is corrected by combining surface roughness features, thereby achieving accurate analysis of coal flow moisture under variable conveyor speed conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the process for acquiring synchronous frame groups for coal flow moisture measurement according to the present invention; Figure 3 This is a flowchart illustrating the process of obtaining the water time-series segmented coding results of the present invention. Figure 4 This is a flowchart illustrating the process of obtaining the combined distribution map block group of water film roughness according to the present invention. Figure 5 This is a flowchart illustrating the process of obtaining the coal moisture fusion feature vector according to the present invention. Figure 6 This is a flowchart illustrating the process of obtaining the coal moisture content measurement results according to the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0019] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0020] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] Please see Figure 1 This invention provides a technical solution: a rapid method for measuring the moisture content of coal based on convolutional neural networks, comprising the following steps: S1: Obtain the hyperspectral reflectance frame series of the coal flow area above the conveyor belt, mark the acquisition time of each frame, record the corresponding wavelength arrangement order, obtain the visible light image stream and near-infrared image stream of the same coal flow area, mark the acquisition time of each image, read the encoder pulse stream corresponding to the movement state of the conveyor belt and associate it with the time mark, align the hyperspectral frames, image frames and pulse records according to the acquisition time and remove missing frames, check the temporal continuity of the aligned frame sequence and perform segmentation, record the start and end times and frame number range of each segment, and generate a synchronous frame group for coal flow moisture measurement. S2: Based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, for each coal flow segment, select pixels or grid units composed of multiple pixels in the coal flow area as analysis units, extract the reflectance of the moisture-sensitive band and construct the reflectance time series, analyze the reflectance time series fluctuation frame by frame, mark the rising, falling or stable state, identify the position where the adjacent state reverses as the turning point candidate point, compare the amplitude and persistence characteristics of the turning point candidate point with the preset evaporation stage discrimination threshold, determine the stage boundary point, divide the reflectance time series into the initial evaporation, stable evaporation and tail drying stages, use the sliding window to aggregate the trajectory of each stage to generate the change contour, stitch the contour of each stage according to time and merge them according to the band order to establish an index, and obtain the moisture time series segmented coding result; S3: Call the visible light and near-infrared image frames in the synchronous frame group for coal flow moisture measurement, analyze the neighborhood grayscale differences in the visible light image, identify the locations of abrupt changes in brightness and dark and map them as height level markers within the corresponding analysis unit, analyze the brightness changes of adjacent pixels in the near-infrared image, extract the brightness transition band and map it as gradient level markers within the corresponding analysis unit, form feature pairs with the height level and gradient level markers within the same analysis unit, divide the image into multiple block regions according to the preset size, count the frequency of occurrence of each feature pair combination within each block region, construct a two-dimensional distribution matrix reflecting the texture of the block region, and generate a joint distribution map block group of water film roughness; S4: Based on the two-dimensional distribution matrix in the joint distribution map of water film roughness, scan the two-dimensional distribution matrix corresponding to each block area one by one, perform neighborhood aggregation on adjacent matrix elements, extract local pattern sequences, and splice the local pattern sequences in spatial arrangement order in each block area to form corresponding map descriptors. Concatenate all map descriptors covering the coal flow area along the coal flow transport direction and lateral arrangement order to construct the whole frame spatial description sequence. Call the segmented coding vectors of each analysis unit in the moisture temporal segmented coding result to form the whole frame temporal feature sequence. Align the whole frame temporal feature sequence with the whole frame spatial description sequence according to the time mark and concatenate them. Rearrange the concatenated sequence according to the preset rules and add a length mark to obtain the coal moisture fusion feature vector. S5: Based on the conveyor belt speed marker and pixel dwell frame number marker in the coal flow moisture measurement synchronization frame group, compare the pixel dwell frame number with the preset frame number threshold, filter out the effective coal flow time segments that meet the dwell requirements, select the time scaling rule according to the conveyor belt speed marker corresponding to each effective coal flow time segment, and perform time scaling, rearrangement and interpolation filling on the moisture measurement value sequence in the moisture time sequence segmentation coding result within the effective coal flow time segment to obtain the time-scaled segment moisture time sequence, align and jointly aggregate it with the corresponding feature sub-vector in the coal moisture fusion feature vector to generate the coal moisture content measurement result.

[0023] The coal flow moisture measurement synchronization frame group includes a unified time reference for multi-source data, effective coal flow segment identifiers, frame sequence continuity markers, and spatial correspondence indexes. The moisture temporal segmentation coding results include evaporation stage type identifiers, stage boundary location codes, stage duration parameters, and band association indexes. The water film roughness joint distribution map block group includes a block texture distribution matrix, roughness level combination features, local structure statistical results, and map block spatial numbers. The coal moisture fusion feature vector includes temporal moisture feature components, spatial texture feature components, spatiotemporal association identifiers, and vector structure description information. The coal moisture content measurement results specifically include the moisture content value corresponding to the coal flow, the time continuous output identifier, the correspondence of effective measurement segments, and the online measurement result marker.

[0024] Please see Figure 2 The steps to obtain S1 are as follows: S101: Obtain the hyperspectral reflectance frame series of the coal flow area above the conveyor belt, record the acquisition time mark for each frame and record the corresponding wavelength arrangement order, and simultaneously acquire the visible light image stream and near-infrared image stream in the same coal flow area and mark the acquisition time of each image. Read the encoder pulse stream corresponding to the motion state of the conveyor belt and associate it with the time mark to generate a multi-source time mark dataset. To acquire hyperspectral reflectance frames of the coal flow area above the conveyor belt, an existing pushbroom hyperspectral imager was used to acquire light intensity signals in the spectral range of 400 nm to 2500 nm. The acquired light intensity signals were converted from analog to digital to generate a digital grayscale matrix, and a band index list was established in ascending order of wavelength. ,in The value is 224, corresponding to 224 spectral channels. The value of the hardware clock counter latched by the FPGA inside the spectrometer when the acquisition is triggered is read, and this value is converted into a microsecond-level timestamp. The data is then appended to the corresponding hyperspectral frame header file. Simultaneously, an industrial-grade CMOS visible light camera and an InGaAs near-infrared camera installed in the same field of view are started to perform continuous exposure acquisition. The Bayer array data of the visible light camera is read and interpolated into an RGB image matrix using a demosaicing algorithm. The single-channel brightness data of the near-infrared camera is read to generate a near-infrared grayscale image. The timestamps generated by the hardware interrupt signals of the two cameras at the moment of exposure end are read and recorded as visible light timestamps. Near-infrared timestamp The pulse signal output by the incremental A / B phase rotary encoder installed on the conveyor belt roller shaft end is monitored by a high-speed counter card, and the current pulse count value is obtained by accumulating the number of pulses. A mapping relationship between pulses and time is established by triggering and latching the pulse count value at the current moment through hardware every time a hyperspectral acquisition completion signal is received. For example, when the first is detected Time of completion of frame hyperspectral data acquisition At that time, read the current cumulative pulse count. All the collected data streams are written into a cache queue in the order of generation, and a set containing hyperspectral timestamp sequences, visible light timestamp sequences, near-infrared timestamp sequences and pulse time mapping tables is constructed to generate a multi-source time stamp dataset.

[0025] S102: Based on the multi-source time-marked dataset, perform frame-by-frame comparison according to the acquisition time of the hyperspectral frame and the acquisition time of the visible light image frame and the near-infrared image frame. Perform same-time window matching for encoder pulse records, remove records with time deviations exceeding the alignment tolerance value, and rearrange the retained frame order according to the acquisition time order to obtain the aligned frame order index table. Based on a multi-source time stamp dataset, a hyperspectral frame is set as the baseline time axis, and each time point in the hyperspectral timestamp sequence is traversed. Read the preset alignment tolerance value This tolerance value is based on the sampling frequency of the hyperspectral camera. Configure the settings; the calculation formula is as follows: If the sampling frequency is 50 Hz, then the sampling interval is 20 milliseconds, and the alignment tolerance value is calculated. It is 10 milliseconds; Calculate each time point in the visible light timestamp sequence. Compared to the current reference time point absolute difference Similarly, calculate each time point in the near-infrared timestamp sequence. Compared to the current reference time point absolute difference Execute conditional judgment, and only retain those that meet the conditions. milliseconds and millisecond image frame index and For combinations that meet the conditions, search for the corresponding combinations in the pulse time mapping table. The closest pulse recording time point is used to read the corresponding cumulative pulse value as the displacement marker for that moment. For a given reference time point... If a corresponding frame with an absolute difference of less than or equal to 10 milliseconds cannot be found in the visible or near-infrared sequence, or if the time difference of the most recent frame found is 15 milliseconds (greater than the tolerance), then the multi-source data corresponding to the reference hyperspectral frame is determined to be incomplete, and a rejection operation is performed to remove the hyperspectral frame and its associated data from the processing queue. For example, when processing the 50th frame of hyperspectral data, the timestamp is 2000 milliseconds, while the most recent timestamp in the visible light sequence is 2018 milliseconds. The difference of 18 milliseconds is greater than the tolerance of 10 milliseconds, so the 50th frame combination is discarded. For the remaining valid matching groups, the combination index is rearranged according to the order of hyperspectral acquisition time, and each row of records containing the hyperspectral frame index, visible light frame index, near-infrared frame index and corresponding pulse count value is constructed to obtain the aligned frame order index table.

[0026] S103: Based on the aligned frame sequence index table, the continuity of the acquisition time difference between adjacent frames is judged, and the position where the difference exceeds the continuity threshold is taken as the segmentation point. The frame sequence is divided into segments, and the start and end times and frame sequence number range of each segment are recorded to generate a coal flow moisture measurement synchronization frame group. Based on the aligned frame order index table, read the consecutively arranged reference hyperspectral acquisition time markers in the table, and calculate the time difference between two adjacent rows of records. Read the preset continuity threshold This threshold is set based on the maximum permissible data loss time under normal conveyor belt operating speed, and is set to 3 times the sampling interval. That is, when the sampling interval is 20 milliseconds, the continuity threshold is... The time difference sequence is calculated by scanning line by line and performing numerical comparisons. When a time difference in a certain line is detected... In milliseconds, it is determined that the current frame and the previous frame belong to the same continuous acquisition segment, and the index of the current segment remains unchanged. When the time difference of a certain row is detected... In milliseconds, for example, if the record time of line 100 is 5000 milliseconds and the record time of line 101 is 5150 milliseconds, the difference of 150 milliseconds exceeds the 60 millisecond threshold, indicating a data interruption. Line 100 is marked as the end frame of the current segment, and line 101 is marked as the start frame of the next new segment. The timestamp corresponding to the first line of each segment is extracted as the segment start time, and the timestamp corresponding to the last line of the segment is extracted as the segment end time. The frame sequence number range contained in the segment is calculated. For example, the range of segment 1 is frame sequence 1 to 100, and the range of segment 2 is frame sequence 101 to 250. The original hyperspectral data blocks, visible light image blocks, near-infrared image blocks, and pulse data corresponding to each segment are packaged to generate a coal flow moisture measurement synchronization frame group.

[0027] Please see Figure 3 The steps to obtain S2 are as follows: S201: Based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, for each coal flow segment, select pixels or grid units composed of multiple pixels in the coal flow area as analysis units, collect the reflectance values ​​of the moisture-sensitive bands corresponding to the analysis units at each collection time, arrange them in the order of collection time, record the difference of reflectance values ​​between adjacent frames, and generate a reflectance time series numerical sequence. Based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, each hyperspectral data cube within the frame group is extracted, and the resolution parameters of the data cube in the spatial dimension are read, expressed in pixel coordinates. Based on this, a grid with a step size of 5 pixels is set within the image area covered by coal flow. The center position of each grid is selected as the anchor point, and a grid with a size of [missing information] is constructed. The rectangular region of a pixel is used as the analysis unit. For each analysis unit, the specific wavelength index corresponding to the moisture absorption characteristics is retrieved from the band index list. 1450 nm and 1940 nm are selected as the moisture-sensitive bands, and the analysis unit is read frame by frame at each time step. The reflectance data is used to calculate The arithmetic mean of the reflectance of the nine pixels within the region under the sensitive band is denoted as the unit reflectance at that moment. This value is a dimensionless physical quantity between 0 and 1, for example, in The mean value of the region at 1450 nm wavelength was 0.12 at any given time. The time-series read value is 0.118. The reflectance values ​​corresponding to the continuously acquired frames are stored in an array according to time sequence. A difference operation is performed between adjacent frame values, calculated using the following formula: For example, calculated It iterates through all frames within the entire sampling time segment, calculates the difference sequence containing positive and negative signs in sequence, and binds the original absolute reflectance sequence with the difference sequence as key-value pairs to generate a reflectance time-series numerical sequence.

[0028] S202: Based on the reflectance time-series numerical sequence, the reflectance difference between adjacent frames is judged frame by frame. The rising state, falling state or stable state is marked according to the change of the sign of the difference. The position where the sign of adjacent states reverses is located. The corresponding reflectance change amplitude and consecutive frame number are extracted and compared with the evaporation stage discrimination threshold to obtain the stage boundary position sequence. Based on the time-series numerical values ​​of reflectance, a noise tolerance threshold is set. The value is 0.001, and the reflectance difference between adjacent frames is scanned one by one. Execute the three-state judgment logic, if Mark the current frame state as "rising". Marked as "decline", if The state is labeled as "stable". A sequence of state labels is generated, and this sequence is scanned to identify nodes where the state label changes, such as the moment when it changes from "decline" to "stable" or "rise", and these are recorded as candidate turning points. For each candidate turning point, backtrack the number of frames the previous state lasted. And calculate the cumulative change in reflectivity during this state. Read the preset evaporation stage discrimination threshold, including the amplitude threshold. With persistence threshold Frame, perform numerical comparison; Only when the candidate point corresponds and When the inflection point is determined to be a valid stage boundary point, for example, if a segment continuously decreases within the first 50 frames and the cumulative decrease is 0.15, and the condition is met, then the position is confirmed as the boundary point between "initial evaporation" and "stable evaporation". If the condition is not met, the inflection point is ignored and the state is regarded as a continuation of the previous stage. Similarly, the boundary point from "stable evaporation" to "tail drying" is identified. All the determined boundary point frame numbers are arranged in chronological order to obtain the stage boundary position sequence.

[0029] S203: Based on the phase boundary position sequence, the reflectance time series is divided into phases according to the boundary position to form the initial evaporation segment, stable evaporation segment and tail drying segment. The reflectance change trajectory of each phase is summarized by sliding window and the phase change contour is generated. The contours of each phase are spliced ​​according to the acquisition time sequence and merged according to the wavelength order to establish an index and generate the moisture time series segmented coding result. Based on the stage boundary position sequence, read the frame sequence number nodes recorded in the sequence. and Using this as a cutting point, the original reflectivity time series The physical segment is divided into three sub-sequences, which are defined as the initial evaporation segment. (corresponding frame) to ), stable evaporation section (corresponding frame) to and the tail drying section (corresponding frame) (Until the end), a sliding window of length 5 is set, and a moving average operation is performed within each subsequence. The mean of the data within the window is calculated as the aggregate feature value at the center time of that window. For example, for a certain window value within the stable evaporation section... The mean was calculated to be 0.0972. The smoothed trajectory curve, i.e. the stage change profile, was generated by moving the window with a step size of 1. The smoothed profile data of the three stages were concatenated end to end according to the time logic order of "initial-stable-drying". The above segmentation and concatenation operations were performed on the 1450 nm and 1940 nm bands respectively to obtain two complete time-series profile vectors. These two vectors were spliced ​​in order of wavelength to construct a one-dimensional long vector containing multi-band time-series features. The corresponding original time index was assigned to each data point to generate the moisture time-series segmented coding result.

[0030] Please see Figure 4 The steps to obtain S3 are as follows: S301: Call the visible light image frame in the coal flow moisture measurement synchronization frame group, perform difference calculation on the gray values ​​of adjacent pixels in the analysis unit, judge the gray value difference and the brightness change threshold, locate the gray value change position, and map the corresponding position to the height level value in the analysis unit to generate a height level mark sequence. The visible light image frames in the coal flow moisture measurement synchronization frame group are retrieved frame by frame. The raw RGB format image data acquired by the industrial-grade CMOS camera is read sequentially. The color images are converted into a single-channel grayscale image matrix using the standard NTSC weighted average method. The weighting formula is set as follows: Traverse each predefined section in the image The pixel analysis unit extracts the grayscale values ​​of 9 pixels within each unit. Using the center pixel as a reference point, the absolute value of the grayscale difference between the center pixel and its 8 neighboring pixels is calculated, resulting in a difference set containing 8 values. A threshold for determining abrupt changes in brightness is then set. The threshold is set to 35 (grayscale range 0-255) based on the shadow depth characteristics of the coal surface under diffuse reflection illumination, and is applied to the maximum value in the difference set. Perform interval mapping judgment, if The value is in the range Inside, the surface of the unit is determined to be flat with no obvious protrusions, and the marking height level is 0. The value is in the range Within the cell, it is determined that there are minor texture undulations, and the height level is marked as 1. If the value is greater than 35, the unit is determined to have significant particle accumulation or deep shadows, and is marked with a height level of 2. For example, if the gray level of the center pixel of an analysis unit is 120 and the gray level of the pixel to its right is 80, the difference is 40, which exceeds the threshold of 35. Therefore, the unit is marked with a height level of 2. The above calculation and judgment operation is performed on all analysis units in the image in sequence, and the obtained level values ​​are arranged according to the row and column scanning order of the units in the image to generate a height level label sequence.

[0031] S302: Based on the near-infrared image frames in the synchronous frame group for coal flow moisture measurement, perform difference judgment on the brightness values ​​of adjacent pixels in the analysis unit, filter continuously changing segments according to the brightness change amplitude and transition judgment threshold, and map the segments to gradient level values ​​in the analysis unit to obtain gradient level labeling sequence. Based on the near-infrared image frames in the coal flow moisture measurement synchronous frame group, the 10-bit depth luminance value matrix (value range 0-1023) acquired by the InGaAs near-infrared camera was read, and for each value corresponding to the spatial position in the visible light image... The analysis unit extracts three columns of pixel data arranged along the conveyor belt's direction of movement within the unit. It then calculates the absolute value of the average difference in brightness between adjacent columns of pixels, using this as a gradient index characterizing the continuity of moisture distribution. Set a threshold for brightness transition detection; Including low threshold With high threshold The calculated gradient index A three-level comparison is performed with the threshold, when When the area is determined to have uniform moisture distribution or be in a completely dry state, the mapping gradient level is 0. When the region is determined to be at the dry-wet transition boundary, the mapping gradient level is 1. When the area is determined to have the edge of a water-accumulated patch or the boundary of a heterogeneous component, the gradient level is mapped to 2. For example, if the average brightness of adjacent columns of a certain unit is calculated to be 600 and 650 respectively, and the difference of 50 is greater than the high threshold of 40, then the gradient level of the unit is marked as 2. By traversing all frames and all analysis units within the frame, the gradient level value corresponding to each unit is recorded to obtain the gradient level marking sequence.

[0032] S303: Based on the height level label sequence and the gradient level label sequence, pair them together in the same analysis unit to form feature pairs. Perform block processing on the image region according to the preset block size. Statistically count the frequency of feature pair combinations in each block region and arrange them into a two-dimensional numerical matrix to generate a joint distribution map block group of water film roughness. Based on the height level label sequence and gradient level label sequence, read the height level values ​​under the same spatial coordinate index. With gradient level values Pair them together to construct feature pairs Set the image block size parameters to divide the entire image into multiple non-overlapping blocks. Pixel sub-regions, each containing approximately 450 analysis units (based on...) (Unit size conversion), establish a unit size conversion within each segmented region. A zero-dimensional matrix is ​​used as a counter. The row indices of the matrix correspond to height levels 0, 1, and 2, and the column indices correspond to gradient levels 0, 1, and 2. Feature pairs within each block are scanned sequentially, and their values ​​are read and incremented at the corresponding positions in the matrix. For example, if a pair is... Then, the value of the element in the 3rd row and 2nd column of the matrix is ​​increased by 1. After the statistics are completed, the frequency distribution of various surface textures and moisture distribution combinations in this block area is obtained. For example, the matrix display elements of a certain block The frequency is 300, element A frequency of 5 indicates that the region is mainly composed of a smooth surface with uniform moisture. The above statistics are performed on all segmented regions and the generated frequency matrix is ​​retained to generate a joint distribution map of water film roughness.

[0033] Please see Figure 5 The steps to obtain S4 are as follows: S401: Based on the two-dimensional distribution matrix in the block group of the joint distribution map of water film roughness, read the matrix element values ​​one by one in the block region order, perform neighborhood aggregation operation on adjacent matrix elements, extract the local numerical arrangement pattern according to the combination relationship of adjacent element values, and record the sequence according to the scanning order to generate a local pattern sequence set. The specific limitations for performing neighborhood aggregation operations on adjacent matrix elements are as follows: Centered on any element in the two-dimensional distribution matrix, a neighborhood window is formed by selecting a fixed number of matrix elements, including that element, and combining the values ​​of the matrix elements within the neighborhood window in a row-major scanning order. The extraction of local numerical arrangement patterns based on the numerical combination relationships of adjacent elements is specifically limited to: The matrix element values ​​within the neighborhood window are arranged according to a preset rule to form a non-repeating numerical sequence, and the numerical sequence is used as the confirmation condition for the local pattern sequence. Based on the two-dimensional distribution matrix in the joint distribution map of water film roughness, the corresponding data for each block are loaded sequentially according to the spatial arrangement order of the block regions. Frequency matrix Set one The local neighborhood window, in terms of matrix elements The starting point of the scan, where The range of values ​​is The window is slid within the matrix, and each slide extracts the values ​​of the four adjacent elements covered by the window. For example, during the first scan, the window is extracted... The frequency value of the location, assuming the extracted value is Arrange these four values ​​into a one-dimensional temporary vector according to row priority. The vector is used to perform numerical comparisons of its elements, generating local binary pattern codes based on these relationships. For example, the first element can be set as the baseline; if subsequent elements are greater than the baseline, they are recorded as 1, otherwise as 0, generating a three-bit binary code. Alternatively, the original numerical sequence can be directly retained as the pattern feature, and the entire sequence can be iterated over. The matrix was executed a total of 4 times. The window sliding operation (corresponding to the four quadrants of upper left, upper right, lower left, and lower right) extracts four sets of numerical sequences. These four sets of sequences are then stored in a list in the order of window scanning. The above operation is repeated for each block region to generate a local pattern sequence set.

[0034] S402: Based on the local pattern sequence set, the local pattern sequences are spliced ​​together in each block region according to the spatial arrangement order. The splicing results are numbered and stored according to the block number to form the numerical description vector of the corresponding block region, thus obtaining the block description subset. The splicing results are stored by numbering the blocks, specifically limited to: Numerical description vector indices for the block regions are generated using a numbering rule consistent with the spatial numbering of the block regions in the joint distribution map of water film roughness. Based on the local pattern sequence set, read the four sets of local pattern numerical sequences generated within the current block region, set a fixed splicing rule, and connect these four sets of sequences end to end according to the spatial quadrant order of "top left-top right-bottom left-bottom right" to construct a sequence of length 16 ( A one-dimensional numerical vector, which fully characterizes the local texture structure of the water film roughness distribution within the segmented region, is used to read the spatial coordinate index of the segment in the original image. Convert the coordinate index into a unique linear ID using the following formula: ,in This represents the total number of horizontal blocks in the image. For example, if the image is divided into 10 rows and 10 columns, totaling 100 blocks, then for the block in the 3rd row and 4th column (index starts from 0, i.e., row 2, column 3), the ID is calculated as follows: The generated one-dimensional numerical vector is bound to the ID and stored as the exclusive descriptor for that block. All block regions are traversed to generate a set containing all block descriptor vectors and their ID indices, thus obtaining the block descriptor subset.

[0035] S403: Based on the block description subset, the block description vectors are concatenated along the coal flow transport direction and the lateral arrangement order to construct a spatial arrangement numerical sequence covering the coal flow area. At the same time, the segmented coding vectors of each analysis unit in the moisture time series segmented coding result are called and sorted according to the acquisition time to generate a whole frame time series feature sequence. Based on the set of tile descriptors, and in ascending order of tile ID, the numerical description vectors of each tile are read sequentially. A vector concatenation operation is then performed to merge the description vectors of all tiles into a single, very long one-dimensional spatial feature vector. For example, if each block description vector has a length of 16, and there are 100 blocks in total, then the merged... With a length of 1600, this vector logically covers the entire coal flow region according to the physical spatial distribution. Simultaneously, it retrieves the moisture temporal segmentation encoding result generated in step S2 from memory, extracting the temporal profile vectors for the 1450 nm and 1940 nm bands. These two profile vectors (assuming each has a length of...) are then... The features are concatenated to construct a one-dimensional temporal feature vector. , length is Read the original acquisition time marker corresponding to the time series feature vector. and And the image acquisition time corresponding to the spatial feature vector. Generate a full-frame temporal feature sequence.

[0036] S404: Based on the acquisition time corresponding to the whole frame temporal feature sequence and the spatially arranged numerical sequence, perform time consistency judgment and complete sequence alignment, perform concatenation and order rearrangement of the aligned sequence, and add sequence length identifier to generate coal moisture fusion feature vector; The appended sequence length identifier is specifically limited to: The effective length values ​​of the entire frame temporal feature sequence and the spatially arranged numerical sequence are recorded in the coal moisture fusion feature vector and stored as part of the vector structure description information; Based on the acquisition times corresponding to the whole frame temporal feature sequence and the spatially arranged numerical sequence, the time corresponding to the spatial feature vector is read. Searching for and matching time series features on the time axis If the absolute time difference between the two closest time points is less than a preset synchronization threshold (e.g., 20 milliseconds), they are determined to be data from the same moment, and alignment confirmation is performed. If the time difference exceeds the threshold, linear interpolation is performed to supplement the data according to the nearest neighbor principle, or the data set is discarded if the difference is too large. After alignment confirmation, a new feature container is created, and the spatial feature vector is written into it. Write all elements into the time series feature vector. Perform a concatenation operation on all elements to form a joint feature vector. The number of elements in the eigenvectors of the statistical space (e.g., 1600) and the number of elements in the time series feature vector (For example, 300), append these two length values ​​as metadata headers. The front-end format is defined as follows: The resulting vector structure contains explicit length information for subsequent parsing, generating a coal moisture fusion feature vector.

[0037] Please see Figure 6 The steps to obtain S5 are as follows: S501: Based on the conveyor belt speed mark and pixel dwell frame number mark in the coal flow moisture measurement synchronization frame group, read the pixel dwell frame number value for each coal flow time segment one by one, compare the dwell frame number with the set dwell frame number threshold, record the index of the time segment that meets the threshold condition, and synchronously associate it with the corresponding conveyor belt speed mark to obtain the valid coal flow time segment index set. Based on the conveyor belt speed marker and pixel dwell frame number marker in the coal flow moisture measurement synchronization frame group, all segmented coal flow time segments are traversed, and the pixel dwell frame number corresponding to each segment is read. This value is derived from the total number of frames experienced by a specific coal flow point within the field of view from entry to exit within the segment, and a threshold for the number of dwell frames is set. This threshold is set based on the shortest dwell time at the conveyor belt's highest operating speed. For example, with a maximum belt speed of 4 m / s, a field of view of 2 meters, and a sampling rate of 50 frames / second, the shortest dwell time is 0.5 seconds, or 25 frames. Therefore, the threshold is set accordingly. For each segment, Compare the values ​​with 25, if The data segment was deemed insufficient to support a complete time series analysis and was therefore discarded. like The segment is determined to be valid observation data, its unique index ID is recorded, and the average conveyor belt speed marker corresponding to the segment is extracted from the synchronization frame group. (Unit: m / s), this speed mark is derived from the measurement value of the photoelectric encoder, and the ID is compared with... The key-value pairs are stored in a list. For example, if the segment ID is 105, the number of frames is 30, and the average belt speed is 3.2 m / s, then the following conditions are met: Add the data to the list, iterate through all the segments, and then summarize all key-value pairs that meet the conditions to obtain the valid coal flow time segment index set.

[0038] S502: Based on the effective coal flow time segment index set, select the time scaling rules according to the conveyor belt speed mark corresponding to each time segment, call the moisture measurement value sequence of the corresponding time segment in the moisture time series segment coding result, scale the time axis of the sequence proportionally, and perform sequential rearrangement and numerical interpolation filling on the scaled time points to obtain the time-scaled moisture time series. Based on the effective coal flow time segment index set, for each index ID in the set, the associated conveyor belt speed marker is read. Calculate the time scaling factor The formula is ,in The reference speed for system calibration is set to 2.0 m / s. For example, if the current... meters per second; but ,like meters per second, then The physical significance of this calculation lies in normalizing the time scale at the current velocity to the time scale at the reference velocity, calling the water time series segmented encoding result generated in step S2, and extracting the original water measurement value sequence and its original time axis corresponding to the segment. ; Apply a scaling factor to perform a transformation calculation on the original time axis. The transformed time coordinates are obtained, and a new standard time grid is constructed based on the standard sampling interval at the reference velocity (e.g., 0.02 seconds per frame). The transformed non-uniform time points Mapping to standard mesh For time points with missing data in the grid, the moisture measurements corresponding to the two nearest transformation time points before and after that point are selected, and a linear interpolation formula is used. Calculate the moisture value at this grid point, for example, at a standard grid point. There is no data at this point, while the most recent transformation time points are 0.058 (corresponding value 0.12) and 0.065 (corresponding value 0.13). The interpolation result is The interpolation operation is performed on the entire sequence to generate a standardized sequence with uniform length and consistent sampling interval, resulting in a time-scaling moisture time series.

[0039] S503: Based on the time-scaling moisture time series, the corresponding feature sub-vectors in the coal moisture fusion feature vector are aligned point by point according to the collection time sequence. The aligned values ​​are then subjected to joint aggregation operation and a single numerical result is output to generate the coal moisture content measurement result. Based on the time-scaling water time series, it is used as the observation vector to be corrected. The temporal feature component is extracted from the coal moisture fusion feature vector output in step S4 and used as a reference vector. Read the length information of both and set the alignment window length to 1. (For example, with 100 data points), starting from the beginning of both sequences, simultaneously extract sub-vectors of length 100, and calculate the Pearson correlation coefficient between the two sub-vectors. ,like (Set a correlation threshold) Determine if the two are well aligned within the current window, and extract... mean within this window As the basic measurement value, extract The spatial texture feature components contained therein (i.e., the tile descriptor aggregation result) are input into a pre-trained multiple linear regression model or neural network model based on existing technologies. Calculate texture correction coefficients For example, function output This indicates that the surface roughness is causing the measured value to be too low and needs to be adjusted upwards by 5%. Perform a joint aggregation operation. The formula is Substitute into numerical calculation ,like Then, according to the principle of maximum relevance, Fine-tune the cutoff window position within the range of data points until the correlation is maximized, and then perform the above calculation. Perform the aggregation operation on the entire time series in segments, and take the arithmetic mean of the values ​​calculated in each segment as the final percentage value of the moisture content of the coal flow segment, thus generating the coal moisture content measurement result.

[0040] A rapid coal moisture content measurement system based on convolutional neural networks, comprising: The multi-source frame synchronization module acquires hyperspectral frames of coal flow and records the time and wavelength order, synchronously acquires visible light and near-infrared images and marks the time, reads encoder pulses and associates them with time, aligns them according to time and removes missing frames, checks the continuity of time and then segments them and records the range, and generates a coal flow moisture measurement synchronization frame group. The spectral time series segmentation module, based on the hyperspectral data in the synchronous frame group of coal flow moisture measurement, selects the analysis unit, extracts the moisture-sensitive band to construct the reflectance time series, identifies candidate turning points, divides the evaporation stage according to the set threshold, and generates the moisture time series segmentation coding result. The texture roughness modeling module calls the visible light and near-infrared images in the synchronous frame group of coal flow moisture measurement, analyzes the brightness of the visible light neighborhood and near-infrared, extracts the height and gradient feature pairs formed by gray-level abrupt changes and brightness transitions, constructs a two-dimensional distribution matrix according to the block statistical frequency, and generates a joint distribution map block group of water film roughness. The spatiotemporal feature fusion module scans the two-dimensional matrix in the joint distribution map of water film roughness, aggregates adjacent elements to extract local patterns, splices them to form map descriptors and concatenates them into a spatial description sequence, combines the moisture temporal segmentation coding results to rearrange and mark the length, and obtains the coal moisture fusion feature vector. The timing correction estimation module filters valid coal flow time segments that meet the threshold based on the conveyor belt speed marker and pixel dwell frame number marker in the coal flow moisture measurement synchronization frame group. It selects time scaling rules according to the conveyor belt speed and performs time scaling, rearrangement and interpolation filling on the moisture measurement value sequence in the moisture time segment based on the moisture time-series segmented coding result. It aligns with the corresponding sub-vectors of the fused feature vector and aggregates them together to generate the coal moisture content measurement result.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for rapid measurement of moisture content of coal based on convolutional neural network, characterized in that, Comprise the following steps: S1: Obtain the coal flow hyperspectral frame and record the time and wavelength order, synchronize the collection of visible and near-infrared images and mark the time, read the encoder pulse and associate the time, align the time and remove the missing frames, check the time continuity, then cut the segments and record the range, generate the coal flow moisture measurement synchronous frame group; S2: Based on the hyperspectral data in the coal flow moisture measurement synchronous frame group, select the analysis unit, extract the moisture sensitive band to construct the reflectivity time sequence, identify the turning candidate points, divide the evaporation stage according to the set threshold, and generate the moisture time sequence segmentation coding result; S3: Call the visible and near-infrared images in the coal flow moisture measurement synchronous frame group, analyze the visible neighborhood and near-infrared brightness, extract the gray scale mutation and brightness transition to form the height and gradient feature pair, construct the two-dimensional distribution matrix according to the block frequency statistics, and generate the water film roughness joint distribution tile group; S4: Scan the two-dimensional matrix in the water film roughness joint distribution tile group, aggregate adjacent elements to extract local patterns, splice to form tile descriptors and concatenate into a spatial description sequence, rearrange and identify the length combined with the moisture time sequence segmentation coding result, and obtain the coal moisture fusion feature vector. 2.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The coal flow moisture measurement synchronous frame group includes a unified time reference of multi-source data, an effective coal flow segment identifier, a frame sequence continuity marker, and a spatial correspondence index. The moisture time sequence segmentation coding result includes evaporation stage type identification, stage boundary position coding, stage duration parameters, and band association index. The water film roughness joint distribution tile group includes block texture distribution matrix, roughness level combination feature, local structure statistical result, and tile space number. The coal moisture fusion feature vector includes time sequence moisture feature component, spatial texture feature component, time-space correlation identifier, and vector structure description information. 3.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The acquisition step of S1 is: S101: Obtain the hyperspectral reflectance frame column of the coal flow area above the conveyor belt, record the acquisition time marker for each frame and record the corresponding wavelength arrangement order, simultaneously collect the visible image stream and near-infrared image stream in the same coal flow area and mark the acquisition time of each image, read the encoder pulse stream corresponding to the conveyor belt motion state and associate the time marker, and generate a multi-source time marker dataset; S102: Based on the multi-source time marker dataset, perform frame-by-frame comparison according to the hyperspectral frame acquisition time and the visible image frame and near-infrared image frame acquisition time, match the same time window for the encoder pulse record, remove records with time deviation exceeding the alignment tolerance value, and rearrange the frame sequence according to the acquisition time order to obtain an alignment frame sequence index table; S103: According to the alignment frame sequence index table, judge the continuity of the acquisition time difference of adjacent frames, take the position with a difference value exceeding the continuity threshold as a cutting point, perform segment division on the frame sequence, record the corresponding start and end time and frame number range of each segment, and generate a coal flow moisture measurement synchronous frame group. 4.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The acquisition step of S2 is: S201: Based on the hyperspectral data in the coal flow moisture measurement synchronous frame group, for each coal flow segment, a pixel point or a grid cell composed of multiple pixels in the coal flow area is selected as an analysis unit, the reflectance value of the analysis unit in the corresponding moisture sensitive wave band at each collection time is collected, the reflectance values of adjacent frames are arranged in sequence according to the collection time, and the difference value of the reflectance values of adjacent frames is recorded to generate a reflectance time sequence value sequence; S202: According to the reflectance time sequence value sequence, the reflectance difference value of adjacent frames is judged frame by frame, the sign change is marked according to the difference value, the rising state, the falling state or the stable state is marked, the position of the sign inversion of adjacent states is located, the corresponding reflectance change amplitude and the number of continuous frames are extracted, and the phase discrimination threshold is compared to obtain a phase boundary position sequence; S203: Based on the phase boundary position sequence, the reflectance time sequence is divided into phases according to the boundary position to form an initial evaporation section, a stable evaporation section and a tail drying section, the reflectance change trajectory of each phase is summarized according to a sliding window, and a phase change profile is generated. According to the collection time sequence, each phase profile is spliced and combined according to the wavelength arrangement order to establish an index, and a moisture time sequence segmentation coding result is generated. 5.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The acquisition step of S3 is: S301: Call the visible light image frame in the coal flow moisture measurement synchronous frame group, perform difference operation on the gray values of adjacent pixels in the analysis unit, judge according to the gray value difference and the light and dark mutation judgment threshold, locate the gray value change position, and map the corresponding position to the height level value in the analysis unit to generate a height level mark sequence; S302: Based on the near-infrared image frame in the coal flow moisture measurement synchronous frame group, difference judgment is performed on the brightness values of adjacent pixels in the analysis unit, continuous change sections are selected according to the brightness change amplitude and the transition judgment threshold, and the sections are mapped to the gradient level value in the analysis unit to obtain a gradient level mark sequence; S303: According to the height level mark sequence and the gradient level mark sequence, a feature pair is formed by pairing and combining in the same analysis unit, the image area is divided into blocks according to the preset block size, the frequency of feature pair combination in each block area is counted and arranged into a two-dimensional numerical matrix, and a water film roughness joint distribution block group is generated. 6.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The acquisition step of S4 is: S401: Based on the two-dimensional distribution matrix in the water film roughness joint distribution block group, read the matrix element values in sequence according to the block area order, perform neighborhood aggregation operation on adjacent matrix elements, extract the local numerical arrangement mode according to the combination relationship of adjacent element values, and record the sequence in scanning order to generate a local mode sequence set; S402: According to the local mode sequence set, the local mode sequence is spliced in each block area according to the spatial arrangement order, the spliced results are stored according to the block number to form a numerical description vector of the corresponding block area, and a block descriptor set is obtained; S403: Based on the tile descriptor set, concatenate each sub-block descriptor vector along the coal flow conveying direction and transverse arrangement order to construct a spatial arrangement numerical sequence covering the coal flow area, and simultaneously call each analysis unit segmented coding vector in the moisture time sequence segmented coding result and sort it according to the acquisition time to generate an entire frame time sequence feature sequence; S404: According to the corresponding acquisition time of the entire frame time sequence feature sequence and the spatial arrangement numerical sequence, perform time consistency judgment and complete sequence alignment, concatenate and reorder the aligned sequence, and add sequence length identifier to generate coal moisture fusion feature vector.

7. The method of claim 6, wherein the method is a convolutional neural network based rapid coal moisture content measurement method. The neighborhood aggregation operation performed on adjacent matrix elements is specifically defined as: Taking any matrix element in the two-dimensional distribution matrix as the center, a neighborhood window composed of a fixed number of matrix elements including the matrix element is selected, and the matrix element values in the neighborhood window are combined in a row-first scanning order; The local numerical arrangement pattern is extracted according to the combination relationship of adjacent element numerical values, which is specifically defined as: The matrix element values in the neighborhood window are combined according to a preset arrangement rule to form a non-repeating numerical sequence, and the numerical sequence is used as a confirmation condition for the local pattern sequence; The storage of the splicing result according to the block number is specifically defined as: The numerical descriptor vector index of the block region is generated using the same numbering rule as the block region space number in the water film roughness joint distribution tile group; The additional sequence length identifier is specifically defined as: The effective length values of the entire frame time sequence feature sequence and the spatial arrangement numerical sequence are recorded in the coal moisture fusion feature vector, and are stored as part of the vector structure description information. 8.The coal moisture content rapid measurement method based on a convolutional neural network according to claim 1, characterized in that: The method further comprises: S5: According to the belt speed marker and the pixel residence frame number marker in the coal flow moisture measurement synchronous frame group, filter the effective coal flow time segments that meet the threshold, select the time stretching rule according to the belt speed, perform time stretching, rearrangement and interpolation filling on the moisture measurement value sequence in the moisture time sequence segmented coding result based on the effective coal flow time segments, align and aggregate the corresponding sub-vector of the fusion feature vector, and generate the coal moisture content measurement result; The coal moisture content measurement result is specifically the moisture content value of the coal flow, the time continuous output identifier, the effective measurement segment corresponding relationship and the online measurement result marker.

9. The method of claim 8, wherein the method further comprises: The acquisition step of S5 is: S501: According to the belt speed marker and the pixel residence frame number marker in the coal flow moisture measurement synchronous frame group, read the pixel residence frame number value for each coal flow time segment, compare the residence frame number with the set residence frame number threshold, record the index of the time segment that meets the threshold condition, and synchronously associate the corresponding belt speed marker to obtain the index set of the effective coal flow time segment; S502: Based on the index set of the effective coal flow time segment, select the time stretching rule according to the belt speed marker corresponding to each time segment, call the moisture measurement value sequence in the corresponding time segment of the moisture time sequence segmented coding result, perform proportional stretching on the time axis of the sequence, and perform sequential rearrangement and numerical interpolation filling on the stretched time points to obtain the time stretched moisture time sequence; S503: According to the time stretching water time sequence, the corresponding feature sub-vector in the coal water fusion feature vector is aligned point by point in the order of collection time, the aligned numerical value is executed joint aggregation operation and output single numerical value result, and coal water content measurement result is generated.

10. A coal moisture content rapid measurement system based on a convolutional neural network, characterized in that, The system is used for the coal water content rapid measurement method based on convolutional neural network in any one of claims 1-9, and the system comprises: A multi-source frame synchronization module acquires coal flow hyperspectral frames and records time and wavelength order, synchronously collects visible light and near-infrared images and marks time, reads encoder pulses and associates time, aligns by time and eliminates missing frames, checks time continuity, cuts segments and records ranges, and generates coal flow water measurement synchronization frame group. A spectral time sequence segmentation module selects an analysis unit based on the hyperspectral data in the coal flow water measurement synchronization frame group, extracts a water-sensitive waveband to construct reflectivity time sequence, identifies turning candidate points, divides evaporation stages according to a set threshold, and generates water time sequence segmentation coding result. A texture roughness modeling module calls the visible light and near-infrared images in the coal flow water measurement synchronization frame group, analyzes visible light neighborhood and near-infrared brightness, extracts gray mutation and brightness transition to form height and gradient feature pairs, constructs two-dimensional distribution matrix according to block statistics frequency, and generates water film roughness joint distribution tile group. A space-time feature fusion module scans the two-dimensional matrix in the water film roughness joint distribution tile group, aggregates adjacent elements to extract local patterns, concatenates tile descriptors to form spatial description sequence, rearranges and identifies length combined with water time sequence segmentation coding result, and obtains coal water fusion feature vector. A time sequence correction estimation module selects effective coal flow time segments that meet the threshold according to the conveyor belt speed marker and pixel residence frame number marker in the coal flow water measurement synchronization frame group, selects time stretching rules according to the conveyor belt speed, executes time stretching, rearrangement and interpolation filling on the water measurement value sequence in the water time sequence segmentation coding result in the effective coal flow time segments, aligns and jointly aggregates the corresponding sub-vector of the fusion feature vector, and generates coal water content measurement result.

Images