A rapid non-destructive identification method, determination method, system, equipment and medium for poplar species and pulp material

Abstract

The present invention relates to a method, determination method, system, equipment and medium for rapid and non-destructive identification of poplar varieties and pulpwood. In terms of the method, log segments are screened by sensors, and near-infrared spectral data is collected after being processed into samples, and chemical composition tests are performed. Then, adaptive filtering and inverse wavelet transform are used to purify the spectrum, and features are screened according to wavelength selection strategies. Models are constructed and trained, and classification results output by serial or parallel combinations of multiple models are integrated. In terms of the system, sample processing, testing, data purification, model training and classification function modules are integrated to achieve full-process automation and intelligent management, and the above method is collaboratively executed on the device through a database and a memory to ensure efficient operation of the method; executable instructions are stored on the medium to facilitate wide deployment and application. The present invention can automatically and accurately output the classification results of poplar wood and pulpwood, which not only significantly improves the identification efficiency, but also reduces human errors.

Description

Technical Field

[0001] The invention relates to the technical field of wood identification, and in particular to a method, a determination method, a system, a device and a medium for rapid non-destructive identification of poplar varieties and pulpwood. Background Art

[0002] As a widely planted artificial forest tree species in my country, poplar has become an important raw material for wood products such as pulp fiber, artificial fiber board and plywood due to its fast growth, strong adaptability and excellent stress resistance. The quality of different poplar varieties varies, resulting in large differences in their wood uses and economic value. For example, some poplar varieties have a low lignin content and are suitable for pulp production, while other varieties may be unsuitable as pulp raw materials due to their high lignin content. Therefore, the accurate identification of poplar varieties, especially judging whether they are suitable for use as pulp raw materials based on key material properties such as lignin content, is of great practical significance.

[0003] Traditional methods for identifying poplar species usually rely on manual observation and empirical judgment, or determine the suitability of wood by chemical experiments to determine the lignin content in the wood. Although these methods are effective, they usually require high experimental costs and time, and are often difficult to fully adapt to large-scale forestry resource management needs. Especially in large-scale poplar species selection and resource allocation, the inefficiency of traditional methods prevents the best use of wood from being maximized.

[0004] In recent years, near-infrared spectroscopy, as a non-destructive and rapid phenotypic analysis tool, has shown great potential in the field of wood identification. Through near-infrared spectroscopy, its chemical composition, especially the lignin content, can be quickly analyzed without destroying the wood. This provides us with a new method for the screening of poplar varieties and the precise matching of wood uses. Since the construction and improvement of spectral databases is a long and complex process, the current wood identification technology based on spectral data still faces the problems of insufficient number of databases and uneven quality. In addition, the accuracy and reliability of identification schemes based on spectral data still need to be further improved.

[0005] Therefore, how to use near-infrared spectroscopy technology to quickly identify poplar varieties and accurately determine the lignin content and other key wood property indicators in order to achieve the appropriate use of poplar wood has become an urgent problem to be solved. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method, a determination method, a system, an equipment and a medium for rapid and non-destructive identification of poplar varieties and pulp materials, which solves the technical problems of insufficient database, uneven data quality and the need to further improve the accuracy and reliability of identification when the existing scheme uses near-infrared spectroscopy technology to rapidly identify poplar varieties.

[0008] (II) Technical solution

[0009] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, an embodiment of the present invention provides a method for rapid and non-destructive identification of poplar varieties and pulp materials, comprising: using a sensor to detect and screen poplar log segments that meet specific conditions, controlling processing equipment to process the poplar log segments into wood blocks or wood powder samples, and using a near-infrared spectrometer to capture near-infrared spectral data of the wood blocks or wood powder samples; performing point-by-point or segment-by-segment adaptive filtering on the obtained near-infrared spectral data, and combining with wavelet inverse transform to obtain purified spectral data; according to a spectral wavelength selection strategy constructed based on a preset target, performing spectral wavelength screening on the purified spectral data and optimizing the wavelength combination, and inputting the optimized spectral data and corresponding label data determined by the chemical composition content of the wood powder sample obtained in advance through detection into a corresponding data file to divide it into a training set and a test set; constructing at least one deep learning model as a recognition model, and performing model training based on the training set and the test set, and introducing a serial or parallel combination mechanism to fuse the output results of different models, and automatically outputting the poplar variety and pulp material classification results.

[0011] Optionally, using sensors to detect and screen poplar logs that meet specific conditions, processing the poplar logs into wood blocks or wood powder samples through an integrated processing line, and using a near-infrared spectrometer to capture near-infrared spectral data of the wood blocks or wood powder samples include: using a sensor array to detect and identify defects in poplar logs; screening and marking poplar logs at least one location on the poplar logs according to preset specific conditions; sending instructions to processing equipment to control the processing equipment to process the selected poplar logs into blank samples of predetermined sizes according to a preset processing flow, and perform air drying; controlling the processing equipment to The blank sample after air drying is processed into several wood blocks of the same specification, and the several wood blocks are automatically divided into a first part of samples and a second part of samples; the crusher is controlled to crush the second part of samples into wood powder, and the filtering equipment is controlled to filter the wood powder into a sieve with a specific mesh number to obtain a wood powder sample; the near-infrared spectrometer is used to capture the near-infrared spectral data of the wood blocks and the wood powder samples of the first part of samples respectively, and in the process of capturing the near-infrared spectral data, the camera deployed at the optical fiber probe of the near-infrared spectrometer is used to identify defects, blue discoloration, heartwood or sapwood again to avoid collecting erroneous data.

[0012] Optionally, performing point-by-point or segment-by-segment adaptive filtering processing on the obtained near-infrared spectral data, and combining with inverse wavelet transform to obtain purified spectral data includes: decomposing the obtained near-infrared spectral data through wavelet transform according to the distribution characteristics of the spectral data to obtain multiple layers of detail coefficients and an approximate coefficient; for each layer of detail coefficients, determining a reference threshold based on the noise characteristics of the current layer, and dynamically adjusting the reference threshold by multiplying by a multiple of the standard deviation of the current layer to construct an adaptive filter; using the constructed adaptive filter, performing point-by-point or segment-by-segment filtering processing on each layer of detail coefficients after decomposition to remove noise and outliers; reconstructing the filtered detail coefficients and approximate coefficients through inverse wavelet transform to obtain purified spectral data;

[0013] in,

[0014] ;

[0015] In the formula, X is the spectral signal, D l is the multi-layer detail coefficient, A L is an approximate coefficient.

[0016] Optionally, according to the spectral wavelength selection strategy constructed based on preset goals, the spectral wavelength screening and wavelength combination optimization of the purified spectral data include: quantifying the preset demand goals including maximizing classification accuracy and minimizing prediction error into an objective function for evaluating different wavelength combinations; constructing a spectral wavelength selection strategy that can dynamically adjust the wavelength combination according to the objective function value; using the purified spectral data as the initial spectral data set after processing, and searching for the optimal combination in the wavelength space of the initial spectral data set using the pre-constructed algorithm model; in each iteration, generating a new candidate wavelength set based on the current spectral data set, and using the objective function to select the optimal combination. The new candidate wavelength set is evaluated by the number; if the value of the objective function of the candidate wavelength set in the current iteration is greater than the value of the objective function of the candidate wavelength set in the previous iteration plus the preset improvement threshold τ, the current new candidate wavelength set is accepted as the spectral data after the wavelength selection is completed with a decreasing acceptance probability; if the value of the objective function in the current iteration is not greater than the value of the objective function in the previous iteration plus the preset improvement threshold τ, the current new candidate wavelength set is rejected, wavelengths are added and / or deleted based on prior knowledge or the contribution of each wavelength to the objective function, and the next round of iterative processing is performed; after the iteration is completed, the current optimal wavelength combination and the corresponding objective function value are output;

[0017] Among them, the acceptance probability is:

[0018] ;

[0019] In the formula, exp is the natural exponential function, The value of the objective function for the set of candidate wavelengths in the current iteration, The value of the objective function for the set of candidate wavelengths in the previous iteration, k B The rate of deceleration of progress, J is a progress parameter, configured to be multiplied by the progress factor Gradually from the initial value J 0 gradually decreases to the end value J f .

[0020] Optionally, the optimized spectral data and the corresponding label data determined by the chemical component content of the wood powder sample obtained in advance through testing are input into the corresponding data files to divide the training set and the test set, including: determining the corresponding grade label according to the set standard value range based on the obtained standard value of the chemical component content; pairing the spectral data with the corresponding chemical component content grade label to form a data table to be sorted containing multiple composite data units; each composite data unit contains spectral data, label data and an initial sequence number; treating the data table to be sorted as the root node of the tree, starting from the root node, continuously splitting the data table into child nodes, repeating the splitting for each newly generated child node until each child node contains only one composite data unit; internally sorting all child nodes, merging two sorted child nodes into a new ordered node, and during the merging process, if If the elements in two child nodes have the same label, a secondary sort is performed according to the initial serial number until all child nodes are merged into an ordered data table; in the ordered data table, a unique serial number is generated for each paired spectral data and label, and the unique serial number of each paired data is spliced ​​with the timestamp to form a unique identifier through hashing; using the file operation library or API, a new empty data file is created and the header information of the file is initialized; the paired spectral data and level labels are traversed, and the spectral data are converted into a format suitable for writing according to a predetermined format, and the level labels are converted into corresponding data representations, and the unique identifier is used as a key field to form a complete data record together with the converted spectral data and level labels, and is written into the data file according to the unique identifier; random sampling or stratified sampling is used for the data stored in the data file to divide it into a training set and a test set.

[0021] Optionally, at least one deep learning model is constructed as a recognition model, and model training is performed based on a training set and a test set, and a serial or parallel combination mechanism is introduced to fuse the output results of different models, and the classification results of poplar species and pulpwood are automatically output, including: according to the needs of classification and recognition, a deep learning model including a NIR-Net model, a NIR-ResNet model and a NIR-Inception model is selected and constructed; the structure of the selected model is adjusted and optimized, including adjusting the number of convolution layers, the size of the convolution kernel and the setting of the pooling layer, so as to adapt to the specific task of poplar wood and pulpwood classification; a serial or parallel combination mechanism is introduced by analyzing the characteristics of the selected model; in the serial combination mechanism, the output of the NIR-Net model is used as the input of the NIR-ResNet model through a feature conversion layer in a predetermined order, and then the NIR-ResNet model is used as the input of the NIR-ResNet model. The output of the Net model is used as the input of the NIR-Inception model through a dimensionality reduction layer, and the end-to-end joint training of the entire serial combination model is completed based on the training set and the test set; in the parallel combination mechanism, the NIR-Net model, the NIR-ResNet model and the NIR-Inception model are first trained independently, and then the same spectral data are simultaneously input into the NIR-Net model, the NIR-ResNet model and the NIR-Inception model, and a unary learner is constructed and trained to minimize the prediction error after fusion. The trained meta-learner is used to receive and fuse the outputs of the NIR-Net model, the NIR-ResNet model and the NIR-Inception model; based on the combined prediction results obtained from the serial or parallel combination mechanism, the classification results of poplar and pulpwood are automatically output.

[0022] In the second aspect, an embodiment of the present invention provides a method for determining the chemical composition content of a wood powder sample, which is applied to the rapid non-destructive identification method of poplar varieties and pulp materials as described above, comprising: weighing a wood powder sample with a mass of 0.1~0.5g, and adding sulfuric acid with a volume between 2~4mL and a concentration of 68%~76%, reacting at a temperature range of 27±3°C, and setting the reaction time to 50~70min; after the reaction is completed, using ultrapure water with a volume between 74~94mL, the reaction solution is transferred to a first blue-capped transparent glass bottle with a capacity of 200~300mL; by placing a solution with a concentration of 3%~5% The first and second blue-capped transparent glass bottles were filled with 80-120 mL of sulfuric acid and 40-60 mg of the corresponding monosaccharides were added to prepare calibration solutions of glucose, xylose and arabinose, respectively, and the concentration of the calibration solutions was 0.5 to 1.5 g / L; the samples in the first and second blue-capped transparent glass bottles were placed in an autoclave and reacted for 50-70 min at a temperature range of 118°C to 124°C; after the reaction was completed, the reaction product was filtered through a filtration device, and the filtrate was washed with distilled water preheated to the set temperature until the filtration was complete, and then the mixture was sterilized at a temperature range of 100°C to 110°C. ℃ oven for 3~5h, and calculate the content of acid-insoluble lignin by weighing; according to the filtrate obtained by filtration, use ultraviolet-visible spectrophotometry, use 3%~5% sulfuric acid as a reference, and measure the absorbance between 200~210nm. If the absorbance value exceeds the range of 0.5~1.5, dilute it with sulfuric acid of the corresponding concentration and re-measure it, and select a molar absorption coefficient between 108~118L·g-1·cm-1 to calculate the content of acid-soluble lignin; take a filtrate with a volume of 15~25mL and place it in a beaker with a capacity of 40~60mL, use CaCO3 to neutralize the pH value to between 4 and 6 by magnetic stirring, and when the pH value is close to 4, adjust the pH value to between 5 and 6. After sufficient stirring, use a syringe filter to load the obtained sample; use HPLC technology to determine the contents of glucose, xylose and arabinose in the filtrate, and calculate the contents of cellulose and hemicellulose based on the determination results;

[0023] in,

[0024] , C is the content;

[0025] ;

[0026] Lignin content = acid soluble lignin content + acid insoluble lignin content.

[0027] In a third aspect, an embodiment of the present invention provides a rapid non-destructive identification system for poplar varieties and pulpwood, including: a sample processing and spectrum acquisition module, which is used to use a sensor to detect and screen poplar log segments that meet specific conditions, control processing equipment to process the poplar log segments into wood blocks or wood powder samples, and use a near-infrared spectrometer to capture near-infrared spectrum data of the wood blocks or wood powder samples; a material property testing module, which is used to perform standard value detection of at least one chemical component content on the wood powder sample; a data purification module, which is used to perform point-by-point or segment-by-segment adaptive filtering processing on the obtained near-infrared spectrum data, and obtain purified spectrum data in combination with inverse wavelet transform; The data set output module is used to screen the spectral wavelengths of the purified spectral data and optimize the wavelength combination according to the spectral wavelength selection strategy constructed based on the preset target, and input the optimized spectral data and the corresponding label data determined by the chemical composition content of the wood powder sample obtained in advance through detection into the corresponding data file to divide it into a training set and a test set; the wood classification module is used to construct at least one deep learning model as a recognition model, and perform model training based on the training set and the test set, and introduce a serial or parallel combination mechanism to fuse the output results of different models, and automatically output the classification results of poplar species and pulpwood.

[0028] In a fourth aspect, an embodiment of the present invention provides a device for rapid and non-destructive identification of poplar varieties and pulp materials, comprising: at least one database; and a memory communicatively connected to the at least one database; wherein the memory stores instructions executable by the at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the method for rapid and non-destructive identification of poplar varieties and pulp materials as described above.

[0029] In a fifth aspect, an embodiment of the present invention provides a computer-readable medium having computer-executable instructions stored thereon, which, when executed by a processor, implement the method for rapid and non-destructive identification of poplar species and pulpwood as described above.

[0030] (III) Beneficial effects

[0031] The beneficial effect of the present invention is that by implementing the scheme of the present invention, not only the problem of imperfect spectral database construction in the existing poplar wood property identification scheme is effectively solved, but also the accuracy and reliability of subsequent identification schemes based on spectral data are significantly improved.

[0032] Specifically, the advanced sensor technology is first used to accurately detect and screen poplar logs that meet specific conditions, laying a solid foundation for subsequent processing. Then, the poplar logs are finely processed into wood blocks or wood powder samples by controlling the processing equipment, and the near-infrared spectrometer is used to efficiently capture the near-infrared spectral data of these samples in a non-destructive manner, which not only ensures the accuracy of the data, but also achieves zero damage to the sample material, thereby greatly enriching the content of the spectral database.

[0033] Furthermore, the captured near-infrared spectral data is subjected to point-by-point or segment-by-segment adaptive filtering, and combined with the wavelet inverse transform technology, the spectral data is effectively purified and noise interference is reduced.

[0034] Subsequently, according to the preset spectral wavelength selection strategy, the purified spectral data is scientifically screened for spectral wavelengths to ensure that the selected spectral wavelengths can best reflect the characteristics of poplar wood properties. The screened spectral data and the corresponding label data are sequentially input into the data file, and the training set and test set are reasonably divided, providing high-quality data support for the construction and training of the deep learning model.

[0035] Finally, at least one deep learning model is constructed as the recognition model, and rigorous model training is performed based on the training set and the test set. By introducing a serial or parallel combination mechanism, the output results of different models are effectively integrated, and the complementary advantages of multiple models and multi-dimensional and multi-level information integration are achieved, thereby completing the rapid, accurate and non-destructive classification and identification of poplar varieties and pulpwood. This series of innovative measures not only significantly improved the accuracy and reliability of poplar variety and pulpwood classification, but also provided strong technical support for the efficient classification and rational use of poplar wood, and promoted scientific and technological progress in related industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of a method for implementing the present invention;

[0037] Figure 2 A specific flow chart of step S1 of the method provided for the implementation of the present invention;

[0038] Figure 3 A specific flow chart of step S2 of the method provided for the implementation of the present invention;

[0039] Figure 4 A specific flow chart of a part of the steps of step S3 of the method provided for the implementation of the present invention;

[0040] Figure 5 A schematic diagram of a specific flow chart of another part of the steps of step S3 of the method provided for the implementation of the present invention;

[0041] Figure 6A specific flow chart of step S4 of the method provided for the implementation of the present invention;

[0042] Figure 7 The method provided for implementing the present invention Figure 2 The near infrared spectra of I107 Yang, Zhongxiong No. 1, Bofeng No. 3, Zhonghe No. 1, Lulin No. 3, and Huanghuai No. 4 at 450-2400nm;

[0043] Figure 8 The method provided for the implementation of the present invention is the 450-2400nm near-infrared spectra of Lulin No. 1, Lulin No. 2, Lulin No. 3, Lulin No. 9, and Lulin No. 16. DETAILED DESCRIPTION

[0044] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0045] like Figure 1 As shown, a method for rapid and non-destructive identification of poplar varieties and pulpwood proposed in an embodiment of the present invention includes: using a sensor to detect and screen poplar log segments that meet specific conditions, controlling processing equipment to process the poplar log segments into wood blocks or wood powder samples, and using a near-infrared spectrometer to capture near-infrared spectral data of the wood blocks or wood powder samples; performing point-by-point or segment-by-segment adaptive filtering processing on the obtained near-infrared spectral data, and combining with wavelet inverse transform to obtain purified spectral data; according to a spectral wavelength selection strategy constructed based on a preset target, the purified spectral data is subjected to spectral wavelength screening and wavelength combination optimization, and the optimized spectral data and corresponding label data determined by the chemical composition content of the wood powder sample obtained by detection in advance are input into corresponding data files to divide them into training sets and test sets; constructing at least one deep learning model as a recognition model, and performing model training based on the training set and the test set, and introducing a serial or parallel combination mechanism to fuse the output results of different models, and automatically outputting the classification results of poplar varieties and pulpwood.

[0046] By implementing the scheme of the present invention, not only the problem of imperfect spectral database construction in the existing poplar wood property identification scheme is effectively solved, but also the accuracy and reliability of subsequent identification schemes based on spectral data are significantly improved.

[0047] Specifically, the advanced sensor technology is first used to accurately detect and screen poplar logs that meet specific conditions, laying a solid foundation for subsequent processing. Then, the poplar logs are finely processed into wood blocks or wood powder samples by controlling the processing equipment, and the near-infrared spectrometer is used to efficiently capture the near-infrared spectral data of these samples in a non-destructive manner, which not only ensures the accuracy of the data, but also achieves zero damage to the sample material, thereby greatly enriching the content of the spectral database.

[0048] Furthermore, the captured near-infrared spectral data is subjected to point-by-point or segment-by-segment adaptive filtering, and combined with the wavelet inverse transform technology, the spectral data is effectively purified and noise interference is reduced.

[0049] Subsequently, according to the preset spectral wavelength selection strategy, the purified spectral data is scientifically screened for spectral wavelengths to ensure that the selected spectral wavelengths can best reflect the characteristics of poplar wood properties. The screened spectral data and the corresponding label data are sequentially input into the data file, and the training set and test set are reasonably divided, providing high-quality data support for the construction and training of the deep learning model.

[0050] Finally, at least one deep learning model is constructed as the recognition model, and rigorous model training is performed based on the training set and the test set. By introducing a serial or parallel combination mechanism, the output results of different models are effectively integrated, and the complementary advantages of multiple models and multi-dimensional and multi-level information integration are achieved, thereby completing the rapid, accurate and non-destructive classification and identification of poplar varieties and pulpwood. This series of innovative measures not only significantly improved the accuracy and reliability of poplar variety and pulpwood classification, but also provided strong technical support for the efficient classification and rational use of poplar wood, and promoted scientific and technological progress in related industries.

[0051] In order to better understand the above technical solution, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer and more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0052] Specifically, an embodiment of the present invention provides a method for rapid and non-destructive identification of poplar species and pulp materials, which comprises:

[0053] S1. Use sensors to detect and screen poplar logs that meet specific conditions, control processing equipment to process the poplar logs into wood blocks or wood powder samples, and use a near-infrared spectrometer to capture near-infrared spectral data of the wood blocks or wood powder samples.

[0054] Furthermore, if Figure 2 As shown, step S1 includes:

[0055] S11. Use sensor arrays to detect and identify defects in poplar wood.

[0056] In one embodiment, a sensor array is provided, which is composed of high-resolution optical sensors, infrared sensors, and ultrasonic sensors in a specific spatial layout, and each sensor in the array is connected to a central processing unit via a flexible connection line. Among them, the high-resolution optical sensor is used to capture subtle texture changes on the surface of the wood and to perform preliminary identification of surface defects such as blue discoloration. The infrared sensor is used to penetrate the surface of the wood and detect possible defects inside, such as decay, cracks, etc. The ultrasonic sensor is used to analyze the density and structural changes inside the wood by emitting and receiving ultrasonic signals, and further confirm the type and location of the defect.

[0057] Therefore, the sensor array can identify defects in the wood, such as cracks, knots, etc., and can also identify whether the wood has blue discoloration, that is, the color change of the wood due to moisture or fungal infection. At the same time, it can also distinguish the heartwood and sapwood of the wood. Based on the above identification process, the coordinates of the predetermined sampling area are obtained, and the required log sections can be screened and marked based on the coordinates of the predetermined sampling area.

[0058] S12. Screening and marking at least one poplar log segment on the poplar log according to preset specific conditions.

[0059] S13, sending instructions to the processing equipment to control the processing equipment to process the selected poplar log segments into blank samples of predetermined sizes according to a preset processing flow, and perform air drying.

[0060] Select different varieties of poplars that are growing normally and have no obvious defects, control the processing equipment to automatically cut the selected poplar logs into log test materials of a certain length, record the information and mark each sample, then transport it back to the designated processing location to be processed into blank samples, and then the processed blank samples enter the air-drying stage.

[0061] S14, controlling the processing equipment to process the air-dried blank sample into a number of wood blocks of the same specification, and automatically dividing the number of wood blocks into a first part of samples and a second part of samples.

[0062] S15, controlling the pulverizer to pulverize the second part of the sample into wood powder, and controlling the filtering device to filter the wood powder into a sieve with a specific mesh number to obtain a wood powder sample, wherein the specific number is 40 to 60 meshes.

[0063] S16. Use a near-infrared spectrometer to capture near-infrared spectral data of the wood blocks and wood powder samples of the first part of the sample. During the process of capturing the near-infrared spectral data, use a camera deployed at the optical fiber probe of the near-infrared spectrometer to identify defects, blue discoloration, heartwood or sapwood again to avoid collecting erroneous data.

[0064] After starting the NIR spectrometer and performing necessary warm-up and calibration, the camera deployed at the fiber optic probe is activated to capture images of the wood surface.

[0065] The camera continuously shoots the wood surface in real time, capturing detailed image information including texture, color, defects (such as cracks, knots, blue discoloration) and material differentiation (heartwood / sapwood). Once it is confirmed that the collection point is defect-free, blue-discolor-free and meets specific material requirements (such as giving priority to heartwood or sapwood), the data collection function of the near-infrared spectrometer is immediately started to accurately align the fiber optic probe with the sampling area to capture the near-infrared spectrum data of the point. If any situation that does not meet the preset conditions is found during the collection process (such as probe offset, new defects, etc.), the data collection will be immediately interrupted and repositioned to a new qualified sampling point.

[0066] It should be emphasized that in order to control the single factor, the poplar wood block samples were placed in a constant temperature and humidity chamber for equilibrium treatment to achieve mass balance. The poplar wood block samples were placed in the same environmental conditions to collect near-infrared spectra of different surfaces of the samples.

[0067] S2. Perform point-by-point or segment-by-segment adaptive filtering on the obtained near-infrared spectral data, and obtain purified spectral data in combination with inverse wavelet transform.

[0068] Furthermore, if Figure 3 As shown, step S2 includes:

[0069] S21. According to the distribution characteristics of the spectral data, the obtained near-infrared spectral data are decomposed by wavelet transform to obtain multi-layer detail coefficients and an approximate coefficient.

[0070] S22, for each layer of detail coefficients, a reference threshold is determined based on the noise characteristics of the current layer, and the reference threshold is multiplied by a multiple of the standard deviation of the current layer k To dynamically adjust to construct an adaptive filter.

[0071] S23. Using the constructed adaptive filter, perform point-by-point or segment-by-segment filtering on the detail coefficients of each layer after decomposition to remove noise and outliers.

[0072] S24. Reconstruct the detail coefficients and approximate coefficients after filtering through inverse wavelet transform to obtain purified spectral data.

[0073] In one embodiment, the spectral data is first subjected to multiplicative scatter correction (MSC) processing to reduce the influence of scattering effects on baseline drift and inter-spectral variation. The MSC processing process is as follows:

[0074] ;

[0075] in, X is the original spectral data, a and b Specifically, they are the slope and intercept of the linear regression equation used to correct each spectrum.

[0076] The data processed by MSC are then transformed by the first-order derivative (Savitsky-Golay derivate1, SG1st). The SG1st method calculates the smoothed derivative by fitting a polynomial to the spectral data within a specified window. The process can be expressed as:

[0077] ;

[0078] in, is the spectral data after derivative transformation, is the first i Points, c i To fit the polynomial coefficients, k is half the width of the window (i.e. half the number of smoothing points). By properly selecting the window size and derivative order, the main feature information can be retained and the interference of noise can be reduced.

[0079] Next, according to the distribution characteristics of the spectral data or the actual situation, the outlier detection threshold is adaptively set through wavelet transform to construct an adaptive filter to filter the spectral data point by point or segment by segment. Specifically, the spectral signal X is first decomposed into multiple layers of detail coefficients D l and an approximate coefficient A L ,as follows:

[0080] ;

[0081] In the formula, X is the spectral signal, D l is the multi-layer detail coefficient, A L is an approximate coefficient.

[0082] Next, an adaptive threshold is set on each layer detail coefficient based on the noise characteristics of the current layer, by multiplying the standard deviation of the layer Multiples of k To dynamically adjust the threshold: .

[0083] The above-mentioned adaptive filtering method can perform denoising point by point or segment by segment according to the actual noise distribution of the spectral data, and finally reconstruct a smooth spectral signal through inverse wavelet transform to effectively remove outliers and noise.

[0084] S3. According to the spectral wavelength selection strategy constructed based on the preset target, the purified spectral data is spectrally screened and the wavelength combination is optimized, and the optimized spectral data is input into the corresponding data file according to the serial number and the corresponding label data determined by the standard value of the chemical component content, so as to divide it into a training set and a test set.

[0085] Furthermore, if Figure 4 As shown, in step S3, according to the spectral wavelength selection strategy constructed based on the preset target, the spectral wavelength screening of the purified spectral data and the optimization of the wavelength combination include:

[0086] S31. quantify the pre-set requirement objectives including maximizing classification accuracy and minimizing prediction error into an objective function for evaluating different wavelength combinations.

[0087] Specifically, the classification accuracy maximization objective is defined as: f 1(W) = number of correctly classified samples / total number of samples, where W is the wavelength combination;

[0088] The prediction error minimization objective is defined as; , where we transform it into a maximization problem by subtracting the prediction error from 1, is the true value, is the predicted value.

[0089] Combining the two defined objectives, a comprehensive objective function is constructed: ;in, is the weight coefficient, which is used to balance the relative importance of the two objectives.

[0090] S32. Construct a spectral wavelength selection strategy that can dynamically adjust the wavelength combination according to the objective function value.

[0091] S33, the purified spectral data is processed as an initial spectral data set, and a pre-built algorithm model (genetic algorithm or particle swarm optimization) is used to search for an optimal combination in the wavelength space of the initial spectral data set.

[0092] In a specific embodiment, if the genetic algorithm model is selected, the following steps are performed:

[0093] (1) Binary encode each wavelength to form a chromosome. For example, M The wavelength can be encoded as a length of MInitialize the population and randomly generate N chromosomes as the initial solution.

[0094] (2) Use a fitness function related to the objective function to evaluate the quality of the chromosome. For example, the inverse or negative number of the objective function can be directly used as the fitness function. The higher the fitness function value, the better the corresponding wavelength combination.

[0095] (3) Select chromosomes to enter the next generation based on the fitness function value. Using the roulette wheel selection method, the probability of each chromosome being selected is proportional to its fitness function value.

[0096] (4) Randomly select two chromosomes to perform a crossover operation to generate a new chromosome. The crossover point can be randomly selected or at a fixed position. For example, in a single-point crossover, a crossover point is randomly selected and all genes of the two chromosomes after that point are exchanged.

[0097] (5) Randomly mutate certain genes in the chromosome to increase the diversity of the population. The probability of mutation is usually low. For example, in binary coding, genes at certain positions can be changed from 0 to 1 or from 1 to 0.

[0098] S34. In each iteration, a new candidate wavelength set is generated based on the current spectral data set, and the new candidate wavelength set is evaluated using the objective function.

[0099] S35, if the value of the objective function of the candidate wavelength set in the current iteration is greater than the value of the objective function of the candidate wavelength set in the previous iteration plus the preset improvement threshold τ, then accepting the current new candidate wavelength set as the spectral data for which wavelength selection has been completed with a decreasing acceptance probability;

[0100] In a specific embodiment, a random number τ between 0 and 1 is generated. If the value of the objective function of the candidate wavelength set in the current iteration is greater than the value of the objective function of the candidate wavelength set in the previous iteration plus the preset improvement threshold τ, the current new candidate wavelength combination can be accepted as a new wavelength combination. At the same time, in order to introduce randomness and avoid the algorithm from converging to a local optimal solution too early, an acceptance probability is introduced in the embodiment of the present invention, and then the current new candidate wavelength combination is selectively accepted as a new wavelength combination. The specific acceptance probability is:

[0101] ;

[0102] In the formula, exp is the natural exponential function, The value of the objective function for the set of candidate wavelengths in the current iteration, The value of the objective function for the set of candidate wavelengths in the previous iteration, k BThe rate of deceleration of progress, J is the progress parameter, which is multiplied by the progress coefficient Gradually from the initial value J 0 gradually decreases to a certain end value J f .

[0103] S36. If the value of the objective function in the current iteration is not greater than the value of the objective function in the previous iteration plus the preset improvement threshold τ, the current new set of candidate wavelengths is rejected, wavelengths are added and / or deleted based on prior knowledge or the contribution of each wavelength to the objective function, and the next round of iterative processing is performed.

[0104] Among them, the strategy of adding wavelengths is: randomly selecting wavelengths from unselected wavelengths or selecting wavelengths based on prior knowledge to add to the current new candidate wavelength set; the strategy of deleting wavelengths is: removing the wavelength with the smallest contribution to the objective function from the current new candidate wavelength set, or deleting it based on prior knowledge.

[0105] It should be clear that after each addition or removal of a wavelength, the value of the objective function is recalculated to verify whether these changes have brought improvements. If the value of the objective function has improved and exceeds the preset improvement threshold τ, these changes are accepted and the next round of iterative processing continues. If the value of the objective function has not improved significantly or even decreased, roll back to the wavelength set of the previous round and reconsider the added or removed wavelengths.

[0106] S37, after the iteration is completed, the current optimal wavelength combination and the corresponding objective function value are outputted. Preferably, the optimal wavelength combination selects data in the wavelength ranges of 450-2400nm, 780-2000nm, 780-1600nm, and 900-1100nm for modeling and analysis.

[0107] And, further, if Figure 5 In step S3, the optimized spectral data is input into the corresponding data file according to the sequence number and the corresponding label data determined by the standard value of the chemical component content, so as to divide the training set and the test set into:

[0108] P31. According to the obtained standard values ​​of chemical composition content, the corresponding grade label is determined according to the set standard value range. For example, the standard values ​​of chemical composition content are divided into high and low grades.

[0109] P32. Pair the spectral data with the wavelength selection completed with the corresponding chemical component content level label to form a data table to be sorted containing multiple composite data units; each composite data unit contains spectral data, label data and an initial serial number.

[0110] P33. Treat the data table to be sorted as the root node of the tree. Starting from the root node, continuously split the data table into child nodes, and repeat the splitting for each newly generated child node until each child node contains only one composite data unit.

[0111] P34. Internally sort all child nodes and merge two sorted child nodes into a new ordered node. During the merging process, if the elements in the two child nodes have the same label, perform secondary sorting according to the initial sequence number until all child nodes are merged into an ordered data table.

[0112] P35. In the ordered data table, a unique serial number is generated for each paired spectral data and label, and the unique serial number of each paired data is concatenated with the timestamp to form a unique identifier through hash processing.

[0113] P36. Use the file operation library or API to create a new empty data file and initialize the file header information (such as column name, data description, etc.).

[0114] P37. Traverse the paired spectral data and level labels, convert the spectral data into a format suitable for writing according to a predetermined format, convert the level label into a corresponding data representation, and use the unique identifier as the key field to form a complete data record together with the converted spectral data and level label, and write it into the data file according to the unique identifier.

[0115] Specifically, according to the predetermined format, convert the spectral data into a format suitable for writing (such as array, string, etc.). Convert the level label into the corresponding data representation (such as integer, enumeration value, etc.). Use the unique identifier as the key field to form a complete data record together with the spectral data and the level label. Write this data record into the data file, ensuring that each record is arranged in the same format and order. When all data is successfully written to the file, close the file handle to ensure that the data is correctly saved and will not be lost.

[0116] P38. Use random sampling or stratified sampling to divide the data stored in the data file into training sets and test sets to ensure that the data of each category is evenly distributed in the training set and the test set.

[0117] S4. Construct at least one deep learning model as a recognition model, and perform model training based on the training set and the test set. In addition, introduce a serial or parallel combination mechanism to fuse the output results of different models, and automatically output the poplar variety and pulpwood classification results.

[0118] Furthermore, if Figure 6 As shown, step S4 includes:

[0119] S41. According to the requirements of classification and recognition, select and build deep learning models including NIR-Net model, NIR-ResNet model and NIR-Inception model.

[0120] After an in-depth analysis of the specific requirements for poplar classification, such as the precision of classification, the requirements for real-time performance, and the acceptable error range, suitable candidate models were selected from models such as NIR-Net, NIR-ResNet, and NIR-Inception according to the requirements.

[0121] Specifically, the NIR-Net, NIR-ResNet, and NIR-Inception models are introduced respectively:

[0122] (1) The NIR-Net model consists of: 1 input layer, 8 one-dimensional convolutional layers, 2 maximum pooling layers, 1 fully connected layer and 1 output layer.

[0123] Its output layer directly takes the spectral data of wood samples preprocessed as input, and its shape is ( N ,1, L ),in N is the number of samples, L is the number of wavelengths.

[0124] The one-dimensional convolution layer is a one-dimensional convolutional network used to extract features. The sizes of the convolution kernels in the first three layers are 1×9, 1×7, and 1×5, respectively, and the sizes of the convolution kernels in the last five layers are all 1×3. The step size in each convolution kernel is set to 1. The first seven convolution layers of the model use 8 convolution kernels, and the eighth convolution layer uses 16 convolution kernels.

[0125] The maximum pooling layer has a pooling size of 1×3 and a stride of 2, and is located after the first and eighth convolutional layers, respectively.

[0126] The fully connected layer is used to classify the flattened feature vector, and the output layer uses the softmax activation function to output the probability distribution of each tree category.

[0127] (2) The NIR-ResNet model consists of: 1 input layer, 1 one-dimensional convolutional layer, 3 residual blocks, 2 maximum pooling layers, 1 fully connected layer and 1 output layer.

[0128] The convolution kernel size used in the initial one-dimensional convolution layer is 1×9, 8 convolution kernels are used, and the tanh activation function is adopted.

[0129] The residual block consists of two one-dimensional convolutional layers and a residual connection. The convolution kernel size of the convolutional layer is 1×3, 8 convolution kernels are used, the step size is set to 1, and the tanh activation function is adopted.

[0130] The parameters of the maximum pooling layer are consistent with the NIR-Net model.

[0131] (3) The NIR-Inception model consists of 1 input layer, 2 Inception modules, 4 one-dimensional convolutional layers, 2 maximum pooling layers, 1 fully connected layer and 1 output layer.

[0132] The Inception module includes five parallel one-dimensional convolution branches and one pooling branch. The one-dimensional convolution kernel sizes are 1×1, 1×3, 1×5, 1×7, and 1×9, respectively. The number of channels is 8. The pooling size of the pooling layer is 1×3, and the step size is 1. The outputs of the above parallel branches are spliced ​​in the channel dimension to form a feature matrix with 48 channels (8 channels × 6 branches). When splicing in the channel dimension, the following formula is used for splicing:

[0133] ;

[0134] in, Indicates i The output features of the convolution branches, the splicing operation can improve the feature expression ability of the model.

[0135] The maximum pooling layer parameters are consistent with the NIR-Net model.

[0136] S42. Structural adjustment and optimization of the selected model, including adjusting the number of convolution layers, convolution kernel size, and pooling layer settings, are performed to adapt to the specific tasks of poplar wood and pulpwood classification.

[0137] S43. By analyzing the characteristics of the selected model, a serial or parallel combination mechanism is introduced.

[0138] S44. In the serial combination mechanism, in a predetermined order, the NIR-Net model is first used for preliminary feature extraction, then passed to the NIR-ResNet model for in-depth processing, and finally the NIR-Inception model performs complex feature fusion and classification. The output of the NIR-Net model is used as the input of the NIR-ResNet model through the feature conversion layer, and then the output of the NIR-ResNet model is used as the input of the NIR-Inception model through the dimensionality reduction layer, and the end-to-end joint training of the entire serial combination model is completed based on the training set and the test set. Preferably, enhancement techniques such as feature selection, feature transformation or attention mechanism can be added in the feature transfer process to emphasize important features and suppress noise.

[0139] Through serial combination, each model can perform more refined classification based on the output of the previous model. For example, NIR-Net may be responsible for the initial distinction between poplar and other woods, NIR-ResNet further subdivides within the poplar category, and NIR-Inception finally determines the specific poplar species.

[0140] Specifically, the above-mentioned model training can use SGD (Stochastic Gradient Descent, SGD) or Adam optimizer.

[0141] S45. In the parallel combination mechanism, the NIR-Net model, the NIR-ResNet model and the NIR-Inception model are first trained independently, and then the same spectral data are simultaneously input into the NIR-Net model, the NIR-ResNet model and the NIR-Inception model, a unary learner is constructed and trained to minimize the prediction error after fusion, and the outputs of the NIR-Net model, the NIR-ResNet model and the NIR-Inception model are received and fused using the trained meta-learner.

[0142] In the parallel combination mechanism, the same spectral data is simultaneously input into the three models of NIR-Net, NIR-ResNet and NIR-Inception. Each model independently performs feature extraction and classification prediction to generate its own output. A meta-learner or fusion layer is designed to receive and fuse the outputs of the three models. Fusion strategies can include weighted averaging, voting mechanisms, learning weight allocation, etc. to ensure that the prediction results of different models can be effectively integrated.

[0143] S56. Automatically output the classification results of poplar and pulpwood based on the combined prediction results obtained from the serial or parallel combination mechanism.

[0144] Specifically, based on the combined prediction results obtained from the serial or parallel combination mechanism, the species of poplars are identified. And at the same time, whether the poplars are suitable for pulpwood is identified. The key to identifying suitable pulpwood is to determine the lignin content. For this purpose, the level of lignin value can be used as a label y , and the spectral data is used as input x In this way, the problem is transformed into a binary classification problem, that is, judging whether poplar is suitable for pulpwood based on the level of lignin content. This automated classification method can quickly and accurately identify poplar species and whether they are suitable for pulpwood, greatly improving work efficiency and identification accuracy.

[0145] In a specific embodiment, poplars are used as research objects, with a total of 6 varieties, and the poplar variety codes are: I107 Yang, Zhongxiong No. 1, Bofeng 3, Zhonghe No. 1, Lulin No. 3, Huanghuai 4. The sample information of the 6 varieties of poplars is shown in Table 1, and the Figure 7 Spectral data are shown.

[0146] Table 1 Sample information of six poplar species

[0147]

[0148] After the above steps S1-S5 are processed, accuracy, recall and F1-score are used as the model classification performance evaluation criteria. The results of this implementation case model on the test set are shown in Table 2.

[0149] Table 2 Classification results of different deep learning models with the same parameters

[0150]

[0151] In another specific embodiment, this implementation case uses poplar as the research object, with a total of 5 varieties, and the poplar variety codes are: Lulin No. 1, Lulin No. 2, Lulin No. 3, Lulin No. 9, and Lulin No. 16. The Lulin sample information is shown in Table 3. Figure 8 ,A total of 1062 near-infrared spectral data were collected in this case.

[0152] Table 3 Lulin sample information table

[0153]

[0154] After the above steps S1-S5 are processed, the results of the implementation case model on the test set are shown in Table 4.

[0155] Table 4 Classification results of different deep learning models with the same parameters

[0156]

[0157] Table 5 Results of qualitative identification of poplar wood quality classification

[0158]

[0159] In addition, an embodiment of the present invention further provides a method for determining the chemical composition content of a wood powder sample, which is applied to the rapid non-destructive identification method of poplar varieties and pulp materials as described above, comprising: weighing a wood powder sample with a mass of 0.1-0.5 g, and adding sulfuric acid with a volume of 2-4 mL and a concentration of 68%-76%, reacting at a temperature range of 27±3°C, and setting the reaction time to 50-70 min; after the reaction is completed, using ultrapure water with a volume of 74-94 mL, transferring the reaction solution to a first blue-capped transparent glass bottle with a capacity of 200-300 mL; by placing a solution with a concentration of 3%-5% The first and second blue-capped transparent glass bottles were filled with 80-120 mL of sulfuric acid and 40-60 mg of the corresponding monosaccharides were added to prepare calibration solutions of glucose, xylose and arabinose, respectively, and the concentration of the calibration solutions was 0.5 to 1.5 g / L; the samples in the first and second blue-capped transparent glass bottles were placed in an autoclave and reacted for 50-70 min at a temperature range of 118°C to 124°C; after the reaction was completed, the reaction product was filtered through a filtration device, and the filtrate was washed with distilled water preheated to the set temperature until the filtration was complete, and then the mixture was sterilized at a temperature range of 100°C to 110°C. ℃ oven for 3-5h, and calculate the content of acid-insoluble lignin by weighing; according to the filtrate obtained by suction filtration, the absorbance is measured between 200-210nm by ultraviolet-visible spectrophotometry with 3%-5% sulfuric acid as a reference. If the absorbance value exceeds the range of 0.5-1.5, it is diluted with sulfuric acid of the corresponding concentration and re-measured, and the molar absorption coefficient is selected between 108-118L·g-1·cm-1 to calculate the content of acid-soluble lignin; the filtrate with a volume of 15-25mL is placed in a beaker with a capacity of 40-60mL, and the pH value is neutralized to between 4-6 with CaCO3 through magnetic stirring. When the pH value is close to 4, the pH value is adjusted to between 5-6. After sufficient stirring, the obtained sample is loaded with a syringe filter; the contents of glucose, xylose and arabinose in the filtrate are determined by HPLC technology, and the contents of cellulose and hemicellulose are calculated according to the determination results.

[0160] In a specific embodiment, the specific process of the method for determining the chemical composition content of the wood powder sample is as follows:

[0161] (1) First, weigh 0.3 g of sample, add 3 mL of 72% sulfuric acid, and react at 30±3 °C for 60 min. After the reaction, transfer the reaction solution in step 3 to a 250-mL blue-capped transparent glass bottle with 84 mL of ultrapure water.

[0162] (2) Prepare calibration solutions of glucose, xylose, and arabinose, all set to 1 g / L. (Prepare in a 100-mL blue-capped transparent glass bottle, add 50 mg of monosaccharide to 50 mL of 4% sulfuric acid)

[0163] (3) Place the samples in (1) and (2) in an autoclave at 121 °C for 60 min.

[0164] (4) Filter the reacted 3.3 (the filter paper must be weighed in advance), wash it with hot distilled water until it is completely filtered out, dry it in an oven at 105 °C for 4 h, and weigh it to calculate the content of acid-insoluble lignin (including ash).

[0165] (5) Collect the filtrate from (4) and calculate the content of acid-soluble lignin and sugar.

[0166] (6) Determine the content of acid-soluble lignin by UV-visible spectrophotometry. Use 4% sulfuric acid as a reference and measure the absorbance at 205 nm. If the value exceeds the range of 0.7-1.0, dilute with 4% sulfuric acid and re-measure. Select 113 L·g-1·cm-1 as the molar absorption coefficient and calculate the content.

[0167] (7) Take 20 mL of the filtrate into a 50-mL beaker and neutralize with CaCO3 to a pH of 5-6 (while stirring with a magnetic bar). When the pH is around 4, fine-tune the pH to 5-6, stir thoroughly, and load the sample with a syringe filter. Determine the content of glucose, xylose, and arabinose by HPLC, and calculate the content of cellulose and hemicellulose.

[0168] in,

[0169] , C is the content;

[0170] ;

[0171] Lignin content = acid soluble lignin content + acid insoluble lignin content. Preferably, the acid soluble lignin content is less than 5% and can be ignored.

[0172] Furthermore, an embodiment of the present invention provides a rapid non-destructive identification system for poplar varieties and pulpwood, including: a sample processing and spectral acquisition module, which is used to use sensors to detect and screen poplar log segments that meet specific conditions, control processing equipment to process poplar log segments into wood blocks or wood powder samples, and use a near-infrared spectrometer to capture near-infrared spectral data of wood blocks or wood powder samples. A material property testing module, which is used to perform standard value detection of at least one chemical component content on the wood powder sample. A data purification module, which is used to perform point-by-point or segmented adaptive filtering processing on the obtained near-infrared spectral data, and obtain purified spectral data in combination with inverse wavelet transformation. A data set output module, which is used to perform spectral wavelength screening and optimize wavelength combinations on the purified spectral data according to a spectral wavelength selection strategy constructed based on a preset target, and input the optimized spectral data and the corresponding label data determined by the chemical component content of the wood powder sample obtained by detection in advance into the corresponding data file to divide it into a training set and a test set. The wood classification module is used to build at least one deep learning model as a recognition model, and to train the model based on the training set and the test set, and to introduce a serial or parallel combination mechanism to fuse the output results of different models, and automatically output the poplar species and pulpwood classification results.

[0173] Furthermore, an embodiment of the present invention provides a device for rapid and non-destructive identification of poplar varieties and pulp materials, comprising: at least one database; and a memory communicatively connected to the at least one database; wherein the memory stores instructions executable by the at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the method for rapid and non-destructive identification of poplar varieties and pulp materials as described above.

[0174] Furthermore, an embodiment of the present invention provides a computer-readable medium having computer-executable instructions stored thereon, and when the executable instructions are executed by a processor, the method for rapid and non-destructive identification of poplar species and pulpwood as described above is implemented.

[0175] It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0176] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments after knowing the basic creative concept. Therefore, the present invention should be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present invention.

[0177] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of the present invention and its equivalent technology, the present invention should also include these modifications and variations.

Claims

1. A method for rapid and non-destructive identification of poplar species and pulpwood, characterized in that: include: Using sensors to detect and screen poplar logs that meet specific conditions, controlling processing equipment to process the poplar logs into wood blocks or wood powder samples, and using a near-infrared spectrometer to capture near-infrared spectrum data of the wood blocks or wood powder samples; The obtained near-infrared spectral data is subjected to point-by-point or segment-by-segment adaptive filtering, and the purified spectral data is obtained by combining the inverse wavelet transform; According to the spectral wavelength selection strategy constructed based on the preset target, the spectral wavelengths of the purified spectral data are screened and the wavelength combination is optimized, and the optimized spectral data and the corresponding label data determined by the chemical composition content of the wood powder sample obtained by testing in advance are input into the corresponding data file to divide the training set and the test set; Construct at least one deep learning model as a recognition model, and perform model training based on the training set and the test set. In addition, introduce a serial or parallel combination mechanism to fuse the output results of different models, and automatically output the poplar species and pulpwood classification results, including: According to the requirements of classification and recognition, deep learning models including NIR-Net model, NIR-ResNet model and NIR-Inception model are selected and constructed; The selected model was structurally adjusted and optimized, including adjusting the number of convolutional layers, convolution kernel size, and pooling layer settings, to suit the specific tasks of poplar and pulpwood classification; By analyzing the characteristics of the selected model, a serial or parallel combination mechanism is introduced; In the serial combination mechanism, according to the predetermined order, the output of the NIR-Net model is used as the input of the NIR-ResNet model through the feature conversion layer, and then the output of the NIR-ResNet model is used as the input of the NIR-Inception model through the dimension reduction layer, and the end-to-end joint training of the entire serial combination model is completed based on the training set and the test set; In the parallel combination mechanism, the NIR-Net model, NIR-ResNet model and NIR-Inception model are first trained independently, and then the same spectral data is simultaneously input into the NIR-Net model, NIR-ResNet model and NIR-Inception model, a unary learner is constructed and trained to minimize the prediction error after fusion, and the trained meta-learner is used to receive and fuse the outputs of the NIR-Net model, NIR-ResNet model and NIR-Inception model; Based on the combined prediction results obtained from the serial or parallel combination mechanism, the classification results of poplar and pulpwood are automatically output.

2. The method for rapid and non-destructive identification of poplar species and pulpwood according to claim 1, characterized in that: The sensor is used to detect and screen the poplar logs that meet specific conditions, and the poplar logs are processed into wood blocks or wood powder samples through an integrated processing line. The near-infrared spectrometer is used to capture the near-infrared spectrum data of the wood blocks or wood powder samples, including: Using sensor arrays to detect and identify defects in poplar logs; Screening and marking at least one poplar log segment on the poplar log according to a preset specific condition; Sending instructions to the processing equipment to control the processing equipment to process the selected poplar log segments into blank samples of predetermined sizes according to a preset processing flow, and to perform air drying; Controlling the processing equipment to process the blank sample after air drying into a number of wood blocks of the same specification, and automatically dividing the number of wood blocks into a first part of the sample and a second part of the sample; Controlling the pulverizer to pulverize the second part of the sample into wood powder, and controlling the filtering device to filter the wood powder into a sieve with a specific mesh number to obtain a wood powder sample; The near-infrared spectrometer is used to capture the near-infrared spectral data of the wood blocks and wood powder samples of the first part of the samples. During the process of capturing the near-infrared spectral data, the camera deployed at the fiber optic probe of the near-infrared spectrometer is used to identify defects, heartwood or sapwood again to avoid collecting erroneous data.

3. The method for rapid non-destructive identification of poplar species and pulpwood according to claim 1, characterized in that: The obtained near-infrared spectral data is subjected to point-by-point or segment-by-segment adaptive filtering, and the purified spectral data obtained by combining the inverse wavelet transform includes: According to the distribution characteristics of spectral data, the obtained near-infrared spectral data are decomposed by wavelet transform to obtain multi-layer detail coefficients and an approximate coefficient. For each layer of detail coefficients, a reference threshold is determined based on the noise characteristics of the current layer, and the reference threshold is dynamically adjusted by multiplying the multiple of the standard deviation of the current layer to construct an adaptive filter; Using the constructed adaptive filter, the detail coefficients of each layer after decomposition are filtered point by point or segment by segment to remove noise and outliers; Through inverse wavelet transform, the detail coefficients and approximate coefficients after filtering are reconstructed to obtain the purified spectral data.

4. The method for rapid non-destructive identification of poplar species and pulpwood according to claim 1, characterized in that: According to the spectral wavelength selection strategy built based on the preset target, the spectral wavelengths of the purified spectral data are screened and the wavelength combination is optimized, including: The pre-set requirement objectives including maximizing classification accuracy and minimizing prediction error are quantified into an objective function for evaluating different wavelength combinations; Construct a spectral wavelength selection strategy that can dynamically adjust the wavelength combination according to the objective function value; The purified spectral data are processed as an initial spectral data set, and a pre-built algorithm model is used to search for an optimal combination in the wavelength space of the initial spectral data set; In each iteration, a new set of candidate wavelengths is generated based on the current set of spectral data, and the new set of candidate wavelengths is evaluated using the objective function; If the value of the objective function of the candidate wavelength set in the current iteration is greater than the value of the objective function of the candidate wavelength set in the previous iteration plus the preset improvement threshold τ, the current new candidate wavelength set is accepted as the spectral data for which wavelength selection has been completed with a decreasing acceptance probability; If the value of the objective function in the current iteration is not greater than the value of the objective function in the previous iteration plus the preset improvement threshold τ, the current new set of candidate wavelengths is rejected, wavelengths are added and / or deleted based on prior knowledge or the contribution of each wavelength to the objective function, and the next round of iterative processing is performed; After the iteration is completed, the current optimal wavelength combination and the corresponding objective function value are output; Among them, the acceptance probability is: ； In the formula, exp is the natural exponential function, The value of the objective function for the set of candidate wavelengths in the current iteration, The value of the objective function for the candidate wavelength set in the previous iteration, k B The rate of deceleration of progress, J is a progress parameter, configured to be multiplied by the progress factor Gradually from the initial value J 0 gradually decreases to the end value J f .

5. The method for rapid non-destructive identification of poplar species and pulpwood according to claim 1, characterized in that: The optimized spectral data and the corresponding label data determined by the chemical composition content of the wood powder sample obtained by testing in advance are input into the corresponding data files to divide the training set and the test set into: According to the obtained standard value of chemical composition content, the corresponding grade label is determined according to the set standard value range; Pairing the spectrum data after wavelength selection with the corresponding chemical component content level label to form a data table to be sorted containing multiple composite data units; each composite data unit contains spectrum data, label data and an initial sequence number; The data table to be sorted is regarded as the root node of the tree. Starting from the root node, the data table is continuously divided into child nodes, and each newly generated child node is repeatedly divided until each child node contains only one composite data unit. Internally sort all child nodes, merge two sorted child nodes into a new ordered node. During the merging process, if the elements in the two child nodes have the same label, perform secondary sorting according to the initial sequence number until all child nodes are merged into an ordered data table; In the ordered data table, a unique serial number is generated for each paired spectral data and label, and the unique serial number of each paired data is concatenated with the timestamp to form a unique identifier through hash processing; Use the file operation library or API to create a new empty data file and initialize the file header information; Traversing the paired spectral data and grade labels, converting the spectral data into a format suitable for writing according to a predetermined format, converting the grade label into a corresponding data representation, and using the unique identifier as a key field to form a complete data record together with the converted spectral data and grade label, and writing it into the data file according to the unique identifier; The data stored in the data file is divided into training set and test set by random sampling or stratified sampling.

6. A method for determining the chemical composition content of a wood powder sample, applied to the rapid non-destructive identification method for poplar species and pulpwood as claimed in any one of claims 1 to 5, characterized in that: include: Weigh 0.1-0.5g of wood powder sample, add 2-4mL of sulfuric acid with a concentration of 68%-76%, react at a temperature range of 27±3℃, and set the reaction time to 50-70min; After the reaction is completed, use ultrapure water with a volume of 74-94 mL to transfer the reaction solution to a first blue-capped transparent glass bottle with a capacity of 200-300 mL; By placing 3% to 5% sulfuric acid in a second blue-capped transparent glass bottle with a capacity of 80 to 120 mL and adding 40 to 60 mg of the corresponding monosaccharide, glucose, xylose and arabinose calibration solutions are prepared respectively, and the concentration of the calibration solution is 0.5 to 1.5 g / L; The samples in the first blue-capped transparent glass bottle and the second blue-capped transparent glass bottle were placed in an autoclave and reacted at a temperature ranging from 118°C to 124°C for 50 to 70 minutes; After the reaction is completed, the reaction product is filtered through a filtering device, and the filtrate is washed with distilled water preheated to a set temperature until the filtration is complete, and then dried in an oven at a temperature range of 100°C to 110°C for 3 to 5 hours, and the content of acid-insoluble lignin is calculated by weighing; According to the filtrate obtained by filtration, the absorbance is measured between 200 and 210 nm by UV-visible spectrophotometry, with 3% to 5% sulfuric acid as a reference. If the absorbance value exceeds the range of 0.5 to 1.5, it is diluted with sulfuric acid of the corresponding concentration and re-measured, and the molar absorption coefficient is selected between 108 and 118 L·g-1·cm-1 to calculate the content of acid-soluble lignin; Take 15-25 mL of the filtrate and place it in a beaker with a capacity of 40-60 mL. Use CaCO3 to neutralize the pH value to between 4 and 6 by magnetic stirring. When the pH value is close to 4, adjust the pH value to between 5 and 6. After sufficient stirring, use a syringe filter to load the obtained sample. The contents of glucose, xylose and arabinose in the filtrate were determined by HPLC technology, and the contents of cellulose and hemicellulose were calculated based on the determination results; in, , C is the content; ； Lignin content = acid soluble lignin content + acid insoluble lignin content.

7. A rapid non-destructive identification system for poplar species and pulpwood, characterized in that: include: A sample processing and spectrum acquisition module, which is used to detect and screen poplar logs that meet specific conditions using sensors, control processing equipment to process the poplar logs into wood blocks or wood powder samples, and use a near-infrared spectrometer to capture near-infrared spectrum data of the wood blocks or wood powder samples; A material property testing module, used to perform a standard value test on at least one chemical component content of a wood powder sample; A data purification module is used to perform point-by-point or segment-by-segment adaptive filtering on the obtained near-infrared spectral data, and obtain the purified spectral data in combination with inverse wavelet transform; The data set output module is used to screen the spectral wavelengths of the purified spectral data and optimize the wavelength combination according to the spectral wavelength selection strategy constructed based on the preset target, and input the optimized spectral data and the corresponding label data determined by the chemical composition content of the wood powder sample obtained by testing in advance into the corresponding data file to divide it into a training set and a test set; The wood classification module is used to build at least one deep learning model as a recognition model, and to train the model based on the training set and the test set, and to introduce a serial or parallel combination mechanism to fuse the output results of different models, and automatically output the classification results of poplar species and pulpwood, including: According to the requirements of classification and recognition, deep learning models including NIR-Net model, NIR-ResNet model and NIR-Inception model are selected and constructed; The selected model was structurally adjusted and optimized, including adjusting the number of convolutional layers, convolution kernel size, and pooling layer settings, to suit the specific tasks of poplar and pulpwood classification; By analyzing the characteristics of the selected model, a serial or parallel combination mechanism is introduced; In the serial combination mechanism, according to the predetermined order, the output of the NIR-Net model is used as the input of the NIR-ResNet model through the feature conversion layer, and then the output of the NIR-ResNet model is used as the input of the NIR-Inception model through the dimension reduction layer, and the end-to-end joint training of the entire serial combination model is completed based on the training set and the test set; In the parallel combination mechanism, the NIR-Net model, NIR-ResNet model and NIR-Inception model are first trained independently, and then the same spectral data is simultaneously input into the NIR-Net model, NIR-ResNet model and NIR-Inception model, a unary learner is constructed and trained to minimize the prediction error after fusion, and the trained meta-learner is used to receive and fuse the outputs of the NIR-Net model, NIR-ResNet model and NIR-Inception model; Based on the combined prediction results obtained from the serial or parallel combination mechanism, the classification results of poplar and pulpwood are automatically output.

8. A rapid and non-destructive identification device for poplar species and pulpwood, characterized in that: include: at least one database; and a memory communicatively connected to the at least one database; wherein the memory stores instructions executable by the at least one database, and the instructions are executed by the at least one database so that the at least one database can execute the method for rapid and non-destructive identification of poplar varieties and pulpwood as described in any one of claims 1 to 5.

9. A computer-readable medium having computer-executable instructions stored thereon, characterized in that: When the executable instructions are executed by the processor, the method for rapid and non-destructive identification of poplar species and pulpwood as described in any one of claims 1 to 5 is implemented.

Images