Nondestructive testing method for moisture content of Chinese art paper based on near infrared reflection spectrum

Abstract

The invention relates to the field of substance moisture content detection, and particularly discloses a Xuan paper moisture content nondestructive testing method based on a near infrared reflection spectrum, which comprises the following steps: acquiring near infrared spectrum data of to-be-tested Xuan paper, inputting the near infrared spectrum data into a detection model, and then outputting a moisture content value of the to-be-tested Xuan paper, the method for constructing the detection model comprises the following steps: a, sample preparation: generating a multi-gradient water content Chinese art paper sample set through humidity control; b, spectrum acquisition: acquiring a reflection spectrum of the Chinese art paper sample in a set wavelength range; c, characteristic wavelength screening: carrying out characteristic wavelength extraction on the original spectrum data by adopting a competitive self-adaptive reweighting algorithm; and d, model training: establishing a partial least squares regression prediction model as a detection model based on the screened characteristic wavelengths. According to the method disclosed by the invention, the water content of the Chinese art paper is measured by utilizing the online nondestructive measurement characteristic of the NIRS technology, so that the rapid and nondestructive measurement of the water content of the Chinese art paper cultural relic body is effectively realized, and the preventive protection of the paper cultural relics is realized.

Description

Technical Field

[0001] The present invention relates to the field of moisture content detection of substances, and particularly to a non-destructive method for detecting the moisture content of rice paper based on near-infrared reflectance spectroscopy. Background Art

[0002] Paper cultural relics are extremely vulnerable to damage during the preservation process, especially the influence of environmental humidity is particularly significant. As a hygroscopic material, paper cultural relics are prone to absorbing water in a high-humidity environment, resulting in moisture absorption, microbial growth, and ultimately mildew; in a low-humidity environment, they will dehydrate, leading to deformation and cracking. Therefore, the change of environmental humidity directly affects the moisture content of paper cultural relics, and further affects their preservation status and lifespan. However, there is no simple corresponding relationship between environmental humidity and the moisture content of the paper cultural relics themselves. Therefore, it is crucial to explore the moisture content of paper cultural relics themselves. By monitoring and controlling the moisture content of paper cultural relics themselves, the damage caused by environmental humidity to cultural relics can be effectively prevented, ensuring that cultural relics are properly protected, extending their lifespan, and maintaining their historical and cultural value.

[0003] Traditional paper moisture detection mainly relies on the current national standard GB / T 462-2023 ("Determination of moisture content of analytical samples of paper, board and pulp"). Although the results of this drying measurement method are accurate, it has significant limitations for precious paper cultural relics. First of all, the drying measurement method will damage paper cultural relics. After high-temperature heating, the paper will show obvious discoloration and embrittlement, and even cause cracking, which is not applicable to non-renewable paper cultural relics; secondly, the drying method requires a large amount of samples. Most calligraphy and painting cultural relics are single-layer papers before being mounted, and the drying method cannot be used for measurement; in addition, there is currently no method specifically for measuring the moisture content of single-layer paper cultural relics themselves.

[0004] Near-infrared spectroscopy (NIRS) mainly measures the moisture content in objects non-destructively through the absorption of the O-H group. In recent years, NIRS has achieved remarkable results in detecting the moisture content of objects. For example, Peng et al. (Peng D, Liu Y, Yang J, et al. Journal of Spectroscopy, 2021, 2021: 1-9) used the method of combining near-infrared spectroscopy with chemometrics to successfully predict the moisture content of walnut kernels through the SNV-FD-PLSR model. Wang et al. (Wang S, Wu Z, Cao C, et al. Sensors, 2023, 23(2): 666) established an on-line detection system for the moisture content of fresh tea leaves by combining near-infrared spectroscopy with partial least squares (PLS), realizing rapid and non-destructive measurement of the moisture content of tea leaves. There is also Li et al. (Li Y, Xia H, Liu Y, et al. Forests, 2023, 14(5): 883) who successfully predicted the moisture content of Masson pine seedlings by combining support vector regression (SVR)-adaptive boosting (AdaBoost)-partial least squares regression (PLSR)-AdaBoost model.

[0005] The above-mentioned existing technologies show that the NIRS technology has high feasibility and potential in measuring the moisture content of samples. However, although NIRS has been widely used in moisture content measurement, there are few reports on the measurement of the moisture content of rice paper. Therefore, if the on-line non-destructive measurement characteristics of NIRS technology can be used to measure the moisture content of rice paper, it can effectively achieve rapid and non-destructive measurement of the moisture content of the rice paper cultural relics themselves, so as to realize the preventive protection of rice paper cultural relics. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a non-destructive detection method for the moisture content of rice paper based on near-infrared reflectance spectroscopy, which uses the on-line non-destructive measurement characteristics of NIRS technology to measure the moisture content of rice paper, effectively realizing rapid and non-destructive measurement of the moisture content of the rice paper cultural relics themselves, so as to realize the preventive protection of paper cultural relics.

[0007] To achieve the above purpose, the present invention provides a non-destructive detection method for the moisture content of rice paper based on near-infrared reflectance spectroscopy, including obtaining the near-infrared spectral data of the rice paper to be measured, and inputting the near-infrared spectral data into the detection model to output the moisture content value of the rice paper to be measured;

[0008] Among them, constructing the detection model includes the following steps:

[0009] a. Sample preparation: Generate a multi-gradient moisture content rice paper sample set through humidity control;

[0010] b. Spectral acquisition: Obtain the reflection spectrum of the rice paper sample within a set wavelength range;

[0011] c. Characteristic wavelength screening: Use the competitive adaptive reweighted sampling algorithm to extract characteristic wavelengths from the original spectral data;

[0012] d. Model training: Establish a partial least squares regression prediction model based on the screened characteristic wavelengths as the detection model.

[0013] Furthermore, in step a, the rice paper samples are placed in a closed environmental chamber, and the humidity of the environmental chamber is set to several humidity environments within 37% - 97% RH. The same number of samples are placed in each environmental chamber, and the samples are left standing in the closed environmental chamber for 5 - 9 days.

[0014] Furthermore, the actual moisture content of the samples is measured by the drying method.

[0015] Furthermore, in step b, collect the original near-infrared spectrum of the rice paper sample in the 900 - 1700 nm band.

[0016] Furthermore, in step b, the sample collection is carried out in a closed dark box. After the sample to be measured is taken out of the closed environmental chamber, the spectral data is collected within 50 s. Each sample to be measured is continuously collected M times. Subsequently, take the average value of the spectral data collected M times as the spectral data of this sample.

[0017] Furthermore, in step c, the original spectral data is the spectral data obtained without preprocessing operations.

[0018] Furthermore, in step c, perform 50 ± 5 Monte Carlo samplings to dynamically screen out 55 - 61 characteristic wavelengths.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects:

[0020] A non-destructive detection method for the moisture content of rice paper based on near-infrared reflectance spectroscopy provided by the present invention verifies the feasibility of non-destructively detecting the moisture content of rice paper by near-infrared spectroscopy, utilizes the online non-destructive measurement characteristics of NIRS technology to measure the moisture content of rice paper, effectively realizes the rapid and non-destructive measurement of the moisture content of the rice paper cultural relics body, provides a reliable detection technical means for measuring the moisture content of rice paper cultural relics, and realizes the preventive protection of paper cultural relics.

[0021] The advantages of the additional aspects of the present invention will be partly given in the following description, partly will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic flow chart of the present invention;

[0023] Figure 2 is the original spectral curve of the Xuan paper's near-infrared spectrum;

[0024] Figure 3 is the process diagram of screening the characteristic wavelengths of the Xuan paper spectrum by the SPA algorithm;

[0025] Figure 4 is the process diagram of screening the characteristic wavelengths by the CARS algorithm;

[0026] Figure 5 is the comparison diagram of the actual value and the predicted value of SNV-PLSR;

[0027] Figure 6 is the comparison diagram of the actual value and the predicted value of the best prediction model of PLSR under the characteristic wavelengths of SPA and CARS;

[0028] Figure 7 is the comparison diagram of the actual value and the predicted value of the best prediction model of DL-BPNN under the characteristic wavelengths of SPA and CARS;

[0029] Figure 8 is the comparison diagram of the five groups of best combination validation sets. Specific implementation mode

[0030] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Embodiment

[0032] A non-destructive detection method for the moisture content of Xuan paper based on near-infrared reflectance spectroscopy provided in this embodiment includes obtaining the near-infrared spectral data of the Xuan paper to be measured, and inputting the near-infrared spectral data into the detection model to output the moisture content value of the Xuan paper to be measured;

[0033] Among them, constructing the detection model includes the following steps:

[0034] a. Sample preparation: Generating a multi-gradient moisture content Xuan paper sample set through humidity control;

[0035] b. Spectral acquisition: Obtaining the reflectance spectrum of the Xuan paper sample within a set wavelength range;

[0036] c. Characteristic wavelength screening: Using the competitive adaptive reweighted sampling algorithm to extract the characteristic wavelengths from the original spectral data;

[0037] d. Model training: Establishing a partial least squares regression prediction model based on the screened characteristic wavelengths as the detection model.

[0038] Specifically, the experimental content corresponding to the method shown in this embodiment is as follows.

[0039] 1. Experimental Section

[0040] 1.1 Paper Samples

[0041] The paper samples were cotton four-foot single Xuan paper provided by the Chongqing China Three Gorges Museum. They were cut into strips approximately 6 cm × 6 cm, and each sample weighed about 2 g. The paper samples were placed in a closed environmental chamber. The humidity in the environmental chamber was set to 7 humidity environments: 37%RH (silica gel), 47%RH (anhydrous potassium carbonate), 57%RH (sodium bromide), 67%RH (urea), 77%RH (sodium chloride), 87%RH (potassium chloride), and 97%RH (potassium sulfate). 30 samples were placed in each environmental chamber, for a total of 210 samples. The samples were left standing in the closed environmental chamber for 7 days to ensure that the paper reached an equilibrium state in their respective humidity environments.

[0042] The actual moisture content of each sample was measured by drying according to the national standard GB / T 462-2023. After the samples were taken out of the closed environmental chamber, they were placed in a moisture tester within 50 s for moisture content determination. The equipment used was the XY-P120G moisture meter. The measured moisture content range of all samples was 6.35%~15.55%.

[0043] 1.2 Spectral Data Acquisition System

[0044] The spectral data acquisition system consisted of a cooled near-infrared spectrometer (LISpec-NIR4000-1.7TEC) from Leisen Optics (Shenzhen) Co., Ltd., a reflective optical fiber (IFR-7IR200-2-GS), a tungsten halogen light source, an optical fiber probe holder, a stage, and a computer.

[0045] The spectral acquisition software was LispecView supporting the spectrometer. The integration time was set to 900 ms, the average number of measurements was 2 times, and the spectral acquisition wavelength range was 900 nm~1700 nm. To ensure the accuracy of the experimental data, all sample acquisitions were carried out in a closed dark box. After the samples to be measured were taken out of the closed environmental chamber, the spectral data acquisition was completed within 50 s. Each sample to be measured was continuously acquired 5 times; subsequently, the average value of the spectral data acquired 5 times was taken as the spectral data of this sample. For example, the original spectral curve of Xuan paper near-infrared spectrum is as Figure 2 shown.

[0046] 1.3 Processing Method of Spectral Data

[0047] 1.3.1 Dataset Division

[0048] To effectively predict the moisture content of Xuan paper, this embodiment adopts a sample set partitioning method based on joint X-Y distances (SPXY) to ensure that the training set and the validation set are representative and diverse in terms of spectral data (X) and moisture content (Y), thereby improving the generalization ability and stability of the model. Specifically, 210 paper samples are divided into 168 training sets and 42 validation sets according to a ratio of 4:1. The maximum, minimum, average, and standard deviation of the moisture content of the training set and the validation set partitioned by the SPXY method are very close, as shown in Table 1. This indicates that the sample partitioning has good representativeness and uniformity, which is conducive to establishing a stable prediction model.

[0049] Table 1

[0050]

[0051] 1.3.2 Data preprocessing

[0052] According to the conventional understanding, due to the influence of instrument noise and external environment, there may be interference signals in the collected spectral data, resulting in baseline drift, scattering effect, and noise in the original spectrogram. Through analysis, it is found that the noise is relatively large in the ranges of 900nm - 954nm and 1667nm - 1700nm at both ends of the instrument, the signal-to-noise ratio of the data is low, which affects the data quality and is not suitable for further analysis. In addition, the characteristic absorption peaks of the O-H stretching vibration and bending vibration of water molecules are concentrated near 1000nm and 1500nm, and these absorption characteristics can effectively reflect the change of moisture content in the paper. Therefore, the band of 954nm - 1667nm is selected for subsequent analysis. This band not only covers the key absorption characteristics of moisture but also effectively avoids the high-noise area to improve the robustness and accuracy of the model.

[0053] Some prediction models in this embodiment adopt the following spectral preprocessing methods: Standard Normal Variate (SNV), Baseline Correction (BC), and Normalization (Norm). In addition, during joint preprocessing, wavelet transform (WT) is combined with other methods, including WT-SNV, WT-BC, and WT-Norm.

[0054] 1.3.3 Successive projection algorithm

[0055] The Successive Projections Algorithm (SPA) aims to reduce multicollinearity and improve the prediction performance of the model. SPA reduces redundant information by selecting variables with the smallest projection similarity, thereby selecting the most representative characteristic wavelengths. This embodiment demonstrates the feature selection process of the un-preprocessed full-band. The same SPA method is used for other preprocessing methods. As Figure 3 shown, the selection process of characteristic wavelengths is determined by the Root Mean Square Error (RMSE), that is, a lower RMSE indicates better prediction performance of the model. When the RMSE reaches the minimum, the corresponding number of characteristic wavelengths is the optimal number of characteristic wavelengths. Figure 3 In (a), it shows that when the RMSE is 0.84113, the optimal number of characteristic wavelengths is 56, and the corresponding characteristic wavelength points in the original spectrum are as Figure 3 shown in (b).

[0056] 1.3.4 Competitive Adaptive Reweighted Sampling

[0057] Competitive Adaptive Reweighted Sampling (CARS) is based on the Monte Carlo method and screens out the most representative characteristic wavelengths through multiple samplings and reweightings. This embodiment demonstrates the feature selection process of the un-preprocessed original wavelengths. The same method is used for other preprocessing methods.

[0058] As Figure 4 shown, the number of Monte Carlo samplings is 50 times. Figure 4 In (a), it shows the changing trend of RMSECV with the number of samplings. RMSECV first slowly decreases to the minimum value of 0.58845 and then gradually increases. This indicates that at the minimum value of RMSECV, the selected set of characteristic wavelengths is the most relevant, and further increasing the number of samplings will eliminate the wavelengths related to the moisture content, resulting in an increase in the RMSECV value. Figure 4 In (b), it shows the change of regression coefficients with the number of samplings. The red marked point indicates that the optimal number of samplings is 22 times. Figure 4 In (c), it shows the change of the number of selected wavelengths with the number of samplings. The red marked point indicates that the optimal number of selected wavelength points is 58.

[0059] 1.4 Spectral Information Modeling Methods and Evaluation Criteria

[0060] 1.4.1 Partial Least Squares Model

[0061] Partial Least Squares Regression (PLSR) is used to process highly correlated and multicollinear spectral data. PLSR establishes a linear relationship between them by considering the predictor variables (X) and the response variable (Y). Finding the optimal number of principal components is a crucial step in PLSR modeling. Too few principal components may fail to capture the main information in the data, while too many may introduce noise and redundant information, leading to overfitting of the model. In this embodiment, the leave-one-out method is used in combination with RMSE and the coefficient of determination (R2) to determine the optimal number of principal components. Taking the un-preprocessed full-band data as an example, as the number of principal components increases, RMSE first decreases and then levels off, while R2 gradually increases and stabilizes, thereby determining the optimal number of principal components.

[0062] 1.4.2 Double-Layer Backpropagation Neural Network Model

[0063] The Double-Layer Backpropagation Neural Network Model (DL-BPNN) is an extension of the traditional BPNN, consisting of an input layer, two hidden layers, and an output layer. The neurons in each layer are fully connected to the neurons in the next layer. The two hidden layers perform non-linear transformations on the data respectively to capture the deep features of the input data, thus better fitting complex non-linear data relationships. It adjusts the network weights through the forward propagation and error backpropagation algorithms to minimize the prediction error.

[0064] In this embodiment, the training function of DL-BPNN uses Levenberg-Marquardt (trainlm), and the node transfer functions are selected as tansig and purelin. To optimize the network structure, the Grid Search method is used to select the point with the minimum RMSE as the optimal number of neurons in the hidden layer. This method ensures the best performance of the model, enabling it to more accurately reflect complex non-linear relationships when predicting the moisture content of rice paper.

[0065] 1.4.3 Evaluation Criteria

[0066] The quality of a model is measured by certain metrics. In this embodiment, the root mean square errors (RMSEC and RMSEP) of the training set and the validation set, and the coefficient of determination To evaluate the prediction accuracy of the model. RMSE measures the average difference between the predicted value and the actual value, that is, the smaller the RMSE, the better, because a lower RMSE indicates a small prediction error and high prediction accuracy of the model. R2 is used to evaluate the data interpretation ability of the model. The closer R2 is to 1, the stronger the data interpretation ability of the model and the better the fitting effect. Through the comprehensive evaluation of these indicators, it can be ensured that the established PLSR and DL-BPNN models have good prediction performance and reliability.

[0067] The calculation formulas of RMSE and R2 are shown in Equations (1) to (2).

[0068]

[0069] Among them, X i is the actual value, Y i is the predicted value, and n is the number of samples.

[0070]

[0071] Among them, X i is the actual value, Y i is the predicted value, is the average value of the actual values.

[0072] 2. Results and Discussion

[0073] In this embodiment, the PLSR model of the full band, the PLSR model of the characteristic band, and the DL-BPNN model are compared. The amount of data in the full band is large and not suitable for the processing of the DL-BPNN model. Therefore, the DL-BPNN model is not considered in the full-band modeling. Finally, the best prediction model for the moisture content of rice paper is selected by comparing these three models.

[0074] 2.1 PLSR Prediction Model of the Full Band

[0075] Use the 6 preprocessing methods in 1.3.2 to establish the PLSR prediction model of the full band. Different preprocessing methods have a significant impact on the prediction performance of the model. By comparing these preprocessing methods, the most suitable preprocessing method is found to improve the prediction accuracy of the model. Table 2 lists the results of the moisture content PLSR model of the near-infrared spectral data under different preprocessing methods.

[0076] Table 2

[0077]

[0078] It can be seen from Table 2 that different preprocessing methods have a significant impact on the prediction ability of the full-band PLSR model. When the original spectral data is not preprocessed, and is relatively high, while RMSEC and RMSEP are relatively low, indicating high data quality and good prediction results.

[0079] Since there was a large baseline drift in the original spectrum, SNV, BC, and Norm preprocessing were used to remove the baseline drift. The results showed that SNV preprocessing significantly improved the prediction performance of the model. and values were both relatively high, and RMSEC and RMSEP values were relatively low, indicating that SNV effectively removed the drift and scattering phenomena in the spectrum. After BC preprocessing, the and values decreased, RMSEC and RMSEP values were relatively high, and the prediction performance was medium. It may be that the baseline correction had a limited impact on the spectral data, but it could remove some low-frequency noise. After Norm preprocessing, the and values were relatively high, RMSEC and RMSEP values were relatively low, and the prediction effect was also good. It shows that Norm processing brought the data to the same scale, which was helpful for the model stability and prediction accuracy.

[0080] After the three preprocessing methods, noise was still found. Therefore, WT was introduced in combination with the three preprocessing methods to reduce noise. However, the performance of the model after the hybrid preprocessing decreased, probably because WT also removed some useful information when removing noise, resulting in a decrease in the model performance.

[0081] In summary, different preprocessing methods had a significant impact on the prediction effect of the PLSR model. Among them, the SNV preprocessing method performed the best. The comparison of the actual values and predicted values of its training set and validation set is Figure 5 shown as follows, verifying its effectiveness in removing baseline drift, scattering effects, and improving the prediction accuracy of the model.

[0082] 2.2 Establishing the PLSR model for characteristic bands

[0083] The SPA and CARS methods for feature extraction were used to extract the characteristic wavelengths from the full band, and then the PLSR prediction model was established. The prediction results are shown in Table 3 (PLSR results based on the SPA characteristic wavelength screening method) and Table 4 (PLSR results based on the CARS characteristic wavelength screening method) below.

[0084] Table 3

[0085]

[0086] Table 4

[0087]

[0088]

[0089] By comparing the results in Table 3, Table 4 and Table 2 respectively, it can be seen that after the characteristic band extraction by CARS, the prediction accuracy of the model has been significantly improved. This shows that feature extraction can effectively extract useful bands and reduce the complexity of data. In particular, the results of Norm-SPA-PLSR and unpreprocessed-CARS-PLSR are relatively prominent, and the comparison graphs of the actual values and predicted values of their training sets and validation sets are as shown in Figure 6 Figure (a) in Figure 6 Figure (b) in. In addition, from the perspective of single preprocessing and mixed preprocessing, the prediction ability of the mixed preprocessing model under the two feature extraction methods is significantly lower than that of the single preprocessing. It may be that the mixed preprocessing removes too much useful information, which is also reflected in Table 2. Therefore, in the extraction of characteristic bands, the accuracy is also not as good as that of single preprocessing.

[0090] By comparing Table 3 and Table 4, it can be seen that whether it is unpreprocessed or preprocessed, the prediction ability of the model with SPA feature extraction is slightly lower than that of the model with CARS feature extraction. The possible reasons are as follows: ① The number of characteristic wavelengths is different. The number of wavelength points extracted by the SPA method is less than that of CARS. SPA may have lost some useful information. ② The algorithm mechanisms are different. SPA reduces multicollinearity through stepwise projection, but may not be able to capture all relevant information. While CARS retains more important wavelengths related to the target variable through competitive weighted sampling, improving the model prediction accuracy. ③ Information redundancy. The CARS method effectively reduces redundant information through competitive weighted sampling and retains more important bands related to the target variable.

[0091] In summary, although both SPA and CARS are effective feature extraction methods, CARS shows better prediction ability when dealing with spectral data. This is mainly because CARS can extract more important wavelengths related to the target variable and effectively reduce information redundancy, thus being superior to the SPA method in feature extraction and model prediction.

[0092] 2.3 Establishing a DL-BPNN Model for Characteristic Bands

[0093] Since the characteristic bands have higher effectiveness and lower complexity compared to the full bands, we first established a PLSR model on the characteristic bands. To explore whether capturing the complex nonlinear relationships in spectral data can further improve the prediction accuracy, we also established a DL-BPNN model. Specifically, under the two characteristic wavelength extraction methods of SPA and CARS, DL-BPNN prediction models for characteristic bands were established respectively, and the prediction results are shown in Table 5 (DL-BPNN results based on the SPA characteristic wavelength screening method) and Table 6 (DL-BPNN results based on the CARS characteristic wavelength screening method).

[0094] Table 5

[0095]

[0096] Table 6

[0097]

[0098]

[0099] From Tables 5 and 6, the model prediction ability under the CARS feature extraction method is slightly better than that under the SPA feature extraction method. Comparing the PLSR prediction models in Tables 3 and 4, there is no significant change in the SPA-DL-BPNN model compared to the SPA-PLSR model, while the result of the CARS-DL-BPNN model is slightly lower than that of the CARS-PLSR model. The best prediction models in Tables 5 and 6 are Unpreprocessed-SPA-DL-BPNN and WT+Norm-CARS-DL-BPNN, and the comparison graphs of the actual values and predicted values of their training sets and validation sets are as shown in Figure 7 Figure (a) and Figure 7 Figure (b).

[0100] These results indicate that although the DL-BPNN model can theoretically capture more complex nonlinear relationships, its prediction accuracy is not significantly higher than that of the PLSR model. This may be due to the following reasons: ① Model complexity: As a linear model, PLSR has a lower complexity and is easier to train and obtain stable results in the case of small samples. Relatively speaking, DL-BPNN is a nonlinear model. Although it has strong fitting ability, it is prone to overfitting when the sample size is small. ② Parameter tuning: The PLSR model has fewer parameters, and the tuning is relatively simple, making it easy to obtain good prediction performance. However, the DL-BPNN model has more parameters, and the tuning process is complex. If the tuning is improper, it may lead to a decline in model performance. ③ Data characteristics: For spectral data, when the linear relationship dominates, the PLSR model can better capture these relationships, thereby providing higher prediction accuracy. Although the DL-BPNN model has an advantage in capturing nonlinear relationships, if the nonlinear characteristics in the data are not significant, the advantage of the model is difficult to reflect.

[0101] In summary, among the models established in the full band and feature bands respectively, the best prediction combinations are SNV-PLSR, Norm-SPA-PLSR, Unpreprocessed-CARS-PLSR, Unpreprocessed-SPA-DL-BPNN, and WT+Norm-CARS-DL-BPNN. Figure 8 shows the prediction accuracy of these 5 combinations on the validation set, using RMSE and R 2 for comparison.

[0102] 3. Conclusion

[0103] In this embodiment, near-infrared spectroscopy analysis was carried out on the four-foot single Xuan paper of cotton material in different humidity environments, multiple moisture content prediction models were established, and their prediction capabilities were evaluated. By using the SPXY division method, the samples were divided into 168 training sets and 42 validation sets according to a ratio of 4:1. Subsequently, the near-infrared spectral data was preprocessed, and the characteristic wavelengths were extracted from the original wavelengths by the SPA and CARS methods. Finally, a full-band PLSR prediction model, as well as PLSR and DL-BPNN prediction models based on characteristic bands, were established.

[0104] Evaluated by comparing different preprocessing methods and feature extraction methods, the results show that the unpreprocessed-CARS-PLSR model established after CARS feature extraction performs best, and the determination coefficient of its validation set is 0.9438, and the root mean square error RMSEP is 0.5707, which indicates that the CARS feature extraction method has significant advantages in retaining important features and removing redundant information. In contrast, although the DL-BPNN model can theoretically capture more complex non-linear relationships, in this embodiment, its prediction accuracy is not significantly higher than that of the PLSR model.

[0105] Generally speaking, in this embodiment, the relationship between the moisture content of Xuan paper and near-infrared spectroscopy was successfully established, the feasibility of non-destructively detecting the moisture content of Xuan paper by near-infrared spectroscopy was verified, the online non-destructive measurement characteristics of NIRS technology were utilized to measure the moisture content of Xuan paper, effectively realizing the rapid and non-destructive measurement of the moisture content of the Xuan paper cultural relic itself, providing a reliable detection technical means for measuring the moisture content of Xuan paper cultural relics, so as to achieve the preventive protection of paper cultural relics.

[0106] Finally, it should be specially noted that when the object to be measured is Xuan paper, the optimal detection model is the unpreprocessed-CARS-PLSR model (R2 = 0.9438, RMSEP = 0.5707), and this phenomenon is contrary to the conventional cognition (preprocessing can improve the model performance). The main reasons include:

[0107] First, CARS retains the wavelengths highly correlated with the target variable (moisture content) through adaptive weighting and iterative screening, while removing noise and redundant bands; this process itself has the functions of noise suppression, feature enhancement and data dimensionality reduction. If the preprocessing (such as SNV, normalization) overcorrects the spectral morphology, it may destroy the non-linear associations or weak but important characteristic peaks related to the moisture content in the original data;

[0108] Second, the change in the moisture content of Xuan paper directly leads to the change in the vibration mode of fiber hydroxyl groups (-OH), and this information may exist in the original spectrum in a specific proportional relationship. Pretreatment (such as normalization) may destroy the relative weights of absorbances at different wavelengths, resulting in the PLSR model being unable to accurately capture the moisture-sensitive bands;

[0109] Third, in the case of Xuan paper, there may be an antagonistic effect between pretreatment and CARS. The rapid moisture absorption characteristics of the thin-layer fibers of Xuan paper result in a small spectral dynamic range, and SNV correction may compress the effective differences; the reflection characteristics of the fillers in Xuan paper (such as calcium carbonate) may be misinterpreted as baseline drift and eliminated;

[0110] Fourth, the aforementioned experiments were completed in a closed environmental chamber (with strict temperature and humidity control), and the spectral noise level was low; the feature selection of CARS has been sufficient to suppress the residual noise, and additional pretreatment may instead introduce unnecessary adjustments;

[0111] Fifth, the moisture content distribution and spectral characteristics of the training set samples are linearly separable, and PLSR can be modeled without pretreatment; pretreatment may destroy the linear boundary (such as normalization changing the distance in the feature space).

[0112] Finally, it should be noted that specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the core idea of the present invention. Without departing from the principles of the present invention, several improvements and modifications can still be made to the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A nondestructive detection method for moisture content of rice paper based on near infrared reflectance spectroscopy, characterized in that: The method comprises obtaining near infrared spectrum data of the Xuan paper to be tested, inputting the near infrared spectrum data into a detection model, and then outputting a moisture content value of the Xuan paper to be tested; Wherein, constructing the detection model comprises the following steps: a. Sample preparation: Generate a set of rice paper samples with multiple gradient moisture contents by humidity control; b. Spectrum acquisition: obtaining the reflectance spectrum of the rice paper sample in a set wavelength range; c. Characteristic wavelength screening: Use competitive adaptive re-weighting algorithm to extract characteristic wavelengths from raw spectral data; d. Model training: A partial least squares regression prediction model is established based on the screened characteristic wavelengths as a detection model.

2. The method for nondestructive detection of moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 1, characterized in that: In step a, the rice paper samples are placed in a closed environmental chamber, the humidity of the environmental chamber is set to several humidity environments ranging from 37% to 97% RH, and the same number of samples are placed in each environmental chamber. The samples are left to stand in the closed environmental chamber for 5 to 9 days.

3. The nondestructive detection method for moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 2, characterized in that: The actual moisture content of the sample was determined by the drying method.

4. The nondestructive detection method for moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 1, characterized in that: In step b, the original near-infrared spectrum of the rice paper sample in the 900-1700 nm band is collected.

5. The method for nondestructive detection of moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 4, characterized in that: In step b, sample collection is carried out in a closed dark box. After the sample to be tested is taken out of the closed environmental box, the spectral data collection is completed within 50 seconds. Each sample to be tested is collected M times continuously. Then, the average value of the spectral data collected M times is taken as the spectral data of the sample.

6. The nondestructive detection method for moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 4, characterized in that: In step c, the original spectral data is spectral data obtained without any preprocessing operation.

7. The nondestructive detection method for moisture content of rice paper based on near infrared reflectance spectroscopy according to claim 1, characterized in that: In step c, 50±5 Monte Carlo samplings are performed to dynamically select 55 to 61 characteristic wavelengths.

Images