Rhodiola rosea quality identification method, terminal equipment and storage medium

Abstract

The invention discloses a Rhodiola rosea quality identification method, a terminal device and a storage medium, the identification method is a Rhodiola rosea quality rapid identification method based on the combination of a three-dimensional fluorescence technology and deep learning, the method performs feature learning on a convolution auto-encoder by establishing a training set, and in model training, the Rhodiola rosea quality is rapidly identified. According to the method, error threshold reconstruction is combined with a dual discrimination mechanism of a local abnormal factor model, misjudgment caused by single judgment is avoided, the identification accuracy is improved, the good and bad identification accuracy of a training set and a test set is 97% or above, and the specificity of abnormal sample identification is 100%.

Description

Technical Field

[0001] This invention relates to the field of Rhodiola rosea quality identification technology using deep learning, and more specifically, to a Rhodiola rosea quality identification method, terminal device, and storage medium. Background Technology

[0002] Rhodiola rosea is a herbaceous plant belonging to the Crassulaceae family. Its medicinal value mainly stems from its various active ingredients, including rhodioloside, tyrosol, phenolic acids, and flavonoids. These components play important roles in anti-inflammation, anti-tumor, anti-fatigue, and immune regulation. While Rhodiola rosea resources are abundant and diverse, there are significant differences in chemical composition and medicinal value among different varieties. The Pharmacopoeia of the People's Republic of China (2025 edition) uses rhodioloside content as the sole standard for evaluating its quality, stipulating that its content must not be less than 0.5%. However, due to strong market demand and high prices for Rhodiola rosea, the market is rife with the practice of counterfeiting or adulterating it with other Rhodiola rosea varieties (such as Rhodiola angustifolia and Rhodiola rosea var. rosae). These varieties generally have rhodioloside content lower than the pharmacopoeia standard, seriously infringing on consumer rights and disrupting market order. Therefore, it is necessary to establish a rapid and accurate method for identifying the quality of Rhodiola rosea to combat such fraudulent activities.

[0003] Currently, the identification and quality control of Rhodiola rosea mainly rely on techniques such as chromatography and mass spectrometry. While these methods can achieve precise quantification of components and identification of varieties, they typically require complex sample pretreatment processes, rely on expensive large-scale instruments and professional operators, resulting in high analysis costs and long cycles, making it difficult to meet the needs of rapid on-site screening in market supervision. As a simple, low-cost, and highly sensitive analytical technique, three-dimensional fluorescence spectroscopy is an extension of traditional fluorescence spectroscopy under a single excitation wavelength. It acquires excitation-emission matrix fluorescence spectra (three-dimensional fluorescence spectroscopy) by collecting emission spectra at different excitation wavelengths with a constant step size. It offers diverse acquisition methods, provides richer spectral fingerprint information, significantly improves the qualitative and quantitative analysis capabilities of samples, and shows great potential in drug analysis. The active components in Rhodiola rosea, such as rhodioloside and tyrosol, possess endogenous fluorescence properties, providing a theoretical basis for using three-dimensional fluorescence spectroscopy for quality analysis. However, traditional fluorescence analysis methods struggle to fully extract subtle feature information from complex three-dimensional fluorescence data, and research combining three-dimensional fluorescence spectroscopy with deep learning algorithms for the quality identification of Rhodiola rosea is still in its infancy. Therefore, developing a new method that combines the rapid analysis advantages of three-dimensional fluorescence spectroscopy with the powerful feature extraction capabilities of deep learning algorithms is of great practical significance for achieving rapid and accurate identification of Rhodiola rosea quality and regulating market order.

[0004] Publication No. CN120741405A discloses an artificial intelligence-based method for detecting rhodioloside content in Rhodiola rosea extract, comprising the following steps: S1, performing spectral scanning of the extract sample in different bands within multiple near-infrared wavelength ranges; S2, performing multi-dimensional interference suppression processing and signal purification; S3, obtaining the rhodioloside characteristic intensity factor through spectral feature compression calculation; S4, performing rhodioloside characteristic signal-to-noise ratio enhancement and purity index calculation; S5, predicting the rhodioloside content and outputting the content prediction score; S6, mapping the content prediction score to the final rhodioloside content. This scheme uses "near-infrared wavelength range band scanning," which belongs to two-dimensional spectroscopy. Two-dimensional spectroscopy is difficult to distinguish subtle feature differences, which can lead to inaccurate feature factor extraction and ultimately result in misjudgment. Summary of the Invention

[0005] To address the technical shortcomings of existing Rhodiola rosea quality testing methods, such as complex pretreatment, long testing cycle, high cost, and difficulty in achieving rapid screening, this invention provides a rapid Rhodiola rosea quality identification method, terminal equipment, and storage medium.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for quality identification of Rhodiola rosea, characterized by the following steps: S1. Pretreatment of Rhodiola rosea samples: Obtain Rhodiola rosea powder samples from different brands, prepare Rhodiola rosea sample filtrate and determine the content of rhodioloside in the filtrate. Samples with rhodioloside content ≥0.5% are marked as normal samples, and samples <0.5% are marked as abnormal samples. They are then randomly divided into training set and test set. S2. Obtain three-dimensional fluorescence spectral data of training and test sets: Use a fluorescence spectrophotometer to scan the filtrate of all Rhodiola rosea samples to obtain the three-dimensional fluorescence spectral data of each sample, and preprocess the three-dimensional fluorescence spectral data. S3. Train a convolutional autoencoder using the training set to obtain a Rhodiola rosea quality discrimination model: The convolutional autoencoder adopts a symmetrical encoder-decoder structure. The encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and two fully connected layers connected in sequence after feature flattening. The fully connected layer at the end of the encoder outputs a 32-dimensional feature vector, which constitutes the 32-dimensional hidden layer feature space of the model. This space contains the essential feature information of the Rhodiola rosea spectrum. The decoder is symmetrical to the encoder. The two fully connected layers restore the 32-dimensional feature vector to the feature map size, and perform two upsampling and deconvolution operations in sequence. The last layer outputs the reconstructed spectral data. The model is trained using normal samples and its network structure and hyperparameters are adjusted for abnormal samples. The trained model judges the quality of the input unknown sample spectral data based on the reconstruction error and the hidden layer feature values.

[0007] Furthermore, in step S1, the content of rhodioloside in the Rhodiola rosea sample filtrate was determined by high performance liquid chromatography.

[0008] Furthermore, the preparation process of the Rhodiola rosea sample filtrate in step S1 includes the following steps: the Rhodiola rosea sample is crushed and passed through a 40-mesh sieve to obtain Rhodiola rosea powder. 200.00 mg of Rhodiola rosea powder is weighed and dissolved in 10.00 mL of methanol. The mixture is ultrasonically extracted at room temperature for 30 min. Then, the extract is filtered through a nylon filter membrane to obtain the Rhodiola rosea sample filtrate.

[0009] Further, the specific implementation process in step S2 includes: diluting with methanol to a suitable factor and performing fluorescence scanning to obtain three-dimensional fluorescence spectral data; the excitation wavelength range of the fluorescence spectrophotometer is 200~340 nm, the emission wavelength range is 260~470 nm, the wavelength interval is 2 nm, the slit width is 5 nm, the voltage is 700 V, and the scanning speed is 30000 nm·min. -1 .

[0010] Furthermore, the preprocessing of the three-dimensional fluorescence spectral data in step S2 includes: subtracting slight Raman scattering using blank, fitting the subtracted scattering using interpolation, and then normalizing the intensity of the obtained data so that the intensity range and data size of the three-dimensional fluorescence spectrum are 0-1 and 80×120, respectively.

[0011] Further, step S3 includes the following steps: inputting all normal samples in the training set into the convolutional autoencoder for modeling, and optimizing the number of iterations of the convolutional autoencoder through the training set. In model training, a dual discrimination mechanism is adopted, setting the upper edge of the box plot as the reconstruction error threshold for normal and abnormal samples. At the same time, in the 32-dimensional hidden layer feature space of the encoder, the local anomaly factor algorithm is used, setting the contamination rate to 0.01, to establish a density model based on the hidden layer feature values ​​of normal samples. This continues until a box plot that contains as much of the reconstruction error of all normal samples as possible and a local anomaly factor model that contains the hidden layer feature values ​​of all normal samples are obtained, thus obtaining the Rhodiola rosea quality identification model.

[0012] Furthermore, after step S3, the method further includes: S4, evaluating the classification performance of the Rhodiola rosea quality discrimination model using the training set and test set.

[0013] Further, the specific steps of step S4 include: inputting the test set into the trained Rhodiola rosea quality identification model, obtaining the model's prediction results, comparing the prediction results with the labeling results in step S1, and calculating the model's correct classification rate and specificity; demonstrating the model's classification performance by reconstructing the error distribution histogram and the hidden layer feature t-SNE visualization; designing ablation experiments to construct simplified models that rely only on reconstruction errors and those that rely only on hidden layer features, and comparing the classification performance of the complete model and the simplified model.

[0014] As an inventive concept, the present invention also provides a terminal device, comprising: a memory, a processor, and a display module; the memory stores a computer program for implementing the above method; the processor runs the computer program stored in the memory; and the display module is used to display the output results after execution to the user.

[0015] As an inventive concept, the present invention also provides a computer storage medium storing a computer program / instructions, the computer program / instructions being executed by a processor using the steps of the above-described Rhodiola rosea quality identification method.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The proposed Rhodiola rosea quality identification method is a rapid method for identifying Rhodiola rosea quality based on three-dimensional fluorescence technology combined with deep learning. This method learns features from a convolutional autoencoder using a training set. During model training, a dual discrimination mechanism combining a reconstruction error threshold and a local anomaly factor model is used to avoid misjudgments caused by single judgment, thus improving the accuracy of identification. The accuracy of quality identification on both the training and test sets is above 97%, and the specificity for identifying abnormal samples is 100%. At the same time, it has a low operational threshold, requiring only basic experimental operations to complete sample preparation and spectral acquisition, enabling rapid identification and improving identification efficiency. Attached Figure Description

[0017] Figure 1 A flowchart for identifying the quality of Rhodiola rosea using three-dimensional fluorescence spectroscopy combined with deep learning; Figures 2(a) to 2(c) show the high performance liquid chromatograms of rhodioloside standard and rhodiola sample (detection wavelength 275 nm), where Figure 2(a) is rhodioloside standard (60 µg / mL), Figure 2(b) is normal rhodiola sample (10 mg / mL), and Figure 2(c) is abnormal rhodiola sample (10 mg / mL). Figures 3(a) to 3(c) are schematic diagrams of the fitting effect of Rhodiola rosea samples processed by interpolation. Figure 3(a) is the original data result diagram, Figure 3(b) is the result diagram after removing scattering, and Figure 3(c) is the result diagram after processing by interpolation. Figures 4(a) to 4(d) show the three-dimensional fluorescence spectra of Rhodiola rosea of ​​different qualities after interpolation. Figure 4(a) is the spectrum of normal Rhodiola rosea, Figure 4(b) is the spectrum of abnormal Rhodiola rosea, Figure 4(c) is the spectrum of rhodioloside, and Figure 4(d) is the spectrum of gallic acid. Figure 5 The specific structure of the CAE in the Rhodiola rosea quality identification model; Figure 6 The specific process of the quality identification model algorithm for Rhodiola rosea; Figure 7(a) is a histogram of the reconstruction error distribution, and Figure 7(b) is a visualization of tSNE. Figures 8(a) to 8(d) show the original and reconstructed spectra of Rhodiola rosea. Figure 8(a) shows the original spectrum of normal Rhodiola rosea, Figure 8(b) shows the original spectrum of abnormal Rhodiola rosea, Figure 8(c) shows the reconstructed spectrum of normal Rhodiola rosea, and Figure 8(d) shows the reconstructed spectrum of abnormal Rhodiola rosea. Detailed Implementation To clearly illustrate the technical features of the present invention, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below. In the present invention, unless otherwise expressly specified and limited, the first feature "on" or "below" the second feature may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0018] Example 1 Please see Figure 1 As shown in Figure 8, this embodiment provides a method for identifying the quality of Rhodiola rosea, including the following steps: S1. Pretreatment of Rhodiola rosea samples: Obtain Rhodiola rosea powder samples from different brands, prepare Rhodiola rosea sample filtrate and determine the content of rhodioloside in the filtrate. Samples with rhodioloside content ≥0.5% are marked as normal samples, and samples <0.5% are marked as abnormal samples. They are then randomly divided into training set and test set. Rhodiola rosea samples from different brands were collected, and 120 independent samples were prepared. After pulverizing and passing through a 40-mesh sieve, 200.00 mg of Rhodiola rosea powder was accurately weighed and dissolved in 10.00 mL of methanol. The mixture was ultrasonically extracted at room temperature for 30 min. The extract was filtered through a 0.45 μm nylon filter to obtain the Rhodiola rosea sample filtrate, which was stored in a refrigerator at 4℃ protected from light for later use. The content of rhodioloside in the filtrate was determined by high performance liquid chromatography (HPLC). The chromatographic conditions were as follows: reversed-phase WondaSil™ C18 column (250 mm × 4.6 mm, 5 µm), column temperature 30℃, mobile phase ultrapure water:methanol = 85:15 (v / v), flow rate 1.0 mL / min, detection wavelength 275 nm, and injection volume 20 µL.

[0019] A standard curve was established using rhodioloside standards, and the rhodioloside content in the samples was calculated. According to the Chinese Pharmacopoeia (2025 edition) standard (with a rhodioloside content ≥0.5% considered normal), 120 samples were labeled as "normal samples" or "abnormal samples," and randomly divided into training and test sets at a ratio of 3:1.

[0020] S2. Obtain three-dimensional fluorescence spectral data of training and test sets: Use a fluorescence spectrophotometer to scan the filtrate of all Rhodiola rosea samples to obtain the three-dimensional fluorescence spectral data of each sample, and preprocess the three-dimensional fluorescence spectral data. Three-dimensional fluorescence spectra of the samples were acquired using a Hitachi F-7000 fluorescence spectrophotometer. Parameters were set as follows: excitation wavelength range 200-340 nm, emission wavelength range 260-470 nm, step size 2 nm, excitation and emission slit width 5 nm, scan speed 30000 nm / min, voltage 700 V. All Rhodiola rosea sample filtrates were diluted with methanol to an appropriate factor to ensure the sample fluorescence intensity was within the instrument's optimal detection range. Fluorescence scanning was performed in a 1 cm quartz cuvette to obtain three-dimensional fluorescence spectra, with the detection repeated three times.

[0021] Interpolation was used to remove Rayleigh and Raman scattering interference. The intensity of the spectral data after scattering was normalized to the range of 0-1 and uniformly scaled to a matrix size of 80×120 pixels as the training input for the model.

[0022] S3. Train a convolutional autoencoder using the training set to obtain a Rhodiola rosea quality discrimination model: The convolutional autoencoder adopts a symmetrical encoder-decoder structure. The encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and two fully connected layers connected in sequence after feature flattening. The fully connected layer at the end of the encoder outputs a 32-dimensional feature vector, which constitutes the 32-dimensional hidden layer feature space of the model. This space contains the essential feature information of the Rhodiola rosea spectrum. The decoder is symmetrical to the encoder. The two fully connected layers restore the 32-dimensional feature vector to the feature map size, and perform two upsampling and deconvolution operations in sequence. The last layer uses the Tanh activation function to output the reconstructed 80×120 spectral data. The model is trained using normal samples and its network structure and hyperparameters are adjusted for abnormal samples. The trained model can jointly judge the quality of the input unknown sample spectral data based on the reconstruction error and the hidden layer feature values.

[0023] A boxplot is a statistical chart consisting of a box and two edges. The top and bottom edges of the box represent the upper quartile (Q3) and lower quartile (Q1) of a data set, respectively. The middle line of the box represents the median (Xm), and the upper and lower edges represent the values ​​of Q3 + 1.5(Q3 - Q1) and Q1 - 1.5(Q3 - Q1), respectively. A boxplot can reflect the distribution of one or more data sets. Five characteristic values ​​can be used to visually identify outliers in the dataset and their degree of anomalousness. Data points above the upper edge or below the lower edge of the boxplot are considered outliers. Because boxplots are unaffected by outliers, they are widely used in outlier detection to describe the dispersion of data and can effectively explain some asymmetric data types and extreme outlier samples.

[0024] The Local Outlier Factor (LOF) algorithm is a density-based outlier detection method implemented using the `sklearn.neighbors.LocalOutlierFactor` function in Python. It quantifies the anomaly level of each data point by calculating its LOF value, which represents the ratio of the density of each data point to the density of its neighboring data points. The LOF algorithm considers both local and global attributes of the dataset, effectively addressing the challenge of outlier detection in datasets with uneven data density distribution. A higher LOF value than 1 indicates a lower density than surrounding data points, making an anomaly more likely. The LOF calculation formula is as follows:

[0025] in LOFk(o) represent o Local anomalous factors, Nk(o) represent o Pointed k-Neighbor ,lrdk(p) and lrdk(o) Represent o Points and o Within the domain of a point p Local reachability density of points; all normal Rhodiola rosea samples in the training set are input into a constructed convolutional autoencoder for modeling, and the number of iterations of the convolutional autoencoder is optimized using the training set. During model training, a dual discrimination mechanism is adopted, setting the upper edge of the Boxplot as the reconstruction error threshold for normal and abnormal samples. Histogram analysis of the reconstruction error distribution shows that the reconstruction error of normal samples is concentrated at 3.16 × 10⁻⁶. -5 ~9.81×10 -5 The low-value range exhibits a compact normal distribution; while the reconstruction error of outlier samples is generally greater than 9.81 × 10⁻⁶. -5 Based on this distribution characteristic, a value of 9.81 × 10⁻⁶ is set. -5 The upper edge threshold of the boxplot is set, and this optimized value can effectively distinguish between the two types of samples. Simultaneously, in the 32-dimensional hidden layer feature space, the LOF algorithm is used with a contamination rate of 0.01 to establish a density model based on the hidden layer feature values ​​of normal samples. The hidden layer feature space refers to the space formed by the 32-dimensional feature vectors output by the second fully connected layer (FC2) at the end of the encoder. This process continues until a boxplot that includes as much of the reconstruction error as possible from normal samples and a local anomaly factor model that includes the hidden layer feature values ​​of all normal samples are obtained, thus yielding the trained Rhodiola rosea quality discrimination model.

[0026] Experiments have shown that when the upper edge of the box plot and the hyperplane of the LOF model contain as many normal samples as possible from the data augmentation training set, the correct classification rate of both the training and test sets reaches over 97%, and the specificity reaches 100%. This fully demonstrates that the rapid identification method for Rhodiola rosea quality based on three-dimensional fluorescence technology combined with deep learning proposed in this embodiment can quickly and reliably identify the quality of Rhodiola rosea samples.

[0027] S4. Evaluate the classification performance of the Rhodiola Rosea quality identification model using the training and test sets.

[0028] The test set is input into the trained Rhodiola rosea quality identification model to obtain the model's prediction results. The prediction results are compared with the labeling results in step S1 to calculate the model's correct classification rate and specificity. The classification performance of the model is demonstrated by reconstructing the error distribution histogram and the hidden layer feature t-SNE visualization. An ablation experiment is designed to construct simplified models that rely only on reconstruction error and those that rely only on hidden layer features, and the classification performance of the complete model and the simplified model is compared.

[0029] The following is a detailed description of the process by which the quality identification model of Rhodiola rosea, based on three-dimensional fluorescence spectroscopy combined with a single-class convolutional autoencoder algorithm, established in Embodiment 1 of the present invention, can quickly and reliably identify the quality of Rhodiola rosea.

[0030] 1. Experimental Instruments and Materials Instruments: fluorescence spectrophotometer; liquid chromatograph; quartz dishes; data analysis and the programs used were run in MATLAB 2009b and PyTorch environments.

[0031] Materials: The Rhodiola rosea samples used in the experiment were from 17 different brands of Rhodiola rosea products. Depending on the weight of each package, each package was divided into 5 or 10 portions, resulting in a total of 120 Rhodiola rosea samples. All the Rhodiola rosea samples were pulverized in a grinder and passed through a 40-mesh sieve to obtain Rhodiola rosea powder. 200.00 mg of Rhodiola rosea powder was accurately weighed and dissolved in 10.00 mL of methanol. The mixture was ultrasonically extracted at room temperature for 30 min. The extract was then filtered through a 0.45 µm nylon filter membrane, and the filtrate was stored in a refrigerator at 4°C in the dark for later use.

[0032] Rhodioloside standards were purchased from Solarbio (Beijing, China), and gallic acid standards were purchased from Yuanye Biotechnology Co., Ltd. (Shanghai, China). These standard solutions were prepared for three-dimensional fluorescence scanning to serve as a reference for assigning characteristic peaks and to explain the source of characteristic signals in the intrinsic fluorescence spectrum of Rhodiola rosea samples. The purity of both standards was greater than 98%. Chromatographic grade methanol was purchased from Oceanpak (Sweden). All ultrapure water used was prepared daily using a Milli-Q ultrapure water purification system. An appropriate amount of rhodioloside standard was weighed and dissolved in the mobile phase (water:methanol = 85:15) to prepare a stock solution (480 µg / mL). The stock solution was then diluted appropriately with the mobile phase to obtain a series of standard solutions. All working solutions were filtered through a 0.45 µm nylon membrane before chromatographic analysis.

[0033] 2. Experimental Methods 2.1 Parameter Settings The liquid chromatograph was set with a reversed-phase WondaSil™ C18 column (size: 250 mm × 4.6 mm, 5 µm), a column temperature of 30 °C, a mobile phase of ultrapure water:methanol = 85:15 (v / v), a flow rate of 1.0 mL / min, a detection wavelength of 275 nm, and an injection volume of 20 µL.

[0034] The excitation wavelength (Ex) range of the fluorescence spectrophotometer was set to 200 nm-340 nm, and the emission wavelength (Em) range was set to 260 nm-470 nm, with a wavelength interval of 5 nm for both wavelengths; the step size was 2 nm for both wavelengths; the excitation and emission slit widths were 5 nm; the scanning speed was 30,000 nm / min; and the scanning voltage was 700 V.

[0035] 2.2 Data Processing 1) Chromatographic data acquisition and authenticity labeling Weigh out rhodioloside standards and prepare a series of standard working solutions of known concentrations using a mobile phase (water:methanol = 85:15). Under the established chromatographic conditions, the standard working solutions were analyzed sequentially, and a typical chromatogram is shown in Figure 2. As shown in Figure 2(a), the rhodioloside standard elutes at approximately 15.9 minutes, with a symmetrical and sharp peak shape, indicating good chromatographic separation. Comparing Figure 2(b) and Figure 2(c), it can be seen that the normal Rhodiola rosea sample ( Figure 2b The samples showed significant rhodioloside chromatographic peaks at the same retention time, with peak areas larger than the threshold corresponding to the pharmacopoeia standard; while abnormal rhodiola samples ( Figure 2c The chromatographic peak response at this location was significantly low. A linear regression was performed on the peak area (y) of rhodioloside against its concentration (x, μg / mL), and a standard curve was plotted to obtain the regression equation and correlation coefficient (R²), ensuring good linearity within the defined concentration range. A typical regression equation was y = 4941.58x - 3775.80, with a linear range covering 7.5 μg / mL to 480 μg / mL, and R² greater than 0.999, indicating that this method meets the accuracy requirements for quantitative analysis.

[0036] All Rhodiola rosea sample solutions were analyzed under the same chromatographic conditions. The chromatographic peaks of rhodioloside were qualitatively identified by comparing retention times, and their peak areas were recorded. The obtained peak area data were substituted into the aforementioned standard curve equation to calculate the concentration of rhodioloside in each test solution. Then, based on factors such as sampling amount, final volume, and dilution factor, the percentage content of rhodioloside in each Rhodiola rosea sample was accurately calculated. Finally, the calculated content results were compared with the standards specified in the Pharmacopoeia of the People's Republic of China (2025 edition) (with a rhodioloside content ≥0.5% as normal), thereby assigning each sample a "normal" (meets the standard) or "abnormal" (below the standard) authenticity label.

[0037] 2) Obtain the three-dimensional fluorescence spectroscopy dataset All Rhodiola rosea sample filtrates were diluted 200-fold with methanol and subjected to fluorescence scanning in a 1 cm quartz dish to obtain three-dimensional fluorescence spectra. This was repeated three times. Each Rhodiola rosea sample, after three-dimensional fluorescence scanning, yielded a two-dimensional matrix of size 71 × 10⁶ (number of excitation wavelengths × number of emission wavelengths). However, the original fluorescence spectral data contained numerous scattering interferences, such as first-order Rayleigh scattering, second-order Rayleigh scattering, and Raman scattering, which disrupt the trilinear structure. In this embodiment, the interpolation method proposed by M. Bahram et al. (Bahram M, Journal of Chemometrics, 2006, 20(3-4): 99-105) was used. The signal in the scattering region was first cleared, and then the data in that region was fitted and completed using an interpolation algorithm based on the surrounding effective fluorescence signals. In this embodiment, -15 to +15, -5 to +5, and -30 to +20 were selected as the widths for first-order Rayleigh scattering, Raman scattering, and second-order Rayleigh scattering, respectively. The specific effects are shown in Figures 3(a) to 3(c). Finally, the intensity of all the three-dimensional fluorescence spectra after scattering was normalized and the size was scaled to make the intensity range and data size of the three-dimensional fluorescence spectra 0~1 and 80×120, respectively, thus obtaining the three-dimensional fluorescence spectrum dataset. The comparison of Figures 4(a)-4(d) clearly shows that Figures 4(a) and 4(b) show the three-dimensional fluorescence spectra of normal and abnormal Rhodiola rosea. As shown in the figure, both have strong fluorescence response values ​​at excitation / emission of 366 / 270 nm, which is caused by the endogenous fluorescent components of Rhodiola rosea. These are components that are common to different types of Rhodiola rosea and have similar contents. According to the literature, it is speculated that this fluorescence peak is the effect of the combined action of gallic acid and other fluorescent components. By comparing the spectrum of gallic acid standard ( Figure 4d The fluorescence peaks at approximately 366 / 270 nm were confirmed to be produced by gallic acid components common to Rhodiola rosea samples, providing direct spectral evidence for the Rhodiola rosea quality identification model to distinguish samples of different quality.

[0038] 2.3 Establishment of a quality identification model for Rhodiola rosea Autoencoders, as a typical deep learning algorithm, learn features from input data through an encoder and a decoder. The encoder maps high-dimensional input data to low-dimensional encoding, achieving feature extraction and compression, while the decoder attempts to reconstruct the original input data from the low-dimensional encoding, achieving data reconstruction. Autoencoders can effectively reduce data dimensionality and extract features, and are widely used in the field of anomaly detection. The convolutional autoencoder (CAE) model built in Embodiment 1 of this invention is an autoencoder-based model, employing a symmetrical encoder-decoder structure. The encoder part includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and two fully connected layers in series. The decoder part adopts a structure symmetrical to the encoder, reconstructing the original-size spectral data from the feature vector through deconvolution and upsampling operations. The model parameters are optimized based on the classification accuracy results of the training set. The convolutional kernel size of the convolutional layers is set to 3×3, the convolutional kernel size of the pooling layers is set to 2×2, the stride of the convolutional and pooling layers is 1 and 2 respectively, and the output dimensions of the two fully connected layers are 120 and 32 respectively. The specific structure is as follows. Figure 5 As shown in Table 1, the specific overview of each part is as follows. The Rhodiola rosea quality identification model determines the upper edge of the boxplot based on the reconstruction error distribution of normal samples. The upper edge is set as the reconstruction error threshold for normal and abnormal samples. Through histogram analysis of the reconstruction error distribution, it is determined that the reconstruction error of normal samples is concentrated at 3.16 × 10⁻⁶. -5 ~9.81×10 -5 The interval is set to 9.81 × 10. -5 To optimize the threshold, a classification model based on the hidden layer feature space is established using the Local Outlier Factor (LOF) algorithm. The hidden layer feature space refers to the space formed by the 32-dimensional feature vectors output from the second fully connected layer (FC2) at the end of the encoder. The LOF algorithm is implemented using the `sklearn.neighbors.LocalOutlierFactor` function in Python, with a contamination rate set to 0.01. A density distribution model of normal samples is constructed in this 32-dimensional feature space. Ultimately, only samples that pass both the reconstruction error test and the hidden feature difference test are considered acceptable in quality. The specific process is as follows: Figure 6 As shown.

[0039] Table 1: Convolutional Autoencoder Structure

[0040] 2.4 Model Evaluation Correct classification rate (CCR) and specificity (SPE) are commonly used to evaluate the classification performance of a model. Correct classification rate is the ratio of correctly predicted samples to the total number of samples, and is often used to evaluate the overall classification performance of the model. Specificity is the ratio of correctly predicted outliers to the total number of outliers, and can be used to measure the model's ability to identify outliers. The closer the correct classification rate and specificity are to 100%, the better the model's classification performance. The formulas for calculating correct classification rate and specificity are as follows: Correct classification rate =

[0041] Specificity =

[0042] Among them, TP represents true positive; TN represents true negative; FN represents false negative; and FP represents false positive.

[0043] 3. Experimental Results By calculating the classification evaluation metrics of the model on the training and test sets, the results show that the Rhodiola rosea quality identification model exhibits excellent classification ability. Table 2 shows the correct classification rate and specificity of each sample in the corresponding datasets. In the training set of 90 samples, the model achieved a correct classification rate of 99.3% and a specificity of 100%, indicating that all abnormal samples were accurately identified. In the test set of 30 samples, the model maintained a correct classification rate of 97.8% and a specificity of 100%, demonstrating the model's good identification ability. Furthermore, the confusion matrices of the training and test sets are shown in Table 3. This table shows that the Rhodiola rosea quality identification model not only has excellent overall classification accuracy but also possesses good abnormal sample detection ability, which is crucial for ensuring the reliability of Rhodiola rosea quality control.

[0044] Table 2: Classification results of the Rhodiola Rosea quality identification model

[0045] Table 3: Confusion matrix of training and test sets obtained by the Rhodiola Rosea quality identification model

[0046] The histogram of reconstruction error distribution and the tSNE visualization are shown in Figure 7. The histogram of reconstruction error distribution clearly shows the significant differences between normal and abnormal samples. The reconstruction error of normal samples is concentrated at 3.16 × 10⁻⁶. -5 ~9.81×10 -5 The low-value range exhibits a compact normal distribution; while the reconstruction error of outlier samples is generally greater than 9.81 × 10⁻⁶. -5 The distribution range is wide and significantly deviates from the normal sample area. The discrimination threshold determined by box plot analysis is 9.81 × 10⁻⁶. -5The threshold effectively distinguishes between the two types of samples, verifying the effectiveness of the optimized threshold.

[0047] The t-SNE dimensionality reduction visualization results provide intuitive evidence for the model's discrimination mechanism. In the two-dimensional feature space, normal samples are tightly clustered to form a clear cluster structure, while abnormal samples are significantly deviated from the main cluster and distributed in the peripheral regions of the feature space. This clear separation phenomenon proves that the Rhodiola rosea quality discrimination model can effectively learn the essential characteristics of the Rhodiola rosea spectrum and map samples of different quality to different regions of the feature space.

[0048] To further verify the reconstruction capability of the Rhodiola rosea quality identification model, we compared and analyzed the original spectra and reconstruction results of normal and abnormal Rhodiola rosea samples, as shown in Figure 8. A comparison of Figure 8(a) and Figure 8(c) clearly shows that the original spectrum of the normal Rhodiola rosea sample can be reconstructed with high accuracy after processing by the Rhodiola rosea quality identification model. The reconstructed spectrum maintains a high degree of consistency with the original spectrum in terms of characteristic peak positions, contours, and intensities. This indicates that the model has fully learned and mastered the spectral characteristic patterns of the normal Rhodiola rosea sample. Conversely, a comparison of Figure 8(b) and Figure 8(d) reveals a significant difference between the reconstructed spectrum and the original spectrum of the abnormal Rhodiola rosea sample, and the model cannot accurately reconstruct its spectral characteristics. This stark contrast in reconstruction results provides an intuitive experimental basis for the reconstruction error discrimination criterion described in this invention.

[0049] Ablation experiments were designed to verify the necessity of the dual-judgment mechanism. When only reconstruction error (box plot) was used for judgment, normal samples in the test set were misclassified as abnormal. When only hidden layer features (LOF algorithm) were used for judgment, abnormal samples in the test set were misclassified as normal. The complete Rhodiola rosea quality identification model (dual-judgment) effectively avoided the above errors and maintained 100% specificity in all cases. The ablation experiment results are shown in Table 4.

[0050] Table 4: Ablation Experiment Results

[0051] In summary, Embodiment 1 of this invention proposes a method for rapidly identifying the quality of Rhodiola rosea based on three-dimensional fluorescence technology. The Rhodiola rosea quality identification model based on convolutional autoencoder effectively classifies the quality of Rhodiola rosea, with classification accuracy exceeding 97% on both the training and test sets. Furthermore, the specificity for identifying abnormal samples is 100%. This fully demonstrates that the Rhodiola rosea quality identification model based on three-dimensional fluorescence technology combined with a single-class convolutional autoencoder proposed in this invention can rapidly and reliably identify the quality of Rhodiola rosea samples, providing a new approach for the quality monitoring of traditional Chinese medicinal materials.

[0052] Example 2 Embodiment 2 of the present invention provides a terminal device corresponding to Embodiment 1 above. The terminal device can be a processing device for a client, such as a mobile phone, a laptop, a tablet computer, a desktop computer, etc., to execute the method of the above embodiments.

[0053] The terminal device in this embodiment includes a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in Embodiment 1 described above.

[0054] In some implementations, the memory may be high-speed random access memory (RAM), and may also include non-volatile memory, such as at least one disk storage device.

[0055] In other implementations, the processor can be any type of general-purpose processor, such as a central processing unit (CPU) or a digital signal processor (DSP), and there is no limitation here.

[0056] Example 3 Embodiment 3 of the present invention provides a computer storage medium corresponding to Embodiment 1 above, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, it implements the steps of the method of Embodiment 1 above.

[0057] Computer storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0058] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for quality identification of Rhodiola rosea, characterized in that, Includes the following steps: S1. Pretreatment of Rhodiola rosea samples: Obtain Rhodiola rosea powder samples from different brands, prepare Rhodiola rosea sample filtrate and determine the content of rhodioloside in the filtrate. Samples with rhodioloside content ≥0.5% are marked as normal samples, and samples <0.5% are marked as abnormal samples. They are then randomly divided into training set and test set. S2. Obtain three-dimensional fluorescence spectral data of training and test sets: Use a fluorescence spectrophotometer to scan the filtrate of all Rhodiola rosea samples to obtain the three-dimensional fluorescence spectral data of each sample, and preprocess the three-dimensional fluorescence spectral data. S3. Train a convolutional autoencoder using the training set to obtain a Rhodiola rosea quality discrimination model: The convolutional autoencoder adopts a symmetrical encoder-decoder structure. The encoder includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and two fully connected layers connected in sequence after feature flattening. The fully connected layer at the end of the encoder outputs a 32-dimensional feature vector, which constitutes the 32-dimensional hidden layer feature space of the model. This space contains the essential feature information of the Rhodiola rosea spectrum. The decoder is symmetrical to the encoder. The two fully connected layers restore the 32-dimensional feature vector to the feature map size, and perform two upsampling and deconvolution operations in sequence. The last layer outputs the reconstructed spectral data. The model is trained using normal samples and its network structure and hyperparameters are adjusted for abnormal samples. The trained model judges the quality of the input unknown sample spectral data based on the reconstruction error and the hidden layer feature values.

2. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, In step S1, the content of rhodioloside in the Rhodiola rosea sample filtrate was determined by high performance liquid chromatography.

3. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, The preparation process of Rhodiola rosea sample filtrate in step S1 includes the following steps: Rhodiola rosea sample is crushed and passed through a 40-mesh sieve to obtain Rhodiola rosea powder. 200.00 mg of Rhodiola rosea powder is weighed and dissolved in 10.00 mL of methanol. The mixture is ultrasonically extracted at room temperature for 30 min. The extract is then filtered through a nylon filter membrane to obtain Rhodiola rosea sample filtrate.

4. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, The specific implementation process in step S2 includes: diluting with methanol 200 times and performing fluorescence scanning to obtain three-dimensional fluorescence spectrum data; the excitation wavelength range of the fluorescence spectrophotometer is 200~340 nm, the emission wavelength range is 260~470 nm, the wavelength interval is 2 nm, the slit width is 5 nm, the voltage is 700 V, and the scanning speed is 30000 nm / min.

5. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, The preprocessing of the three-dimensional fluorescence spectral data in step S2 includes: subtracting slight Raman scattering using blank, fitting the subtracted scattering using interpolation, and then normalizing the intensity of the obtained data so that the intensity range and data size of the three-dimensional fluorescence spectrum are 0~1 and 80×120, respectively.

6. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, Step S3 includes the following steps: inputting all normal samples from the training set into a convolutional autoencoder for modeling, optimizing the number of iterations of the convolutional autoencoder using the training set, employing a dual discrimination mechanism during model training, setting the upper edge of the box plot as the reconstruction error threshold for normal and abnormal samples; simultaneously, in the 32-dimensional hidden layer feature space of the encoder, using a local anomaly factor algorithm with a contamination rate of 0.01, establishing a density model based on the hidden layer feature values ​​of normal samples; until a box plot that contains as much of the reconstruction error of all normal samples as possible and a local anomaly factor model that contains the hidden layer feature values ​​of all normal samples are obtained, thus obtaining the Rhodiola rosea quality identification model.

7. The method for quality identification of Rhodiola rosea according to claim 1, characterized in that, Step S3 is followed by: S4, evaluating the classification performance of the Rhodiola Rosea quality identification model using the training set and test set.

8. The method for quality identification of Rhodiola rosea according to claim 7, characterized in that, The specific steps of step S4 include: inputting the test set into the trained Rhodiola rosea quality identification model, obtaining the model's prediction results, comparing the prediction results with the labeling results in step S1, and calculating the model's correct classification rate and specificity; demonstrating the model's classification performance by reconstructing the error distribution histogram and the hidden layer feature t-SNE visualization; designing ablation experiments to construct simplified models that rely only on reconstruction errors and those that rely only on hidden layer features, and comparing the classification performance of the complete model and the simplified model.

9. A terminal device, characterized in that, include: Memory, processor, display module; The memory stores a computer program that implements the method according to any one of claims 1 to 8; The processor runs computer programs stored in the memory; The display module is used to show the user the output results after the process is completed.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program / instruction, which is executed by a processor according to the steps of the Rhodiola rosea quality identification method of any one of claims 1 to 8.

Images