A multi-index detection system for traditional Chinese medicine injection based on hyperspectral

Abstract

The present invention relates to the application field of hyperspectral technology, and in particular to a detection system for simultaneously measuring multiple indicators of traditional Chinese medicine injections. It includes a traditional Chinese medicine injection ampoule bottle transmission module, a hyperspectral image acquisition module, and a hyperspectral image processing module; the ampoule bottle transmission module transmits the traditional Chinese medicine injection to be tested to the hyperspectral image acquisition area, the hyperspectral image acquisition module acquires the hyperspectral image of the injection, and the hyperspectral image processing module performs data processing on the acquired hyperspectral image, analyzes and outputs the chemical composition prediction results, activity prediction results, chromaticity detection results, and visible foreign matter detection results in the traditional Chinese medicine injection. It is only necessary to scan the traditional Chinese medicine injection bottle on the bottle body transmission module to form a hyperspectral image, and the extracted spectral information can predict multiple chemical indicators and multiple activity indicators and measure the chromaticity; the extracted image information can determine whether there are visible foreign matters in the injection; the detection method and detection equipment are greatly simplified, saving time and effort.

Description

Technical Field

[0001] The invention relates to the application field of hyperspectral technology, and in particular to a detection system for simultaneously measuring multiple indicators of traditional Chinese medicine injections. Background Art

[0002] Chinese medicine injection is a product that combines traditional Chinese medicine with modern preparations. It is widely used in the treatment of cardiovascular and cerebrovascular diseases, viral infections, inflammation and other indications. However, the frequent occurrence of adverse drug reactions makes the use of Chinese medicine injection controversial, so it is crucial to control the quality of Chinese medicine injection.

[0003] At present, the quality inspection standards for Chinese medicine injections include two physical indicators: color and visible foreign matter. Color is one of the indicators for enterprises to control the consistency between batches, and the color standards of different enterprises may be different. Visible foreign matter can cause clinical adverse reactions such as thrombosis and inflammation. Controlling the two indicators of color and visible foreign matter is conducive to ensuring the consistency and safety of medication. Some Chinese medicine injections also need to be tested for chemical indicators and activity indicators (the efficacy of Chinese medicine injections is the result of the combined effect of its effective ingredients. Testing the activity indicators and evaluating the quality of Chinese medicine injections can better ensure the effectiveness of Chinese medicine injections). For example, for Shuxuening injection, the 2015 Chinese Pharmacopoeia stipulates that the two major chemical indicators of total flavonol glycosides and ginkgo lactones need to be tested. These two major categories of components are considered to be the main active ingredients of ginkgo extract; in addition, based on its indications, it is generally believed that anticoagulant activity and antioxidant activity are closely related to its biological activity, and they are also quality indicators that need attention.

[0004] The detection of chemical indicators of injection products at home and abroad mainly adopts sampling for offline detection, generally thin layer chromatography identification, high performance liquid chromatography, etc., and mainly detects the index components of traditional Chinese medicine injection. Although these methods are stable and reliable, they also have disadvantages such as slow detection speed, sample destruction, and environmental pollution caused by the generated chemical and biological reagents. These problems also exist in the conventional detection methods of biological activity.

[0005] There are also a series of reports on other non-destructive chemical index detection methods, such as applying near-infrared spectroscopy and ultraviolet spectroscopy to the detection of chemical components in liquids. For example, a Chinese patent discloses a near-infrared spectroscopy analysis method based on a one-dimensional convolutional neural network, with the application number CN201710780270.2. This type of patent usually includes the following steps: collecting spectral data of training set samples and preprocessing them; using the preprocessed training set data to establish a spectral correction model; collecting spectral data of test set samples and preprocessing them; substituting the preprocessed test set data into the spectral correction model to obtain prediction results. The establishment of quantitative correction modeling methods generally requires spectral preprocessing and extraction of characteristic bands. Although this analysis method simplifies the detection process to a certain extent, there are still the following problems: the collected spectral data must first be spectrally preprocessed, which will lead to unnecessary computing power burden. Moreover, the preprocessing method often varies with environmental factors (temperature, humidity, ambient light) and subjective factors of the operator, and the generalization performance of the model is limited. Incorrect use of the preprocessing method may also distort the spectral signal, resulting in a decrease in model accuracy. Although the algorithm for extracting characteristic bands selects key bands in the spectrum, it may also cause loss of effective information. Neural Network (NN) can theoretically achieve the purpose of modeling without preprocessing and extracting characteristic bands, but due to too many parameter variables in the neural network, it leads to slow training speed and the risk of overfitting.

[0006] Visible foreign matter detection is an inspection item stipulated in the general rules for injections. It is generally detected by manual light inspection, which is inefficient and highly subjective. It can also be detected by light inspection machines, which can quickly and efficiently remove unqualified samples containing visible foreign matter on the production line. However, the light inspection machine market in China is monopolized by three major groups: Japan's Eisai, Italy's Beverett and Germany's Sennende, and their quotations can be as high as 10 million RMB. This has resulted in the vast majority of Chinese pharmaceutical companies being unable to use light inspection machines to inspect their products, and instead using inefficient and subjective manual inspection methods.

[0007] In addition, when Chinese medicine injections are actually shipped out of the factory, the quality of the products between batches is actually not uniform, mainly because: 1. Raw material differences: The quality of Chinese patent medicine products depends on the chemical composition of Chinese medicinal materials. Although different batches of medicinal materials follow strict planting and harvesting specifications, the origin, weather and other non-human factors still cause differences in chemical composition; 2. Fluctuations introduced in the production process: Although the production and manufacturing of Chinese patent medicines strictly follow the declared process regulations, there are still differences that need to rely on the judgment of the worker supervisor or have not yet been clarified. Different batches of medicines produced from the same raw materials still have certain quality differences; 3. Even the same batch of products will have slight differences in product quality due to differences in filling time, sterilization location, etc. However, the current quality inspection of Chinese medicine injections is basically carried out by random inspection, and the possibility of missed inspection is very high, and the safety during actual use cannot be effectively guaranteed.

[0008] In summary, in the prior art, the chemical, biological activity, and physical indicators of Chinese medicine injections are divided into different detection systems for measurement, and offline sampling detection is generally used. This leads to defects such as complex and expensive existing detection equipment, cumbersome and time-consuming detection process, large workload, low efficiency, and low reliability of product detection results. Summary of the invention

[0009] In view of the problems in the prior art such as complicated and time-consuming detection process, high price, and low safety assurance due to the random inspection method, the present invention provides a detection system for simultaneously measuring multiple indicators of traditional Chinese medicine injection.

[0010] A multi-index detection system for Chinese medicine injection based on hyperspectrum, comprising a Chinese medicine injection ampoule conveying module, a hyperspectral image acquisition module, and a hyperspectral image processing module; the ampoule conveying module conveys the Chinese medicine injection to be inspected to a hyperspectral image acquisition area, the hyperspectral image acquisition module acquires a hyperspectral image of the injection, and the hyperspectral image processing module processes the acquired hyperspectral image, analyzes and outputs chemical component detection results, activity detection results, chromaticity detection results, and visible foreign matter detection results in the Chinese medicine injection.

[0011] Furthermore, the hyperspectral image processing module includes a chemical / activity index processing unit, a chromaticity index processing unit, and a visible foreign matter index processing unit. The chemical / activity index processing unit analyzes and obtains the chemical composition and activity detection results in the traditional Chinese medicine injection, the chromaticity index processing unit analyzes and obtains the chromaticity detection results, and the visible foreign matter index processing unit analyzes and obtains the visible foreign matter detection results.

[0012] Furthermore, the chemical / activity index processing unit uses a prediction model established by a convolutional neural network (i.e., CNN network) method to process the collected hyperspectral images through the CNN network method; the CNN network includes an input layer, three convolutional pooling layers, a fully connected layer, and N parallel output modules, each output module contains two fully connected layers and an output layer; the input layer uses the original spectrum of the input sample, and the output layer simultaneously outputs N chemical component quantitative results and biological activity detection results.

[0013] Furthermore, in the chemical / activity index processing unit, the pixel with the largest RGB value in the full hyperspectral image is selected, and a rectangular area is delineated on the bottle body. The RGB value of the pixel in the rectangular area should be greater than or equal to 70% of the maximum value. The rectangular area is different from other areas and represents the bottle body of the Chinese medicine injection as the hyperspectral sampling and processing area. Within the width and height range of the bottle body sampling and processing area, a number of sampling and processing points for each sample are evenly selected in a matrix form.

[0014] Furthermore, the average spectrum of the pixel blocks covered by the sampling and processing points and the volume of the sample liquid covered by the sampling and processing points and the optical path (length × width × optical path depth of the sampling point area) are calculated, and then the obtained average spectrum of the pixel blocks is used as the input of the quantitative correction model, and the value obtained by multiplying the obtained volume with the chemical composition content and activity data of the sample determined by conventional methods is used as the output of the quantitative correction model.

[0015] Furthermore, the chromaticity index processing unit calculates the average spectrum of the hyperspectral image in the hyperspectral sampling processing area, selects the light intensity at 700nm (red), 546.1nm (green) and 435.8nm (blue), and converts it into the H value in the HSV color space, which is the chromaticity.

[0016] Furthermore, the visible foreign matter index processing unit selects the hyperspectral image collected at the bottom of the Chinese medicine injection bottle as the image processing area; the bottle conveying module gathers the visible foreign matter of the conveyed Chinese medicine injection to the bottom edge of the injection bottle, and conveys it to the hyperspectral image collection area, and the visible foreign matter index processing unit extracts the bottom area contour of the collected hyperspectral image to detect the large-scale and small-scale particles deposited in the injection bottle.

[0017] Furthermore, for the bottom area contour of the injection bottle, the visible foreign matter index processing unit uses the chain code technology to extract the bottom edge of the side containing the visible foreign matter as the lower curve; the lower curve is subjected to large-scale detection, and the points where the curvature suddenly changes are detected on the lower curve, and then the adjacent corner points are connected with line segments to determine the possible foreign matter area, and the foreign matter area is judged whether it is a false alarm based on the root mean square standard deviation of the spectrum in the foreign matter area. If the root mean square standard deviation is less than 0.2, it is a false alarm; the part of the lower curve where the large-scale detection method fails to detect foreign matter continues to be inspected at a small scale, and the small-scale inspection directly checks the curvature of the lower curve. When the curvature of the pixel point on the curve is greater than 0.3, it is judged that small particles of foreign matter exist in the sample.

[0018] Furthermore, the bottle conveying module includes an inclined vibrating conveying platform, an inclined conveying platform, and a fixed bracket for conveying the bottles upright and in an orderly manner. The inclined conveying platform and the inclined vibrating conveying platform are in an inclined state with an inclination angle of 5-15°. The injection bottles are conveyed forward from the inclined vibrating conveying platform and the inclined conveying platform in sequence.

[0019] Furthermore, the hyperspectral image acquisition module includes a halogen lamp light source, a hyperspectral camera and a light shield, wherein the halogen lamp light source and the hyperspectral camera are arranged on both sides of the bottle conveying module, wherein the hyperspectral camera collects wavelengths in a range including all wavelength combinations within 400-1700nm, and a wavelength resolution of 5nm and below, and the hyperspectral camera scans the Chinese medicine injection bottles on the inclined conveying platform to obtain hyperspectral images.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0021] (1) The multi-index detection system for Chinese medicine injection based on hyperspectral provided by the present invention has a hyperspectral image acquisition module that only needs to scan the Chinese medicine injection bottle on the bottle conveying module to form a hyperspectral image, and the hyperspectral image processing module performs data processing on the acquired hyperspectral image. The extracted spectral information can predict multiple chemical indicators, multiple activity indicators and chromaticity; the extracted image information can be used to determine whether there are visible foreign matter in the injection; after the modeling is completed, no sample processing is required, and multi-index simultaneous detection can be achieved non-destructively and quickly, which greatly simplifies the detection method and detection equipment, saves time and effort, and adopts the system provided by the present invention to improve the detection efficiency of quality indicators of Chinese medicine injection.

[0022] (2) The present invention can be applied to online detection on the production line, and can detect all injection products with high throughput, thereby achieving full product inspection, overcoming the problem of missed inspections due to random inspections, better ensuring product quality, and effectively improving the safety and effectiveness of Chinese medicine injections. In addition, the index results can be predicted without damaging the injection samples, and the method is clean and environmentally friendly.

[0023] (3) The chemical / activity index processing unit of the present invention uses a convolutional neural network (i.e., CNN network) method to establish a prediction model. Due to the use of a deep neural network and a convolutional layer structure, the number of variables that need to be trained is greatly reduced, thereby increasing the scalability of the model while reducing the amount of calculation and ensuring the accuracy of the model. In addition, the CNN model of this invention uses a common convolutional layer module for N indicators, that is, it extracts common spectral features. The CNN model of the present invention can predict N indicators by establishing a model, which only takes 1 / N of the time, and does not require spectral preprocessing and extraction of characteristic bands and manual adjustment of model parameters. It is an end-to-end quantitative correction model with strong versatility.

[0024] (4) The hyperspectral-based visible foreign matter detection method for injections of the present invention is different from the detection principle of the traditional light inspection machine. In the present invention, the visible foreign matter is gathered at the bottom of the bottle, and then the bottle bottom is inspected in large and small scales, with higher detection accuracy; and the device used in the method has a simple structure and is easy to operate, which can greatly reduce the detection cost.

[0025] (5) This system can be widely used in the field of quality inspection of traditional Chinese medicine injections. It only needs to make corresponding modeling adjustments based on the specific chemical composition and activity indicators of different preparations. Chromatography and visible foreign matter are common inspection items for water injections and can be directly applied. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The following is a simplified flow chart of the detection system provided by the present invention.

[0027] Figure 2 It is a schematic diagram of the detection system provided by the present invention.

[0028] Figure 3 It is a detection flow chart of the embodiment provided by the present invention.

[0029] Figure 4 It is a structural diagram of a convolutional neural network according to an embodiment of the present invention.

[0030] Figure 5 It is a sampling schematic diagram of the biochemical indicator processing unit of the embodiment provided by the present invention.

[0031] Figure 6 It is a schematic diagram of sampling by the visible foreign matter indicator processing unit of the embodiment provided by the present invention.

[0032] Figure 7 It is the chromaticity distribution diagram of 7 manufacturers of the embodiments provided by the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] See also Figure 1 The present embodiment provides a multi-index detection system for Chinese medicine injection based on hyperspectrum. The Chinese medicine injection is in the shape of a cylindrical bottle, and includes a Chinese medicine injection ampoule bottle transmission module, a hyperspectral image acquisition module, and a hyperspectral image processing module; the ampoule bottle transmission module transmits the Chinese medicine injection to be inspected to the hyperspectral image acquisition area, the hyperspectral image acquisition module acquires the hyperspectral image of the injection, and the hyperspectral image processing module processes the acquired hyperspectral image. The hyperspectral image processing module includes a chemical / activity index processing unit, a chromaticity index processing unit, and a visible foreign matter index processing unit. The chemical / activity index processing unit analyzes and obtains the chemical composition and activity detection results in the Chinese medicine injection, the chromaticity index processing unit analyzes and obtains the chromaticity detection result, and the visible foreign matter index processing unit analyzes and obtains the visible foreign matter detection result.

[0035] Since different substances have different performances under different band spectral signals, hyperspectral imaging technology is based on a large number of narrow-band image data technology. It combines imaging technology with spectral technology to detect the two-dimensional geometric space and spectral information of the target and obtain high-resolution continuous, narrow-band image data, which provides the possibility for the implementation of the detection of N indicators in the present invention.

[0036] See also Figure 2 In this system, the bottle conveying module includes an inclined vibration conveying platform 4, an inclined conveying platform 3, and a fixed bracket 2 for making the bottle stand upright and tilted in an orderly manner. The inclination angles of the inclined vibration conveying platform 4 and the inclined conveying platform 3 are 5-15°, preferably 5°, so as to achieve a better balance between conveying and depositing foreign particles. The hyperspectral image acquisition module includes a halogen lamp light source 1, a hyperspectral camera 7, and a light shield (not shown in the figure), wherein the hyperspectral camera 7 is a visible-shortwave near-infrared hyperspectral camera, and the hyperspectral camera 7 collects wavelengths in a range of all wavelengths within 400-1700nm, with a wavelength resolution of 5nm and below. The hyperspectral camera 7 scans the Chinese medicine injection bottle on the inclined conveying platform 3 to obtain a hyperspectral image; the hyperspectral image processing module is a computer 6.

[0037] The halogen light source 1 is a surface light source with adjustable light intensity. The fixed bracket 2 is used to prevent the injection 5 from tipping over. The Chinese medicine injection bottle is sequentially conveyed forward by the inclined vibration conveying platform 4 and the inclined conveying platform 3 to pass in front of the hyperspectral camera lens. The vibration conveying platform 4 and the inclined conveying platform 3 are used to collect visible foreign matter on the bottom edge of the injection. The hyperspectral camera 7 obtains the hyperspectral image of the injection, and the computer 6 performs subsequent data processing on the obtained image. The hyperspectral image acquisition module is placed in a light-shielding cover or a dark box to prevent interference from ambient light.

[0038] In the present invention, Shuxuening injection is taken as an example. Shuxuening injection is a sterile aqueous solution processed from ginkgo leaves, which has the effects of dilating blood vessels and improving microcirculation. It is mainly used clinically for ischemic cardiovascular and cerebrovascular diseases, coronary heart disease, angina pectoris, cerebral embolism, and cerebral vasospasm. It is one of the major varieties of Chinese medicine injection. The quality standard stipulates the quantitative detection of the content of two major chemical components, total flavonol glycosides and ginkgo lactones, so these two quantitative indicators are selected as chemical indicators; in addition, from its indications, its anticoagulant activity and antioxidant activity indicators are closely related to its drug efficacy, so these two indicators are selected as activity indicators. At the same time, chromaticity and visible foreign matter are selected as physical detection indicators, and a total of 6 quality indicators are selected.

[0039] At present, the testing of chemical and activity indicators of injection products at home and abroad mainly adopts sampling to conduct offline testing. Figure 3 As shown, the total flavonol glycoside concentration is detected by liquid chromatography, the ginkgolide concentration is detected by evaporative light heat dissipation, the anticoagulant activity (thrombin activity inhibition rate) is detected by fluorescence, and the antioxidant activity (DPPH free radical scavenging rate) is detected by colorimetry. Although these methods are stable and reliable, they also have disadvantages such as slow detection speed, sample damage, and environmental pollution caused by the generated chemical and biological reagents. In the prior art, the chemical, active, and physical indicators of Chinese medicine injections are all divided into different detection systems for measurement, which leads to complex detection equipment, cumbersome and time-consuming detection process, large workload, and low efficiency. There are a series of reports on the application of near-infrared spectroscopy and ultraviolet spectroscopy to the detection of chemical components in liquids in the prior art, but traditional spectroscopy has two main disadvantages: (1) Traditional spectroscopy technology only collects spectral information but not image information, so it cannot handle the spatial heterogeneity of cylindrical ampoules; (2) Visible foreign matter detection requires spatial information, which cannot be achieved by spectroscopy. In the prior art, expensive light inspection machines are mostly used for the detection of visible foreign matter, which makes the overall detection cost very high.

[0040] When Shuxuening injection is applied to this system, it is only necessary to scan the bottle of Chinese medicine injection on the bottle transmission module to form a hyperspectral image, the hyperspectral image processing module performs data processing on the collected hyperspectral image, the biochemical index processing unit of the hyperspectral image processing module analyzes and obtains the chemical index detection results and biological activity index detection results of the Chinese medicine injection, the colorimetric index processing unit analyzes and obtains the colorimetric index detection results, and the visible foreign matter index processing unit analyzes and obtains the visible foreign matter index detection results.

[0041] The testing of Shuxuening injection, when used in practice, includes sampling, calibration, data processing and other processes. The specific process is as follows:

[0042] 1. Simultaneous quantitative detection of chemical / active indicators

[0043] 1) Image acquisition: A total of 60 batches of samples from 7 companies were collected with hyperspectral images, including 11 bottles of colorless 2ml, 37 bottles of colorless 5ml and 12 bottles of brown 5ml ampoules. Since this technology uses the method of collecting transmission spectra and the cylindrical liquid has a focusing effect, it is necessary to first collect hyperspectral images of purified water samples of the same volume filled with ampoules of the same specifications as the background for whiteboard calibration of subsequent samples. In order to eliminate the errors caused by different bottle colors, images of three different specifications of water ampoules were used as calibration images.

[0044] 2) Image correction: In this embodiment, the spectra of 60 sample sampling points are directly divided by the spectra of 3 calibration image sampling points, and the obtained result is used as the corrected spectrum of each sample.

[0045] 3) Image processing: For these 63 hyperspectral images, first select the pixel with the largest RGB value in the full hyperspectral image, and define a rectangular area on the bottle. The RGB value of the pixel in the rectangular area should be greater than or equal to 70% of the maximum value. The rectangular area is different from other areas and represents the bottle of Chinese medicine injection as the hyperspectral sampling and processing area. The existing hyperspectral detection technology usually sets the entire sample image area as the image processing area, calculates the average spectrum and uses it for quantitative correction modeling. However, due to the characteristics of the cylindrical ampoule bottle, the optical path at different positions in the sample area is different, and the above method will cause large errors. Therefore, the present invention uniformly collects 3×5 small pixel blocks in each selected sampling and processing area, and applies to the 63 hyperspectral images for chemical / activity detection. According to the width and height of the selected sampling and processing area, 15 sampling and processing points are uniformly obtained (in the form of a 3*5 matrix). Its specific form is: Figure 5 As shown, let the size of the hyperspectral data be L W ×L H ×L B , where L W ×L Hare the length and width of the spatial dimension, and L B is the number of bands. Set five vertical lines in the rectangular area of ​​the bottle: Set three horizontal lines at the same time The intersection of the horizontal and vertical lines is the sampling and processing point. Calculate the average spectrum of the pixel blocks covered by the sampling and processing points and the volume of the sample liquid covered by the sampling and processing points and the optical path (length × width × optical path depth of the sampling and processing point area). Randomly divide the average spectra of all sampling and processing points and the corresponding chemical / activity indicators into a calibration set and a validation set in a ratio of 3:1, and then use the average spectrum of the pixel blocks in the calibration set as the input of the quantitative calibration model, and use the corresponding volume multiplied by the chemical composition content and activity data of the sample measured by conventional methods as the output of the quantitative calibration model. The chemical / activity indicator processing unit uses a convolutional neural network (CNN) method to establish a prediction model, such as Figure 4 As shown in the figure, the CNN network contains an input layer, three convolutional pooling layers, a fully connected layer and four parallel output modules. Each module contains two fully connected layers and an output layer. The input layer inputs the original spectrum of the sample without preprocessing, and the output layer simultaneously outputs four quantitative results: two chemical component contents (total flavonol glycosides, ginkgo lactones), and two activity indicators (antioxidant activity, anticoagulant activity).

[0046] The validation set was tested on the prediction model obtained after training, and better results were obtained than the traditional algorithm - partial least squares regression. The model evaluation and comparison results with traditional methods are shown in the following table:

[0047]

[0048]

[0049] The model established by the present invention can effectively output the test results of four quality indicators, and the determination coefficient (R 2 The prediction performance of the established model was evaluated by three model evaluation indicators: root mean square error (RMSE, the smaller the better), root mean square error (RMSE, the smaller the better), and residual prediction deviation (RPD, the larger the better). Compared with the traditional partial least squares method, each quality indicator achieved better model performance on both the calibration set and the validation set.

[0050] The conventional CNN network model models the four indicators by establishing four separate models. The CNN model of the present invention can predict the four indicators by establishing one model, which only takes 1 / 4 of the time. In addition, the CNN model of the present invention uses a common convolution layer for the four indicators, that is, it extracts common spectral features. The characteristic variables in the spectrum can be inferred based on the weight coefficients of the convolution layer and the input layer, which increases the interpretability of CNN. At the same time, the hyperspectral system can be transformed into a multi-spectral system, which can greatly improve the detection speed.

[0051] The CNN network model in the present invention modifies the output module of a fully connected layer into multiple parallel ones to achieve the goal of multi-output.

[0052] In order to consider the prediction accuracy of the four indicators at the same time, the sum of the mean square error (MSE) of the four indicators is used as the model loss for training, so as to comprehensively evaluate the prediction effect of the model for the four indicators. In order to eliminate the bias of the gradient descent direction during model training caused by dimensional differences, the output data needs to be normalized before model training, and denormalized when used for prediction.

[0053] Regarding the preprocessing problem: Traditional spectral quantitative correction modeling methods usually include three steps: spectral preprocessing, characteristic band extraction and quantitative modeling.

[0054] The existing spectral preprocessing methods are: a smoothing method; b derivative method; c standard normal variable transformation method; d multivariate scattering correction. The more common method is to extract the characteristic bands in a supervised or unsupervised manner after preprocessing the spectrum. After the methods of preprocessing and extracting characteristic bands are fully combined, they are screened according to the modeling effect. This traversal algorithm combination method will lead to unnecessary computing power burden. Moreover, the preprocessing methods often vary with environmental factors (temperature, humidity, ambient light) and subjective factors of the operator, and the generalization performance of the model is limited. Incorrect use of preprocessing methods may also distort spectral signals and reduce model accuracy. Although the algorithm for extracting characteristic bands selects key bands in the spectrum, it may also cause loss of effective information. Neural Network (NN) can theoretically achieve the purpose of completing modeling without preprocessing and extracting characteristic bands, but due to the excessive number of parameter variables in the neural network, the training speed is slow and there is also the risk of overfitting.

[0055] This CNN technology does not require preprocessing and extracting feature bands. At the same time, due to the use of deep neural network and convolution layer structure, the number of variables is greatly reduced, which increases the scalability of the model while reducing the amount of calculation and ensuring the accuracy of the model. The convolutional neural network model contains convolutional layers and fully connected layers. The convolutional layer is essentially a feature extraction process that can extract feature information from the spectrum and filter out some irrelevant information. Therefore, a more ideal modeling effect can be obtained without spectral preprocessing.

[0056] 2. Detection of chromaticity indicators

[0057] 1) Image acquisition: Hyperspectral images of the bottle body area of ​​63 ampoules were taken for “simultaneous quantitative detection of chemical / activity indicators”.

[0058] 2) Image processing: For these 63 hyperspectral images, first select the pixel with the largest RGB value in the full hyperspectral image, and define a rectangular area on the bottle. The RGB value of the pixel in the rectangular area should be greater than or equal to 70% of the maximum value. The rectangular area is different from other areas and represents the bottle of Chinese medicine injection as the hyperspectral sampling and processing area; the chromaticity index processing unit calculates the average spectrum of the hyperspectral image in the hyperspectral sampling and processing area, and selects the light intensity at 700nm (red), 546.1nm (green) and 435.8nm (blue). The chromaticity distribution of the seven manufacturers is shown in Figure 2. Figure 7 .

[0059] Specifically, the light intensity of the three bands is set as R', G' and B'. The calculation formula of chromaticity is as follows:

[0060] Δ=max(R,G,B′)-min(R,G,B)

[0061]

[0062] 3. Detection of visible foreign matter indicators

[0063] 1) Image acquisition: A total of 100 bottles of samples (32 bottles were judged to be abnormal when leaving the factory) were collected for hyperspectral images. In addition, in order to test whether this method can be applied to the detection of particles of tens of microns, hyperspectral images of 60μm standard particles were collected. For the hyperspectral images of these 101 ampoules, a small window of 3*3 pixels was used to traverse all the pixels of the hyperspectral image, except for a circle of pixels at the outermost edge of the image. The variance of the spectrum of each pixel in the small window was calculated, and the variance value was used as the pixel value of the new image to obtain a single-layer grayscale image. The watershed algorithm was used to obtain the bottom mask of the grayscale image, thereby obtaining the region of interest at the bottom.

[0064] 2) Image processing: Extract the bottom area contour, and use the chain code technology to extract the half of the bottom edge close to the lens, which is called the lower curve. First, use a large-scale detection method to detect possible corner points (points with sudden changes in curvature) on the lower curve, and then connect adjacent corner points with line segments, such as Figure 6 As shown in (1), the possible foreign body area is determined. Based on the root mean square standard deviation of the spectrum in the foreign body area, the foreign body area is judged whether it is a false alarm. If the root mean square standard deviation is less than 0.2, it is a false alarm. These false alarms may be caused by dust attached to the outer wall or spectral noise points.

[0065] The formula for calculating the root mean square standard deviation is as follows:

[0066]

[0067] where n i and RSD(i) is the number of pixels and the root mean square difference of the i-th possible impurity area, PI j,k is the light intensity of the kth band at the jth pixel.

[0068] In the lower curve part where the large-scale detection method fails to detect foreign matter, a small-scale inspection is continued. The small-scale inspection directly checks the curvature of the curve. If there is a pixel point on the curve with a curvature greater than 0.3, such as Figure 6 (2) shows that there are small foreign particles in the sample. This method can detect visible foreign matter with a particle size as low as 60μm under static conditions.

[0069] There are 100 samples for visible foreign matter inspection, of which 32 are abnormal samples. Figure 6 As shown in (3), L represents large-scale particles and S represents small-scale particles. The large-scale inspection and small-scale inspection of this method were run sequentially, of which 33 bottles were detected by large-scale inspection and 12 bottles were detected by small-scale inspection. Combining the large-scale and small-scale results, a total of 35 bottles were detected. The light inspection machine detected 32 bottles, so this method can detect all abnormal samples, but there are "false alarms" for 3 bottles. For some samples that were detected abnormally by large-scale inspection, it detected additional impurity particles, and small-scale inspection can effectively avoid the occurrence of missed detection. In addition, the small-scale inspection method used in this method can detect 60μm particles, which is close to the pharmacopoeia's detection standard that cannot detect 50μm particles.

[0070] Different from the "light inspection machine" detection: the current more mature light inspection machine uses a high-speed CCD camera to collect images of the drug solution during the flipping process, and determines whether there are visible particles by analyzing the changes in the grayscale of the pixels in the image. Its mechanical structure is complex and has high requirements on the shooting speed and imaging accuracy of the CCD camera, so it is expensive. This system does not require the sample to be flipped, which reduces the complexity of the mechanical structure. In addition, due to the combination of spectral information, it can reduce the requirements for pixel resolution to a certain extent, thereby reducing costs.

[0071] In general, when the multi-index detection system of traditional Chinese medicine injection is applied to Shuxuening injection, Figure 3 As shown in the figure, after the Shuxuening injection forms a hyperspectral image, the bottle body area is tested for chromaticity indicators, and the bottle bottom area is tested for visible foreign matter indicators based on the judgment of spectral variance and edge curvature. A quantitative correction model is made for the matrix sampling point pixel blocks based on the convolutional neural network to predict the chemical and activity indicator results.

[0072] The multi-index detection system for traditional Chinese medicine injection provided by the present invention has a hyperspectral image acquisition module that only needs to scan the traditional Chinese medicine injection bottle on the bottle transmission module to form a hyperspectral image, and a hyperspectral image processing module performs data processing on the acquired hyperspectral image, and extracts spectral information to predict multiple chemical indicators and multiple activity indicators and obtain chromaticity detection results; extracting image information can determine whether there are visible foreign matter in the injection; after the modeling is completed, multiple indicators can be detected simultaneously without any physical and chemical indicators, which greatly simplifies the detection method and detection equipment, saves time and effort, and adopts the system provided by the present invention to improve the detection efficiency of traditional Chinese medicine injection.

[0073] The multi-index detection system of the present invention can be applied to online detection on the production line, and can detect all injection products with high throughput, realize full product inspection, overcome the problem of missed inspections caused by random inspections, better ensure product quality, effectively improve the safety of drug use, and measure indicator results without damaging the injection samples, which is clean and environmentally friendly.

[0074] In addition, this system can also be applied to other liquid preparations and adjusted accordingly according to the specific chemical composition and activity indicators of different preparations. Chromatography and visible foreign matter are common inspection items for water injections and can be directly applied, making them suitable for promotion.

[0075] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principle of the present invention shall be equivalent replacement modes and shall be included in the protection scope of the present invention.

Claims

1. A multi-index detection system for traditional Chinese medicine injection based on hyperspectral, It is characterized in that It includes a Chinese medicine injection ampoule conveying module, a hyperspectral image acquisition module, and a hyperspectral image processing module; the ampoule conveying module conveys the Chinese medicine injection to be inspected to the hyperspectral image acquisition area, the hyperspectral image acquisition module acquires the hyperspectral image of the injection, and the hyperspectral image processing module processes the acquired hyperspectral image, analyzes and outputs the chemical composition detection results, activity detection results, chromaticity detection results, and visible foreign matter detection results in the Chinese medicine injection; Among them, the hyperspectral image processing module includes a chemical / activity index processing unit, a chromaticity index processing unit, and a visible foreign matter index processing unit; The chemical / activity index processing unit uses a prediction model established by a CNN network method to process the collected hyperspectral images through the CNN network method; the CNN network includes an input layer, three convolutional pooling layers, a fully connected layer, and N parallel output modules, each of which contains two fully connected layers and an output layer; the input layer uses the original spectrum of the input sample, and the output layer simultaneously outputs N chemical component quantitative results and biological activity detection results; In the chemical / activity index processing unit, the pixel with the largest RGB value in the full hyperspectral image is selected to delineate a rectangular area on the bottle. The RGB value of the pixel in the rectangular area should be greater than or equal to 70% of the maximum value. This rectangular area is different from other areas and represents the bottle of Chinese medicine injection as the hyperspectral sampling and processing area. Within the width and height of the bottle sampling and processing area, several sampling and processing points for each sample are evenly selected in a matrix form.

2. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 1, It is characterized in that The average spectrum of the pixel blocks covered by the sampling and processing points and the volume of the sample liquid covered by the sampling and processing points and the optical path are calculated, and the obtained average spectrum of the pixel blocks is used as the input of the quantitative correction model. The value obtained by multiplying the obtained volume with the chemical composition content and activity data of the sample determined by conventional methods is used as the output of the quantitative correction model.

3. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 1, It is characterized in that The chromaticity index processing unit calculates the average spectrum of the hyperspectral image in the hyperspectral sampling processing area, selects the light intensity at 700 nm, 546.1 nm and 435.8 nm, and converts it into the H value in the HSV color space, which is the chromaticity.

4. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 1, It is characterized in that The visible foreign matter index processing unit selects the hyperspectral image collected at the bottom of the Chinese medicine injection bottle as the image processing area; the bottle conveying module gathers the visible foreign matter of the conveyed Chinese medicine injection to the bottom edge of the injection bottle, and conveys it to the hyperspectral image collection area, and the visible foreign matter index processing unit extracts the bottom area contour of the collected hyperspectral image to detect the large-scale and small-scale particles deposited in the injection bottle.

5. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 4, It is characterized in that For the bottom area contour of the injection bottle, the visible foreign matter index processing unit uses the chain code technology to extract the bottom edge of the side containing the visible foreign matter as the lower curve; the lower curve is subjected to large-scale detection, and the points with sudden changes in curvature are detected on the lower curve, and then the adjacent corner points are connected with line segments to determine the possible foreign matter area, and the foreign matter area is judged whether it is a false alarm based on the root mean square standard deviation of the spectrum in the foreign matter area. If the root mean square standard deviation is less than 0.2, it is a false alarm; the part of the lower curve where the large-scale detection method fails to detect foreign matter continues to be inspected at a small scale, and the small-scale inspection directly checks the curvature of the lower curve. When the curvature of the pixel point on the curve is greater than 0.3, it is judged that small particles of foreign matter exist in the sample.

6. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 1, It is characterized in that The ampoule bottle conveying module comprises an inclined vibration conveying platform (4), an inclined conveying platform (3), and a fixed bracket (2) for allowing the bottles to be transported upright and in an orderly manner. The surfaces of the inclined conveying platform (3) and the inclined vibration conveying platform (4) are in an inclined state with an inclination angle of 5-15 degrees. The injection bottles are conveyed and moved forward from the inclined vibration conveying platform (4) and the inclined conveying platform (3) in sequence.

7. The multi-index detection system for traditional Chinese medicine injection based on hyperspectral according to claim 1, It is characterized in that The hyperspectral image acquisition module comprises a halogen lamp light source (1), a hyperspectral camera (7) and a light shield, wherein the halogen lamp light source (1) and the hyperspectral camera (7) are arranged on both sides of the bottle conveying module, wherein the hyperspectral camera (7) collects wavelengths in a range of all wavelength combinations within a range of 400-1700 nm, and has a wavelength resolution of 5 nm or less, and the hyperspectral camera (7) scans the Chinese medicine injection bottle on the inclined conveying platform (3) to obtain a hyperspectral image.

Images