Artemisinin extraction intelligent tracking and identification method

Abstract

The application discloses an intelligent tracking and identification method for artemisinin extraction, and relates to the technical field of image data processing; the method cooperatively constructs an integrated closed-loop system through four core technical means: type filtering and bimodal adaptive segmentation based on foam stability difference, accurate removal of dynamic foam and optimization of preliminary interface extraction; attention-enhanced U-Net network embedding channel attention module and edge-enhanced branch, realizing accurate positioning and fluctuation tracking of layered interfaces; attention mechanism dynamically adjusts bimodal weight, combined with partition strategy and improved network to complete multi-index collaborative tracking; the crystallization growth model is introduced to correct particle weight, hierarchical resampling is adopted to optimize tracking effect, and a closed-loop feedback mechanism of crystallization state and process parameters is established. Deep cooperation of each technical means realizes intelligent and accurate control of the whole extraction process, and provides reliable technical support for artemisinin extraction production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, in particular to an intelligent tracking and identification method for artemisinin extraction. BACKGROUND

[0002] As an important raw material for antimalarial drugs, the stability and precision of the extraction production process of artemisinin directly determine the purity and yield of the product. The conventional pilot-scale artemisinin extraction production line mostly uses solvent extraction method to realize the full contact of raw materials and solvent through stirring under normal temperature and pressure environment. However, a large amount of dynamic foam is generated in the stirring process, which is easy to collapse and merge and is accompanied by rotary motion. Not only does it cover the interface of the extraction liquid layer, but also it blocks the initial crystalline particles, which brings core interference to the process monitoring. At present, the monitoring of the extraction process in the industry mostly relies on single image acquisition and processing technology, such as using visible light image to identify the shape of raw materials or using near-infrared image to analyze the composition of extraction liquid. There is no integrated control system formed by the cooperation of multiple technologies, resulting in fragmented monitoring data and the inability to fully reflect the real state of the extraction process.

[0003] In the prior art, although single technical means such as image denoising, interface segmentation and index identification for the extraction process have appeared, the technologies are independent of each other and lack of cooperation and linkage, which makes it difficult to solve the core problems of multi-index accurate identification, real-time tracking of crystalline morphology and out-of-sync of process parameter closed-loop feedback under dynamic interference. Specifically, the existing foam denoising technology does not combine the dynamic characteristics of the interface, and there is still interface positioning deviation after denoising. The interface recognition technology cannot accurately capture the small fluctuations caused by stirring, resulting in inaccurate multi-index identification area demarcation. Multi-index tracking is mostly based on single modal image, and the information is one-sided and the quantitative error is large. The crystalline tracking algorithm is prone to particle degradation, and is disconnected with process parameter adjustment, and feedback lag can easily cause over-extraction or insufficient extraction. These problems are superimposed on each other, resulting in low intelligent control precision and slow response speed of the conventional artemisinin extraction process, which cannot meet the needs of industrial production for efficient, stable and accurate control, and has become a key bottleneck restricting the improvement of artemisinin extraction production efficiency and product quality.

[0004] Therefore, the present application is proposed. SUMMARY

[0005] The present application aims to provide an intelligent tracking and identification method for artemisinin extraction to solve the problems in the background art.

[0006] To solve the above technical problems, the present application provides an intelligent tracking and identification method for artemisinin extraction, comprising the following steps:

[0007] S1 collects visible light and near-infrared dual-mode images in the artemisinin extraction process, removes dynamic foam interference and preliminarily segments the extraction liquid layer interface through adaptive spatio-temporal joint denoising and interface segmentation processing;

[0008] S2 based on the dual-mode image processed in step S1, adopts an attention-enhanced U-Net network to dynamically identify the extraction liquid layer interface, the attention-enhanced U-Net network embeds a channel attention module and an edge enhancement branch, the channel attention module realizes feature enhancement by calculating the importance weight of each channel of the feature map, the weight calculation formula is , is the channel attention weight matrix, is the network input feature map, is the Sigmoid activation function, , , , is a learnable parameter, is the average pooling result of the feature map, is the maximum pooling result of the feature map; the edge enhancement branch uses an edge loss function:

[0009] ;

[0010] optimizes the interface contour extraction, wherein is the edge loss value, is the number of pixels, is the pixel real label, is the pixel predicted label, is the edge pixel weight coefficient; outputs the sub-pixel level interface coordinate and interface fluctuation amplitude data;

[0011] S3 fuses the dual-mode image features of step S1 and the interface region information of step S2, and synchronously identifies and quantifies the core indicators in the artemisinin extraction process through multi-modal collaborative tracking technology; the multi-modal collaborative tracking technology includes synchronous acquisition and target identification of key parameters and material state in each link of raw material pretreatment, extraction, purification and solvent recovery;

[0012] S4 based on the interface coordinates of step S2 and the indicator data of step S3, adopts an optimized particle filtering technology to track the change of crystalline morphology, and feeds back in real time to adjust the extraction process parameters to form a closed loop control; the optimized particle filtering technology includes dynamic filtering and trend prediction of multi-modal data, construction of cross-link parameter correlation database and deep learning prediction model; through the unique design of the attention-enhanced U-Net network and the multi-technology collaborative closed loop, the precise identification and dynamic control of the artemisinin extraction whole process are realized, the interface recognition precision and the indicator tracking accuracy are significantly improved, the parameter adjustment response time is shortened, and the stability and continuity of the extraction process are ensured.

[0013] Further, the S1 specifically comprises the following sub-steps: S11 adopts an industrial camera to collect visible light and near-infrared dual-mode images in the extraction kettle in real time; S12 obtains foam sub-blocks by sliding window traversal of the images, and analyzes the motion stability of each sub-block; S13 adopts motion compensation time domain filtering processing for stable foam sub-blocks, and adopts spatial correlation filtering processing for non-stable foam sub-blocks to remove dynamic foam interference; S14 preliminarily extracts the extraction liquid layering interface based on the filtered images through an adaptive threshold segmentation algorithm, and outputs the dual-mode images without foam interference and the preliminary interface region; for dynamic foam interference in the extraction process, type-specific filtering processing is adopted to improve the denoising pertinence, and the adaptive threshold segmentation is combined to realize preliminary interface extraction, thereby providing a high-quality data basis for subsequent accurate recognition and enhancing the anti-interference capability of image processing.

[0014] Further, the S2 further comprises: taking the preliminary interface region output by the step S1 as a key attention region of an attention-enhanced U-Net network, and strengthening extraction of texture features and grayscale features of the region through the network to output time sequence data of interface fluctuation amplitude changing with time; the key attention region is delimited to focus on interface recognition, core feature extraction is strengthened, and time sequence data of interface fluctuation is obtained, thereby providing a dynamic basis for subsequent multi-index collaborative analysis and further improving the accuracy and integrity of interface recognition.

[0015] Further, the S3 specifically comprises the following sub-steps: S31 extracts the raw material morphology features and crystallization density features in the visible light images, and extracts the extraction liquid purity features and impurity content features in the near-infrared images; S32 dynamically adjusts the weight distribution of the dual-mode features through an attention mechanism, and constructs a fusion feature map; S33 delimits the raw material area and the extraction liquid area based on the interface coordinates output by the step S2, adopts an improved YOLO-V5 network to identify core indexes in the two areas respectively, and forms a multi-index real-time data matrix; multi-dimensional information complementation is realized through dual-mode feature fusion, the pertinence of index identification is improved through partition identification, comprehensive index data is quickly output, reliable decision basis is provided for process parameter adjustment, and the efficiency and accuracy of multi-index tracking are improved.

[0016] Further, S4 specifically includes the following sub-steps: S41 determines the crystallization generation region based on the interface coordinates of step S2, and initializes the crystal particle tracking particle set using optimized particle filtering technology; S42 introduces a crystal growth model to correct particle weights, updates the particle set through a resampling strategy, and tracks the growth trajectory, particle size change, and morphological stability of the crystal particles; S43 combines the multi-index data from step S3, and when the index deviates from the preset range, generates a process parameter adjustment command and feeds it back to the extraction control system; S44 after parameter adjustment, repeats steps S1 to S4 to form a closed-loop optimization; by combining particle filtering optimization and the crystal growth model, the stability and accuracy of crystal morphology tracking are improved, real-time linkage between index changes and parameter adjustments is achieved, ensuring that the extraction process is always in the optimal state, and improving product quality and production efficiency.

[0017] Furthermore, in step S32, the weight adjustment of the attention mechanism is based on the recognition confidence of each modality feature. The higher the confidence, the greater the corresponding feature weight. The weight adjustment formula is:

[0018] ;

[0019] in Let m be the feature weights of the m-th mode. For the first The confidence level for the identification of each modality These correspond to the visible light mode and the near-infrared mode, respectively. By dynamically adjusting the mode weights based on the recognition confidence level, the fused features are made to better fit the actual recognition needs, improving the rationality and effectiveness of multimodal fusion and further enhancing the accuracy of core indicator recognition.

[0020] Furthermore, in S42, the resampling strategy adopts a hierarchical resampling method, which divides particles into different levels according to particle weights, extracts particles according to the level ratio to form a new particle set, and ensures the effective retention of high-weight particles; the hierarchical resampling strategy optimizes the particle set update, reduces particle degradation and depletion, ensures the continuity and accuracy of crystal morphology tracking, and provides accurate crystal state data for parameter feedback.

[0021] Furthermore, the core indicators in S33 include raw material consumption rate, extract purity, interface fluctuation amplitude, number of crystal particles, and impurity distribution density. By improving the detection head of the YOLO-V5 network, each indicator is quantitatively output. The quantification method of the core indicators is clarified, and the synchronous and accurate acquisition of multi-dimensional indicators is achieved, providing a complete basis for the comprehensive evaluation of the extraction process and parameter adjustment, and improving the comprehensiveness of intelligent control.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1、Dynamic foam denoising and preliminary interface segmentation: based on the type of filter strategy based on the difference in foam stability, accurate removal of dynamic foam and retention of effective features, combined with dual-mode adaptive segmentation optimization interface extraction, to avoid false interface interference. This technical means solves the interference problem of foam covering the core target, improves the image clarity and segmentation accuracy, lays a reliable data foundation for subsequent processing, and significantly enhances the anti-interference ability.

[0024] 2、Extraction liquid stratified interface dynamic identification: attention enhanced U-Net enhances features through channel attention module, optimizes contour through edge branch, and improves identification relevance through focus on key areas. This technical means overcomes interface fluctuation and distortion problems, realizes sub-pixel level positioning and fluctuation capture, reduces identification error, and defines accurate range for index tracking, greatly improves identification accuracy.

[0025] 3、Multi-modal multi-index collaborative tracking: attention mechanism dynamically adjusts dual-mode weight to realize information complementation, and combines partition strategy and improved network to identify indicators. This technical means breaks the single mode limitation, solves the problem of multi-index identification inaccuracy, realizes comprehensive collaborative tracking, improves identification reliability, and provides accurate basis for parameter adjustment.

[0026] 4、Crystalline morphology tracking and process parameter closed-loop feedback: the particle weight is corrected by the crystalline growth model, and the particle degradation is avoided by hierarchical resampling, and a closed-loop feedback mechanism is established to link parameter adjustment. This technical means solves the problems of unstable tracking and feedback lag, improves the accuracy of crystalline tracking, ensures timely parameter adjustment, and significantly improves product purity, yield and production stability.

[0027] 5、Overall coordination: The integrated closed-loop system realizes the deep coordination of the front-end supported by the subsequent and the subsequent feeding back to the front-end, and integrates the technical advantages to solve the fragmentation problem of existing technologies. This architecture realizes multi-link synchronous control, improves the overall control accuracy and stability, reduces manual intervention and invalid consumption, promotes the production to intelligent and adaptive upgrade, and supports industrial application. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A flowchart of an artemisinin extraction intelligent tracking and identification method. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely in the embodiments of the present application combined with the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0030] Please refer to Figure 1The application provides a technical solution: an artemisinin extraction intelligent tracking and identification method. The method constructs an integrated closed-loop system of "foam denoising-interface positioning-multi-index tracking-crystal feedback", and realizes intelligent and accurate control of the extraction process through the collaborative application of four core technical means. The embodiment is aimed at the scene of a pilot-scale conventional artemisinin extraction production line. The scene uses solvent extraction method, maintains a normal temperature and pressure environment in the extraction kettle, and stabilizes the temperature at 25 to 35 degrees Celsius, without needing to cope with extreme conditions such as high temperature, high pressure and strong corrosion. However, a large amount of dynamic foam is generated during the stirring operation in the extraction process. Such foam is easy to collapse and merge and is accompanied by rotary motion, which not only obscures the interface of the extraction liquid layer, but also blocks the initial crystalline particles, making it difficult to position the interface, identify the index and track the crystal. At the same time, the production process needs to monitor multiple core indexes such as raw material consumption, extraction purity and crystal growth at the same time, and adjust the process parameters in real time according to the changes in the indexes, so as to ensure the extraction efficiency and product quality.

[0031] In the prior art, the foam treatment mostly uses a general static denoising method, without considering the dynamic characteristics of artemisinin extraction foam, and the denoising is not complete and easy to damage the interface characteristics; the interface identification relies on traditional segmentation algorithms, which cannot accurately capture the interface fluctuations caused by stirring; the multi-index tracking is mostly based on single modal images, and the one-sided information leads to judgment deviation; the crystal tracking and parameter adjustment are disconnected, and the feedback lag easily causes over-extraction or insufficient extraction problems.

[0032] S1 dual-mode image acquisition and dynamic foam denoising and preliminary interface segmentation: In the pilot-scale conventional artemisinin extraction process, the dynamic foam generated by stirring is the core interference factor affecting image quality. The foam will obscure the interface of the extraction liquid layer and the initial crystalline particles, causing the data basis of subsequent interface identification and index tracking to be distorted. If the original image is directly used for subsequent processing, there will be problems of large interface positioning deviation and high crystal missed detection rate. Therefore, image data that can fully reflect the extraction state need to be collected first, and then the dynamic foam characteristics are accurately denoised, and the extraction liquid layer interface is preliminarily segmented, providing clean and effective data support for subsequent high-precision processing.

[0033] It should be noted here that: in the prior art, the methods for image denoising in the extraction process mainly include static spatial filtering, general time domain filtering, and the like. The extraction image denoising method based on Gaussian filtering can only process static noise and cannot cope with the dynamic motion characteristics of the foam in the artemisinin extraction process. The single threshold value interface is used in the disclosed document, and the preprocessing of the foam interference is not combined, resulting in low interface segmentation accuracy. Based on the motion stability difference of the dynamic foam, a type filtering strategy is adopted, and self-adaptive threshold segmentation is combined to realize accurate foam removal and interface preliminary extraction. The theoretical basis is that the stable sub-blocks and non-stable sub-blocks of the dynamic foam have significant feature differences in the time and space domains. The type processing can retain the effective features of the interface and the crystal while denoising. The self-adaptive threshold segmentation can adapt to the gray scale changes of the interface in different extraction stages. The specific technical means are as follows:

[0034] S11 dual-mode image acquisition: an industrial-grade visible light camera and a near-infrared camera are coaxially installed on the observation window at the top of the extraction kettle, with the lens facing the extraction area in the kettle to ensure that the collection range covers the entire liquid level and the raw material distribution area. The shooting wavelength range of the visible light camera is 400 to 760 nanometers, the frame rate is set to 30 frames per second, and the resolution is adjusted to 1920x1080 pixels, which is used to capture the appearance characteristics of the raw material and the crystal particles. The shooting wavelength range of the near-infrared camera is 780 to 2500 nanometers, the frame rate is consistent with that of the visible light camera, and the resolution is the same, which is used to extract the composition characteristics such as purity and impurity content inside the extraction liquid. The two cameras are triggered synchronously to collect images, each frame of image is attached with a time stamp to ensure the spatio-temporal alignment of the dual-mode images, and the collected image data is stored in the local hard disk of the industrial control computer in BMP format for subsequent algorithm calling.

[0035] S12 analysis of motion stability of foam sub-blocks: the synchronously collected visible light and near-infrared dual-mode images are processed by sliding window traversal. The sliding window size is set to 32x32 pixels, and the sliding window step is set to 16 pixels to ensure that there is a 50% overlap area between adjacent sliding windows to avoid missing the foam features. For each sliding window corresponding to the image sub-block, the center coordinates of the sub-block in the continuous 5 frames of images are extracted, and the variance of the coordinate change is calculated . The stability judgment threshold is set, when the coordinate change variance of the sub-block is less than the stability judgment threshold, it is determined that the sub-block is a stable foam sub-block, and the motion trajectory of such sub-block is not easy to collapse; when , it is determined that the sub-block is a non-stable foam sub-block, and such sub-block moves violently and is easy to merge or collapse.

[0036] S13 Subtype Dynamic Foam Denoising: For stable foam sub-blocks, a motion compensation temporal filtering algorithm is used for processing. First, the motion vector of the stable sub-block in the continuous 5 frames of images is calculated, and based on the motion vector, the position of the sub-block in the previous 4 frames of images is corrected to align the sub-block in the spatial and temporal domains; then the pixel values of the aligned 5 frames of sub-blocks are weighted and averaged, and the weight coefficient decreases with the frame number, and the weight coefficient of the latest frame is the largest to highlight the image information of the current frame, and the filtering formula is:

[0037] ;

[0038] wherein, is the pixel value of the filtered sub-block at coordinates , is the weight coefficient of the th frame, which satisfies:

[0039] ;

[0040] and , is the pixel value of the th frame at coordinates after motion compensation, and are the , direction offset amounts of the th frame sub-block relative to the current frame.

[0041] For non-stable foam sub-blocks, a spatial correlation filtering algorithm is used for processing. The gray correlation of each pixel in the sub-block with the surrounding 3x3 neighborhood pixels is calculated, and the correlation calculation formula is:

[0042] ;

[0043] wherein, is the gray correlation of the pixel at coordinates , is the gray value of the pixel, is the gray value of the neighborhood pixel. When is greater than the set correlation threshold , it is determined that the pixel is a foam pixel, and the median value of the neighborhood pixels is used to replace it; when , the original gray value of the pixel is retained to avoid damaging the effective features of the interface and the crystal.

[0044] S14 Adaptive threshold preliminary interface segmentation: Calculate the gray level histogram of the denoised bimodal image, respectively, and use the maximum between-class variance method (Otsu algorithm) to adaptively determine the segmentation threshold. For visible light image, the interface between raw material area and extractant area is mainly segmented, and the between-class variance of foreground (extractant area) and background (raw material area) in the gray level histogram is calculated. When the between-class variance is maximum, the corresponding gray value is the segmentation threshold of the visible light image ; for near-infrared image, the interface between extractant area and bottom area is mainly segmented, and the segmentation threshold is obtained in the same way . Based on and , the visible light and near-infrared images are binarized, the interface area is extracted, and the interface area of the bimodal image is logically ANDed to remove the false interface segmented by a single mode, to obtain the preliminary extractant stratification interface, and output the bimodal image without foam interference and the preliminary interface area coordinate data.

[0045] Example: This example is applied to a certain pilot-scale conventional artemisinin solvent extraction production line, the extraction kettle volume is 50 liters, paddle type stirrer is used, stirring speed is 100 revolutions per minute, extraction solvent is ethanol, and raw material is crushed artemisia leaf.

[0046] S11 Start the visible light camera and the near-infrared camera, and synchronously collect the images in the kettle. The collection time is from the 10th minute to the 60th minute after the extraction starts, and each frame of image is attached with a time stamp accurate to milliseconds. In the collected visible light image, the raw material area appears dark green, the extractant area appears light yellow, and the foam area is white and unevenly distributed; in the near-infrared image, the gray value of the extractant area is higher than that of the raw material area, and the gray value of the foam area is between the two.

[0047] S12 Set the sliding window size to 32x32 pixels, the step size to 16 pixels, and the stability judgment threshold pixel 2 . Perform sliding window traversal on a visible light image, extract the center coordinates of each sub-block in the continuous 5 frames of images, and calculate the coordinate change variance pixel 2 ( , stable sub-block), pixel 2 ( , unstable sub-block), to complete the stability classification of the foam sub-block.

[0048] S13 For the stable sub-block of pixel 2 , set the weight coefficients , , , , , the motion compensation time domain filtering formula is substituted to obtain the filtered sub-block image, and the foam interference is significantly reduced; and pixels 2 The correlation threshold of the unstable sub-blocks of the pixels is set , the foam pixels are calculated , the effective pixels are replaced by 3*3 neighborhood median value

[0049] S14 calculates the gray level histogram of the denoised visible light image to obtain the segmentation threshold corresponding to the maximum inter-class variance After the image is binarized, the interface between the raw material area and the extraction liquid area is extracted; the segmentation threshold of the near-infrared image is , the interface between the extraction liquid area and the kettle bottom area is extracted; the logical AND operation is performed on the bimodal interface area to remove the false interface of the raw material area edge in the visible light image, and the preliminary interface area coordinates are obtained pixels (vertical direction of the image), and the bimodal image without foam interference and the coordinate data are output.

[0050] The existing public documents use single Gaussian filtering to process the extraction image noise, without distinguishing the dynamic characteristics of the foam, resulting in uneven denoising effect of stable foam and unstable foam, and easy to blur the interface features; the public documents use fixed threshold to segment the interface, which cannot adapt to the gray level changes in different extraction stages.

[0051] The unique technical means of this step is to use type-specific filtering strategy based on the motion stability difference of the foam sub-blocks, to preserve the spatiotemporal continuity of the stable foam sub-blocks through motion compensation time domain filtering, and to accurately remove the unstable foam sub-blocks through spatial correlation filtering, while combining the bimodal adaptive threshold segmentation and logical AND operation to optimize the interface extraction. This technical means not only solves the problem of incomplete denoising of dynamic foam and easy damage to effective features in the prior art, but also realizes the preliminary accurate positioning of the interface, significantly reduces the interference of the denoised image, and greatly improves the clarity of the interface area, laying a high-quality data foundation for subsequent accurate interface recognition and multi-index tracking. Compared with the prior art, the foam removal rate is significantly improved, and the accuracy of the preliminary interface segmentation is obviously improved.

[0052] ​​S2 Dynamic recognition of the liquid-liquid interface based on attention-enhanced U-Net: After the preliminary segmentation in S1, although the liquid-liquid interface is obtained, due to the continuous disturbance of the stirred fluid, the interface will produce slight dynamic fluctuations. The preliminary segmented interface coordinates cannot accurately reflect the true interface position, and the interface contour is distorted with jagged edges. If the preliminary interface data is directly used for subsequent multi-index tracking, it will lead to the deviation of the region demarcation of index recognition, affecting the accuracy of index quantification. Therefore, a high-precision image segmentation network needs to be introduced to focus on optimizing the preliminary interface region, accurately recognize the dynamic fluctuation interface, and output the sub-pixel level interface coordinates and fluctuation amplitude data to provide accurate region demarcation basis for multi-index tracking.

[0053] It should be noted that: in the existing public technology, the network used for interface segmentation mainly includes traditional U-Net network, FCN network, etc. For example, the public document "Extraction interface segmentation method based on U-Net" does not introduce attention mechanism, which cannot focus on the core features of the interface region, and the recognition accuracy of the dynamic fluctuation interface is insufficient. The public document "Liquid interface segmentation algorithm based on deep learning" uses FCN network to segment the interface, which has weak edge extraction capability and serious interface contour distortion. In this step, an attention-enhanced U-Net network is used, which embeds a channel attention module and an edge enhancement branch to strengthen the interface feature extraction and contour optimization. The theoretical basis is that the texture and gray features of the liquid-liquid interface are significantly different from the surrounding region. The channel attention module can adaptively strengthen the weight of the interface feature channel and suppress irrelevant feature interference. The edge enhancement branch can optimize the smoothness and accuracy of the interface contour through a special loss function, realizing the sub-pixel level recognition of the dynamic fluctuation interface. The specific technical means are as follows:

[0054] S21 Network structure construction: The attention-enhanced U-Net network is improved based on the traditional U-Net architecture, which is divided into encoder, decoder and attention module. The encoder is composed of 4 convolution blocks, each of which contains 2 3x3 convolution layers, 1 batch normalization layer and 1 ReLU activation function. The convolution step is 2, which realizes the feature map downsampling and feature extraction. The decoder is composed of 4 deconvolution blocks, each of which contains 1 2x2 deconvolution layer, 2 3x3 convolution layers, 1 batch normalization layer and 1 ReLU activation function, which realizes the feature map upsampling and feature recovery. Between each corresponding level of the encoder and the decoder, a channel attention module is embedded, and an edge enhancement branch is embedded at the last level of the decoder.

[0055] S22 Channel attention module calculation: The channel attention module receives the feature map output by the encoder , the dimension of the feature map is , where is the channel number, is the height, The width is first globally averaged and globally maximum-pooled to obtain two feature vectors , respectively. ; Then the two feature vectors are input into two independent fully connected layers, the weight of the first fully connected layer is (dimension , is the dimension reduction coefficient), and the weight of the second fully connected layer is (dimension ), and is processed to obtain a feature vector ; Similarly, by weight (dimension ) and (dimension ), the is processed to obtain a feature vector ; After and are added and passed through the Sigmoid activation function, the channel attention weight matrix is obtained, and the formula is:

[0056] ;

[0057] wherein is the Sigmoid activation function, which maps the weight value to 0 to 1. Multiply with the original feature map channel by channel to obtain the enhanced feature map , which realizes the strengthening of the interface core features and the suppression of irrelevant features.

[0058] S23 edge enhancement branch optimization: the edge enhancement branch receives the feature map of the last level of the decoder, first reduces the channel number of to 1 through a 1x1 convolutional layer to obtain a single-channel feature map ; Then the Canny operator is used to extract the edge profile of to obtain an edge mask ; an edge loss function is constructed, and by optimizing the loss function, the extraction accuracy of the interface profile is improved, and the formula of the edge loss function is:

[0059] ;

[0060] wherein is the total number of pixels of the single-channel feature map , is the first​​ real edge label of a pixel (1 for edge pixel, 0 for non-edge pixel), predicted edge label of the th pixel, edge weight coefficient of the th pixel, the value of edge region pixel is greater than that of non-edge region to highlight the importance of edge optimization.

[0061] S24 interface dynamic recognition and data output: the bubble-free bimodal image output by S1 is channel spliced as the input of the attention-enhanced U-Net network, and the preliminary interface region obtained by S1 is the key focus area of the network. In the training and inference process, the loss value of this area is given a higher weight. The network output is a binary interface segmentation map, and the sub-pixel level coordinates of the interface are extracted through a sub-pixel edge detection algorithm; based on the interface coordinates of the last 30 frames of images, the interface offset of each frame of image and the previous frame of image is calculated, and the time series data of the interface fluctuation amplitude changing with time is obtained, and the sub-pixel level interface coordinates and fluctuation amplitude time series data are output.

[0062] Example: continue the pilot production line example of S1, the training data set of the attention-enhanced U-Net network uses 10,000 frames of bimodal images collected by the production line, of which 8,000 frames are used as the training set and 2,000 frames are used as the verification set. The image resolution is uniformly adjusted to 1920x1080 pixels.

[0063] S21 constructs an attention-enhanced U-Net network, and the four convolution blocks of the encoder output feature maps with channel numbers of 64, 128, 256 and 512 respectively, and the four deconvolution blocks of the decoder output feature maps with channel numbers of 256, 128, 64 and 32 respectively, and the dimension reduction coefficients , the edge weight coefficient is set to 2 in the edge region and 1 in the non-edge region.

[0064] S22 performs global average pooling on the feature map (channel number , height , width ) of a certain frame of input image to obtain , , and global maximum pooling to obtain , ; the weight dimensions of the fully connected layer are , , , , , , ,​ ; the channel attention weight matrix is obtained by substituting the sigmoid function , , the channel attention weight matrix is obtained by substituting the sigmoid function , wherein the channel weight value corresponding to the interface feature is 0.8-0.9, and the channel weight value corresponding to the irrelevant feature is 0.1-0.2; and is multiplied by channel by channel to obtain the enhanced feature map , and the interface feature is significantly enhanced.

[0065] The last level output feature map of the S23 decoder (channel number 32) is reduced to a single channel feature map through a 1x1 convolutional layer ; the edge profile of is extracted using a Canny operator to obtain an edge mask ; the edge loss function is substituted, and the value of a certain frame image is calculated to be 0.05, and after multiple rounds of iterative optimization, the value is reduced to below 0.01, and the smoothness of the interface profile is significantly improved.

[0066] S24 inputs the double-modal image output by S1 after splicing into the network, and sets the key attention area as the preliminary interface coordinates within a range of 50 pixels around the pixel, the network outputs an interface segmentation map, and through a sub-pixel edge detection algorithm, the sub-pixel level interface coordinates are extracted as pixels; based on the coordinate data of 30 consecutive frames of images, the interface fluctuation amplitude is calculated to be between 0.1-0.3 pixels, and the coordinate and fluctuation amplitude time series data are output.

[0067] The existing public documents use the traditional U-Net network to segment and extract the interface, do not introduce the attention mechanism, cannot focus on the core features of the interface, and have low recognition accuracy for dynamic fluctuating interfaces, and the edge profile is distorted; the liquid interface segmentation algorithm based on deep learning in the public literature uses the FCN network, lacks a special edge optimization branch, and the smoothness of the interface profile is insufficient.

[0068] The unique technical means of this step is to embed the channel attention module into the encoder and decoder levels of the U-Net network, to strengthen the interface features by adaptively calculating channel weights, and to add an edge enhancement branch at the end of the decoder to optimize the interface contour through a special edge loss function, and to take the preliminary interface area as the key attention area to improve the identification relevance. This technical means overcomes the defects of the existing network in the identification accuracy of dynamic fluctuating interface and edge distortion, realizes sub-pixel level interface positioning and fluctuation tracking, greatly improves the accuracy of interface identification, and more comprehensively captures the fluctuation amplitude. Compared with the prior art, the interface identification error is significantly reduced, which provides a precise basis for subsequent multi-index tracking area demarcation, and the area deviation of index identification is significantly reduced.

[0069] S3 Multi-modal multi-index collaborative tracking: The control of the artemisinin extraction process needs to comprehensively grasp multiple core indexes such as raw material consumption, extract purity, and crystal growth. Single modal image can only reflect part of the information, for example, visible light image is good at capturing appearance features, near-infrared image is good at reflecting component features. If only based on single modal for index identification, it will lead to one-sided information and inaccurate index quantification. At the same time, without combining the identification of precise interface area, it will lead to fuzzy index tracking range, further exacerbating the deviation. Therefore, it is necessary to fuse the feature advantages of double modal images, combine the precise interface coordinates obtained in S2 step to demarcate the identification area, realize the synchronous collaborative tracking of multiple indexes, and output comprehensive and accurate index data to provide decision basis for process parameter adjustment.

[0070] It should be noted that: in the existing public technology, multi-index identification mostly uses single modal image or simple modal splicing, for example, the extraction index identification method based on visible light image proposed in the public document can only identify the shape of raw materials and the number of crystals, and cannot obtain component indexes such as extract purity; the application of multi-modal image fusion in industry detection in the public document uses a simple pixel-level fusion method, without considering the weight difference of different modal features, and the fusion effect is poor. This step dynamically adjusts the feature weights of double modal based on attention mechanism, and performs partition identification based on precise interface area. The theoretical basis is that different modal images have different contributions in different index identification, and attention mechanism can adaptively allocate weights to highlight effective modal features; based on the demarcation of raw material area and extract area by interface coordinates, the index identification is more targeted, the interference of irrelevant area is reduced, and the quantification accuracy is improved. The specific technical means are as follows:

[0071] S31 Double modal feature extraction: The foam-free visible light image and near-infrared image output by S1 step are subjected to feature extraction respectively. For the visible light image, ResNet50 network is used as the feature extraction backbone, the last fully connected layer is removed, and the convolution layer part is retained to extract a feature map with a dimension of ​This feature map primarily reflects the appearance characteristics such as raw material morphology and crystal density; for near-infrared images, the same ResNet50 network structure is used to extract dimensions of... Feature map This feature map primarily reflects the component characteristics of the extract, such as purity and impurity content.

[0072] S32 Attention Mechanism Dynamic Weight Fusion: This mechanism calculates the recognition confidence of the bimodal feature map to dynamically adjust the weights. For appearance-related indicators such as raw material consumption rate and crystal particle count, it utilizes feature maps from visible light images. Calculate confidence level For classification indicators such as extract purity and impurity content, feature maps based on near-infrared images are used. Calculate confidence level The confidence level is calculated based on the response intensity of the feature map. A higher response intensity results in a higher confidence level. The formula for calculating the confidence level is:

[0073] ;

[0074] in, Time corresponds to visible light mode ( ), Time corresponds to near-infrared mode ( ), For feature map In coordinates eigenvalues ​​at that location For feature map The largest eigenvalue.

[0075] Fusion weights for bimodal features calculated based on confidence level. The formula is:

[0076] ;

[0077] in, The weights of visible light features, The weights of the near-infrared features satisfy... .Will and The fused feature map is obtained by multiplying each feature map by its corresponding weight and then summing the results. .

[0078] S33 Improves YOLO-V5 Partition Index Recognition: Based on the sub-pixel level interface coordinates output from step S2, two recognition regions are defined: the raw material region (the area above the interface) and the extract region (the area below the interface). The detection head of the YOLO-V5 network is improved by adding two dedicated detection branches, corresponding to index recognition in the raw material region and the extract region, respectively.

[0079] The raw material area detection branch targets two indicators: raw material consumption rate and impurity distribution density, quantified by detecting changes in the area of ​​the raw material region and the proportion of areas with uneven grayscale. The extract liquid area detection branch targets three indicators: extract liquid purity, interface fluctuation amplitude, and number of crystal particles, quantified by the characteristic response intensity of the extract liquid region, the fluctuation range of interface coordinates, and the number of detection frames for crystal particles. The fused feature map... The improved YOLO-V5 network is input, and the network outputs quantitative data of five core indicators, forming a multi-indicator real-time data matrix.

[0080] Example: Continuing with the pilot production line example above, the ResNet50 network is initialized with pre-trained weights and fine-tuned on the dataset of this production line. The training dataset of the improved YOLO-V5 network contains 5000 labeled fused feature maps, with the annotations being the true values ​​of five core metrics.

[0081] S31 inputs the bubble-free visible light image output by S1 into the ResNet50 network to extract the feature map. (dimension) The feature map shows significantly higher characteristic response intensity in the raw material area and crystal grain region than in other areas; similarly, the feature map is extracted from the near-infrared image. The characteristic response intensity is high in the extract region.

[0082] S32 calculates the visible light modal confidence level of appearance-related indicators. Regarding the raw material consumption rate index, of Calculations yielded Near-infrared modal confidence Substituting into the weight formula, we get , , fusion feature map Regarding the purity of the extract as a component classification indicator, , Weight , , .

[0083] S33 is based on the interface coordinates output by S2. Pixels define the raw material area as The pixel area, the extract area is The improved YOLO-V5 network's feed area detection branch detects the feed area and compares it with the initial feed area to obtain the feed consumption rate; it also detects the proportion of uneven grayscale regions in the feed area to obtain the impurity distribution density. The extract region detection branch... The purity of the extract liquid is calculated based on the characteristic response intensity of the extract liquid area; the interface fluctuation amplitude is calculated by the fluctuation range of the continuous frame interface coordinates; and the number of crystalline particles is obtained by detecting the number of detection boxes of the crystalline particles. A multi-index data matrix at a certain moment is output: raw material consumption rate 0.02 g / s, extract liquid purity 85%, interface fluctuation amplitude 0.2 pixels, number of crystalline particles 120, and impurity distribution density 0.05.

[0084] The existing disclosure documents only identify extraction indicators based on visible light images, cannot obtain classification indicators, and the information is one-sided; the application of multi-modal image fusion in industrial detection in the disclosed documents uses simple pixel-level fusion, does not dynamically adjust the modal weight, and the fused features cannot adapt to the recognition needs of different indicators.

[0085] The unique technical means of this step is to dynamically calculate the recognition confidence of the dual-modal features based on the indicator type, allocate the fusion weight through the attention mechanism, and combine the accurate interface coordinates to demarcate the raw material area and the extract liquid area, and use the improved YOLO-V5 network to identify the indicators in the partition. This technical means realizes the synchronous and collaborative tracking of appearance and classification indicators, overcomes the defects of single-modal information being one-sided and fixed weight fusion being poor, greatly improves the comprehensiveness and accuracy of multi-indicator recognition, significantly reduces the indicator quantization error compared to the existing technology, provides a comprehensive and reliable decision basis for process parameter adjustment, and significantly enhances the pertinence of parameter adjustment.

[0086] S4 Particle filtering optimized crystalline morphology tracking and process parameter closed-loop feedback: The growth state of the crystalline morphology directly determines the product purity and yield of artemisinin, and it is necessary to track the growth trajectory, particle size change and morphology stability of the crystalline particles in real time. The existing tracking algorithm is prone to particle degradation and depletion, which leads to tracking interruption or deviation; at the same time, the crystalline tracking data is disconnected with the process parameter adjustment, which cannot optimize the process conditions in time according to the crystalline state, affecting the product quality. Therefore, it is necessary to optimize the particle filtering algorithm, improve the tracking accuracy combined with the crystalline growth law, establish a closed-loop feedback mechanism between the tracking data and the process parameter adjustment, realize the dynamic control of crystalline growth, and ensure the optimal state of the extraction process.

[0087] It should be noted here that: in the prior art, crystal tracking is mostly based on traditional particle filtering algorithms, such as the crystal tracking method based on standard particle filtering proposed in the public document, which does not combine the crystal growth model, the particle weight update is unreasonable, and degeneration is prone to occur; the application of particle filtering in particle tracking in the public document adopts a random resampling strategy, resulting in loss of high-weight particles and poor tracking stability. This step introduces a crystal growth model to correct the particle weight and uses a hierarchical resampling strategy to optimize the particle set. The theoretical basis is as follows: the growth of crystal particles follows a specific rule, and the growth model can accurately predict the particle state, correct the weight, and improve the tracking accuracy; hierarchical resampling can retain high-weight particles according to the weight proportion, reduce degeneration and impoverishment, and ensure tracking continuity; the establishment of a linkage feedback between tracking data and process parameters can realize dynamic optimization of crystal growth. The specific technical means are as follows:

[0088] S41 crystal region determination and particle set initialization: based on the sub-pixel level interface coordinates output in the S2 step, the crystal generation region is determined as the range of 50 to 200 pixels from the interface in the extract liquid area (this area is the main generation area of artemisinin crystals). A uniform distribution sampling strategy is used to initialize particles in the crystal generation area, each particle representing a potential crystal particle center coordinate , , the initial weight of the particle , ensuring uniform distribution of the initial weight.

[0089] S42 particle state prediction and weight correction: the next frame state of the particle is predicted based on the crystal growth model. The crystal growth model uses a spherical growth model, which assumes that the radius of the crystal particle increases linearly with time, and the radius growth formula is , where is the crystal radius at time t, is the initial radius, is the growth rate constant. Based on the current coordinates and radius of the particle, the coordinates of the next frame are predicted , where and are random offsets based on fluid flow characteristics, following a Gaussian distribution , is the standard deviation of the offset. The observation likelihood of the particle is calculated based on the fusion feature map output in the S3 step

[0090] , the feature vector of the region corresponding to each particle is extracted, and the cosine similarity with the template feature vector of the crystal particle is calculated. The higher the similarity, the greater the observation likelihood. The observation likelihood formula is:

[0091] ​ ;

[0092] wherein, is the observation data of the frame (fusion feature map), is the predicted state of the th particle, is the adjustment coefficient, is the cosine similarity of the region feature corresponding to the particle and the template feature.

[0093] The particle weight is corrected in combination with the crystal growth model, and the correction formula is:

[0094] ;

[0095] wherein, is the prior probability of the crystal growth model, is the predicted crystal radius of the particle, if the growth rule of is met, , otherwise the weight of the particle that does not meet the growth rule is inhibited.

[0096] S43 stratified resampling and state estimation: calculating the normalization factor of particle weight , the particle weight is normalized . Set the resampling threshold , when the normalized weight of the particle , it is determined as a low weight particle and needs to be resampled.

[0097] The stratified resampling strategy is adopted, and the particles are sorted according to the normalized weight from large to small, and divided into levels, and the sum of the particle weights in each level is equal. In each level, particles are randomly extracted according to the proportion of particle weight, and the number of extraction is proportional to the proportion of particle weight in the level, which ensures that high weight particles are fully retained, and low weight particles are replaced by the copies of high weight particles, forming a new particle set.

[0098] Based on the new particle set, the state of the crystal particle is estimated by using the weighted average method, and the center coordinate estimation value is: ;

[0099] The radius estimation value is: ;

[0100] The growth trajectory, particle size change and morphology stability data (morphology stability is measured by the variance of radius change) of the crystal particle are output.

[0101] S44 Process parameter closed-loop feedback: Set the preset range of five core indicators. When a certain indicator in the multi-index data matrix output by S3 deviates from the preset range, generate process parameter adjustment instructions in combination with the crystallization state data output by S43. For example, when the purity of the extraction liquid is lower than the preset lower limit and the crystalline particle size increases slowly, generate instructions to increase the stirring speed and improve the extraction temperature; when the number of crystalline particles is too large and the impurity distribution density is higher than the preset upper limit, generate instructions to reduce the stirring speed and adjust the solvent ratio.

[0102] The adjustment instructions are sent to the extraction control system. After the control system executes parameter adjustment, the extraction state in the extraction kettle changes. The changed state is re-acquired by the camera in S1 to enter the next round of "denoising-interface recognition-indicator tracking-crystallization feedback" closed-loop process, realizing self-adaptive optimization.

[0103] Example: Continue the pilot production line example in the previous section. The crystallization generation area is set to the extraction liquid area in the extraction kettle to The range of pixels (based on interface coordinates pixels), the initial number of particles , the initial weight , the initial radius of the crystal pixels, the growth rate constant pixels / frame, the adjustment coefficient , the offset standard deviation pixels, the resampling threshold , the number of levels .

[0104] S41 uniformly samples 200 particles in the range of to pixels. The particle coordinates are distributed in this area, and the initial weight is 0.005.

[0105] S42 based on the crystallization growth model, predicts the next frame radius of a certain particle pixels, coordinates ( , , subject to ); extract the feature vector of the area corresponding to the particle, and the cosine similarity of the template feature vector , substitute into the observation likelihood formula to get ; since conforms to the growth law, , the corrected weight .

[0106] S43 calculates the normalization factor , the normalized weight ( ), high-weight particles; 200 particles are sorted by weight and divided into 5 levels, and the sum of the weight of each level is 0.2; particles are extracted in proportion in each level, and high-weight particles are extracted multiple times to form a new particle set; the estimated center coordinates of the crystalline particles are calculated based on the new particle set , radius 5.3 pixels, radius variance 0.02, good shape stability.

[0107] S44 sets the extraction liquid purity preset range to 88% to 95%, generates an adjustment instruction when the extraction liquid purity output by S3 is 85% (lower than the lower limit) and the crystalline particle size grows slowly (from 5 pixels to 5.3 pixels, growth rate 0.3 pixels / 10 frames): the stirring speed is increased from 100 rpm to 110 rpm, and the extraction temperature is increased from 30°C to 32°C; after the instruction is sent to the control system for execution, the new extraction state is collected by the camera of S1, and the next round of closed-loop process is entered.

[0108] The existing public documents use traditional particle filtering to track the crystalline, without combining the crystalline growth model, the particle weight update lacks theoretical basis, the tracking accuracy is low, and particle degradation is easy to occur; the particle filtering in the application of particle tracking in the public literature uses random resampling, which leads to the loss of high-weight particles, poor tracking stability, and no closed-loop feedback with process parameters is established.

[0109] The unique technical means of this step is to introduce a spherical crystalline growth model to correct the particle weight, ensure that the particle state conforms to the crystalline growth rule, use a layered resampling strategy to extract particles by weight level, retain high-weight particles, and establish a closed-loop feedback mechanism between the crystalline tracking data and the process parameter adjustment. This technical means overcomes the degradation and depletion problems of existing particle filtering, improves the accuracy and stability of crystalline morphology tracking, realizes real-time linkage optimization of crystalline state and process parameters, significantly improves the success rate of crystalline tracking, speeds up the response speed of process parameter adjustment, significantly improves the purity and yield of artemisinin, and greatly enhances the stability of the production process.

[0110] This method constructs an integrated closed-loop system of "foam denoising-interface positioning-multi-index tracking-crystalline feedback" through the synergistic linkage of S1 to S4 steps, and each step supports and promotes each other: the foam denoising and preliminary interface segmentation of S1 provide a clean data basis for the accurate interface recognition of S2; the sub-pixel level interface coordinates of S2 define the accurate range for the partition index recognition of S3; the multi-index data of S3 provide comprehensive basis for the crystalline tracking and parameter feedback of S4; the parameter adjustment results of S4 are fed back to the front-end steps through the closed loop to realize adaptive optimization.

[0111] Compared with the superposition application of existing single technology, the synergistic system of the method not only solves multiple technical pain points such as dynamic foam interference, interface fluctuation, one-sided multi-index information, crystallization tracking deviation, feedback lag, etc., but also produces unexpected synergistic effects, significantly improves the intelligent control precision, response speed and stability of the extraction process, provides reliable technical support for efficient and high-quality production of the pilot-level conventional artemisinin extraction production line, and is easy to migrate to industrial production lines.

Claims

1. A method for intelligent tracking and identification of artemisinin extraction, characterized in that: Includes the following steps: S1 acquires visible and near-infrared dual-modal images of artemisinin extraction process. Through adaptive spatiotemporal joint denoising and interface segmentation processing, dynamic foam interference is removed and the layered interface of the extract is initially segmented. S2, based on the bimodal image processed in step S1, uses an attention-enhanced U-Net network to dynamically identify the layered interface of the extract. The attention-enhanced U-Net network embeds a channel attention module and an edge enhancement branch. The channel attention module enhances features by calculating the importance weights of each channel in the feature map. The weight calculation formula is as follows: ； wherein is a channel attention weight matrix, is a network input feature map, is a Sigmoid activation function, , , , is a learning parameter, is a feature map average pooling result, is a feature map maximum pooling result; the edge enhancement branch passes through an edge loss function: ; Optimize the interface profile extraction, wherein is an edge loss value, is a pixel point number, is the true label of the first pixel, is the predicted label of the first pixel, is the first edge pixel weight coefficient; output sub-pixel level interface coordinates and interface fluctuation amplitude data; S3 integrates the dual-modal image features of step S1 with the interface region information of step S2, and synchronously identifies and quantifies the core indicators in the artemisia annua extraction process using multimodal collaborative tracking technology. This multimodal collaborative tracking technology includes the synchronous acquisition and target identification of key parameters and material states in each stage of raw material pretreatment, extraction, purification, and solvent recovery; specifically, it includes: S31, extracting features of raw material morphology and crystalline density in the visible light image output by S1 and features of extract purity and impurity content in the near-infrared image ; S32, dynamically adjust the weight distribution of the bimodal features through the attention mechanism, and construct a fusion feature map ; S33. Based on the interface coordinates output in step S2, delineate the raw material area and the extract area, identify the core indicators in the two areas respectively, and form a multi-indicator real-time data matrix. In S32, the weight adjustment of the attention mechanism is based on the recognition confidence of each modality feature. The higher the confidence, the greater the corresponding feature weight. The weight adjustment formula is: ; wherein is a feature weight of the first modal, is a recognition confidence of the first modal, respectively corresponding to the visible light modal and the near-infrared modal; is a weight of the visible light feature, is a weight of the near-infrared feature, satisfying ; and the fused feature map is: ; Based on the interface coordinates from step S2 and the index data from step S3, S4 uses optimized particle filtering technology to track changes in crystal morphology, providing real-time feedback to adjust extraction process parameters and form a closed-loop control. The optimized particle filtering technology includes dynamic filtering and trend prediction of multimodal data, and the construction of a cross-stage parameter association database and a deep learning prediction model.

2. The intelligent tracking and identification method for artemisinin extraction as described in claim 1, characterized in that: S1 specifically includes the following sub-steps: S11 uses an industrial camera to acquire real-time images of the extraction vessel in both visible and near-infrared modes; S12 obtains foam sub-blocks by traversing the image through a sliding window and analyzes the motion stability of each sub-block; S13 uses motion-compensated time-domain filtering for stable foam sub-blocks and spatial correlation filtering for unstable foam sub-blocks to remove dynamic foam interference. S14 uses an adaptive threshold segmentation algorithm to initially extract the layered interface of the extractant based on the filtered image, and outputs a dual-modal image and preliminary interface region without foam interference.

3. The intelligent tracking and identification method for artemisinin extraction as described in claim 1, characterized in that: S2 further includes: taking the preliminary interface region output in step S1 as the key focus area of ​​the attention-enhancing U-Net network, and using the network to enhance and extract the texture features and grayscale features of the region, and outputting time-series data of the interface fluctuation amplitude changing over time.

4. The intelligent tracking and identification method for artemisinin extraction as described in claim 1, characterized in that: Step S33 employs an improved YOLO-V5 network to identify core indicators in two regions: the improved YOLO-V5 network involves modifying the detection head of the YOLO-V5 network and adding two dedicated detection branches, corresponding to the indicator identification in the raw material region and the extract region, respectively.

5. The intelligent tracking and identification method for artemisinin extraction as described in claim 1, characterized in that: S4 specifically includes the following sub-steps: S41 determines the crystallization region based on the interface coordinates of step S2, and initializes the crystallization particle tracking particle set using optimized particle filtering technology; S42 introduces a crystal growth model to correct particle weights and updates the particle set through a resampling strategy to track the growth trajectory, particle size change, and morphological stability of crystal particles. S43 combines the multi-index data from step S3. When an index deviates from the preset range, a process parameter adjustment instruction is generated and fed back to the extraction control system. After adjusting parameter S44, repeat steps S1 to S4 to form a closed-loop optimization.

6. The intelligent tracking and identification method for artemisinin extraction as described in claim 5, characterized in that: In S42, the resampling strategy adopts a hierarchical resampling method, which divides particles into different levels according to particle weights, and extracts particles according to the level ratio to form a new particle set, ensuring the effective retention of high-weight particles.

7. The intelligent tracking and identification method for artemisinin extraction as described in claim 4, characterized in that: The core indicators in S33 include raw material consumption rate, extract purity, interface fluctuation amplitude, number of crystal particles, and impurity distribution density. Each indicator is quantified and output by improving the detection head of the YOLO-V5 network.

Images