Intelligent tracking and identification method for artemisinin extraction

Abstract

The invention discloses an artemisinin extraction intelligent tracking and identification method, and relates to the technical field of image data processing. According to the method, an integrated closed-loop system is constructed through cooperation of four core technical means: based on classified filtering of foam stability difference and bimodal adaptive segmentation, dynamic foam is accurately removed, and preliminary interface extraction is optimized; the attention enhancement U-Net network is embedded into a channel attention module and an edge enhancement branch to realize accurate positioning and fluctuation tracking of a layered interface; the attention mechanism dynamically adjusts the bimodal weight, and completes multi-index cooperative tracking in combination with a partitioning strategy and an improved network; a crystal growth model is introduced to correct particle weight, a tracking effect is optimized by adopting layered resampling, and a closed-loop feedback mechanism of a crystal state and process parameters is established. All technical means are deeply coordinated, intelligent and accurate management and control of the whole extraction process are achieved, and reliable technical support is provided 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 there are single technical means such as image denoising, interface segmentation and index identification for the extraction process, 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 demarcation of multi-index identification area. Multi-index tracking is mostly based on single modal image, which is one-sided and has large quantitative error. The crystalline tracking algorithm is prone to particle degradation, and is disconnected with process parameter adjustment, which may cause over-extraction or insufficient extraction due to feedback lag. 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 high-efficiency, stable and accurate control in industrial production, and becomes 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, which comprises the following steps: S1: Collecting visible light and near-infrared dual-modal images in the artemisinin extraction process, removing the dynamic foam interference and preliminarily segmenting the interface of the extraction liquid layer through adaptive spatio-temporal joint denoising and interface segmentation processing; S2 based on the bimodal image processed in step S1, using attention enhanced U-Net network to extract liquid interface for dynamic identification, attention enhanced U-Net network embedded channel attention module and edge enhancement branch, channel attention module through the calculation of the importance weight of each channel of feature map realizes feature enhancement, the weight calculation formula is , , , , , , , , , , , , , , , , , , S3 fusion step S1 bimodal image features and step S2 interface region information, through multi-modal collaborative tracking technology synchronous identification and quantification of the core indicators in the process of artemisinin extraction; Multi-modal collaborative tracking technology includes the key parameters and material state of raw material pretreatment, extraction, purification, solvent recovery each link are synchronous acquisition and target identification;

[0007] Further, S1 specifically includes the following sub-steps: S11 uses an industrial camera to acquire visible light and near-infrared dual-modal images of the extraction vessel in real time; S12 obtains foam sub-blocks by traversing the images through a sliding window and analyzes the motion stability of each sub-block; S13 applies motion compensation temporal filtering to stable foam sub-blocks and spatial correlation filtering to unstable foam sub-blocks to remove dynamic foam interference; S14, based on the filtered image, uses an adaptive threshold segmentation algorithm to initially extract the layered interface of the extraction liquid, outputting a dual-modal image without foam interference and a preliminary interface area; for dynamic foam interference during the extraction process, categorized filtering is used to improve the denoising targeting, and adaptive threshold segmentation is combined to achieve preliminary interface extraction, providing a high-quality data foundation for subsequent accurate identification and enhancing the anti-interference capability of image processing.

[0008] Furthermore, S2 also includes: using 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 the extraction of texture features and grayscale features of this region, outputting time-series data of interface fluctuation amplitude changing over time; by defining the key focus area to focus on interface recognition, the core feature extraction is enhanced, and time-series data of interface fluctuation is obtained, providing a dynamic basis for subsequent multi-indicator collaborative analysis, and further improving the accuracy and completeness of interface recognition.

[0009] Further, S3 specifically includes the following sub-steps: S31 Extracting raw material morphology and crystal density features from visible light images, and extracting extract purity and impurity content features from near-infrared images; S32 Dynamically adjusting the weight allocation of dual-modal features through an attention mechanism to construct a fused feature map; S33 Delineating the raw material area and extract area based on the interface coordinates output in step S2, and using an improved YOLO-V5 network to identify core indicators in the two areas respectively, forming a multi-indicator real-time data matrix; Achieving multi-dimensional information complementarity through dual-modal feature fusion, and improving the targeting of indicator identification by combining partition recognition, quickly outputting comprehensive indicator data, providing a reliable decision basis for process parameter adjustment, and improving the efficiency and accuracy of multi-indicator tracking.

[0010] 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.

[0011] 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: ; 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.

[0012] 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.

[0013] 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.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. Dynamic Bubble Denoising and Preliminary Interface Segmentation: Based on a categorized filtering strategy using differences in bubble stability, dynamic bubbles are accurately removed while retaining effective features. Combined with bimodal adaptive segmentation to optimize interface extraction, false interface interference is avoided. This technique addresses the problem of bubbles obscuring core targets, improving image clarity and segmentation accuracy, laying a reliable data foundation for subsequent processing, and significantly enhancing anti-interference capabilities.

[0015] 2. Dynamic Recognition of Extraction Liquid Layered Interface: Attention-Enhanced U-Net strengthens features through channel attention modules and optimizes contours through edge branches, while focusing on key areas to improve recognition accuracy. This technique overcomes interface fluctuations and distortion issues, achieving sub-pixel-level positioning and fluctuation capture, reducing recognition errors, defining a precise range for indicator tracking, and significantly improving recognition accuracy.

[0016] 3. Multimodal and Multi-indicator Collaborative Tracking: An attention mechanism dynamically adjusts the weights of the two modalities to achieve information complementarity, combined with a partitioning strategy and an improved network for targeted indicator identification. This technique overcomes the limitations of single-modality tracking, solves the problem of inaccurate multi-indicator identification, achieves comprehensive collaborative tracking, improves identification reliability, and provides accurate basis for parameter adjustment.

[0017] 4. Crystallization Morphology Tracking and Closed-Loop Feedback of Process Parameters: The crystal growth model corrects particle weights, stratified resampling avoids particle degradation, and a closed-loop feedback mechanism is established to link parameter adjustments. This technique solves the problems of unstable tracking and delayed feedback, improves the accuracy of crystallization tracking, ensures timely parameter adjustments, and significantly improves product purity, yield, and production stability.

[0018] 5. Overall Collaboration: The integrated closed-loop system achieves deep collaboration between upstream and downstream processes, integrating technological advantages to solve the fragmentation problem of existing technologies. This architecture enables synchronous control of multiple stages, improving the overall control accuracy and stability, reducing manual intervention and ineffective consumption, promoting the intelligent and adaptive upgrade of production, and supporting industrial applications. Attached Figure Description

[0019] Figure 1 This is a flowchart of an intelligent tracking and identification method for artemisinin extraction. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1This invention provides a technical solution: an intelligent tracking and identification method for artemisinin extraction. This method constructs an integrated closed-loop system of "foam denoising, interface positioning, multi-index tracking, and crystallization feedback," achieving intelligent and precise control of the extraction process through the synergistic application of four core technologies. This embodiment targets a pilot-scale conventional artemisinin extraction production line scenario. This scenario employs solvent extraction, maintaining a normal temperature and pressure environment within the extraction vessel, with the temperature stable between 25 and 35 degrees Celsius. It does not require handling extreme conditions such as high temperature, high pressure, and strong corrosion. However, the stirring operation during extraction generates a large amount of dynamic foam. This foam is prone to collapse and is accompanied by rotational motion, not only obscuring the layered interface of the extract but also blocking initial crystallization particles, making interface positioning, index identification, and crystallization tracking difficult. Simultaneously, the production process needs to monitor multiple core indicators such as raw material consumption, extraction purity, and crystal growth, and adjust process parameters in real time based on changes in these indicators to ensure extraction efficiency and product quality.

[0022] In existing technologies, foam treatment often uses general static denoising methods, which do not take into account the dynamic characteristics of artemisinin extraction foam. This results in incomplete denoising and easy damage to interface features. Interface recognition relies on traditional segmentation algorithms, which cannot accurately capture interface fluctuations caused by stirring. Multi-index tracking is mostly based on single-modal images, and the one-sided information leads to judgment bias. Crystallization tracking is disconnected from parameter adjustment, and the feedback lag can easily lead to over-extraction or insufficient extraction.

[0023] S1 Dual-Modal Image Acquisition, Dynamic Foam Denoising, and Preliminary Interface Segmentation: During the pilot-scale conventional artemisinin extraction process, dynamic foam generated by stirring is the core interference factor affecting image quality. Foam can obscure the layered interfaces of the extract and initial crystallization particles, leading to distortion of the data foundation for subsequent interface recognition and indicator tracking. If the original images are used directly for subsequent processing, problems such as large interface positioning deviations and high crystallization false negative rates will occur. Therefore, it is necessary to first acquire image data that can comprehensively reflect the extraction state, then perform precise denoising based on the characteristics of dynamic foam, and simultaneously perform preliminary segmentation of the layered interfaces of the extract to provide clean and effective data support for subsequent high-precision processing.

[0024] It should be noted that existing publicly available technologies for denoising extraction process images mainly include static spatial filtering and general temporal filtering. Gaussian filtering-based extraction image denoising methods can only handle static noise and cannot address the dynamic motion characteristics of foam during artemisinin extraction. Publicly available literature uses a single threshold for interface segmentation without preprocessing for foam interference, resulting in low interface segmentation accuracy. This step, based on the differences in the motion stability of dynamic foam, employs a categorized filtering strategy combined with adaptive threshold segmentation to achieve accurate foam removal and preliminary interface extraction. The theoretical basis is that stable and unstable sub-blocks of dynamic foam exhibit significant differences in spatiotemporal characteristics. Categorized processing can preserve effective features of the interface and crystallization while denoising, and adaptive threshold segmentation can adapt to interface grayscale changes at different extraction stages. Specific technical methods are as follows: S11 Dual-Modal Image Acquisition: Employing an industrial-grade visible light camera and a near-infrared camera, both coaxially mounted on the observation window at the top of the extraction vessel, with their lenses facing the extraction area inside the vessel, ensuring the acquisition range covers the entire surface of the extractant and the raw material distribution area. The visible light camera has a shooting wavelength range of 400 to 760 nanometers, a frame rate of 30 frames per second, and a resolution of 1920×1080 pixels, used to capture the morphology of the raw materials and the appearance characteristics of the crystal particles. The near-infrared camera has a shooting wavelength range of 780 to 2500 nanometers, with the same frame rate and resolution as the visible light camera, used to extract the purity, impurity content, and other component characteristics within the extractant. Both cameras are synchronously triggered for acquisition, with each frame accompanied by a timestamp to ensure spatiotemporal alignment of the dual-modal images. The acquired image data is stored on the local hard drive of the industrial control computer in BMP format for easy retrieval by subsequent algorithms.

[0025] S12 Foam Sub-block Motion Stability Analysis: Sliding window traversal processing was performed on the synchronously acquired visible light and near-infrared dual-modal images. The sliding window size was set to 32×32 pixels, and the sliding window step size was 16 pixels, ensuring a 50% overlap between adjacent sliding windows to avoid missing foam features. For each image sub-block corresponding to a sliding window, the center coordinates of that sub-block were extracted from five consecutive frames of images, and the variance of the coordinate changes was calculated. Set a stability judgment threshold. When the variance of the coordinate change of the sub-block When the sub-block is determined to be a stable foam sub-block, such sub-blocks have regular movement trajectories and are not easily collapsed; when When the condition is met, it is determined to be an unstable foam sub-block. Such sub-blocks move violently and are prone to merging or collapsing.

[0026] S13 Classified Dynamic Foam Denoising: For stable foam sub-blocks, a motion-compensated temporal filtering algorithm is used. First, the motion vector of the stable sub-block in five consecutive frames is calculated. Based on the motion vector, the sub-blocks in the first four frames are positionally corrected to align them in the spatiotemporal domain. Then, the pixel values ​​of the aligned sub-blocks in the five frames are weighted and averaged. The weight coefficient decreases with the frame number, with the latest frame having the largest weight coefficient to highlight the image information of the current frame. The filtering formula is as follows: ; in, For the filtered sub-block in coordinates Pixel value at that location, For the first The weighting coefficients of the frame satisfy: ; and , For the first Frame coordinates after motion compensation Pixel value at that location, and For the first Frame sub-blocks relative to the current frame , Directional offset.

[0027] For unstable foam sub-blocks, a spatial correlation filtering algorithm is used. The gray-level correlation between each pixel within the sub-block and its surrounding 3×3 neighboring pixels is calculated using the following formula: ; in, coordinates Gray-level correlation of pixels. This is the grayscale value of the pixel. This represents the grayscale value of the neighboring pixels. When Greater than the set correlation threshold When the pixel is identified as a bubble pixel, it is replaced by the median of its neighboring pixels; when... At the same time, the original grayscale value of the pixel is preserved to avoid damaging the interface and the effective features of crystallization.

[0028] S14 Adaptive Thresholding Preliminary Interface Segmentation: For the denoised bimodal image, its gray-level histogram is calculated separately, and the segmentation threshold is adaptively determined using the Otsu algorithm (maximum inter-class variance method). For the visible light image, the interface between the raw material area and the extract area is segmented. The inter-class variance of the foreground (extract area) and background (raw material area) in the gray-level histogram is calculated. When the inter-class variance is maximized, the corresponding gray-level value is the segmentation threshold for the visible light image. For near-infrared images, the interface between the extract liquid region and the bottom of the vessel is segmented, and the segmentation threshold is obtained similarly. .based on and The visible light and near-infrared images are binarized separately to extract the interface region. Then, a logical AND operation is performed on the interface region of the dual-modal image to remove the false interface of the single-modal segmentation, and a preliminary layered interface of the extract is obtained. The dual-modal image without foam interference and the preliminary interface region coordinate data are output.

[0029] Example: This example is applied to a pilot-scale conventional artemisinin solvent extraction production line. The extraction vessel has a volume of 50 liters, uses a paddle stirrer with a stirring speed of 100 rpm, uses ethanol as the extraction solvent, and uses crushed artemisia leaves as the raw material.

[0030] S11 activates the visible light camera and near-infrared camera to simultaneously acquire images of the vessel interior. Acquisition time ranges from 10 to 60 minutes after extraction begins, with each frame accompanied by a timestamp accurate to milliseconds. In the acquired visible light images, the raw material area appears dark green, the extract area appears pale yellow, and the foam area is white and unevenly distributed. In the near-infrared images, the grayscale value of the extract area is higher than that of the raw material area, while the grayscale value of the foam area falls between the two.

[0031] The S12 sets the sliding window size to 32×32 pixels, the step size to 16 pixels, and the stability judgment threshold. Pixels 2 A sliding window is used to traverse a given frame of a visible light image, extracting the center coordinates of each sub-block in five consecutive frames, and calculating the variance of the coordinate changes of some sub-blocks. Pixels 2 ( (stable sub-blocks) Pixels 2 ( (Instantaneous sub-blocks), complete the stability classification of foam sub-blocks.

[0032] S13 to Pixels 2 Stable sub-blocks, with weighted coefficients set. , , , , Substituting the values ​​into the motion-compensation time-domain filtering formula, the filtered sub-block image is obtained, and the foam interference is significantly reduced; for Pixels 2 For unstable sub-blocks, set a correlation threshold. Calculate the partial pixels ( (foam pixels), replaced with the median of a 3×3 neighborhood. ( (effective pixels), retaining the original grayscale values.

[0033] S14 calculates the grayscale histogram of the denoised visible light image to obtain the segmentation threshold corresponding to the maximum inter-class variance. After image binarization, the interface between the raw material area and the extract area is extracted; the segmentation threshold of the near-infrared image is determined. The interface between the extract liquid region and the bottom region of the vessel is extracted; a logical AND operation is performed on the bimodal interface region to remove false interfaces at the edge of the raw material region in the visible light image, yielding preliminary interface region coordinates. The output is a bimodal image with no bubble interference, consisting of pixels (vertical direction of the image) and the corresponding coordinate data.

[0034] Existing publicly available documents use a single Gaussian filter to extract image noise, without distinguishing the dynamic characteristics of foam, resulting in uneven denoising effects for stable and unstable foam, and easily blurring interface features; the published literature uses a fixed threshold to segment the interface, which cannot adapt to grayscale changes in different extraction stages.

[0035] The unique technique in this step lies in its categorized filtering strategy based on the differences in motion stability of foam sub-blocks. Stable foam sub-blocks retain their spatiotemporal continuity through motion-compensated temporal filtering, while unstable foam sub-blocks are precisely removed through spatial correlation filtering. Simultaneously, dual-modal adaptive threshold segmentation and logical AND operations are combined to optimize interface extraction. This technique not only solves the problems of incomplete dynamic foam denoising and easy damage to effective features in existing technologies, but also achieves preliminary accurate interface localization. The denoised image interference is significantly reduced, and the clarity of the interface area is greatly improved, laying a high-quality data foundation for subsequent accurate interface recognition and multi-indicator tracking. Compared with existing technologies, the foam removal rate is significantly improved, and the accuracy of preliminary interface segmentation is markedly enhanced.

[0036] S2 Dynamic Recognition of Extraction Liquid Layered Interface Based on Attention-Enhanced U-Net: After the initial segmentation in step S1, although the layered interface of the extract liquid was obtained, the interface experienced minute dynamic fluctuations due to continuous disturbance from the stirred fluid. The interface coordinates obtained from the initial segmentation could not accurately reflect the true interface location, and the interface contour exhibited jagged distortion. Directly using the initial interface data for subsequent multi-indicator tracking would lead to deviations in the region delineation for indicator recognition, affecting the accuracy of indicator quantification. Therefore, a high-precision image segmentation network is needed to focus on optimizing the initial interface region, accurately identifying the dynamically fluctuating interface, and outputting sub-pixel-level interface coordinates and fluctuation amplitude data, providing a precise region segmentation basis for multi-indicator tracking.

[0037] It should be noted that existing publicly available technologies for interface segmentation mainly use traditional U-Net networks and FCN networks. For example, publicly available U-Net-based extraction interface segmentation methods do not incorporate an attention mechanism, failing to focus on the core features of the interface region and resulting in insufficient accuracy in recognizing dynamically fluctuating interfaces. Publicly available deep learning-based liquid interface segmentation algorithms use FCN networks for interface segmentation, exhibiting weak edge extraction capabilities and severe interface contour distortion. This step utilizes an attention-enhanced U-Net network, embedding channel attention modules and edge enhancement branches to strengthen interface feature extraction and contour optimization. The theoretical basis is that the texture and grayscale features of the layered interface of the extracted liquid differ significantly from the surrounding area. The channel attention module can adaptively enhance the weights of interface feature channels, suppressing interference from irrelevant features. The edge enhancement branch, optimized through a dedicated loss function, can improve the smoothness and accuracy of the interface contour, achieving sub-pixel-level recognition of dynamically fluctuating interfaces. Specific technical methods are as follows: S21 Network Structure: The attention-enhanced U-Net network is an improvement on the traditional U-Net architecture, consisting of three parts: an encoder, a decoder, and an attention module. The encoder comprises four convolutional blocks, each containing two 3×3 convolutional layers, one batch normalization layer, and one ReLU activation function, with a stride of 2, achieving feature map downsampling and feature extraction. The decoder comprises four deconvolutional blocks, each containing one 2×2 deconvolutional layer, two 3×3 convolutional layers, one batch normalization layer, and one ReLU activation function, achieving feature map upsampling and feature recovery. Channel attention modules are embedded between each corresponding layer of the encoder and decoder, and an edge enhancement branch is embedded in the last layer of the decoder.

[0038] S22 Channel Attention Module Calculation: The channel attention module receives the feature map output by the encoder. The feature map dimension is ,in For the number of channels, For height, For width. First, for... Performing global average pooling and global max pooling respectively yields two results. eigenvectors and Then, the two feature vectors are input into two independent fully connected layers, with the weights of the first fully connected layer being... (dimension) , (where is the dimensionality reduction coefficient), the weights of the second fully connected layer are (dimension) ),right The process is performed to obtain the feature vector. Similarly, through weights (dimension) )and (dimension) )right The process is performed to obtain the feature vector. ;Will and After summing, the results are processed through a Sigmoid activation function to obtain the channel attention weight matrix. The formula is: ; in, The sigmoid activation function maps weight values ​​to the range of 0 to 1. Compared with the original feature map Perform channel-by-channel multiplication to obtain the enhanced feature map. This enables the enhancement of core interface features and the suppression of irrelevant features.

[0039] S23 Edge Enhancement Branch Optimization: Edge enhancement branch receives the feature map of the last layer of the decoder. First, a 1×1 convolutional layer is used to... The number of channels is reduced to 1, resulting in a single-channel feature map. Then, the Canny operator is used to extract... The edge contour is used to obtain the edge mask. Constructing the edge loss function By optimizing this loss function, the accuracy of interface contour extraction is improved. The formula for the edge loss function is: ; in, Single-channel feature map The total number of pixels, For the first The true edge label of each pixel (edge ​​pixels are 1, non-edge pixels are 0). For the first Predicted edge labels for pixels. For the first The edge weight coefficients of each pixel, and the pixels in the edge region. The value is greater than that of non-edge regions to highlight the importance of edge optimization.

[0040] S24 Interface Dynamic Recognition and Data Output: The bubble-free bimodal image output from step S1 is channel-stitched and used as input to the attention-enhanced U-Net network. Simultaneously, the preliminary interface region obtained in S1 is designated as the network's focus region, receiving higher weights in the loss value during training and inference. The network output is a binarized interface segmentation map. Subpixel-level coordinates of the interface are extracted using a subpixel edge detection algorithm. Based on the interface coordinates of 30 consecutive frames, the interface offset between each frame and the previous frame is calculated, yielding temporal data of interface fluctuation amplitude over time. Subpixel-level interface coordinates and fluctuation amplitude temporal data are then output.

[0041] Example: Continuing with the pilot production line example in S1, the training dataset for the attention-enhanced U-Net network uses 10,000 frames of bimodal images historically collected from this production line, of which 8,000 frames are used as the training set and 2,000 frames are used as the validation set. The image resolution is uniformly adjusted to 1920×1080 pixels.

[0042] S21 constructs an attention-enhanced U-Net network. The encoder's four convolutional blocks output feature maps with 64, 128, 256, and 512 channels, respectively, while the decoder's four deconvolutional blocks output feature maps with 256, 128, 64, and 32 channels, respectively. Dimensionality reduction coefficients... Marginal weight coefficient Set the value to 2 for edge regions and 1 for non-edge regions.

[0043] S22 Feature map of a certain frame of input image (Number of channels) ,high ,width ), and perform global average pooling to obtain ( Global max pooling yields ( ); Fully connected layer weights Dimensions , Dimensions , Dimensions , Dimensions ;Calculation obtained , Substituting into the Sigmoid function yields the channel attention weight matrix. The channel weights for interface features are 0.8-0.9, and the channel weights for irrelevant features are 0.1-0.2. and Channel-by-channel multiplication yields the enhanced feature map. The interface features have been significantly enhanced.

[0044] S23 decoder output feature map of the last layer (32 channels) Reduced to a single-channel feature map using a 1×1 convolutional layer. Extraction using the Canny operator The edge contour is used to obtain the edge mask. Substitute into the edge loss function Calculate the value of a certain frame of image The value was 0.05, and after multiple rounds of iterative optimization, When the value drops below 0.01, the smoothness of the interface outline is significantly improved.

[0045] S24 stitches the bimodal images output by S1 and inputs them into the network, setting the region of interest as the initial interface coordinates. Within a 50-pixel radius of the pixel, the network outputs a segmentation map of the interface. Using a sub-pixel edge detection algorithm, sub-pixel-level interface coordinates are extracted. Pixel; Based on the coordinate data of 30 consecutive frames of images, the interface fluctuation amplitude is calculated to be between 0.1 and 0.3 pixels, and the time series data of the coordinates and fluctuation amplitude are output.

[0046] Existing publicly available documents use traditional U-Net networks to segment and extract interfaces, but do not introduce an attention mechanism, making it impossible to focus on the core features of the interface. This results in low recognition accuracy for dynamically fluctuating interfaces and distorted edge contours. The liquid interface segmentation algorithms based on deep learning in the published literature use FCN networks, but lack dedicated edge optimization branches, resulting in insufficient smoothness of the interface contours.

[0047] The unique technique in this step lies in embedding the channel attention module into the encoder and decoder layers of the U-Net network. It enhances interface features through adaptive calculation of channel weights, adds an edge enhancement branch at the decoder end, optimizes the interface contour using a dedicated edge loss function, and focuses on the initial interface region as a key area of ​​interest to improve recognition specificity. This technique overcomes the shortcomings of existing networks in terms of insufficient accuracy and edge distortion in recognizing dynamically fluctuating interfaces, achieving sub-pixel-level interface localization and fluctuation tracking. The accuracy of interface recognition is significantly improved, and the capture of fluctuation amplitude is more comprehensive. Compared to existing technologies, the interface recognition error is significantly reduced, providing a precise basis for subsequent multi-indicator tracking region delineation, and significantly reducing the regional deviation of indicator recognition.

[0048] S3 Multimodal Multi-Indicator Collaborative Tracking: Controlling the artemisinin extraction process requires a comprehensive understanding of multiple core indicators, including raw material consumption, extract purity, and crystal growth. Single-modal images can only reflect partial information; for example, visible light images excel at capturing appearance features, while near-infrared images are better at reflecting component characteristics. Relying solely on a single modality for indicator identification leads to incomplete information and inaccurate indicator quantification. Furthermore, failing to combine precise interface region identification results in a blurred tracking range, further exacerbating deviations. Therefore, it is necessary to integrate the advantages of dual-modal images and combine them with the precise interface coordinates obtained in step S2 to delineate the identification area, achieving synchronous collaborative tracking of multiple indicators. This outputs comprehensive and accurate indicator data, providing a basis for decision-making in adjusting process parameters.

[0049] It should be noted that existing publicly available technologies for multi-index recognition often employ single-modal images or simple modal stitching. For example, the extraction index recognition method based on visible light images proposed in published documents can only identify the raw material morphology and crystal quantity, but cannot obtain component indicators such as the purity of the extract. The application of multi-modal image fusion in industrial testing in published literature uses a simple pixel-level fusion method, which does not consider the weight differences of different modal features, resulting in poor fusion performance. This step dynamically adjusts the weights of dual-modal features based on an attention mechanism, combined with precise interface regions for partitioned recognition. The theoretical basis is that different modal images contribute differently to the recognition of different indicators, and the attention mechanism can adaptively allocate weights to highlight effective modal features. The raw material region and extract region delineated based on interface coordinates make index recognition more targeted, reduce interference from irrelevant regions, and improve quantification accuracy. Specific technical means are as follows: S31 Dual-Modal Feature Extraction: Feature extraction is performed on the bubble-free visible light image and near-infrared image output from step S1. For the visible light image, a ResNet50 network is used as the feature extraction backbone. The last fully connected layer is removed, and the convolutional layer portion is retained, resulting in an extracted feature vector of dimension [dimension missing]. Feature map 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.

[0050] 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: ; 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.

[0051] Fusion weights for bimodal features calculated based on confidence level. The formula is: ; 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. .

[0052] 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.

[0053] 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.

[0054] 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.

[0055] 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.

[0056] 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 , , .

[0057] 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 is calculated by the characteristic response intensity in the extract region; the interface fluctuation amplitude is calculated by the fluctuation range of the interface coordinates in consecutive frames; and the number of crystal particles is obtained by the number of detection frames used to detect crystal particles. A multi-index data matrix is ​​output at a specific moment: raw material consumption rate 0.02 g / s, extract purity 85%, interface fluctuation amplitude 0.2 pixels, number of crystal particles 120, and impurity distribution density 0.05.

[0058] Existing publicly available documents only extract indicators based on visible light image recognition, and cannot obtain classification indicators, resulting in incomplete information. The application of multimodal image fusion in industrial inspection in publicly available literature adopts simple pixel-level fusion without dynamically adjusting modal weights, and the fusion features cannot adapt to the recognition needs of different indicators.

[0059] The unique technical approach of this step lies in dynamically calculating the recognition confidence of bimodal features based on indicator types, allocating fusion weights through an attention mechanism, and simultaneously delineating the raw material and extract areas using precise interface coordinates. An improved YOLO-V5 network is then employed for zoned indicator recognition. This technique achieves synchronous and collaborative tracking of appearance and component classification indicators, overcoming the shortcomings of single-modal information being incomplete and the poor fusion effect of fixed weights. The comprehensiveness and accuracy of multi-indicator recognition are significantly improved. Compared to existing technologies, indicator quantification errors are significantly reduced, providing a comprehensive and reliable decision-making basis for process parameter adjustments, and significantly enhancing the targeting of parameter adjustments.

[0060] S4 Particle Filter Optimization for Crystal Morphology Tracking and Closed-Loop Feedback of Process Parameters: The growth state of crystal morphology directly determines the purity and yield of artemisinin, requiring real-time tracking of the crystal growth trajectory, particle size changes, and morphological stability. Existing tracking algorithms are prone to particle degradation and depletion, leading to tracking interruptions or deviations. Simultaneously, crystal tracking data is disconnected from process parameter adjustments, failing to optimize process conditions promptly based on crystal state, thus impacting product quality. Therefore, it is necessary to optimize the particle filter algorithm, improve tracking accuracy by incorporating crystal growth patterns, and establish a closed-loop feedback mechanism between tracking data and process parameter adjustments to achieve dynamic control of crystal growth and ensure optimal extraction conditions.

[0061] It should be noted that existing publicly available technologies for crystallization tracking often employ traditional particle filtering algorithms. For example, the crystallization tracking method based on standard particle filtering proposed in published documents does not incorporate a crystallization growth model, resulting in unreasonable particle weight updates and a tendency for degradation. Furthermore, the application of particle filtering in particle tracking in published literature uses a random resampling strategy, leading to the loss of high-weight particles and poor tracking stability. This step introduces a crystallization growth model to correct particle weights and employs a hierarchical resampling strategy to optimize the particle set. The theoretical basis for this is that the growth of crystallized particles follows specific laws, and combining this with a growth model can accurately predict particle states and correct weights to improve tracking accuracy. Hierarchical resampling can retain high-weight particles according to their weight ratio, reducing degradation and depletion phenomena and ensuring tracking continuity. Establishing a linkage feedback between tracking data and process parameters enables dynamic optimization of crystallization growth. Specific technical methods are as follows: S41 Crystallization Region Determination and Particle Set Initialization: Based on the sub-pixel-level interface coordinates output in step S2, the crystallization region is determined to be the range of 50 to 200 pixels from the interface in the extract liquid area (this region is the main crystallization region of artemisinin). A uniform distribution sampling strategy is used to initialize the particle set within the crystallization region. Each tracking particle represents the coordinates of a potential crystal grain center. , Initial weights of particles This ensures that the initial weights are evenly distributed.

[0062] S42 Particle State Prediction and Weight Correction: The next frame state of particles is predicted based on a crystal growth model. This model uses a spherical growth model, assuming that the radius of the crystal particles grows linearly with time. The radius growth formula is as follows: ,in for The crystallization radius at time, Let the initial radius be , This is the growth rate constant. Combined with the particle's current coordinates... and radius Predict the coordinates of the next frame ,in and The random offset is based on the fluid flow characteristics and follows a Gaussian distribution. , This represents the standard deviation of the offset.

[0063] Calculate the observation likelihood of particles based on the fused feature map output from step S3. The feature vector of the region corresponding to each particle is extracted, and the cosine similarity is calculated with the template feature vector of the crystallized particle. The higher the similarity, the greater the observation likelihood. The formula for observation likelihood is: ; in, for Frame observation data (fused feature map). For the first The predicted state of each particle. To adjust the coefficient, The cosine similarity between the features of the region corresponding to the particle and the features of the template.

[0064] The particle weights are adjusted based on the crystal growth model, and the adjustment formula is as follows: ; in, Let be the prior probability of the crystal growth model. Given the predicted crystallization radius of the particle, if conform to The growth pattern ,otherwise This suppresses particle weights that do not conform to the growth pattern.

[0065] S43 Hierarchical Resampling and State Estimation: Calculating the Normalization Factor for Particle Weights Normalize the particle weights. Set the resampling threshold. When the normalized weight of the particle When a particle is identified as having low weight, it needs to be resampled.

[0066] A hierarchical resampling strategy is adopted, in which particles are sorted from largest to smallest according to their normalized weights and divided into groups. There are three levels, and the sum of particle weights in each level is equal. Within each level, particles are randomly selected according to their weight proportions. The number of particles selected is proportional to the weight proportion of particles in the level, ensuring that high-weight particles are fully preserved and low-weight particles are replaced with copies of high-weight particles to form a new particle set.

[0067] Based on the new particle set, the state of the crystallized particles is estimated using a weighted average method, and the estimated center coordinates are: ; The radius estimate is: ; Output data on the growth trajectory, particle size variation, and morphological stability of the crystallized particles (morphological stability is measured by the variance of the radius variation).

[0068] S44 Process Parameter Closed-Loop Feedback: Preset ranges are established for five core indicators. When any indicator in the multi-indicator data matrix output from step S3 deviates from the preset range, a process parameter adjustment command is generated based on the crystallization state data output from S43. For example, when the purity of the extract is below the preset lower limit and the particle size of the crystals grows slowly, a command to increase the stirring speed and raise the extraction temperature is generated; when the number of crystal particles is excessive and the impurity distribution density is above the preset upper limit, a command to decrease the stirring speed and adjust the solvent ratio is generated.

[0069] The adjustment command is sent to the extraction control system. After the control system adjusts the parameters, the extraction state inside the extraction vessel changes. The changed state is then re-imaged by the camera in step S1, and the process enters the next closed loop of "denoising-interface recognition-index tracking-crystallization feedback" to achieve adaptive optimization.

[0070] Example: Continuing with the pilot production line example above, the crystallization region is set as the extract liquid region. to Pixel range (based on interface coordinates) (pixels), initial particle count Initial weights Initial radius of crystallization Pixel, growth rate constant Pixels / frame, adjustment factor Offset standard deviation Pixel, resampling threshold Number of levels .

[0071] S41 in to Within a pixel area, 200 particles are uniformly sampled, with particle coordinates distributed within the area and initial weights of 0.005 for each particle.

[0072] S42 predicts the radius of a particle in the next frame based on a crystal growth model. pixels, coordinates ( , ,obey Extract the feature vector of the region corresponding to the particle and calculate its cosine similarity with the template feature vector. Substituting into the observation likelihood formula, we get ;because In accordance with growth patterns Corrected weights .

[0073] S43 Calculate the normalization factor Normalized weights ( The 200 particles are sorted by weight and divided into 5 levels, with the sum of the weights in each level being 0.2. Particles are then extracted proportionally within each level, with high-weight particles being extracted multiple times to form a new particle set. The estimated center coordinates of the crystal grains are calculated based on these new particle sets. With a radius of 5.3 pixels and a radius variation variance of 0.02, it exhibits good morphological stability.

[0074] S44 sets the extraction purity to a preset range of 88% to 95%. When the extraction purity output by S3 is 85% (below the lower limit) and the crystal particle size increases slowly (from 5 pixels to 5.3 pixels, with a growth rate of 0.3 pixels / 10 frames), an adjustment command is generated: the stirring speed is increased from 100 rpm to 110 rpm, and the extraction temperature is increased from 30 degrees Celsius to 32 degrees Celsius. After the command is sent to the control system for execution, the new extraction state is captured by the camera of S1, and the next closed-loop process begins.

[0075] Existing public documents use traditional particle filtering to track crystal growth without incorporating a crystal growth model. The particle weight update lacks a theoretical basis, resulting in low tracking accuracy and a tendency for particle degradation. The application of particle filtering in particle tracking in 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.

[0076] The unique technical approach in this step lies in introducing a spherical crystal growth model to correct particle weights, ensuring that the particle state conforms to the crystal growth law. A hierarchical resampling strategy is employed to extract particles according to their weight levels, retaining high-weight particles. Simultaneously, a closed-loop feedback mechanism is established between crystal tracking data and process parameter adjustments. This technique overcomes the degradation and depletion problems of existing particle filters, improves the accuracy and stability of crystal morphology tracking, and achieves real-time linkage optimization between crystal state and process parameters. Compared to existing technologies, the success rate of crystal tracking is significantly improved, the response speed of process parameter adjustments is faster, the purity and yield of artemisinin are significantly improved, and the stability of the production process is greatly enhanced.

[0077] This method constructs an integrated closed-loop system of "foam denoising - interface localization - multi-index tracking - crystallization feedback" through the coordinated linkage of steps S1 to S4. Each step supports and promotes the others: foam denoising and preliminary interface segmentation in S1 provide a clean data foundation for accurate interface recognition in S2; sub-pixel-level interface coordinates in S2 define the precise range for partition index recognition in S3; multi-index data in S3 provide a comprehensive basis for crystallization tracking and parameter feedback in S4; and the parameter adjustment results in S4 feed back to the front-end steps through the closed loop to achieve adaptive optimization.

[0078] Compared to the existing application of single technologies, the collaborative system of this method not only solves multiple technical pain points such as dynamic foam interference, interface fluctuations, one-sided information of multiple indicators, crystallization tracking deviation, and feedback lag, but also produces unexpected synergistic effects. It significantly improves the intelligent control accuracy, response speed and stability of the extraction process, provides reliable technical support for the efficient and high-quality production of pilot-scale conventional artemisinin extraction production lines, 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: ； in Here is the channel attention weight matrix. Input feature maps to the network, It is the Sigmoid activation function. , , , For learning parameters, The result of average pooling of feature maps. The feature map is max-pooled; the edge enhancement branch uses the edge loss function: ; Optimize interface outline extraction, among which This represents the edge loss value. The number of pixels. For pixel-based true labels, Predict labels for pixels. The edge pixel weight coefficients are used; the sub-pixel level interface coordinates and interface fluctuation amplitude data are output. 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 through multimodal collaborative tracking technology. The multimodal collaborative tracking technology includes synchronous acquisition and target identification of key parameters and material states in each stage of raw material pretreatment, extraction, purification, and solvent recovery. 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: S3 specifically includes the following sub-steps: S31 extracts the morphological and crystal density characteristics of raw materials from visible light images, and extracts the purity and impurity content characteristics of the extract from near-infrared images; S32 dynamically adjusts the weight allocation of dual-modal features through an attention mechanism to construct a fused feature map; S33 delineates the raw material zone and the extract zone based on the interface coordinates output in step S2, and uses an improved YOLO-V5 network to identify core indicators in the two zones respectively, forming a multi-indicator real-time data matrix.

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 4, characterized in that: 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: ; 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.

7. 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.

8. 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