Salvia miltiorrhiza root section automatic quantitative feeding system

Abstract

The invention discloses an automatic quantitative feeding system for radix salviae miltiorrhizae root segments, and relates to the technical field of traditional Chinese medicine planting, and the system comprises a vibration conveying module which realizes directional conveying of the radix salviae miltiorrhizae root segments through controllable vibration and a volume-variable groove body, and obtains three-dimensional point cloud data based on synchronous laser scanning; the self-adaptive metering module fuses machine vision and near infrared spectrum, rejects invalid salvia miltiorrhiza root segments through double verification and outputs an actual effective number; the intelligent feeding module controls an opening and closing device to realize accurate quantitative feeding by taking the actual effective quantity as a reference; the vibration control module predicts an initial vibration frequency through LSTM migration, and generates an actual vibration frequency containing an anti-blocking mechanism based on multi-modal sensing fusion; the volume control module dynamically adjusts the volume of the tank body according to voxelization and spectral compensation. According to the quantitative feeding device, the problems that counting is inaccurate, feeding is prone to being blocked and the radix salviae miltiorrhizae root sections are damaged due to the fact that the characteristics of single root sections are different in the conveying process of the radix salviae miltiorrhizae root sections are solved, and automation and lossless quantitative feeding of the radix salviae miltiorrhizae root sections are achieved.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine cultivation technology, specifically to an automatic quantitative feeding system for Salvia miltiorrhiza root segments. Background Technology

[0002] Salvia miltiorrhiza is a perennial erect herb belonging to the genus Salvia of the Lamiaceae family. Its root is used medicinally and has effects such as promoting blood circulation and regulating menstruation. It is an important medicine in gynecology. The feeding process generally relies on manual grasping and feeding. Operators need to carry heavy loads for a long time in a dusty environment, which is extremely labor-intensive. Moreover, because the root segments of Salvia miltiorrhiza are of uneven thickness and covered with mud, manual estimation is very easy to result in over- or under-feeding, which leads to fluctuations in planting density per unit area, uneven emergence, and affects the grade and yield stability of the medicinal material.

[0003] While existing agricultural or food industries have adopted solutions that combine vibratory conveying with rotary metering for quantitative feeding of seeds and granular materials, the root segments of Salvia miltiorrhiza are irregularly long and strip-shaped, with a large aspect ratio, are prone to tangling and sticking, and have a high coefficient of friction. Traditional equipment's volumetric metering and counting metering modes are unable to cope with the irregular shape and tangling characteristics of Salvia miltiorrhiza root segments, and cannot balance accuracy and efficiency. Furthermore, fluctuations in the water and mud content of the root segments can easily cause adhesion and jamming, which cannot meet the needs of large-scale Salvia miltiorrhiza cultivation. Summary of the Invention

[0004] This invention provides an automatic quantitative feeding system for Salvia miltiorrhiza root segments, which solves the problems of inaccurate counting, easy blockage during feeding, and damage to Salvia miltiorrhiza root segments caused by the different characteristics of individual root segments during transportation, and realizes the automation and non-destructive quantitative feeding of Salvia miltiorrhiza root segments.

[0005] This invention provides an automatic quantitative feeding system for Salvia miltiorrhiza root segments, comprising:

[0006] The vibration conveying module is used to directionally convey the root segments of Salvia miltiorrhiza through controllable vibration and a variable volume tank, and to acquire the three-dimensional point cloud data of the root segments of Salvia miltiorrhiza in real time based on laser contour scanning.

[0007] The adaptive measurement module is used to detect the root segments of Salvia miltiorrhiza through machine vision and near-infrared spectroscopy, and obtain the image information, spectral information and actual effective quantity of the root segments.

[0008] The intelligent feeding module is used to control the opening and closing device to feed the Salvia miltiorrhiza root segments quantitatively based on the actual effective quantity of the root segments.

[0009] The vibration control module is used to predict the initial vibration frequency of controllable vibration based on historical transmission data using an LSTM network. Based on the image information of the Salvia miltiorrhiza root segments, the spectral information of the Salvia miltiorrhiza root segments, and the actual effective number of Salvia miltiorrhiza root segments, the initial vibration frequency is corrected to obtain the actual vibration frequency of controllable vibration.

[0010] The volume control module is used to dynamically adjust the volume of the Salvia miltiorrhiza delivery tank based on the three-dimensional point cloud data and spectral information of the Salvia miltiorrhiza root segments, and deliver the Salvia miltiorrhiza root segments through the variable volume tank after volume adjustment.

[0011] This invention addresses the problems of inaccurate counting, easy clogging, and damage to Salvia miltiorrhiza root segments during transportation due to the varying characteristics of individual root segments, such as irregular shape and fluctuating water and mud content. It provides an automatic quantitative feeding system for Salvia miltiorrhiza root segments. The system achieves directional transport of root segments through controllable vibration of the vibration conveying module and a variable-volume trough. It utilizes laser contour scanning to generate real-time 3D point cloud data of the Salvia miltiorrhiza root segments. An adaptive metering module integrates machine vision and near-infrared spectroscopy to perform image and spectral recognition of the Salvia miltiorrhiza root segments, obtaining accurate data. The system measures the appearance integrity and internal material of each root segment, discarding invalid Salvia miltiorrhiza root segments and outputting the actual effective quantity, thus achieving accurate counting of Salvia miltiorrhiza root segments. An intelligent feeding module controls the opening and closing device based on the actual effective quantity to achieve precise quantitative feeding of Salvia miltiorrhiza root segments. A vibration control module, based on historical transmission data including vibration frequency, temperature, humidity, and root segment size, enables controllable vibration frequency selection. A volume control module, based on the three-dimensional point cloud data and spectral information of the Salvia miltiorrhiza root segments, enables the selection of the volume of the variable-volume tank.

[0012] Furthermore: the vibration transmission module includes:

[0013] The vibration generating unit is used to receive vibration frequency correction signals and generate controllable mechanical vibrations based on the vibration frequency correction signals.

[0014] The elastic support unit is used to transmit controllable mechanical vibration to the variable volume trough that carries the Salvia miltiorrhiza root segments, and to convert the controllable mechanical vibration into directional vibration of the variable volume trough along the conveying direction, so as to directionally convey the Salvia miltiorrhiza root segments through the variable volume trough.

[0015] The laser scanning unit is used to synchronously collect vibration data of the Salvia miltiorrhiza root segments within the variable volume tank, thereby acquiring three-dimensional point cloud data of the Salvia miltiorrhiza root segments.

[0016] This invention receives a vibration frequency correction signal through a vibration generating unit to generate controllable vibration matching the currently transported Salvia miltiorrhiza root segments. An elastic support unit converts this vibration into directional vibration waves along the transport direction of the trough, completing the directional transport of the Salvia miltiorrhiza root segments. By receiving the vibration frequency correction signal, the vibration frequency of the vibration transport is adjusted in real time to reduce the occurrence of Salvia miltiorrhiza root segment entanglement and adhesion. The laser scanning unit is rigidly connected to the elastic support unit and vibrates synchronously, enabling high-precision acquisition of three-dimensional point cloud data under vibration conditions. This invention solves the problem of Salvia miltiorrhiza root segments being easily affected by root segment entanglement and adhesion during traditional vibration transport, resulting in disordered arrangement and blockage of the root segments and unstable feeding.

[0017] Furthermore: the adaptive metering module includes:

[0018] The multispectral synchronous acquisition unit is used to simultaneously acquire images and spectra of the root segments of Salvia miltiorrhiza, and obtain the visible light images and raw infrared spectral data of the root segments of Salvia miltiorrhiza.

[0019] The feature fusion preprocessing unit is used to process the visible light image to obtain the image information of the Salvia miltiorrhiza root segment, process the raw infrared spectrum data to obtain the spectral information of the Salvia miltiorrhiza root segment, and register the image information and spectral information to obtain the registered image and registered spectrum of the Salvia miltiorrhiza root segment.

[0020] The material morphology recognition unit is used to segment the contour of the Salvia miltiorrhiza root segment in the registration image using an image segmentation network to obtain the morphological integrity signal of each Salvia miltiorrhiza root segment, extract the material features of the Salvia miltiorrhiza root segment from the registration spectrum to obtain the material qualification signal of each Salvia miltiorrhiza root segment, and mark the Salvia miltiorrhiza root segment as a valid Salvia miltiorrhiza root segment when both the morphological integrity signal and the material qualification signal of the Salvia miltiorrhiza root segment exist.

[0021] The Salvia miltiorrhiza root segment counting unit is used to count the number of valid Salvia miltiorrhiza root segment identifiers and obtain the actual effective number of Salvia miltiorrhiza root segments.

[0022] To address the challenges of irregular morphology, material variations, and easy adhesion and breakage of Salvia miltiorrhiza root segments, leading to large measurement errors and the inclusion of invalid root segments, this invention employs a multispectral synchronous acquisition unit to acquire visible light and near-infrared spectra of Salvia miltiorrhiza root segments. A feature fusion preprocessing unit then registers the visible light and infrared images. A material and morphology recognition unit uses an image segmentation network to accurately segment the Salvia miltiorrhiza root segments and identifies them as valid segments. Finally, a root segment counting unit counts these valid segments to determine the actual number of valid segments, eliminating unqualified segments. This achieves effective identification and counting of Salvia miltiorrhiza root segments, reducing the incorporation of hollow or damaged segments.

[0023] Furthermore: the vibration control module includes:

[0024] The historical database unit is used to store a data matrix of Salvia miltiorrhiza root segments, including historical vibration frequencies, ambient temperature and humidity, average root segment size, and actual feeding quantity.

[0025] The LSTM transfer prediction unit is used to adapt the Salvia miltiorrhiza root segment delivery scenario through transfer learning using a pre-trained LSTM network. It takes the Salvia miltiorrhiza root segment data matrix as input and outputs the initial vibration frequency.

[0026] The multimodal perception fusion unit is used to extract the bulk density of Salvia miltiorrhiza root segments based on image information, extract the water content gradient of Salvia miltiorrhiza root segments based on spectral information, and generate frequency correction coefficients based on the bulk density, water content gradient, and actual effective quantity of Salvia miltiorrhiza root segments.

[0027] A frequency synthesizer is used to obtain the actual vibration frequency of controllable vibration based on the initial vibration frequency and the frequency correction coefficient.

[0028] This invention constructs a data matrix of Salvia miltiorrhiza root segments using a historical database unit to store multi-dimensional data. An LSTM transfer prediction unit then transfers a pre-trained LSTM network to the Salvia miltiorrhiza root segment recognition scenario. Based on the multi-dimensional data of the root segments, an initial vibration frequency is output. A multi-modal perception fusion unit then corrects the initial vibration frequency by incorporating the root segments' packing density, moisture content gradient, and actual effective quantity, resulting in an actual vibration frequency suitable for the currently transported root segments. This invention solves the problem of traditional vibration conveying methods using a fixed frequency, which cannot adapt to temperature and humidity fluctuations, root segment size differences, and changes in root segment packing, causing blockages during root segment feeding. It achieves matching of the conveyed vibration frequency with the state of the Salvia miltiorrhiza root segments, reducing the risk of blockage and breakage.

[0029] Furthermore: the pre-trained LSTM network is configured as follows:

[0030] An industrial vibration dataset containing vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate is used to pre-train an LSTM network. The LSTM network includes:

[0031] The input layer consists of 5 neurons, which receive vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate, respectively.

[0032] The hidden layer comprises a first LSTM layer, a Danshen feature embedding subnetwork, a feature fusion layer, and a second LSTM layer connected in sequence. Both the first LSTM layer and the second LSTM layer contain 64 neurons.

[0033] The output layer consists of one linear neuron, which outputs the predicted initial vibrational frequency.

[0034] The method of adapting to the delivery scenario of Salvia miltiorrhiza root segments through transfer learning specifically includes: freezing the parameters and pre-trained weights of the first LSTM layer, performing supervised fine-tuning on the Salvia miltiorrhiza feature embedding sub-network, feature fusion layer and second LSTM layer, and outputting the vibration frequency of the Salvia miltiorrhiza root segment scenario.

[0035] The pre-trained LSTM network in this invention is trained using industrial vibration data. By freezing the weights of the first LSTM layer to retain industrial vibration knowledge and reduce overfitting caused by insufficient Salvia miltiorrhiza data, a Salvia miltiorrhiza feature embedding sub-network and a feature fusion layer are inserted into the hidden layers to fuse the vibration features of the industrial data with the features of the Salvia miltiorrhiza data. This achieves complementary knowledge in the transfer learning domain, enabling the pre-trained LSTM network to focus on the vibration frequency of the Salvia miltiorrhiza root segment. This invention solves the problem of a large gap between the original knowledge neighborhood and the Salvia miltiorrhiza root segment recognition scenario after transfer learning, which leads to a decrease in model accuracy. At the same time, it reduces the data requirements of the Salvia miltiorrhiza root segment field and shortens the learning cycle of transfer learning.

[0036] Furthermore: the Danshen feature embedding subnetwork includes an input layer and a fully connected layer;

[0037] The input layer is used to receive image information of Salvia miltiorrhiza root segments, spectral information of Salvia miltiorrhiza root segments, and the actual effective quantity of Salvia miltiorrhiza root segments.

[0038] The fully connected layer is used to map the image information of the Salvia miltiorrhiza root segments, the spectral information of the Salvia miltiorrhiza root segments, and the actual effective number of Salvia miltiorrhiza root segments to obtain the feature vector of the Salvia miltiorrhiza root segments.

[0039] The feature fusion layer is used to fuse the output feature vector of the first LSTM layer with the feature vector of the Salvia miltiorrhiza root segment to obtain the Salvia miltiorrhiza root segment fused feature vector.

[0040] This invention addresses the problem of significant semantic discrepancies between the pre-trained LSTM network's original knowledge domain and the Salvia miltiorrhiza root segment recognition scenario, which makes it difficult for the transferred LSTM network to effectively capture the coupling relationship between the Salvia miltiorrhiza root segment material and vibration frequency. It employs a fully connected layer to perform nonlinear transformation and dimensionality reduction on the image information, spectral information, and actual effective quantity of the Salvia miltiorrhiza root segments, resulting in semantically unified feature vectors and eliminating dimensional differences and redundant noise between the original data. Finally, a feature fusion layer fuses the output feature vector of the first LSTM layer with the Salvia miltiorrhiza root segment feature vector, obtaining a fused feature vector that retains the original feature knowledge while adapting to the Salvia miltiorrhiza root segment. This ensures that the output initial vibration frequency encompasses both the pre-trained LSTM network's original knowledge domain and the Salvia miltiorrhiza recognition knowledge domain, guaranteeing the adaptability of the Salvia miltiorrhiza root segment delivery process.

[0041] Furthermore, the expression for the actual vibration frequency is as follows:

[0042]

[0043]

[0044]

[0045] in, The actual vibration frequency, The initial vibration frequency, This is the frequency correction factor. To prevent blockage in the root segments of Salvia miltiorrhiza, For the anti-blocking coefficient, For time, For indicator functions, For the frequency correction coefficient threshold, This refers to the bulk density of Salvia miltiorrhiza root segments. This refers to the actual effective quantity of Salvia miltiorrhiza root segments.

[0046] This invention solves the problem of the inability to adjust the vibration frequency in real time to adapt to changes in the operating conditions of Salvia miltiorrhiza root segments during transportation by using an expression for the actual vibration frequency. The invention constructs a frequency correction coefficient based on the actual effective quantity of Salvia miltiorrhiza root segments, their bulk density, and moisture content gradient. When the root segments are piled up haphazardly or the moisture content increases, the initial vibration frequency is increased to reduce blockage. When the actual effective number of root segments is large, the initial vibration frequency is decreased to suppress amplitude and reduce damage to the root segments. An indicator function triggers an anti-blockage frequency. When the bulk density of the root segments is too high, a sinusoidal modulation wave is added to the initial vibration frequency, i.e., the anti-blockage frequency, to reduce congestion by increasing the vibration frequency in a short time. This invention achieves real-time matching between the vibration frequency and the transportation conditions, improving the robustness and adaptability of directional transportation of Salvia miltiorrhiza root segments.

[0047] Further: the volume control module includes:

[0048] Voxelization unit is used to voxelize the three-dimensional point cloud data of Salvia miltiorrhiza root segments to obtain the density distribution matrix of Salvia miltiorrhiza root segments.

[0049] The CNN prediction unit is used to input the density distribution matrix into a pre-trained convolutional neural network volume prediction model and output the target volume value.

[0050] The PID control unit is used to generate a volume adjustment command signal based on the deviation between the target volume value and the current measured value of the volume sensor.

[0051] The spectral compensation unit is used to compensate the volume adjustment command signal for moisture content based on the spectral information of the Salvia miltiorrhiza root segment, so as to obtain a corrected volume adjustment command signal.

[0052] The volume control unit is used to adjust the volume of the Salvia miltiorrhiza delivery tank according to the modified volume control command signal.

[0053] To address the problem of uneven root segment size and fluctuating moisture content leading to excessively large or insufficient transport tank volume, resulting in root segments squeezing or clogging, this invention uses a voxelization unit to compress a three-dimensional point cloud into a density distribution matrix in real time. A CNN prediction unit then rapidly outputs the target volume based on this matrix. A PID control unit generates preliminary instructions based on the deviation between the target and measured volumes. A spectral compensation unit then corrects these instructions based on near-infrared spectral moisture content. Finally, a volume execution unit completes closed-loop volume adjustment of the tank. This invention can predict the target volume based on three-dimensional point cloud data and determine the specific volume based on moisture content, accurately controlling the volume, reducing root segment accumulation and clogging, and minimizing the waste of vibration energy caused by excessive expansion. It achieves adaptive volume control based on the moisture content and clogging status of the Salvia miltiorrhiza root segments.

[0054] Furthermore, the expression for the modified volume adjustment command signal is as follows:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] in, For the corrected volume adjustment command signal, This is a volume adjustment command signal. The water content is obtained based on the spectral information of the Salvia miltiorrhiza root segment. For spectral moisture content deviation, , , All are PID control adjustment coefficients. This represents the current response deviation. To accumulate historical bias, The rate of change of deviation For the target volume value, This is the current measured value from the volume sensor. For a pre-trained convolutional neural network volume prediction model, The density distribution matrix is The x-coordinate of the raster. The vertical coordinate of the raster is... This is the three-dimensional point cloud data of the root segment of Salvia miltiorrhiza. This is a voxelization operation.

[0061] To address the problem of inaccurate PID volume control commands and inaccurate tank adjustment caused by fluctuations in 3D point cloud density due to changes in root segment size and moisture content, this invention superimposes the moisture content deviation measured in real-time by the spectrum onto the PID command. When the moisture content is high, the PID command is amplified, and the volume tank immediately expands; when the moisture content is low, the PID command is reduced, and the volume tank contracts. This achieves accurate adjustment of the variable volume tank, reduces tank volume error, and minimizes jamming and excessive vibration during the directional transport of Salvia miltiorrhiza root segments.

[0062] The automatic quantitative feeding system for Salvia miltiorrhiza root segments provided by this invention has at least the following technical effects or advantages:

[0063] This invention utilizes a variable-volume trough in the vibration conveying module and synchronous laser scanning to adapt to the irregular shape and moisture content fluctuations of Salvia miltiorrhiza root segments, reducing the squeezing and jamming issues inherent in traditional vibration conveying. An adaptive metering module integrates machine vision and near-infrared spectral analysis to eliminate invalid Salvia miltiorrhiza root segments such as broken or hollow segments. The vibration control module, based on LSTM migration prediction and multimodal perception fusion, adjusts the nonlinear correction coefficient according to the conveying conditions of the Salvia miltiorrhiza root segments, achieving dynamic optimization of the vibration frequency. An anti-blocking frequency triggering mechanism is also added to reduce the risk of blockage and breakage in high-moisture-content root segments. The volume control module combines three-dimensional point cloud voxelization processing and spectral moisture content compensation to adjust the volume of the variable-volume trough according to the characteristics of the Salvia miltiorrhiza root segments, ensuring smooth conveying while avoiding waste caused by over-expansion. Attached Figure Description

[0064] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0065] Figure 1 This is a schematic diagram of the structure of an automatic quantitative feeding system for Salvia miltiorrhiza root segments according to the present invention. Detailed Implementation

[0066] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0067] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0068] Example 1

[0069] like Figure 1 As shown, the present invention provides an automatic quantitative feeding system for Salvia miltiorrhiza root segments, comprising:

[0070] The vibration conveying module is used to directionally convey the root segments of Salvia miltiorrhiza through controllable vibration and a variable volume tank, and to acquire the three-dimensional point cloud data of the root segments of Salvia miltiorrhiza in real time based on laser contour scanning.

[0071] The adaptive measurement module is used to detect the root segments of Salvia miltiorrhiza through machine vision and near-infrared spectroscopy, and obtain the image information, spectral information and actual effective quantity of the root segments.

[0072] The intelligent feeding module is used to control the opening and closing device to feed the Salvia miltiorrhiza root segments quantitatively based on the actual effective quantity of the root segments.

[0073] The vibration control module is used to predict the initial vibration frequency of controllable vibration based on historical transmission data using an LSTM network. Based on the image information of the Salvia miltiorrhiza root segments, the spectral information of the Salvia miltiorrhiza root segments, and the actual effective number of Salvia miltiorrhiza root segments, the initial vibration frequency is corrected to obtain the actual vibration frequency of controllable vibration.

[0074] The volume control module is used to dynamically adjust the volume of the Salvia miltiorrhiza delivery tank based on the three-dimensional point cloud data and spectral information of the Salvia miltiorrhiza root segments, and deliver the Salvia miltiorrhiza root segments through the variable volume tank after volume adjustment.

[0075] In one embodiment of the present invention, this invention addresses the problems of entanglement and adhesion caused by the varying shapes and high surface friction coefficients of *Salvia miltiorrhiza* root segments during transport, as well as the disordered accumulation and blockage caused by differences in root segment dimensions in traditional vibratory conveying. The present invention designs a vibratory conveying module, which includes:

[0076] The vibration generating unit is used to receive vibration frequency correction signals and generate controllable mechanical vibration based on the vibration frequency correction signals. Through controllable mechanical vibration, it can adapt to the specific working conditions of the current Salvia miltiorrhiza root segment vibration conveying, and reduce the occurrence of blockage and entanglement. Among them, an electromagnetic exciter can be used to receive the vibration frequency correction signal and then perform mechanical vibration, or a motor can be used to perform reciprocating linear motion based on the vibration frequency correction signal and other conventional technical means to achieve controllable mechanical vibration.

[0077] The elastic support unit is used to transmit controllable mechanical vibration to the variable volume tank that carries the Salvia miltiorrhiza root segments, and to convert the controllable mechanical vibration into directional vibration of the variable volume tank along the conveying direction. The Salvia miltiorrhiza root segments are then conveyed directionally through the variable volume tank. The specific structural design of the variable volume tank can refer to existing technologies, such as CN114763615B - a sodium hypochlorite generator with variable volume, CN209089650U - a pig feed trough for breeding, CN206368198U - a variable volume pickling tank for lining, and CN102481992B - a variable volume corrugated bottle. The elastic support unit converts the vertical mechanical vibration of the vibration generating unit into directional vibration waves in the conveying direction, thereby realizing the directional conveying of the Salvia miltiorrhiza root segments.

[0078] A laser scanning unit is used to synchronously collect vibration data of the Salvia miltiorrhiza root segments within a variable volume tank, acquiring three-dimensional point cloud data of the root segments. The scanning head of the laser scanning unit is fixedly connected to the variable volume tank via a rigid bracket to achieve synchronous vibration data acquisition of the Salvia miltiorrhiza root segments. Obtaining the current dimensions of the Salvia miltiorrhiza root segments from the three-dimensional point cloud data is a pre-existing technical method, such as CN112581621A - A method for online extraction of three-dimensional point cloud spatial dimensions of steel plates, and CN107504917B - A method and device for measuring three-dimensional dimensions. These methods can measure the volume and dimensions of objects using their three-dimensional point cloud data, facilitating the adjustment of the volume of the variable volume tank.

[0079] The variable volume trough dynamically adjusts its cross-sectional area according to the size distribution of the Salvia miltiorrhiza root segments. When large Salvia miltiorrhiza root segments are detected to be clustered, the volume is expanded to avoid compression. When small Salvia miltiorrhiza root segments are detected to be clustered, the volume is contracted to prevent disorderly rolling of single root segments, thus maintaining a stable conveying state of the Salvia miltiorrhiza root segments.

[0080] In one embodiment of the present invention, to address the problem that the actual effective quantity of *Salvia miltiorrhiza* root segments is inaccurately counted due to defects such as irregular shape, easy adhesion and breakage, and internal hollow damage that cannot be visually identified, an adaptive measurement module achieves accurate and effective measurement of *Salvia miltiorrhiza* root segments through a multimodal perception and decision fusion mechanism using visible light images and near-infrared light images. The adaptive measurement module includes:

[0081] The multispectral synchronous acquisition unit is used to simultaneously acquire images and spectra of the root segments of Salvia miltiorrhiza, and obtain the visible light images and raw infrared spectral data of the root segments of Salvia miltiorrhiza.

[0082] The feature fusion preprocessing unit is used to process the visible light image to obtain the image information of the Salvia miltiorrhiza root segment, process the raw infrared spectral data to obtain the spectral information of the Salvia miltiorrhiza root segment, and register the image information and spectral information to obtain the registered image and registered spectrum of the Salvia miltiorrhiza root segment. Among them, adaptive histogram equalization can be performed on the visible light image to enhance the details of the electro-optical image of the Salvia miltiorrhiza root segment, suppress noise contrast, and improve the visual quality of the image. Canny edge detection can be performed on the raw infrared spectral data to extract the initial contour value of the Salvia miltiorrhiza root segment.

[0083] The material morphology recognition unit is used to segment the contour of the Salvia miltiorrhiza root segment in the registration image using an image segmentation network to obtain the morphological integrity signal of each Salvia miltiorrhiza root segment, extract the material features of the Salvia miltiorrhiza root segment from the registration spectrum to obtain the material qualification signal of each Salvia miltiorrhiza root segment, and mark the Salvia miltiorrhiza root segment as a valid Salvia miltiorrhiza root segment when both the morphological integrity signal and the material qualification signal of the Salvia miltiorrhiza root segment exist.

[0084] Among them, the image segmentation network can use the U-Net segmentation network for semantic segmentation, identify broken and incomplete Salvia miltiorrhiza root segments and complete Salvia miltiorrhiza root segments, record the morphological integrity signal of the Salvia miltiorrhiza root segments, extract the absorbance of characteristic bands from the registration spectrum, and judge the tissue health of the Salvia miltiorrhiza root segments by the absorbance. When a Salvia miltiorrhiza root segment has both morphological integrity signal and material qualification signal, this Salvia miltiorrhiza root segment is marked as a valid Salvia miltiorrhiza root segment identifier.

[0085] The Salvia miltiorrhiza root segment counting unit is used to count the number of valid Salvia miltiorrhiza root segment identifiers and obtain the actual effective number of Salvia miltiorrhiza root segments.

[0086] The adaptive metering module of this invention can identify external fractures and internal damage of Salvia miltiorrhiza root segments from visible light and near-infrared light, and obtain the actual effective quantity of Salvia miltiorrhiza root segments. When invalid Salvia miltiorrhiza root segments are detected, they can be removed during the transportation process, reducing the mixing of invalid Salvia miltiorrhiza root segments, realizing the identification and effective counting of Salvia miltiorrhiza root segments, and reducing the mixing of hollow and damaged Salvia miltiorrhiza root segments into the feed.

[0087] In one embodiment of the present invention, the intelligent feeding module controls the opening and closing device to quantitatively feed the Salvia miltiorrhiza root segments according to the actual effective quantity of the segments. This enables precise control of the fed Salvia miltiorrhiza root segments, achieving quantitative feeding. Specifically, this includes:

[0088] The actual effective quantity of Salvia miltiorrhiza root segments is received, and the quantity of effective Salvia miltiorrhiza root segments, ineffective Salvia miltiorrhiza root segments, and the number of times the opening and closing device is opened and closed are obtained. Among them, the number of times the opening and closing device is opened and closed is equal to the actual effective quantity of Salvia miltiorrhiza root segments.

[0089] The passage of the Salvia miltiorrhiza root segment is detected by a photoelectric sensor installed in the opening and closing device. When the opening and closing device detects the passage of the Salvia miltiorrhiza root segment, a passage signal is obtained. When the opening and closing device does not detect the passage of the Salvia miltiorrhiza root segment, an error signal is obtained.

[0090] Increase the number of times the opening and closing device is activated based on the number of error signals.

[0091] Determine whether the number of signals passed is equal to the actual effective number. If so, it means that all effective Salvia miltiorrhiza root segments have been fed, and the quantitative feeding of Salvia miltiorrhiza root segments has been completed. Otherwise, there are cases where effective Salvia miltiorrhiza root segments have not been fed or where ineffective Salvia miltiorrhiza root segments have been fed.

[0092] In one embodiment of the present invention, addressing the problem of unadjustable vibration frequency during the vibration transport of Salvia miltiorrhiza root segments due to fluctuations in moisture content, size differences, and changes in stacking state, leading to transport blockages and breakage, the present invention achieves dynamic real-time optimization of the vibration frequency through transfer learning and multimodal perception fusion of the vibration control module. The vibration control module includes:

[0093] The historical database unit is used to store a data matrix of Salvia miltiorrhiza root segments, including historical vibration frequency, ambient temperature and humidity, average root segment size, and actual feeding quantity. Data can be obtained from historical Salvia miltiorrhiza root segment quantitative feeding operations.

[0094] The LSTM transfer prediction unit is used to adapt the pre-trained LSTM network to the Salvia miltiorrhiza root segment delivery scenario through transfer learning. It takes the Salvia miltiorrhiza root segment data matrix as input and outputs the initial vibration frequency. Transfer learning is an existing technology and can be found in CN114358190B-Transfer Learning Method and Apparatus. This invention transfers the pre-trained LSTM network in the source domain to the target domain, namely the Salvia miltiorrhiza root segment scenario, through transfer learning. This enables the LSTM network after transfer learning to have the prediction accuracy of the source domain and the adaptability of the target domain Salvia miltiorrhiza root segment scenario, thereby achieving accurate prediction of the initial vibration frequency.

[0095] The multimodal perception fusion unit is used to extract the bulk density of *Salvia miltiorrhiza* root segments based on image information, extract the moisture content gradient of *Salvia miltiorrhiza* root segments based on spectral information, and generate frequency correction coefficients based on the bulk density, moisture content gradient, and actual effective quantity of *Salvia miltiorrhiza* root segments. Specifically, the bulk density of *Salvia miltiorrhiza* root segments can be obtained by referring to existing technology CN112767343A - an online calculation method for coal bulk density on a belt conveyor, and the moisture content gradient of *Salvia miltiorrhiza* root segments can be obtained by referring to existing technology CN113533248A - a near-infrared spectral analysis method for crude oil moisture content in refining enterprises.

[0096] A frequency synthesizer is used to obtain the actual vibration frequency of controllable vibration based on the initial vibration frequency and the frequency correction coefficient.

[0097] The pre-trained LSTM network can be trained using an industrial vibration dataset, which includes vibration frequency, temperature, humidity, material size, and material flow rate. Through the LSTM migration prediction unit, the working vibration data can be transferred to the scenario of Salvia miltiorrhiza root segment transportation. It can accurately output the initial vibration frequency of Salvia miltiorrhiza root segments even with limited sample data. Then, through the multimodal perception fusion unit, the bulk density of Salvia miltiorrhiza root segments, the moisture content gradient of Salvia miltiorrhiza root segments, and the actual effective quantity of Salvia miltiorrhiza root segments are fused with the initial vibration frequency to improve the adaptability of the vibration frequency to the current Salvia miltiorrhiza root segment transportation conditions. This achieves the matching of the vibration frequency of the vibration transportation with the state of the Salvia miltiorrhiza root segments, reducing the risk of blockage and breakage of the Salvia miltiorrhiza root segments.

[0098] In one embodiment of the present invention, a pre-trained LSTM network is transferred to the Salvia miltiorrhiza root segment scenario using an LSTM transfer prediction unit. However, the physical characteristics of industrial materials differ from those of Salvia miltiorrhiza root segments, leading to a semantic gap in the domain, and there is also the problem of model overfitting due to insufficient data in the Salvia miltiorrhiza root segment transportation scenario. To address this, the present invention adds a Salvia miltiorrhiza feature embedding sub-network and a feature fusion layer, enabling the initial vibration frequency output by the transferred LSTM model to possess both the robustness of an industrial model and the adaptability to the Salvia miltiorrhiza root segment transportation scenario. The pre-trained LSTM network is configured as follows:

[0099] An industrial vibration dataset containing vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate is used to pre-train an LSTM network. The LSTM network includes:

[0100] The input layer, consisting of 5 neurons, receives vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate, respectively. This can be denoted as the Danshen root segment dataset, which corresponds to the vibration frequency, temperature, humidity, material size, and material flow rate in the industrial vibration dataset.

[0101] The hidden layer includes a first LSTM layer, a Salvia miltiorrhiza feature embedding sub-network, a feature fusion layer, and a second LSTM layer connected in sequence. The first LSTM layer and the second LSTM layer each contain 64 neurons and can extract the temporal features of the industrial vibration dataset. The Salvia miltiorrhiza feature embedding sub-network can encode the visible light image information and near-infrared spectral information of the Salvia miltiorrhiza root segment into a semantic vector of the same dimension as the industrial features output by the first LSTM layer. The feature fusion layer fuses the industrial features output by the first LSTM layer with the semantic vector output by the Salvia miltiorrhiza feature embedding sub-network.

[0102] The output layer consists of one linear neuron, which outputs the predicted initial vibrational frequency.

[0103] The pre-trained LSTM network of this invention is adapted to the Salvia miltiorrhiza root segment delivery scenario through transfer learning. Specifically, it includes: freezing the parameters and pre-trained weights of the first LSTM layer, which can lock the weights of the first LSTM layer and prevent fine-tuning of the Salvia miltiorrhiza root segment dataset from destroying the universality; performing supervised fine-tuning on the Salvia miltiorrhiza feature embedding sub-network, feature fusion layer and second LSTM layer; and outputting the vibration frequency of the Salvia miltiorrhiza root segment scenario. Specifically, the supervised fine-tuning on the Salvia miltiorrhiza feature embedding sub-network, feature fusion layer and second LSTM layer includes: training the Salvia miltiorrhiza feature embedding sub-network, feature fusion layer and second LSTM layer using the Salvia miltiorrhiza root segment dataset; and using the mean squared error loss function to optimize the vibration frequency prediction and improve the prediction accuracy.

[0104] This invention reduces overfitting caused by the scarcity of Salvia miltiorrhiza root segment datasets by freezing the first LSTM layer to retain general vibration frequency knowledge; the Salvia miltiorrhiza feature embedding subnetwork encodes the visible light image information and near-infrared spectral information of Salvia miltiorrhiza root segments into semantic vectors of the same dimension as industrial features; the feature fusion layer dynamically balances industrial laws and the characteristics of Salvia miltiorrhiza root segments through trainable weights, so that the initial frequency prediction value has both the robustness of industrial models and the adaptability of Salvia miltiorrhiza root segment scenarios.

[0105] In one embodiment of the present invention, in order to address the problem that the original knowledge domain of the pre-trained LSTM network has a large semantic gap with the Salvia miltiorrhiza root segment recognition scenario, making it difficult for the transferred LSTM network to effectively capture the coupling relationship between the Salvia miltiorrhiza root segment material and the vibration frequency, the present invention integrates the original knowledge domain with the Salvia miltiorrhiza recognition knowledge domain through a Salvia miltiorrhiza feature embedding sub-network and a feature fusion layer, thereby realizing the knowledge transfer of the original knowledge domain of the LSTM network.

[0106] The Danshen feature embedding subnetwork includes an input layer and a fully connected layer;

[0107] The input layer is used to receive the image information, spectral information, and actual effective quantity of the Salvia miltiorrhiza root segments. At the same time, data normalization processing is required. Min-max normalization can be used to normalize the image information, spectral information, and actual effective quantity of the Salvia miltiorrhiza root segments to between 0 and 1, which facilitates feature transformation by the fully connected layer.

[0108] The fully connected layer is used to map the image information, spectral information and actual effective number of Salvia miltiorrhiza root segments to obtain the feature vector of Salvia miltiorrhiza root segments. The fully connected layer adopts non-linear transformation and is set to 64 neurons to obtain a 64-dimensional feature vector of Salvia miltiorrhiza root segments, so as to adapt to the 64-dimensional feature vector of the first LSTM layer and the second LSTM layer.

[0109] The feature fusion layer is used to fuse the output feature vector of the first LSTM layer with the feature vector of the Salvia miltiorrhiza root segment to obtain the fused feature vector of the Salvia miltiorrhiza root segment. The fusion weights can be obtained by training the Salvia miltiorrhiza root segment dataset to balance the contribution of industrial vibration frequency and Salvia miltiorrhiza root segment scene.

[0110] In this invention, the fully connected layer of the Salvia miltiorrhiza feature embedding subnetwork maps the image information, spectral information, and actual effective quantity of Salvia miltiorrhiza root segments into a feature vector of Salvia miltiorrhiza root segments with the same dimension as the output feature vector of the first LSTM layer through nonlinear transformation, thus eliminating dimensional differences and modal gaps. The feature fusion layer can fuse the feature vector of Salvia miltiorrhiza root segments with the output feature vector of the first LSTM layer to obtain a fused feature vector of Salvia miltiorrhiza root segments that retains the original feature knowledge and is adapted to the Salvia miltiorrhiza root segments. This ensures that the output initial vibration frequency has both the original knowledge domain of the pre-trained LSTM network and the knowledge domain of Salvia miltiorrhiza recognition, thus guaranteeing the adaptability of the Salvia miltiorrhiza root segment delivery process.

[0111] In one embodiment of the present invention, addressing the problem of clogged or damaged Salvia miltiorrhiza root segments due to the inability to adjust the vibration frequency in real time to adapt to changes in the operating conditions of Salvia miltiorrhiza during transportation, the present invention combines the bulk density of Salvia miltiorrhiza root segments, the moisture content gradient of Salvia miltiorrhiza, and the actual effective quantity of Salvia miltiorrhiza root segments to adjust the actual vibration frequency in real time. This achieves real-time matching between the vibration frequency and the transportation conditions, improving the robustness and adaptability of directional transportation of Salvia miltiorrhiza root segments. The expression for the actual vibration frequency is as follows:

[0112]

[0113]

[0114]

[0115] in, The actual vibration frequency, The initial vibration frequency, This is the frequency correction factor. To prevent blockage in the root segments of Salvia miltiorrhiza, For the anti-blocking coefficient, For time, For indicator functions, For the frequency correction coefficient threshold, This refers to the bulk density of Salvia miltiorrhiza root segments. Water content gradient of Salvia miltiorrhiza root segments This refers to the actual effective quantity of Salvia miltiorrhiza root segments.

[0116] When the bulk density of *Salvia miltiorrhiza* root segments increases and the moisture content changes abruptly, the frequency correction coefficient increases to enhance the intensity of vibration and reduce root segment blockage. Conversely, when the actual effective number of *Salvia miltiorrhiza* root segments increases, the frequency correction coefficient decreases to reduce excessive vibration damage. When the frequency correction coefficient exceeds the threshold, it indicates an increase in the volume and moisture content of *Salvia miltiorrhiza* root segments in the visible light image, making root segment blockage more likely. The value is 1. An anti-blocking frequency is added to the *Salvia miltiorrhiza* root segment, instantaneously increasing the vibration frequency to prevent large-volume *Salvia miltiorrhiza* root segments from blocking transport and causing adhesion. When the frequency correction coefficient is less than the frequency correction coefficient threshold, The frequency is set to 0, and the current vibration frequency is maintained for conveying. This invention reduces the risk of blockage and breakage of Salvia miltiorrhiza root segments by using the anti-blockage frequency of Salvia miltiorrhiza root segments, while ensuring feeding accuracy.

[0117] In one embodiment of the present invention, to address the problem of excessively large or insufficient volume of the transport tank caused by uneven size and fluctuating moisture content of Salvia miltiorrhiza root segments, leading to mutual compression or blockage of the root segments, the present invention implements adaptive volume control of the moisture content and blockage status of the Salvia miltiorrhiza root segments through a volume control module. The volume control module includes:

[0118] Voxelization units are used to voxelize the 3D point cloud data of Salvia miltiorrhiza root segments to obtain the density distribution matrix of the Salvia miltiorrhiza root segments, as shown in the following expression:

[0119]

[0120] in, The density distribution matrix is The x-coordinate of the raster. The vertical coordinate of the raster is... This is the three-dimensional point cloud data of the root segment of Salvia miltiorrhiza. For voxelization operations;

[0121] The CNN prediction unit is used to input the density distribution matrix into a pre-trained convolutional neural network volume prediction model and output the target volume value. The pre-trained convolutional neural network volume prediction model is a commonly used technique. This model can be configured with three convolutional layers and two fully connected layers. All three convolutional layers have 3×3 kernels with a stride of 1 and ReLU activation. The two fully connected layers have 128 to 64 neurons and use Sigmoid activation. The expression for the target volume value is as follows:

[0122]

[0123] in, For the target volume value, A pre-trained convolutional neural network volume prediction model;

[0124] The PID control unit is used to generate a volume adjustment command signal based on the deviation between the target volume value and the current measured value from the volume sensor. The expression is as follows:

[0125]

[0126]

[0127] in, This is a volume adjustment command signal. , , All are PID control adjustment coefficients. This represents the current response deviation. To accumulate historical bias, The rate of change of deviation This is the current measured value from the volume sensor;

[0128] The spectral compensation unit is used to compensate for the moisture content of the volume adjustment command signal based on the spectral information of the Salvia miltiorrhiza root segment, resulting in a corrected volume adjustment command signal, expressed as follows:

[0129]

[0130] in, For the corrected volume adjustment command signal, The water content is obtained based on the spectral information of the Salvia miltiorrhiza root segment. This refers to the deviation in spectral moisture content.

[0131] The volume control unit is used to adjust the volume of the Salvia miltiorrhiza delivery tank according to the modified volume control command signal.

[0132] The voxelization unit of this invention voxelizes three-dimensional point cloud data into a density matrix characterizing the spatial distribution of Salvia miltiorrhiza root segments; the density matrix of Salvia miltiorrhiza root segments is received by the CNN prediction unit, and the target volume value is output; the PID control unit generates a preliminary volume adjustment command signal based on the deviation between the target volume value and the measured volume value; the spectral compensation unit corrects the volume adjustment command signal based on the near-infrared spectral moisture content; and the volume execution unit completes the volume adjustment of the variable volume tank according to the corrected volume adjustment command signal.

[0133] This invention can estimate the target volume based on three-dimensional point cloud data, and then determine the specific volume based on the moisture content, so as to accurately control the volume, reduce the occurrence of root segment accumulation and blockage, and reduce the waste of vibration energy caused by excessive expansion, thereby achieving adaptive volume control based on the moisture content and blockage of Salvia miltiorrhiza root segments.

[0134] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0135] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An automatic quantitative feeding system for Salvia miltiorrhiza root segments, characterized in that, include: The vibration conveying module is used to directionally convey the root segments of Salvia miltiorrhiza through controllable vibration and a variable volume tank, and to acquire the three-dimensional point cloud data of the root segments of Salvia miltiorrhiza in real time based on laser contour scanning. The adaptive measurement module is used to detect the root segments of Salvia miltiorrhiza through machine vision and near-infrared spectroscopy, and obtain the image information, spectral information and actual effective quantity of the root segments. The intelligent feeding module is used to control the opening and closing device to feed the Salvia miltiorrhiza root segments quantitatively based on the actual effective quantity of the root segments. The vibration control module is used to predict the initial vibration frequency of controllable vibration based on historical transmission data using an LSTM network. Based on the image information of the Salvia miltiorrhiza root segments, the spectral information of the Salvia miltiorrhiza root segments, and the actual effective number of Salvia miltiorrhiza root segments, the initial vibration frequency is corrected to obtain the actual vibration frequency of controllable vibration. The volume control module is used to dynamically adjust the volume of the Salvia miltiorrhiza delivery tank based on the three-dimensional point cloud data and spectral information of the Salvia miltiorrhiza root segments, and deliver the Salvia miltiorrhiza root segments through the variable volume tank after volume adjustment.

2. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 1, characterized in that, The vibration transmission module includes: The vibration generating unit is used to receive vibration frequency correction signals and generate controllable mechanical vibrations based on the vibration frequency correction signals. The elastic support unit is used to transmit controllable mechanical vibration to the variable volume trough that carries the Salvia miltiorrhiza root segments, and to convert the controllable mechanical vibration into directional vibration of the variable volume trough along the conveying direction, so as to directionally convey the Salvia miltiorrhiza root segments through the variable volume trough. The laser scanning unit is used to synchronously collect vibration data of the Salvia miltiorrhiza root segments within the variable volume tank, thereby acquiring three-dimensional point cloud data of the Salvia miltiorrhiza root segments.

3. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 1, characterized in that, The adaptive metering module includes: The multispectral synchronous acquisition unit is used to simultaneously acquire images and spectra of the root segments of Salvia miltiorrhiza, and obtain the visible light images and raw infrared spectral data of the root segments of Salvia miltiorrhiza. The feature fusion preprocessing unit is used to process the visible light image to obtain the image information of the Salvia miltiorrhiza root segment, process the raw infrared spectrum data to obtain the spectral information of the Salvia miltiorrhiza root segment, and register the image information and spectral information to obtain the registered image and registered spectrum. The material morphology recognition unit is used to segment the contour of the Salvia miltiorrhiza root segment in the registration image using an image segmentation network to obtain the morphological integrity signal of each Salvia miltiorrhiza root segment, extract the material features of the Salvia miltiorrhiza root segment from the registration spectrum to obtain the material qualification signal of each Salvia miltiorrhiza root segment, and mark the Salvia miltiorrhiza root segment as a valid Salvia miltiorrhiza root segment when both the morphological integrity signal and the material qualification signal of the Salvia miltiorrhiza root segment exist. The Salvia miltiorrhiza root segment counting unit is used to count the number of valid Salvia miltiorrhiza root segment identifiers and obtain the actual effective number of Salvia miltiorrhiza root segments.

4. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 1, characterized in that, The vibration control module includes: The historical database unit is used to store a data matrix of Salvia miltiorrhiza root segments, including historical vibration frequencies, ambient temperature and humidity, average root segment size, and actual feeding quantity. The LSTM transfer prediction unit is used to adapt the Salvia miltiorrhiza root segment delivery scenario through transfer learning using a pre-trained LSTM network. It takes the Salvia miltiorrhiza root segment data matrix as input and outputs the initial vibration frequency. The multimodal perception fusion unit is used to extract the bulk density of Salvia miltiorrhiza root segments based on image information, extract the water content gradient of Salvia miltiorrhiza root segments based on spectral information, and generate frequency correction coefficients based on the bulk density, water content gradient, and actual effective quantity of Salvia miltiorrhiza root segments. A frequency synthesizer is used to obtain the actual vibration frequency of controllable vibration based on the initial vibration frequency and the frequency correction coefficient.

5. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 4, characterized in that, The pre-trained LSTM network is configured as follows: An industrial vibration dataset containing vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate is used to pre-train an LSTM network. The LSTM network includes: The input layer consists of 5 neurons, which receive vibration frequency, ambient temperature, ambient humidity, average material size, and material flow rate, respectively. The hidden layer comprises a first LSTM layer, a Danshen feature embedding subnetwork, a feature fusion layer, and a second LSTM layer connected in sequence. Both the first LSTM layer and the second LSTM layer contain 64 neurons. The output layer consists of one linear neuron, which outputs the predicted initial vibrational frequency. The method of adapting to the delivery scenario of Salvia miltiorrhiza root segments through transfer learning specifically includes: freezing the parameters and pre-trained weights of the first LSTM layer, performing supervised fine-tuning on the Salvia miltiorrhiza feature embedding sub-network, feature fusion layer and second LSTM layer, and outputting the vibration frequency of the Salvia miltiorrhiza root segment scenario.

6. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 5, characterized in that, The Danshen feature embedding subnetwork includes an input layer and a fully connected layer; The input layer is used to receive image information of Salvia miltiorrhiza root segments, spectral information of Salvia miltiorrhiza root segments, and the actual effective quantity of Salvia miltiorrhiza root segments. The fully connected layer is used to map the image information of the Salvia miltiorrhiza root segments, the spectral information of the Salvia miltiorrhiza root segments, and the actual effective number of Salvia miltiorrhiza root segments to obtain the feature vector of the Salvia miltiorrhiza root segments. The feature fusion layer is used to fuse the output feature vector of the first LSTM layer with the feature vector of the Salvia miltiorrhiza root segment to obtain the Salvia miltiorrhiza root segment fused feature vector.

7. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 4, characterized in that, The expression for the actual vibration frequency is as follows: in, The actual vibration frequency, The initial vibration frequency, This is the frequency correction factor. To prevent blockage in the root segments of Salvia miltiorrhiza, For the anti-blocking coefficient, For time, For indicator functions, For the frequency correction coefficient threshold, This refers to the bulk density of Salvia miltiorrhiza root segments. This refers to the actual effective quantity of Salvia miltiorrhiza root segments.

8. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 1, characterized in that, The volume control module includes: Voxelization unit is used to voxelize the three-dimensional point cloud data of Salvia miltiorrhiza root segments to obtain the density distribution matrix of Salvia miltiorrhiza root segments. The CNN prediction unit is used to input the density distribution matrix into a pre-trained convolutional neural network volume prediction model and output the target volume value. The PID control unit is used to generate a volume adjustment command signal based on the deviation between the target volume value and the current measured value of the volume sensor. The spectral compensation unit is used to compensate the volume adjustment command signal for moisture content based on the spectral information of the Salvia miltiorrhiza root segment, so as to obtain a corrected volume adjustment command signal. The volume control unit is used to adjust the volume of the Salvia miltiorrhiza delivery tank according to the modified volume control command signal.

9. The automatic quantitative feeding system for Salvia miltiorrhiza root segments according to claim 8, characterized in that, The expression for the modified volume adjustment command signal is as follows: in, For the corrected volume adjustment command signal, This is a volume adjustment command signal. The water content is obtained based on the spectral information of the Salvia miltiorrhiza root segment. For spectral moisture content deviation, , , All are PID control adjustment coefficients. This represents the current response deviation. To accumulate historical bias, The rate of change of deviation For the target volume value, This is the current measured value from the volume sensor. For a pre-trained convolutional neural network volume prediction model, The density distribution matrix is The x-coordinate of the raster. The vertical coordinate of the raster is... This is the three-dimensional point cloud data of the root segment of Salvia miltiorrhiza. This is a voxelization operation.

Images