Liquid drop volume identification method based on YOLOv11 instance segmentation and closed-loop two-dimensional sample application system

Abstract

The invention discloses a liquid drop volume identification method based on YOLOv11 instance segmentation and a closed-loop two-dimensional sample application system. The method comprises the following steps: recording a liquid drop growth video under the condition that a liquid drop at a nozzle naturally drips, dividing the liquid drop growth video into a plurality of pictures according to a frame rate, marking masks of the liquid drop and the nozzle in the pictures, and dividing a training set, a test set and a verification set; establishing an improved YOLOv11 model, and after training, processing the collected image of the liquid drop at the nozzle to obtain the data of the liquid drop and the nozzle area; a liquid drop area-volume conversion module is built, the liquid drop area is converted into liquid drop volumes, and the liquid drop volumes at different times are obtained; and establishing a time feature LSTM model for obtaining prediction data corresponding to the historical data, and predicting the volume of the liquid drop according to the image of the liquid drop at the nozzle, which is acquired in real time. According to the invention, the sample application efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of microfluidics and liquid spotting technology, and in particular, it is a droplet volume identification method and a closed-loop two-dimensional spotting system based on YOLOv11 instance segmentation. Background Technology

[0002] Traditional manual spotting methods suffer from low efficiency, large operational errors, and difficulty in scaling up, making them unsuitable for the demands of modern high-throughput experiments. Against this backdrop, automated spotting equipment combining microfluidics, precision mechanical control, and computer vision is gradually becoming a core tool for standardized laboratory operations. Furthermore, current cutting-edge technologies such as PCR and gene sequencing often involve multiple rounds of precise spotting operations, placing even higher demands on the automation, accuracy, and stability of the equipment.

[0003] Current mechanical spotting systems mostly mimic manual operation. During the spotting process, the pipette tip needs to be replaced after each spotting position on a microplate to avoid liquid remaining on the tip affecting the spotting volume, thus increasing the spotting cost. In addition, the minimum pipetting volume of a pipette is only about a microliter, which limits the minimum spotting volume and leads to an increase in reagent consumption. If the reagent is diluted in advance, it introduces variables in reagent homogeneity and makes the operation more cumbersome.

[0004] In addition, there are acoustic pipetting systems, which cannot be used with reagents of high viscosity. Furthermore, each time a different reagent is used, the acoustic intensity needs to be recalibrated, resulting in additional time and reagent waste. The single-pipette volume of acoustic pipetting systems is approximately 2.5-25 nanoliters, which also limits their loading speed. Piezoelectric pipetting systems also require recalibration for different reagents; moreover, the high voltage required limits their application in biological fields.

[0005] The Chinese invention patent with patent application number "CN202310830177.3" entitled "A Droplet Ejection Method Based on Pulsed Airflow" proposes a method to achieve droplet ejection using pulsed airflow shearing. However, due to the long response time of the fluid pump, this method can only achieve continuous spotting of droplets of the same size. Adjusting the droplet size takes a long time and cannot fully utilize the advantage of the wide droplet generation range of this method. Summary of the Invention

[0006] The purpose of this invention is to provide a droplet volume recognition method based on YOLOv11 instance segmentation, which can be applied to a spotting system to achieve closed-loop control of droplet size during spotting operations and improve spotting speed.

[0007] The technical solution to achieve the purpose of this invention is as follows:

[0008] A droplet volume recognition method based on YOLOv11 instance segmentation includes:

[0009] Record a video of droplet growth under the condition of droplets falling naturally at the nozzle, divide it into multiple images according to the frame rate, label the droplets and nozzle masks in the images, and divide them into training set, test set and validation set;

[0010] An improved YOLOv11 model was established and trained. The collected droplet images at the nozzle were then processed to obtain droplet and nozzle region data.

[0011] A droplet region-volume conversion module was built to convert the droplet region into the droplet volume, thus obtaining the droplet volume at different times;

[0012] A time-feature LSTM model is established to obtain prediction data corresponding to historical data. Based on the droplet images collected in real time at the nozzle, the droplet volume is predicted.

[0013] A closed-loop two-dimensional spotting system, comprising:

[0014] Nozzle module;

[0015] XY mobile platform, used for installing liquid receiving devices;

[0016] The liquid micro-injection module is used to inject liquid into the nozzle module;

[0017] The air path module is used to generate airflow to break up the suspended droplets at the nozzle module;

[0018] The control module is used to control the XY moving platform, the liquid micro-injection module and the gas path module. By controlling the growth rate of the droplets at the nozzle module, it can adjust the time required for droplet growth to produce droplets of the required size.

[0019] The host computer is used to run the droplet volume recognition method described above.

[0020] The significant advantages of this invention compared to existing technologies are:

[0021] (1) The active droplet preparation method using pulsed gas flow avoids the trouble of recalibration required every time the spotting liquid is changed in the existing spotting system, and at the same time, it can spot liquids with higher viscosity, thus broadening the range of spotting liquids.

[0022] (2) The droplet volume is observed at the nozzle by the droplet volume recognition module, and the droplet volume is fed back in near real time. In open-loop control, the droplet size needs to be controlled by the pulse time interval. After the introduction of this module, the liquid volume flow control parameters are freed up, and the advantage of the wide droplet volume range of the droplet jet module is fully utilized, which improves the sampling efficiency.

[0023] (3) By modifying the neck structure of the original YOLOv11 model, data from the P2 layer with a downsampling rate of 4 is added and mixed with the original layer according to the feature pyramid structure, so that the detection head part can acquire higher resolution images, thereby enhancing the model's performance in small droplet detection.

[0024] (4) The nozzle and droplet profiles are identified separately using the YOLOv11 model. The nozzle profile is used to determine the axis direction, which is then used as the droplet axis direction. This avoids the difficulty of determining the droplet's main axis direction using complex algorithms when only the droplet profile is extracted, thus improving the detection speed. It also reduces the requirements for image acquisition equipment installation and improves detection adaptability. Furthermore, using the nozzle diameter to calibrate the image scale avoids the need for readjustment after replacing or adjusting the image acquisition equipment.

[0025] (5) By using the temporal feature LSTM, the time factor of each volume data point is taken into account, avoiding the impact of uneven distribution of data points over time caused by fluctuations in detection time on volume prediction, thus improving the accuracy of volume prediction. At the same time, multiple data points are predicted at once, reducing the computing power requirements of the equipment and improving the real-time performance of volume detection.

[0026] (6) Through the nozzle structure design, the nozzle is further encapsulated as a module, so that when different sampling ranges are required, only the module needs to be replaced, without disassembling the entire nozzle, thus improving efficiency. By using two screws to cooperate with the nozzle module, and the nozzle module to cooperate with the nozzle housing, the vertical position of the nozzle can be easily adjusted to keep it in the center of the image acquisition lens, avoiding droplet volume errors caused by position, and maintaining good positional accuracy and darkroom. Attached Figure Description

[0027] Figure 1 This is a structural diagram of the instance segmentation model used in this invention.

[0028] Figure 2 This is a flowchart of the droplet volume recognition module used in this invention.

[0029] Figure 3 This is a schematic diagram of the dimension transformation layer of the time feature LSTM model used in this invention.

[0030] Figure 4 This is a structural diagram of the prediction layer of the LSTM model used in this invention.

[0031] Figure 5 This is a schematic diagram of the closed-loop two-dimensional spotting system of the present invention.

[0032] Figure 6 This is a three-dimensional isometric drawing of the two-dimensional point sampling system of the present invention.

[0033] Figure 7 This is a three-dimensional isometric view of the nozzle used in this invention.

[0034] Figure 8 This is a cross-sectional view of the nozzle used in this invention.

[0035] Figure 9 These are the front view and partial sectional view of the nozzle used in this invention.

[0036] Figure 10 This is a top view of the nozzle used in this invention. Detailed Implementation

[0037] The following description, with reference to the accompanying drawings, further describes specific technical embodiments of the present invention to enable those skilled in the art to further understand the present invention, without constituting a limitation on its rights.

[0038] Combination Figure 1 This embodiment provides a droplet volume identification method based on YOLOv11 instance segmentation, including the following steps:

[0039] Step S1: Droplet image acquisition to obtain the original image.

[0040] Using an image acquisition lens, video was recorded showing droplet growth under conditions where no pulsed airflow was applied at the nozzle, allowing the droplets to fall naturally. Software was then used to divide the recorded video into multiple images based on frame rate.

[0041] Step S2: Manually annotate the images to obtain the YOLOv11 dataset.

[0042] The masks for droplets and nozzles in the images were manually labeled, and the labeled data were randomly divided into training set, test set, and validation set in a ratio of 7:2:1.

[0043] Step S3: Establish the improved YOLOv11 model. This model requires converting the acquired images into mask data of droplets and nozzles in the image. The following is a detailed explanation:

[0044] The original model consists of a backbone structure and a neck structure, see [link / reference]. Figure 1 .

[0045] In the backbone structure, starting from the image data input, there are sequentially one Conv convolutional module, four Conv convolutional modules and a C3k2 module, one SPPF module and a C2PSA module. For ease of representation, the nth Conv convolutional module through which the image data passes is called the Pn layer; the C3k2 module after the Pn layer is called the Pn feature extraction layer; the data after the Pn layer is processed by the C3k2 module is called the Pn layer feature extraction data.

[0046] The neck structure is divided into upward fusion and downward fusion sections.

[0047] The neck-down fusion section: The C2PSA structure connects to the neck-up fusion section on one hand, and through the Upsample module, it connects to a set of Concat and C3k2 modules along with the P4 feature extraction layer. Afterward, the C3k2 module, through the Upsample module, connects to the set of Concat and C3k2 modules along with the P3 feature extraction layer. For simplicity, the set of Concat and C3k2 modules connected to the Pn feature extraction layer is called the Pn-down feature fusion layer, and its corresponding output data is the Pn-layer down feature fusion data.

[0048] The upward fusion section of the neck: The P3 downward feature fusion layer, on one hand, connects to a set of Concat modules and C3k2 modules via the Conv module and the P4 downward feature fusion layer, and on the other hand, connects to the Detect head. After this, the C3k2 module, on the other hand, connects to a set of Concat modules and C3k2 modules via the Conv module and the C2PSA module, and on the other hand, connects to the Detect head. Finally, the C3k2 module also connects to the Detect head.

[0049] The neck structure of the original YOLOv11 model was modified by mixing the data from layer P2 with the original data from layers P3, P4, and P5 according to a feature pyramid structure. This allows the detection head to acquire higher-resolution images, thereby enhancing the model's performance in small droplet detection. See [link to documentation]. Figure 1 The area within the dashed box. Specific modifications are as follows:

[0050] In the neck section, two new Concat modules and C3k2 modules are added. The first module receives feature extraction data from layer P2 and feature fusion data from layer P3 after upsampling. Simultaneously, the output data is passed through a Conv convolution module and fed into the second Concat and C3k2 modules. The second module receives data from the first module and feature fusion data from layer P3. It replaces the P3 feature fusion layer, passes the output data through a Conv convolution module, and sends it to a Concat and C3k2 module connected to the P4 feature extraction layer, as well as to the Detect head.

[0051] Step S4: Use the YOLOv11 dataset obtained in step S2 to train the improved YOLOv11 model.

[0052] Step S5: Build the droplet region-volume conversion module. This module needs to process the droplet and nozzle region data given after image processing by the improved YOLOv11 model and convert it into the droplet volume at that moment. Specifically, it can be divided into the following steps:

[0053] Step S51, Contour Extraction. The droplet and nozzle region data (also known as droplet masks and nozzle masks) given after image processing by the improved YOLOv11 model are converted into nozzle contours and droplet contours. See [link to relevant documentation]. Figure 2 .

[0054] Assuming the original image has a pixel height of h and a width of w, the data format of the droplet mask and nozzle mask is an h×w 01 matrix, where the part that does not contain droplets or nozzles is 0, and the part where droplets or nozzles are located is 1.

[0055] The horizontal gradient G is computed using the Sobel operator. x and vertical gradient G y :

[0056]

[0057] Where B is the droplet mask and nozzle mask matrix, and in the following text, B at coordinates i and j in a pixel is denoted as Bij. ij .

[0058] Edge strength E ij :

[0059]

[0060] Among them, E ij Let G(i,j) be the edge intensity at coordinates i and j in the pixel. x Let G(i,j) be the horizontal gradient at coordinates i and j. y Let be the vertical gradient at coordinate points i and j.

[0061] The contour point is E. ij >0 and B ij =1 pixel.

[0062] Step S52: Perform principal component analysis on the nozzle profile to obtain the directions of its major axis (first principal component) and minor axis (second principal component), see [link to relevant documentation]. Figure 2 First, the contour point set is centered:

[0063]

[0064] Where N is the total number of contour points. Let x be the average of the x and y coordinates of N contour points in the contour point set. n ,y nLet P be the coordinates of the nth pixel in the contour points, and P be the set of contour vectors after centering. Then, construct the covariance matrix Cov:

[0065]

[0066] By calculating the eigenvectors of the covariance matrix Cov, we can obtain its first principal component direction v1 and second principal component direction v2, which are the axial and radial directions of the nozzle.

[0067] Step S53: Based on the direction of the second principal component, use the truncated mean method to calibrate the reference object according to the nozzle diameter. See [link to relevant documentation]. Figure 2 First, calculate the projection value l of the nth profile point in the radial direction of the nozzle. n :

[0068] l n =p n ·v2 (7)

[0069] Where, p n The nth element in the contour vector set P. The projection value l of the N contour points. n Create a set L, sort the elements in L in ascending order, truncate the middle 50% of the values ​​to remove outliers, and calculate the average. The scaling factor k is obtained by comparing it with the nozzle diameter:

[0070]

[0071] Among them l real This is the actual diameter of the nozzle.

[0072] Step S54: Calculate the droplet volume data using the volume integration method based on the direction of the first principal component. (See [link to relevant documentation]). Figure 2 First, calculate the projection value z of the nozzle along the axial direction at the nth contour point. n and the projection value r in the radial direction n :

[0073]

[0074] The volume V of the droplet in the image was calculated as follows:

[0075]

[0076] Step S55: Perform volume calibration to obtain the true volume V real :

[0077] V real =V·k 3 (11)

[0078] Step S6: Use an image acquisition lens to monitor droplets at the nozzle and use the improved YOLOv11 model to extract droplet regions at different times.

[0079] Step S7: Use the droplet region-volume conversion module to convert the droplet region into the droplet volume to obtain the droplet volume at different times.

[0080] Step S8: Establish a time-feature LSTM model. This model needs to process historical data points and predict droplet volume data points to improve the real-time performance of the detection. Here, a dimension transformation layer and a result decoding and output layer are added to the original LSTM model, mainly including the following structure:

[0081] Dimension Transformation Layer: The number of historical data points (M+1) input to the model is determined based on actual needs. Volumetric data and temporal data are separated; volumetric data is directly treated as a single dimension, while temporal data undergoes additional processing. Temporal data processing is as follows: The absolute time data is converted to a time interval Δt through subtraction, and then fed into two fully connected layers (see...). Figure 3 The number of neurons J in each fully connected layer is set according to the actual situation; here it is set to 8. Finally, the data is concatenated with the volumetric data and fed into the LSTM model. To enhance the model's ability to represent temporal information nonlinearly, the fully connected layer data is calculated as follows:

[0082] h j =ReLU(W1Δt+b1) (12)

[0083] Among them, h j The output of the j-th neuron is represented by ReLU, the activation function is ReLU, W1 is the neuron input weight term, and b1 is the bias term. During training, the data dimension is [b, M, J+1], where b is the batch size, which can be set according to requirements.

[0084] LSTM Model Prediction Layer: The LSTM model is used; see the structure below. Figure 4 The number of hidden neurons, I, is set according to requirements; here, we assume it to be 64, meaning 64 LSTM neuron structures are used for state transitions, providing sufficient capacity to remember long-term patterns. The number of hidden state outputs that can be collected at each time step is M, but here we only retain the hidden state output of the last time step for subsequent prediction. The data dimension at this point is [b, 1, I]. The update process for the LSTM unit is as follows:

[0085]

[0086] Where σ is the Sigmoid activation function, tanh is the Tanh activation function, and x t Given the input sequence, c t and c t-1These represent the memory unit states at the previous and current time points, respectively, h. t and h t-1 The hidden states of the previous and current time steps are respectively, W. f W i W c W o For the weights of each part, b f b i b c b o For each part of the bias term. f t i t , o t These represent the forget gate output, input gate output, memory cell candidate vector value, and output gate output at time t, respectively.

[0087] The result decoding and output layer: The data output by the LSTM model is mapped from the hidden state to the target dimension N through a fully connected layer. The data structure is then [b, 1, N], where N is the number of predicted data points output by the model, which can be set according to actual needs. Here, it is set to 5, corresponding to the data at 5 equally spaced time points within 0.1s intervals. This avoids the accumulation of errors caused by directly looping the LSTM model output.

[0088] Step S9: Prepare the dataset for the time feature LSTM model. Assuming the time feature LSTM model established in step S8 has M input data points and N output data points, and the total time for N data points is 0.1s, the steps are as follows:

[0089] Step S91: Obtain historical dataset. Using the droplet volume data at different times obtained in step S7, extract M+1 consecutive data points as one set of historical data. Assuming that the number of data points obtained in step S7 is S, then SM-2 sets of historical data can be extracted.

[0090] Step S92: Obtain the prediction dataset. First, using linear interpolation, the data obtained in step S7 is interpolated over time to obtain droplet volume data for every 0.1 / N(s), which is called the interpolation dataset. Then, based on the time point of the last data point in the historical dataset in step S91, the volume data of the 5 data points after that time point in the interpolation dataset are taken as the prediction data, for a total of SM-2 groups, which correspond one-to-one with the SM-2 groups of historical data in step S91.

[0091] Step S10: Train the temporal feature LSTM model.

[0092] Step S10: Deploy the trained improved YOLOv11 model, the droplet region-volume conversion module, and the trained temporal feature LSTM model to detect the image from the image acquisition lens and return the droplet volume.

[0093] Combination Figure 5 This embodiment provides a closed-loop two-dimensional spotting system based on a droplet volume recognition method using YOLOv11 instance segmentation. The system consists of a nozzle module, an image acquisition module, a liquid micro-injection module, a gas path module, an XY motion platform, a control module, and a host computer. An isometric drawing of the two-dimensional spotting system can be found here. Figure 6 ;

[0094] The nozzle module includes a nozzle 1, a left nozzle clamping block 2-1, a right nozzle clamping block 2-2, a rear nozzle housing 8-1, a front nozzle housing 8-2, a manual lead screw slide 4, and an adjusting screw 9. The left and right nozzle clamping blocks 2-1 and 2-2 are connected by screws, forming a single nozzle clamping block. This clamping block is pressure-connected to the nozzle 1 and screw-connected to the cover plate 7. The rear and front nozzle housings 8-1 and 8-2 are also connected by screws, forming the nozzle housing. The nozzle housing is screw-connected to the lead screw slide 4, which is fixed to the frame and used to adjust the height of the nozzle 1 relative to the XY moving platform. The left and right nozzle clamping blocks 2-1 and 2-2 are connected to the nozzle housing via the adjusting screw 9. The nozzle clamping blocks slide up and down within the grooves of the nozzle housing, and the threads on the outer side of the clamping blocks engage with the threads of the adjusting screw 9. See the completed isometric drawing. Figure 7 .

[0095] The image acquisition module includes an image acquisition lens 3, a diffuser 5, and a dot matrix LED 6. The image acquisition lens 3, the dot matrix LED 6, the lead screw slide 4, and the nozzle housing are connected by screws. The diffuser 5 is located between the nozzle 1 and the dot matrix LED 6, with the dot matrix LED 6 serving as a supplementary light source, and the diffuser 5 ensuring uniform light distribution. During installation, the adjusting screws 9 are placed at the same height on both sides of the nozzle clamping block, and together with the nozzle clamping block, they are placed in the slide groove of the nozzle housing for use. When adjusting the height, due to gravity and the support of the bottom of the slide groove of the nozzle housing, the relative position of the adjusting screws 9 and the nozzle housing remains unchanged. Simultaneously rotating both adjusting screws 9 adjusts the relative position of the image acquisition lens 3 and the nozzle 1.

[0096] The control module, connected to the liquid micro-injection module, gas path module, and XY moving platform, employs a microcontroller. The liquid micro-injection module uses a fluid pump, and the microcontroller connects to the fluid pump via a data interface. The fluid pump is connected to the nozzle's liquid phase inlet via a liquid path pipeline. The microcontroller controls the flow rate of the fluid pump, thereby controlling the droplet growth rate at the nozzle and adjusting the time required for droplet growth. In conjunction with the gas path module (described later), it generates droplets of the desired size within the required time. The gas path module includes a solenoid valve, a gas pump, and a relay. The microcontroller connects to the relay via a circuit, and the relay is also connected to the solenoid valve via a circuit. The gas pump connects to the solenoid valve via a gas path pipeline, and then to the nozzle's gas phase inlet. The microcontroller generates pulse signals to control the relay, which are then converted into voltage signals to control the opening and closing of the solenoid valve. This generates a pulsed airflow at the nozzle, causing the suspended droplets at the nozzle to break up prematurely. In conjunction with the micro-injection module, this generates droplets of the desired size. The microcontroller is connected to the XY moving platform via a circuit, which is used to mount the liquid receiving device. The microcontroller controls the XY mobile platform to move to the desired position by sending a certain number of pulse signals. This module works in conjunction with the modules described above to achieve sampling at different locations.

[0097] The host computer is used to run the aforementioned droplet volume recognition method, including an image processing module, a droplet region-volume conversion module, and a droplet volume prediction module. The image processing module obtains droplet and nozzle region data by running the improved YOLOv11 model. The droplet region-volume conversion module converts the droplet and nozzle region data obtained after image processing by the improved YOLOv11 model into droplet volume at the time of acquisition. The droplet volume prediction module predicts the droplet volume at future corresponding times by running the aforementioned temporal feature LSTM model and controls the sampling volume through a controller. Its motion control strategy is as follows:

[0098] Step S1: Start the image processing module, droplet region-volume conversion module, and droplet volume prediction module in the host computer.

[0099] Step S2: Calculate the time required for the XY mobile platform to move to the next location.

[0100] Step S3: Based on the time required for movement and the droplet size required for the next position, while ensuring time redundancy (here, the redundancy time is selected as 50ms), the liquid micro-injection module is configured with the fluid pump flow rate, and the host computer is set with the droplet volume threshold.

[0101] Step S4: The XY moving platform moves to the designated position under the control of the microcontroller and waits for the droplet volume prediction module to return that the droplet volume exceeds the threshold.

[0102] Step S5: The gas path module generates a pulsed airflow, which in turn generates the required droplets. After the droplets fall (a waiting time of 50ms is selected here), the process repeats to step S2 to execute the next point.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for droplet volume recognition based on YOLOv11 instance segmentation, characterized in that, include: Record a video of droplet growth under the condition of droplets falling naturally at the nozzle, divide it into multiple images according to the frame rate, label the droplets and nozzle masks in the images, and divide them into training set, test set and validation set; An improved YOLOv11 model was established and trained. The collected droplet images at the nozzle were then processed to obtain droplet and nozzle region data. A droplet region-volume conversion module was built to convert the droplet region into the droplet volume, thus obtaining the droplet volume at different times; A time-feature LSTM model is established to obtain prediction data corresponding to historical data. Based on the droplet images collected in real time at the nozzle, the droplet volume is predicted.

2. The droplet volume recognition method based on YOLOv11 instance segmentation according to claim 1, characterized in that, The improved YOLOv11 model modifies the neck structure of the original YOLOv11 model. The neck down-fusion part adds two sets of Concat modules and C3k2 modules: The first set receives feature extraction data from the P2 layer and down-fusion data from the P3 layer after the Upsample module. At the same time, the output data is passed through the Conv convolution module and sent to the second set of Concat and C3k2 modules. The second set receives data from the first set and down-fusion data from the P3 layer. It replaces the P3 down-fusion layer and sends the output data through the Conv convolution module to a set of Concat and C3k2 modules connected to the P4 feature extraction layer, as well as to the Detect head.

3. The droplet volume recognition method based on YOLOv11 instance segmentation according to claim 1, characterized in that, The processing steps of the droplet region-volume conversion module include: (1) The droplet and nozzle region data given after processing the image by the improved YOLOv11 model are converted into nozzle profile and droplet profile. (2) Perform principal component analysis on the nozzle profile to obtain the directions of the first principal component and the second principal component; (3) Based on the direction of the second principal component, the truncated mean method is used to calibrate the reference object according to the nozzle diameter to obtain the scaling ratio; (4) The volume of the droplet in the image is calculated using the volume integration method based on the direction of the first principal component; (5) Obtain the actual volume based on the droplet volume and scaling ratio in the image.

4. The droplet volume recognition method based on YOLOv11 instance segmentation according to claim 3, characterized in that, The actual volume is: V real =V·k 3 Where V is the calculated droplet volume in the image. The projection value l of N contour points n Sort the elements in set L in ascending order, truncate the middle 50% of the values ​​to remove outliers, and calculate the average of the resulting values; z n and r n Let p be the projection values ​​of the nozzle in the axial direction and the radial direction, respectively, for the nth contour point. n v1 and v2 are the nth elements in the centered contour vector set, where v1 and v2 are the directions of the first and second principal components, respectively; where l n =p n v2.

5. The droplet volume recognition method based on YOLOv11 instance segmentation according to claim 1, characterized in that, The temporal feature LSTM model adds a dimension transformation layer and a result decoding and output layer to the original LSTM model; Dimension transformation layer: Separates volume data from time data. Volume data is directly used as a dimension, and time data is converted from absolute time to time interval Δt by subtraction. Then it is fed into two fully connected layers, and finally concatenated with volume data and fed into the LSTM model. Result Decoding and Output Layer: The data output by the LSTM model is mapped to the target dimension N through a fully connected layer. At this time, the data structure is [b,1,N], where b is the batch size and N is the number of predicted data points output by the model.

6. The droplet volume recognition method based on YOLOv11 instance segmentation according to claim 5, characterized in that, The process by which the time-feature LSTM model obtains the predicted data corresponding to historical data is as follows: (a) Use the droplet region-volume conversion module to obtain droplet volume data at different times, and extract M+1 consecutive data points as a set of historical data; where M is the number of input data for the time feature LSTM model; (b) The data obtained by the droplet region-volume conversion module is interpolated according to time to obtain droplet volume data of 0.1 / N, which is called the interpolation dataset; then, according to the time point of the last data in the historical dataset in step (a), the volume data of the 5 data points after the time point in the interpolation dataset are taken as the prediction data, a total of SM-2 groups, which correspond one-to-one with the SM-2 groups of historical data in step (a).

7. A closed-loop two-dimensional sampling system, characterized in that, include: Nozzle module; XY mobile platform, used for installing liquid receiving devices; The liquid micro-injection module is used to inject liquid into the nozzle module; The air path module is used to generate airflow to break up the suspended droplets at the nozzle module; The control module is used to control the XY moving platform, the liquid micro-injection module and the gas path module. By controlling the growth rate of droplets at the nozzle module and the duration of the pulsed gas flow at the nozzle module, droplets of the required size are generated, thereby realizing the active preparation of droplets. A host computer is used to run the droplet volume recognition method as described in any one of claims 1-6.

8. The closed-loop two-dimensional sampling system according to claim 7, characterized in that, The nozzle module includes a nozzle, a nozzle clamping block, a nozzle housing, a manual lead screw slide, and an adjustment screw. The nozzle clamping block is pressure-connected to the nozzle, and the nozzle housing is connected to the lead screw slide. The lead screw slide is fixed on the frame and is used to adjust the height of the nozzle relative to the XY moving platform. The nozzle clamping block slides up and down in the slide groove of the nozzle housing. The thread on the outside of the nozzle clamping block is engaged with the thread of the adjusting screw. The nozzle clamping block is connected to the nozzle housing through the adjusting screw.

9. The closed-loop two-dimensional sampling system according to claim 7, characterized in that, The gas path module includes a solenoid valve, a gas pump, and a relay. The control module is connected to the relay via a circuit, and the relay is connected to the solenoid valve via a circuit. The gas pump is connected to the solenoid valve via a gas path pipeline, and then to the gas phase inlet of the nozzle. The control module generates a pulse signal to control the relay, which is then converted into a voltage signal to control the opening and closing of the solenoid valve. This generates a pulsed airflow at the nozzle, causing the suspended droplets at the nozzle to break up. In conjunction with the micro-injection module, this produces droplets of the desired size, achieving active droplet preparation.

Images