Tobacco power conveying speed measuring method and device based on ghost image migration deep learning

Abstract

The application discloses a method and device for measuring the speed of tobacco air conveying based on a trailing image migration deep learning, and relates to the fields of image speed measurement, industrial visual detection and tobacco intelligent manufacturing technology.The method comprises the following steps: constructing a special image sample library for tobacco movement trailing; constructing and training a single-frame trailing feature extraction and speed measurement integrated deep learning model; collecting single-frame trailing images in real time and pre-processing; extracting the trailing pixel length by the model, and calculating the speed in combination with a calibration scale coefficient and an exposure time length; and monitoring the pipeline pressure in real time to realize blockage early warning and air volume closed-loop regulation and control.The application realizes non-contact speed measurement through single-frame imaging, does not need to add tracer particles in the industrial speed measurement stage, has high speed measurement precision and strong real-time performance, integrates blockage early warning and air volume regulation and control functions, and meets the online speed measurement and intelligent monitoring requirements of the tobacco industry air conveying.

Description

Technical Field

[0001] This invention relates to the fields of image speed measurement, industrial visual inspection and intelligent tobacco manufacturing technology, and in particular to a method and device for measuring the speed of pneumatic conveying of tobacco shreds based on deep learning of motion blur image transfer. Background Technology

[0002] Domestic cigarette manufacturers generally use pneumatic conveying technology to transport tobacco shreds over long distances. As a loose and flexible material, tobacco shreds form a complex gas-solid two-phase flow with compressed air within the pipeline. In actual industrial operations, the conveying speed of tobacco shreds is concentrated in the high-speed range of 8-20 m / s, and the flow patterns are subdivided into three types: uniform flow, non-uniform flow, and sliding flow. The distribution density, size, and trajectory of the tobacco shred bundles differ significantly under different patterns. Accurate online detection of the tobacco shred conveying speed is the core basis for the airflow balance control of the pneumatic conveying system, directly determining the tobacco breakage rate, moisture loss, and production loss rate. It is a key technological link in improving the quality and efficiency of intelligent tobacco manufacturing.

[0003] Existing tobacco conveying speed measurement technologies suffer from numerous insurmountable drawbacks, failing to meet the high-speed, continuous, and non-contact speed measurement requirements of industrial settings: First, contact-based speed measurement equipment directly interferes with the flow pattern of the tobacco, leading to mechanical wear over long-term use, continuous degradation of measurement accuracy, and easy contamination of the tobacco. Second, traditional optical speed measurement is affected by the loose shape of the tobacco, pipe shadows, and light reflections, resulting in poor image quality and measurement errors as high as 2-3 m / s, making it difficult to meet process accuracy requirements. Third, traditional particle image velocimetry (PIV) technology requires the addition of tracer particles, making it difficult to operate in industrial settings and prone to accidents. Fourth, existing deep learning speed measurement solutions mostly use two consecutive frames of images to compare and extract the pixel displacement of the same tobacco clump to calculate the speed. This method relies on strict inter-frame synchronization, which is prone to matching failures caused by tobacco clump deformation and occlusion. In addition, it requires caching multiple frames of images for time-series processing, resulting in high inference latency and failing to meet the real-time requirements of industrial production lines. Fifth, existing speed measurement devices generally lack a dedicated tobacco image sample library, have poor model generalization ability, and do not integrate industrially practical functions such as material blockage warning and air volume feedback control. They are functionally limited and have weak adaptability.

[0004] For example, Chinese patent CN113086653A discloses a method and device for real-time monitoring of tobacco shreds based on high-speed imaging of dilute phase transport in a curved tube. It uses high-speed imaging technology to acquire images of the tobacco shreds and obtains the tobacco shred velocity distribution after image processing. This solution belongs to the related technical exploration in the field of tobacco shred transport speed measurement, but its core still relies on high-speed cameras to continuously acquire multiple frames of images for time-series analysis. It has not deviated from the technical paradigm of inter-frame matching and also faces problems such as matching failure due to tobacco shred deformation and high processing latency due to buffering multiple frames. Summary of the Invention

[0005] One of the objectives of this invention is to address the problems in the prior art by providing a method for measuring the speed of pneumatic tobacco conveying based on deep learning of motion blur image transfer, so as to achieve high-precision, high-real-time, and high-practicability online speed measurement in tobacco pneumatic conveying scenarios.

[0006] To address the aforementioned technical problems, this invention employs the following technical solution: a method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning for motion blur image transfer. The method and its core principle are as follows: using a global shutter industrial camera, a high-speed tobacco shred bundle is continuously backlit within an extremely short fixed exposure time Δt. The tobacco shred bundle moves at a uniform speed during the exposure period, forming a clear linear motion blur on a single-frame image. The length L of the motion blur pixel in the single-frame image is accurately extracted using a deep learning model. Combined with a pre-calibrated image scale coefficient k (unit: mm / pixel), the speed of the tobacco shred bundle is directly calculated using the formula v=(k×L) / Δt. The entire speed measurement can be completed using only a single-frame image, with no inter-frame synchronization errors, significantly improving inference efficiency.

[0007] Specifically, the following steps are included:

[0008] S1. Construct a dedicated image sample library for tobacco motion blur:

[0009] In a transparent conveying pipeline experimental section simulating actual industrial conditions, a standardized motion trailing image platform was built. Using a global shutter camera and a constant backlight source, single-frame motion trailing image sequences were acquired for four flow patterns: uniform flow, clustered flow, scattered flow, and transitional flow, within the 8-20 m / s velocity range of wind-driven silk conveying. To ensure annotation accuracy, no more than 10% of the samples were selected for auxiliary ground truth annotation using PIV technology. A tracer particle was added to the conveying medium, and sub-pixel-level accurate velocity field was obtained through PIV analysis. The actual trailing length and actual velocity of the single-frame trailing image were then calibrated to form high-precision auxiliary labels. The remaining samples were semi-automatically annotated using the auxiliary labels. The samples were generalized and expanded through methods such as rotation, translation, brightness adjustment, and Gaussian noise addition. Finally, a dedicated sample library with no less than 50,000 sets of total samples and no less than 5,000 sets of samples for each flow pattern was constructed. The samples were divided into training dataset, validation dataset, and test dataset in a 7:2:1 ratio to provide highly adaptable data support for model training.

[0010] S2. Construct and train a deep learning model that integrates single-frame motion blur feature extraction and speed measurement:

[0011] Abandoning traditional temporal recurrent neural networks, a lightweight deep learning model adapted for single-frame motion blur velocity measurement is built. ResNet-18 is used as the pre-trained backbone feature extraction network, transferring its general image feature extraction capabilities to reduce the computational cost and training time of the model. A two-layer 3×3 convolutional neural network module is connected after the backbone network to enhance the local detail features such as the edge, contour, and length of the tobacco blob motion blur and remove background interference information. The tail module is connected to the motion blur segmentation and length accurate calculation module, which directly outputs the tobacco blob identifier, motion blur pixel length and corresponding motion speed, realizing the integrated output of target recognition, motion blur extraction and speed calculation.

[0012] The core objective of model training is to minimize the error between the model's prediction speed and the actual speed of PIV-assisted annotation. The Adam optimizer is used, with an initial learning rate of 1e-4, a batch size of 32, and 200 training epochs. The learning rate is decayed at epochs 100 and 150. The loss function is a combination of mean absolute error (MAE) and mean squared error (MSE) to balance the accuracy of tobacco blob recognition and positioning with the accuracy of trail length measurement, thereby improving the overall robustness and generalization ability of the model. After training, the model is quantized into RKNN format to meet the high-speed inference requirements of edge computing units.

[0013] S3. Real-time single-frame motion blur image acquisition and preprocessing:

[0014] In the transparent observation section of the actual tobacco conveying pipeline, a global shutter industrial camera and a constant backlight source are symmetrically arranged to completely eliminate shadow and reflection interference and ensure clear trailing image imaging. The camera is set to an extremely short fixed-duration continuous exposure mode, with the exposure time adjustable from 0.2 to 1 ms to adapt to the trailing image formation requirements of different tobacco flow rates. This ensures that the linear trailing image is of moderate length and has a clear outline, while avoiding trailing images that are too long and overlapping or too short and difficult to identify. The light source and camera are triggered simultaneously. After acquiring a single frame of trailing image, noise reduction, contrast enhancement, and scale normalization preprocessing are performed to remove interference such as impurities on the inner wall of the pipeline and light fluctuations, highlighting the trailing characteristics of the tobacco clump.

[0015] S4. Single-frame motion blur speed measurement and result output:

[0016] The preprocessed single-frame motion blur image is directly input into the integrated deep learning model deployed on the edge computing unit. The model quickly completes tobacco blob recognition, motion blur contour segmentation and accurate pixel length extraction through forward inference, automatically substitutes it into the speed calculation formula, and directly outputs the unique identifier and real-time motion speed of the corresponding tobacco blob. The entire process does not require caching multiple frames of images or temporal feature matching. Speed ​​measurement is completed by single-frame inference, which greatly reduces processing latency.

[0017] S5. Blockage warning and feedback closed-loop control:

[0018] A pressure sensing unit is installed downstream of the transparent observation section of the conveying pipeline to monitor changes in air pressure in the pipeline in real time. The preset material blockage warning threshold is 5 kPa. When the pipeline pressure is detected to continuously exceed the threshold, an audible and visual material blockage warning is immediately triggered, and the upstream tobacco feeding equipment is simultaneously linked to suspend material supply. Combined with the real-time output of tobacco conveying speed data from the model, a closed-loop control signal is generated to automatically adjust the air volume of the tobacco feeding fan, stabilizing the tobacco speed within the process setting range, thus realizing the integration of speed measurement and operating condition control.

[0019] In addition, this invention also provides a tobacco pneumatic conveying speed measurement device based on motion blur image transfer deep learning to implement the above-mentioned method. The device adopts an integrated industrial housing design, is highly integrated, easy to install, and does not require modification of existing pneumatic conveying pipes. It mainly includes three parts: an image acquisition unit, a pressure sensing unit, and a data processing and execution unit.

[0020] 1. Image acquisition unit:

[0021] Composed of a constant backlight source, a global shutter industrial camera, and a transparent observation square tube section, the transparent square tube section has a square cross-section with a side length of 107~140mm, which can be directly and seamlessly connected to the Φ120~150mm industrial tobacco feeding tube to ensure stable tobacco flow and unobstructed imaging. The light source and camera are symmetrically arranged on both sides of the transparent square tube section, adopting a bottom backlight layout to completely solve the interference of shadows and reflections on the motion blur imaging. The industrial camera uses a CCD or CMOS global shutter camera with a photosensitive area size of 2 / 3 inch, equipped with a 6~8mm fixed focus lens, and an adjustable working distance of 400mm or 600mm. It has a dedicated continuous exposure mode adapted for motion blur imaging, with an exposure time that is precisely adjustable from 0.2~1ms, avoiding the rolling shutter effect caused by line-by-line exposure and ensuring that the motion blur has a standard linear shape.

[0022] 2. Pressure sensing unit:

[0023] Employing an industrial-grade pressure sensor installed on the inner wall of the downstream pipe of a transparent square tube section, it collects real-time air pressure data within the pipe with a sampling frequency of no less than 100Hz. It accurately captures pressure surge signals in the early stages of material blockage, with a warning threshold set at 5kPa. Upon triggering, it quickly outputs a warning signal, enabling automatic start-up without suction wire and immediate alarm for material blockage, thus meeting the unmanned operation requirements of the tobacco industry.

[0024] 3. Data Processing and Execution Unit:

[0025] Employing an embedded edge computing device based on the RK3576 chip, it features a built-in NPU computing unit, supports high-speed inference of the quantized RKNN model, and is electrically connected to the image acquisition unit and pressure sensing unit. Internally, it integrates a model storage module, an image preprocessing module, a motion blur velocity measurement module, and a control output module. The model storage module carries a trained integrated deep learning model, the image preprocessing module performs noise reduction and enhancement of single-frame images, the motion blur velocity measurement module performs motion blur extraction and speed calculation, and the control output module integrates speed and pressure data to complete the issuance of blockage warning and fan airflow closed-loop control commands. It features short response time, stable operation, and is suitable for long-term continuous operation in industrial sites.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. Innovative speed measurement principle with no synchronization error: It abandons the traditional two-frame image time-matching speed measurement and adopts a single-frame extremely short continuous exposure to form motion blur speed measurement, which completely eliminates the problems of inter-frame synchronization error and tobacco shred deformation matching failure. The speed measurement stability is greatly improved, with a speed measurement error of ≤±0.4m / s, which is superior to traditional PIV and optical speed measurement technology.

[0028] 2. Significantly improved real-time performance: Speed ​​measurement can be completed with a single frame image, eliminating the need for multi-frame buffering and timing processing. Edge inference latency is <200ms, and the overall device response time is <1s, which is far superior to traditional two-frame speed measurement and PIV speed measurement solutions, fully meeting the real-time requirements of high-speed industrial production lines.

[0029] 3. Dedicated sample library and strong model generalization: For the first time, a dedicated image sample library for tobacco motion trails has been built, covering all flow speeds and all flow patterns. Combined with transfer learning to optimize the model, the accuracy of tobacco clump recognition is ≥96%, which can adapt to the differentiated needs of different cigarette factories and different tobacco feeding conditions, and solve the pain points of traditional models that have no dedicated samples and poor adaptability.

[0030] 4. Highly applicable to industrial use and free from media contamination: PIV is used only as an auxiliary annotation method during the training phase. No tracer particles need to be added for actual industrial speed measurement, avoiding tobacco contamination. The equipment can be operated as soon as it is powered on, without the need for professional personnel. Installation and calibration are convenient and it can be directly deployed on a large scale.

[0031] 5. Integrated functions to meet industrial needs: It integrates single-frame high-precision speed measurement, automatic start-up without suction, audible and visual warning of material blockage, and closed-loop control of fan air volume, breaking through the limitations of traditional speed measurement devices that only measure speed. It becomes an intelligent monitoring node of the wind-powered tobacco feeding system, helping to upgrade tobacco intelligent manufacturing.

[0032] 6. Lightweight structure and controllable cost: The device adopts mature industrial-grade hardware, is integrated as a whole, has a small size, is flexible in installation, does not require large-scale modification of existing wire feeding pipes, has low maintenance costs, and has the potential to be promoted to other fields of pneumatic conveying of loose materials. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of the device in an embodiment of the present invention;

[0034] Figure 2 This is a schematic diagram of the single-frame motion blur imaging and velocity measurement logic in this invention. Detailed Implementation

[0035] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0036] This invention provides a method and apparatus for measuring the pneumatic conveying speed of tobacco shreds based on deep learning for image migration. In the industrial speed measurement stage, this method does not require the addition of any tracer particles; instead, it uses PIV technology to assist in ground truth labeling of a subset of samples during sample library construction, thus avoiding tobacco shred contamination.

[0037] Figure 1 This is a schematic diagram of the overall structure and principle of the motion blur velocity measuring device of the present invention. The device adopts an integrated industrial box design, which is highly integrated and easy to install without modifying the existing pneumatic wire feeding pipeline. It mainly consists of three parts: an image acquisition unit, a pressure sensing unit, and a data processing and execution unit.

[0038] The image acquisition unit consists of a constant backlight source, a global shutter industrial camera, and a transparent observation square tube section. The transparent observation square tube section has a square cross-section, with a side length selectable within the range of 107~140mm. In this embodiment, an acrylic square tube with a side length of 140mm is specifically selected, which can be directly and seamlessly connected to the Φ120~150mm tobacco feeding tube in the industrial environment, ensuring stable tobacco flow and unobstructed imaging. The light source and camera are symmetrically arranged on both sides of the transparent square tube section, using a bottom backlight layout to completely eliminate the interference of shadows and reflections on motion blur imaging, ensuring clear and distortion-free motion blur imaging. The industrial camera is a 2 / 3-inch global shutter CMOS camera equipped with a fixed-focus lens with a focal length of 6~8mm. In this embodiment, an 8mm fixed-focus lens is specifically selected, with a working distance set at 400mm and pre-focused on the center of the pipe. The camera-exclusive continuous exposure mode is adapted for motion blur imaging. The exposure time is precisely adjustable within the range of 0.2~1ms. It adopts a global shutter method to avoid the rolling shutter effect caused by line-by-line exposure and ensures that the motion blur is in a standard linear shape.

[0039] During device deployment, image scale coefficient calibration is performed using a standard calibration board. In this embodiment, the final calibration coefficient k = 0.1 mm / pixel, which will be used for subsequent speed calculations.

[0040] The pressure sensing unit uses an industrial-grade pressure sensor, installed on the inner wall of the pipe approximately 20cm downstream of the transparent square tube section. The warning threshold is preset to 5kPa. The pressure sensor can accurately capture pressure surges in the early stages of blockage and quickly output a warning signal upon triggering.

[0041] The data processing and execution unit utilizes an embedded edge computing device based on the RK3576 chip, with a built-in NPU computing unit. Internally, it integrates a model storage module, an image preprocessing module, a trailing image speed measurement module, and a control output module. The model storage module carries a trained integrated deep learning model, which is quantized to RKNN format with INT8 precision. Leveraging the NPU's computing power, it achieves high-speed forward inference with a single-frame inference latency of less than 200ms. The image preprocessing module performs denoising and contrast enhancement on single-frame images. The trailing image speed measurement module identifies tobacco clumps, extracts trailing image pixel length, and calculates speed. The control output module integrates speed and pressure data to determine material blockage warnings and issue closed-loop control commands for fan airflow. The device automatically enters standby mode upon power-on and automatically activates imaging speed measurement when tobacco is detected, enabling automatic start-up when no tobacco is being fed, thus meeting the unmanned operation requirements of the tobacco industry.

[0042] In constructing a dedicated image sample library for tobacco motion trails, a standardized trailing image imaging platform was built in a transparent conveying pipeline experimental section simulating actual industrial conditions, employing a global shutter camera and a constant backlight source. Single-frame motion trailing image sequences were acquired for wind-driven tobacco conveying speeds of 8–20 m / s and for four flow patterns: uniform flow, clustered flow, scattered flow, and transitional flow. To ensure annotation accuracy, trailing image samples representing no more than 10% of the total sample size were selected for auxiliary ground truth annotation using PIV technology. During the acquisition period corresponding to these samples, tracer particles were added to the conveying medium, and sub-pixel-level velocity field data was obtained through PIV analysis. Based on this, the true trailing length and actual movement speed of the tobacco clump in each single-frame trailing image were calibrated, forming high-precision auxiliary labels. The remaining samples were semi-automatically annotated using these auxiliary labels, eliminating the need to add tracer particles frame by frame. To further enhance sample diversity and model generalization ability, the labeled samples were generalized and expanded through rotation, translation, brightness adjustment, and Gaussian noise addition. The rotation angle range was 0~360°, the translation range was ±10 pixels, and the brightness adjustment range was ±20%. After generalization and expansion, a dedicated sample library with a total sample size of no less than 50,000 groups and no less than 5,000 groups of samples for each flow type was finally constructed. In this embodiment, 8,000 original trailing images of four flow types with flow velocities of 8~20 m / s under actual cigarette feeding conditions were collected. 700 samples were selected and auxiliary ground truth annotation was performed using PIV technology. After generalization and expansion, the total sample library reached 52,000 groups, including 13,000 groups of uniformly distributed flow samples, 14,000 groups of clustered flow samples, 12,000 groups of scattered flow samples, and 13,000 groups of transitional flow samples, meeting the requirement of no less than 5,000 groups of samples for each flow type. The sample library is divided into training dataset, validation dataset and test dataset in a ratio of 7:2:1, corresponding to 36,400 training sets, 10,400 validation sets and 5,200 test sets, to provide data support for model training.

[0043] In the construction and training of the integrated deep learning model, the traditional temporal recurrent neural network is abandoned, and a lightweight deep learning model adapted to single-frame motion blur velocity measurement is built. The model uses ResNet-18 as the pre-trained backbone feature extraction network, transferring its general image feature extraction capabilities to reduce the computational cost and training time of the model. After the backbone network, a two-layer 3×3 convolutional neural network module is connected to enhance the local detail features such as the edge, contour, and length of the tobacco blob motion blur, and remove background interference information. The tail module is connected to the motion blur segmentation and accurate length calculation module, which directly outputs the tobacco blob identifier, motion blur pixel length and corresponding motion speed, realizing the integrated output of target recognition, motion blur extraction and speed calculation. The model training focuses on minimizing the error between the model's predicted speed and the actual speed labeled by PIV (Picture-in-the-Video). The Adam optimizer is used with an initial learning rate of 1e-4, a batch size of 32, and 200 training epochs. Learning rate decay is implemented at epochs 100 and 150. The loss function combines mean absolute error and mean squared error to balance the accuracy of tobacco blob recognition and trail length calculation, improving the overall robustness and generalization ability of the model. The validation set loss is monitored during training, and training stops when the validation set loss does not decrease for 10 consecutive epochs. After training, the model is quantized into an INT8 precision RKNN format to meet the high-speed inference requirements of edge computing units. In this embodiment, the trained model achieves a tobacco blob recognition accuracy of 96.2% on the test set, with a trail length calculation error of no more than 2 pixels, a speed measurement error of no more than ±0.4 m / s, a single-frame inference latency of 160 ms, and an overall device response time of less than 1 second, meeting the requirements for speed measurement accuracy and real-time performance in industrial wind-powered tobacco feeding scenarios.

[0044] In real-time single-frame trailing image acquisition and preprocessing, a global shutter industrial camera and a constant backlight source are symmetrically arranged in the transparent observation section of the actual tobacco conveying pipeline. The light source adopts a bottom-backlit layout, and the light source and camera are triggered synchronously. The camera is set to a fixed-duration continuous exposure mode, with the exposure time set within the range of 0.2~1ms according to the tobacco flow rate under the current operating conditions. This ensures the formation of a linear trailing image of moderate length and clear outline, while avoiding trailing images that are too long and overlapping or too short and difficult to identify. After acquiring the single-frame trailing image, the image preprocessing module performs noise reduction, contrast enhancement, and scale normalization preprocessing on the image to remove interference such as impurities on the inner wall of the pipeline and light fluctuations, highlighting the trailing characteristics of the tobacco clump.

[0045] Figure 2This is a schematic diagram of the single-frame motion blur imaging and velocity measurement logic in this invention. The tobacco blob moves at a constant speed along the pipe axis within an extremely short, fixed exposure time Δt, forming a linear motion blur on a single-frame image due to continuous exposure. The pre-processed single-frame motion blur image is directly input into an integrated deep learning model deployed on the edge computing unit. The model quickly completes tobacco blob identification and motion blur contour segmentation through forward inference, accurately extracting the pixel length L of the motion blur. Combining the pre-calibrated image scale coefficient k and the fixed exposure time Δt, the tobacco blob's velocity is directly calculated according to the formula v=(k×L) / Δt, outputting the tobacco blob identifier and real-time velocity. The entire process requires no caching of multiple frames or inter-frame temporal feature matching; velocity measurement can be completed with single-frame inference, with an inference latency of less than 200ms.

[0046] In the closed-loop control of material blockage early warning and airflow feedback, an industrial-grade pressure sensing unit is installed on the inner wall of the downstream pipeline of the transparent observation section to monitor changes in air pressure within the pipeline in real time. The preset material blockage early warning threshold is 5 kPa. When the pressure sensing unit detects that the pipeline pressure continuously exceeds this threshold, the control output module immediately triggers an audible and visual material blockage early warning, and simultaneously links the upstream tobacco feeding equipment to suspend material supply, realizing automatic start-up without tobacco suction and immediate alarm for material blockage, with a response time of no more than 1 second. At the same time, the control output module combines the real-time tobacco conveying speed data output by the model to generate a closed-loop control signal, automatically adjusting the airflow of the tobacco feeding fan to stabilize the tobacco speed within the process setting range, achieving the integration of speed measurement and operating condition control.

[0047] To verify the comprehensive advantages of the present invention over existing technologies, the device of the present invention was compared with traditional PIV speed measurement devices, two-frame image deep learning speed measurement devices, and traditional optical speed measurement devices under the same industrial conditions. In terms of speed measurement accuracy, the error of the present invention does not exceed ±0.4 m / s, while the error of traditional PIV is ±0.8~1.0 m / s, the error of two-frame speed measurement is ±1.0~1.5 m / s, and the error of optical speed measurement is ±1~3 m / s. Regarding real-time performance, the inference delay of the present invention is 160 ms, and the overall response time is less than 1 s, while the processing time of traditional PIV is 2~8 s, the response time of two-frame speed measurement is 1~2 s, and the response time of optical speed measurement is 1~3 s. In terms of industrial operability, the present invention requires no tracer particles, has no frame synchronization requirements, and is ready to use upon power-up, while other devices suffer from complex operation, tobacco contamination, or poor stability. Functionally, the present invention integrates speed measurement, early warning, and control functions, while other devices only have a single speed measurement function, fully verifying the comprehensive advantages of the present invention.

[0048] To facilitate understanding by those skilled in the art of the improvements of this invention over the prior art, some of the accompanying drawings and descriptions have been simplified. The above embodiments are preferred implementations of this invention. In addition, this invention can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this invention.

Claims

1. A method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning for image migration, characterized in that, Speed ​​measurement is achieved by using a single-frame extremely short continuous exposure to create motion blur, including the following steps: S1. Construct a dedicated image sample library for tobacco motion trails: In a transparent conveying pipeline section simulating actual working conditions, single-frame motion trail images with multiple flow patterns and multiple flow velocity ranges are acquired using a global shutter camera and constant backlight; some samples are selected to obtain the true velocity values ​​using PIV technology as auxiliary annotations, and the remaining samples are semi-automatically annotated using auxiliary labels. After generalization and expansion, a sample library is formed and divided into training, validation, and test datasets according to proportions. S2. Construct and train a deep learning model that integrates single-frame motion blur feature extraction and speed measurement: The model uses a pre-trained convolutional neural network as the backbone feature extraction network, followed by a local feature enhancement module and a motion blur segmentation length measurement module. Transfer learning is used for training, with the goal of minimizing the predicted speed and PIV-assisted true value error, and the model parameters are optimized. S3. Real-time single-frame motion image acquisition and preprocessing: In the transparent observation section of the actual transport pipeline, a global shutter camera is used to acquire single-frame motion images with a fixed duration of continuous exposure, and then preprocessing is performed for noise reduction and contrast enhancement. S4. Single-frame trailing image speed measurement output: Input the pre-processed single-frame image into the trained model, extract the pixel length of the tobacco blob trailing image, combine the calibration scale coefficient and exposure time to calculate the speed, and directly output the tobacco blob identifier and real-time speed. S5. Blockage Warning and Feedback Control: Real-time monitoring of pipeline pressure; triggering a blockage warning when the pressure exceeds a preset threshold; and adjusting the airflow of the wire feeding fan in real-time to achieve closed-loop control.

2. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 1, characterized in that: The various flow patterns mentioned in step S1 include uniform flow, clustered flow, scattered flow and transitional flow. The sample generalization and expansion methods include rotation, translation, brightness adjustment and Gaussian noise addition.

3. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 1, characterized in that: In step S1, PIV-assisted annotation obtains sub-pixel-level velocity field data by adding a small number of tracer particles, and calibrates the true length of the trail and the actual velocity accordingly, which serves as the ground truth label for model training.

4. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of motion blur image transfer according to claim 1, characterized in that: The pre-trained convolutional neural network mentioned in step S2 is ResNet-18, followed by a two-layer 3×3 convolutional neural network module. The model training uses the Adam optimizer with an initial learning rate of 1e-4, a batch size of 32, 200 training epochs, and a loss function that is a combination of mean absolute error and mean squared error.

5. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 1, characterized in that: The fixed duration mentioned in step S3 is 0.2~1ms. The camera and the constant backlight source are symmetrically arranged, and a bottom backlight layout is adopted to eliminate shadow and reflection interference.

6. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 1, characterized in that: In step S4, the speed calculation formula is v=(k×L) / Δt, where k is the image scale calibration coefficient, L is the length of the trailing pixel, and Δt is the fixed exposure time of the camera.

7. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 1, characterized in that: In step S5, the sampling frequency of the pressure sensing unit is not less than 100Hz, the material blockage warning threshold is set to 5kPa, the material blockage warning response time is not more than 1s, and the air volume control is a real-time closed-loop adjustment.

8. A device for measuring the pneumatic conveying speed of tobacco shreds based on deep learning for image migration, used to implement the method described in any one of claims 1-7, characterized in that: It includes an image acquisition unit, a pressure sensing unit, and a data processing and execution unit; The image acquisition unit includes a constant backlight source, a global shutter industrial camera, and a transparent observation square tube section, which are symmetrically arranged on both sides of the pipe and are adapted for single-frame trailing continuous exposure imaging. The pressure sensing unit is an industrial-grade pressure sensor, installed downstream of the transparent square tube, to monitor the pipeline pressure in real time. The data processing and execution unit is an embedded edge computing device that integrates model storage, image preprocessing, motion blur speed measurement, and control output modules to achieve integrated single-frame speed measurement, early warning, and control.

9. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of motion blur image transfer according to claim 8, characterized in that: The transparent observation square tube section has a square cross-section with a side length of 107~140mm, and is seamlessly connected to a Φ120~150mm wire feeding round tube; the industrial camera has a photosensitive area of ​​2 / 3 inch, is equipped with a 6~8mm fixed-focus lens, and has an adjustable working distance of 400mm or 600mm.

10. The method for measuring the pneumatic conveying speed of tobacco shreds based on deep learning of image migration according to claim 8, characterized in that: The embedded edge computing device is equipped with a quantized integrated deep learning model, with a single-frame inference latency of less than 200ms and an overall device response time of less than 1s.

Images