Screen plate cleaning device and method for tobacco redrying machine

Abstract

The invention provides a device and a method for cleaning a screen plate of a tobacco redrying machine, and belongs to the technical field of redrying machine cleaning equipment. The control cabinet is arranged on the edge of the top of the base; the PLC is arranged in the control cabinet; the communication module is arranged in the control cabinet and is connected with the PLC; the industrial camera is arranged at the top end of the base, the shooting direction of the industrial camera is upward, and the industrial camera is connected with the PLC; the lifting module is arranged at the top end of the base and is connected with the PLC; the cleaning module is mounted on the lifting module and connected with the PLC; the sundry collecting module is arranged at the top end of the base, located under the lifting module and the sweeping module and connected with the PLC; the stroke induction sensor is arranged on the lifting module, the induction direction of the stroke induction sensor faces the sweeping module, and the stroke induction sensor is connected with the PLC; the power output end of the driving module is connected with the lifting module and the sweeping module, and the control end of the driving module is connected with the PLC. The cleaning device has the advantages that the cleaning efficiency and quality of the screen plate of the tobacco redrying machine are greatly improved, and the cleaning cost is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of re-drying machine cleaning equipment, and particularly to a screen plate cleaning device and method for a tobacco leaf re-drying machine. Background Art

[0002] The tobacco leaf processing technological process generally includes key processes such as harvesting and airing, grading and sorting, biochemical fermentation, re-drying, cutting process, quality inspection, and packaging and storage. Among them, the re-drying process is the core technological link, and its technical key lies in implementing secondary drying treatment on the fermented tobacco leaves, accurately regulating the moisture content of the tobacco leaves to the range of 8 - 12%, so as to achieve the setting of the tobacco leaf fiber tissue and the directional conversion of aroma components.

[0003] For tobacco leaf re-drying, a tobacco leaf re-drying machine is required. In order to improve the penetration efficiency of hot air and enable the hot air to evenly pass through the tobacco leaves, thereby improving the drying effect, the tobacco leaf re-drying machine uses a screen plate (a perforated stainless steel screen plate) to convey the tobacco leaves. However, under the process conditions of high temperature and high humidity, organic components such as reducing sugars (glucose, fructose, etc.), nitrogen-containing compounds (nicotine, protein), and pectin contained in the tobacco leaves will generate high-viscosity tar by-products through the Maillard reaction and caramelization, blocking the holes of the screen plate. Through experimental measurement, after the tobacco leaf re-drying machine operates continuously for 72 hours, the tar deposition amount on the surface of the screen plate can reach 1.2 - 1.8 g / cm 2 , resulting in a 45 - 60% reduction in the effective through-hole area and a decrease in the heat conduction efficiency to below 78%, seriously affecting the quality of tobacco leaves (the aroma permeability decreases by 2 - 3 grades, and the amount of miscellaneous gas increases by 1.5 times). Therefore, it is necessary to regularly clean the screen plate of the tobacco leaf re-drying machine.

[0004] Regarding the cleaning of the screen plate of the tobacco leaf re-drying machine, the traditional three-step cleaning method of manual disassembly, soaking, and brushing has the following disadvantages: 1. Manual cleaning has low efficiency, high labor costs, and there are problems of incomplete cleaning and poor consistency; 2. The tobacco leaf re-drying machine needs to be shut down for operation. Due to the low efficiency of manual cleaning, it will directly affect the production continuity.

[0005] Therefore, how to provide a screen plate cleaning device and method for a tobacco leaf re-drying machine to improve the cleaning efficiency and quality of the screen plate of the tobacco leaf re-drying machine and reduce the cleaning cost has become an urgent technical problem to be solved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a screen plate cleaning device and method for a tobacco leaf re-drying machine to improve the cleaning efficiency and quality of the screen plate of the tobacco leaf re-drying machine and reduce the cleaning cost.

[0007] In a first aspect, the present invention provides a screen plate cleaning device for a tobacco leaf re-drying machine, including:

[0008] A base with three sets of rollers provided at the bottom;

[0009] A control cabinet provided at the edge of the top of the base;

[0010] A PLC provided inside the control cabinet;

[0011] A communication module provided inside the control cabinet and connected to the PLC;

[0012] At least one industrial camera provided at the top of the base, with the shooting direction facing upwards and connected to the PLC;

[0013] A lifting module provided at the top of the base and connected to the PLC;

[0014] A cleaning module installed on the lifting module and connected to the PLC;

[0015] A sundry collection module provided at the top of the base, directly below the lifting module and the cleaning module, and connected to the PLC;

[0016] Two travel induction sensors provided on the lifting module, with the induction direction facing the cleaning module and connected to the PLC;

[0017] A driving module, with the power output end connected to the lifting module and the cleaning module, and the control end connected to the PLC.

[0018] Further, the rollers are universal wheels.

[0019] Further, the communication module is at least one of a 2G communication module, 3G communication module, 4G communication module, 5G communication module, NB-IOT communication module, LORA communication module, WIFI communication module, Bluetooth communication module, ZigBee communication module or a wired communication module.

[0020] Further, the lifting module includes:

[0021] Two lifting columns vertically provided at the top of the base, spaced apart from each other by a certain distance;

[0022] A cross beam horizontally provided at the top of the two lifting columns, and a horizontal sliding groove is provided on the side; the cleaning module is installed on the cross beam through the sliding groove;

[0023] The cleaning module includes:

[0024] A sliding seat slidably connected to the sliding groove;

[0025] A moving base, with one end of the side connected to the sliding seat;

[0026] A nylon cleaning brush head is provided at the top of the moving base, and a circular suction nozzle is provided in the middle;

[0027] A first motor is provided at the bottom of the moving base, with the power output end facing upward, and the power output end passes through the moving base and is connected to the nylon cleaning brush head, and the control end is connected to the PLC;

[0028] A centrifugal fan, the suction end of which is communicated with the circular suction nozzle, and the control end is connected to the PLC.

[0029] Further, the debris collection module includes:

[0030] A belt conveyor is provided at the top of the base, located directly below the lifting module and the cleaning module, and is connected to the PLC;

[0031] A debris collection enclosure is provided around the belt conveyor, and a notch is provided at the transmission end of the belt conveyor.

[0032] Further, the two travel induction sensors are symmetrically provided inside the lifting columns.

[0033] Further, the drive module includes:

[0034] A second motor, the power output end of which is connected to the lifting module, and the control end is connected to the PLC;

[0035] A third motor, the power output end of which is connected to the cleaning module, and the control end is connected to the PLC.

[0036] In a second aspect, the present invention provides a method for cleaning the wire mesh plate of a tobacco leaf re-drying machine, including the following steps:

[0037] Step S1, create a tar detection model based on a multi-scale feature extraction module, a candidate region localization module, an attention enhancement module, a dynamic ROI alignment module, and an output module, and set the loss function of the tar detection model;

[0038] Step S2, obtain a large number of historical wire mesh plate images, preprocess and annotate each of the historical wire mesh plate images to construct a data set, divide the data set into a training set, a validation set, and a test set, and train, validate, and test the tar detection model through the training set, the validation set, and the test set in sequence, and deploy the tar detection model that passes the test to the server;

[0039] Step S3: Move the cleaning device to the lower part of the wire mesh plate of the tobacco leaf re-dryer through the rollers, and the PLC obtains the cleaning instruction carrying the cleaning parameters; the cleaning parameters at least include the working height of the cross beam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity, and the cleaning path;

[0040] Step S4: Based on the cleaning instruction, the PLC drives the lifting column to rise through the second motor, and then drives the cross beam to rise in linkage until it rises to the working height of the cross beam. Then, the PLC drives the cleaning module to translate in the sliding groove through the third motor until it moves to the starting end of the cleaning path, and calibrates the position of the cleaning module through the travel induction sensor to complete the initialization;

[0041] Step S5: The PLC controls the first motor to work based on the rotation speed of the brush head to rotate the nylon cleaning brush head to perform a cleaning operation on the wire mesh plate. During the cleaning process, the PLC drives the third motor to work based on the translation speed of the cleaning module, so that the nylon cleaning brush head displaces based on the translation speed of the cleaning module and the cleaning path; during the cleaning process, start the centrifugal fan and the belt conveyor, so that the centrifugal fan sucks the dust on the wire mesh plate based on the negative pressure intensity, and the cleaned oil dirt falls on the belt conveyor and is transported out through the belt conveyor until the cleaning of the cleaning path is completed;

[0042] Step S6: The PLC collects the real-time wire mesh plate image of the wire mesh plate through the industrial camera, and uploads the real-time wire mesh plate image to the server through the communication module; after preprocessing the received real-time wire mesh plate image, the server inputs it into the deployed tar detection model, obtains the tar detection report and sends it to the PLC;

[0043] Step S7: The PLC controls the nylon cleaning brush head to displace based on the tar position carried in the tar detection report, and dynamically adjusts the rotation speed of the brush head and the negative pressure intensity to perform secondary cleaning on the wire mesh plate at the tar position. After the secondary cleaning is completed, collect the real-time wire mesh plate image through the industrial camera for inspection.

[0044] Further, in the step S1, the multi-scale feature extraction module is constructed based on the ResNet50 network and the feature pyramid network; the ResNet50 network is used to collect the underlying texture features, edge features, and high-level semantic features from the input wire mesh plate image, and the feature pyramid network is used to fuse the underlying texture features, edge features, and high-level semantic features to obtain multi-scale features;

[0045] The candidate region localization module is constructed based on deformable convolutions of size 3×3, a classification branch, and a regression branch; the deformable convolutions are used to adaptively match meshes with irregular shapes and generate candidate regions of the meshes based on multi-scale features; the classification branch is used to determine whether there are meshes to screen the candidate regions; the regression branch is used to optimize the coordinates of the candidate regions;

[0046] The attention enhancement module is constructed based on a spatial attention unit, a channel attention unit, and a fusion unit; the spatial attention unit is used to extract spatial features from the candidate regions; the channel attention unit is used to extract channel features from the candidate regions; the fusion unit is used to fuse the spatial features and the channel features and suppress background noise to obtain attention features;

[0047] The dynamic ROI alignment module is used to unify the sizes of the candidate regions through bilinear interpolation and dynamically adjust the sampling grid density of the candidate regions;

[0048] The output module is constructed based on a fully connected layer and is used to calculate the probability of tar existing in each candidate region according to the attention features, and output a tar detection report carrying the tar position based on the probability and the candidate regions;

[0049] The formula of the loss function is:

[0050] L_total = L_rpn + λ1L_att + L_cls;

[0051] Among them, L_total represents the loss value of the loss function; L_rpn represents the candidate region localization loss; λ1 represents the balance coefficient; L_att represents the attention focusing loss; L_cls represents the tar classification loss.

[0052] Further, the specific content of step S2 is as follows:

[0053] Obtain a large number of historical mesh images containing different positions, different angles, different lighting conditions, and different degrees of dirt. Perform preprocessing on each of the historical mesh images, including at least image cropping, image scaling, image rotation, image flipping, noise reduction, and color jitter. Mark the positions of tar existing in each of the preprocessed historical mesh images, and construct a dataset based on the marked historical mesh images;

[0054] Divide the dataset into a training set, a validation set, and a test set according to a ratio of 7:2:1. Train the tar detection model through the training set. During the training process, continuously optimize the hyperparameters of the tar detection model, including at least the learning rate, the random dropout rate, the batch size, and the optimizer, until the loss value of the loss function is less than a preset loss threshold;

[0055] Calculate the detection accuracy through the validation set to verify the trained tar detection model. If the verification fails, expand the training set and continue training; if the verification passes, then:

[0056] Calculate the confidence level through the test set to test the tar detection model that has passed the verification. If the test fails, expand the training set and continue training; if the test passes, end the training;

[0057] Deploy the tar detection model that has passed the test to the server;

[0058] The step S7 further includes:

[0059] If the secondary cleaning does not meet the standard, perform the cleaning operation again until the cleaning meets the standard, and record the cleaning log including at least the cleaning instruction, cleaning time, number of repeated cleaning times, and repeated cleaning position. Dynamically adjust the cleaning parameters based on the cleaning log, and encrypt and back up the cleaning log.

[0060] The advantages of the present invention are as follows:

[0061] 1. By setting a base, a PLC, a communication module, an industrial camera, a lifting module, a cleaning module, a travel induction sensor, and a driving module, the PLC is respectively connected to the communication module, the industrial camera, the lifting module, the cleaning module, the travel induction sensor, and the driving module. The cleaning module is installed on the base through the lifting module. The travel induction sensor is used to sense the position of the cleaning module, and the driving module is used to drive the lifting module to lift and drive the cleaning module to translate. And the bottom end of the base is provided with rollers; when it is necessary to clean the mesh plate, only need to move the cleaning device to the cleaning position of the tobacco leaf re-drying machine through the rollers, then lift the cleaning module through the lifting module to contact the mesh plate, and rotate the nylon cleaning brush head of the cleaning module to perform the cleaning operation. Adjust the cleaning position by translating the cleaning module. Subsequently, the industrial camera can be used to check the cleaning situation. If the cleaning does not meet the standard, repeated cleaning can be carried out, that is, the PLC controls the lifting module, the cleaning module, and the driving module to perform automatic cleaning operations, without the need for manual disassembly, soaking, and brushing as in the traditional way, thus greatly improving the cleaning efficiency and quality of the mesh plate of the tobacco leaf re-drying machine and greatly reducing the cleaning cost.

[0062] 2. By setting the rollers as universal wheels, the convenience of transplanting the cleaning device is greatly improved.

[0063] 3. By setting the communication module to at least one of a 2G communication module, a 3G communication module, a 4G communication module, a 5G communication module, a NB-IOT communication module, a LORA communication module, a WIFI communication module, a Bluetooth communication module, a ZigBee communication module or a wired communication module, it can not only meet the requirements of wired network access, but also adapt to wireless remote control in areas without network coverage, thereby expanding the deployment scenarios of the cleaning device.

[0064] 4. By setting a horizontal sliding groove on the side of the beam, the nylon cleaning brush head is slidably connected to the sliding groove through the movable base and the slide seat, so as to realize the horizontal displacement of the nylon cleaning brush head, that is, to realize the flexible adjustment of the horizontal cleaning position of the nylon cleaning brush head.

[0065] 5. A circular suction nozzle is arranged in the middle of the nylon cleaning brush head, and the suction end of the centrifugal fan is connected to the circular suction nozzle. That is, during the rotating cleaning process of the nylon cleaning brush head, dust can be sucked through the circular suction nozzle, thereby greatly improving the cleaning quality.

[0066] 6. By setting up a belt conveyor directly under the lifting module and the cleaning module, and setting up debris collection enclosures around the belt conveyor, the debris dropped during the cleaning process can fall on the belt conveyor and be collected through the transmission of the belt conveyor. No manual cleaning is required, which further improves efficiency.

[0067] 7. With PLC as the core controller, integrated drive module, industrial camera, lifting module and stroke sensing sensor, the cleaning process is fully automated. The modules are electrically linked to form a closed-loop system, significantly reducing the need for manual intervention.

[0068] 8. The dual composite detection mechanism equipped with industrial cameras (visual detection) and travel sensing sensors can not only locate the contaminated area of the stencil through image recognition, but also monitor the displacement stroke of the cleaning module in real time to ensure operation accuracy and safety.

[0069] 9. By adopting the composite structure of lifting column + beam (sliding groove), the cleaning module can achieve dual degrees of freedom of vertical lifting and horizontal sliding. With the universal wheel, it can form X / Y / Z three-axis movement capability, which can meet the three-dimensional cleaning needs of screens of different specifications.

[0070] 10. The nylon cleaning brush head is integrated with a circular suction nozzle, and the centrifugal fan is used to achieve "cleaning-collecting" simultaneous operation. While the nylon cleaning brush head rotates to peel off impurities, the centrifugal fan (negative pressure system) instantly sucks, effectively preventing secondary pollution, and significantly improving efficiency compared to traditional step-by-step operations.

[0071] 11. A redundant detection system is formed by symmetrically setting the travel sensing sensors on the inner side of the lifting column, which can accurately identify the out-of-bounds movement of the cleaning module and prevent mechanical jamming or collision accidents.

[0072] 12. By using the ResNet50 network to extract the underlying texture features, edge features, and high-level semantic features of the mesh plate image, and combining with the Feature Pyramid Network (FPN) to achieve cross-scale feature fusion, it effectively captures the detailed features of tar regions of different sizes and improves the detection robustness in complex scenarios.

[0073] 13. By using Deformable Convolution to adaptively match the irregular mesh structure, combining with the classification branch to screen candidate regions, and the regression branch to optimize the coordinates of candidate regions, it significantly improves the localization accuracy of tar regions.

[0074] 14. By introducing dual units of spatial attention and channel attention in the attention enhancement module, extracting the spatial position sensitivity and channel feature importance respectively, and suppressing background noise through the fusion unit, it enhances the focusing ability of the model on tar regions.

[0075] 15. By setting the dynamic ROI alignment module to unify the candidate region size using bilinear interpolation and dynamically adjust the sampling grid density, it solves the feature misalignment problem caused by fixed partitioning in traditional ROI pooling and improves the feature alignment accuracy.

[0076] 16. By combining the candidate region localization loss (L_rpn), attention focusing loss (L_att), and classification loss (L_cls), and adjusting the weights of each loss through the balance coefficient (λ1), it effectively optimizes the training stability and detection performance of the model.

[0077] 17. By collecting images at different positions, angles, illuminations, and degrees of dirtiness, and combining preprocessing means such as image cropping, rotation, and noise reduction, it enhances data diversity and improves the generalization ability of the model.

[0078] 18. By randomly searching to dynamically adjust hyperparameters such as the learning rate and batch size, it avoids the subjectivity of manual hyperparameter tuning, accelerates model convergence, and improves detection accuracy.

[0079] 19. By adopting a three-stage verification of training set → validation set → test set, and evaluating the model performance by combining the dual indicators of detection accuracy and confidence, it ensures the reliability of the model and the deployment effect.

[0080] 20. By integrating the adjustment of multiple parameters such as the working height of the crossbeam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity, and the cleaning path, it adapts to the requirements of different working conditions and realizes the fine cleaning of the mesh plate.

[0081] 21. By using a travel induction sensor to calibrate the position of the cleaning module, it ensures the initial positioning accuracy and avoids the accumulation of mechanical errors.

[0082] 22. The industrial camera is used to collect the stencil images in real time and upload them to the server. The tar detection model quickly provides a tar detection report, driving the PLC to dynamically adjust the brush head speed, negative pressure intensity, and cleaning path, achieving a closed-loop control of "detection → cleaning → re-inspection → re-cleaning".

[0083] 23. The unqualified areas are cleaned a second time based on the tar detection report, and the cleaning logs (time, number of times, location, etc.) are recorded. The cleaning parameters are dynamically adjusted to form an intelligent system with self-iterative optimization.

[0084] 24. Through the decoupled design of the tar detection model and the cleaning control module, it is convenient for the model to be independently upgraded or replaced to meet the needs of technological iteration.

[0085] 25. By deeply integrating deep learning and automation control technologies, through intelligent tar detection, data-driven optimization, closed-loop feedback regulation, and efficient resource management, the accuracy, efficiency, and reliability of stencil cleaning are significantly improved. Description of the Drawings

[0086] The present invention will be further described below with reference to the drawings in conjunction with the embodiments.

[0087] Figure 1 It is the front view of a stencil cleaning device for a tobacco redrying machine according to the present invention.

[0088] Figure 2 It is the side view of a stencil cleaning device for a tobacco redrying machine according to the present invention.

[0089] Figure 3 It is the top view of a stencil cleaning device for a tobacco redrying machine according to the present invention.

[0090] Figure 4 It is the cross-sectional view of a stencil cleaning device for a tobacco redrying machine according to the present invention.

[0091] Figure 5 It is the circuit principle block diagram of a stencil cleaning device for a tobacco redrying machine according to the present invention.

[0092] Figure 6 It is the flowchart of a stencil cleaning method for a tobacco redrying machine according to the present invention.

[0093] Marking Explanation:

[0094] 100 - Tobacco leaf redrying machine screen cleaning device, 1 - Base, 2 - Control cabinet, 3 - PLC, 4 - Communication module, 5 - Industrial camera, 6 - Lifting module, 7 - Cleaning module, 8 - Debris collection module, 9 - Travel induction sensor, 10 - Driving module, 11 - Roller, 61 - Lifting column, 62 - Cross beam, 621 - Sliding groove, 71 - Slide base, 72 - Moving base, 73 - Nylon cleaning brush head, 74 - First motor, 75 - Centrifugal fan, 731 - Circular suction nozzle, 81 - Belt conveyor, 82 - Debris collection enclosure, 811 - Belt shaft, 101 - Second motor, 102 - Third motor. Detailed implementation manner

[0095] The technical solution in the embodiment of the present application has the following general idea: The PLC 3 controls the lifting module 6, the cleaning module 7 and the driving module 10 to perform automatic cleaning operations. Combining the position sensing of the travel induction sensor 9 and the cleaning verification of the industrial camera 5, it ensures that the cleaning position is correct and there are no dead corners, eliminating the need for manual disassembly, soaking, and brushing as in the traditional method, thereby improving the efficiency and quality of the tobacco leaf redrying machine screen cleaning and reducing the cleaning cost.

[0096] Please refer to Figures 1 to 6 As shown, a preferred embodiment of a tobacco leaf redrying machine screen cleaning device 100 of the present invention includes:

[0097] A base 1 with three groups of rollers 11 at the bottom; the base 1 is used to carry the cleaning device 100 and serves as a bracket; the rollers 11 are used to transplant the cleaning device 100;

[0098] A control cabinet 2 is provided at the edge of the top of the base 1;

[0099] A PLC 3 is provided in the control cabinet 2 and is used to control the operation of the cleaning device 100. In specific implementation, any PLC that can achieve this function can be selected from the prior art, and there is no limitation on the model. Moreover, the control program is well-known to those skilled in the art and can be obtained by those skilled in the art without creative labor;

[0100] A communication module 4 is provided in the control cabinet 2 and is connected to the PLC 3 for the communication between the cleaning device 100 and the outside world;

[0101] At least one industrial camera 5 is provided at the top of the base 1 with the shooting direction upward and is connected to the PLC 3 for shooting the screen image to judge whether it is cleaned;

[0102] A lifting module 6 is provided at the top of the base 1 and is connected to the PLC 3 for the lifting and translation of the cleaning module 7;

[0103] A cleaning module 7, installed on the lifting module 6 and connected to the PLC 3, is used to perform a cleaning operation on a screen plate (not shown).

[0104] A sundry collection module 8 is provided at the top of the base 1, directly below the lifting module 6 and the cleaning module 7, and is connected to the PLC 3, for collecting sundries dropped during the cleaning process of the screen plate.

[0105] Two travel induction sensors 9 are provided on the lifting module 6, with the induction direction facing the cleaning module 7, and are connected to the PLC 3; the travel induction sensor 9 is an optoelectronic sensor, used to sense the position of the cleaning module 7.

[0106] A driving module 10, with its power output end connected to the lifting module 6 and the cleaning module 7, and its control end connected to the PLC 3, is used to drive the lifting and translation of the cleaning module 7.

[0107] The roller 11 is a universal wheel, effectively improving the convenience of movement.

[0108] The communication module 4 is at least one of a 2G communication module, 3G communication module, 4G communication module, 5G communication module, NB-IOT communication module, LORA communication module, WIFI communication module, Bluetooth communication module, ZigBee communication module or a wired communication module.

[0109] The lifting module 6 includes:

[0110] Two lifting columns 61 are vertically provided at the top of the base 1, spaced apart from each other by a certain distance; the lifting columns 61 are used to support the cross beam 62.

[0111] A cross beam 62 is horizontally provided at the top of the two lifting columns 61, and a horizontal sliding groove 621 is provided on the side; the cleaning module 7 is installed on the cross beam 62 through the sliding groove 621 to achieve horizontal limit sliding.

[0112] The cleaning module 7 includes:

[0113] A sliding seat 71 is slidably connected to the sliding groove 621, and is used to drive the moving base 72 to displace.

[0114] A moving base 72 is connected to one end of the sliding seat 71 on the side, and is used to drive the nylon cleaning brush head 73 to displace.

[0115] A nylon cleaning brush head 73 is provided at the top of the moving base 72, and a circular suction nozzle 731 is provided in the middle; the nylon cleaning brush head 73 is used to clean the mesh plate by rotation; the circular suction nozzle 731 is used to suck the dust generated during the cleaning process;

[0116] A first motor 74 is provided at the bottom end of the moving base 72, with the power output end facing upward, and the power output end passes through the moving base 72 and is connected to the nylon cleaning brush head 73, and the control end is connected to the PLC 3; the power output end of the first motor 74 is connected to the nylon cleaning brush head 73 through a spherical joint (not shown) or a universal joint (not shown) to realize the angle adjustment of the nylon cleaning brush head 73;

[0117] A centrifugal fan 75, the suction end of which is communicated with the circular suction nozzle 731, and the control end is connected to the PLC 3, is used to generate negative pressure to suck dust through the circular suction nozzle 731.

[0118] The sundry collection module 8 includes:

[0119] A belt conveyor 81 is provided at the top of the base 1, located directly below the lifting module 6 and the cleaning module 7, and is connected to the PLC 3; the belt conveyor 81 operates through a belt shaft 811;

[0120] A sundry collection enclosure 82 is provided around the belt conveyor 81, and a notch (not shown) is provided at the transmission end of the belt conveyor 81.

[0121] Two of the travel induction sensors 9 are symmetrically provided inside the lifting columns 61 for sensing the position of the cleaning module 7.

[0122] The drive module 10 includes:

[0123] A second motor 101, the power output end of which is connected to the lifting column 61 of the lifting module 6, and the control end is connected to the PLC 3, is used to drive the lifting module 6 to lift; specifically, in implementation, the second motor 101 can drive the lifting module 6 to lift through a gear (not shown) and a rack (not shown);

[0124] A third motor 102, with its power output end connected to the cleaning module 7 and its control end connected to the PLC 3, is used to drive the cleaning module 7 to move horizontally. In specific implementation, the third motor 102 can drive a belt (not shown) to rotate to link the cleaning module 7 to make a reciprocating motion. That is, one end of the belt is sleeved on a rotating shaft (not shown), the other end is sleeved on the power output end of the third motor 102, and the belt is parallel to the sliding groove 621 and connected to the cleaning module 7. When the third motor 102 drives the belt to rotate, it links the cleaning module 7 to displace. The first motor 74, the second motor 101, and the third motor 102 are all servo motors.

[0125] A preferred embodiment of a method for cleaning the wire mesh plate of a tobacco leaf re-drying machine according to the present invention includes the following steps:

[0126] Step S1: Create a tar detection model based on a multi-scale feature extraction module, a candidate region localization module, an attention enhancement module, a dynamic ROI alignment module, and an output module, and set the loss function of the tar detection model;

[0127] Step S2: Obtain a large number of historical wire mesh plate images, preprocess and annotate each of the historical wire mesh plate images to construct a data set, divide the data set into a training set, a validation set, and a test set, and train, validate, and test the tar detection model through the training set, the validation set, and the test set in sequence, and deploy the tar detection model that passes the test to the server;

[0128] Through the decoupled design of the tar detection model and the cleaning control module, it is convenient for the model to be independently upgraded or replaced to meet the requirements of technology iteration.

[0129] Step S3: Move the cleaning device to the lower part of the wire mesh plate of the tobacco leaf re-drying machine through rollers, and the PLC obtains a cleaning instruction carrying cleaning parameters; the cleaning parameters at least include the working height of the cross beam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity, and the cleaning path;

[0130] By integrating the adjustment of multiple parameters such as the working height of the cross beam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity, and the cleaning path, it can meet the requirements of different working conditions and achieve fine cleaning of the wire mesh plate.

[0131] Step S4: Based on the cleaning instruction, the PLC drives the lifting column to rise through the second motor, and then links the cross beam to rise until it reaches the working height of the cross beam. Then, the third motor drives the cleaning module to translate in the sliding groove until it moves to the starting end of the cleaning path, and the position of the cleaning module is verified by the travel induction sensor to complete the initialization;

[0132] Step S5: The PLC controls the first motor to operate based on the brush head rotation speed to rotate the nylon cleaning brush head to perform a cleaning operation on the mesh plate. During the cleaning process, the third motor is driven to operate based on the translation speed of the cleaning module, so that the nylon cleaning brush head is displaced based on the translation speed of the cleaning module and the cleaning path; during the cleaning process, the centrifugal fan and the belt conveyor are started, so that the centrifugal fan sucks the dust on the mesh plate based on the negative pressure intensity, and the cleaned oil dirt falls on the belt conveyor and is conveyed externally through the belt conveyor until the cleaning of the cleaning path is completed;

[0133] Step S6: The PLC acquires the real-time mesh plate image of the mesh plate through an industrial camera and uploads the real-time mesh plate image to the server through the communication module; after preprocessing the received real-time mesh plate image, the server inputs it into the deployed tar detection model, obtains a tar detection report and issues it to the PLC;

[0134] Step S7: The PLC controls the displacement of the nylon cleaning brush head based on the tar position carried in the tar detection report, and dynamically adjusts the brush head rotation speed and the negative pressure intensity to perform secondary cleaning on the mesh plate at the tar position. After the secondary cleaning is completed, the real-time mesh plate image is acquired through the industrial camera for inspection.

[0135] By deeply integrating deep learning and automation control technologies, through intelligent tar detection, data-driven optimization, closed-loop feedback regulation, and efficient resource management, the accuracy, efficiency, and reliability of mesh plate cleaning are significantly improved.

[0136] In the step S1, the multi-scale feature extraction module is constructed based on the ResNet50 network and the feature pyramid network; the ResNet50 network is used to acquire the underlying texture features, edge features, and high-level semantic features from the input mesh plate image, and the feature pyramid network is used to fuse the underlying texture features, edge features, and high-level semantic features to obtain multi-scale features;

[0137] By using the ResNet50 network to extract the underlying texture features, edge features, and high-level semantic features of the mesh plate image, and combining the feature pyramid network (FPN) to achieve cross-scale feature fusion, the detailed features of tar regions of different sizes are effectively captured, and the detection robustness in complex scenarios is improved.

[0138] The candidate region localization module is constructed based on a deformable convolution with a size of 3×3, a classification branch, and a regression branch; the deformable convolution is used to adaptively match the irregularly shaped mesh holes and generate candidate regions of the mesh holes based on multi-scale features; the classification branch is used to determine whether there are mesh holes to screen the candidate regions; the regression branch is used to optimize the coordinates of the candidate regions;

[0139] Adaptive matching of irregular mesh structures through deformable convolution, screening candidate regions through the classification branch, and optimizing the coordinates of candidate regions through the regression branch significantly improve the positioning accuracy of tar regions.

[0140] The attention enhancement module is constructed based on a spatial attention unit, a channel attention unit, and a fusion unit; the spatial attention unit is used to extract spatial features from candidate regions; the channel attention unit is used to extract channel features from candidate regions; the fusion unit is used to fuse the spatial features and channel features and suppress background noise to obtain attention features;

[0141] By introducing dual units of spatial attention and channel attention in the attention enhancement module, the spatial position sensitivity and the importance of channel features are respectively extracted, and background noise is suppressed by the fusion unit, enhancing the model's focusing ability on tar regions.

[0142] The dynamic ROI alignment module is used to unify the sizes of each candidate region through bilinear interpolation and dynamically adjust the sampling grid density of the candidate regions;

[0143] By setting the dynamic ROI alignment module to use bilinear interpolation to unify the candidate region sizes and dynamically adjust the sampling grid density, the problem of feature misalignment caused by fixed partitioning in traditional ROI pooling is solved, and the feature alignment accuracy is improved.

[0144] The output module is constructed based on a fully connected layer, used to calculate the probability of tar existence in each candidate region according to the attention features, and output a tar detection report carrying the tar position based on the probability and the candidate regions;

[0145] The formula of the loss function is:

[0146] L_total = L_rpn + λ1L_att + L_cls;

[0147] Among them, L_total represents the loss value of the loss function; L_rpn represents the candidate region positioning loss; λ1 represents the balance coefficient; L_att represents the attention focusing loss; L_cls represents the tar classification loss.

[0148] By combining the candidate region positioning loss (L_rpn), the attention focusing loss (L_att), and the classification loss (L_cls), and adjusting the weights of each loss through the balance coefficient (λ1), the training stability and detection performance of the model are effectively optimized.

[0149] L_rpn = L_cls_rpn + λ2L_reg_rpn;

[0150] Among them, L_cls_rpn represents the cross-entropy loss; λ2 represents the balance coefficient; L_reg_rpn represents the smooth L1 loss;

[0151] L_cls_rpn = -Σ[p_i log(q_i) + (1 - p_i)log(1 - q_i)];

[0152] L_reg_rpn = Σsmooth_L1(t_i - t_i^*);

[0153] Among them, p_i represents whether the candidate region contains the true label of the mesh; q_i represents the classification branch prediction probability; t_i and t_i^* respectively represent the coordinate parameters of the predicted box and the true box;

[0154] L_att = Σ||A_f ⊙ (M_gt - M_pred)||2;

[0155] Among them, A_f represents the fused attention feature; M_gt represents the true tar region mask; M_pred represents the region response map generated by the attention feature; ⊙ represents the Hadamard product. The attention focusing loss is used to strengthen the spatial alignment between the attention feature and the true tar region.

[0156] L_cls = -Σ[y_j log(σ(s_j)) + (1 - y_j)log(1 - σ(s_j))];

[0157] Among them, y_j represents the tar presence label; s_j represents the logits output by the fully connected layer; σ represents the sigmoid function.

[0158] L_rpn: The candidate regions are screened through the classification branch (L_cls_rpn), the region coordinates are optimized by the regression branch (L_reg_rpn), and the offsets of the deformable convolution are implicitly optimized in the regression loss. L_att: Using the L2 distance loss between the attention feature and the true annotation, forcing the spatial attention and channel attention fusion features to suppress background noise and enhancing the representation ability for irregular tar regions. L_cls: The final classification loss strengthens the learning of difficult samples in the form of focal loss, and the feature consistency of the dynamic ROI alignment is implicitly optimized in the gradient propagation through the coordinate transformation of bilinear interpolation.

[0159] The specific steps of step S2 are as follows:

[0160] Obtain a large number of historical stencil images with different positions, angles, lighting conditions, and degrees of dirt. Perform preprocessing on each of the historical stencil images, including at least image cropping, image scaling, image rotation, image flipping, noise reduction, and color jitter. Mark the positions with tar in each of the preprocessed historical stencil images, and construct a dataset based on the marked historical stencil images;

[0161] By collecting images with different positions, angles, lighting, and degrees of dirt, and combining preprocessing means such as image cropping, rotation, and noise reduction, enhance data diversity and improve the generalization ability of the model.

[0162] Divide the dataset into a training set, a validation set, and a test set based on a ratio of 7:2:1. Train the tar detection model using the training set. During the training process, continuously optimize at least the hyperparameters of the tar detection model, including the learning rate, random dropout rate, batch size, and optimizer, through random search until the loss value of the loss function is less than a preset loss threshold;

[0163] Dynamically adjust hyperparameters such as the learning rate and batch size through random search, avoid the subjectivity of manual hyperparameter tuning, accelerate model convergence, and improve detection accuracy.

[0164] Calculate the detection accuracy through the validation set to verify the trained tar detection model. If the verification fails, expand the training set and continue training; if the verification passes, then:

[0165] Calculate the confidence level through the test set to test the tar detection model that has passed the verification. If the test fails, expand the training set and continue training; if the test passes, end the training;

[0166] Deploy the tar detection model that has passed the test to the server;

[0167] By adopting a three-stage verification of training set → validation set → test set, and evaluating the model performance with dual indicators of detection accuracy and confidence level, ensure the reliability of the model and the deployment effect.

[0168] The step S7 further includes:

[0169] If the secondary cleaning does not meet the standard, perform the cleaning operation again until the cleaning meets the standard, and record the cleaning log including at least the cleaning instruction, cleaning time, number of repeated cleanings, and repeated cleaning positions. Dynamically adjust the cleaning parameters based on the cleaning log, and encrypt and backup the cleaning log.

[0170] Perform secondary cleaning on the non-compliant area through the tar detection report, record the cleaning log (time, number of times, position, etc.), dynamically adjust the cleaning parameters, and form an intelligent system with self-iterative optimization.

[0171] The encryption storage and backup of the cleaning log are specifically as follows:

[0172] The cleaning log is encrypted by the RC2 algorithm to obtain first-level encrypted data. A random string of a specified length is generated, and the random string is inserted after the 9th character of the first-level encrypted data to obtain second-level encrypted data. The second-level encrypted data is encrypted by the 3DES algorithm to obtain third-level encrypted data. The third-level encrypted data is segmented based on a 7:2 segmentation ratio and the order of the front and back is swapped to obtain fourth-level encrypted data. The fourth-level encrypted data is converted into hexadecimal data, and the numbers 8 and the letter B in the hexadecimal data are swapped, and the number 1 and the letter C are swapped to obtain fifth-level encrypted data. The fifth-level encrypted data is encrypted by the SM9 algorithm to obtain an encrypted log, and the encrypted log is stored and distributed for backup. By combining multiple encryption algorithms and data transformation rules, the security of the cleaning log storage is effectively improved. If any one of the encryption algorithms or data transformation rules is not known, the encrypted log cannot be cracked.

[0173] In summary, the advantages of the present invention are as follows:

[0174] 1. By setting a base, a PLC, a communication module, an industrial camera, a lifting module, a cleaning module, a travel induction sensor, and a driving module, the PLC is respectively connected to the communication module, the industrial camera, the lifting module, the cleaning module, the travel induction sensor, and the driving module. The cleaning module is installed on the base through the lifting module. The travel induction sensor is used to sense the position of the cleaning module, and the driving module is used to drive the lifting module to lift and drive the cleaning module to translate. And the bottom end of the base is provided with rollers. When it is necessary to clean the mesh plate, only need to move the cleaning device to the cleaning position of the tobacco leaf re-drying machine through the rollers, then lift the cleaning module by the lifting module to contact the mesh plate, and rotate the nylon cleaning brush head of the cleaning module to perform the cleaning operation. The cleaning position can be adjusted by translating the cleaning module. Subsequently, the industrial camera can be used to check the cleaning situation. If the cleaning is not up to standard, repeated cleaning can be carried out, that is, the PLC controls the lifting module, the cleaning module, and the driving module to perform automatic cleaning operations, without the need for manual disassembly, soaking, and brushing as in the traditional method, thereby greatly improving the efficiency and quality of the mesh plate cleaning of the tobacco leaf re-drying machine and greatly reducing the cleaning cost.

[0175] 2. By setting the rollers as universal wheels, the convenience of transplanting the cleaning device is greatly improved.

[0176] 3. By setting the communication module as at least one of a 2G communication module, 3G communication module, 4G communication module, 5G communication module, NB-IOT communication module, LORA communication module, WIFI communication module, Bluetooth communication module, ZigBee communication module or wired communication module, it not only meets the wired network access but also can adapt to the wireless remote control in areas without network coverage, expanding the deployment scenarios of the cleaning device.

[0177] 4. By setting a horizontal sliding groove on the side of the cross beam, the nylon cleaning brush head is slidably connected to the sliding groove through a moving base and a sliding seat, realizing the horizontal displacement of the nylon cleaning brush head, that is, realizing the flexible adjustment of the horizontal cleaning position of the nylon cleaning brush head.

[0178] 5. By setting a circular suction nozzle in the middle of the nylon cleaning brush head and connecting the suction end of the centrifugal fan to the circular suction nozzle, that is, during the rotation and cleaning process of the nylon cleaning brush head, dust can also be sucked through the circular suction nozzle, thus greatly improving the cleaning quality.

[0179] 6. By setting a belt conveyor directly below the lifting module and the cleaning module and setting a debris collection enclosure around the belt conveyor, the debris falling during the cleaning process can fall on the belt conveyor and be collected through the transmission of the belt conveyor, eliminating the need for manual cleaning and further improving the efficiency.

[0180] 7. With the PLC as the core controller, integrating the drive module, industrial camera, lifting module and travel induction sensor, the full-automatic control of the cleaning process is realized. The modules form a closed-loop system through electrical linkage, significantly reducing the need for manual intervention.

[0181] 8. By equipping a dual composite detection mechanism of an industrial camera (visual inspection) and a travel induction sensor, it can not only identify the contaminated area of the stencil through image recognition but also monitor the displacement travel of the cleaning module in real time to ensure the operation accuracy and safety.

[0182] 9. By adopting the composite structure of a lifting column + cross beam (sliding groove), the vertical lifting and horizontal sliding of the cleaning module are realized with two degrees of freedom. Cooperating with the universal wheels, the X / Y / Z three-axis movement ability is formed, which can meet the three-dimensional cleaning requirements of different specifications of stencils.

[0183] 10. By integrating a circular suction nozzle into the nylon cleaning brush head and realizing the synchronous operation of "cleaning - collection" through a centrifugal fan, while the nylon cleaning brush head rotates to peel off impurities, the centrifugal fan (negative pressure system) immediately sucks, effectively preventing secondary pollution and significantly improving the efficiency compared with the traditional step-by-step operation.

[0184] 11. By symmetrically setting travel induction sensors inside the lifting column, a redundant detection system is formed, which can accurately identify the out-of-bounds movement of the cleaning module and prevent mechanical jamming or collision accidents.

[0185] 12. By using the ResNet50 network to extract the underlying texture features, edge features, and high-level semantic features of the mesh plate image, and combining with the Feature Pyramid Network (FPN) to achieve cross-scale feature fusion, effectively capture the detailed features of tar regions of different sizes, and improve the detection robustness in complex scenarios.

[0186] 13. By using deformable convolution to adaptively match the irregular mesh structure, combining with the classification branch to screen candidate regions, and the regression branch to optimize the coordinates of candidate regions, significantly improve the positioning accuracy of tar regions.

[0187] 14. By introducing dual units of spatial attention and channel attention in the attention enhancement module, respectively extracting the spatial position sensitivity and channel feature importance, and suppressing background noise through the fusion unit, enhance the focusing ability of the model on tar regions.

[0188] 15. By setting the dynamic ROI alignment module to unify the candidate region size using bilinear interpolation and dynamically adjust the sampling grid density, solve the feature misalignment problem caused by fixed partitioning in traditional ROI pooling, and improve the feature alignment accuracy.

[0189] 16. By combining the candidate region localization loss (L_rpn), attention focusing loss (L_att), and classification loss (L_cls), and adjusting the weights of each loss through the balance coefficient (λ1), effectively optimize the training stability and detection performance of the model.

[0190] 17. By collecting images at different positions, angles, illuminations, and degrees of dirtiness, and combining preprocessing means such as image cropping, rotation, and noise reduction, enhance data diversity and improve the generalization ability of the model.

[0191] 18. By randomly searching to dynamically adjust hyperparameters such as the learning rate and batch size, avoid the subjectivity of manual parameter tuning, accelerate model convergence, and improve detection accuracy.

[0192] 19. By adopting a three-stage verification of training set → validation set → test set, and evaluating the model performance by combining the dual indicators of detection accuracy and confidence, ensure the reliability and deployment effect of the model.

[0193] 20. By integrating the adjustment of multiple parameters such as the working height of the crossbeam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity, and the cleaning path, adapt to the requirements of different working conditions, and achieve fine cleaning of the mesh plate.

[0194] 21. By using a travel induction sensor to calibrate the position of the cleaning module, ensure the initial positioning accuracy, and avoid the accumulation of mechanical errors.

[0195] 22. The industrial camera is used to collect the stencil images in real time and upload them to the server. The tar detection model quickly provides a tar detection report, driving the PLC to dynamically adjust the brush head speed, negative pressure intensity, and cleaning path to achieve a closed-loop control of "detection → cleaning → re-inspection → re-cleaning".

[0196] 23. The areas that do not meet the standards are cleaned a second time based on the tar detection report, and the cleaning logs (time, number of times, location, etc.) are recorded. The cleaning parameters are dynamically adjusted to form an intelligent system that self-iterates and optimizes.

[0197] 24. Through the decoupled design of the tar detection model and the cleaning control module, it is convenient for the model to be independently upgraded or replaced to meet the needs of technological iteration.

[0198] 25. By deeply integrating deep learning and automation control technologies, through intelligent tar detection, data-driven optimization, closed-loop feedback regulation, and efficient resource management, the accuracy, efficiency, and reliability of stencil cleaning are significantly improved.

[0199] Although the specific implementation manners of the present invention have been described above, those skilled in the art of this technology should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.

Claims

1. A screen cleaning device for a tobacco leaf redrying machine, characterized in that: Including: A base with three sets of rollers provided at the bottom; A control cabinet provided at the edge of the top of the base; A PLC provided in the control cabinet; A communication module provided in the control cabinet and connected to the PLC; At least one industrial camera provided at the top of the base, with the shooting direction facing upward and connected to the PLC; A lifting module provided at the top of the base and connected to the PLC; A cleaning module installed on the lifting module and connected to the PLC; A sundry collection module provided at the top of the base, directly below the lifting module and the cleaning module, and connected to the PLC; Two travel induction sensors provided on the lifting module, with the induction direction facing the cleaning module and connected to the PLC; A driving module, with the power output end connected to the lifting module and the cleaning module, and the control end connected to the PLC.

2. The screen cleaning device for a tobacco redrying machine according to claim 1, wherein: The rollers are universal wheels.

3. The screen cleaning device for a tobacco leaf redrying machine according to claim 1, characterized in that: The communication module is at least one of a 2G communication module, 3G communication module, 4G communication module, 5G communication module, NB-IOT communication module, LORA communication module, WIFI communication module, Bluetooth communication module, ZigBee communication module or a wired communication module.

4. A screen cleaning device for a tobacco leaf re-drying machine according to claim 1, characterized in that: The lifting module includes: Two lifting columns vertically provided at the top of the base, spaced a certain distance apart from each other; A cross beam horizontally provided at the top of the two lifting columns, and a horizontal sliding groove is provided on the side; the cleaning module is installed on the cross beam through the sliding groove; The cleaning module includes: A sliding seat slidably connected to the sliding groove; A moving base, with one end of the side connected to the sliding seat; A nylon cleaning brush head provided at the top of the moving base, and a circular suction nozzle is provided in the middle; A first motor provided at the bottom of the moving base, with the power output end facing upward, and the power output end passes through the moving base and is connected to the nylon cleaning brush head, and the control end is connected to the PLC; A centrifugal fan, with the suction end communicated with the circular suction nozzle, and the control end is connected to the PLC.

5. The screen cleaning device for a tobacco leaf redrying machine according to claim 1, wherein: The sundry collection module includes: A belt conveyor provided at the top of the base, directly below the lifting module and the cleaning module, and connected to the PLC; A sundry collection enclosure provided around the belt conveyor, and a notch is provided at the transmission end of the belt conveyor.

6. The screen cleaning device for a tobacco leaf redrying machine according to claim 4, wherein: The two travel induction sensors are symmetrically provided inside the lifting columns.

7. The wire mesh cleaning device for a tobacco redrying machine according to claim 1, characterized in that: The driving module includes: A second motor, with the power output end connected to the lifting module, and the control end connected to the PLC; A third motor, with the power output end connected to the cleaning module, and the control end connected to the PLC.

8. A method for cleaning the wire mesh plate of a tobacco leaf re-drying machine, characterized in that: The method needs to use the tobacco leaf redrying machine screen cleaning device according to any one of claims 1 to 7, and includes the following steps: Step S1, create a tar detection model based on a multi-scale feature extraction module, a candidate region localization module, an attention enhancement module, a dynamic ROI alignment module and an output module, and set the loss function of the tar detection model; Step S2, obtaining a large number of historical stencil images, constructing a data set after preprocessing and annotating each of the historical stencil images, dividing the data set into a training set, a validation set and a test set, training, validating and testing the tar detection model through the training set, validation set and test set in turn, and deploying the tar detection model that passes the test to the server; Step S3, moving the cleaning device to the bottom of the screen of the tobacco redrying machine through the roller, and the PLC obtains the input cleaning instruction carrying cleaning parameters; the cleaning parameters at least include the working height of the beam, the translation speed of the cleaning module, the rotation speed of the brush head, the negative pressure intensity and the cleaning path; Step S4, based on the cleaning instruction, the PLC drives the lifting column to rise through the second motor, and then links the crossbeam to rise until it rises to the working height of the crossbeam, and then drives the cleaning module to translate in the sliding groove through the third motor until it moves to the starting end of the cleaning path, and verifies the position of the cleaning module through the stroke sensing sensor to complete the initialization; Step S5, the PLC controls the first motor to work based on the brush head speed, so as to rotate the nylon cleaning brush head to perform a cleaning operation on the mesh plate, and drives the third motor to work based on the translation speed of the cleaning module during the cleaning process, so that the nylon cleaning brush head moves based on the translation speed of the cleaning module and the cleaning path; the centrifugal fan and the belt conveyor are started during the cleaning process, so that the centrifugal fan sucks the dust on the mesh plate based on the negative pressure intensity, and the cleaned grease falls on the belt conveyor and is transported to the outside through the belt conveyor until the cleaning of the cleaning path is completed; Step S6, the PLC collects a real-time screen image of the screen through an industrial camera, and uploads the real-time screen image to a server through a communication module; After preprocessing the received real-time screen image, the server inputs the deployed tar detection model, obtains the tar detection report and sends it to the PLC; Step S7, the PLC controls the displacement of the nylon cleaning brush head based on the tar position carried in the tar detection report, and dynamically adjusts the brush head rotation speed and negative pressure intensity to perform secondary cleaning of the mesh at the tar position. After the secondary cleaning is completed, the real-time mesh image is collected by the industrial camera for inspection.

9. A method for cleaning the screen plate of a tobacco redrying machine according to claim 8, characterized in that: In step S1, the multi-scale feature extraction module is constructed based on the ResNet50 network and the feature pyramid network; the ResNet50 network is used to collect underlying texture features, edge features and high-level semantic features from the input stencil image, and the feature pyramid network is used to fuse the underlying texture features, edge features and high-level semantic features to obtain multi-scale features; The candidate region positioning module is constructed based on a deformable convolution with a size of 3×3, a classification branch, and a regression branch; The deformable convolution is used to adaptively match the irregularly shaped mesh and generate a candidate region of the mesh based on multi-scale features; The classification branch is used to determine whether there is a mesh to screen the candidate area; The regression branch is used to optimize the coordinates of the candidate region; The attention enhancement module is constructed based on a spatial attention unit, a channel attention unit, and a fusion unit; The spatial attention unit is used to extract spatial features from the candidate regions; The channel attention unit is used to extract channel features from the candidate regions; The fusion unit is used to fuse the spatial features and the channel features, and suppress background noise to obtain attention features; The dynamic ROI alignment module is used to unify the sizes of the candidate regions by bilinear interpolation and dynamically adjust the sampling grid density of the candidate regions; The output module is constructed based on a fully connected layer, and is used to calculate the probability of tar existing in each candidate region according to the attention features, and output a tar detection report carrying the tar position based on the probability and the candidate regions; The formula of the loss function is: L_total = L_rpn + λ1L_att + L_cls; Wherein, L_total represents the loss value of the loss function; L_rpn represents the candidate region localization loss; λ1 represents the balance coefficient; L_att represents the attention focusing loss; L_cls represents the tar classification loss.

10. A method for cleaning the wire mesh plate of a tobacco leaf re-drying machine according to claim 8, characterized in that: The specific steps of step S2 are as follows: Obtain a large number of historical stencil images with different positions, different angles, different illumination conditions, and different degrees of dirt, perform preprocessing on each of the historical stencil images at least including image cropping, image scaling, image rotation, image flipping, noise reduction, and color jitter, label the positions where tar exists in each of the preprocessed historical stencil images, and construct a dataset based on the labeled historical stencil images; Divide the dataset into a training set, a validation set, and a test set according to a ratio of 7:2:1, train the tar detection model through the training set, and continuously optimize at least the hyperparameters of the tar detection model including the learning rate, the random dropout rate, the batch size, and the optimizer during the training process by random search until the loss value of the loss function is less than a preset loss threshold; Calculate the detection accuracy through the validation set to verify the trained tar detection model. If the verification fails, expand the training set and continue training; if the verification passes, then: Calculate the confidence through the test set to test the tar detection model that has passed the verification. If the test fails, expand the training set and continue training; If the test passes, end the training; Deploy the tar detection model that has passed the test to the server; Step S7 further includes: If the secondary cleaning does not meet the standard, perform the cleaning operation again until the cleaning meets the standard, record the cleaning log at least including the cleaning instruction, the cleaning time, the number of repeated cleaning times, and the repeated cleaning position, dynamically adjust the cleaning parameters based on the cleaning log, and encrypt and back up the cleaning log.

Images