A method and system for intelligent control of negative pressure adsorption conveying based on machine learning

Abstract

The application provides a kind of intelligent control method and system of negative pressure adsorption conveying based on machine learning.The system includes conveying belt assembly, partition vacuum adsorption module, visual recognition module, pressure sensor array, flexible adaptive structure and main control unit.Through visual recognition and pressure array fusion, the size, position and distribution information of photovoltaic silicon wafer are obtained, and a material distribution mapping matrix is constructed.The main control unit generates a time-space adsorption window map based on trajectory prediction, and issues dynamic control instructions to the partition vacuum adsorption module, realizing on-demand opening and negative pressure intensity adjustment of different partitions.The flexible adaptive structure realizes the compensation of the silicon wafer bottom surface through the micro-deformation adjustment of the air bag array.The application significantly reduces energy consumption while ensuring stable silicon wafer conveying, improves the flexibility, intelligence and robustness of the system, and is suitable for high-speed conveying and sorting of photovoltaic silicon wafers.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic silicon wafer conveying technology, and in particular to a machine learning-based intelligent control method and system for negative pressure adsorption conveying. Background Technology

[0002] As a key material in photovoltaic module manufacturing, photovoltaic silicon wafers typically undergo processes including cutting, cleaning, inspection, and sorting. Between these processes, the silicon wafers require high-speed, stable transport via conveyor systems. Existing technologies have proposed using negative pressure belt conveyor systems to transport silicon wafers, preventing slippage and misalignment during transport.

[0003] However, existing negative pressure belt conveyor systems mostly adopt a fixed adsorption zone design, lacking independent zone control and dynamic adjustment capabilities, resulting in low adsorption efficiency and high energy consumption during multi-specification mixed-flow production. Secondly, existing negative pressure belt conveyor systems lack feedback optimization mechanisms based on indicators such as adsorption success rate, energy consumption, and deviation rate. Therefore, we propose a machine learning-based intelligent control method and system for negative pressure adsorption conveying. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a machine learning-based intelligent control method and system for negative pressure adsorption and delivery, thereby solving the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A machine learning-based intelligent control method for negative pressure adsorption and delivery includes the following steps:

[0007] S1. Obtain the size, shape and position distribution information of the material to be conveyed on the conveyor belt, wherein the information is jointly obtained by the visual recognition module and the belt surface pressure sensor to construct a material spatial distribution map and a contact state map.

[0008] S2. Based on the material spatial distribution map and contact state map, combined with the conveyor belt running speed, cycle time information and material form, a corresponding temporal and spatial adsorption window map is generated to indicate the activation state and required adsorption intensity of each negative pressure zone at a specific time.

[0009] S3. Based on the adsorption window diagram, dynamically control the opening status of vacuum valves and the negative pressure output intensity of multiple independent negative pressure zones below the conveyor belt to perform adsorption on demand.

[0010] S4. During the material adsorption process, based on the flexible structure of the belt and the real-time contact state of the material, the deformation of the adsorption surface is dynamically adjusted to improve the adhesion effect and reduce sliding offset.

[0011] S5. Based on historical material adsorption feedback data, energy consumption data and offset rate, construct a mapping model between material type and adsorption parameters to form an adsorption parameter memory library.

[0012] S6. By using reinforcement learning or deep neural network models deployed in the main control unit, the current feedback data is compared and analyzed with the memory model to dynamically update the adsorption window diagram and partition control parameters, thereby achieving continuous adaptive optimization of the adsorption behavior of multiple types of materials.

[0013] S1 specifically includes:

[0014] Two-dimensional image information of the material is obtained by a vision recognition module installed above the conveyor belt, and the size, boundary and shape characteristics of the material are identified.

[0015] Information on the contact pressure distribution between the material and the belt is collected by pressure sensors installed on the surface of the conveyor belt.

[0016] The two-dimensional image information is fused with the contact pressure distribution information to generate material distribution mapping data that includes material location, size, and contact state.

[0017] S2 specifically includes:

[0018] Based on the material distribution mapping data, combined with the real-time operating speed and cycle time parameters of the conveyor belt, the trajectory of the material's position change during the conveying process is calculated.

[0019] Based on the position change trajectory, the required adsorption area position of the material on the conveying path is determined;

[0020] By combining the size and contact state of different materials at the adsorption region location, a corresponding time-space adsorption window diagram is generated to indicate the activation state and negative pressure intensity of each negative pressure zone at a specific moment.

[0021] S3 specifically includes:

[0022] Based on the aforementioned temporal and spatial adsorption window diagram, control commands are sent to the partitioned vacuum adsorption modules below the conveyor belt.

[0023] After receiving the control command, the partitioned vacuum adsorption module opens the vacuum valve of the corresponding partition as needed and sets the negative pressure intensity.

[0024] During the transport process, the opening status of each negative pressure zone is dynamically switched according to the timing requirements of the adsorption window diagram to achieve adsorption on demand and reduce ineffective air extraction.

[0025] S4 specifically includes:

[0026] During the process of material adsorption, the adhesion status between the material and the surface of the conveyor belt is monitored in real time.

[0027] When local unevenness in bonding is detected, the elastic layer or micro-convex structure of the flexible adapter belt structure is deformed.

[0028] The deformation of the elastic layer or micro-convex structure allows the surface of the conveyor belt to automatically adjust to a conforming shape that matches the bottom surface of the material, thereby enhancing adsorption stability.

[0029] S5 specifically includes:

[0030] Collect data on the adsorption success rate, energy consumption, and material offset rate of materials during the transportation process;

[0031] The adsorption success rate data, energy consumption data, and material offset rate data are correlated with the material distribution mapping data;

[0032] Based on the correlation analysis results, a mapping relationship between material type and adsorption parameters is generated, and an adsorption memory model is constructed.

[0033] S6 specifically includes:

[0034] The current adsorption operation data of the material is compared with the adsorption memory model to identify parameter deviations;

[0035] Based on the identified parameter deviations, new negative pressure zoning control parameters are generated through reinforcement learning algorithms or neural network models deployed in the main control unit.

[0036] Update the spatiotemporal adsorption window diagram to form a new adsorption strategy;

[0037] During the next material transport, the partitioned vacuum adsorption module is controlled according to the updated adsorption strategy to achieve continuous adaptive optimization.

[0038] A machine learning-based intelligent control system for negative pressure adsorption and delivery includes:

[0039] The conveyor belt assembly is used to carry and transport photovoltaic silicon wafers. The surface of the conveyor belt is embedded with a flexible adapter structure and a resistive pressure sensor array to collect the contact pressure distribution between the material and the belt.

[0040] The visual recognition module, located above the conveyor belt assembly, is used to acquire two-dimensional images and depth information of the photovoltaic silicon wafers, and, combined with the output of the pressure sensor array, generate a material distribution mapping matrix.

[0041] The partitioned vacuum adsorption module is arranged below the conveyor belt, which divides the conveying area into multiple independently controllable negative pressure zones. Each zone is equipped with a vacuum valve and a vacuum pump, which can be opened and adjusted as needed.

[0042] The main control unit is electrically connected to the visual recognition module, the pressure sensor array, and the partitioned vacuum adsorption module.

[0043] The beneficial effects of this invention are as follows:

[0044] This invention integrates a visual recognition module with a pressure sensor array to simultaneously acquire the size, boundary shape, and contact pressure distribution of photovoltaic silicon wafers with the conveyor belt. An extended Kalman filter is then used to generate a material distribution mapping matrix. Compared to traditional single-visual recognition methods, this approach significantly improves the accuracy of identifying silicon wafers of different specifications and thicknesses, avoiding recognition errors caused by factors such as illumination and reflection.

[0045] This invention constructs a time-space adsorption window diagram, allowing the main control unit to precisely indicate the opening and closing times and target negative pressure intensity of each negative pressure zone, thus achieving dynamic scheduling of the zoned vacuum valves. This mechanism avoids the energy waste of continuous global vacuuming, reducing energy consumption by approximately 30%–50%, while ensuring the stability of materials during high-speed transport.

[0046] This invention introduces a flexible adapting structure on the surface of the conveyor belt. Through micro-deformation compensation via a micro-airbag array or elastic layer, it can actively adjust the belt surface morphology when uneven adhesion is detected in real time, thereby ensuring uniform adhesion between the bottom surface of the silicon wafer and the belt surface. This measure effectively avoids edge lifting and localized stress concentration caused by belt vibration or surface unevenness, significantly reducing the silicon wafer breakage rate.

[0047] This invention collects operational data such as adsorption success rate, energy consumption, and offset rate to construct a mapping relationship between material type and adsorption parameters, forming an adsorption parameter memory library. In multi-specification mixed-flow production, the system can directly call upon historically optimal parameter combinations, shortening the debugging cycle and improving the first-time adsorption success rate.

[0048] This invention utilizes a reinforcement learning or neural network model deployed in the main control unit. The system can perform error analysis between predicted values ​​and actual operating data, and execute online retraining and parameter updates when necessary. This mechanism enables the system to maintain a high level of adsorption accuracy and stability even under long-term operation and complex conditions, achieving adaptive optimization of the silicon wafer transport process. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of a machine learning-based intelligent control method for negative pressure adsorption and delivery according to the present invention.

[0050] Figure 2 This is a schematic diagram of the framework of a machine learning-based intelligent control system for negative pressure adsorption and delivery according to the present invention.

[0051] Figure 3This is a schematic diagram of an existing negative pressure belt conveyor assembly. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1: As Figure 1 As shown, this embodiment provides a machine learning-based intelligent control method for negative pressure adsorption and delivery, including the following steps:

[0054] S1: Material multimodal recognition: Obtain the size, shape and position distribution information of the material (photovoltaic silicon wafer) to be conveyed on the conveyor belt. The information is obtained by the visual recognition module and the belt surface pressure sensor to construct a material spatial distribution map and a contact state map.

[0055] S2: Generate time-space adsorption window diagram: Based on the material distribution map, combined with the conveyor belt running speed, cycle time information and material form, generate the corresponding "time-space adsorption window diagram" to indicate the activation state and required adsorption intensity of each negative pressure zone at a specific time.

[0056] S3: Dynamic execution of adsorption strategy: Based on the adsorption window diagram, dynamically control the opening status of vacuum valves and the negative pressure output intensity of multiple independent negative pressure zones below the conveyor belt to perform adsorption on demand and avoid ineffective pumping;

[0057] S4: Flexible adaptive bonding control: During the material adsorption process, based on the flexible structure of the belt and the real-time contact state of the material, the deformation of the adsorption surface is dynamically adjusted to improve the bonding effect and reduce slippage.

[0058] S5: Construct an adsorption memory model: Based on historical material adsorption feedback data, energy consumption data and offset rate, construct a mapping model between material type and adsorption parameters to form an adsorption parameter memory library;

[0059] S6: Machine Learning Feedback Tuning: By using reinforcement learning or deep neural network models deployed in the main control unit, the current feedback data is compared and analyzed with the memory model to dynamically update the adsorption window diagram and partition control parameters, thereby achieving continuous adaptive optimization of the adsorption behavior of multiple types of materials.

[0060] S1: Material Multimodal Recognition: Acquire the size, shape, and positional distribution information of the material to be conveyed on the conveyor belt. This information is jointly acquired by a visual recognition module and a belt surface pressure sensor to construct a material spatial distribution map and a contact state map. Specifically, this includes the following sub-steps:

[0061] S110: Two-Dimensional Image Information Acquisition: A structured light 3D camera positioned above the conveyor belt acquires two-dimensional images and depth information of the material. The camera is vertically mounted 300 mm above the conveyor belt, covering a 250 mm × 250 mm field of view, with a resolution of 1920 × 1080, a frame rate of 30 Hz, and a depth measurement accuracy of ±0.1 mm. To ensure image quality, the camera is equipped with a ring-shaped LED supplementary light with a color temperature of 5000 K and an illuminance maintained between 500-800 lx to avoid recognition errors caused by fluctuations in external lighting, shadows, and reflections. After installation, the camera is calibrated using a checkerboard calibration board to establish the transformation relationship between pixel coordinates and physical coordinates. Calibration uses a projection matrix. The formula is:

[0062]

[0063] in,( (X,Y) represents the image pixel coordinates, and (X,Y) represents the physical coordinates. The projection matrix is ​​3×3. The calibration error is controlled within 0.05mm. During image processing, the Canny edge detection operator is used to extract the material boundary curve; then, the minimum bounding rectangle fitting algorithm is used to calculate the boundary rectangle of the material, obtaining the material's length L, width W, and shape feature parameters.

[0064] The formulas for calculating the length and width of the material are as follows:

[0065]

[0066] In the formula, These represent the maximum and minimum coordinate values ​​of the material boundary points in the image coordinate system.

[0067] S120: Contact Pressure Distribution Information Acquisition: A 10 × 10 resistive pressure sensor array is uniformly arranged along the transverse and longitudinal directions on the conveyor belt surface, with a total coverage area of ​​200 mm × 200 mm and a center-to-center spacing of 20 mm between each sensor. The sensor range is 0–10 kPa, the detection resolution is ±5 Pa, and the sampling frequency is 50 Hz. The sensors are embedded 0.5 mm below the belt surface via a flexible substrate to ensure accurate measurement without affecting belt flexibility and service life. The sensor surface is covered with a 0.2 mm thick polyurethane protective film to improve wear resistance, dustproofing, and waterproofing, meeting the IP54 protection rating. All sensors undergo zero-point calibration and full-scale calibration after installation, with a calibration error of less than ±1%FS. Sensor signals are acquired at a frequency of 1 kHz via a 16-bit A / D conversion module and input to the main control unit. Before entering the main control unit, the data is processed by a low-pass filter with a cutoff frequency set to 10 Hz to eliminate high-frequency noise caused by conveyor belt vibration and electromagnetic interference. When the material coverage area exceeds the boundary of the pressure array, the main control unit uses an extrapolation algorithm to estimate the pressure in the edge area, thereby avoiding the fitting recognition error caused by missing data.

[0068] S130, Data Fusion and Material Distribution Mapping Generation: The material contour coordinates obtained in S110 and the contact pressure distribution data obtained in S120 are fused together, and the extended Kalman filter algorithm (EKF) is used for dynamic fusion.

[0069] In EKF: the state vector is defined as ,in Indicates the center position of the material. The velocity components of the material in the transverse and longitudinal directions are represented by the state transition equation:

[0070]

[0071] in, The sampling period; This represents the target's coordinate position at the current time k; , This represents the predicted coordinate position of the target at the next time step k+1; the observation vector is defined as the coordinates of the material center obtained from visual recognition and the coordinates of the contact point of the pressure array; the observation equation is:

[0072]

[0073] in The measurement noise is represented by the following initial conditions: the state vector is initialized based on the initial center position of the material obtained from visual recognition, and the initial value of the covariance matrix is ​​set to the identity matrix I; the noise model is as follows: the process noise matrix Q is set to 0.01I, and the observation noise matrix R is set to 0.1I, which were obtained through experimental optimization. The EKF refresh rate is 50Hz, that is, the state estimate is updated every 20ms. If visual data is lost, the system automatically recalls the previous prediction value and uses the contact point information of the pressure array for correction to avoid recognition errors caused by data interruption.

[0074] The fusion output is a material distribution mapping matrix:

[0075] in, This indicates the spatial distribution of materials on the conveyor belt. This represents the position coordinates of the material in the physical coordinate system. This represents the contact pressure intensity at the corresponding location. The matrix resolution is 5mm × 5mm cells, with the origin set at the left front reference point of the conveyor belt. The matrix is ​​stored in the main control unit's memory as a two-dimensional array and refreshed every 20ms. For example, in a certain sampling, if the contact pressure at coordinate point (50mm, 100mm) is 120Pa, then the corresponding matrix cell is... .

[0076] S2: Generate a time-space adsorption window diagram: Based on the material distribution diagram, combined with the conveyor belt speed, cycle time information, and material morphology, generate a corresponding "time-space adsorption window diagram" to indicate the activation state and required adsorption intensity of each negative pressure zone at a specific time; specifically including the following sub-steps:

[0077] S210: Material location trajectory calculation: based on the material distribution mapping matrix generated in S130 By combining the conveyor belt running speed parameter v and the running cycle parameter T, the trajectory of the material's position change during the conveying process is calculated.

[0078] The belt speed is obtained by an optical encoder mounted on the drive pulley shaft, with a resolution of 0.01 m / s. In cases using a servo motor drive, the speed information can be fed back from the motor controller as input. If acceleration disturbances need to be considered, a MEMS accelerometer is installed under the belt to collect ±2 m / s² data. 2 Real-time acceleration signals within the specified range. This method is applicable to operating conditions with conveyor speeds ranging from 0.1 to 2.0 m / s.

[0079] The coordinates of the center position of the material at time t are calculated by the following formula:

[0080]

[0081] in, , This represents the predicted position (two-dimensional coordinates) of the material at time t. The coordinates of the center of the material at the initial moment. This refers to the angle of the conveyor belt's running direction. To correct for track jitter caused by belt vibration, the main control unit uses a second-order polynomial fitting:

[0082]

[0083] in, , , The polynomial coefficients are used to fit the trajectory in the X direction; , , The polynomial coefficients for fitting the trajectory in the Y direction; coefficient vector Solve using the least squares method:

[0084]

[0085] Where, vector These are the trajectory fitting coefficients in the X direction; vectors These are the trajectory fitting coefficients in the Y direction. T is the time sample matrix, containing the t values ​​at each time point; X and Y are the sampled material observation vectors in the X and Y directions, respectively. The least squares normal equation solution matrix; the trajectory refresh period is 20ms. When speed signal packet loss or abnormal fluctuations are detected, the system calls the predicted speed from the previous moment and performs linear interpolation compensation to ensure trajectory continuity.

[0086] S220: Determination of adsorption region location: Based on the trajectory calculation results, combined with the material distribution mapping matrix This determines the coverage area of ​​the material along the conveying path. The partitioned vacuum adsorption module under the conveyor belt is divided into several physical zones. Each partition has a physical size of 50mm × 50mm, and the system supports 20–100 partitions. When the material covers a partition area greater than or equal to 30%, the partition is determined to be a core area; if the coverage area is less than 30%, it is determined to be an edge area. The 30% threshold can be configured by software, ranging from 20% to 40%.

[0087] Within each zone, the set of target zones to be activated is calculated by combining the material size parameters (L, W) and the contact pressure distribution C(x, y). The applicable material size range is 30 mm × 30 mm to 200 mm × 200 mm.

[0088] For partitions with edge coverage, negative pressure is allocated using a weighted method:

[0089]

[0090] in, The effective adsorption pressure of the partition, The target pressure for the core area of ​​the material. The target pressure is the pressure at the edge of the material. This is a weighting coefficient, ranging from 0.5 to 0.9. The target negative pressure value is adjustable within the range of -2 kPa to -6 kPa.

[0091] S230: Temporal-spatial adsorption window generation: After determining the target partition to be activated, the conveyor belt speed v, cycle time T, and material coverage time interval are combined. Generate a time-space snapping window for each partition. Defined as follows: ,in, Indicates partition Adsorption window, and These are the start and end times when the partition needs to maintain negative pressure. The target negative pressure intensity value for this zone during the stated time period.

[0092] The final output is a temporal and spatial adsorption window plot: The window graph is stored in the main control unit database in JSON format with a timestamp precision of 1 ms. Each window object contains a partition number, start time, end time, and target negative pressure value. The system caches the most recent 100 window objects and refreshes them every 20 ms.

[0093] For example: when materials cover zones At this time, the window object is: {"Zone":3, "Start":120, "End":280,"Pressure":-3.5}. In the case of multiple materials operating in parallel, the main control unit allows a maximum of 5 materials to exist simultaneously and employs a priority queue scheduling mechanism: the default priority is based on the order in which the materials arrive; when materials arrive simultaneously, the material window with the larger area is scheduled first; other material windows are scheduled sequentially. If window data in a certain partition is lost or abnormal, the system automatically inherits the window configuration from the previous cycle and recalculates it in the next refresh cycle to avoid adsorption interruptions.

[0094] S3: Dynamic Execution of Adsorption Strategy: Based on the adsorption window diagram, dynamically control the opening status of vacuum valves and the negative pressure output intensity of multiple independent negative pressure zones below the conveyor belt to perform adsorption on demand and avoid ineffective pumping; specifically including the following sub-steps:

[0095] S310: Issue adsorption control command: based on the time-space adsorption window diagram generated by S230. The main control unit parses each window object according to a refresh cycle of 20ms and generates control commands.

[0096] The instruction includes: partition number Adsorption start time Timestamp accuracy 1ms; Adsorption off time Timestamp accuracy 1ms; target negative pressure intensity value The pressure range is -2 kPa to -6 kPa, with a minimum adjustment step of 0.1 kPa. Commands are issued via the CAN bus, with each message being 8 bytes long and formatted as follows: Bytes 0-1: Partition number (2 bytes, supports 0-65535); Bytes 2-3: Start timestamp (2 bytes, in ms); Bytes 4-5: End timestamp (2 bytes, in ms); Bytes 6-7: Target negative pressure value (2 bytes, in 0.01 kPa). The main control unit and each partition control module use a synchronous clock mechanism, periodically sending time synchronization frames to ensure a clock deviation of less than ±1 ms. The communication delay does not exceed 5 ms; if it does, the message is discarded and retransmitted.

[0097] S320: Partitioned Vacuum Adsorption Module Execution: After receiving the instruction, the partitioned vacuum adsorption module performs the following operation within the designated interval: Internal actuator control: Actuator: adopts a proportional solenoid valve, with an action response time of less than 20 ms; Vacuum pump: speed adjustment range of 1000–3000 rpm, adjustment resolution of 50 rpm, response delay <5 ms; Sensor sampling frequency: 50 Hz.

[0098] The closed-loop control uses a PID control algorithm, and the deviation is defined as follows: The control law is:

[0099]

[0100] in, , , The proportional, integral, and derivative parameters were set to these values ​​after experimental optimization. It is 2.0. It is 0.5. It is 0.1. When When the pressure is low, the PID controller initiates compensation, restoring the negative pressure to the target range by adjusting the vacuum pump speed or the solenoid valve opening. The adjustment cycle is 10 ms.

[0101] S330: Dynamic timing switching execution: During the conveying process, the main control unit executes according to the window diagram. The timing requirement is to dynamically switch the partition status: when a material enters a partition: immediately trigger the window. When a zone valve is open, or when material leaves a zone, the zone valve is closed to prevent ineffective air extraction. For uncovered zones, the valve remains closed to minimize energy consumption. To avoid pressure surges, the valve operates in a linear, gradual change mode: it gradually changes from 0% opening to the target opening within 10 ms, or reverses to close.

[0102] The system supports a maximum of 5 materials simultaneously. When two materials request the same partition at the same time, the material that arrives first is prioritized. If the arrival times are the same, the material with the larger area is prioritized. Delayed materials are re-queued in the next cycle to avoid loss. If no material covers a partition for 500 ms, the vacuum pump branch for that partition is automatically shut down. The partition is pre-vacuumed 50 ms before material enters to ensure response speed. In case of lost instructions or message errors, the partition automatically inherits the parameters from the previous cycle; it receives new instructions again in the next refresh cycle to avoid adsorption interruption. For example: when defined in the window diagram... At t=120 ms, the main control unit sends a message. , The zone valves open gradually over 10 ms, while the PID controller maintains a pressure of -3.5 ± 0.05 kPa. The valves then gradually close at t = 280 ms.

[0103] S4: Flexible Adaptive Adhesion Control: During the material adsorption process, based on the flexible structure of the belt and the real-time contact state of the material, the deformation of the adsorption surface is dynamically adjusted to improve the adhesion effect and reduce slippage; specifically including the following sub-steps:

[0104] S410: Adhesion Status Monitoring: During the process of material being adsorbed by the partitioned vacuum adsorption module, the adhesion status between the material and the conveyor belt surface is monitored in real time. Two high-speed cameras are arranged above the conveyor belt, with a frame rate of 60Hz, covering the entire width of the belt. The cameras complete the mapping between pixel coordinates and physical height through a calibration plate. The conversion formula is as follows:

[0105]

[0106] in, For physical height difference, For pixel difference, The calibration coefficient (obtained experimentally as 0.067) is used. When a physical height difference greater than 0.2 mm and a pixel difference greater than 3 px are detected, a gap is determined to exist. The contact pressure distribution is collected using a 10×10 pressure sensor array in the S120. The sampling frequency is 50Hz. If the local contact pressure is below the threshold... If the visual detection and pressure array detection results are inconsistent, the system will prioritize the pressure array results, with visual detection serving as redundancy.

[0107] S420: Flexible Adaptor Structure Drive: When uneven adhesion is detected, the main control unit sends an adjustment command to the flexible adapter belt structure. The flexible adapter belt structure includes: an elastic base layer: 2 mm thick, made of TPU, with a Young's modulus of 15 MPa; micro-protrusion airbag units: 5 mm in diameter and 1 mm in height, arranged in an array at 10 mm × 10 mm intervals, covering the entire adsorption area; each airbag unit has a built-in micro-electric control valve with a unit response time of less than 10 ms. The air supply system is provided by a centralized air source with a pressure range of 0.1-0.5 bar and an adjustment accuracy of 0.01 bar. The system can simultaneously drive up to 20 airbag units for compensation adjustment. During adjustment, the main control unit selects the target airbag unit based on the coordinates of the abnormal area and controls its inflation / deflation within a range of ±0.5 mm. The adjustment resolution is 0.05 mm, and the single response time is less than 20 ms.

[0108] S430: Automatic Surface Adjustment and Stable Adsorption: After the airbag unit undergoes slight deformation, the surface of the conveyor belt automatically adjusts to a conforming shape that matches the bottom surface of the material. During this process, the pressure array continuously collects real-time contact pressure distribution data. And calculate the uniformity index:

[0109]

[0110] in, The standard deviation of contact pressure, This represents the average contact pressure. When all local contact pressures... ,and Upon reaching the set point, the system determines that the fit adjustment is complete. If the above conditions are not met within 200 ms, the system performs a secondary adjustment: increasing the adjustment range by 20% and repeating the inflation / deflation process. If the secondary adjustment still fails to meet the standard, an alarm is triggered and the abnormal zone is marked.

[0111] After the material leaves the zone, the airbag unit automatically deflates to zero pressure within 50ms, with idle power consumption below 0.1W. The system periodically (every 1000 cycles) performs airbag pressure self-checks to ensure a leakage rate of less than 2%. If an airbag unit fails to achieve the expected deformation (deviation greater than 0.1 mm) after three consecutive adjustments, the system automatically disables that unit, and surrounding units compensate. If pressure sensor data is lost, visual inspection results are redundantly used for judgment. Each airbag unit is designed for a lifespan of greater than 10 years. 6 During the initial inflation / deflation cycle, the main control unit issues a maintenance prompt when the material exceeds its lifespan. For example, if visual inspection detects that the material edge has risen 0.4 mm (corresponding coordinates (60mm, 120mm)), the main control unit sends an inflation command to the two airbag units in the corresponding area, inflating them by 0.4 mm. The pressure array monitors that the contact pressure has recovered to 150 Pa, and the uniformity index... The result is 7%, indicating that the fit adjustment is complete.

[0112] S5: Constructing an Adsorption Memory Model: Based on historical material adsorption feedback data, energy consumption data, and offset rate, construct a mapping model between material types and adsorption parameters to form an adsorption parameter memory library; this includes the following sub-steps:

[0113] S510: Adsorption Operation Data Acquisition: During the material conveying process, the main control unit collects the following operational data in real time: Adsorption success rate data: The visual recognition module monitors the material's offset and detachment during the conveying process, and the calculation formula is:

[0114]

[0115] in, To ensure a stable quantity of materials being transported, This represents the total quantity of material conveyed. Energy consumption data: Simultaneously sampled by a voltage sensor (range 0–240 V, accuracy ±0.5%FS) and a current sensor (range 0–10A, accuracy ±0.5%FS), with a sampling frequency of 1kHz. The energy consumption calculation formula is as follows:

[0116]

[0117] Where U(t) is voltage, I(t) is current, and the unit of E is... .

[0118] Material offset rate data: The visual recognition module detects the offset between the material's center point and the desired trajectory. The average offset rate is calculated using the following formula:

[0119]

[0120] in, Let be the offset distance of the i-th material. This represents the total number of statistical samples, i.e., the number of observed material location points. All data is sampled at a frequency of 50 Hz, and data integrity is ensured through CRC checksum verification. In case of packet loss, linear interpolation compensation is used.

[0121] S520: Data Correlation Analysis: Mapping the collected data with the material distribution matrix output by S130 Perform correlation analysis. Input features include: material size parameters (L, W); contact pressure distribution C(x, y); adsorption window parameters. Output metrics include: adsorption success rate. Energy consumption E; Average offset rate .

[0122] Establish association function:

[0123] Processing flow: Normalization:

[0124] Outlier handling: If a data point deviates from the mean by more than [a certain value]... If an outlier is found, it is considered an outlier and removed. Modeling method: A hybrid modeling approach using multivariate linear regression and random forest regression is employed. The random forest parameters are 100 trees and a maximum depth of 5. Weight allocation: The weights for linear regression and random forest are 0.4 and 0.6, respectively. Calculation cycle: Updated every 1 second, with each calculation based on the most recent 50 data sets.

[0125] S530: Mapping Relationships and Model Construction: Based on the correlation analysis results, the main control unit generates an adsorption memory model, which includes: Input layer: material size (L,W), pressure distribution characteristics C(x,y), target negative pressure intensity. Hidden layers: 3 fully connected neural network layers, with sizes of 64, 32, and 16 neurons respectively, activation functions... It is combined with a random forest model to form a hybrid learner; output layer: predicts the success rate of adsorption. Energy consumption Average offset .

[0126] Model training parameters: learning rate 0.001, batch size 32, number of iterations 1000; training set to validation set ratio 80%:20%; validation accuracy requirement greater than 95%; incremental training is triggered after every 500 delivery tasks.

[0127] Storage and Lifespan: Models are stored in TensorFlow Lite format in the master control unit database; the 10 most recent versions are retained, managed using a circular FIFO strategy; model lifespan is greater than 10. 6 The system has a storage redundancy rate of 200% for each call. If the model's deviation exceeds a threshold in three consecutive predictions (e.g., the success rate prediction error is greater than 10% or the energy consumption prediction error is greater than 15%), the system automatically rolls back to the previous stable version; if the database storage is insufficient, the oldest version is deleted and a maintenance prompt is displayed. For example, for a material with a size of 80×100 mm, a peak pressure distribution of 120 Pa, a target negative pressure of -3.5 kPa, the model predicts an adsorption success rate of 98%, energy consumption of 0.12 Wh, and an average offset rate of 0.4 mm, all of which meet the process requirements.

[0128] S6: Machine Learning Feedback Tuning: By deploying reinforcement learning or deep neural network models in the main control unit, and comparing and analyzing the current feedback data with the memory model, the adsorption window diagram and partition control parameters are dynamically updated to achieve continuous adaptive optimization of the adsorption behavior of multiple types of materials. This includes the following sub-steps:

[0129] S610: Operational Feedback Data Acquisition: During the adsorption and conveying process, the main control unit collects the following feedback data in real time: Actual adsorption success rate: The material offset and detachment are detected by the vision module, and the calculation formula is the same as that of S510.

[0130] Actual energy consumption Data is collected by voltage and current sensors, and the calculation formula is as follows:

[0131]

[0132] Wherein, the integration interval Aligned with the adsorption window time, ensuring a one-to-one correspondence between predicted and actual values. Sensor delay is controlled within 5ms.

[0133] Actual average offset rate: The offset between the material center point and the desired trajectory is calculated by the vision module, using the window start time as the alignment reference. All data is sampled at a frequency of 50Hz and CRC check is performed. If data packet loss or anomalies occur, linear interpolation compensation is used.

[0134] S620: Error Analysis and Optimization Strategy Generation: The main control unit generates predicted values. Compared with actual value Compare and calculate the error:

[0135]

[0136] This represents the prediction error of the k-th indicator; These are the actual observed values ​​(true values). These are model predictions; statistical indicators include: mean error.

[0137]

[0138] is the average error of the k-th index; N is the number of observation samples (e.g., 50 frames or 100 measurements). It is the prediction error of index k in the i-th observation.

[0139] Mean Square Error (MSE):

[0140]

[0141] is the mean squared error of the k-th indicator, and N is the number of samples. Threshold setting: Greater than 5% or A value greater than 0.05 triggers tuning. The threshold is derived from the statistical average of laboratory benchmark tests and 1000 runs.

[0142] Adjustment strategy: Weight fine-tuning: Linear regression weights are in the range of 0.3–0.7, with a step size of 0.0 each time; Adaptive learning rate: If the error decreases slowly, reduce to 0.002; if the error oscillates, reduce to 0.0005; Feature importance update: Feature importance is calculated based on the output of the random forest, and features with less than 1% importance are removed. If the error exceeds 5% after removal, the feature is restored.

[0143] S630: Online parameter update and model retraining: When the tuning condition is triggered, the main control unit performs online retraining: Data source: the most recent 1000 sets of running data; Split ratio: 80% for training set and 20% for validation set; Optimization: Adam; Number of iterations: maximum 500; Early stopping condition: no improvement in validation accuracy for 10 consecutive rounds; Computing platform: The main control unit has an embedded ARM Cortex-A72 CPU, and the GPU module is called for acceleration when necessary. The training time for a single session is controlled within 1 second to ensure real-time performance.

[0144] The formula for updating new data based on the old model parameters is:

[0145]

[0146] in, For the old model parameter vector, For learning rate, This is the updated model parameter vector; It is a loss function For parameters The gradient is calculated; after the model is updated, it is immediately written to the database, and the 10 most recent versions are retained, using a FIFO strategy. If the accuracy of the new model's validation set is less than 95%, it is automatically rolled back to the old version.

[0147] S640: Application of tuning results and fault tolerance mechanism: The updated model is directly used in the next round of delivery tasks, outputting new predicted values. ;

[0148] Fault Tolerance Mechanism: Prediction Deviation Monitoring: If the prediction deviation exceeds the threshold for three consecutive tasks (success rate error greater than 10% or energy consumption error greater than 15%), the system immediately rolls back to the old model. Caching Mechanism: The system caches the most recent 100 optimization results, stored in JSON format, including timestamps, optimization strategies, and error metrics for easy traceability. Safety Fallback Strategy: When the model is unavailable, the system switches to default parameters (adsorption pressure – 3.5 kPa, window length 200 ms), which can be manually configured in the maintenance interface. Manual Intervention Conditions: If the number of consecutive rollbacks exceeds 5, the system issues an alarm, prompting manual inspection. Example: In a delivery task, the original model predicted a success rate of 95%, while the actual success rate was 88%, with an error of 7%. The main control unit triggered optimization, adjusting the weights to linear regression 0.35 and random forest 0.65, and reducing the learning rate to 0.0008. After retraining, the error between the new model's prediction and the actual success rate decreased to 2%, meeting the requirements.

[0149] Example 2: Figure 2 As shown, this embodiment provides a machine learning-based intelligent control system for negative pressure adsorption and delivery, including:

[0150] The conveyor belt assembly is used to carry and transport photovoltaic silicon wafers. The surface of the conveyor belt is embedded with a flexible adapter structure and a resistive pressure sensor array to collect the contact pressure distribution between the material and the belt.

[0151] A visual recognition module, located above the conveyor belt, is used to acquire two-dimensional images and depth information of the photovoltaic silicon wafers, and generate a material distribution mapping matrix by combining the output of the pressure sensor array.

[0152] The partitioned vacuum adsorption module is arranged below the conveyor belt, which divides the conveying area into multiple independently controllable negative pressure zones. Each zone is equipped with a vacuum valve and a vacuum pump, which can be opened and adjusted as needed.

[0153] The main control unit is electrically connected to the visual recognition module, the pressure sensor array, and the partitioned vacuum adsorption module.

[0154] This conveying system is used to: calculate the material trajectory and generate a time-space adsorption window diagram based on the material distribution mapping matrix and the conveyor belt speed;

[0155] Based on the aforementioned temporal and spatial adsorption window diagram, adsorption control commands are sent to each zone vacuum adsorption module to dynamically switch the opening state and negative pressure intensity of the zone vacuum valves, thereby achieving adsorption on demand.

[0156] During the material adsorption process, the flexible adaptor structure is driven to undergo micro-deformation based on pressure sensor and visual detection feedback, so as to improve the material adhesion stability.

[0157] Collect data on adsorption success rate, energy consumption, and offset rate during material transportation, establish an adsorption memory model, and realize the mapping between material type and adsorption parameters;

[0158] By deploying a machine learning model and comparing it with the adsorption memory model based on operational feedback data, the time-space adsorption window diagram and partition control parameters are dynamically optimized to achieve continuous adaptive optimization for photovoltaic silicon wafers of different sizes and shapes.

[0159] like Figure 3 The diagram shows an existing negative pressure belt conveyor assembly, including a work frame and a belt, preferably improved by adding corresponding modules.

[0160] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0161] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0162] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0163] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0164] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0165] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0166] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0167] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0168] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0169] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based intelligent control method for negative pressure adsorption and transport, characterized in that, Includes the following steps: S1. Obtain the size, shape and position distribution information of the material to be conveyed on the conveyor belt, wherein the information is jointly obtained by the visual recognition module and the belt surface pressure sensor to construct a material spatial distribution map and a contact state map. S2. Based on the material spatial distribution map and contact state map, combined with the conveyor belt running speed, cycle time information and material form, a corresponding temporal and spatial adsorption window map is generated to indicate the activation state and required adsorption intensity of each negative pressure zone at a specific time. S3. Based on the adsorption window diagram, dynamically control the opening status of vacuum valves and the negative pressure output intensity of multiple independent negative pressure zones below the conveyor belt to perform adsorption on demand. S4. During the material adsorption process, based on the flexible structure of the belt and the real-time contact state of the material, the deformation of the adsorption surface is dynamically adjusted to improve the adhesion effect and reduce sliding offset. S5. Based on historical material adsorption feedback data, energy consumption data and offset rate, construct a mapping model between material type and adsorption parameters to form an adsorption parameter memory library. S6. By using reinforcement learning or deep neural network models deployed in the main control unit, the current feedback data is compared and analyzed with the memory model to dynamically update the adsorption window diagram and partition control parameters, thereby achieving continuous adaptive optimization of the adsorption behavior of multiple types of materials. S3 specifically includes: sending a control command to the partitioned vacuum adsorption module below the conveyor belt according to the time-space adsorption window diagram; after receiving the control command, the partitioned vacuum adsorption module opens the vacuum valve of the corresponding partition as needed and sets the negative pressure intensity; during the conveying process, according to the timing requirements of the adsorption window diagram, the opening state of each negative pressure partition is dynamically switched to achieve adsorption on demand and reduce ineffective pumping. S4 specifically includes: during the adsorption process, real-time monitoring of the adhesion state between the material and the surface of the conveyor belt; when uneven adhesion is detected locally, driving the elastic layer or micro-convex structure of the flexible adapter belt structure to deform; through the deformation of the elastic layer or micro-convex structure, the surface of the conveyor belt is automatically adjusted to an adhesion form that matches the bottom surface of the material, thereby enhancing adsorption stability.

2. The intelligent control method for negative pressure adsorption and transport based on machine learning according to claim 1, characterized in that, S1 specifically includes: Two-dimensional image information of the material is obtained by a vision recognition module installed above the conveyor belt, and the size, boundary and shape characteristics of the material are identified. Information on the contact pressure distribution between the material and the belt is collected by pressure sensors installed on the surface of the conveyor belt. The two-dimensional image information is fused with the contact pressure distribution information to generate material distribution mapping data that includes material location, size, and contact state.

3. The intelligent control method for negative pressure adsorption and transport based on machine learning according to claim 2, characterized in that, S2 specifically includes: Based on the material distribution mapping data, combined with the real-time operating speed and cycle time parameters of the conveyor belt, the trajectory of the material's position change during the conveying process is calculated. Based on the position change trajectory, the required adsorption area position of the material on the conveying path is determined; By combining the size and contact state of different materials at the adsorption region location, a corresponding time-space adsorption window diagram is generated to indicate the activation state and negative pressure intensity of each negative pressure zone at a specific moment.

4. The intelligent control method for negative pressure adsorption and transport based on machine learning according to claim 1, characterized in that, S5 specifically includes: Collect data on the adsorption success rate, energy consumption, and material offset rate of materials during the transportation process; The adsorption success rate data, energy consumption data, and material offset rate data are correlated with the material distribution mapping data; Based on the correlation analysis results, a mapping relationship between material type and adsorption parameters is generated, and an adsorption memory model is constructed.

5. The intelligent control method for negative pressure adsorption and transport based on machine learning according to claim 1, characterized in that, S6 specifically includes: The current adsorption operation data of the material is compared with the adsorption memory model to identify parameter deviations; Based on the identified parameter deviations, new negative pressure zoning control parameters are generated through reinforcement learning algorithms or neural network models deployed in the main control unit. Update the spatiotemporal adsorption window diagram to form a new adsorption strategy; During the next material transport, the partitioned vacuum adsorption module is controlled according to the updated adsorption strategy to achieve continuous adaptive optimization.

6. A machine learning-based intelligent control system for negative pressure adsorption and conveying, based on the machine learning-based intelligent control method for negative pressure adsorption and conveying according to any one of claims 1-5, characterized in that, include: The conveyor belt assembly is used to carry and transport photovoltaic silicon wafers. The surface of the conveyor belt is embedded with a flexible adapter structure and a resistive pressure sensor array to collect the contact pressure distribution between the material and the belt. The visual recognition module, located above the conveyor belt assembly, is used to acquire two-dimensional images and depth information of the photovoltaic silicon wafers, and, combined with the output of the pressure sensor array, generate a material distribution mapping matrix. The partitioned vacuum adsorption module is arranged below the conveyor belt, dividing the conveying area into multiple independently controllable negative pressure zones. Each zone is equipped with a vacuum valve and a vacuum pump, which can be opened and adjusted as needed. The main control unit is electrically connected to the visual recognition module, the pressure sensor array, and the partitioned vacuum adsorption module.

Images