Negative pressure adsorption conveying intelligent control method and system based on machine learning

Abstract

The invention provides a negative pressure adsorption conveying intelligent control method and system based on machine learning. The system comprises a conveying belt assembly, a partition vacuum adsorption module, a visual identification module, a pressure sensor array, a flexible adaptation structure and a main control unit. Obtaining the size, position and distribution information of the photovoltaic silicon wafer through visual identification and pressure array fusion, and constructing a material distribution mapping matrix; and the main control unit generates a time-space adsorption window graph based on track prediction, and issues a dynamic control instruction to the partition vacuum adsorption module to realize on-demand opening and negative pressure intensity adjustment of different partitions. And the flexible adaptive structure realizes fitting compensation on the bottom surface of the silicon wafer through micro-deformation adjustment of the air bag array. The energy consumption is remarkably reduced while stable conveying of the silicon wafers is ensured, the flexibility, intelligence and robustness of the system are improved, and the system is suitable for being popularized and applied to high-speed conveying and sorting links of the photovoltaic silicon wafers.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic silicon wafer conveying, in particular to a negative pressure adsorption conveying intelligent control method and system based on machine learning. BACKGROUND

[0002] As a key basic material in the manufacture of photovoltaic modules, photovoltaic silicon wafers usually include cutting, cleaning, detection and sorting processes in the production process. Between the above processes, the silicon wafer needs to be transferred by a conveying device at high speed and stability. In the prior art, a negative pressure belt conveying system has been proposed to transport the silicon wafer to avoid slipping and deviation of the silicon wafer during conveying.

[0003] However, the existing negative pressure belt conveying system adopts a fixed adsorption zone design, lacks independent control and dynamic adjustment of the partition, resulting in low adsorption efficiency and high energy consumption in mixed flow production of multiple specifications. Secondly, the existing negative pressure belt conveying system lacks a feedback optimization mechanism based on adsorption success rate, energy consumption and deviation rate, so we propose a negative pressure adsorption conveying intelligent control method and system based on machine learning. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a negative pressure adsorption conveying intelligent control method and system based on machine learning to solve the technical problems mentioned in the background.

[0005] To achieve the above purpose, the present application provides the following technical scheme: A negative pressure adsorption conveying intelligent control method based on machine learning, comprising the following steps: S1, obtaining the size, shape and position distribution information of the material to be conveyed on the conveying belt, wherein the information is obtained by a visual recognition module and a belt surface pressure sensor, and a material space distribution map and a contact state map are constructed; S2, based on the material space distribution map and the contact state map, combining the conveying belt running speed, the beat information and the material form, generating a corresponding time-space adsorption window map for indicating the activated state and the required adsorption strength of each negative pressure partition at a specific time; S3, according to the adsorption window map, dynamically controlling the opening state of the vacuum valve and the negative pressure output intensity of the multiple independent negative pressure partitions under the conveying belt to perform on-demand adsorption; S4, during the material adsorption process, based on the belt flexible structure and the real-time contact state of the material, dynamically adjusting the adsorption surface deformation to improve the fitting effect and reduce the sliding deviation; S5, based on the historical material adsorption feedback data, the energy consumption data and the deviation rate, constructing a mapping model between the material type and the adsorption parameters to form an adsorption parameter memory bank; S6, by deploying a reinforcement learning or deep neural network model in the master unit, using current feedback data and memory model for comparative analysis, dynamically updating the adsorption window map and partition control parameters, realizing continuous self-adaptive optimization of the adsorption behavior of multiple types of materials.

[0006] S1 specifically includes: The two-dimensional image information of the material is obtained by the visual recognition module arranged above the conveying belt, and the size, boundary and shape characteristics of the material are recognized; The contact pressure distribution information between the material and the belt is collected by the pressure sensor arranged on the surface of the conveying belt; The two-dimensional image information and the contact pressure distribution information are fused to generate material distribution mapping data containing material position, size and contact state.

[0007] S2 specifically includes: According to the material distribution mapping data, the real-time running speed and running beat parameters of the conveying belt are combined to calculate the position change trajectory of the material in the conveying process; Based on the position change trajectory, the adsorption area position required by the material on the conveying path is determined; According to the size and contact state of different materials, the corresponding time-space adsorption window map is generated at the adsorption area position, which is used to indicate the activation state and negative pressure intensity of each negative pressure partition at a specific time.

[0008] S3 specifically includes: According to the time-space adsorption window map, control instructions are sent to the partition vacuum adsorption module below the conveying belt; After receiving the control instructions, the partition vacuum adsorption module opens the vacuum valve of the corresponding partition as needed and sets the negative pressure intensity; During the conveying process, the opening state of each negative pressure partition is dynamically switched according to the time sequence requirement of the adsorption window map, realizing on-demand adsorption and reducing invalid air suction.

[0009] S4 specifically includes: During the adsorption of the material, the fit state of the material and the surface of the conveying belt is monitored in real time; When local uneven fit is detected, the elastic layer or micro-convex structure of the flexible adaptive belt structure is deformed; Through the deformation of the elastic layer or micro-convex structure, the surface of the conveying belt is automatically adjusted to match the fit form of the material bottom surface, enhancing the adsorption stability.

[0010] S5 specifically includes: The adsorption success rate data, energy consumption data and material deviation rate data of the material in the conveying process are collected; Correlate the adsorption success rate data, energy consumption data and material deviation rate data with the material distribution mapping data for analysis; Based on the correlation analysis result, a mapping relationship between the material type and the adsorption parameter is generated, and an adsorption memory model is constructed.

[0011] S6 specifically includes: Compare the adsorption operation data of the current material with the adsorption memory model to identify parameter deviation; Based on the identified parameter deviation, a new negative pressure partition control parameter is generated through a reinforcement learning algorithm or a neural network model deployed in the main control unit; Update the time-space adsorption window graph to form a new adsorption strategy; In the next material conveying, the partition vacuum adsorption module is controlled according to the updated adsorption strategy to realize continuous self-adaptive optimization.

[0012] A negative pressure adsorption conveying intelligent control system based on machine learning, comprising: A conveying belt assembly for carrying and conveying photovoltaic silicon wafers, the conveying belt surface is embedded with a flexible adaptive structure and an array of resistive pressure sensors for collecting the contact pressure distribution between the material and the belt; A visual recognition module arranged above the conveying belt assembly for acquiring two-dimensional images and depth information of the photovoltaic silicon wafers, and generating a material distribution mapping matrix in combination with the output of the pressure sensor array; A partition vacuum adsorption module arranged below the conveying belt, dividing the conveying area into a plurality of independently controllable negative pressure partitions, each partition being provided with a vacuum valve and a vacuum pump, and being capable of opening and adjusting the negative pressure intensity as needed; A main control unit electrically connected with the visual recognition module, the pressure sensor array and the partition vacuum adsorption module.

[0013] The beneficial effects of the present application are: Through the fusion of the visual recognition module and the pressure sensor array, the size, boundary shape and contact pressure distribution of the photovoltaic silicon wafers with the belt can be acquired simultaneously, and a material distribution mapping matrix is generated through extended Kalman filtering. Compared with the traditional single visual recognition method, this method greatly improves the recognition accuracy of different specifications and thicknesses of silicon wafers, and avoids recognition errors caused by factors such as light and reflection.

[0014] By constructing a time-space adsorption window graph, the main control unit can accurately indicate the opening time, closing time and target negative pressure intensity of each negative pressure partition, and realize dynamic scheduling of the partition vacuum valve. This mechanism avoids the energy waste of global continuous vacuum pumping, reduces energy consumption by about 30%-50%, and at the same time ensures the stability of the material in the high-speed conveying process.

[0015] This invention incorporates a flexible, adaptive structure on the conveyor belt surface. By using a micro-airbag array or elastic layer to compensate for micro-deformation, the belt surface can be proactively adjusted based on real-time detection of uneven fit, ensuring uniform fit between the bottom surface of the silicon wafer and the belt surface. This measure effectively prevents edge warping and localized stress concentration caused by belt vibration or surface unevenness, significantly reducing the risk of silicon wafer breakage.

[0016] The system collects operational data such as adsorption success rate, energy consumption, and drift rate to map material types to adsorption parameters, creating a memory library of adsorption parameters. When producing mixed-flow products with multiple specifications, the system can directly retrieve the optimal parameter combination from history, shortening the debugging cycle and improving the first-time adsorption success rate.

[0017] By leveraging a reinforcement learning or neural network model deployed in the main control unit, the system analyzes the difference between predicted values ​​and actual operating data, and performs online retraining and parameter updates when necessary. This mechanism enables the system to maintain high levels of adsorption accuracy and stability even during long-term operation and under complex operating conditions, enabling adaptive optimization of the wafer transport process. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of an intelligent control method for negative pressure adsorption delivery based on machine learning in the present invention; Figure 2 This is a schematic diagram of the framework of a negative pressure adsorption and delivery intelligent control system based on machine learning in the present invention; Figure 3 The figure is a schematic diagram of the structure of an existing negative pressure belt conveyor assembly. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Example 1: Figure 1 As shown, this embodiment provides a method for intelligent control of negative pressure adsorption delivery based on machine learning, comprising the following steps: S1: Multimodal Material Recognition: Acquires the size, shape, and position distribution information of the material (photovoltaic silicon wafer) to be conveyed on the conveyor belt. This information is acquired jointly by the visual recognition module and the belt surface pressure sensor to construct a material spatial distribution map and contact state map. S2: Generate a time-space adsorption window diagram: Based on the material distribution diagram, combined with the conveyor belt running speed, beat information and material form, a corresponding "time-space adsorption window diagram" is generated to indicate the activation state and required adsorption intensity of each negative pressure zone at a specific time; S3: Dynamic execution of adsorption strategy: Based on the adsorption window diagram, the vacuum valve opening status and negative pressure output intensity of multiple independent negative pressure zones below the conveyor belt are dynamically controlled to perform adsorption on demand and avoid ineffective extraction; S4: Flexible Adaptive Fit Control: During the material adsorption process, the belt's flexible structure and the material's real-time contact state dynamically adjust the adsorption surface deformation to improve the fit and reduce sliding offset. S5: Constructing an adsorption memory model: Based on historical material adsorption feedback data, energy consumption data, and offset rate, a mapping model between material type and adsorption parameters is constructed to form an adsorption parameter memory library; S6: Machine learning feedback tuning: Through the reinforcement learning or deep neural network model deployed in the main control unit, the current feedback data is compared and analyzed with the memory model, and the adsorption window map and partition control parameters are dynamically updated to achieve continuous adaptive optimization of the adsorption behavior of multiple types of materials.

[0021] S1: Multimodal Material Recognition: Acquire the size, shape, and position distribution information of the material to be conveyed on the conveyor belt. This information is obtained jointly by the visual recognition module and the belt surface pressure sensor to construct a material spatial distribution map and contact state map. This specifically includes the following sub-steps: S110: Acquisition of two-dimensional image information: The two-dimensional image and depth information of the material are acquired through a structured light three-dimensional camera installed above the conveyor belt. The camera is installed vertically 300 mm above the conveyor belt, with a field of view covering an area of ​​250 mm × 250 mm, a resolution of 1920 × 1080, a frame rate of 30 Hz, and a depth measurement accuracy of ±0.1 mm. To ensure imaging quality, the camera is equipped with a ring LED fill light with a color temperature of 5000 K and an illumination maintained at 500-800 lx to avoid recognition errors caused by external light fluctuations, shadows, and reflections. After installation, the camera is calibrated using a checkerboard calibration plate to establish a conversion relationship between pixel coordinates and physical coordinates. The calibration uses a projection matrix , the formula is: in,( ) is the image pixel coordinate, (X, Y) is the physical coordinate, is a 3x3 projection matrix. The calibration error is controlled within 0.05 mm. In the image processing process, the Canny edge detection operator is used to extract the material boundary curve; then the minimum circumscribed rectangle fitting algorithm is used to calculate the boundary rectangular frame of the material to obtain the length L, width W and shape feature parameters of the material.

[0022] wherein the calculation formula of the length and width of the material is: wherein, and are the maximum and minimum coordinate values of the material boundary points in the image coordinate system, respectively.

[0023] S120: Contact pressure distribution information acquisition: 10x10 arrays of resistive pressure sensors are uniformly arranged on the surface of the conveying belt in the transverse and longitudinal directions, with a total coverage area of 200 mm x 200 mm and a center-to-center spacing of 20 mm for each sensor. The sensor range is 0-10 kPa, the detection resolution is ±5 Pa, and the sampling frequency is 50 Hz. The sensor is embedded 0.5 mm below the surface layer of the belt through a flexible substrate to ensure accurate measurement without affecting the flexibility and service life of the belt. A 0.2 mm thick polyurethane protective film is applied to the surface of the sensor to improve wear resistance, dust resistance, and water resistance, meeting the IP54 protection level. All sensors are calibrated to zero and full scale after installation, with a calibration error of less than ±1% FS. The sensor signal is collected at a frequency of 1 kHz through a 16-bit A / D conversion module and input into the main control unit. The data is processed through a low-pass filter before entering the main control unit, with a cutoff frequency of 10 Hz, to eliminate high-frequency noise caused by the running vibration of the conveying belt 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 data missing.

[0024] S130, data fusion and material distribution mapping generation: fuse the material profile coordinates obtained in S110 with the contact pressure distribution data obtained in S120, and use the extended Kalman filter algorithm (EKF) for dynamic fusion.

[0025] In EKF: the state vector is defined as wherein represents the center position of the material, represents the velocity components of the material in the transverse and longitudinal directions; the state transition equation is: wherein, is the sampling period; represents the coordinate position of the target at the current time k; , represents the predicted coordinate position of the target at the next moment k+1; the observation vector is defined as the coordinate of the contact point between the material center coordinate obtained by visual recognition and the pressure array; the observation equation is: in =(\textbf{"a"," ...

[0026] The fusion output is the material distribution mapping matrix: in, Indicates the spatial distribution of materials on the conveyor belt, Indicates the position coordinates of the material in the physical coordinate system. Indicates the contact pressure intensity at the corresponding position. The matrix resolution is 5mm×5mm cells, and the coordinate origin is set at the left front end reference point of the conveyor belt. The matrix is ​​stored in the main control unit memory in the form of a two-dimensional array and is refreshed every 20ms. For example: in a certain sampling, the contact pressure at the coordinate point (50mm, 100mm) is 120Pa, then the corresponding matrix cell is .

[0027] S2: Generate a time-space adsorption window diagram: Based on the material distribution diagram, combined with the conveyor belt running speed, beat information and material form, 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, the following sub-steps are included: S210: Material location trajectory calculation: based on the material distribution mapping matrix generated in S130 , combined with the conveyor belt running speed parameter v and the running beat parameter T, calculate the position change trajectory of the material during the conveying process.

[0028] The belt speed is obtained by a photoelectric encoder installed on the driving wheel shaft with a resolution of 0.01m / s. In the case of a servo motor drive, the speed information can be fed back by the motor controller as input. If acceleration disturbance needs to be considered, a MEMS accelerometer is installed under the belt to collect ±2 m / s 2 This method is applicable to the working range of conveying speed between 0.1 and 2.0 m / s.

[0029] The center position coordinate of the material at time t is calculated by the following formula: wherein, , represents the predicted position (two-dimensional coordinates) of the material at time t; is the center coordinate of the material at the initial time, is the running direction angle of the conveying belt. To correct the trajectory jitter caused by belt vibration, the host control unit adopts a second-order polynomial fitting: wherein, , , is the polynomial coefficient of the fitted X direction trajectory; , , is the polynomial coefficient of the fitted Y direction trajectory; the coefficient vector is solved by the least square method: wherein, the vector is the X direction trajectory fitting coefficient; the vector is the Y direction trajectory fitting coefficient. T is the time sample matrix, containing the t value at each time; X, Y are the observation value vectors of the material in the X, Y directions obtained by sampling; : the normal equation solution matrix of the least square method; the trajectory refresh period is 20 ms. When the speed signal packet loss or abnormal fluctuation is detected, the system calls the predicted speed at the previous time and performs linear interpolation compensation to ensure the continuity of the trajectory.

[0030] S220: Adsorption area position determination: based on the trajectory calculation result, combined with the material distribution mapping matrix , the coverage area range of the material on the conveying path is determined. The partitioned vacuum adsorption module below the conveying belt is divided into a plurality of physical partitions , each physical partition has a physical size of 50 mm × 50 mm, and the system supports 20-100 partitions. When the material covers the partition area greater than or equal to 30%, the partition is determined as the core area; if the coverage area is less than 30%, it is determined as the edge area. The threshold value 30% can be configured by software, ranging from 20-40%.

[0031] In each partition, combined with the material size parameters (L, W) and the contact pressure distribution C(x, y), the target partition set that needs to be activated is calculated. The material size is applicable to the range of 30 mm × 30 mm to 200 mm × 200 mm.

[0032] For the edge covered partition, a weighted method is used to distribute the negative pressure: wherein, is the effective adsorption pressure of the partition, is the target pressure of the core region of the material, is the target pressure of the edge region of the material, is the weight coefficient, ranging from 0.5 to 0.9. The target negative pressure value is adjustable within the range of -2 kPa to -6 kPa.

[0033] S230: Time-space adsorption window map generation: After determining the target partition to be activated, combine the conveying belt speed v, the running cycle T, and the material coverage time interval to generate a time-space adsorption window for each partition. The definition is as follows: wherein, represents the adsorption window of the partition , and and are the start time and end time, respectively, at which the partition needs to maintain negative pressure, is the target negative pressure value of the partition within the time period.

[0034] The final output is a time-space adsorption window map: The window map is stored in the master control unit database in JSON format, with a timestamp accuracy of 1 ms. Each window object contains the partition number, start time, end time, and target negative pressure value. The system caches the last 100 window objects and refreshes every 20 ms.

[0035] For example, when the material covers the partition , the window object is: {"Zone":3, "Start":120, "End":280,"Pressure":-3.5}. In the case of multiple materials in parallel, the master control unit allows a maximum of 5 materials to exist simultaneously and uses a priority queue scheduling mechanism: the default priority is based on the order of material arrival; when arriving simultaneously, the window of the material with larger area is prioritized; other material windows are delayed. If the window data of a certain partition is lost or abnormal, the system automatically inherits the window configuration of the previous period and recalculates in the next refresh period to avoid adsorption interruption.

[0036] S3: Dynamic execution of adsorption strategy: According to the adsorption window map, dynamically control the opening state of the vacuum valve under the conveying belt and the negative pressure output intensity of the multiple independent negative pressure partitions, execute on-demand adsorption, and avoid invalid pumping; specific steps include: S310: Issue adsorption control instructions: Based on the time-space adsorption window map generated in S230 The master control unit parses each window object according to a refresh period of 20 ms and generates a control instruction.

[0037] The instruction content includes: partition number ; adsorption start time , timestamp precision 1 ms; adsorption end time , timestamp precision 1 ms; target negative pressure value , range -2 kPa to -6 kPa, minimum adjustment step 0.1 kPa. The instruction is issued through the CAN bus, and the length of each message is 8 bytes, and the format is as follows: bytes 0-1: partition number (2 bytes, support 0-65535); bytes 2-3: start timestamp (2 bytes, unit ms); bytes 4-5: end timestamp (2 bytes, unit ms); bytes 6-7: target negative pressure value (2 bytes, unit 0.01 kPa). The master control unit and each partition control module use a synchronous clock mechanism to periodically send time synchronization frames to ensure that the clock deviation is less than ±1 ms. The communication delay is not more than 5 ms, if it exceeds, the message is discarded and re-sent.

[0038] S320: The partition vacuum adsorption module executes: After receiving the instruction, the partition vacuum adsorption module executes control in the interval : actuator: proportional electromagnetic valve, action response time less than 20 ms; vacuum pump: speed regulation range 1000-3000 rpm, regulation resolution 50 rpm, response delay <5 ms; sensor sampling frequency: 50 Hz.

[0039] The closed-loop control uses a PID control algorithm, and the deviation is defined as ; the control law is: wherein, , , are proportional, integral, and differential parameters, which are experimentally optimized and set to is 2.0, is 0.5, is 0.1. When , the PID controller starts compensation, and the negative pressure is restored to the target range by adjusting the speed of the vacuum pump or the opening of the electromagnetic valve. The adjustment period is 10 ms.

[0040] S330: Dynamic timing switching execution: during the conveying process, the master control unit dynamically switches the partition state according to the timing requirements of the window : when the material enters the partition: immediately trigger the window Zone valve opens; When material leaves the zone: Zone valve closes to prevent invalid pumping; Uncovered zone: Keep closed to ensure minimum energy consumption. To avoid pressure shock, the valve opening and closing adopts linear gradual mode: Gradually changes from 0% opening to target opening within 10 ms, or reversely closes.

[0041] The system supports up to 5 materials existing simultaneously; When two materials request the same zone simultaneously, the material that arrives first is preferentially served; If the arrival time is the same, the material with larger area is preferentially served; The delayed material is queued again in the next period to avoid loss. When a zone is not covered by material for 500 ms, the zone vacuum pump branch is automatically closed; 50 ms before the material enters, the zone is pre-charged to ensure response speed. Command loss or message error → Zone automatically inherits the parameters of the last period; Receive new instructions in the next refresh period to avoid adsorption interruption. For example: When the window diagram is defined as , , The zone valve opens in 10 ms gradual mode, the PID controller maintains -3.5±0.05 kPa, and the valve gradually closes at t=280 ms.

[0042] S4: Flexible adaptive fitting control: During the material adsorption process, based on the flexible structure of the belt and the real-time contact state of the material, the adsorption surface deformation is dynamically adjusted to improve the fitting effect and reduce the sliding deviation; Specifically, the following sub-steps are included: S410: Fitting state monitoring: During the process of material being adsorbed by the zone vacuum adsorption module, the fitting state of the material and the surface of the conveying belt is monitored in real time. Two high-speed cameras are arranged above the conveying belt, with a frame rate of 60 Hz, covering the full width of the belt. The camera completes the mapping of pixel coordinates and physical height through a calibration board, and the conversion formula is: Wherein, is the physical height difference, is the pixel difference, is the calibration coefficient (0.067 measured by experiment). When the physical height difference is greater than 0.2 mm and the pixel difference is greater than 3 px, it is determined that there is a gap. The 10x10 pressure sensor array in S120 is used to collect the contact pressure distribution , with a sampling frequency of 50 Hz. If the local contact pressure is lower than the threshold , it is determined that the fitting is insufficient. Multi-modal fusion: If the visual detection and pressure array detection results are inconsistent, the system preferentially uses the pressure array result, and the visual detection is used as redundancy.

[0043] S420: Flexible adaptation structure driving: When it is determined that there is uneven fitting, the main control unit sends an adjustment instruction to the flexible adaptation belt structure. The flexible adaptation belt structure comprises: an elastic base layer with a thickness of 2 mm and a material of TPU and a Young's modulus of 15 MPa; a micro-boss air bag unit with a diameter of 5 mm and a height of 1 mm, arranged in an array with a spacing of 10 mm x 10 mm, covering the entire adsorption area; each air bag unit is provided with a micro electric control valve, and the response time of the unit is less than 10 ms. The gas supply system is provided by a centralized gas source, and the gas source pressure range is 0.1-0.5 bar, and the adjustment accuracy is 0.01 bar. The system can simultaneously drive up to 20 air bag units for compensation adjustment. When adjusting, the main control unit selects the target air bag unit according to the abnormal area coordinates, and controls it to inflate / deflate within ±0.5 mm. The adjustment resolution is 0.05 mm, and the single response time is less than 20 ms.

[0044] S430: Surface automatic adjustment and stable adsorption: After the micro-deformation of the air bag unit, the surface of the conveying belt is automatically adjusted to a fitting shape matching the bottom surface of the material. In this process: the pressure array continuously collects real-time contact pressure distribution , and calculates the uniformity index: wherein, is the standard deviation of the contact pressure, is the average value of the contact pressure. When all local contact pressures , and , it is determined that the fitting adjustment is completed; if the above conditions are not met within 200 ms, the system performs secondary adjustment: increases the adjustment range by 20%, and repeats the inflation / deflation process. If the secondary adjustment still does not meet the standard, an alarm is triggered and the abnormal partition is marked.

[0045] When the material leaves the partition, the air bag unit automatically deflates to zero pressure state within 50 ms, and the idle power consumption is less than 0.1 W; the system regularly (every 1000 cycles) performs air bag pressure self-checking to ensure that the leakage rate is less than 2%, if a certain air bag unit continuously adjusts for 3 times without reaching the expected deformation (deviation greater than 0.1 mm), the system automatically shields the unit and is compensated by the surrounding units; if the pressure sensor data is lost, the visual detection result is used for determination; the designed service life of each air bag unit is greater than 10 6 times of inflation / deflation cycles, and the main control unit sends a maintenance prompt when the service life is exceeded. For example: when the visual detection finds that the material edge is floating up by 0.4 mm, the corresponding coordinates are (60 mm, 120 mm), the main control unit sends an inflation instruction to the two air bag units in the corresponding area, so that they expand by 0.4 mm. The pressure array monitors that the contact pressure returns to 150 Pa, and the uniformity index is 7%, and it is determined that the fitting adjustment is completed.

[0046] S5: Constructing adsorption memory model: based on historical material adsorption feedback data, energy consumption data and deviation rate, a mapping model between material type and adsorption parameters is constructed to form an adsorption parameter memory library; the following sub-steps are included: S510: Adsorption operation data acquisition: in the material conveying process, the main control unit collects the following operation data in real time: adsorption success rate data: the visual recognition module monitors the deviation and shedding of materials in the conveying process, and the calculation formula is: Wherein, is the stable conveying material quantity, is the total conveying material quantity. Energy consumption data: sampled by voltage sensor (range 0-240 V, accuracy ±0.5% FS) and current sensor (range 0-10 A, accuracy ±0.5% FS) synchronously, sampling frequency 1 kHz, energy consumption calculation formula: Wherein, U(t) is the voltage, I(t) is the current, and the unit of E is .

[0047] Material deviation rate data: the visual recognition module detects the deviation of the material center point from the expected trajectory, and the average deviation rate calculation formula is: Wherein, is the deviation distance of the ith material, is the total number of statistical samples, that is, the number of observed material position points. The sampling frequency of all data is 50 Hz, and the data integrity is ensured by CRC check. If there is packet loss, linear interpolation compensation is used.

[0048] S520: Data correlation analysis: the collected data and the material distribution mapping matrix output by S130 are analyzed. The input features include: material size parameters (L, W); contact pressure distribution C(x, y); adsorption window parameters . The output indicators include: adsorption success rate ; energy consumption E; average deviation rate .

[0049] Establish the correlation function: Processing flow: normalization: Outlier processing: if a data point deviates from the mean value by more than , then it is recorded as an outlier and discarded. Modeling method: mixed modeling of multivariate linear regression and random forest regression, with random forest parameters of 100 tree numbers and a maximum depth of 5. Weight distribution: the weights of linear regression and random forest are 0.4 and 0.6, respectively. Calculation period: updated every 1s, and each calculation is based on the last 50 groups of data.

[0050] S530: mapping relationship and model building: based on the correlation analysis result, the master control unit generates the adsorption memory model, including: input layer: material size (L, W), pressure distribution characteristic C(x, y), target negative pressure intensity ; hidden layer: 3-layer fully connected neural network, with scales of 64, 32, and 16 neurons, respectively, and activation function ; combined with a random forest model as a hybrid learner; output layer: predicted adsorption success rate , energy consumption , average deviation rate .

[0051] Model training parameters: learning rate 0.001, batch size 32, iteration number 1000; training set and validation set ratio 80%:20%; verification accuracy requirement greater than 95%; incremental training triggered once every 500 conveying tasks.

[0052] Storage and life: the model is stored in the master control unit database in TensorFlow Lite format; the last 10 versions are retained, and a ring FIFO strategy is used for management; the model life is greater than 10 6 times of calling, with a storage redundancy rate of 200%. If the model deviates by more than the threshold value (e.g., success rate prediction error greater than 10% or energy consumption prediction error greater than 15%) in continuous 3 predictions, the system automatically rolls back to the last stable version; when the database storage is insufficient, the oldest version is deleted and maintenance is prompted. For example: for a material with size 80x100 mm, pressure distribution peak value 120 Pa, target negative pressure -3.5kPa, the model predicts the adsorption success rate to be 98%, the energy consumption to be 0.12Wh, and the average deviation rate to be 0.4mm, all of which meet the process requirements.

[0053] S6: machine learning feedback optimization: through the reinforcement learning or deep neural network model deployed in the master control unit, the current feedback data and memory model are compared and analyzed to dynamically update the adsorption window diagram and partition control parameters, realizing continuous self-adaptive optimization of adsorption behavior of multiple types of materials, including the following sub-steps: S610: running feedback data acquisition: during the adsorption conveying process, the master control unit real-time collects the following feedback data: actual adsorption success rate: the material deviation and shedding are detected by the vision module, and the calculation formula is the same as S510.

[0054] Actual energy consumption : Collected by voltage sensor and current sensor, formula is: Where, integral interval Aligns with adsorption window time, ensures that predicted value and actual value correspond one by one. Sensor delay is controlled within 5 ms.

[0055] Actual average deviation rate: deviation of material center point from expected trajectory is calculated by vision module, window start time is used as alignment reference. All data sampling frequency is 50 Hz, and CRC check is performed. If data is lost or abnormal, linear interpolation compensation is used.

[0056] S620: Error analysis and tuning strategy generation: main control unit compares predicted value with actual value , calculates error: represents the prediction error of the kth index; is the actual observation value (true value), is the model prediction value; statistical indicators include: average error: is the average error of the kth index; N is the number of observation samples (e.g. 50 frames or 100 measurements); is the prediction error of index k in the i-th observation.

[0057] Mean square error (MSE): is the mean square error of the kth index, N: sample size. Threshold setting: greater than 5% or greater than 0.05 triggers tuning. Threshold is derived from laboratory benchmarking and statistical average of 1000 runs.

[0058] Tuning strategy: weight fine-tuning: linear regression weight is in the range of 0.3-0.7, with a step size of 0.0 each time; learning rate adaptation: if error decreases slowly, 0.002; if error oscillates, reduce by 0.0005; feature importance update is based on random forest output to calculate feature importance, if less than 1%, remove it. If error exceeds 5% after removal, restore the feature.

[0059] S630: Online parameter update and model retraining: When the tuning conditions are triggered, the main control unit performs online retraining: Data source: the most recent 1000 sets of running data; Partition ratio: 80% for training set, 20% for validation set; Optimization: Adam; Number of iterations: maximum 500; Early stopping condition: No improvement after 10 consecutive rounds of validation accuracy; Computing platform: The main control unit has an embedded ARM Cortex-A72 CPU, and the GPU module is called for acceleration when necessary. The single training time is controlled within 1 second to ensure real-time performance. Update new data based on the old model parameters. The formula is: in, is the old model parameter vector, is the learning rate, is the updated model parameter vector; is the loss function Parameters After the model is updated, it is immediately written to the database and the last 10 versions are retained, using a FIFO strategy. If the new model validation set accuracy is less than 95%, it will automatically roll back to the old version.

[0060] S640: Application of optimization results and fault tolerance mechanism: The updated model is directly put into use in the next round of transportation tasks and outputs new prediction values ; Fault-Tolerance Mechanism: Prediction Deviation Monitoring: If the prediction deviation exceeds a threshold (success rate error greater than 10% or energy consumption error greater than 15%) for three consecutive times, the system immediately rolls back to the old model. Caching: The system caches the last 100 tuning results, storing them in JSON format with timestamps, tuning strategies, and error metrics for easy traceability. A Safety Fallback: If a model becomes 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 Condition: If the number of consecutive rollbacks exceeds five, the system issues an alarm, prompting manual inspection. Example: For a delivery mission, the original model predicted a success rate of 95%, but the actual rate was 88%, with an error of 7%. The main control unit triggered tuning, adjusting the weights to 0.35 for linear regression and 0.65 for random forest, and reducing the learning rate to 0.0008. After retraining, the error between the new model's predicted and actual success rates dropped to 2%, meeting the requirement.

[0061] Example 2: Figure 2 As shown, this embodiment provides a negative pressure adsorption delivery intelligent control system based on machine learning, including: The conveyor belt assembly is used to carry and transport photovoltaic silicon wafers. The conveyor belt surface is embedded with a flexible adaptor structure and a resistive pressure sensor array to collect the contact pressure distribution between the material and the belt; A visual recognition module is arranged above the conveying belt to acquire a two-dimensional image and depth information of the photovoltaic silicon wafer, and generate a material distribution mapping matrix in combination with the output of the pressure sensor array; A partitioned vacuum suction module is arranged below the conveying belt to divide the conveying area into a plurality of independently controllable negative pressure partitions, each of which is provided with a vacuum valve and a vacuum pump to open and adjust the negative pressure intensity as needed; A main control unit is electrically connected to the visual recognition module, the pressure sensor array, and the partitioned vacuum suction module.

[0062] The conveying system is used to calculate the material trajectory and generate a time-space suction window map based on the material distribution mapping matrix and the conveying belt speed; According to the time-space suction window map, suction control instructions are issued to each partitioned vacuum suction module to dynamically switch the opening state and negative pressure intensity of the partitioned vacuum valve to achieve on-demand suction; During the material suction process, based on the feedback of the pressure sensor and the visual detection, the flexible adaptive structure is driven to deform slightly to improve the material fitting stability; The suction success rate, energy consumption, and deviation rate data during the material conveying process are collected to establish a suction memory model to realize the mapping of the material type and the suction parameters; Through the deployed machine learning model, the running feedback data and the suction memory model are compared and analyzed to dynamically optimize the time-space suction window map and the partitioned control parameters, realizing continuous adaptive optimization of photovoltaic silicon wafers of different sizes and shapes.

[0063] As Figure 3 shown is an existing negative pressure belt conveying assembly, which includes a workbench and a belt, and the corresponding modules are preferably improved on this basis.

[0064] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the actual situation. The preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0065] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of 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, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0066] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those of ordinary skill 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 the present application.

[0067] Those of ordinary skill in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0068] In several embodiments provided in the present application, it should be understood that the disclosed system, device, and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed ones can be indirect coupling or communication connection through some interfaces, devices, or modules, which can be electrical, mechanical, or other forms.

[0069] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0070] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0071] If the functions are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of software products, and the computer software products are stored in a storage medium, including a plurality of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various program code storage media.

[0072] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0073] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A method for intelligent control of a pressure swing adsorption system based on machine learning, the method comprising: receiving a plurality of inputs; determining a plurality of outputs based on the plurality of inputs; and outputting the plurality of outputs. The method comprises the following steps: S1, obtaining the size, shape and position distribution information of the material to be conveyed on the conveying belt, wherein the information is obtained by a visual recognition module and a belt surface pressure sensor, and a material space distribution map and a contact state map are constructed; S2, based on the material space distribution map and the contact state map, combining the conveying belt running speed, the beat information and the material form, a corresponding time-space adsorption window map is generated to indicate the activation state and the required adsorption strength of each negative pressure partition at a specific time; S3, according to the adsorption window map, dynamically control the opening state and negative pressure output strength of the vacuum valve of the multiple independent negative pressure partitions under the conveying belt to perform on-demand adsorption; S4, during the material adsorption process, based on the real-time contact state of the belt flexible structure and the material, dynamically adjust the adsorption surface deformation to improve the fitting effect and reduce the sliding deviation; S5, based on the historical material adsorption feedback data, energy consumption data and deviation rate, a mapping model between the material type and the adsorption parameters is constructed to form an adsorption parameter memory library; S6, through the reinforcement learning or deep neural network model deployed in the main control unit, the current feedback data and the memory model are compared and analyzed to dynamically update the adsorption window map and the partition control parameters, and the continuous adaptive optimization of the adsorption behavior of multiple types of materials is realized.

2. The method of claim 1, wherein, S1 specifically includes: Obtain the two-dimensional image information of the material through the visual recognition module arranged above the conveying belt to identify the size, boundary and shape characteristics of the material; Collect the contact pressure distribution information between the material and the belt through the pressure sensor arranged on the surface of the conveying belt; Fuse the two-dimensional image information and the contact pressure distribution information to generate material distribution mapping data containing material position, size and contact state.

3. The method of claim 2, wherein the method further comprises: S2 specifically includes: According to the material distribution mapping data, combining the real-time running speed and running beat parameters of the conveying belt, the position change trajectory of the material in 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; In the adsorption area position, combined with the size and contact state of different materials, a corresponding time-space adsorption window map is generated to indicate the activation state and negative pressure intensity of each negative pressure partition at a specific time.

4. The method of claim 1, wherein, S3 specifically includes: According to the time-space adsorption window map, control instructions are sent to the partition vacuum adsorption module under the conveying belt; After receiving the control instructions, the partition vacuum adsorption module opens the vacuum valve of the corresponding partition as needed and sets the negative pressure intensity; During the conveying process, the opening state of each negative pressure partition is dynamically switched according to the time sequence requirement of the adsorption window map to realize on-demand adsorption and reduce invalid air suction.

5. The method of claim 1, wherein, S4 specifically includes: During the material adsorption process, the fitting state of the material and the conveying belt surface is monitored in real time; When local uneven fitting is detected, the elastic layer or micro-convex structure of the flexible adaptive belt structure is deformed; Through the deformation of the elastic layer or micro-convex structure, the surface of the conveying belt is automatically adjusted to match the fitting form of the material bottom surface to enhance the adsorption stability.

6. The method of claim 1, wherein, S5 specifically includes: Collecting adsorption success rate data, energy consumption data and material deviation rate data of the material in the conveying process; Correlation analysis of the adsorption success rate data, energy consumption data and material deviation rate data and material distribution mapping data; Based on the correlation analysis results, generate the mapping relationship between the material type and the adsorption parameters, and construct the adsorption memory model.

7. The method of claim 1, wherein the method is based on machine learning. S6 specifically includes: Comparing the adsorption operation data of the current material with the adsorption memory model to identify parameter deviation; Based on the identified parameter deviation, generate new negative pressure partition control parameters through the reinforcement learning algorithm or neural network model deployed in the main control unit; Update the time-space adsorption window diagram to form a new adsorption strategy; In the next material conveying, control the partition vacuum adsorption module according to the updated adsorption strategy to realize continuous adaptive optimization.

8. A machine learning based intelligent control system for pressure swing adsorption transport, based on the machine learning based intelligent control method for pressure swing adsorption transport according to any one of claims 1-7, characterized in that, It includes: A conveying belt assembly for carrying and conveying photovoltaic silicon wafers, the conveying belt surface is embedded with flexible adaptive structure and resistive pressure sensor array for collecting the contact pressure distribution between the material and the belt; A visual recognition module is arranged above the conveying belt assembly for obtaining two-dimensional image and depth information of the photovoltaic silicon wafer, combining the pressure sensor array output to generate a material distribution mapping matrix; A partition vacuum adsorption module is arranged below the conveying belt, which divides the conveying area into multiple independently controllable negative pressure partitions, each partition is provided with a vacuum valve and a vacuum pump, which can open and adjust the negative pressure intensity as needed; A main control unit is electrically connected with the visual recognition module, pressure sensor array and partition vacuum adsorption module.

Images