Intelligent auxiliary material bonding system based on cloud-edge collaboration

Abstract

This invention relates to the field of cloud-edge collaboration technology, specifically to an intelligent auxiliary material bonding system based on cloud-edge collaboration. A cloud-based process optimization server receives environmental temperature and humidity data, auxiliary adhesive batch numbers, and curing agent ratio data. It then uses a built-in benchmark pressure prediction model to calculate and output a benchmark pressure-time curve, which is sent to the edge-end pressure controller. The edge-end pressure controller collects torque signals from piezoelectric force sensors in real time, performs differential calculations on the torque signals to extract a high-frequency pressure residual sequence at the moment of bonding contact, and superimposes this high-frequency pressure residual sequence onto the corresponding time node of the benchmark pressure-time curve to generate an actual pressure control command sequence. Based on this actual pressure control command sequence, it outputs an analog adjustment signal to a proportional valve to control the cylinder pressure. This solution dynamically generates a benchmark pressure curve based on environmental and material properties and superimposes local residual corrections, eliminating viscosity interference caused by environmental and material fluctuations and overcoming defects such as adhesive overflow and air bubbles.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaboration technology, and more specifically to an intelligent auxiliary material bonding system based on cloud-edge collaboration. Background Technology

[0002] In the bonding process between glass panels and decorative components, conventional bonding equipment uses a programmable logic controller (PLC) at the edge to control pressure. Operators pre-set fixed pressure values ​​or segmented pressure curves in the controller according to the product model. During bonding, the edge controller reads the pressure signal from the force sensor in real time, compares the read signal with the preset fixed pressure value, and outputs an adjustment signal to the proportional valve based on the comparison result. The proportional valve controls the cylinder air pressure to maintain a constant pressure output from the bonding mechanism. This type of equipment only records basic product specifications before operation and does not collect data on the bonding environment or batch properties of the adhesive.

[0003] The aforementioned constant pressure control method has shortcomings in addressing fluctuations in the physical properties of auxiliary adhesives. The viscosity of the auxiliary adhesive is affected by ambient temperature and humidity, batch differences in adhesive, and variations in the hardener ratio. When the ambient temperature rises or the hardener ratio changes, the adhesive's fluidity increases, causing the constant pressure output to result in adhesive overflowing towards the edges of the decorative part. When the adhesive viscosity increases, the constant pressure cannot ensure uniform spread of the adhesive, leading to air bubbles in the bonding area. Existing edge controllers have limited computing power and do not incorporate material and environmental property parameters, making it impossible to adjust the reference pressure based on fluctuations in material and environmental properties; they can only maintain a preset, rigid pressure output. A core technical problem arising from this is that the existing constant pressure control method cannot dynamically adjust the bonding reference pressure based on environmental fluctuations and differences in auxiliary adhesive properties, resulting in adhesive overflow or air bubble defects during the bonding process. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent auxiliary material bonding system based on cloud-edge collaboration, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The cloud-edge collaborative intelligent auxiliary material bonding system includes a cloud-based process optimization server and an edge-based pressure controller. The cloud-based process optimization server receives current environmental temperature and humidity data, auxiliary material adhesive batch numbers, and curing agent ratio data uploaded by the edge gateway. The cloud-based process optimization server has a built-in benchmark pressure prediction model. Based on the current environmental temperature and humidity data, the auxiliary material adhesive batch numbers, and the curing agent ratio data, the benchmark pressure prediction model outputs a benchmark pressure-time curve for the bonding process through fully connected layer calculations and sends the benchmark pressure-time curve to the edge-based pressure controller. The edge-based pressure controller collects torque signals from piezoelectric force sensors in real time. The edge-based pressure controller has a built-in residual calculation module. The residual calculation module performs differential calculations on the torque signals to extract a high-frequency pressure residual sequence at the moment of bonding contact. The high-frequency pressure residual sequence is superimposed on the corresponding time node of the benchmark pressure-time curve to generate an actual pressure control command sequence. The edge-based pressure controller outputs an analog adjustment signal to a proportional valve to control the cylinder pressure based on the actual pressure control command sequence.

[0006] Preferably, the benchmark pressure prediction model includes a feature mapping layer and a fluid viscosity fitting layer. The feature mapping layer receives the current ambient temperature and humidity data, the batch number of the auxiliary adhesive, and the curing agent ratio data. It maps the current ambient temperature and humidity data into an initial viscosity feature vector of the adhesive and maps the batch number of the auxiliary adhesive and the curing agent ratio data into a curing reaction rate feature vector of the adhesive. The fluid viscosity fitting layer receives the initial viscosity feature vector of the adhesive and the curing reaction rate feature vector of the adhesive. It performs matrix multiplication based on a preset adhesive shear thinning constitutive equation and outputs a dynamic viscosity change curve during the bonding process. The dynamic viscosity change curve is linearly transformed by the fully connected layer of the benchmark pressure prediction model to form the benchmark pressure-time curve.

[0007] Preferably, the residual calculation module includes a sliding window interception submodule and a differential operation submodule. The sliding window interception submodule takes the time point when the suction head contacts the glass panel as the starting point and intercepts multiple continuous signal segments backward on the torque signal according to a preset time step. The differential operation submodule performs a first-order backward differential calculation on the discrete data points in each of the continuous signal segments, obtains multiple differential result sequences, extracts the set of discrete points in the differential result sequences whose absolute values ​​are greater than a preset differential threshold, and rearranges the set of discrete points according to the timestamp order to generate the high-frequency pressure residual sequence.

[0008] Preferably, the edge pressure controller further includes a residual correction processor. The residual correction processor receives the high-frequency pressure residual sequence and the reference pressure-time curve, locates the corresponding time node on the reference pressure-time curve based on the timestamps of each discrete point in the high-frequency pressure residual sequence, and incorporates a moving average filter. The moving average filter performs a sliding window mean calculation on the high-frequency pressure residual sequence to generate a smoothed residual sequence. The residual correction processor amplifies the smoothed residual sequence according to a preset residual weight coefficient and linearly superimposes the amplified smoothed residual sequence onto the corresponding time node of the reference pressure-time curve.

[0009] Preferably, the edge pressure controller has a built-in proportional-integral-derivative (PID) arithmetic unit and a digital-to-analog converter (DAC). The PID receives the reference pressure-time curve superimposed with the smoothed residual sequence as the pressure setpoint and receives the torque signal from the piezoelectric force sensor as the pressure feedback value. The PID performs proportional-integral-derivative logic operations based on the deviation between the pressure setpoint and the pressure feedback value and outputs a digital control quantity. The DAC receives the digital control quantity and converts it into an analog adjustment signal with a corresponding voltage amplitude. The analog adjustment signal is transmitted to the electrical control interface of the proportional valve via a shielded cable.

[0010] Preferably, when generating the reference pressure-time curve, the cloud-based process optimization server assigns a cloud-based absolute timestamp to the first data point of the reference pressure-time curve, and encapsulates the cloud-based absolute timestamp and the reference pressure-time curve into a data packet and sends it down. The edge pressure controller receives the data packet, parses it to obtain the cloud-based absolute timestamp and the reference pressure-time curve, aligns the local clock of the edge pressure controller with the cloud-based absolute timestamp, and caches the reference pressure-time curve in the circular buffer of the edge pressure controller according to the aligned time reference.

[0011] Preferably, the feature mapping layer embeds a chemical reaction kinetic equation solver. The chemical reaction kinetic equation solver receives the curing agent ratio data to extract the curing agent concentration parameter, receives the batch number of the auxiliary adhesive to retrieve the corresponding main agent molecular chain length parameter and crosslinking density parameter, and substitutes the curing agent concentration parameter, the main agent molecular chain length parameter, and the crosslinking density parameter into the Arrhenius equation to perform iterative solution, outputting a multi-node conversion rate array of the auxiliary adhesive within the bonding time window. The feature mapping layer converts the multi-node conversion rate array into a feature vector of the adhesive curing reaction rate.

[0012] Preferably, the residual correction processor stores bonding stage division rules, which divide the bonding process into an initial contact stage, an intermediate extrusion stage, and a pressure holding stage. The residual correction processor determines the current bonding stage based on the interval on the reference pressure-time curve where the current time node is located. In the initial contact stage, the residual correction processor sets the residual weight coefficient to a first value; in the intermediate extrusion stage, it switches the residual weight coefficient to a second value; and in the pressure holding stage, it switches the residual weight coefficient to a third value. The residual correction processor amplifies the smoothed residual sequence based on the value corresponding to the current bonding stage.

[0013] Preferably, the proportional-integral-differential (PID) arithmetic unit is connected to a fuzzy inference engine. The fuzzy inference engine receives the deviation between the pressure setpoint and the pressure feedback value, as well as the rate of change of the deviation, as input linguistic variables. The fuzzy inference engine performs fuzzification, fuzzy inference, and defuzzification processing on the input linguistic variables according to a preset fuzzy rule base, and outputs the proportional coefficient adjustment, integral coefficient adjustment, and differential coefficient adjustment. The PID arithmetic unit receives the proportional coefficient adjustment, integral coefficient adjustment, and differential coefficient adjustment, updates the initial proportional coefficient, initial integral coefficient, and initial differential coefficient in real time, and performs the PID logical operation.

[0014] Preferably, the circular buffer includes a front-end buffer queue and a back-end buffer queue. When the edge-end pressure controller receives the data packet sent by the cloud-based process optimization server, it writes the parsed baseline pressure-time curve into the back-end buffer queue. After writing is completed, a buffer switching signal is triggered. In response to the buffer switching signal, the edge-end pressure controller swaps the read and write pointers of the front-end buffer queue and the back-end buffer queue. The proportional-integral-differential (PI-DI) arithmetic unit of the edge-end pressure controller reads the data points of the baseline pressure-time curve point by point according to the read and write pointers of the swapped front-end buffer queue in ascending order of timestamps.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The cloud-based process optimization server acquires environmental temperature and humidity data, adhesive batch numbers, and hardener ratio data. It then uses a built-in benchmark pressure prediction model to calculate and output a benchmark pressure-time curve, which is sent to the edge-end pressure controller. The edge-end pressure controller, based on the high-frequency pressure residual sequence extracted from the torque signal at the moment of contact, superimposes this residual sequence onto the corresponding time node of the benchmark pressure-time curve to generate the actual pressure control command sequence. This control logic transforms the bonding pressure from a fixed, rigid curve into a dynamically generated benchmark curve based on environmental and material properties. Furthermore, it incorporates local pressure residual correction at the edge end at the moment of contact, eliminating the interference of viscosity fluctuations caused by changes in environmental temperature and humidity and adhesive batch changes on the bonding pressure, and overcoming adhesive overflow and bubble defects.

[0016] 2. By setting up a chemical reaction kinetic equation solver and a fluid viscosity fitting layer in the cloud-based feature mapping layer, physical properties such as curing agent concentration and main agent molecular chain length are transformed into dynamic viscosity change curves, improving the matching degree between the reference pressure curve and the physical state of the adhesive. At the edge, a proportional-integral-differential (PID) arithmetic unit with a fuzzy inference engine is used to update control parameters in real time based on the deviation between the pressure setpoint and the feedback value and the rate of change of the deviation, improving the responsiveness of pressure control. By assigning a cloud-based absolute timestamp to the reference pressure time curve and aligning it with the edge's local clock, combined with a data read / write pointer swapping mechanism between the front-end and back-end dual-loop buffer queues, the time synchronization and control continuity of pressure curve data transmission and edge reading during cloud-edge collaboration are ensured. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall interaction of the cloud-edge collaborative intelligent auxiliary material bonding system of the present invention. Figure 2 This is a flowchart of the cloud-based baseline pressure prediction model processing of the present invention; Figure 3 This is a flowchart of the solution process for the chemical reaction kinetic equations of this invention; Figure 4 This is a flowchart of the high-frequency pressure residual sequence extraction process at the edge of the present invention. Figure 5 This is a flowchart of the edge residual correction process of the present invention; Figure 6 This is a flowchart of the proportional-integral-differential operation and fuzzy inference control of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides an intelligent auxiliary material bonding system based on cloud-edge collaboration, including a cloud-based process optimization server and an edge-end pressure controller. The cloud-based process optimization server establishes a wide area network (WAN) communication connection with an edge gateway. The edge gateway is deployed in the field control unit of the bonding production line and establishes a local area network (LAN) communication connection with the environmental parameter acquisition unit, auxiliary material information input unit, and edge-end pressure controller at the bonding station. The environmental parameter acquisition unit is deployed in the working area of ​​the bonding station to collect the current environmental temperature and humidity data. The auxiliary material information input unit obtains the batch number of the auxiliary adhesive and the ratio data of the curing agent. The edge gateway uploads the collected current environmental temperature and humidity data, auxiliary adhesive batch number, and curing agent ratio data to the cloud-based process optimization server.

[0020] The cloud-based process optimization server incorporates a benchmark pressure prediction model. This model receives current ambient temperature and humidity data, adhesive batch numbers, and curing agent ratios uploaded from the edge gateway as input parameters. It then outputs a benchmark pressure-time curve for the bonding process through fully connected layer calculations. The cloud-based process optimization server distributes this benchmark pressure-time curve to the edge-end pressure controller via the edge gateway. The edge-end pressure controller is deployed within the local control module of the bonding equipment and establishes real-time communication connections with the piezoelectric force sensor, proportional valve, and cylinder actuator of the bonding equipment. The piezoelectric force sensor is installed at the connection point of the bonding suction head, collecting the torque signal between the bonding suction head and the workpiece in real time during the bonding process. The edge-end pressure controller acquires the torque signal collected by the piezoelectric force sensor in real time.

[0021] The edge-end pressure controller incorporates a residual calculation module. This module receives real-time acquired torque signals, performs differential calculations on the torque signals, and extracts the high-frequency pressure residual sequence at the moment of contact. The residual calculation module then superimposes the generated high-frequency pressure residual sequence onto the corresponding time nodes of the reference pressure-time curve according to the timestamp correspondence, generating an actual pressure control command sequence. Based on the generated actual pressure control command sequence, the edge-end pressure controller outputs an analog adjustment signal of corresponding amplitude to a proportional valve. Upon receiving the analog adjustment signal, the proportional valve adjusts the output air pressure, controlling the extension and retraction of the cylinder through changes in output air pressure, thereby adjusting the output pressure of the bonding suction head and achieving closed-loop pressure control during the bonding process.

[0022] In this embodiment, the reference pressure-time curve for the bonding process is a continuous sequence of pressure setpoints throughout the entire bonding process, covering the entire process from the initial contact between the bonding suction head and the workpiece to the completion of pressure holding. The discrete data points of the reference pressure-time curve are processed using cubic spline interpolation to generate continuous curve data, ensuring a smooth and abrupt change in the pressure setpoints. Table 1 shows the core discrete parameters of the reference pressure-time curve for the bonding process. This parameter table is used to generate the basic data for the reference pressure-time curve. By interpolating the discrete data points in the table, a continuous reference pressure-time curve for the entire process can be obtained.

[0023] Table 1. Core Discrete Parameters of the Base Pressure-Time Curve for the Bonding Process ; In Table 1, t0 to t5 represent continuous time nodes in the bonding process. The interval between these time nodes can be adjusted according to the requirements of the bonding process. P0 to P5 represent the reference pressure settings for the corresponding time nodes. The magnitude of the reference pressure settings matches the specifications of the workpiece and the type of adhesive. During the generation of the reference pressure-time curve, based on the discrete data points in Table 1, reference pressure values ​​for any time node are generated through cubic spline interpolation, forming a continuous reference pressure-time curve.

[0024] Before uploading data to the cloud-based process optimization server, the edge gateway preprocesses the collected ambient temperature and humidity data, auxiliary adhesive batch numbers, and hardener ratio data. The preprocessing includes outlier removal and format validation. Outlier removal uses the Raida criterion to remove abnormal data points in the ambient temperature and humidity data that exceed a preset reasonable range. Format validation ensures that the format of the auxiliary adhesive batch numbers and hardener ratio data conforms to the input requirements of the baseline pressure prediction model, avoiding model calculation anomalies caused by invalid data input.

[0025] After acquiring the torque signal, the edge-end pressure controller first preprocesses the signal. This preprocessing includes signal amplification, power frequency interference filtering, and analog-to-digital conversion. The preprocessed torque signal is then converted into a digital signal and input to the residual calculation module. The filtering process uses a low-pass filter to retain the high-frequency effective components of the torque signal while filtering out random noise and power frequency interference, ensuring that subsequent differential calculations can accurately extract the pressure change characteristics at the moment of contact.

[0026] The instant of contact during bonding is determined by the abrupt change in the amplitude of the torque signal. When the amplitude of the pre-processed torque signal exceeds the preset contact determination threshold, the edge pressure controller determines that the current time point is the instant of contact between the bonding suction head and the workpiece to be bonded. The residual calculation module uses this instant of contact as the starting point to perform high-frequency pressure residual sequence extraction. The contact determination threshold can be adjusted according to the requirements of the bonding process to avoid misjudgments caused by environmental vibration or equipment shaking.

[0027] The high-frequency pressure residual sequence generated by the residual calculation module consists of high-frequency components in the torque signal at the moment of bonding that exceed the trend of the reference pressure change. These components characterize the pressure changes caused by abrupt changes in the adhesive flow state during bonding. Each discrete data point in the high-frequency pressure residual sequence has a corresponding timestamp. The accuracy of the timestamps is consistent with the sampling accuracy of the torque signal, ensuring that the residual sequence corresponds to the time nodes of the reference pressure-time curve.

[0028] When the edge-end pressure controller superimposes the high-frequency pressure residual sequence onto the reference pressure-time curve, it superimposes the value of each discrete data point in the residual sequence onto the reference pressure value at the same time in the reference pressure-time curve according to the time stamp correspondence, generating an actual pressure control command sequence. The superimposed actual pressure control command sequence is set with upper and lower pressure thresholds. The upper pressure threshold is the maximum pressure value allowed by the bonding process, and the lower pressure threshold is 0 to prevent the superimposed pressure value from exceeding the process allowable range, which could lead to bonding defects or equipment malfunctions.

[0029] In this embodiment, the cloud-based process optimization server adopts a distributed deployment architecture, including a data receiving unit, a model calculation unit, and a data distribution unit. The data receiving unit receives data uploaded from multiple edge gateways through a load balancing node, enabling parallel data reception across multiple production lines and workstations. The model calculation unit utilizes computing resources to execute calculations on the baseline pressure prediction model, achieving parallel generation of the baseline pressure-time curve. The data distribution unit distributes the generated baseline pressure-time curve to the corresponding edge pressure controller via an encrypted communication channel, ensuring the security and real-time performance of data transmission.

[0030] This embodiment utilizes a cloud-based process optimization server to combine ambient temperature and humidity data from the bonding station, along with adhesive batch number and hardener ratio data, to generate a dynamic baseline pressure-time curve that matches the current working conditions. This replaces the fixed, rigid pressure setpoints used in existing technologies, achieving dynamic adaptation of the baseline pressure. An edge-end pressure controller collects torque signals during the bonding process in real time, extracts the high-frequency pressure residual sequence at the moment of bonding contact, and corrects the baseline pressure-time curve in real time. This generates actual pressure control commands adapted to the real-time flow state of the adhesive, achieving dynamic closed-loop control of the bonding pressure and adapting to fluctuations in environmental and material parameters.

[0031] In a preferred embodiment, reference Figure 2 and Figure 3 The baseline pressure prediction model comprises a feature mapping layer and a fluid viscosity fitting layer. The feature mapping layer receives input data from the edge gateway, including current ambient temperature and humidity data, adhesive batch number, and hardener ratio. Its first output is an initial viscosity feature vector, and its second output is a curing reaction rate feature vector. The fluid viscosity fitting layer receives both the initial viscosity and curing reaction rate feature vectors. It performs matrix multiplication based on a pre-defined adhesive shear-thinning constitutive equation and outputs a dynamic viscosity change curve during the bonding process. This dynamic viscosity change curve is then input to the fully connected layer of the baseline pressure prediction model. After linear transformation by the fully connected layer, it outputs a baseline pressure-time curve for the bonding process.

[0032] The feature mapping layer extracts features from the input current ambient temperature and humidity data, mapping the ambient temperature and relative humidity data into an initial viscosity feature vector for the adhesive. This initial viscosity feature vector characterizes the initial flow properties of the adhesive under the current ambient temperature and humidity conditions. The feature mapping layer also extracts features from the input auxiliary adhesive batch number and hardener ratio data, mapping the physical properties of the main adhesive component corresponding to the auxiliary adhesive batch number and the hardener ratio data into an adhesive curing reaction rate feature vector. This curing reaction rate feature vector characterizes the rate change trend of the adhesive curing reaction during the bonding process.

[0033] The feature mapping layer embeds a chemical reaction kinetics equation solver. The solver's input receives curing agent ratio data and the main agent property parameters corresponding to the batch number of the auxiliary adhesive. The solver extracts the curing agent concentration parameter from the curing agent ratio data and retrieves the corresponding main agent molecular chain length and crosslinking density parameters from the cloud-based process optimization server database based on the auxiliary adhesive batch number. The solver then substitutes the extracted curing agent concentration, main agent molecular chain length, and crosslinking density parameters into the Arrhenius equation for iterative solving, outputting a multi-node conversion rate array of the auxiliary adhesive within the bonding time window. The feature mapping layer then converts this multi-node conversion rate array into a feature vector of the adhesive curing reaction rate.

[0034] The expression for the Arrhenius equation is: (1) In equation (1), Let A be the rate constant of the adhesive curing reaction at thermodynamic temperature T, and let A be the pre-exponential factor. The value of the pre-exponential factor is related to the molecular structure of the adhesive main agent and is determined by the physical property parameters of the main agent corresponding to the adhesive batch number. is the activation energy of the glue curing reaction, R is the universal gas constant with a value of 8.314 J / (mol·K), T is the environmental thermodynamic temperature at the bonding station, and the unit is K.

[0035] The iterative solution equation for the glue curing conversion rate is: (2) In Equation (2), is the glue curing conversion rate at time t during the bonding process, and its value range is from 0 to 1. t is the time variable during the bonding process, and the unit is s. is the integration time variable. The chemical reaction kinetics equation solver divides the bonding time window into multiple equally spaced time nodes, performs numerical integration on Equation (2) to obtain the curing conversion rate values corresponding to each time node, and forms a multi-node conversion rate array.

[0036] When the feature mapping layer maps the environmental temperature and humidity data to the initial viscosity feature vector of the glue, the initial viscosity calculation equation used is: (3) In Equation (3), is the initial viscosity of the glue under the current environmental temperature and humidity conditions, T is the environmental thermodynamic temperature, H is the environmental relative humidity. is the reference viscosity of the glue under the reference temperature and humidity conditions. is the reference thermodynamic temperature. is the reference relative humidity. is the viscous flow activation energy of the glue. is the influence coefficient of humidity on the viscosity of the glue. The values of the reference temperature and humidity conditions, the viscous flow activation energy, and the humidity influence coefficient are determined by the physical properties parameters of the main agent corresponding to the glue batch number.

[0037] The fluid viscosity fitting layer receives the initial viscosity feature vector of the glue and the glue curing reaction rate feature vector, and performs matrix multiplication operations according to the glue shear thinning constitutive equation to generate the dynamic viscosity change curve during the bonding process. The expression of the glue shear thinning constitutive equation is: (4) In Equation (4), is the dynamic viscosity of the glue at time t during the bonding process. is the shear rate received by the glue during the bonding process. n is the influence coefficient of the curing conversion rate on the viscosity of the glue, and m is the shear thinning index of the glue, and the value range is 0 < m < 1. The value of the shear rate is determined by the pressing speed of the bonding suction head and the thickness of the glue layer. The values of the curing conversion rate influence coefficient and the shear thinning index are determined by the physical properties parameters of the main agent corresponding to the glue batch number.

[0038] The fluid viscosity fitting layer combines the dynamic viscosity values ​​corresponding to each time point within the bonding time window to form a dynamic viscosity change curve. This curve represents a continuous sequence of adhesive viscosity changes over time during the bonding process. The dynamic viscosity change curve is then input to the fully connected layer of the benchmark pressure prediction model. The fully connected layer performs a linear transformation on the curve, generating a benchmark pressure-time curve. The linear transformation expression for the fully connected layer is as follows: (5) In equation (5), Let W be the baseline pressure value at time t during the bonding process, W be the weight matrix of the fully connected layer, and b be the bias vector of the fully connected layer. The weight matrix W and the bias vector b are obtained through training on a pre-constructed calibration dataset, which contains qualified bonding pressure curve data of glue under different viscosity conditions. The training process uses mean squared error as the loss function, and the values ​​of the weight matrix and bias vector are iteratively optimized through gradient descent algorithm until the loss function converges to below a preset threshold.

[0039] In this embodiment, the cloud process optimization server's database stores curing reaction kinetic parameters and viscosity characteristic parameters corresponding to different adhesive batches. Table 2 is a table of core physical property parameters corresponding to different adhesive batches. After receiving the batch number of the auxiliary adhesive, the feature mapping layer retrieves the corresponding parameters from the table and substitutes them into the chemical reaction kinetic equation solver and the initial viscosity calculation equation to perform the calculation.

[0040] Table 2. Core physical property parameters corresponding to different batches of adhesive. ; In Table 2, the batch number of the adhesive is a unique identifier for the auxiliary adhesive. Each batch number corresponds to a fixed set of physical property parameters, the values ​​of which are obtained through pre-calibrated laboratory tests. After receiving the batch number of the auxiliary adhesive, the feature mapping layer retrieves the corresponding physical property parameters by matching the batch number in Table 2, ensuring that the input parameters for model calculation are completely matched with the physical properties of the currently used adhesive.

[0041] The initial viscosity feature vector and the curing reaction rate feature vector of the glue generated by the feature mapping layer are both one-dimensional vectors with fixed dimensions. The dimension of the vector is consistent with the number of time nodes within the bonding time window, ensuring that the fluid viscosity fitting layer can perform point-by-point matrix operations on the two feature vectors to generate dynamic viscosity values ​​for the corresponding time nodes.

[0042] When performing matrix multiplication, the fluid viscosity fitting layer converts the initial viscosity feature vector of the glue, the curing conversion rate array and the shear rate parameter into a matrix of the same dimension. It then performs point-by-point numerical calculations according to equation (4) to obtain the dynamic viscosity value corresponding to each time node. Finally, it combines the dynamic viscosity values ​​of all time nodes in chronological order to form a continuous dynamic viscosity change curve.

[0043] When performing linear transformation, the fully connected layer performs Equation (5) on the viscosity value at each time node of the dynamic viscosity change curve to obtain the reference pressure value at the corresponding time node. Then, the reference pressure values ​​of all time nodes are combined in chronological order to form a continuous reference pressure-time curve. The linear transformation operation of the fully connected layer establishes a mapping relationship between the dynamic viscosity of the adhesive and the bonding reference pressure, ensuring that the generated reference pressure-time curve can adapt to the real-time changes in the adhesive viscosity.

[0044] This embodiment maps ambient temperature and humidity, adhesive batch and hardener ratio data to corresponding initial adhesive viscosity and curing reaction rate characteristics through a feature mapping layer. An embedded chemical reaction kinetics equation solver accurately calculates the trend of adhesive curing conversion rate during bonding. A fluid viscosity fitting layer, combined with the adhesive shear-thinning constitutive equation, generates a dynamic viscosity change curve during bonding. Then, through linear transformation calculations in a fully connected layer, a reference pressure-time curve matching the real-time physical properties of the adhesive is obtained, ensuring that the reference pressure can adapt to the dynamic changes in adhesive viscosity and improving the matching degree between the reference pressure curve and the physical state of the adhesive.

[0045] In yet another preferred embodiment, reference is made to Figure 4 and Figure 5 The residual calculation module includes a sliding window interception submodule and a differential operation submodule. The sliding window interception submodule receives the pre-processed torque signal and the time signal at the moment of contact during bonding. Starting from the moment the bonding suction head contacts the workpiece, it intercepts multiple continuous signal segments from the torque signal according to a preset time step. The differential operation submodule receives the multiple continuous signal segments output by the sliding window interception submodule. It performs a first-order backward differential calculation on the discrete data points within each continuous signal segment, obtaining multiple differential result sequences. The differential operation submodule extracts the set of discrete points in the differential result sequences whose absolute values ​​are greater than a preset differential threshold, rearranges the discrete point set in ascending order of timestamps, and generates a high-frequency pressure residual sequence.

[0046] The expression for calculating the first-order backward difference is: (6) In equation (6), for The first-order backward difference result at time 1, for Discrete values ​​of the torque signal acquired at different times. for Discrete values ​​of the torque signal acquired at different times. and The time interval between these intervals is the sampling period of the torque signal, measured in seconds.

[0047] The length of the continuous signal segments captured by the sliding window segmentation submodule is consistent with the preset sliding window length. The length of the sliding window can be adjusted according to the requirements of the bonding process. There is a preset overlap area between each continuous signal segment to ensure that the high-frequency pressure change signal at the moment of bonding contact can be completely captured, avoiding signal omission. When capturing signal segments, the sliding window segmentation submodule retains the corresponding original timestamp for each discrete data point within the signal segment, ensuring that the subsequently generated high-frequency pressure residual sequence corresponds to the time nodes of the reference pressure-time curve.

[0048] The differential operation submodule performs first-order backward differential calculations of equation (6) on discrete data points within each continuous signal segment in ascending order of timestamps to obtain the differential result sequence corresponding to each signal segment. The differential result sequence is used to characterize the rate of change of the torque signal. The sudden change in the flow state of the adhesive at the moment of contact will cause high-frequency abrupt changes in the torque signal, and the corresponding differential result sequence will have a large absolute value. The differential operation submodule extracts the discrete points of these abrupt changes through a preset differential threshold to form a high-frequency pressure residual sequence.

[0049] The edge-end pressure controller also includes a residual correction processor. The input of the residual correction processor receives the high-frequency pressure residual sequence output by the residual calculation module, as well as the baseline pressure-time curve sent by the cloud-based process optimization server. Based on the timestamps of each discrete point in the high-frequency pressure residual sequence, the residual correction processor locates the corresponding time nodes on the baseline pressure-time curve, establishing a correspondence between the discrete points of the residual sequence and the time nodes of the baseline curve.

[0050] The residual correction processor incorporates a moving average filter. The input to the moving average filter receives a high-frequency pressure residual sequence, and a sliding window mean is calculated on this sequence to generate a smoothed residual sequence. The calculation expression for the moving average filter is as follows: (7) In equation (7), for The smoothed residual value at time 10:00. for The discrete values ​​of the high-frequency pressure residual sequence at time t, where N is the sliding window length of the moving average filter, and the value is a positive integer.

[0051] The moving average filter calculates the mean through a sliding window, eliminating random noise components in the high-frequency pressure residual sequence while retaining the effective components characterizing changes in the adhesive flow state, thus avoiding fluctuations in pressure control commands caused by random noise. The value of the sliding window length N can be adjusted according to the requirements of the bonding process. A larger window length results in a more stable smooth residual sequence after filtering, while a smaller window length results in a faster response speed to signal abrupt changes.

[0052] The residual correction processor stores preset residual weight coefficients. It performs numerical amplification operations on the smoothed residual sequence according to these preset residual weight coefficients. The expression for the numerical amplification operation is: (8) In equation (8), for The smoothed residual value after time magnification. This represents the residual weighting coefficient.

[0053] The residual correction processor linearly superimposes the amplified smooth residual sequence onto the corresponding time nodes of the reference pressure-time curve according to the timestamp correspondence, thus completing the residual correction of the reference pressure-time curve. The linear superposition process involves adding the value of each discrete point in the amplified smooth residual sequence to the reference pressure value at the same timestamp in the reference pressure-time curve to obtain the corrected pressure setpoint sequence.

[0054] The residual correction processor stores bonding stage division rules, which divide the entire bonding process into an initial contact stage, an intermediate extrusion stage, and a holding pressure stage. Each bonding stage corresponds to a fixed time interval, which can be adjusted according to the requirements of the bonding process. The residual correction processor determines the current bonding stage based on the time interval of the current time node on the reference pressure-time curve and adjusts the residual weight coefficient value accordingly. In the initial contact stage, the residual weight coefficient is set to the first value; in the intermediate extrusion stage, it is switched to the second value; and in the holding pressure stage, it is switched to the third value. The processor then amplifies the smoothed residual sequence based on the value corresponding to the current bonding stage.

[0055] In this embodiment, the local storage unit of the edge pressure controller stores a control parameter configuration table for each bonding stage. Table 3 is the core control parameter table for each bonding stage. The residual correction processor retrieves the corresponding parameters from the table according to the bonding stage at the current time node and performs filtering, amplification and superposition operations on the residual sequence.

[0056] Table 3. Core Control Parameters for Each Stage ; In Table 3, T0-T1 represents the time interval of the initial contact stage, corresponding to the time when the bonding nozzle and the workpiece just come into contact; T1-T2 represents the time interval of the intermediate extrusion stage, corresponding to the time period of adhesive extrusion and spreading; and T2-T3 represents the time interval of the pressure holding stage, corresponding to the time period of adhesive curing and pressure holding. k1, k2, and k3 are the residual weighting coefficients corresponding to the three bonding stages, respectively. , , These are the differential thresholds corresponding to the three bonding stages, and N1, N2, and N3 are the sliding window lengths of the moving average filter corresponding to the three bonding stages.

[0057] During the bonding process, the residual correction processor obtains the current timestamp in real time, matches the current timestamp with the time interval in Table 3, determines the current bonding stage, retrieves the residual weight coefficient, differential threshold and sliding window length corresponding to the stage, and updates the operating parameters of the differential operation submodule and the moving average filter to ensure that the residual extraction and correction operations can adapt to the process characteristics of different bonding stages.

[0058] The first residual weighting coefficient for the initial contact stage is set to be greater than the second value for the intermediate extrusion stage. During the initial contact stage, a rapid response to the contact impact between the bonding nozzle and the workpiece is required. A larger residual weighting coefficient allows for quick adjustment of the reference pressure value, preventing adhesive splattering caused by excessive pressure at the moment of contact. The second value for the intermediate extrusion stage is set to be greater than the third value for the holding pressure stage. During the intermediate extrusion stage, the adhesive is in a rapid spreading state, requiring appropriate correction for pressure changes caused by adhesive flow. During the holding pressure stage, the adhesive flow tends to stabilize, and a smaller residual weighting coefficient maintains stable pressure output, avoiding bonding defects caused by pressure fluctuations.

[0059] When the differential operation submodule extracts the set of discrete points in the differential result sequence, the differential threshold used is matched with the current bonding stage. The differential threshold in the initial contact stage is set to a smaller value, which can extract the small pressure change signal at the moment of contact. The differential threshold in the pressure holding stage is set to a larger value, which can filter out the small signal fluctuations in the pressure holding process and avoid unnecessary pressure adjustments.

[0060] The sliding window length of the moving average filter is matched with the current bonding stage. The sliding window length in the initial contact stage is set to a smaller value to ensure that the filtered residual sequence can quickly respond to the pressure change at the moment of contact. The sliding window length in the pressure holding stage is set to a larger value to ensure that the filtered residual sequence is stable enough to maintain a stable pressure output in the pressure holding stage.

[0061] After completing the amplification operation of the smoothed residual sequence, the residual correction processor performs amplitude limiting on the amplified smoothed residual sequence. The amplitude limiting threshold matches the maximum value of the reference pressure corresponding to the current bonding stage, so as to avoid the amplified residual value being too large, which would cause the superimposed pressure setting value to exceed the range allowed by the process.

[0062] This embodiment accurately extracts the high-frequency pressure residual sequence at the instant of bonding contact through a sliding window interception submodule and a differential operation submodule. A moving average filter eliminates random noise in the residual sequence, retaining the effective components characterizing the changes in adhesive flow state. By defining bonding stage division rules, corresponding residual weight coefficients, differential thresholds, and filtering parameters are configured for different bonding stages. This allows the residual correction operation to adapt to the process characteristics of different bonding stages. Through staged residual correction, precise adjustment of the reference pressure-time curve is achieved, ensuring a perfect match between the bonding pressure and the real-time flow state of the adhesive.

[0063] In yet another preferred embodiment, reference is made to Figure 6 When generating the baseline pressure-time curve, the cloud-based process optimization server assigns a cloud-based absolute timestamp to the first data point of the curve. This timestamp uses the UTC time standard with millisecond-level precision. The cloud-based process optimization server encapsulates the cloud-based absolute timestamp and the baseline pressure-time curve into a standard-format data packet and sends it to the edge pressure controller via the edge gateway. Upon receiving the data packet, the edge pressure controller parses it to obtain the cloud-based absolute timestamp and the baseline pressure-time curve data. The edge pressure controller then aligns its local clock with the parsed cloud-based absolute timestamp using a network time protocol, ensuring the time error after alignment is within a preset error threshold. Based on the aligned time reference, the edge pressure controller caches the parsed baseline pressure-time curve in its local circular buffer.

[0064] The circular buffer comprises a front-end buffer queue and a back-end buffer queue, both with the same storage capacity, capable of storing complete baseline pressure-time curve data. After receiving and parsing the data packet from the cloud-based process optimization server, the edge pressure controller writes the parsed baseline pressure-time curve data into the back-end buffer queue. This writing process does not affect the normal read operation of the front-end buffer queue. Once the baseline pressure-time curve data is completely written into the back-end buffer queue, the edge pressure controller triggers a buffer switching signal. Responding to this signal, the read / write pointers of the front-end and back-end buffer queues are swapped. After the swap, the original back-end buffer queue becomes the new front-end buffer queue, and vice versa. The edge pressure controller's proportional-integral-derivative (PID) arithmetic unit, based on the swapped read / write pointers of the front-end buffer queue, reads the data points of the baseline pressure-time curve point by point in ascending timestamp order, using these as the setpoint for pressure closed-loop control.

[0065] The edge-end pressure controller integrates a proportional-integral-derivative (PID) arithmetic logic unit (PID) and a digital-to-analog converter (DAC). The first input of the PID receives the reference pressure-time curve data after superimposing a smoothed residual sequence, serving as the pressure setpoint. The second input receives the torque signal collected by a piezoelectric force sensor, serving as the pressure feedback value. The PID calculates the deviation between the pressure setpoint and the pressure feedback value, performs PID logic operations based on the deviation, and outputs a digital control signal. The DAC receives the digital control signal output from the PID and converts it into an analog adjustment signal with a corresponding voltage amplitude. This analog adjustment signal is transmitted via a shielded cable to the proportional valve's electrical control interface, enabling real-time control of the proportional valve.

[0066] The continuous-time operation expression of the proportional-integral-differential operator is: (9) In equation (9), The proportional-integral-differential operation at time t outputs the digital control quantity. This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients. The pressure deviation at time t. The value of is the pressure setpoint minus the pressure feedback value, and τ is the integral time variable.

[0067] The discretization expression for a digital control system is as follows: (10) In equation (10), This is the output digital control quantity for the k-th sampling period. The pressure deviation in the k-th sampling period, The pressure deviation during the (k-1)th sampling period. The sampling period for proportional-integral-differential operations is expressed in seconds (s).

[0068] The proportional-integral-differential (PID) arithmetic unit is connected to a fuzzy inference engine. The fuzzy inference engine receives the deviation between the pressure setpoint and the pressure feedback value, as well as the rate of change of this deviation, as input linguistic variables. The rate of change of deviation is calculated as the deviation of the current sampling period minus the deviation of the previous sampling period. Based on a pre-defined fuzzy rule base, the fuzzy inference engine sequentially performs fuzzification, fuzzy inference, and defuzzification processing on the input linguistic variables, outputting adjustments to the proportional coefficient, integral coefficient, and derivative coefficient. The PID arithmetic unit receives the coefficient adjustments output by the fuzzy inference engine, updates the initial proportional coefficient, initial integral coefficient, and initial derivative coefficient in real time, and performs PID logical operations using the updated coefficients.

[0069] The coefficient update expression for the proportional-integral-differential operator is: (11) In equation (11), , , The proportional coefficient, integral coefficient, and differential coefficient are updated after the k-th sampling period. , , These are the initial proportional coefficient, initial integral coefficient, and initial differential coefficient. , , These are the adjustment amounts for the proportional coefficient, integral coefficient, and differential coefficient output by the fuzzy inference engine.

[0070] The fuzzification process of the fuzzy inference engine uses a triangular membership function to convert precise input values ​​into corresponding fuzzy linguistic variables. The expression for the triangular membership function is: (12) In equation (12), Let x be the membership value of the input variable, and a, b, and c be the three vertex parameters of the triangular membership function, corresponding to the values ​​of the input variable with membership degrees of 0, 1, and 0, respectively.

[0071] The defuzzification process of the fuzzy inference engine uses the centroid method to convert the fuzzy output obtained from fuzzy inference into precise numerical adjustment values. The expression for the centroid method is: (13) In equation (13), To obtain the accurate output value after defuzzification, To output discrete points within the universe of discourse, for The corresponding membership value, where n is the number of discrete points in the output universe.

[0072] In this embodiment, the internal storage unit of the fuzzy inference engine stores a fuzzy rule base parameter mapping table. Table 4 is the core parameter mapping table of the fuzzy rule base. Based on the fuzzification result of the input deviation and the rate of change of deviation, the fuzzy inference engine retrieves the corresponding coefficient adjustment amount from the table, performs defuzzification processing, and outputs it to the proportional-integral-differential operator.

[0073] Table 4. Mapping table of core parameters of fuzzy rule base ; In Table 4, NB, NM, NS, ZO, PS, PM, and PB are subsets of fuzzy linguistic variables, corresponding to seven fuzzy levels: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large. When performing fuzzy inference, the fuzzy inference engine converts the input deviation and deviation change rate into corresponding fuzzy subsets. By matching the rules in Table 4, it obtains the corresponding fuzzy subset of coefficient adjustment amounts. Then, through defuzzification, it obtains the precise coefficient adjustment amounts and outputs them to the proportional-integral-differential (PID) arithmetic unit.

[0074] The clock alignment process for the edge pressure controller is performed each time a reference pressure-time curve data packet is received from the cloud. This ensures that the local clock at the edge is synchronized with the absolute timestamp in the cloud, preventing timing misalignment of pressure control caused by time discrepancies between the cloud and the edge. The clock alignment error threshold is set to 1ms to ensure that the timing accuracy of pressure control meets the requirements of the bonding process.

[0075] Both the front and back buffer queues of the circular buffer adopt a first-in, first-out (FIFO) storage structure, with read and write pointers pointing to the read and write positions of the queues, respectively. During buffer switching, only the addresses of the read and write pointers are swapped; no data copying is performed. This ensures that the buffer switching process can be completed within a single clock cycle, avoiding control interruptions caused by buffer switching and guaranteeing the continuity of pressure control.

[0076] When performing calculations, the proportional-integral-derivative (PID) arithmetic unit (PID) uses an incremental operation method, outputting only the increment of the control quantity between the current sampling period and the previous sampling period to avoid control overshoot caused by integral saturation. The output digital control quantity of the PID is set with upper and lower threshold values ​​to prevent the output control quantity from exceeding the input range of the digital-to-analog converter (DAC), which would cause distortion of the analog signal.

[0077] The conversion accuracy of the digital-to-analog converter matches the control accuracy of the proportional valve. The converted analog control signal is a standard industrial signal, which can directly drive the electronic control interface of the proportional valve to achieve linear air pressure regulation. The analog control signal is transmitted via a shielded cable to avoid signal distortion caused by electromagnetic interference in the production line environment, ensuring the accuracy of proportional valve control.

[0078] This embodiment achieves clock synchronization between the cloud and edge by using absolute timestamps in the cloud, ensuring a perfect match between the timing of the reference pressure-time curve and the control timing at the edge. A read / write pointer swapping mechanism using dual-buffered queues in the foreground and background enables seamless switching of the reference pressure-time curve data, guaranteeing the continuity of pressure curve data transmission and retrieval and preventing control interruptions. A proportional-integral-differential (PID) arithmetic unit with a fuzzy inference engine updates the control coefficients in real time based on pressure deviation and the rate of change of deviation, improving the response speed and stability of the pressure closed-loop control and achieving high-precision closed-loop control of the fitting process pressure.

Claims

1. A cloud-edge collaborative intelligent auxiliary material bonding system, characterized in that, The system includes a cloud-based process optimization server and an edge-end pressure controller. The cloud-based process optimization server receives current environmental temperature and humidity data, auxiliary adhesive batch numbers, and curing agent ratio data uploaded by the edge gateway. The cloud-based process optimization server has a built-in benchmark pressure prediction model. Based on the current environmental temperature and humidity data, the auxiliary adhesive batch numbers, and the curing agent ratio data, the benchmark pressure prediction model outputs a benchmark pressure-time curve for the bonding process through fully connected layer calculations and sends the benchmark pressure-time curve to the edge-end pressure controller. The edge-end pressure controller collects torque signals from piezoelectric force sensors in real time. The edge-end pressure controller has a built-in residual calculation module. The residual calculation module performs differential calculations on the torque signals to extract a high-frequency pressure residual sequence at the moment of bonding contact. The high-frequency pressure residual sequence is superimposed on the corresponding time node of the benchmark pressure-time curve to generate an actual pressure control command sequence. The edge-end pressure controller outputs an analog adjustment signal to a proportional valve to control the cylinder pressure based on the actual pressure control command sequence.

2. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 1, characterized in that, The benchmark pressure prediction model includes a feature mapping layer and a fluid viscosity fitting layer. The feature mapping layer receives the current ambient temperature and humidity data, the batch number of the auxiliary adhesive, and the curing agent ratio data. It maps the current ambient temperature and humidity data into an initial viscosity feature vector of the adhesive, and maps the batch number of the auxiliary adhesive and the curing agent ratio data into a curing reaction rate feature vector of the adhesive. The fluid viscosity fitting layer receives the initial viscosity feature vector of the adhesive and the curing reaction rate feature vector of the adhesive, performs matrix multiplication based on a preset adhesive shear thinning constitutive equation, and outputs a dynamic viscosity change curve during the bonding process. The dynamic viscosity change curve is linearly transformed by the fully connected layer of the benchmark pressure prediction model to form the benchmark pressure-time curve.

3. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 1, characterized in that, The residual calculation module includes a sliding window interception submodule and a differential operation submodule. The sliding window interception submodule takes the time point when the suction head contacts the glass panel as the starting point and intercepts multiple continuous signal segments backward on the torque signal according to a preset time step. The differential operation submodule performs a first-order backward differential calculation on the discrete data points in each continuous signal segment, obtains multiple differential result sequences, extracts the set of discrete points in the differential result sequences whose absolute values ​​are greater than a preset differential threshold, and rearranges the set of discrete points according to the timestamp order to generate the high-frequency pressure residual sequence.

4. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 3, characterized in that, The edge pressure controller further includes a residual correction processor. The residual correction processor receives the high-frequency pressure residual sequence and the reference pressure-time curve. Based on the timestamps of each discrete point in the high-frequency pressure residual sequence, it locates the corresponding time node on the reference pressure-time curve. The residual correction processor has a built-in moving average filter. The moving average filter performs a sliding window mean calculation on the high-frequency pressure residual sequence to generate a smooth residual sequence. The residual correction processor amplifies the smooth residual sequence according to a preset residual weight coefficient and linearly superimposes the amplified smooth residual sequence onto the corresponding time node of the reference pressure-time curve.

5. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 4, characterized in that, The edge-end pressure controller incorporates a proportional-integral-derivative (PID) arithmetic unit and a digital-to-analog converter (DAC). The PID receives the reference pressure-time curve superimposed with the smoothed residual sequence as the pressure setpoint and the torque signal from the piezoelectric force sensor as the pressure feedback value. Based on the deviation between the pressure setpoint and the pressure feedback value, the PID performs a proportional-integral-derivative logic operation to output a digital control quantity. The DAC receives the digital control quantity and converts it into an analog adjustment signal with a corresponding voltage amplitude. The analog adjustment signal is transmitted to the electrical control interface of the proportional valve via a shielded cable.

6. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 1, characterized in that, When generating the baseline pressure-time curve, the cloud-based process optimization server assigns a cloud-based absolute timestamp to the first data point of the baseline pressure-time curve and encapsulates the cloud-based absolute timestamp and the baseline pressure-time curve into a data packet for distribution. The edge pressure controller receives the data packet, parses it to obtain the cloud-based absolute timestamp and the baseline pressure-time curve, aligns its local clock with the cloud-based absolute timestamp, and caches the baseline pressure-time curve in the circular buffer of the edge pressure controller based on the aligned time reference.

7. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 2, characterized in that, The feature mapping layer embeds a chemical reaction kinetic equation solver. The chemical reaction kinetic equation solver receives the curing agent ratio data to extract the curing agent concentration parameter, and receives the batch number of the auxiliary adhesive to retrieve the corresponding main agent molecular chain length parameter and crosslinking density parameter. The chemical reaction kinetic equation solver substitutes the curing agent concentration parameter, the main agent molecular chain length parameter, and the crosslinking density parameter into the Arrhenius equation to perform iterative solution, and outputs a multi-node conversion rate array of the auxiliary adhesive within the bonding time window. The feature mapping layer converts the multi-node conversion rate array into a feature vector of the adhesive curing reaction rate.

8. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 4, characterized in that, The residual correction processor stores bonding stage division rules, which divide the bonding process into an initial contact stage, an intermediate extrusion stage, and a holding pressure stage. The residual correction processor determines the current bonding stage based on the interval on the reference pressure-time curve where the current time node is located. In the initial contact stage, the residual weighting coefficient is set to a first value; in the intermediate extrusion stage, the residual weighting coefficient is switched to a second value; and in the holding pressure stage, the residual weighting coefficient is switched to a third value. The residual correction processor amplifies the smoothed residual sequence based on the value corresponding to the current bonding stage.

9. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 5, characterized in that, The proportional-integral-differential (PID) arithmetic unit is connected to a fuzzy inference engine. The fuzzy inference engine receives the deviation between the pressure setpoint and the pressure feedback value, as well as the rate of change of the deviation, as input linguistic variables. The fuzzy inference engine performs fuzzification, fuzzy inference, and defuzzification processing on the input linguistic variables according to a preset fuzzy rule base, and outputs the proportional coefficient adjustment, integral coefficient adjustment, and differential coefficient adjustment. The PID arithmetic unit receives the proportional coefficient adjustment, integral coefficient adjustment, and differential coefficient adjustment, updates the initial proportional coefficient, initial integral coefficient, and initial differential coefficient in real time, and performs the PID logical operation.

10. The intelligent auxiliary material bonding system based on cloud-edge collaboration according to claim 6, characterized in that, The circular buffer includes a front-end buffer queue and a back-end buffer queue. When the edge-end pressure controller receives the data packet sent by the cloud process optimization server, it writes the parsed baseline pressure-time curve into the back-end buffer queue. After writing is completed, a buffer switching signal is triggered. In response to the buffer switching signal, the edge-end pressure controller swaps the read and write pointers of the front-end buffer queue and the back-end buffer queue. The proportional-integral-differential (PI-DI) arithmetic unit of the edge-end pressure controller reads the data points of the baseline pressure-time curve point by point according to the read and write pointers of the swapped front-end buffer queue in ascending order of timestamps.

Images