Piezoelectric jet dispensing control method, storage medium and equipment

Abstract

The invention discloses a piezoelectric jet dispensing control method, a storage medium and equipment, and belongs to the technical field of control. The invention aims to solve the problems of limited precision and inflexible control of the existing piezoelectric jet dispensing control. Normalization vectors corresponding to driving air pressure, a nozzle taper angle, a nozzle outlet inner diameter and a nozzle inner side gap serve as first input and are fed into an input layer of a first branch, and first branch output is obtained through a first hidden layer and an attention mechanism layer; the normalized vectors of the current target displacement instruction value, the actual measurement displacement value, the current applied voltage value, the voltage change rate, the historical displacement sequence average value, the working frequency and the accumulated working time serve as second input and are sent into an input layer of a second branch, and second branch output is obtained through a second hidden layer; and the first branch output and the second branch output are spliced, and then a compensation voltage increment, a prediction displacement error and a system stability index are finally obtained by utilizing a network model, so that the control of piezoelectric injection dispensing is realized.

Description

Technical Field

[0001] This invention belongs to the field of control technology, specifically relating to a control method, storage medium, and device for piezoelectric jet dispensing. Background Technology

[0002] With the rapid advancement of modern industrial technology, the electronic packaging industry has become one of the key pillars driving national economic development. Currently, from personal electronic devices to automotive electronics, and further to aerospace and defense, higher demands are being placed on electronic products—not only must they possess high performance and multifunctionality, but they also need to achieve low cost, miniaturization, and high reliability. Under this development trend, dispensing technology, which enables precise coating and efficient packaging, is gradually becoming one of the key links in the electronic manufacturing process.

[0003] The core of dispensing technology lies in accurately and evenly distributing adhesive to designated locations to achieve functions such as fixing, encapsulating, and interconnecting electronic components. Based on the application method, dispensing technology is mainly divided into two categories: contact dispensing and non-contact dispensing. Contact dispensing relies on the needle directly contacting the substrate, delivering the adhesive under pressure; while non-contact dispensing utilizes the instantaneous impact of the needle against the nozzle to spray the adhesive onto the substrate surface. In comparison, non-contact dispensing avoids direct contact with the substrate, effectively reducing mechanical wear and contamination problems, while also offering higher dispensing speed and repeatability, thus finding widespread application in the microelectronics packaging field.

[0004] In recent years, the introduction of piezoelectric actuation technology has further improved the performance of non-contact dispensing. Due to the advantages of piezoelectric ceramics, such as fast response speed, strong driving force, and high energy conversion efficiency, they are highly compatible with the needs of high-speed micro-dispensing and have become an important development direction to replace traditional pneumatic and hydraulic dispensing valves. In piezoelectric dispensing valves, the piezoelectric ceramic generates high-frequency vibrations under the drive of an applied voltage, causing the ejector pin to periodically strike the nozzle, thereby achieving rapid ejection of tiny adhesive droplets.

[0005] However, the dispensing process involves complex hydrodynamic behavior. The adhesive, impacted by the ejector pin, generates strong transient flow within the nozzle cavity. This flow is influenced by a combination of factors, including air pressure and changes in cavity volume. Existing control methods do not consider this coupling effect, instead controlling the dispensing volume per unit time based on predefined conditions. This approach limits control accuracy. Furthermore, due to the structure of the dispensing valve, the control signal and dispensing volume are not a standard linear relationship. Therefore, any change in electronic components necessitates designing adjustments to the dispensing volume for each component, requiring a readjustment of the control logic. This reduces control accuracy to some extent, especially with different circuit boards. In some cases, changes to all components necessitate readjusting the dispensing volume and rewriting the control logic for the current board, resulting in inflexible control and excessive time costs. These problems become even more pronounced when the dispensing valve is replaced. Summary of the Invention

[0006] This invention aims to solve the problems of limited accuracy and inflexibility in existing piezoelectric jet dispensing control.

[0007] A method for controlling piezoelectric jet dispensing includes:

[0008] The system acquires environmental data for piezoelectric jet dispensing, including driving air pressure, nozzle cone angle, nozzle outlet inner diameter, and nozzle inner clearance. The normalized vector corresponding to the environmental data is used as the first input and fed into the input layer of the first branch. The first branch output is obtained after passing through the first hidden layer and the attention mechanism layer.

[0009] Acquire control data for piezoelectric jet dispensing, including: current target displacement command value, actual measured displacement value, current applied voltage value, voltage change rate, historical displacement sequence average value, working frequency, and cumulative working time; send the normalized vector of the control data as the second input to the input layer of the second branch, and obtain the output of the second branch through the second hidden layer;

[0010] The outputs of the first branch and the second branch are concatenated to obtain h. 1j Based on h 1j Using the network model, a 3-dimensional vector Y=[y1, y2, y3] is finally obtained, which represents the compensation voltage increment ΔV, the predicted displacement error Δx, and the system stability index S, respectively.

[0011] Control of piezoelectric jet dispensing is achieved based on compensation voltage increment, prediction displacement error, and system stability indicators.

[0012] Furthermore, in the process of controlling piezoelectric jet dispensing by compensating for voltage increment, predicting displacement error, and system stability indicators, the compensating voltage increment is used to compensate for the voltage compensation amount that offsets the hysteresis effect of piezoelectric ceramics, and the predicting displacement error is used to compensate for the difference between the target displacement and the actual displacement.

[0013] Furthermore, in the process of controlling piezoelectric jet dispensing by compensating for voltage increments, predicting displacement errors, and evaluating system stability indicators, the system stability indicators are used to assess the effectiveness of the control.

[0014] Furthermore, the first hidden layer contains 64 neurons and uses the ReLU activation function; the second hidden layer contains 64 neurons and uses the ReLU activation function.

[0015] Furthermore, the calculation process of the normalized value of the historical displacement sequence average in the normalized vector corresponding to the environmental data includes:

[0016] For the historical displacement sequence average x 25 The displacement is normalized using the average displacement of the first three time points; the normalization formula is: ,in This represents the average displacement over the first three time points, which are time t, the time corresponding to t-2 milliseconds, and the time corresponding to t-4 milliseconds, respectively.

[0017] Furthermore, based on h 1j In the process of obtaining the 3D vector Y=[y1, y2, y3] using the network model, h 1j The vector Y is fed into the third hidden layer and then passed through the output layer to obtain a 3D vector.

[0018] Furthermore, the third hidden layer contains 32 neurons and employs the ReLU activation function.

[0019] Furthermore, the network model used to obtain the 3D vector Y, including the first hidden layer, the attention mechanism layer, and the second hidden layer, is pre-trained. The total loss function during training is as follows:

[0020]

[0021] Where: L_voltage is the voltage compensation loss, using mean square error; L_displacement is the displacement error loss, using mean square error; L_stability is the stability index loss, using binary cross-entropy; L_reg is the regularization loss, calculated as the cumulative sum of the weight coefficients of the first and second hidden layers; , , λ is the adjustment weight coefficient of the total loss function.

[0022] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the piezoelectric jet dispensing control method.

[0023] A control device for piezoelectric jet dispensing includes a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the piezoelectric jet dispensing control method.

[0024] Beneficial effects:

[0025] The control method of this invention takes into account the coupling effects of air pressure and cavity size, which in itself can improve control accuracy. More importantly, whether replacing components, circuit boards, or dispensing valves, the control method of this invention can ensure stable control accuracy without the need for recalibration, correction, or readjustment of the control logic for dispensing volume. This makes the invention more flexible to use and saves time on adjusting and updating the control logic, effectively saving time costs. Therefore, the FNN model of this invention shows greater potential. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the spraying model; where 1. firing pin; 2. adhesive; 3. nozzle.

[0027] Figure 2 The image shows the distribution of dispensing performance, where (a) corresponds to the adhesive flow rate and (b) corresponds to the adhesive liquid volume fraction.

[0028] Figure 3 The flow rate cloud diagrams for each cone angle are shown, where (a)-(d) correspond to 30° cone angle, 60° cone angle, 90° cone angle, and 120° cone angle, respectively.

[0029] Figure 4 The following are flow rate cloud diagrams for each inner diameter, where (a)-(d) correspond to inner diameters of 0.15mm, 0.25mm, 0.35mm, and 0.45mm, respectively.

[0030] Figure 5 The flow rate cloud diagrams for each gap are shown, where (a)-(d) correspond to gaps of 0.24 mm, 0.6 mm, 0.96 mm, and 1.2 mm, respectively.

[0031] Figure 6 Comparison chart for compensated voltage increments.

[0032] Figure 7 This is a comparison chart of displacement error control.

[0033] Figure 8 This is a comparison chart of the system stability evolution. Detailed Implementation

[0034] Since the nonlinear processes in the background technology are difficult to describe accurately using analytical models, it is necessary to study them using numerical simulation methods. Based on this, this invention, combining the working principle of a piezoelectric-driven dispensing valve, establishes a two-dimensional geometric model of the collision structure between the ejector pin and the nozzle, and uses COMSOL multiphysics simulation software to numerically simulate the flow and spray characteristics of the adhesive during the dispensing process. By analyzing the influence of factors such as driving air pressure, nozzle geometric parameters, and internal cavity clearance on the spraying speed and adhesive dot size, a foundation is provided for the structural optimization of the piezoelectric dispensing valve and the control process of small-diameter adhesive dots. To further improve dispensing accuracy and optimize the control effect of the dispensing process, this invention introduces a feedforward neural network algorithm. Compared with the traditional PID control method, FNN exhibits higher accuracy in handling complex nonlinear problems and can more effectively control the spraying speed and adhesive dot diameter, thus providing more precise technical support for the optimized design of the dispensing valve. Specific implementation method one:

[0036] The piezoelectric jet dispensing control method described in this embodiment includes the following steps:

[0037] S1. Fluid Dynamics Analysis:

[0038] During the dispensing process, the adhesive undergoes highly complex hydrodynamic motion within a sealed valve chamber. The reciprocating impact of the ejector pin causes the adhesive to flow rapidly and randomly within the chamber. Although this process exhibits unsteady and random characteristics, the flow of the adhesive must still strictly adhere to fundamental physical laws such as the conservation of mass, momentum, and energy. The law of conservation of mass primarily states that within any tiny unit, the mass of fluid flowing in and out per unit time should remain equal; that is, the system will not experience mass loss or accumulation during the flow process.

[0039] (1)

[0040] In the formula, The speeds are along the three directions. Vector components; t is the density of the adhesive fluid, and t is time.

[0041] Because the motion of the firing pin is linearly accelerated, the flow of the adhesive under the pressure of the firing pin remains continuous and consistent. Simultaneously, all field variables involved in the adhesive flow, such as pressure, velocity, density, and temperature, are differentiable. Based on the fundamental principles of conservation of mass, momentum, and energy, the Navier-Stokes (NS) equations can be derived. Therefore, the hydrodynamic behavior of the adhesive can be described by the corresponding mathematical differential equations.

[0042] (2)

[0043] (3)

[0044] (4)

[0045] In the formula, The speeds are along the three directions. The vector components, i.e. , , yes right , , The partial derivatives; These are the external forces acting on the micro-unit; It is the fluid density of the adhesive solution. It is the kinematic viscosity coefficient, and its unit is... .

[0046] Assuming the flow process is incompressible, the energy equation can be neglected, and the flow equation can be solved directly. Clearly, the (NS) equations are a nonlinear system of partial differential equations, and finding an analytical solution is extremely complex, usually only possible under specific conditions. In most cases, we cannot directly obtain the position parameters and three-dimensional orientation through analytical methods. velocity vector and pressure Furthermore, the volume of the sealed cavity changes when the ejector pin squeezes the adhesive, making it more difficult to obtain analytical solutions to the differential equations. COMSOL software discretizes the system of differential equations using the finite element method and solves these equations numerically, ultimately obtaining numerical solutions for all unknown parameters through iterative calculations.

[0047] S2. Numerical Analysis Model Establishment and Validation:

[0048] Considering that air pressure, nozzle structure size, and other factors can affect dispensing performance, they will also affect the control of piezoelectric dispensing valves. Since the piezoelectric dispensing valve is "fixed" in actual production, factors such as air pressure and nozzle structure size can be disregarded during the actual production process. Control can be achieved simply by establishing a relationship between the input control signal and the output dispensing effect. However, research has revealed that as production progresses, issues such as equipment maintenance and aging replacements arise. The replacement of equipment and components involves factors such as procurement standards, funding, shortlisted suppliers, and equipment iteration updates, meaning that the replaced equipment may not necessarily be the same as the original. If control is still performed according to the original control logic, it will inevitably lead to significant errors. Therefore, in actual production, it is necessary for the equipment manufacturer's technicians or experienced technicians to re-adjust the system, which will seriously affect the production process. More importantly, under normal circumstances, the debugging personnel only ensure that the equipment can dispense normally and do not consider it to directly affect the dispensing volume. They usually only control the dispensing volume by the working time of the dispensing valve. However, another problem arises: usually, once this parameter is set, the dispensing valve can operate according to the set setting during the dispensing process, without adjustment in the running control logic. However, research revealed that it can affect the compensation hysteresis effect to some extent, a point often overlooked in existing technologies. Therefore, this invention, after research, decided to incorporate air pressure and nozzle structural dimensions as implicit influences for control. Thus, this invention first analyzes the impact of air pressure and nozzle structural dimensions on dispensing performance, and then designs a control model based on the analysis results.

[0049] This invention establishes a two-dimensional numerical simulation model for the collision structure between the impact pin and the nozzle in a piezoelectric dispensing valve-driven jet dispensing device, as shown below. Figure 1 As shown. In each spraying cycle, the ejector pin undergoes two key motion phases: the impact phase and the separation phase. During the impact phase, the ejector pin and nozzle are in close contact, forming a contact line at the contact point. After the ejector pin strikes the nozzle, it stops moving, at which point the adhesive ejection velocity from the nozzle reaches its maximum. Then, in the separation phase, the ejector pin begins to move away from the nozzle orifice and stops retracting after colliding with the stop block. During the retraction process, the localized instantaneous vacuum effect generated by the ejector pin causes adhesive to rapidly replenish the area, preparing for the next spraying cycle. At this point, the nozzle stops ejecting adhesive, ensuring the stability and continuity of each spraying cycle.

[0050] To improve the accuracy of the analysis and to create a regularly shaped mesh, this invention appropriately simplified the collision structure and created a geometric model in COMSOL for numerical simulation. Considering the dynamic change in the volume of the fluid-sealed cavity during the collision between the impactor and the nozzle, the fluid-structure coupling module in COMSOL was used to simulate the dynamic characteristics of the fluid and the interaction between the structure. This module can handle the fluid boundary changes caused by the impactor's motion, especially the impact of the instantaneous vacuum formed during the separation stage on the flow of the adhesive. Simulation and analysis were performed using COMSOL.

[0051] S2.1 The effect of driving air pressure on dispensing performance:

[0052] To simulate and analyze the influence of driving air pressure on dispensing performance, a simulation model was first established in the COMSOL software. Since this simulation primarily studies the flow state of the adhesive, only the adhesive region between the inner side of the nozzle and the outer side of the ejector pin was modeled. Furthermore, the adhesive viscosity was set to 2000 cps, the ejector pin displacement to 180 μm, and the ejector pin vibration frequency to 100 Hz.

[0053] When a gas pressure P1 of 0.8 MPa is applied at the fluid inlet and the relative pressure P0 at the outlet is defined as 0 MPa, the simulation yields the following distribution cloud map of its dispensing performance: Figure 2 As shown, the velocity distribution exhibits a clear gradient change, with a maximum flow velocity of 10.3 m / s. Figure 2 As shown in (a), after the adhesive is sprayed into the air from the nozzle, the droplet diameter is 0.44 mm, and its volume fraction distribution is as follows. Figure 2 As shown in (b), the volume fraction of 0 to 1 indicates the range from no adhesive present in the unit volume (completely air) to the unit volume being filled with adhesive.

[0054] To investigate the effect of driving pressure on dispensing performance, while keeping other conditions constant, the outlet flow rate and droplet diameter of the adhesive were calculated when P1 was 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, and 0.8 MPa.

[0055] As shown in Table 1, both the outlet velocity of the adhesive and the droplet diameter increase with increasing driving pressure, while the slope of the velocity change remains basically consistent. The reason for this is that when the driving air pressure increases, the initial velocity at the fluid inlet also increases, resulting in a constant increase in the velocity at the nozzle outlet; simultaneously, the amount of adhesive flowing into the nozzle increases with pressure, leading to a larger diameter of the ejected droplets.

[0056] Table 1 Relationship between outlet flow rate, droplet diameter, and driving pressure

[0057]

[0058] S2.2 The effect of nozzle cone angle on dispensing performance:

[0059] Similar to the S2.1 simulation analysis method, the fluid region was used as the simulation model. In the COMSOL software, the viscosity of the adhesive was set to 2000 cps, the driving air pressure to 0.8 MPa, the impact pin displacement to 180 μm, the impact pin vibration frequency to 100 Hz, and the ambient temperature to 20℃. In the model, the impact pin diameter was 1.5 mm, the nozzle outlet diameter was 0.2 mm, and the gap between the impact pin and the nozzle was 0.75 mm. The adhesive flow velocity distribution was analyzed when the nozzle cone angle ranged from 30° to 120°. Figure 3 It can be seen that the maximum flow velocities of the adhesive under the four cone angle conditions are 7.77 m / s, 11.4 m / s, 11.5 m / s, and 7.97 m / s, respectively. To further analyze the relationship between the cone angle and the flow velocity, data calculation points were set at every 10° interval between the cone angles of 30° and 120°, as shown in Table 2.

[0060] Table 2 shows a significant nonlinear relationship between the cone angle and the adhesive flow rate. Specifically, within the cone angle range of 30° to 60°, the adhesive flow rate gradually increases, from 7.51 m / s to 9.72 m / s. This indicates that at smaller cone angles, the adhesive flow is relatively slow; as the cone angle increases, the flow accelerates and the flow rate increases. However, when the cone angle exceeds 60°, the adhesive flow rate begins to decrease, dropping to 4.02 m / s at 120°. This decrease in flow rate is related to the increased flow resistance caused by excessively large cone angles. Excessively large cone angles induce flow separation or intensified turbulence, thus suppressing the flow rate. This trend suggests that the effect of the cone angle on the adhesive flow rate is critical: moderately increasing the cone angle can increase the adhesive flow rate, but beyond a certain angle, the flow rate decreases rapidly. Therefore, in practical applications, selecting an appropriate cone angle is crucial for optimizing the adhesive flow rate.

[0061] Table 2 Relationship between cone angle and adhesive outlet velocity

[0062]

[0063] S2.3, The effect of nozzle outlet inner diameter on dispensing performance:

[0064] The driving conditions are the same as in the aforementioned simulation model. Additionally, the diameter of the impact pin is set to 1.5 mm, the gap between the impact pin and the nozzle to be 0.75 mm, and the nozzle cone angle to be 120°. The flow velocity distribution of the adhesive is analyzed when the outlet inner diameter is 0.15 mm, 0.25 mm, 0.35 mm, and 0.45 mm. Figure 4The paper presents the velocity distribution of the dispensing flow field under different nozzle outlet inner diameters. The results show that as the nozzle outlet inner diameter increases, the overall maximum velocity of the flow field gradually increases, reaching 11.7 m / s, 13.1 m / s, 14.1 m / s, and 15.2 m / s, respectively. This demonstrates that the nozzle outlet size has a significant impact on the fluid motion characteristics during dispensing.

[0065] When the nozzle inner diameter is small, the fluid experiences significant viscous resistance in the narrow channel, resulting in a low overall flow velocity, with the high-velocity region mainly concentrated near the nozzle axis and a large velocity gradient. As the nozzle inner diameter increases, the fluid channel resistance decreases, the flow becomes smoother, the high-velocity region expands significantly, and the velocity distribution becomes more uniform. At this point, the fluid's kinetic energy is fully released, and the overall energy level of the flow field increases.

[0066] Table 3 shows the correlation between the outlet inner diameter and the outlet flow rate of the adhesive. The data shows that as the outlet inner diameter increases, the outlet flow rate of the adhesive gradually decreases, exhibiting a clear negative correlation. The flow rate is highest at 5.81 m / s when the outlet inner diameter is 0.15 mm; however, when the inner diameter increases to 0.45 mm, the flow rate drops to 2.52 m / s. This indicates that under a certain driving pressure, the smaller the nozzle outlet size, the higher the fluid velocity. This is due to the confinement of the fluid in the narrow channel, resulting in concentrated kinetic energy.

[0067] This principle aligns with fluid mechanics: when fluid passes through a smaller cross-section, the velocity increases due to the constraints of the continuity equation. A smaller outlet diameter leads to a higher jet velocity, but may also increase local pressure drop and shear stress. Therefore, nozzle design should comprehensively consider both jet velocity requirements and fluid stability to ensure optimal jet performance while maintaining sufficient flow rate.

[0068] Table 3 Relationship between outlet inner diameter and adhesive outlet flow rate

[0069]

[0070] S2.4 The effect of nozzle inner gap on dispensing performance:

[0071] In the COMSOL software, keeping the driving conditions constant, the impact pin diameter is defined as 1.5mm, the nozzle outlet inner diameter as 0.2mm, and the nozzle cone angle as 120°. The adhesive flow rate distribution is analyzed when the nozzle inner clearance is 0.24mm, 0.6mm, 0.96mm, and 1.2mm. Figure 5The maximum flow velocities were found to be 9.5 m / s, 8.71 m / s, 7.99 m / s, and 7.68 m / s, respectively. Further analysis was conducted on the maximum flow velocity of the adhesive when the gap between the inner side of the nozzle and the outer side of the ejector pin was 0.24–1.2 mm, with each 0.12 mm serving as a data calculation point. The results are shown in Table 4.

[0072] When the gap is small, the velocity gradient increases, and the shear stress of the fluid per unit area also increases, resulting in a higher flow rate. Therefore, when selecting different sized ejector pins, the nozzle should ensure sufficient glue supply within the cavity while appropriately reducing the inner gap to increase the ejection speed.

[0073] Table 4 Relationship between gap and adhesive outlet flow rate

[0074]

[0075] This invention considers that the analysis becomes relatively complex when all factors influence dispensing performance together, especially when both positively and negatively correlated indicators affect dispensing performance, making the control analysis even more complicated. The above process analyzed the impact of different factors on dispensing performance, showing that some influences exhibit linear relationships, some nonlinear relationships, and even some negative correlations. However, it is certain that all these factors will affect dispensing performance and subsequent control effects. Therefore, this invention employs a neural network model for control.

[0076] S3. Build and train a neural network model for control:

[0077] Based on actual control signals and the implicit influences of air pressure, nozzle structure size, etc., this invention designs a dual-channel neural network model. The first channel includes an input layer, a first hidden layer, and an attention mechanism layer. The second channel includes an input layer and a second hidden layer. The outputs of the first and second hidden layers are concatenated and then fed into a third hidden layer, which is then connected to the output layer. It should be noted that the network model used in this embodiment is a feedforward neural network model. However, in actual processing, any neural network model can be built. But the input should be processed separately for the two types of inputs using a dual-channel approach before fusion. The depth of the dual-channel branches before fusion can be determined according to the actual situation, generally not exceeding 3 layers. The depth of the network layers after concatenation can also be determined according to the actual situation, generally not exceeding 3 layers.

[0078] In the feedforward neural network of this embodiment

[0079] The input layer of the first channel receives a 4-dimensional input vector X1 = [x 11 , x 12 , x 13, x 14 The dimensions are as follows:

[0080] x 11 The driving pressure is set to a value ranging from 0 to 1 MPa and normalized to the interval [0,1].

[0081] x 12 The nozzle cone angle is 0-0.5 mm, normalized to the [0,1] interval.

[0082] x 13 The nozzle outlet inner diameter is 0-1.5 mm, normalized to the [0,1] interval;

[0083] x 14 The nozzle inner clearance ranges from 0 to 51.5 mm and is normalized to the [0,1] interval.

[0084] It should be noted that the input value range here is determined based on a relatively common range and normalized accordingly. For other special models or sizes (especially those larger than the above range), the size range can be redefined, normalized again, and the network retrained for subsequent recognition processes.

[0085] Regarding the setting of this channel, it should be noted that in previous actual production processes, it required technical personnel from the equipment manufacturer or experienced technicians to re-adjust the system. This significantly impacted the production process. More importantly, typically, the technicians only ensured that the equipment could dispense glue normally and did not consider its direct impact on the dispensing volume. They usually controlled the dispensing volume solely through the working time of the dispensing valve. However, this presents another problem: once this parameter is set, the dispensing valve operates according to the set settings during the dispensing process, without adjustment in the operational control logic. However, research has revealed that it can affect the compensation hysteresis effect to some extent, a point often overlooked in existing technologies. Therefore, this invention, after research, decided to consider air pressure and nozzle structural dimensions as implicit influences on control. Thus, this invention first analyzes the impact of air pressure and nozzle structural dimensions on dispensing performance, and then designs a control model based on the analysis results.

[0086] The first hidden layer contains 64 neurons and uses the ReLU activation function. The output of the i-th neuron is calculated using the following formula: , i=1,2,...,64, where Weight matrix elements, For bias vector The ReLU function is defined as ReLU(z) = max(0, z) for the elements.

[0087] The output of the first hidden layer is fed into the attention mechanism layer. The input to the attention mechanism layer first undergoes a linear feature transformation and then passes through a softmax function to obtain a weight vector. The output of the attention mechanism layer is obtained by multiplying the input of the attention mechanism layer with the corresponding weight vector. Since changes in air pressure and cavity volume are influenced by multiple coupled factors, meaning that each factor does not directly affect the amount of adhesive dispensing, this invention uses an attention mechanism to address the influence of each factor in the coupled effects, thereby improving the feature extraction performance of the first channel.

[0088] The second channel's input layer receives a 7-dimensional input vector X2=[ x 21 , x 22 , x 23 , x 24 , x 25 , x 26 , x 27 The dimensions are as follows:

[0089] x 21 The current target displacement command value xd, with a value range of 0-50μm, normalized to the interval [0,1];

[0090] x 22 The actual measured displacement value xa, with a range of 0-50μm, is normalized to the [0,1] interval;

[0091] x 23 The current applied voltage value V, ranging from 0 to 150V, is normalized to the interval [0,1].

[0092] x 24 The voltage change rate dV / dt ranges from -1000V / s to 1000V / s and is normalized to the interval [-1,1].

[0093] x 25 x is the average of the historical displacement sequence, which is the average displacement of the first 3 time points, normalized to the interval [0,1]; 25 The normalization formula is: ,in =(1 / 3)×(xa(t)+xa(t-2ms)+xa(t-4ms));

[0094] x 26 The operating frequency f ranges from 10 to 500 Hz and is normalized to the interval [0,1].

[0095] x 27 The cumulative working time t_acc ranges from 0 to 10000h and is normalized to the interval [0,1].

[0096] The second hidden layer contains 64 neurons and uses the ReLU activation function, which is basically the same as the first hidden layer, except for the weight matrix. .

[0097] The outputs of the attention mechanism layer and the second hidden layer are concatenated to obtain h. 1j Then, the signal is fed into the third hidden layer. The third hidden layer contains 32 neurons, using the ReLU activation function. The output of the i-th neuron is calculated using the following formula:

[0098] i=1,2,...,32

[0099] in Weight matrix elements, For bias vector Element;

[0100] The output layer contains three neurons, outputting a 3-dimensional vector Y = [y1, y2, y3], representing the compensation voltage increment ΔV, the predicted displacement error Δx, and the system stability index S, respectively. The calculation formula for the output layer neurons is as follows:

[0101] i=1,2,3

[0102] in Weight matrix elements, For bias vector Element;

[0103] The outputs of the three neurons in the output layer are as follows:

[0104] y1 is the compensation voltage increment ΔV, with a value range of -20V to 20V, used to offset the hysteresis effect of piezoelectric ceramics. The formula for calculating the total driving voltage after compensation is: V_total=V_base+ΔV, where V_base is the base control voltage.

[0105] y2 is the predicted displacement error Δx, which ranges from -5μm to 5μm. It represents the estimated value of the difference between the target displacement xd and the actual displacement xa. The calculation formula is: Δx = xd - xa_predicted;

[0106] y3 is the system stability index S, with a value range of 0-1. It is a normalized value. The closer the value is to 1, the more stable the system is. When S < 0.6, a system warning is triggered.

[0107] It should be noted that: regarding h 1jIn order to ensure efficiency and effective training of the attention mechanism layer, this embodiment of the network model uses one hidden layer. However, other network depths and network structures can be used as needed.

[0108] In this embodiment, the feedforward neural network collects the working data of each dispensing head in real time and optimizes the voltage signal according to the feedforward neural network algorithm to compensate for the hysteresis effect, thereby improving the control accuracy and stability of the system. It should be noted that the compensation voltage increment ΔV is used to counteract the hysteresis effect of the piezoelectric ceramic, and the characteristics of the first channel of this invention also affect it. By influencing the hysteresis effect of the piezoelectric ceramic, it acts on the dispensing control. This is actually a implicit control method, which not only makes the overall control more convenient but also improves the control effect.

[0109] The neural network is controlled based on the training set. The training of the feedforward neural network adopts a weighted multi-task loss function:

[0110]

[0111] Where: L_voltage is the voltage compensation loss, calculated using the mean square error formula:

[0112] ;

[0113] in This is a real label;

[0114] L_displacement represents the displacement error loss, calculated using the mean square error formula:

[0115] ;

[0116] L_stability is the stability index loss, calculated using binary cross-entropy, and the formula is as follows:

[0117] ;

[0118] L_reg is the regularization loss, calculated using the following formula:

[0119] ;

[0120] The adjustment weighting coefficient is set as follows: =0.5, =0.3, =0.2, λ=0.001;

[0121] The Adam optimization algorithm is used for training, with an initial learning rate of... =0.001, first moment estimate of exponential decay rate =0.9, second moment estimate of exponential decay rate =0.999, Batch size=32, Maximum number of training epochs=200.

[0122] S4. Real-time control based on the feedforward neural network module includes the following steps:

[0123] Step S41: Data acquisition. The analog signals from the displacement sensor and voltage sensor are acquired by the ADC analog-to-digital converter and converted into digital quantities. The sampling time is ≤0.2ms. At the same time, the air pressure and nozzle structure size data are acquired and converted into digital quantities.

[0124] Step S42: Feature extraction and normalization;

[0125] Step S43: Perform inference based on neural networks;

[0126] Step S44: Output denormalization, converting the normalized network output into physical quantities, ΔV = y1×20-10, Δx = y2×5-2.5, S = y3, processing time ≤ 0.1ms;

[0127] Step S45: Calculate the compensation voltage. Calculate the total driving voltage V_total = V_base + ΔV based on the output compensation voltage increment. Processing time ≤ 0.1ms.

[0128] Step S46: Safety limiting, limit the output voltage range to ensure 0≤V_total≤150V, processing time≤0.1ms;

[0129] Step S47: PWM signal generation, converting the voltage command into a 100kHz PWM signal to drive the power amplifier, with a generation time ≤0.4ms;

[0130] The control cycle is 2ms, which meets the requirements for real-time control.

[0131] Neural network control results and comparative analysis: To demonstrate its impact on control, PID control is introduced for comparison, such as... Figures 6-8 As shown.

[0132] Figure 6 Comparison chart of voltage increments. Figure 6 The X-axis represents the comprehensive index of latent influencing factors, and the Y-axis represents the compensation voltage increment ΔV (V), i.e.:

[0133] Horizontal axis: Comprehensive index of hidden influencing factors - comprehensively considering the coupled influence of structural parameters such as driving air pressure, nozzle cone angle, outlet inner diameter, and inner clearance on dispensing performance.

[0134] Vertical axis: Compensation voltage increment ΔV(V) - the voltage compensation amount used to counteract the hysteresis effect of piezoelectric ceramics. The more precise the value, the better the control effect.

[0135] From low voltage with a small cone angle to high voltage with a large cone angle, the influence of latent factors gradually weakens, and the required compensation voltage decreases. The FNN can accurately capture the coupling effect of multiple factors with an error of only 0.2-0.3%. Specific data are shown in Table 5.

[0136] Table 5 Comparison of Compensation Voltage Increments

[0137]

[0138] Figure 7 This is a comparison chart of displacement error control. Figure 7 In the diagram, the X-axis represents the comprehensive index of the control signal, and the Y-axis represents the predicted displacement error Δx (μm), i.e.:

[0139] Horizontal axis: Comprehensive index of control signals - the complexity of real-time control signals including target displacement, applied voltage, voltage change rate, and historical displacement sequence.

[0140] Vertical axis: Predicted displacement error Δx (μm) - the difference between the target displacement and the actual displacement. The smaller the error, the higher the dispensing accuracy.

[0141] As the control signal strength increases (large displacement + high voltage), the system response becomes more complete, and the displacement error decreases. The FNN error (Δx) stabilizes at 0.2-0.6μm, which is far better than the 1.2-2.8μm of the PID. Specific data are shown in Table 6.

[0142] Table 6 Comparison of Predicted Displacement Errors

[0143]

[0144]

[0145] Figure 8 This is a comparison chart of system stability evolution. Figure 8 In the diagram, the X-axis represents the operating condition complexity index, and the Y-axis represents the system stability index S, i.e.:

[0146] The horizontal axis represents the operational complexity index, which combines the overall operating frequency and cumulative operating time, reflecting the level of challenge faced by the equipment under high-frequency, long-duration operation.

[0147] The vertical axis represents the system stability index S(0-1), a dimensionless index; the closer it is to 1, the more stable the system.

[0148] Increased operating complexity (high frequency + long duration) poses a severe challenge to system stability. The original system's S-value dropped from 0.68 to 0.48, the PID remained at 0.70-0.82, and the FNN consistently maintained high stability at 0.92-0.95. Specific data are shown in Table 7.

[0149] Table 7 Comparison of System Stability Evolution

[0150]

[0151]

[0152] In summary, the FNN model outperforms the PID model under all test conditions, exhibiting higher control accuracy. Since the control method of this invention considers the coupling effects of air pressure and cavity size, this inherently improves control accuracy. More importantly, regardless of whether components, circuit boards, or dispensing valves are replaced, the control method of this invention can guarantee stable control accuracy without requiring recalibration, correction, or readjustment of the dispensing volume control logic. This makes the invention more flexible and saves time on adjusting and updating the control logic, effectively reducing time overhead. Therefore, the FNN model of this invention demonstrates greater potential. Specific Implementation Method Two:

[0154] This embodiment is a computer storage medium that stores at least one instruction, which is loaded and executed by a processor to implement the piezoelectric jet dispensing control method.

[0155] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0157] This embodiment is a control device for piezoelectric jet dispensing. The device includes a processor and a memory. It should be understood that this includes any device including a processor and a memory described in this invention. The device may also include other units or modules that perform display, interaction, processing, control, and other functions through signals or instructions.

[0158] The memory stores at least one instruction, which is loaded and executed by the processor to implement the piezoelectric jet dispensing control method.

[0159] Those skilled in the art will understand that at least one stored instruction constitutes a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0160] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0164] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0165] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for controlling piezoelectric jet dispensing, characterized in that, include: Acquire environmental data for piezoelectric jet dispensing, including driving air pressure, nozzle cone angle, nozzle outlet inner diameter, and nozzle inner clearance; The normalized vector corresponding to the environmental data is used as the first input and fed into the input layer of the first branch. After passing through the first hidden layer and the attention mechanism layer, the output of the first branch is obtained. Acquire control data for piezoelectric jet dispensing, including: current target displacement command value, actual measured displacement value, current applied voltage value, voltage change rate, historical displacement sequence average value, working frequency, and cumulative working time; send the normalized vector of the control data as the second input to the input layer of the second branch, and obtain the output of the second branch through the second hidden layer; The outputs of the first branch and the second branch are concatenated to obtain h. 1j Based on h 1j Using the network model, a 3-dimensional vector Y=[y1, y2, y3] is finally obtained, which represents the compensation voltage increment ΔV, the predicted displacement error Δx, and the system stability index S, respectively. Control of piezoelectric jet dispensing is achieved based on compensation voltage increment, prediction displacement error, and system stability indicators.

2. The piezoelectric jet dispensing control method according to claim 1, characterized in that, In the process of controlling piezoelectric jet dispensing by compensating for voltage increment, predicting displacement error, and system stability indicators, the compensating voltage increment is used to compensate for the voltage compensation amount to offset the hysteresis effect of piezoelectric ceramics, and the predicting displacement error is used to compensate for the difference between the target displacement and the actual displacement.

3. The piezoelectric jet dispensing control method according to claim 2, characterized in that, In the process of controlling piezoelectric jet dispensing by compensating for voltage increments, predicting displacement errors, and evaluating system stability, the system stability index is used to assess the effectiveness of the control.

4. The piezoelectric jet dispensing control method according to claim 1, characterized in that, The first hidden layer contains 64 neurons and uses the ReLU activation function; the second hidden layer contains 64 neurons and uses the ReLU activation function.

5. The piezoelectric jet dispensing control method according to claim 1, characterized in that, The calculation process for the normalized value of the historical displacement sequence average in the normalized vector corresponding to the environmental data includes: For the historical displacement sequence average x 25 The displacement is normalized using the average displacement of the first three time points; the normalization formula is: ,in This represents the average displacement over the first three time points, which are time t, the time corresponding to t-2 milliseconds, and the time corresponding to t-4 milliseconds, respectively.

6. The piezoelectric jet dispensing control method according to claim 1, characterized in that, Based on h 1j In the process of obtaining the 3D vector Y=[y1, y2, y3] using the network model, h 1j The vector Y is fed into the third hidden layer and then passed through the output layer to obtain a 3D vector.

7. The piezoelectric jet dispensing control method according to claim 6, characterized in that, The third hidden layer contains 32 neurons and uses the ReLU activation function.

8. A method for controlling piezoelectric jet dispensing according to any one of claims 1 to 7, characterized in that, The network model used to obtain the 3D vector Y, including the first hidden layer, the attention mechanism layer, and the second hidden layer, is pre-trained. The total loss function during training is as follows: Where: L_voltage is the voltage compensation loss, using mean square error; L_displacement is the displacement error loss, using mean square error; L_stability is the stability index loss, using binary cross-entropy; L_reg is the regularization loss, calculated as the cumulative sum of the weight coefficients of the first and second hidden layers; , , λ is the adjustment weight coefficient of the total loss function.

9. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the piezoelectric jet dispensing control method according to any one of claims 1 to 8.

10. A control device for piezoelectric jet dispensing, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the piezoelectric jet dispensing control method according to any one of claims 1 to 8.

Images