Dispensing control method and system for discontinuous path

Abstract

The invention discloses a discontinuous path-oriented glue dispensing control method and system, and the method comprises the steps: obtaining a three-dimensional glue dispensing track containing at least one path discontinuity point, and recognizing track segments before and after the path discontinuity point, and recording the track segments as a front track segment and a rear track segment; based on the termination geometric feature of the front track section, the initial geometric feature of the rear track section and the spatial span of the path discontinuity point, the fracture and lap joint behaviors of the glue at the path discontinuity point are predicted, and a special compensation model used for controlling the forms of the glue line at the two ends of the path discontinuity point is established; and before the dispensing head moves to the path discontinuity point along the front track section, a first preset action sequence is executed according to the special compensation model. Intelligent closed-loop control over discontinuous path dispensing is achieved, the problem that the form of glue is out of control when the glue crosses a physical gap is effectively solved, and the dispensing quality and the technology self-adaptive capacity are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor manufacturing, in particular to a point control method and system for discontinuous paths. BACKGROUND

[0002] In the modern motor manufacturing field, precise glue coating of stator assembly is one of the key processes to ensure product structural stability and electrical insulation performance.

[0003] With the popularity of high-performance motors such as brushless DC motors and shielded motors, their stator structures are becoming increasingly complex, often containing core laminations, tooth slots, windings, and insulation skeletons, etc. multiple special-shaped structures, resulting in frequent physical barriers to the glue dispensing path, forming a large number of discontinuous path breakpoints.

[0004] Traditional continuous path glue dispensing methods have inherent limitations when facing such breakpoints. The glue often experiences uncontrolled stringing, dripping when crossing gaps, or difficulty in initial spreading after the breakpoint. This not only affects the appearance, but also can cause serious quality problems, such as contamination of precise electrical areas by excess glue, or increased vibration, stress concentration, and reduced structural strength due to poor glue overlap during high-speed motor operation, directly affecting the reliability, efficiency, and lifespan of the motor.

[0005] Therefore, there is an urgent need in the art for an automated glue dispensing technology that can intelligently adapt to such complex discontinuous paths, ensuring precise, stable, and reliable glue line overlap at each breakpoint. SUMMARY

[0006] To solve the above problems, an embodiment of the present application provides a point control method for discontinuous paths, the method comprising:

[0007] Obtaining a three-dimensional glue dispensing trajectory containing at least one path breakpoint, and identifying the trajectory segments before and after the path breakpoint, denoted as the front trajectory segment and the rear trajectory segment;

[0008] Based on the termination geometric features of the front trajectory segment, the starting geometric features of the rear trajectory segment, and the spatial span of the path breakpoint, predicting the breaking and overlapping behavior of the glue at the path breakpoint, and establishing a special compensation model for controlling the form of the glue line at both ends of the path breakpoint;

[0009] Before the glue dispensing head moves along the front trajectory segment to the path breakpoint, a first preset action sequence is executed according to the special compensation model to form a front glue line with a target ending shape; when the glue dispensing head crosses the path breakpoint and is about to start the rear trajectory segment, a second preset action sequence is executed according to the special compensation model to form a starting shape at the beginning of the rear trajectory segment that matches the ending shape of the front glue line;

[0010] Based on the physical data collected synchronously during the execution of the action sequence and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system, the actual shape is compared with the ideal shape predicted by the dedicated compensation model to generate performance error. Using the performance error, the parameters of the dedicated compensation model are fine-tuned through an incremental learning algorithm, and the process data of this round and the optimized model parameters are archived into the process knowledge base.

[0011] Furthermore, the dedicated compensation model outputs a decision by comprehensively analyzing the dynamic balance between the spatial span of the path discontinuity and the cohesive force and tensile viscosity of the colloid itself. The decision includes an air-drawing overlap mode and a substrate-reserved overlap mode. When the decision is the air-drawing overlap mode, the pulled filament is controlled to cross the discontinuity in a natural parabolic trajectory affected by air resistance and gravity, and its end landing point overlaps with the starting area of ​​the subsequent trajectory segment within a preset range. When the decision is the substrate-reserved overlap mode, the colloid head is controlled to form a colloid anchor point with a predetermined volume and shape at the end of the previous trajectory segment before reaching the path discontinuity. This anchor point serves as the physical basis for overlap at the beginning of the subsequent trajectory segment after crossing the path discontinuity.

[0012] Furthermore, the computational mechanism of the dedicated compensation model includes: taking the spatial vector of the path discontinuity point, the rheological parameters of the colloid, and the dynamic crossing speed of the dispensing head as inputs, and simultaneously calculating the amplitude and speed of the pull-back action, as well as the frequency characteristics and energy level of the high-frequency micro-amplitude vibration, through a built-in dynamic function that characterizes the viscoelastic properties of the colloid.

[0013] Furthermore, the spatial motion path followed by the dispensing head when crossing the discontinuity of the path is a three-dimensional spline curve dynamically generated based on the tension optimization target output by the dedicated compensation model. This curve ensures the tension stability and landing accuracy of the adhesive filament during the crossing process.

[0014] Furthermore, the performance error is decoupled into a fracture morphology error corresponding to the end shape of the front adhesive line and a spreading morphology error corresponding to the beginning shape of the rear adhesive line; the incremental learning algorithm selectively reinforces the sub-modules controlling the pull-back action and high-frequency micro-amplitude vibration in the dedicated compensation model with unequal weights based on the contribution of the fracture morphology error and the spreading morphology error.

[0015] Furthermore, the dedicated compensation model also uses the wetting characteristics of the substrate surface as one of the input parameters for dynamic decision-making; when the model identifies the substrate as a low surface energy material, the energy level of the high-frequency micro-amplitude vibration will be dynamically increased to overcome the energy barrier when the colloid begins to spread in the later trajectory segment.

[0016] Furthermore, after fine-tuning the parameters of the current dedicated compensation model, the incremental learning algorithm initiates a meta-optimization sub-process. This sub-process first encodes the geometric features, colloidal properties, and substrate wetting characteristics of the current path discontinuity into a unified feature vector. Subsequently, based on this feature vector, it retrieves K most similar historical process cases from the process knowledge base, where K is a preset positive integer. Then, by analyzing the correlation between the initial model parameters and the final optimized performance in the cases, it optimizes a globally shared model initialization strategy library. The optimized strategy library is used to directly generate a set of initial model parameters that are better than the default values ​​for new discontinuities encountered subsequently.

[0017] Furthermore, the frequency characteristics of the high-frequency micro-amplitude vibration are set within a specific range, which is designed to excite the colloid to produce a thixotropic effect, thereby achieving faster spreading and shaping at the beginning of the subsequent trajectory segment.

[0018] A dispensing control method for discontinuous paths, the method further includes:

[0019] Establish and maintain a local cache with a discontinuity compensation strategy.

[0020] A dispensing control system for discontinuous paths, the system comprising:

[0021] The trajectory parsing module acquires a three-dimensional dispensing trajectory containing at least one path discontinuity and identifies the trajectory segments before and after the path discontinuity, denoted as the front trajectory segment and the back trajectory segment.

[0022] The behavior modeling module predicts the breaking and overlapping behavior of the colloid at the path discontinuity based on the termination geometric features of the preceding trajectory segment, the starting geometric features of the following trajectory segment, and the spatial span of the path discontinuity, and establishes a dedicated compensation model for controlling the shape of the adhesive line at both ends of the path discontinuity.

[0023] The dynamic execution module executes a first preset action sequence according to the dedicated compensation model before the dispensing head moves along the front trajectory segment to the path discontinuity to form a front adhesive line with a target ending shape; when the dispensing head crosses the path discontinuity and is about to start the rear trajectory segment, it executes a second preset action sequence according to the dedicated compensation model to form a starting shape at the beginning of the rear trajectory segment that matches the ending shape of the front adhesive line.

[0024] The online evolution module compares the physical data collected synchronously during the execution of the action sequence and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system with the ideal shape predicted by the dedicated compensation model to generate performance error. Using the performance error, the dedicated compensation model is fine-tuned through an incremental learning algorithm, and the process data of this round and the optimized model parameters are archived to the process knowledge base.

[0025] The technical effects and advantages of the dispensing control method for discontinuous paths provided by this invention are as follows:

[0026] This invention achieves intelligent closed-loop control for dispensing along discontinuous paths, effectively solving the problem of morphological loss of the adhesive when crossing physical gaps, and significantly improving dispensing quality and process adaptability. Traditional methods can only passively handle consequences such as stringing and dripping when encountering path discontinuities. However, this invention, by establishing a dedicated compensation model, can predict the fracture and overlap behavior of the adhesive in advance and plan precise compensation actions, fundamentally avoiding the generation of quality defects. By dynamically generating crossing paths, coordinating dispensing actions, and introducing online learning and a knowledge base, the system can generate optimal strategies for discontinuities with different characteristics, ensuring stable and reliable adhesive line overlap at each discontinuity, thereby guaranteeing the structural strength and long-term reliability of the final product (such as a motor stator). It not only considers the geometric path but also incorporates complex factors such as the rheological properties of the adhesive and the surface energy of the substrate into the decision-making process, enabling the dispensing process to intelligently adapt to changes in materials and the environment, reducing debugging costs during object switching and dependence on operator experience. Attached Figure Description

[0027] Figure 1 This is a flowchart of a dispensing control method for discontinuous paths in Example 1;

[0028] Figure 2 This is a flowchart of a dispensing control method for discontinuous paths in Example 2;

[0029] Figure 3 This is a schematic diagram of a dispensing control system for discontinuous paths in Example 3. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1:

[0032] Please see Figure 1 As shown, an embodiment of the present invention provides a dispensing control method for discontinuous paths, characterized by comprising the following steps:

[0033] S1: Obtain a three-dimensional dispensing trajectory containing at least one path discontinuity, and identify the trajectory segments before and after the path discontinuity, denoted as the front trajectory segment and the back trajectory segment.

[0034] S2: Based on the termination geometry of the preceding trajectory segment, the starting geometry of the following trajectory segment, and the spatial span of the path discontinuity, predict the breaking and overlapping behavior of the colloid at the path discontinuity, and establish a dedicated compensation model for controlling the shape of the adhesive line at both ends of the path discontinuity.

[0035] S3: Before the dispensing head moves along the previous trajectory segment to the path discontinuity, a first preset action sequence is executed according to the dedicated compensation model to form a front adhesive line with a target ending shape; when the dispensing head crosses the path discontinuity and is about to start the subsequent trajectory segment, a second preset action sequence is executed according to the dedicated compensation model to form a starting shape at the beginning of the subsequent trajectory segment that matches the ending shape of the front adhesive line; wherein, the first preset action sequence includes at least reducing the dispensing amount before moving to the discontinuity and performing a rapid pull-back or lifting action at the precise discontinuity coordinates to actively break the adhesive thread and control its break point shape; the second preset action sequence includes at least establishing dispensing pressure before reaching the starting point of the subsequent trajectory segment and triggering a high-frequency micro-amplitude vibration at the instant of contact with the starting point;

[0036] S4: Based on the physical data (such as pressure, position, and vibration feedback) collected synchronously during the execution of the action sequence in S3, and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system, it is compared with the ideal shape predicted by the model in S2 to generate performance error. Using the performance error, the parameters of the dedicated compensation model are fine-tuned through an incremental learning algorithm. The process data of this round and the optimized model parameters are archived into the process knowledge base to complete an autonomous performance evolution for this specific discontinuity condition.

[0037] The dedicated compensation model in S2 is not a static lookup table, but an intelligent decision-making subsystem that can comprehensively analyze path geometry, material properties, and dynamic processes.

[0038] The dedicated compensation model outputs the final overlap mode decision by comprehensively analyzing the dynamic balance between the spatial span of the path discontinuity and the cohesive force and tensile viscosity of the colloid itself. The dynamic balance here essentially refers to the competition between the colloid's internal resistance to breakage and the external tensile force that causes it to break when it is stretched. This competitive relationship directly determines the yield rate of the motor stator dispensing.

[0039] The spatial span of a path discontinuity refers to the straight-line distance and spatial vector between the end point of the previous trajectory segment and the starting point of the subsequent trajectory segment in three-dimensional space. This is a key geometric constraint.

[0040] The cohesive force of a colloid is the force that attracts molecules within the liquid. It causes the liquid surface to contract and resists the liquid from being broken apart.

[0041] The tensile viscosity of a colloid describes its ability to resist flow deformation when subjected to tensile stress. Colloids with high tensile viscosity are more likely to form long and stable filaments.

[0042] The dedicated compensation model simulates a scenario where, when the dispensing head attempts to cross a given spatial span, the adhesive can be stretched into an unbroken thread and successfully overlapped to the other end. If it can, the air-stretching overlap mode is preferred; if not, the base-reserved overlap mode is required.

[0043] Aerial wire-drawing and overlapping patterns include:

[0044] When the dedicated compensation model determines that the cohesive force and tensile viscosity of the colloid are sufficient to support it across the current discontinuity spatial span, it will decide to adopt the aerial filament splicing mode.

[0045] In this mode, the system precisely controls the deceleration of the dispensing head before the discontinuity, the non-linear reduction of the dispensing volume, and the final lifting motion. The purpose of this series of actions is to actively guide the adhesive to break and control the break point to form a sharp end. Simultaneously, the dispensing head crosses the discontinuity at a specific speed trajectory, causing the pulled adhesive thread to form a natural parabolic trajectory under the combined influence of air resistance and gravity. This trajectory is not randomly formed but is pre-calculated through fluid dynamics simulation to ensure that the end point of the adhesive thread overlaps with the starting area of ​​the subsequent trajectory segment within a preset range. This "preset range" is a relative concept; for example, it can be set to 50% to 150% of the adhesive thread width to ensure sufficient contact area for reliable electrical or mechanical connections.

[0046] For example: When applying sealant to the stator slot of a brushless DC motor, the adhesive path needs to cross from the side of one stator tooth to the adjacent tooth. Assuming the spatial span here is 1.5 mm, the dedicated compensation model determines that it meets the condition of "air stringing" based on this span and the parameters of the epoxy sealant used. The system then executes a predetermined action: before reaching the discontinuity point, the dispensing head nonlinearly reduces the amount of adhesive dispensed and performs a rapid pull-back action at the precise coordinates. This causes the adhesive to be drawn into a thin and uniform thread, which successfully overlaps with the starting application position of the adjacent stator tooth with a natural parabolic trajectory. This mode can achieve efficient and clean crossing.

[0047] The base pre-reserved overlap mode includes:

[0048] When the dedicated compensation model determines that the current spatial span or process conditions are unsuitable for aerial wire drawing—for example, when it needs to cross the top of a motor winding coil or the short-circuit ring fixing post of a shaded-pole motor, where the spatial span may reach 3 mm or more—the model will decide on a substrate pre-reserved overlap mode. The core strategy of this mode is to abandon the high-risk aerial wire drawing and instead ensure the reliability of the connection by increasing the volume of the substrate adhesive. The system will not attempt to allow the adhesive to cross in the air, but will begin preparing for the "ground" connection when the previous trajectory segment is about to end but has not yet reached the discontinuity. Specifically, the model instructs the dispensing head to form a local adhesive anchor point with a predetermined volume and shape at the end of the previous trajectory segment through a short pause and a slight increase in adhesive. This adhesive anchor point acts similarly to the pier of a dock or bridge, serving as a reliable physical foundation. After the dispensing head crosses the discontinuity (at which point there is no need for stringing, it may move directly in the air or be lifted), it begins to dispense adhesive at the beginning of the next trajectory segment, ensuring that the new adhesive line fully contacts and merges with the previously reserved adhesive anchor point, thus physically achieving the continuity of the adhesive line.

[0049] Example: On the stator core of a shaded-pole motor, there is a glue path that needs to bypass the short-circuit ring mounting bracket. Before the glue path is interrupted by the mounting bracket, the model instructs the dispensing head to form a glue anchor point with a predetermined volume and shape at the end of the path. This anchor point serves as a solid physical foundation. The dispensing head then rises and safely passes over the mounting bracket, falls again on the other side to begin dispensing, and fully integrates the new glue line with the reserved glue anchor point.

[0050] The computational mechanism of the dedicated compensation model includes: taking the spatial vector of the discontinuity, the rheological parameters of the colloid, and the dynamic crossing speed of the dispensing head as inputs, and simultaneously calculating the amplitude and speed of the pull-back action, as well as the frequency characteristics and energy level of the high-frequency micro-amplitude vibration, through a built-in dynamic function that characterizes the viscoelastic properties of the colloid.

[0051] This computational mechanism ensures that, under the selected overlap mode, every micro-action can perfectly match the real-time state and spatial geometry of the colloid. The computation of the dedicated compensation model is not performed in isolation, but is closely dependent on a multi-dimensional set of inputs, which together define the current dispensing conditions.

[0052] The spatial vector of a discontinuity is not just a distance scalar, but a three-dimensional vector that includes direction. For example, in the stator of a motor, an upward spatial vector (such as crossing a winding) and a horizontal spatial vector (such as crossing a core tooth slot) require drastically different control strategies even if the distances are the same, because the direction of gravity changes in the process.

[0053] The rheological parameters of colloids are key physical properties that describe the deformation and flow behavior of colloids. They mainly include viscosity (the ability to resist flow) and elasticity (the ability to recover after deformation). These parameters together constitute the viscoelastic properties of colloids, which determine the response of colloids when stretched (pull-back action) and forced to spread (high-frequency micro-amplitude vibration).

[0054] The dynamic crossing speed of the dispensing head refers to the speed at which the dispensing head moves through the discontinuity area of ​​the path. This speed directly affects the stretching rate applied to the adhesive and is one of the core dynamic variables that determine the breakage shape and length of the adhesive filament.

[0055] The core of the dedicated compensation model is a built-in dynamic function that characterizes the viscoelastic properties of the colloid. This function can be understood as a "digital brain" trained with a large amount of data and physical laws. It can simulate the dynamic behavior of the colloid under the above input conditions. Its core task is to perform synchronous calculation, that is, to calculate the control commands of multiple key actuators in one go and in a coordinated manner.

[0056] The model outputs a profile of the "pullback action" rather than a single pullback distance. Here, the profile refers to a sequence of instructions that changes over time. It precisely defines the acceleration, constant velocity phase, and deceleration of the pullback action, thus ensuring that the breakage process of the colloid is "clean" rather than "sluggish." A steep, high-speed pullback action profile is suitable for low-viscosity, easily flowing colloids to quickly cut the filaments. For highly elastic colloids, a profile with short pauses may be needed to fully utilize their elastic recoil.

[0057] The model calculates the parameters of "high-frequency micro-amplitude vibration". While calculating, the model also calculates the vibration parameters for the second preset action sequence, including frequency characteristics (such as the main frequency value and its harmonic distribution) and energy level (such as amplitude or driving power). The core function of high-frequency micro-amplitude vibration is to use vibration energy to instantly reduce the apparent viscosity of the colloid, so that it can spread out quickly when it comes into contact with the substrate, forming a perfect starting shape, and seamlessly integrate with the end of the previous adhesive line or the reserved adhesive anchor point.

[0058] Example: Consider a scenario where adhesive is applied across two adjacent teeth of a brushless DC motor stator.

[0059] Assuming the spatial vector of the discontinuity is a horizontal vector with a length of 2 mm, and the adhesive used is a silicone adhesive with a certain degree of elasticity, after receiving these inputs, the dedicated compensation model begins to calculate its internal dynamic function. It may calculate a pull-back profile as follows: within 5 milliseconds, it first starts with an acceleration of 50 mm / s², and then pulls back 0.4 mm at a speed of 100 mm / s in a very short time. This specific combination can ensure that when the silicone adhesive is pulled apart, it forms a thin and strong filament. At the same time, the model will calculate vibration parameters for high-frequency micro-amplitude vibration with a main frequency of 800 Hz and an amplitude of 10 micrometers. These parameters can effectively overcome the elasticity of the adhesive, so that when it starts on the surface of the stator teeth, it immediately forms a uniform and well-wetted adhesive line.

[0060] Through this highly integrated computing mechanism, the present invention achieves precise control of complex fluids in dynamic processes.

[0061] In S3, the spatial motion path followed by the dispensing head when crossing path discontinuities is a three-dimensional spline curve dynamically generated based on the tension optimization target output by the dedicated compensation model. This curve ensures the tension stability and landing accuracy of the adhesive filament during the crossing process.

[0062] In the air-drawing overlap mode, the entire process from the dispensing head leaving the pre-trajectory segment to arriving at the starting point of the post-trajectory segment is the key stage that determines the success or failure of the overlap; while the dedicated compensation model makes decisions and outputs action sequence parameters, it also outputs a key physical target, namely the tension optimization target.

[0063] The goal of tension optimization is to maintain a relatively stable and optimal internal tension during the stretching of the rubber filament. If the tension is too low, the filament may sag excessively due to gravity and touch the underlying structure (such as motor windings), causing contamination. If the tension is too high, the filament may break before reaching its landing point. The goal is to find a balance point that prevents the filament from breaking or loosening.

[0064] To achieve the above goals, the model calculates and generates a smooth spatial motion path in real time for the dispensing head to execute. This path is usually composed of a three-dimensional spline curve, which is a very smooth curve defined by control points. Its advantage is that it can achieve smooth acceleration and velocity transitions, avoiding additional impact on the glue filament due to the sudden stop and start of the dispensing head.

[0065] This dynamically generated 3D spline curve is not an arbitrary trajectory; it possesses two core physical guarantees:

[0066] By carefully designing the curvature variation of the curve, the speed of the dispensing head during the crossing process can be matched with the rheological properties of the adhesive, thereby ensuring that the tension inside the adhesive filament is maintained within a preset optimal range; stable tension is a prerequisite for forming a uniform adhesive filament diameter and precisely controlling its parabolic shape.

[0067] The endpoint of the curve in space is strictly constrained within the starting area of ​​the subsequent trajectory segment. By comprehensively considering the dynamic crossing speed of the dispensing head, the gravitational acceleration, and the influence of air resistance, the model calculates the path that the dispensing head should follow in reverse, so that when the dispensing head reaches the endpoint, the adhesive filament pulled out behind it can accurately land at the expected overlapping position under the action of gravity and inertia.

[0068] For example, in a scenario where the path needs to cross the stator windings of a motor, the spatial span of the path discontinuity is 3 mm, with a 2 mm drop. If the dispensing head simply moves along a straight line, the adhesive filament will be rapidly stretched due to gravitational acceleration, causing a sudden increase in tension and making it prone to breakage. In this case, a dedicated compensation model calculates, based on the tensile viscosity of the adhesive, a strategy of "rapidly lifting to accumulate the adhesive filament, then slowly lowering to release it." Based on this, the model generates a specific three-dimensional spline curve: the dispensing head first moves upward at a relatively steep angle of about 1 mm, which helps buffer the tension and prevent instantaneous breakage; then, it moves downward along a smooth curve to the starting point of the subsequent trajectory segment. This path ensures that the tension of the adhesive filament changes smoothly throughout the crossing process, and its end ultimately lands accurately in the preset overlap area, with an error controlled within ±0.05 mm, thus achieving a highly reliable aerial overlap.

[0069] By intelligently optimizing the crossing path itself, this invention ensures that the overlapping behavior of adhesive filaments is highly controllable and predictable even in complex three-dimensional space, thereby significantly improving the yield and connection reliability of dispensing for precision components such as brushless DC motors.

[0070] In S4, the performance error is decoupled into the fracture morphology error corresponding to the end shape of the front adhesive line and the spreading morphology error corresponding to the beginning shape of the rear adhesive line. The incremental learning algorithm performs selective reinforcement learning on the sub-modules in the dedicated compensation model that control the pull-back action and the high-frequency micro-amplitude vibration respectively, based on the contribution of the fracture morphology error and the spreading morphology error, with unequal weights.

[0071] In step S4, after comparing the actual 3D morphology of the adhesive line acquired by the system with the ideal morphology predicted by the model, the resulting performance error is not a single evaluation metric. Based on a deep understanding of the adhesive line morphology, the system intelligently decouples the total error into two independent error components, each linked to a core action:

[0072] Breakage morphology error: This error component specifically evaluates the deviation between the end shape of the front rubber line and the target. For example, the ideal end may be a sharp break, while the actual result may be a trailing "tail". This error mainly reflects the control accuracy of the first preset action sequence (especially the pull-back action).

[0073] Spreading shape error: This error component specifically evaluates the deviation between the initial shape of the backing line and the target. For example, the ideal starting point should be a rounded "meniscus" shape, but in reality it may be an irregular sphere. This error mainly reflects the effectiveness of the second preset action sequence (especially high-frequency micro-amplitude vibration).

[0074] Through this decoupling, the system can diagnose the main problem of the current dispensing issue, namely whether the glue filament is not broken cleanly enough or the initial spreading is not sufficient.

[0075] After obtaining the two error components, the incremental learning algorithm implements a selective reinforcement learning strategy. The core logic of this strategy is to optimize the sub-modules controlling the pull-back action and the high-frequency micro-amplitude vibration in the dedicated compensation model with unequal weights based on the contribution of the fracture morphology error and the spreading morphology error. This means that if the fracture morphology error is dominant (i.e., the ending shape is very poor, but the starting shape is acceptable), the algorithm will determine that the problem lies in the breakage of the rubber strand. Therefore, it will allocate the optimization focus and more weight adjustment share to the part of the model responsible for solving the amplitude and velocity profile of the pull-back action. For example, it may significantly increase the initial acceleration weight of the pull-back action to pursue a crisper breakage. Conversely, if the spreading morphology error is dominant (i.e., the ending shape is good, but the starting shape is a mess), the algorithm will focus on optimizing the frequency characteristics and energy level of the high-frequency micro-amplitude vibration in the model, such as trying to increase the vibration frequency to better overcome the elasticity of the colloid.

[0076] For example: In a motor stator dispensing application, the vision system detected a significant defect in the glue line after crossing the boundary. Analysis revealed that the performance error was decoupled: the fracture morphology error was small, indicating a basically successful pull-back action; however, the spreading morphology error was very large, indicating a poor glue line initiation. Based on this, the system determined that the current problem was mainly caused by the poor performance of the second preset action sequence. Therefore, in this optimization loop, the incremental learning algorithm allocated 90% of the learning weight to the vibration control submodule, significantly calibrating its internal parameters, while only fine-tuning the pull-back action submodule. After several production cycles of iteration, the system was specifically enhanced for this type of "difficult spreading" condition, enabling it to directly output better vibration parameters when encountering similar discontinuities, thereby quickly eliminating such defects.

[0077] In S2, the dedicated compensation model also uses the wetting characteristics of the substrate surface as one of the input parameters for dynamic decision-making; when the model identifies the substrate as a low surface energy material, the energy level of high-frequency micro-amplitude vibrations will be dynamically increased to overcome the energy barrier when the colloid begins to spread in the later trajectory segment.

[0078] In practical dispensing applications, the final morphological quality of the adhesive line is greatly limited by the interaction between the adhesive and the substrate material. The wetting characteristics of the substrate surface, in layman's terms, describe the ease with which a liquid adhesive spreads and adheres to a solid substrate surface. It is a physical concept that integrates parameters such as surface tension and surface energy.

[0079] For high surface energy materials (such as clean metals and glass), the colloid is easy to spread, forming a low contact angle and a wide and flat colloid cross section.

[0080] For low surface energy materials (such as many plastics, fluorinated coatings, or substrates with antifouling coatings), the adhesive tends to "shrink" into a spherical shape, making it difficult to spread out, forming a high contact angle, a narrow and high adhesive line cross-section, and poor adhesion.

[0081] In step S2, the model obtains the substrate wetting characteristics parameters of the current dispensing area through a pre-stored material database or an integrated surface sensing module.

[0082] When the model identifies the substrate as a low surface energy material, it determines that the colloid will face a higher energy barrier when it begins to spread in the later trajectory segment. Here, energy barrier is a physicochemical term that refers to the energy barrier that colloidal molecules need to overcome to transition from the initial state to a stable spreading state on the substrate.

[0083] To overcome this obstacle, the model dynamically adjusts the strategy of the second preset action sequence. The core measure is to increase the energy level of the high-frequency micro-amplitude vibration. The energy level here is a comprehensive index, which is usually positively correlated with the amplitude and power of the vibration. The purpose of increasing the energy level is to give the colloidal molecules higher energy through more intense vibration, so that they have enough power to break through the repulsion between the colloidal molecules and the low surface energy substrate. Higher vibration energy can "smash" the gas film or impurities that hinder the contact between the colloidal molecules and the substrate, forcing the colloidal molecules to achieve close molecular-level contact with the substrate in an instant.

[0084] For example: Suppose that dispensing needs to be performed on a low surface energy engineering plastic (such as LCP) commonly used in electronic components. Its water contact angle is greater than 90 degrees, making it a typical difficult-to-wet material. If high-frequency micro-amplitude vibration parameters (e.g., amplitude of 5 micrometers) for metal substrates are used, the starting end of the adhesive line will form an irregular spherical shape due to its inability to spread, resulting in overlapping failure. In this case, after receiving the input that "the substrate is LCP material", the dedicated compensation model will immediately calculate that the vibration energy level needs to be increased by about 30% (e.g., by increasing the amplitude to 6.5 micrometers). This adjustment allows the adhesive to obtain enough extra energy to overcome its high energy barrier at the moment of contact with the substrate, thereby forming a relatively flat starting end with good contact with the substrate, laying a solid foundation for the success of subsequent aerial wire overlapping or pre-set adhesive anchor point overlapping.

[0085] In S4, the incremental learning algorithm is responsible for fine-tuning the dedicated compensation model to adapt to the current task. However, this is only the first layer of system learning. The meta-optimization subprocess launched after this is not aimed at directly optimizing the current process, but at improving the system's initial response capability when facing unforeseen types of discontinuities in the future, i.e., rapid generalization capability.

[0086] This sub-process is a systematic process of experience distillation, including:

[0087] First, all key contextual information of the current discontinuity, including its geometric features (such as span length and turning angle), colloidal properties (such as viscosity and rheology), and substrate wetting characteristics, is standardized and encoded into a unified feature vector. This vector acts as a "digital fingerprint" of the process problem, uniquely defining its challenge. The system then uses this "digital fingerprint" as a query condition to retrieve a certain number (e.g., K=5) of the most similar historical optimization cases in the feature space from the ever-expanding process knowledge base. Next comes the core "learning" step, where the algorithm deeply analyzes this set of similar cases to find a key pattern: from the initial model parameter settings to the final optimized performance. At the peak, what kind of mapping relationship exists? For example, analysis might reveal that for a certain type of corner with specific geometric characteristics, if the energy parameter of high-frequency micro-amplitude vibration in the initial model is preset to be a certain proportion higher than the conventional value, the number of iterations required to converge to the optimal solution will be significantly reduced. Ultimately, the rules obtained from the above analysis are used to optimize a globally shared model initialization strategy library. This strategy library is a more advanced, empirical "scheduling center." Its role is that when the system encounters a new type of discontinuity again, even if it has never been directly processed, the system can first generate a feature vector for it, and then use this optimized strategy library to directly output a set of initial model parameters that is better than the default value.

[0088] For example: Suppose the system first processes a large-span right-angle turn discontinuity on a low surface energy plastic. After several rounds of iteration, the incremental learning algorithm finds the optimal parameters. Subsequently, the meta-optimization subprocess is launched, which finds that the success of this optimization is largely due to the appropriate increase in vibration energy at the beginning. It strengthens and stores the association between the feature combination of "low surface energy + large-span right angle" and the successful strategy of "initial high-energy vibration" in the model initialization strategy library. A week later, a large-span acute-angle turn appears on the production line on another low surface energy coating. Although the specific parameters are different, its feature vector is highly similar to the previous case in the two dimensions of "low wettability" and "large span". At this time, the system no longer starts from the general default parameters, but directly calls the strategy library to recommend a parameter set with a higher initial energy level for this new discontinuity. As a result, the system achieves satisfactory results after only one round of fine-tuning, while it may take four to five rounds if starting from scratch.

[0089] In S3, the frequency characteristics of the high-frequency micro-amplitude vibration are set within a specific range, which is designed to excite the colloid to produce a thixotropic effect, thereby achieving faster spreading and shaping at the beginning of the subsequent trajectory segment. The thixotropic effect is the key physical principle for achieving precise control in this invention. It describes the characteristic that the apparent viscosity of a fluid (i.e., the ease of flow) temporarily decreases with the increase of the shear rate (such as stirring or vibration), and gradually recovers the viscosity after the shearing action stops.

[0090] In this method, a high-frequency shearing action is applied to the colloid by applying high-frequency micro-amplitude vibration, which causes the apparent viscosity of the colloid to decrease significantly in an instant, thereby enhancing its fluidity.

[0091] The frequency characteristics of the high-frequency micro-amplitude vibration are set within a specific range of 100 Hz to 2000 Hz (example). This range is set based on the following multiple considerations:

[0092] Setting a lower limit (e.g., 100 Hz): If the frequency is too low, the thixotropic effect cannot be effectively excited. Only when the vibration frequency is high enough can a continuous and rapid impact be formed on the internal structural network of the colloid, preventing it from rebuilding in time, thereby maintaining a low viscosity and easy-to-spread state. Below this lower limit, the shearing action is discontinuous, the viscosity of the colloid does not decrease significantly, and it is difficult to overcome the spreading energy barrier.

[0093] Setting an upper limit (e.g., 2000 Hz): Too high a frequency can cause a series of problems. First, too high a frequency may cause vibration energy to be dissipated as heat rather than being effectively used to change the colloidal structure. Second, it may cause the colloidal material to splash or generate unnecessary standing waves, which can disrupt the morphological consistency of the adhesive lines. In addition, from an engineering perspective, generating and controlling micro-amplitude vibrations at extremely high frequencies requires demanding equipment and is not cost-effective.

[0094] Therefore, the range of 100 Hz to 2000 Hz is usually an optimized range that can reliably and efficiently excite colloidal thixotropic effects while avoiding side effects.

[0095] At the critical moment of initial spreading in the later trajectory segment, the dispensing needle vibrates at this specific frequency; due to the instantaneous decrease in viscosity of the adhesive, its flow resistance is reduced, and the spreading speed on the substrate surface is greatly increased, which helps to quickly form a stable and wide overlapping substrate.

[0096] When the needle is removed and the vibration stops, the thixotropic effect allows the viscosity of the colloid to recover quickly. This rapid "re-thickening" characteristic enables the adhesive line to quickly fix its shape after it is spread, preventing it from flowing or shrinking due to excessive gravity or surface tension, thus ensuring the precision and stability of the geometry of the overlap point.

[0097] For example, take an epoxy resin adhesive used for filling the bottom of a chip as an example. This adhesive has significant thixotropic properties. When air-stretching and overlapping are required, the dispensing needle is turned on with a high-frequency micro-vibration of 800 Hz and an amplitude of 10 micrometers. At the moment of contact with the substrate, the viscosity of the adhesive drops sharply under the action of vibration, allowing it to spread rapidly to the target width within 50 milliseconds. Subsequently, the vibration stops, and the viscosity of the adhesive returns to near its initial state within 100 milliseconds, "locking" the overlap point in place. This perfectly avoids the risk of collapse or bridging short circuit caused by the adhesive being too thin.

[0098] Example 2:

[0099] like Figure 2 As shown, this embodiment further improves the design based on embodiment one. The difference is that in actual operation of embodiment one, it was found that when there are multiple high-density, rapidly alternating discontinuities in the dispensing path, the computational and decision-making load of the system will increase sharply during the continuous execution of the "perception-decision-compensation-learning" loop from S3 to S4, resulting in system response delay. The physical movement of the dispensing equipment has to be paused to wait for the decision result, and truly smooth, uninterrupted, and efficient production cannot be achieved. Based on this, a dispensing control method for discontinuous paths also includes S5: establishing and maintaining a local cache of a discontinuity compensation strategy.

[0100] This step is activated immediately after S1 and runs asynchronously, independent of the main control loop (S2-S4). Its specific flow is as follows:

[0101] When the system is idle or has sufficient computing resources, a "virtual planning" process is pre-executed. This process identifies all planned dispensing points in advance based on the complete dispensing path obtained from the upstream CAD / CAM system. For each identified dispensing point, the system will simulate and execute the core steps of S3 and S4 in advance. That is, based on the known characteristics of the dispensing point, the corresponding dedicated compensation model and initial parameters are pre-loaded or quickly generated from the process knowledge base and model initialization strategy library. These "pre-compiled" compensation strategies are packaged into a lightweight strategy data package and stored in a local cache of the device controller.

[0102] When the dispensing equipment reaches this breakpoint in actual production, its control logic will no longer rely entirely on real-time, high-load model calculations. Instead, it will prioritize querying this local cache and directly calling the prepared strategy data package, which greatly reduces the calculation delay during critical process windows.

[0103] Meanwhile, the system will maintain the real-time performance of the cache. If, during the incremental learning process of S4, a significant optimization update is made to the compensation model for a certain discontinuity, the system will synchronously update the corresponding policy data package in the cache to ensure that the latest and optimal policy is used in subsequent production.

[0104] Example 3:

[0105] like Figure 3 As shown, based on the same inventive concept as the dispensing control method for discontinuous paths in the foregoing embodiments, this application provides a dispensing control system for discontinuous paths. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0106] The trajectory parsing module acquires a three-dimensional dispensing trajectory containing at least one path discontinuity and identifies the trajectory segments before and after the path discontinuity, denoted as the front trajectory segment and the back trajectory segment.

[0107] The behavior modeling module predicts the breaking and overlapping behavior of the colloid at the path discontinuity based on the termination geometric features of the preceding trajectory segment, the starting geometric features of the following trajectory segment, and the spatial span of the path discontinuity. It also establishes a dedicated compensation model for controlling the shape of the adhesive line at both ends of the path discontinuity.

[0108] The dynamic execution module executes a first preset action sequence according to the dedicated compensation model before the dispensing head moves along the front trajectory segment to the path discontinuity to form a front adhesive line with a target ending shape; when the dispensing head crosses the path discontinuity and is about to start the rear trajectory segment, it executes a second preset action sequence according to the dedicated compensation model to form a starting shape at the beginning of the rear trajectory segment that matches the ending shape of the front adhesive line.

[0109] The online evolution module compares the physical data collected synchronously during the execution of the action sequence and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system with the ideal shape predicted by the dedicated compensation model to generate performance error. Using the performance error, the dedicated compensation model is fine-tuned through an incremental learning algorithm, and the process data of this round and the optimized model parameters are archived to the process knowledge base.

[0110] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0111] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. A dispensing control method for discontinuous paths, characterized in that, The methods include: Obtain a three-dimensional dispensing trajectory containing at least one path discontinuity, and identify the trajectory segments before and after the path discontinuity, denoted as the front trajectory segment and the back trajectory segment. Based on the termination geometry of the preceding trajectory segment, the starting geometry of the following trajectory segment, and the spatial span of the path discontinuity, the breaking and overlapping behavior of the colloid at the path discontinuity is predicted, and a special compensation model is established to control the shape of the adhesive line at both ends of the path discontinuity. Before the dispensing head moves along the front trajectory segment to the path discontinuity, a first preset action sequence is executed according to the dedicated compensation model to form a front adhesive line with a target ending shape. When the dispensing head crosses the path discontinuity and is about to start the subsequent trajectory segment, a second preset action sequence is executed according to the dedicated compensation model to form a starting shape at the beginning of the subsequent trajectory segment that matches the ending shape of the previous adhesive line. Based on the physical data collected synchronously during the execution of the action sequence and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system, the actual shape is compared with the ideal shape predicted by the dedicated compensation model to generate performance error. Using the performance error, the parameters of the dedicated compensation model are fine-tuned through an incremental learning algorithm, and the process data of this round and the optimized model parameters are archived into the process knowledge base.

2. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The dedicated compensation model outputs a decision by comprehensively analyzing the dynamic balance between the spatial span of the path discontinuity and the cohesive force and tensile viscosity of the colloid itself. The decision includes an air-drawing overlap mode and a substrate-reserved overlap mode. When the decision is the air-drawing overlap mode, the drawn filament is controlled to cross the discontinuity in a natural parabolic trajectory affected by air resistance and gravity, and its end landing point overlaps with the starting area of ​​the subsequent trajectory segment within a preset range. When the decision is the substrate-reserved overlap mode, the colloid head is controlled to form a colloid anchor point with a predetermined volume and shape at the end of the previous trajectory segment just before reaching the path discontinuity. This anchor point serves as the physical basis for overlap at the beginning of the subsequent trajectory segment after crossing.

3. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The computational mechanism of the dedicated compensation model includes: taking the spatial vector of the path discontinuity, the rheological parameters of the colloid, and the dynamic crossing speed of the dispensing head as inputs, and simultaneously calculating the amplitude and speed of the pull-back action, as well as the frequency characteristics and energy level of the high-frequency micro-amplitude vibration, through a built-in dynamic function that characterizes the viscoelastic properties of the colloid.

4. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The spatial motion path followed by the dispensing head when crossing the discontinuity of the path is a three-dimensional spline curve dynamically generated based on the tension optimization target output by the dedicated compensation model. This curve ensures the tension stability and landing accuracy of the adhesive filament during the crossing process.

5. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The performance error is decoupled into a fracture morphology error corresponding to the end shape of the front adhesive line and a spreading morphology error corresponding to the beginning shape of the rear adhesive line. The incremental learning algorithm performs selective reinforcement learning on the sub-modules that control the pull-back action and the high-frequency micro-amplitude vibration in the dedicated compensation model with unequal weights based on the contribution of the fracture morphology error and the spreading morphology error.

6. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The dedicated compensation model also uses the wetting characteristics of the substrate surface as one of the input parameters for dynamic decision-making; when the model identifies the substrate as a low surface energy material, the energy level of high-frequency micro-amplitude vibration will be dynamically increased to overcome the energy barrier when the colloid begins to spread in the later trajectory segment.

7. The dispensing control method for discontinuous paths according to claim 1, characterized in that, After fine-tuning the parameters of the current dedicated compensation model, the incremental learning algorithm initiates a meta-optimization sub-process. This sub-process first encodes the geometric features, colloidal properties, and substrate wetting characteristics of the current path discontinuity into a unified feature vector. Subsequently, based on this feature vector, it retrieves the K most similar historical process cases from the process knowledge base, where K is a preset positive integer. Then, by analyzing the correlation between the initial model parameters and the final optimized performance in the cases, it optimizes a globally shared model initialization strategy library. The optimized strategy library is used to directly generate a set of initial model parameters that are better than the default values ​​for new discontinuities encountered subsequently.

8. The dispensing control method for discontinuous paths according to claim 3, characterized in that, The frequency characteristics of the high-frequency micro-amplitude vibration are set within a specific range, which is designed to excite the colloid to produce a thixotropic effect, thereby achieving faster spreading and shaping at the beginning of the latter trajectory segment.

9. The dispensing control method for discontinuous paths according to claim 1, characterized in that, The method also includes: Establish and maintain a local cache with a discontinuity compensation strategy.

10. A dispensing control system for discontinuous paths, characterized in that, The system includes: The trajectory parsing module acquires a three-dimensional dispensing trajectory containing at least one path discontinuity and identifies the trajectory segments before and after the path discontinuity, denoted as the front trajectory segment and the back trajectory segment. The behavior modeling module predicts the breaking and overlapping behavior of the colloid at the path discontinuity based on the termination geometric features of the preceding trajectory segment, the starting geometric features of the following trajectory segment, and the spatial span of the path discontinuity, and establishes a dedicated compensation model for controlling the shape of the adhesive line at both ends of the path discontinuity. The dynamic execution module executes a first preset action sequence according to the dedicated compensation model before the dispensing head moves along the front trajectory segment to the path discontinuity to form a front adhesive line with a target ending shape; when the dispensing head crosses the path discontinuity and is about to start the rear trajectory segment, it executes a second preset action sequence according to the dedicated compensation model to form a starting shape at the beginning of the rear trajectory segment that matches the ending shape of the front adhesive line. The online evolution module compares the physical data collected synchronously during the execution of the action sequence and the actual three-dimensional shape of the adhesive line obtained by the visual inspection system with the ideal shape predicted by the dedicated compensation model to generate performance error. Using the performance error, the dedicated compensation model is fine-tuned through an incremental learning algorithm, and the process data of this round and the optimized model parameters are archived to the process knowledge base.

Images