Intelligent gluing control method and system based on large model

Abstract

The invention relates to the technical field of deep learning, in particular to an intelligent gluing control method and system based on a large model, and the method comprises the steps: obtaining the spectral data of initial glue extruded by a glue gun of a gluing robot before the initial glue contacts with a plastic base; obtaining a material physical property deviation of the initial glue based on the spectral data through a first large model, and triggering a gluing robot to execute a first gluing control strategy when the material physical property deviation exceeds a first threshold value; obtaining time sequence process data in a gluing process and spatial form data of a glued channel section, and discretizing the spatial form data into a graph structure; and obtaining a second gluing control strategy through the second large model based on the material physical property deviation, the time sequence process data and the graph structure, and triggering the gluing robot to execute the second gluing control strategy. Therefore, two-stage intelligent glue coating control can be carried out in the glue coating starting stage and the glue coating process, the glue overflowing problem can be quickly responded, and complex glue overflowing scenes can be accurately dealt with.

Description

Technical Field

[0001] This disclosure relates to the field of deep learning technology, specifically to an intelligent glue application control method and system based on a large model. Background Technology

[0002] The washing machine door consists of a plastic base and a glass decorative cover, glued together. During manufacturing, a glue-applying robot first applies a ring of glue to the plastic base using a glue gun. Then, the glass decorative cover is placed on the plastic base, and finally, pressure is applied to the glass cover to complete the assembly of the washing machine door. However, there is a risk of glue overflow during the glue-applying process. For example… Figure 1 As shown, the starting point of the ring-shaped adhesive ring is in the shape of a water droplet. When the glass decorative cover is applied, the adhesive is squeezed and spreads outwards. Due to the large amount of adhesive at the starting point, the water droplet-shaped adhesive spreads over a wider area and may even overflow the glass decorative cover.

[0003] In related technologies, to avoid glue overflow, the glass decorative cover is usually tilted when manually installed, so that its bottom is close to the starting point of the glue ring. This causes the glue ring to be pressed inward when the glass decorative cover is placed down. However, this method requires manual judgment of the tilt angle of the glass decorative cover. If the manually judged tilt angle is inaccurate, it cannot effectively solve the glue overflow problem. Summary of the Invention

[0004] To overcome the problems existing in the related technologies, this disclosure provides an intelligent glue application control method and system based on a large model to solve the defects in the related technologies.

[0005] According to a first aspect of the present disclosure, an intelligent adhesive application control method based on a large model is provided, comprising: In response to the glue application signal to the plastic base, the spectral data of the initial glue extruded by the glue gun of the glue application robot before contacting the plastic base is acquired; Based on the spectral data, the first model obtains the material property deviation of the initial glue. When the material property deviation exceeds a first threshold, the glue-applying robot is triggered to execute a first glue-applying control strategy. The first glue-applying strategy is at least used to control the glue gun to perform glue back-suction. The timing process data of the glue-applying robot and the spatial morphology data of the glued sections on the plastic base are obtained during the glue-applying process. The spatial morphology data is then discretized into a graph structure, wherein the nodes of the graph structure represent key location points obtained based on the spatial morphology data, and the edges of the nodes are obtained based on the spatial adjacency of the key location points. The second glue application control strategy is obtained by using the second major model based on the material property deviation, the timing process data, and the graph structure, and the glue application robot is triggered to execute the second glue application control strategy.

[0006] In one embodiment, obtaining the material property deviation of the initial adhesive based on the spectral data using a first large model includes: The viscosity of the initial adhesive is obtained based on the spectral data using the first major model, and the viscosity deviation of the viscosity relative to the preset standard viscosity is determined. The first adhesive application control strategy is also used to adjust the adhesive application process parameters of the adhesive application robot in the following ways: when the viscosity is greater than the preset standard viscosity, increase the movement speed of the adhesive application robot and decrease the dispensing pressure of the glue gun; or, when the viscosity is less than or equal to the preset standard viscosity, decrease the movement speed of the adhesive application robot and increase the dispensing pressure of the glue gun.

[0007] In one embodiment, the second adhesive application control strategy, derived from the second large model based on the material property deviation, the time-series process data, and the graph structure, includes: The timing process data is encoded by the timing encoder of the second major model to obtain a timing feature vector, and the material property data is concatenated with the timing feature vector to obtain a first fused feature vector; The spatial morphological data is encoded by the graph encoder of the second major model to obtain a spatial feature vector, and the spatial feature vector is concatenated with the first fusion feature vector to obtain a multimodal fusion feature. The decision layer of the second major model outputs a control strategy identifier based on the multimodal fusion features, and determines a second adhesive application control strategy from a preset strategy library based on the control strategy identifier.

[0008] In one embodiment, the preset strategy library includes at least one of the following adhesive application control strategies: Adjust the dispensing pressure of the glue gun of the glue applicator and control the glue gun of the glue applicator to perform glue back suction action; The laser curing module is controlled to perform localized pre-curing of the initial adhesive on the plastic base; The glue-applying robot is controlled to perform a scraping motion on the initial glue on the plastic base. Adjust the motion path and speed curve of the glue-applying robot.

[0009] In one embodiment, the output of the second large model also includes the predicted volume of water droplets from the glue gun and the probability value of glue overflow risk. The intelligent glue application control method based on the large model further includes: The predicted water droplet volume is discretized to obtain a volume risk component, and the probability value of the overflow risk is discretized to obtain a probability risk component. The higher value between the volume risk component and the probability risk component is taken as the comprehensive risk level. Based on the comprehensive risk level, a set of control strategy identifiers is determined from the preset strategy mapping relationship, and the control strategy identifiers output by the second major model are compared with the set of control strategy identifiers for consistency. When the control strategy identifier output by the second major model does not belong to the set of control strategy identifiers, a target control strategy identifier is selected from the set of control strategy identifiers, and the glue-applying robot is triggered to execute the glue-applying control strategy corresponding to the target control strategy identifier.

[0010] In one embodiment, the intelligent adhesive application control method based on a large model further includes: Determine the volume deviation of the predicted water droplet volume relative to the historical predicted water droplet volume, and determine the risk deviation of the overflow risk probability value relative to the historical overflow risk probability value; The larger of the volume deviation and the risk deviation is used as the decision confidence level; The discretization of the predicted water droplet volume to obtain a volume risk component, and the discretization of the overflow risk probability value to obtain a probability risk component, include: If the decision confidence deviation does not exceed the second threshold, the predicted water droplet volume is discretized to obtain the volume risk component, and the probability value of the overflow risk is discretized to obtain the probability risk component.

[0011] In one embodiment, the intelligent adhesive application control method based on a large model further includes: The second large model is optimized and trained based on the second glue application control strategy and the glue application control strategy corresponding to the target control strategy identifier; or... If the decision confidence deviation exceeds the second threshold, a manual review prompt message for the second glue application control strategy is output. After receiving the manual review result of the second glue application control strategy, the second large model is optimized and trained based on the second glue application control strategy and the manual review result.

[0012] In one embodiment, a microspectrometer is coaxially integrated at the nozzle of the glue gun of the glue-applying robot. The process of acquiring spectral data of the initial glue extruded from the glue gun of the glue-applying robot before contacting the plastic base, in response to a glue-applying signal to the plastic base, includes: In response to the glue application signal to the plastic base, after waiting for a target preset time, a first trigger signal is sent to the micro-spectrometer, so that the micro-spectrometer responds to the first trigger signal to perform spectral acquisition on the initial glue extruded from the glue gun of the glue application robot, and obtains the spectral data of the initial glue before contacting the plastic base. The probe of the micro-spectrometer is coaxially integrated into the end of the glue gun nozzle of the glue application robot, and the target preset time is obtained by adjusting the initial preset time calibrated by the experiment based on the quality of the historically acquired spectral signals.

[0013] In one embodiment, the spatial morphological data is three-dimensional point cloud data, and the discretization of the spatial morphological data into a graph structure includes: Based on the three-dimensional point cloud data, the local centerline corresponding to the glued section is extracted; Based on the real-time pose data of the glue-applying robot, the spatial continuity between the starting point of the local centerline and the end node of the existing graph structure is determined, wherein the existing graph structure is obtained by extracting the local centerline from historical 3D point cloud data. Based on the spatial continuity, the local centerline data is connected to the end nodes of the existing graph structure data through a curve fitting algorithm to obtain a spliced ​​graph structure. The spliced ​​graph structure data is resampled with equal arc lengths to obtain a new graph structure containing a preset number of nodes.

[0014] According to a second aspect of the present disclosure, an intelligent glue application control system based on a large model is provided, including a glue application robot, a spatial data acquisition unit, and a controller. The glue application robot includes a spectral data acquisition unit, a time-series data acquisition unit, and a glue gun. The spectral data acquisition unit is used to acquire the spectrum of the initial glue extruded by the glue gun, obtain the spectral data of the initial glue before it contacts the plastic base, and send the spectral data to the controller. The timing data acquisition unit is used to acquire timing process data of the glue-applying robot during the glue-applying process and send the timing process data to the controller; The spatial data acquisition unit is used to acquire spatial morphological data of the glued track section on the plastic base and send the spatial morphological data to the controller; The glue gun is used to apply glue to the plastic base; The controller is configured to execute the intelligent glue application control method based on a large model as described in any of the first aspects, so as to control the glue application operation of the glue application robot.

[0015] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects: The intelligent glue application control method based on a large model provided in this disclosure can perform two levels of intelligent glue application control: at the beginning stage and during the application process. This not only allows for rapid response to glue overflow issues but also enables precise handling of complex glue overflow scenarios. At the beginning stage, a first large model performs spectral analysis to obtain the material property deviations of the initial glue. If the material property deviation exceeds a first threshold, a first glue application control strategy is executed, thereby quickly addressing glue overflow issues caused by glue material fluctuations and reducing the generation of droplet-like defects in the initial glue from the source, achieving pre-overflow prevention. During the application process, a second large model analyzes multimodal data (i.e., material property deviations, time-series process data, and graph structures) to fully understand the complex coupling relationship between the glue application robot state, glue path geometry, and material physical properties, improving decision-making accuracy and resulting in a more accurate second glue application control strategy, effectively addressing complex glue overflow scenarios. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0017] Figure 1 This is a schematic diagram of the adhesive applied to the plastic base of the washing machine door; Figure 2 This is a flowchart illustrating an exemplary embodiment of the present disclosure of an intelligent glue application control method based on a large model; Figure 3 This is a schematic diagram of the structure of an intelligent adhesive application control system based on a large model, as illustrated in an exemplary embodiment of this disclosure. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure.

[0019] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure. The singular forms “a,” “the,” and “the” used in this disclosure are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items. It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this disclosure, and similarly, second information may also be referred to as first information. Depending on the context, the word “if,” as used herein, can be interpreted as “when,” “in response to a determination,” or “when…”.

[0020] As mentioned in the background section, in related technologies, to avoid adhesive overflow, the glass decorative cover is typically tilted during manual installation, with its bottom close to the starting point of the adhesive ring. This creates an inward pressure on the starting point of the adhesive ring when the glass decorative cover is placed down. However, this method requires manual judgment of the tilt angle of the glass decorative cover; if the manually judged tilt angle is inaccurate, it cannot effectively solve the adhesive overflow problem.

[0021] Based on this, in a first aspect, at least one embodiment of this disclosure provides an intelligent adhesive application control method based on a large model, please refer to the appendix. Figure 2 The diagram illustrates the process of the method, including steps S101 to S104.

[0022] In step S101, in response to the glue application signal to the plastic base, the spectral data of the initial glue extruded from the glue gun of the glue application robot before contacting the plastic base is acquired.

[0023] In step S102, the material property deviation of the initial adhesive is obtained based on spectral data using the first large model. If the material property deviation exceeds a first threshold, the adhesive application robot is triggered to execute the first adhesive application control strategy. The first adhesive application strategy is used at least to control the glue gun to perform the adhesive back-suction action.

[0024] In step S103, the timing process data during the adhesive coating process and the spatial morphology data of the coated sections on the plastic machine base are acquired, and the spatial morphology data is discretized into a graph structure. The nodes of the graph structure represent key location points obtained based on the spatial morphology data, and the edges of the nodes are obtained based on the spatial adjacency of the key location points.

[0025] In step S104, a second adhesive application control strategy is obtained based on material property deviations, timing process data, and graph structure using the second large model, and the adhesive application robot is triggered to execute the second adhesive application control strategy.

[0026] Therefore, two-level intelligent glue application control can be implemented at the beginning and end of the glue application process. This not only allows for rapid response to glue overflow issues but also enables precise handling of complex glue overflow scenarios. At the beginning of the application process, spectral analysis using the first major model reveals the material property deviations of the initial glue. If these deviations exceed a first threshold, the first glue application control strategy is executed. This quickly addresses glue overflow issues caused by material fluctuations, reducing the generation of droplet-like defects in the initial glue and achieving pre-overflow prevention. During the glue application process, the second major model analyzes multimodal data (i.e., material property deviations, time-series process data, and graph structures). This allows for a thorough understanding of the complex coupling relationship between the glue application robot's state, glue path geometry, and material physical properties, improving decision-making accuracy and resulting in a more accurate second glue application control strategy to effectively handle complex glue overflow scenarios.

[0027] To make it easier to understand, the above steps will be further illustrated with examples below.

[0028] For example, the plastic base could be the plastic base of a washing machine door. The adhesive application signal indicates the initiation of adhesive application to the plastic base.

[0029] In some embodiments, a microspectrometer is coaxially integrated at the nozzle of the glue gun of the glue-applying robot. In response to the glue-applying signal to the plastic base, the microspectrometer acquires the spectral data of the initial glue extruded from the glue gun of the glue-applying robot before contacting the plastic base. This includes: in response to the glue-applying signal to the plastic base, after waiting for a target preset time, sending a first trigger signal to the microspectrometer, so that the microspectrometer, in response to the first trigger signal, performs spectral acquisition on the initial glue extruded from the glue gun of the glue-applying robot to obtain the spectral data of the initial glue before contacting the plastic base. The probe of the microspectrometer is coaxially integrated at the end of the glue gun nozzle of the glue-applying robot, and the target preset time is obtained by adjusting the experimentally calibrated initial preset time based on the quality of historically acquired spectral signals.

[0030] The target preset duration is used to ensure that the microspectrometer acquires spectra at the instant the adhesive is stably extruded and forms a complete liquid bridge (or adhesive meniscus). For example, it can be set to 2ms-10ms. In practical applications, a high-speed camera can be used to synchronously record the adhesive application process. By analyzing video frames, the time elapsed from the adhesive application signal to the first complete appearance of the stable adhesive flow in the spectrometer's optical path can be accurately measured, thus obtaining an initial preset duration. Then, fine-tuning is performed based on the quality of historically acquired spectral signals (such as signal-to-noise ratio and signal strength), automatically optimizing the process. For example, if the signal strength is below a threshold, it indicates that the adhesive may not have reached the optimal measurement position, so a preset time step can be added to the initial preset duration. If the signal strength is above the threshold, it indicates that there may be too much adhesive or the adhesive may be too close to the target area, so a preset time step can be reduced from the initial preset duration.

[0031] It should be understood that the initial glue is a segment of glue extruded from the glue gun of the glue applicator at the beginning stage, for example, a 1-2 mm long segment. Assuming the robot's glue applicating speed is 300 mm / s, the spectral acquisition window length of the microspectrometer can be 6-7 ms. Based on this, a microspectrometer with an exposure time between 1 ms and 10 ms can be selected. The probe of the microspectrometer can be coaxially integrated into the end of the glue gun nozzle of the glue applicator. Its optical path is parallel to or at a preset angle to the glue extrusion path, thus ensuring that spectral acquisition and glue extrusion occur simultaneously. This captures the spectral data of the glue about to be applied to the plastic base, rather than the spectral data of the glue that may have changed in the pipe or storage tank, thereby ensuring the accuracy of the first-stage glue applicator control.

[0032] In some embodiments, the material property deviation of the initial adhesive is obtained based on spectral data using a first large model, including: obtaining the viscosity of the initial adhesive based on spectral data using the first large model, and determining the viscosity deviation relative to a preset standard viscosity; the first adhesive application control strategy is also used to adjust the adhesive application process parameters of the adhesive application robot in the following ways: when the viscosity is greater than the preset standard viscosity, increasing the movement speed of the adhesive application robot and decreasing the dispensing pressure of the glue gun, or when the viscosity is less than or equal to the preset standard viscosity, decreasing the movement speed of the adhesive application robot and increasing the dispensing pressure of the glue gun.

[0033] It should be understood that when the physical properties of adhesive change, the intermolecular forces also change, which in turn affects its spectral characteristics. For example, increasing the water content in the adhesive alters the overall molecular interaction network, decreasing the viscosity and enhancing the absorption peaks corresponding to OH bond vibrations in the near-infrared region. Therefore, by analyzing the intensity, shape, or overall spectral pattern of absorption peaks at specific wavelengths in the spectrum using the first major model, the viscosity of the adhesive can be indirectly estimated.

[0034] For example, a lightweight Transformer-based large model can be trained based on a massive spectral-viscosity dataset and deployed in an edge computing unit within the control cabinet of the glue-applying robot. This model can infer the viscosity of the glue from the real-time spectrum within 5ms and calculate its viscosity deviation from a preset standard viscosity. Once the viscosity deviation exceeds a first threshold (e.g., 5%), the first ultra-fast compensation (i.e., the execution of the first glue-applying control strategy) can be triggered at the beginning stage of the ring-shaped glue application.

[0035] For example, the preset standard viscosity can be obtained by accurately measuring a sample of glue from the same batch taken from the glue gun using an offline, high-precision standard instrument (such as a rotational rheometer) during the stabilization phase in the laboratory or early stages of production.

[0036] For example, the first glue application control strategy may be: controlling the glue gun to perform a glue back suction action and, when the viscosity is greater than a preset standard viscosity, increasing the movement speed of the glue application robot and decreasing the glue gun's dispensing pressure. Alternatively, the first glue application control strategy may be: controlling the glue gun to perform a glue back suction action and, when the viscosity is less than or equal to a preset standard viscosity, decreasing the movement speed of the glue application robot and increasing the glue gun's dispensing pressure.

[0037] It should be understood that increased viscosity leads to poorer glue flow. If the glue-applying robot moves slowly, a large amount of glue will accumulate in a small area, resulting in a high and narrow glue line. Therefore, it is necessary to increase the movement speed of the glue-applying robot to spread the same volume of glue over a longer distance. Additionally, reducing the dispensing pressure of the glue gun can prevent excessive total glue dispensing due to increased speed. Thus, the combined effect of back suction, increased speed, and reduced pressure can precisely control the amount of glue at the starting point within the optimal range, reducing glue overflow.

[0038] It should be understood that low-viscosity adhesives have good flowability. If the moving speed is too fast, the adhesive will be dragged away before it can stabilize and form, resulting in excessive spreading due to inertia. This can lead to overly wide adhesive lines, uneven edges, or even adhesive overflow into other areas. Therefore, it is necessary to reduce the movement speed of the adhesive applicator, allowing the adhesive more time to stabilize and deposit under gravity, forming a full and regular adhesive line outline. Additionally, low-viscosity adhesives are more prone to backflow in the pipeline. Increasing the dispensing pressure of the glue gun can ensure sufficient adhesive flow to maintain the necessary height and volume of the adhesive line, thereby guaranteeing the final bond strength and sealing effect.

[0039] By employing the methods described above, back suction can resolve momentary glue overload, while adjusting process parameters (i.e., adjusting the movement speed of the glue applicator and the glue gun's dispensing pressure) can reduce similar problems in subsequent glue application processes. Therefore, the combined action of back suction and process parameter adjustment can quickly address glue overflow issues caused by changes in glue properties, reducing the risk of glue overflow.

[0040] Furthermore, it should be understood that the spectral characteristics are also correlated with the surface tension of the adhesive, thus the first major model can be trained using a surface tension-spectral dataset. Therefore, the viscosity deviation and / or surface tension deviation can be obtained from the first major model based on the spectral data. In practical applications, these deviations can be set according to actual needs, and this disclosure does not limit this.

[0041] In some embodiments, the first adhesive application control strategy may further include: controlling the adhesive application robot to apply or spray the initial adhesive into other adhesive recycling equipment, and then applying a ring-shaped adhesive ring onto the plastic base.

[0042] During the execution of the first adhesive application control strategy, temporal process data and spatial morphology data can be acquired simultaneously. The temporal process data includes the ambient temperature of the adhesive application robot's environment, the pressure and temperature of the glue gun, the robot's pose, speed, acceleration, and angular velocity. The spatial morphology data can be 3D point cloud data, obtained as follows: a second trigger signal is sent to the line laser scanner to emit a laser line onto the adhesive-coated track on the plastic base, and a third trigger signal is sent to the camera to capture the laser deformation image caused by the adhesive-coated track; based on the triangulation principle, the laser deformation image is reconstructed in 3D to generate 3D point cloud data representing the spatial morphology of the adhesive-coated track segment.

[0043] For example, a line laser scanner can be mounted 50-100mm behind the glue gun of an adhesive applicator robot, allowing it to scan the glue run immediately after application. For instance, when applying glue at a speed of 300mm / s, the line laser scanner moves synchronously at the same rate, ensuring that the glue run is scanned immediately after application to acquire real-time 3D point cloud data.

[0044] In some embodiments, the spatial morphology data is three-dimensional point cloud data. Discretizing the spatial morphology data into a graph structure includes: extracting local centerlines corresponding to the glued track segments based on the three-dimensional point cloud data; determining the spatial continuity between the starting point of the local centerline and the end nodes of the existing graph structure based on the real-time pose data of the glue-applying robot, wherein the existing graph structure is obtained by extracting local centerlines from historically obtained three-dimensional point cloud data; connecting the local centerline data to the end nodes of the existing graph structure data using a curve fitting algorithm according to the spatial continuity to obtain a spliced ​​graph structure; and resampling the spliced ​​graph structure data with equal arc lengths to obtain a new graph structure containing a preset number of nodes.

[0045] For example, extracting the local centerline corresponding to an adhesive run can be done by: for each scanned adhesive run, calculating the center point of each cross-section using the weighted centroid method, and then connecting these center points to form a local centerline. For example, for an adhesive run segment with a length of 10 mm, 50 center points are obtained, which represent the center trajectory of that adhesive run segment.

[0046] For example, the existing graph structure is composed of the centerlines of previously scanned glue lane segments. During stitching, real-time pose data (including position and orientation) of the glue applicator robot can be used to determine the spatial continuity between the starting point of the local centerline and the end of the existing graph structure. Then, a curve fitting algorithm (such as B-spline curve fitting) is used to smoothly connect the local centerline to the end nodes of the existing graph structure, ensuring the continuity of the glue lane geometry. For example, if the existing graph structure has 150 nodes and the newly scanned local centerline has 50 nodes, the total number of nodes is maintained at around 200 through stitching and subsequent resampling.

[0047] For example, the preset number can be set to 200. First, the total arc length of the current mosaic structure can be calculated. Then, resampling points are selected at equal arc length intervals, and the three-dimensional coordinates and attributes (such as glue height, glue width, etc.) of these points are obtained through linear interpolation. For example, if the total arc length is 500mm and the preset number is 200, then a resampling point is selected every 2.5mm.

[0048] Therefore, the glued sections can be scanned in real time during the glue application process, and the newly scanned local center line can be spliced ​​with the existing graph structure to achieve dynamic updating of the graph structure, thereby more timely detection of the risk of glue overflow at the starting point during the glue application process.

[0049] It should be understood that after the first stage of back suction and acceleration, slight blockage of the glue gun may prevent the back suction from achieving the expected effect, resulting in glue accumulation at the starting point. Alternatively, the glue properties may be completely normal in the first stage, but the glue-applying robot may suddenly vibrate violently, causing excessive glue to be splashed out at the starting point. To cope with various unexpected glue overflow scenarios, embodiments of this disclosure can also perform a second stage of glue application control. This second stage of glue application control can predict whether glue overflow will occur after the glass decorative cover is placed by real-time analysis of the morphology of the applied glue path and the current process parameters, and take compensatory measures to change the state of the glue (such as reducing the amount of glue, changing the shape of the glue, etc.) before placing the glass decorative cover, thereby avoiding glue overflow.

[0050] For example, a second-stage adhesive application control strategy (i.e., the second adhesive application control strategy) can be obtained through a second large model based on material property deviations, time-series process data, and graph structures. The second large model, through its graph structure, can acquire the shape of the adhesive-coated segments from the starting point to the current time (including the size of the teardrop-shaped areas in the starting area). If teardrop-shaped accumulation has already occurred in the starting area, it can be predicted that the accumulated adhesive will be squeezed out and overflow when the glass decorative cover is applied. Compensation measures (such as back-suction, laser pre-curing, etc.) can then be taken to reduce accumulation or increase adhesive strength, thereby preventing overflow. Time-series process parameters (such as robot speed, acceleration, glue gun pressure, etc.) can reflect the dynamic behavior of the adhesive application process. If abnormalities such as robot vibration or sudden speed changes are detected, it can be predicted that these abnormalities may lead to defects in the applied adhesive (such as adhesive accumulation at the starting point), and subsequent process parameters can be adjusted in advance or the applied adhesive can be directly compensated to prevent further deterioration of defects.

[0051] For example, if the initial glue buildup persists after the first stage of suction and acceleration, the second model can understand the buildup from the graph structure, thereby triggering a suitable second glue application control strategy. If the glue properties are completely normal, but the glue application robot suddenly experiences severe vibration, the second model can capture the vibration from the timing process data, understand the buildup from the graph structure, and confirm the material is correct from the property data, thereby triggering a suitable second glue application control strategy.

[0052] Therefore, the first major model can quickly respond to glue spillage problems, while the second major model can perform multimodal understanding and reasoning to accurately deal with complex glue spillage scenarios, thus effectively addressing glue spillage problems in different scenarios.

[0053] In some embodiments, a second adhesive application control strategy is obtained by using a second large model based on material property deviations, time-series process data, and graph structure. This includes: encoding the time-series process data using a time-series encoder of the second large model to obtain a time-series feature vector, and concatenating the material property data with the time-series feature vector to obtain a first fused feature vector; encoding the spatial morphology data using a graph encoder of the second large model to obtain a spatial feature vector, and concatenating the spatial feature vector with the first fused feature vector to obtain a multimodal fused feature; and outputting a control strategy identifier based on the multimodal fused feature by the decision layer of the second large model, and determining the second adhesive application control strategy in a preset strategy library based on the control strategy identifier.

[0054] For example, a temporal encoder can be based on a Transformer architecture to extract features characterizing robot motion states and process stability from high-frequency time series. A graph encoder can be based on a Graph-Transformer architecture to extract spatial distribution features characterizing anomalies in the geometry of the glue track from a graph structure.

[0055] For example, concatenating material property data with a temporal feature vector to obtain the first fused feature vector can be achieved by directly concatenating glue viscosity deviation and surface tension deviation as additional feature dimensions to the end of the temporal feature vector. Thus, concatenating material property data with the temporal feature vector allows the second model to consider the physical state of the glue, thereby more accurately explaining temporal behavior (e.g., the same robot shaking has different effects on glues of different viscosities). Then, concatenating the concatenated temporal-physical property features with graph features allows the model to simultaneously consider temporal dynamics, spatial morphology, and material properties, thereby making more comprehensive decisions.

[0056] In some embodiments, the preset strategy library includes at least one of the following adhesive application control strategies: Adjust the dispensing pressure of the glue gun in the glue applicator and control the glue gun in the glue applicator to perform the glue back suction action; The laser curing module is controlled to perform localized pre-curing of the initial adhesive on the plastic base; Control the glue-applying robot to scrape the initial glue on the plastic base; Adjust the motion path and speed curve of the glue-applying robot.

[0057] For example, each glue application control strategy in the preset strategy library can correspond to an identifier, such as an integer from 0 to 4, which correspond to no action and the above four glue application control strategies, respectively. Therefore, if the second model outputs control strategy identifier 2 based on multimodal fusion features, the second glue application control strategy can be determined as follows: control the laser curing module to locally pre-cure the initial glue on the plastic base, so that the glue surface in the starting area is slightly cured, changing its rheology and preventing glue overflow.

[0058] In some embodiments, the output of the second major model also includes a predicted droplet volume of glue extruded from the glue gun and a probability value of glue overflow risk. Accordingly, the predicted droplet volume can be discretized to obtain a volume risk component, and the probability value of glue overflow risk can be discretized to obtain a probability risk component. The higher value between the volume risk component and the probability risk component is taken as the comprehensive risk level. Based on the comprehensive risk level, a set of control strategy identifiers is determined from a preset strategy mapping relationship, and the control strategy identifiers output by the second major model are compared with the set of control strategy identifiers. When the control strategy identifier output by the second major model does not belong to the set of control strategy identifiers, a target control strategy identifier is selected from the set of control strategy identifiers. The second major model is then optimized and trained based on the second glue application control strategy and the glue application control strategy corresponding to the target control strategy identifier.

[0059] In some embodiments, determining a second adhesive application control strategy in a preset strategy library based on a control strategy identifier includes: when the control strategy identifier output by the second large model belongs to a set of control strategy identifiers, determining a second adhesive application control strategy in the preset strategy library based on the control strategy identifier.

[0060] For example, the decision layer of the second-largest model may include a multilayer perceptron for performing multi-task learning, simultaneously outputting a predicted droplet volume, a probability value of glue overflow risk, and a control policy identifier. Thus, the rationality of the second-largest model's choice of the control policy identifier can be verified based on the predicted droplet volume and the probability value of glue overflow risk. If the verification fails, the policy selection is deemed unreasonable, triggering a policy correction mechanism.

[0061] For example, the discretization of the predicted water droplet volume can be as follows: when the predicted water droplet volume is <15nL, the volume risk component is set to 0; when 15nL ≤ the predicted water droplet volume is <25nL, the volume risk component is set to 1; when the predicted water droplet volume is ≥25nL, the volume risk component is set to 2.

[0062] For example, the discretization of the probability value of glue overflow risk can be as follows: when the probability value of glue overflow risk is <0.3, the probability risk component is set to 0; when 0.3≤ the probability value of glue overflow risk is <0.6, the probability risk component is set to 1; when the probability value of glue overflow risk is ≥0.6, the probability risk component is set to 2.

[0063] For example, the preset strategy mapping relationship can be: risk level 0 corresponds to control strategy identifier {0}; risk level 1 corresponds to control strategy identifier {1,2}; risk level 2 corresponds to control strategy identifier {3,4}.

[0064] For example, if the control strategy identifier output by the second largest model does not belong to the set of control strategy identifiers, it indicates that the glue application control strategy selected by the second largest model is unreasonable. Therefore, the most conservative compensation strategy can be selected from the set of control strategy identifiers as an alternative strategy. For instance, if the current predicted droplet volume is 30 nL, the probability of glue overflow is 0.8, the risk level is 2, and the set of control strategy identifiers is {3,4}, if the control strategy identifier output by the second largest model is 0 (no action required), the validation layer will determine it as unreasonable and then select the most conservative strategy 3 (robot scraping inside) from {3,4} as an alternative strategy.

[0065] It should be understood that the second major model generates a glue application control strategy based on complex multimodal features. This strategy considers the complex interactions of various factors and then outputs a specific glue application control strategy. The aforementioned validation process, based on rule-based independent checks, ensures that the strategy selection of the second major model conforms to basic risk logic and prevents obvious model decision-making errors.

[0066] In some embodiments, to improve the accuracy of strategy verification, the volume deviation of the predicted water droplet volume relative to historical predicted water droplet volumes can be determined, and the risk deviation of the glue overflow risk probability value relative to historical glue overflow risk probability values ​​can be determined; the larger of the volume deviation and the risk deviation is used as the decision confidence level. Accordingly, if the decision confidence level deviation does not exceed a second threshold, the predicted water droplet volume can be discretized to obtain a volume risk component, and the glue overflow risk probability value can be discretized to obtain a probability risk component.

[0067] Therefore, strategy verification can be performed when the predicted droplet volume and the probability value of glue overflow risk are reliable, thereby improving the verification accuracy of the output strategy of the second major model.

[0068] In some embodiments, the second large model can be optimized and trained based on the second glue application control strategy and the glue application control strategy corresponding to the target control strategy identifier; or, if the decision confidence deviation exceeds the second threshold, a manual review prompt message for the second glue application control strategy is output, and after receiving the manual review result of the second glue application control strategy, the second large model is optimized and trained based on the second glue application control strategy and the manual review result.

[0069] Therefore, boundary cases can be collected by verifying the results or manually reviewing them, and used for the optimization of the second-largest model, thereby improving the accuracy of the strategy selection of the second-largest model.

[0070] According to a second aspect of the embodiments of this disclosure, an intelligent adhesive application control system based on a large model is provided. Please refer to the appendix. Figure 3 The intelligent glue application control system 200 based on a large model includes: a glue application robot 201, a spatial data acquisition unit 202, and a controller 203. The glue application robot 201 includes a spectral data acquisition unit 2011, a time-series data acquisition unit 2012, and a glue gun 2013.

[0071] The spectral data acquisition unit 2011 is used to acquire the spectrum of the initial glue extruded by the glue gun 2013, obtain the spectral data of the initial glue before contacting the plastic base, and send the spectral data to the controller 203. The timing data acquisition unit 2012 is used to collect the timing process data of the glue-applying robot 201 during the glue-applying process and send the timing process data to the controller 203; The spatial data acquisition unit 202 is used to acquire spatial morphological data of the glued track section on the plastic base and send the spatial morphological data to the controller 203. The glue gun 2013 is used for applying glue to plastic bases. The controller 203 is used to execute any of the above-mentioned intelligent glue application control methods based on large models to control the glue application operation of the glue application robot 201.

[0072] For example, the spectral data acquisition unit 2011 may be a microspectrometer. The timing data acquisition unit 2012 may include one or more sensors capable of acquiring timing process parameters of the coating robot, such as temperature sensors, pressure sensors, speed sensors, and acceleration sensors. The spatial data acquisition unit 202 may include a line laser and a camera.

[0073] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments of the method in the first aspect, and will not be elaborated upon here.

[0074] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0075] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0076] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

Claims

1. An intelligent adhesive application control method based on a large model, characterized in that, include: In response to the glue application signal to the plastic base, the spectral data of the initial glue extruded by the glue gun of the glue application robot before contacting the plastic base is acquired; Based on the spectral data, the first model obtains the material property deviation of the initial glue. When the material property deviation exceeds a first threshold, the glue-applying robot is triggered to execute a first glue-applying control strategy. The first glue-applying strategy is at least used to control the glue gun to perform glue back-suction. The timing process data of the glue-applying robot and the spatial morphology data of the glued sections on the plastic base are obtained during the glue-applying process. The spatial morphology data is then discretized into a graph structure, wherein the nodes of the graph structure represent key location points obtained based on the spatial morphology data, and the edges of the nodes are obtained based on the spatial adjacency of the key location points. The second glue application control strategy is obtained by using the second major model based on the material property deviation, the timing process data, and the graph structure, and the glue application robot is triggered to execute the second glue application control strategy.

2. The intelligent adhesive application control method based on a large model according to claim 1, characterized in that, The process of obtaining the material property deviations of the initial adhesive based on the spectral data using the first large model includes: The viscosity of the initial adhesive is obtained based on the spectral data using the first major model, and the viscosity deviation of the viscosity relative to the preset standard viscosity is determined. The first adhesive application control strategy is also used to adjust the adhesive application process parameters of the adhesive application robot in the following ways: when the viscosity is greater than the preset standard viscosity, increase the movement speed of the adhesive application robot and decrease the dispensing pressure of the glue gun; or, when the viscosity is less than or equal to the preset standard viscosity, decrease the movement speed of the adhesive application robot and increase the dispensing pressure of the glue gun.

3. The intelligent glue application control method based on a large model according to claim 1, characterized in that, The second adhesive application control strategy is obtained through the second major model based on the material property deviation, the time-series process data, and the graph structure, including: The timing process data is encoded by the timing encoder of the second major model to obtain a timing feature vector, and the material property data is concatenated with the timing feature vector to obtain a first fused feature vector; The spatial morphological data is encoded by the graph encoder of the second major model to obtain a spatial feature vector, and the spatial feature vector is concatenated with the first fusion feature vector to obtain a multimodal fusion feature. The decision layer of the second major model outputs a control strategy identifier based on the multimodal fusion features, and determines a second adhesive application control strategy from a preset strategy library based on the control strategy identifier.

4. The intelligent glue application control method based on a large model according to claim 3, characterized in that, The preset strategy library includes at least one of the following adhesive application control strategies: Adjust the dispensing pressure of the glue gun of the glue applicator and control the glue gun of the glue applicator to perform glue back suction action; The laser curing module is controlled to perform localized pre-curing of the initial adhesive on the plastic base; The glue-applying robot is controlled to perform a scraping motion on the initial glue on the plastic base. Adjust the motion path and speed curve of the glue-applying robot.

5. The intelligent adhesive application control method based on a large model according to claim 3, characterized in that, The output of the second large model also includes the predicted volume of water droplets from the glue gun and the probability of glue overflow. The intelligent glue application control method based on the large model also includes: The predicted water droplet volume is discretized to obtain a volume risk component, and the probability value of the overflow risk is discretized to obtain a probability risk component. The higher value between the volume risk component and the probability risk component is taken as the comprehensive risk level. Based on the comprehensive risk level, a set of control strategy identifiers is determined from the preset strategy mapping relationship, and the control strategy identifiers output by the second major model are compared with the set of control strategy identifiers for consistency. When the control strategy identifier output by the second major model does not belong to the set of control strategy identifiers, a target control strategy identifier is selected from the set of control strategy identifiers, and the glue-applying robot is triggered to execute the glue-applying control strategy corresponding to the target control strategy identifier.

6. The intelligent adhesive application control method based on a large model according to claim 5, characterized in that, The intelligent glue application control method based on a large model also includes: Determine the volume deviation of the predicted water droplet volume relative to the historical predicted water droplet volume, and determine the risk deviation of the overflow risk probability value relative to the historical overflow risk probability value; The larger of the volume deviation and the risk deviation is used as the decision confidence level; The discretization of the predicted water droplet volume to obtain a volume risk component, and the discretization of the overflow risk probability value to obtain a probability risk component, include: If the decision confidence deviation does not exceed the second threshold, the predicted water droplet volume is discretized to obtain the volume risk component, and the probability value of the overflow risk is discretized to obtain the probability risk component.

7. The intelligent adhesive application control method based on a large model according to claim 6, characterized in that, The intelligent glue application control method based on a large model also includes: The second large model is optimized and trained based on the second glue application control strategy and the glue application control strategy corresponding to the target control strategy identifier; or... If the decision confidence deviation exceeds the second threshold, a manual review prompt message for the second glue application control strategy is output. After receiving the manual review result of the second glue application control strategy, the second large model is optimized and trained based on the second glue application control strategy and the manual review result.

8. The intelligent glue application control method based on a large model according to any one of claims 1-7, characterized in that, The glue-applying robot has a micro-spectrometer coaxially integrated at the glue gun nozzle. In response to a glue-applying signal to the plastic base, the robot acquires spectral data of the initial glue extruded from the glue gun before contacting the plastic base, including: In response to the glue application signal to the plastic base, after waiting for a target preset time, a first trigger signal is sent to the micro-spectrometer, so that the micro-spectrometer responds to the first trigger signal to perform spectral acquisition on the initial glue extruded from the glue gun of the glue application robot, and obtains the spectral data of the initial glue before contacting the plastic base. The probe of the micro-spectrometer is coaxially integrated into the end of the glue gun nozzle of the glue application robot, and the target preset time is obtained by adjusting the initial preset time calibrated by the experiment based on the quality of the historically acquired spectral signals.

9. The intelligent glue application control method based on a large model according to any one of claims 1-7, characterized in that, The spatial morphological data is three-dimensional point cloud data, and the discretization of the spatial morphological data into a graph structure includes: Based on the three-dimensional point cloud data, the local centerline corresponding to the glued section is extracted; Based on the real-time pose data of the glue-applying robot, the spatial continuity between the starting point of the local centerline and the end node of the existing graph structure is determined, wherein the existing graph structure is obtained by extracting the local centerline from historical 3D point cloud data. Based on the spatial continuity, the local centerline data is connected to the end nodes of the existing graph structure data through a curve fitting algorithm to obtain a spliced ​​graph structure. The spliced ​​graph structure data is resampled with equal arc lengths to obtain a new graph structure containing a preset number of nodes.

10. An intelligent adhesive application control system based on a large model, characterized in that, It includes a glue-applying robot, a spatial data acquisition unit, and a controller. The glue-applying robot includes a spectral data acquisition unit, a time-series data acquisition unit, and a glue gun. The spectral data acquisition unit is used to acquire the spectrum of the initial glue extruded by the glue gun, obtain the spectral data of the initial glue before it contacts the plastic base, and send the spectral data to the controller. The timing data acquisition unit is used to acquire timing process data of the glue-applying robot during the glue-applying process and send the timing process data to the controller; The spatial data acquisition unit is used to acquire spatial morphological data of the glued track section on the plastic base and send the spatial morphological data to the controller; The glue gun is used to apply glue to the plastic base; The controller is configured to execute the intelligent glue application control method based on a large model as described in any one of claims 1-9, so as to control the glue application operation of the glue application robot.

Images