Hot melt adhesive bonding defect online identification and closed-loop compensation control method

Abstract

The invention belongs to the technical field of hot melt adhesive control, and discloses a hot melt adhesive bonding defect online identification and closed-loop compensation control method, which comprises the following steps: building a self-adaptive multi-source sensing acquisition system, dynamically matching an optimal sensor combination according to the material, specification and process of a workpiece, fusing environmental interference data and realizing multi-dimensional data space-time accurate alignment; effective information is screened through cross-modal collaborative preprocessing, feature extraction is enhanced in combination with a fusion network of a domain knowledge graph and deep learning, rapid adaptation of new defects of small samples is achieved in cooperation with a meta-learning model, an interpretable report can be generated, and the comprehensiveness, precision and positioning accuracy of defect recognition can be improved; a multi-target optimization algorithm is adopted to iteratively optimize parameter combination, target weight is dynamically adjusted and optimized in combination with defect severity, linkage adjustment of multiple execution mechanisms is realized by means of an industrial bus, and targeted dynamic accurate compensation is realized by matching with a fault pre-judgment and real-time feedback mechanism, so that the bonding quality stability is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of hot melt adhesive quality control technology, specifically a method for online identification and closed-loop compensation control of hot melt adhesive bonding defects. Background Technology

[0002] Hot melt adhesive bonding technology, with its advantages of fast curing speed, environmental friendliness, solvent-free application, and wide bonding range, has been widely used in various industrial fields such as packaging, electronics manufacturing, automotive assembly, and labeling. However, with industrial development and increasingly stringent requirements for precision and efficiency in production, the stable control of hot melt adhesive bonding quality has gradually become a key bottleneck restricting the development of high-end manufacturing. Currently, the main technical problems related to hot melt adhesive bonding quality control are as follows: On the one hand, defect detection methods are relatively backward, mostly relying on a single sensor to carry out detection work, which makes it difficult to fully capture multi-dimensional defect information in the hot melt adhesive bonding process and cannot meet the needs of high-speed production lines for accurate online identification of multi-dimensional defects, especially for complex defects such as adhesive bubbles and local missing adhesive.

[0003] On the other hand, existing technologies mostly adopt open-loop control or simple closed-loop control modes, which fail to establish a precise correspondence between defect types, severity and various control parameters. They can only make coarse adjustments to a single parameter, resulting in slow compensation response and inability to achieve targeted dynamic and precise compensation. This leads to insufficient bonding quality stability and makes it difficult to adapt to the actual needs of mixed production lines for multi-specification workpieces.

[0004] The quality of hot melt adhesive bonding is affected by a combination of factors such as temperature, pressure, and spraying speed, which further increases the difficulty of defect control. Summary of the Invention

[0005] The purpose of this invention is to provide an online identification and closed-loop compensation control method for hot melt adhesive bonding defects, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online identification and closed-loop compensation control method for hot melt adhesive bonding defects, comprising a multi-source acquisition stage, a collaborative preprocessing stage, a knowledge fusion stage, a meta-learning identification stage, a twin compensation stage, an edge execution stage, and a closed-loop optimization stage; Preferably, the multi-source acquisition stage constructs a multi-source sensing dynamic reconstruction acquisition system, integrating a high frame rate industrial camera, ultrasonic sensor, infrared temperature sensor, pressure sensor, servo motor encoder, and sensor type adaptive selection module. The adaptive selection module has a built-in process knowledge base retrieval engine, with the retrieval priority being "workpiece material compatibility > bonding process accuracy requirements > production cycle time compatibility". The matching judgment criterion is that the compatibility score between the sensor combination and the workpiece or process is ≥85 points; the full score is 100 points. Based on the weighted calculation of material compatibility, detection accuracy compliance rate, and response speed matching degree, it can automatically match the optimal sensor combination and acquisition parameters according to the workpiece information and bonding process requirements. The system incorporates a Kalman filter fusion algorithm to fuse data from ambient temperature, humidity, airflow speed, and other interfering factors in real time, dynamically adjusting the sampling frequency and trigger timing of each sensor. Industrial cameras operate at 50-200fps, ultrasonic sensors at 100-500Hz, and infrared temperature sensors at 50-100Hz. Trigger timing is based on the workpiece triggering the photoelectric sensor signal, with all sensors initiating acquisition 10ms in advance. A millisecond-level spatiotemporal synchronization calibration mechanism is employed, using a dual calibration strategy of clock synchronization and image physical feature point matching to achieve spatiotemporal alignment of multi-dimensional data such as image, acoustic, temperature, and mechanical data, generating a multi-dimensional structured dataset with annotations of interfering factors and correlations of process parameters.

[0007] An integrated sensor health monitoring module is used to diagnose accuracy degradation in real time by collecting data such as sensor output signal stability and calibration deviation. When the calibration deviation exceeds ±3%, an adaptive calibration algorithm is triggered, which uses a standard reference source for comparison calibration and adjusts the sensor output coefficient through a linear correction formula. The calibration cycle is automatically performed once every 2 hours by default, and is executed immediately when the deviation exceeds the threshold. When the deviation exceeds ±5%, redundancy switching is performed, triggering the adaptive calibration algorithm or executing redundancy switching.

[0008] Preferably, the collaborative preprocessing stage is based on the multi-dimensional structured dataset obtained in the multi-source acquisition stage, and a cross-modal collaborative preprocessing framework is designed; a composite strategy of illumination adaptive Retinex enhancement combined with neighborhood modality constraint denoising is adopted for the image signal, the type of illumination interference is determined by the synchronously acquired temperature signal, and the denoising schemes such as Gaussian filtering and median filtering and the size of the filter kernel are adaptively selected; A wavelet threshold adaptive denoising scheme combined with temperature acoustic feature correlation calibration is adopted for ultrasonic signals. The threshold adjustment is based on the energy distribution of ultrasonic signals and the temperature gradient change. For every 10°C increase in temperature, the threshold is reduced by 8%. When the signal energy fluctuation exceeds 20%, the threshold is dynamically adjusted to follow the energy peak. The propagation speed deviation of ultrasonic waves in different media is corrected in real time by infrared temperature data, and the signal sampling window is optimized by combining the bonding gap parameter. For time-series signals such as temperature, pressure, and speed, a sliding window smoothing strategy combined with cross-modal redundancy verification is adopted. The dynamic threshold setting rule is as follows: based on the mean μ and standard deviation σ of nearly 100 normal samples, the upper threshold = μ + 3σ and the lower threshold = μ - 3σ, and cross-modal data consistency must be met. For example, when the pressure exceeds the threshold, the fluctuation of temperature and speed signals must be ≤ ±5%, otherwise it is judged as a single signal anomaly, and a dynamic threshold is set to remove abnormal fluctuations. A three-level signal quality assessment model is constructed. The first-level assessment is based on the signal-to-noise ratio (≥20dB is valid) and data integrity (≥98% is valid) for preliminary screening. The second-level assessment is verified by cross-modal data consistency (matching degree of multi-modal signal change trend at the same time node ≥90%). The third-level assessment combines the process knowledge base to identify non-interference inherent signal fluctuations (similarity with historical normal signals ≥95%) and retains valid information through signal enhancement algorithms. An adaptive optimization module for preprocessing parameters is integrated to iteratively update the preprocessing strategy parameters according to the detection effect of different workpiece types, forming a parameter optimization library.

[0009] Preferably, the knowledge fusion stage relies on the multimodal data output from the collaborative preprocessing stage to construct a collaborative feature fusion network combining a knowledge graph of the hot melt adhesive bonding field with deep learning; based on the rheological properties of hot melt adhesives, bonding physical mechanisms, material thermodynamic properties, and historical defect data, an ontological modeling method is used to build a multimodal knowledge graph containing the full-link relationship of materials, processes, defects, features, and mechanisms. The core entities include 5 categories: adhesive types: EVA, POE, etc.; workpiece materials: metal, plastic, etc.; defect types: bubbles, insufficient adhesive, etc.; process parameters: temperature, pressure, etc.; feature parameters: grayscale value, peak amplitude, etc.; entity relationships include 6 categories such as "material-compatible adhesive type", "defect-feature association", and "process-defect influence"; the construction steps are as follows: 1) Data cleaning and annotation of historical defects and process data; 2) Define entity categories and attribute dictionaries; 3) Extracting entity relationships using a method that combines rules and machine learning; 4) Manually review and correct related logic.

[0010] The knowledge graph includes feature weighting rules and association logic for adhesive types, workpiece materials, and typical defects; For image signals, an improved MobileNetV2 network is adopted, which reduces the original network's 17 bottleneck layers to 12, retains the core feature extraction capability and reduces the computational cost. The inflation factor is set to 6. Prior constraints of typical defect morphology from the knowledge graph are introduced into the 5th to 8th convolutional layers of the network. Attention gating units are connected in series with the convolutional layers. By assigning a weight of 0.8-1.0 to suspected defect regions and a weight of 0.1-0.3 to non-suspected defect regions in the feature map, the feature extraction of defect regions is enhanced. For ultrasonic and temperature signals, time-domain analysis uses a 10ms sliding window to extract features such as peak amplitude, time-of-flight difference, and temperature gradient, with a window step size of 5ms. Frequency-domain analysis uses fast Fourier transform, with an analysis frequency range of 20kHz-1MHz. After extracting parameters such as characteristic frequencies and energy distribution, a knowledge graph is used for feature screening and preliminary weight allocation to exclude frequency components and feature parameters that are not related to the current adhesive type and workpiece material characteristics. A dynamic attention fusion mechanism is designed based on the relation weights of the knowledge graph (preset range 0.1-0.9) and the preliminary judgment results of defects in real time. The weights of each modality feature are calculated by the Softmax function, and the sum of the weights is 1. For example, the image modality weight = exp(knowledge graph image feature association weight) / Σexp(knowledge graph association weight of each modality). The weights of each modality feature are adaptively allocated, and finally a fusion feature vector with both data-driven accuracy and physical mechanism-supported reliability is generated.

[0011] Preferably, the meta-learning recognition stage uses the fused feature vector generated in the knowledge fusion stage as input to construct a meta-learning-enhanced interpretable defect recognition hierarchical model. The main body of the model adopts an improved Centernet network, with the backbone network set to 18 convolutional layers and 4 residual blocks. The feature map fusion adopts a bidirectional fusion strategy combining top-down and bottom-up approaches, fusing C3, C4, and C5 scale feature maps. This structure improves the defect localization accuracy. An integrated meta-feature transfer module is used, based on a model-independent meta-learning algorithm. In the pre-training stage, more than 10,000 historical defect samples are used, including more than 30 typical defect types. The sample categories are balanced, and the meta-learning rate is set to 0.001, the base learning rate to 0.01, and the number of iterations to 5,000. General feature transfer capabilities are obtained from existing defect recognition experience. In the fine-tuning stage, only the parameters of the last 3 fully connected layers of the model are adjusted, with a fine-tuning learning rate of 0.0005. Adaptation can be completed based on 3-5 new defect samples without retraining the entire model. In terms of interpretability design, an improved Grad-CAM visualization technology is adopted, selecting the C4 feature layer of the Centernet network as the visualization target layer. A heat map is generated by weighted summation of the feature map of the defect response area. Combined with the domain knowledge graph, a full-chain interpretation report on the correlation between defects, key features, physical mechanisms, and impact levels is generated to assist operators in understanding the decision-making logic and formulating targeted handling solutions. A dynamic matching model of defect classification and hazard level is built, with preset safety thresholds for application scenarios: for automotive safety components (such as body bonding), the threshold for minor defects is ≤0.5mm², moderate defects ≤2mm², and severe defects >2mm²; for ordinary packaging, minor defects ≤3mm², moderate defects ≤8mm², and severe defects >8mm²; for electronic component packaging, minor defects ≤0.3mm², moderate defects ≤1mm², and severe defects >1mm². The classification thresholds are dynamically adjusted according to the workpiece usage scenario to classify defects into three levels: minor, moderate, and severe, and output identification results including defect location, type, level, and hazard assessment.

[0012] Preferably, the twin compensation stage establishes a two-layer compensation strategy system based on the defect identification results output from the meta-learning recognition stage, combining the correspondence between defects and control parameters with digital twin pre-simulation; based on the hot melt adhesive rheological property prediction model, historical compensation cases, and process parameter constraints, the input parameters of the hot melt adhesive rheological property prediction model are adhesive temperature, adhesive type, and ambient humidity, and the output is viscosity value and flow rate, establishing a preliminary correspondence between defects and parameters. The adaptation rule is to match the model parameter coefficients according to the adhesive type (EVA / POE / PU, etc.), and adjust the prediction correction coefficients according to the ambient temperature range (5-15℃ / 15-30℃ / 30-40℃), clarifying the correlation between different defect types and levels and control parameters such as coating pressure, temperature, and speed; A digital twin model of the hot melt adhesive bonding process was built, integrating fluid dynamics simulation, thermodynamics simulation, and mechanical strength simulation modules. The fluid dynamics simulation used a structured mesh with a mesh size of 0.1mm × 0.1mm × 0.2mm, and the boundary conditions were set as pressure outlet (atmospheric pressure) and velocity inlet (consistent with the nozzle spray velocity). The thermodynamics simulation set the thermal conductivity of the hot melt adhesive to 1.2-1.8 W / (m·K), and the thermal conductivity of the workpiece substrate was adapted according to the material (200-400 W / (m·K) for metals and 0.2-0.5 W / (m·K) for plastics). The mechanical strength simulation used a tensile shear strength calculation model to recreate the entire process of adhesive layer spraying, spreading, and curing. Preliminary compensation parameters were input into the model for virtual verification, simulating the adhesive layer flow state, temperature distribution, bonding strength, and defect elimination effect under different parameter combinations, and outputting a simulation error analysis report. An improved NSGAⅢ multi-objective optimization algorithm was adopted, with a population size of 100, 50 iterations, a crossover probability of 0.8, and a mutation probability of 0.05. The core objective was quantified as follows: Defect elimination rate = (Number of defects before compensation - Number of defects after compensation) / Number of defects before compensation × 100%. The bond strength was quantified based on the tensile shear strength test results from digital twin simulation, in MPa. Energy consumption was quantified as the power consumption of the adhesive coating system per unit time, in kWh. Equipment wear was quantified as the cumulative wear of the actuator, in μm. The algorithm input consisted of defect type, level, and location data from the defect identification results, as well as the process parameter constraints. The adhesive coating pressure was 0.3-0.8M. The parameters are: Pa, coating temperature: 150-200℃, nozzle moving speed: 50-200mm / s. The output is an optimized combination of process parameters such as coating pressure, temperature, and nozzle position. Combining simulation results with real-time production cycle requirements, the compensation parameter combination is iteratively optimized. For glue shortage defects, the coating pressure, nozzle moving speed, nozzle orifice diameter, and pipeline heating temperature are optimized simultaneously. For glue line misalignment defects, the nozzle position, workpiece conveying speed, and coating sequence are optimized collaboratively. A compensation strategy priority mechanism is constructed to dynamically adjust the optimization target weight according to the severity of defects and production cycle requirements. Severe defects prioritize defect elimination and bonding strength, while minor defects balance quality, efficiency, and energy consumption, generating compensation instructions.

[0013] Preferably, the edge execution stage is based on the EtherCAT industrial bus to build a real-time collaborative control architecture at the edge; the compensation parameter optimization algorithm is deployed on the edge computing node, and the nozzle position, glue application pressure and workpiece conveying speed are synchronously adjusted through the servo drive system to ensure that the time synchronization accuracy of the linkage of multiple actuators is ≤5ms; The pressure control module adopts a composite control algorithm combining PID and reinforcement learning. The reinforcement learning uses "minimum absolute value of pressure deviation and lowest parameter adjustment frequency" as the reward function. The state space includes the current pressure value, target pressure value, and pressure change rate. The action space is the adjustment step size of the PID parameters (±0.01~±0.1). Parameter evaluation and optimization are completed every 10ms. The PID parameters are optimized in real time through reinforcement learning. The temperature control module adopts a segmented temperature control strategy, dividing the glue application process into a preheating section, a working section, and a heat preservation section. It dynamically adjusts the heating power based on infrared temperature feedback data, achieving independent control of the hot melt adhesive nozzle temperature and the glue tank temperature, with a temperature fluctuation range of ≤±2℃. The preheating section nozzle temperature is 120-140℃, and the glue tank temperature is 110-130℃; the working section nozzle temperature is 150-200℃, and the glue tank temperature is 140-190℃, adapted to different adhesive types; the heat preservation section nozzle temperature is 130-150℃, and the glue tank temperature is 120-140℃. An integrated actuator health monitoring module collects data such as servo motor current (normal range 5-15A), speed (normal range 1000-3000rpm), heating module power (normal range 500-1500W), and air pressure signal (normal range 0.4-0.8MPa) to establish a fault prediction model: when the current exceeds the normal range by ±20% for 3 consecutive seconds, the speed fluctuation exceeds ±10%, the power deviation exceeds ±15%, or the air pressure drop rate exceeds 0.05MPa / s, it is judged as a fault precursor; a three-level warning is triggered: mild precursors only provide a warning; moderate precursors adjust compensation parameters and provide a warning; severe precursors trigger a shutdown warning, identifying fault precursors such as nozzle blockage, pipeline leakage, and motor attenuation in advance, triggering a warning and adjusting compensation parameters; A real-time feedback channel for compensation effects is constructed, and the compensation effect is dynamically verified through synchronously collected images and sensor data, forming a closed loop of command issuance, execution, and feedback.

[0014] Preferably, the closed-loop optimization stage combines the compensation effect feedback data from the edge execution stage with the full-process production data to construct a full-link multi-level closed-loop iterative mechanism for data collection, model optimization, and knowledge updating; the first-level feedback is used to optimize the feature fusion and defect recognition model, and the compensated defect elimination effect data and new defect samples are input into the model in reverse. The data screening rule is: retain samples with a defect recognition accuracy of less than 95%, new defect samples, and samples corresponding to parameter combinations whose compensation effect does not meet the preset standard; incremental learning adopts the method of freezing the first 10 layers of the backbone network and only updating the parameters of the subsequent convolutional layers and fully connected layers, and updates the model parameters and feature weight rules in the knowledge graph through the incremental learning algorithm. A lightweight iteration is completed after each batch of production, and a model generalization ability verification and parameter solidification are performed after every 10 batches of production. The secondary feedback trigger condition is that the same type of defect compensation is ≥5 times, and it is iterated once after every 10 batches of production. The compensation data from multiple times is statistically analyzed, the optimal parameter range is extracted to form a process library, and the simulation parameters of the digital twin model are corrected at the same time. The three-level feedback is based on the accuracy of defect identification and the results of data quality assessment (specific indicators include data consistency ≥95%, signal stability fluctuation ≤±2%, cross-modal data matching degree ≥90%, and invalid data ratio ≤3%), and dynamically adjusts the sensor configuration scheme, sampling parameters and preprocessing strategies.

[0015] The beneficial effects of this invention are as follows: 1. This invention constructs an adaptive multi-source sensing acquisition system that can dynamically match the optimal sensor combination according to the workpiece material, specifications, and process, fuse environmental interference data, and achieve precise spatiotemporal alignment of multi-dimensional data; it filters effective information through cross-modal collaborative preprocessing, and then enhances feature extraction by combining a fusion network of domain knowledge graph and deep learning, and uses a meta-learning model to achieve rapid adaptation to new defects in small samples. It can also generate interpretable reports, improve the comprehensiveness, accuracy, and positioning accuracy of defect identification, and better adapt to the defect detection requirements of multi-scenario production.

[0016] 2. This invention constructs a two-layer compensation strategy system. First, it establishes a preliminary correspondence between defects and parameters by using a hot melt adhesive rheological property prediction model and historical cases. Then, it verifies the feasibility of the scheme through digital twin model pre-simulation. A multi-objective optimization algorithm is used to iteratively optimize the parameter combination, dynamically adjusting the optimization target weights based on the severity of defects. It relies on an industrial bus to achieve coordinated adjustment of multiple actuators, coupled with fault prediction and real-time feedback mechanisms, to achieve targeted, dynamic, and precise compensation, effectively improving the stability of bonding quality and adapting to the needs of mixed-line production of multi-specification workpieces.

[0017] 3. This invention constructs a full-link multi-level closed-loop iterative mechanism. The first-level feedback inputs the compensation effect data and new defect samples back into the model, updating the model parameters and knowledge graph rules through incremental learning. The second-level feedback statistically analyzes multiple compensation data, extracts the optimal parameters to form a process library, and corrects the digital twin simulation parameters. The third-level feedback dynamically adjusts the sensor configuration, sampling parameters, and preprocessing strategies based on the recognition accuracy and data quality. The full-link data connection forms a cycle of acquisition, recognition, compensation, and optimization, which not only continuously improves system performance but also reduces the cost of adapting to new defects and adjusting to new operating conditions, and reduces defect rework and equipment wear. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the precise identification process for hot melt adhesive bonding defects according to the present invention. Figure 3 This is a flowchart of the dynamic and precise compensation control process for defects in this invention. Figure 4 This is a flowchart of the full-link closed-loop iterative optimization process of the present invention. Detailed Implementation

[0019] 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.

[0020] As Figures 1 to 4 shown, an embodiment of the present invention provides an online identification and closed-loop compensation control method for hot-melt adhesive bonding defects, including a multi-source acquisition stage, a collaborative preprocessing stage, a knowledge fusion stage, a meta-learning identification stage, a twin compensation stage, an edge execution stage, and a closed-loop optimization stage. The specific implementation of each stage is as follows: Among them, in the multi-source acquisition stage, a multi-source sensing dynamic reconstruction acquisition system is constructed, integrating a high-frame-rate industrial camera, an ultrasonic sensor, an infrared temperature sensor, a pressure sensor, a servo motor encoder, and a sensor type adaptive selection module; this module is built with a process knowledge base retrieval engine, which can automatically match the optimal sensor combination and acquisition parameters according to the workpiece material (such as metal, plastic, composite material), specification size, and bonding process requirements. For example, for light-transmitting materials, the infrared and ultrasonic collaborative acquisition mode is strengthened, and for dark and highly reflective materials, the camera exposure dynamic range is increased to 120dB and an anti-glare filter is enabled.

[0021] The system is built with a Kalman filter fusion algorithm to fuse data of interference factors such as environmental temperature, humidity, and air flow speed in real time, dynamically adjust the sampling frequency and triggering timing of each sensor, and ensure the pertinence of data acquisition under complex working conditions. A millisecond-level spatio-temporal synchronization calibration mechanism is adopted, and through a dual calibration strategy of high-precision clock synchronization and image physical feature point matching, spatio-temporal precise alignment of multi-dimensional data such as images, acoustics, temperature, and mechanics is achieved, and a multi-dimensional structured data set with interference factor annotations and process parameter associations is generated.

[0022] An integrated sensor health status monitoring module is integrated to collect data such as the stability of the sensor output signal and calibration deviation in real time to diagnose the accuracy attenuation situation, trigger an adaptive calibration algorithm or perform redundant switching, effectively solving the technical problems of traditional sensors with fixed configuration, weak anti-interference ability, and easy accuracy attenuation.

[0023] Among them, in the collaborative preprocessing stage, based on the multi-dimensional structured data set obtained in the multi-source acquisition stage, a cross-modal collaborative preprocessing framework is designed to break through the limitation of single-modal independent preprocessing and achieve the联动 optimization of multi-modal data. For the image signal, a composite strategy of illumination adaptive Retinex enhancement combined with neighborhood mode constraint denoising is adopted, and the temperature signal collected synchronously is used to accurately judge the type of illumination interference. For example, for situations such as high-temperature radiation interference, sudden change of ambient light, and local shadow occlusion, denoising schemes such as Gaussian filtering and median filtering and the filter kernel size are adaptively selected; for the ultrasonic signal, a scheme of wavelet threshold adaptive denoising combined with temperature-acoustic feature correlation calibration is adopted, and the propagation speed deviation of ultrasonic waves in different media is corrected in real time through infrared temperature data, and the signal sampling window is optimized in combination with the bonding gap parameters to improve the detection accuracy of defects such as air bubbles and debonding in the adhesive layer.

[0024] For time-series signals such as temperature, pressure, and speed, a sliding window smoothing strategy combined with cross-modal redundancy verification is employed. Dynamic thresholds are set to eliminate abnormal fluctuations; for example, when pressure changes abruptly, the speed and temperature signals are linked for verification to ensure synchronous changes, avoiding misjudgment based on a single signal. A multi-level signal quality assessment model is constructed, which can not only determine signal validity based on indicators such as signal-to-noise ratio and data integrity, but also identify non-interference inherent signal fluctuations based on a process knowledge base, such as inherent signal fluctuations caused by specific workpiece materials. Valid information is preserved through signal enhancement algorithms, avoiding false triggering of sensor adjustments. An integrated preprocessing parameter adaptive optimization module iteratively updates preprocessing strategy parameters based on the detection results of different workpiece types, forming a parameter optimization library.

[0025] In the knowledge fusion stage, relying on the high-quality multimodal data output from the collaborative preprocessing stage, a collaborative feature fusion network combining a knowledge graph of the hot melt adhesive bonding field and deep learning is constructed to achieve a deep integration of data-driven approaches and physical mechanisms. Based on the rheological properties of hot melt adhesives, bonding physical mechanisms, material thermodynamic properties, and a large amount of historical defect data, an ontological modeling method is used to build a multimodal knowledge graph containing the full-link relationships of materials, processes, defects, features, and mechanisms, covering feature weight rules and association logic for adhesive types, workpiece materials, and typical defects.

[0026] For image signals, an improved MobileNetV2 network is employed. Prior constraints based on typical defect morphologies from a knowledge graph are introduced into the network's convolutional layers. Attention gating units enhance defect region feature extraction, improving the feature recognition and extraction efficiency for small defects such as bubbles smaller than 0.5 mm and adhesive defects as narrow as 0.3 mm. For ultrasonic and temperature signals, features such as peak amplitude, time-of-flight difference, and temperature gradient are extracted through time-domain analysis. After extracting parameters such as feature frequency and energy distribution through frequency-domain analysis, the knowledge graph is used for feature filtering and preliminary weight allocation, eliminating frequency components and feature parameters irrelevant to the current adhesive type and workpiece material characteristics.

[0027] A dynamic attention fusion mechanism is designed, which adaptively allocates the feature weights of each modality based on the relation weights of the knowledge graph and the preliminary judgment results of defects in real time. For example, when detecting interlayer bonding of composite materials, the weight of ultrasonic features is increased while the weight of surface visual features is weakened. When detecting the offset of adhesive lines on the surface, the weight of visual features is strengthened. Finally, a fusion feature vector with both data-driven accuracy and physical mechanism-supported reliability is generated, which solves the problems of traditional feature extraction being one-sided, having poor generalization ability, and being insufficiently adaptable to complex working conditions.

[0028] The meta-learning identification stage uses the fused feature vector generated in the knowledge fusion stage as input to construct a meta-learning-enhanced interpretable defect identification hierarchical model, enabling rapid adaptation and transparent decision-making for new defects in small samples. The main body of the model adopts an improved Centernet network, which improves defect localization accuracy by optimizing the backbone network depth and feature map fusion strategy. It integrates a meta-feature transfer module and, based on a model-independent meta-learning algorithm, can obtain general feature transfer capabilities from existing defect identification experience. Based on a small number of new defect samples (no more than 5), it can quickly fine-tune the model parameters to achieve rapid adaptation to new defect types without retraining the entire model, significantly reducing model update costs.

[0029] In terms of interpretability design, an improved Grad-CAM visualization technology is adopted, combined with a domain knowledge graph to generate a full-chain interpretation report linking defects, key features, physical mechanisms, and impact levels. For example, it marks the abnormal ultrasonic amplitude areas and abnormal image grayscale areas corresponding to bubble defects, and simultaneously links them to the mechanism of hot melt adhesive vaporization caused by excessively high temperature, quantifying the impact of bubbles on bond strength. This helps operators quickly understand the decision-making logic and formulate targeted handling solutions. A dynamic matching model for defect classification and hazard level is built, with preset safety thresholds for different application scenarios. The classification thresholds are dynamically adjusted according to the usage scenarios of workpieces such as automotive safety components, general packaging parts, and electronic component packaging, classifying defects into three levels: mild, moderate, and severe. The output includes identification results containing defect location, type, level, and hazard assessment, adapting to the mixed-line production needs of workpieces with multiple specifications and scenarios.

[0030] The twin compensation stage, based on the defect identification results output from the meta-learning recognition stage, establishes a two-layer compensation strategy system combining the correspondence between defects and control parameters with digital twin pre-simulation, achieving dual assurance of the accuracy and reliability of the compensation scheme. Based on the hot melt adhesive rheological property prediction model, combined with historical compensation cases and process parameter constraints in the domain knowledge graph, a preliminary correspondence between defects and parameters is constructed, clarifying the correlation between different defect types and levels and control parameters such as adhesive application pressure, temperature, and speed. A digital twin model of the hot melt adhesive bonding process is built, integrating fluid dynamics simulation, thermodynamics simulation, and mechanical strength simulation modules to recreate the entire process of adhesive spraying, spreading, and curing. Preliminary compensation parameters are input into the model for virtual verification, simulating the adhesive flow state, temperature distribution, bonding strength, and defect elimination effect under different parameter combinations, and outputting a simulation error analysis report.

[0031] An improved NSGAⅢ multi-objective optimization algorithm is adopted, with defect elimination, optimal bond strength, minimum energy consumption, and minimum equipment wear as core objectives. Combining simulation results with real-time production cycle requirements, the algorithm iteratively optimizes the combination of compensation parameters. For example, for insufficient adhesive defects, the algorithm simultaneously optimizes the application pressure, nozzle movement speed, nozzle orifice diameter, and pipeline heating temperature to ensure accurate adhesive replenishment. For adhesive line misalignment defects, the algorithm collaboratively optimizes the nozzle position, workpiece conveying speed, and application sequence. A compensation strategy priority mechanism is constructed, dynamically adjusting the weight of optimization objectives based on the severity of defects and production cycle requirements. For example, severe defects prioritize defect elimination and bond strength, while minor defects balance quality, efficiency, and energy consumption, generating directly executable, refined compensation instructions to improve the practicality and economy of the compensation strategy.

[0032] In the edge execution phase, a real-time collaborative control architecture is built on the EtherCAT industrial bus to enable rapid issuance of compensation commands and coordinated adjustment of multiple actuators. The compensation parameter optimization algorithm is deployed on edge computing nodes to avoid cloud transmission delays. A servo drive system enables synchronous and coordinated adjustment of nozzle position, adhesive pressure, and workpiece conveying speed. For example, when adjusting nozzle lateral offset compensation, the adhesive pressure and nozzle movement speed are simultaneously corrected to prevent secondary defects caused by adhesive volume fluctuations.

[0033] The pressure control module employs a composite control algorithm combining PID and reinforcement learning, optimizing PID parameters in real time through reinforcement learning. The temperature control module uses a segmented, precise temperature control strategy, dynamically adjusting heating power based on infrared temperature feedback data to achieve independent control of the hot melt adhesive nozzle temperature and the adhesive tank temperature. An integrated actuator health monitoring module collects data such as servo motor current, speed, heating module power, and air pressure signals to establish a fault prediction model. This model identifies early signs of faults such as nozzle blockage, pipeline leaks, and motor attenuation, triggering warnings and adjusting compensation parameters to avoid secondary defects.

[0034] A real-time feedback channel for compensation effects is constructed, and the compensation effect is dynamically verified by synchronously collected images and sensor data, forming a short-term closed loop of instruction issuance, execution, and feedback to ensure effective elimination of defects.

[0035] The closed-loop optimization phase combines the compensation effect feedback data from the edge execution phase with full-process production data to construct a multi-level closed-loop iterative mechanism encompassing data acquisition, model optimization, and knowledge updates, enabling continuous evolution of system performance. The first-level feedback focuses on feature fusion and defect identification model optimization. Compensated defect elimination effect data and new defect samples are input back into the model, and incremental learning algorithms update model parameters and feature weight rules in the knowledge graph. A lightweight iteration is completed after each batch of production, improving defect identification accuracy and generalization ability.

[0036] Secondary feedback optimizes the compensation strategy and digital twin model. It statistically analyzes multiple compensation data for defects of the same type and level, examines the correlation between compensation parameters and defect elimination effects, extracts the optimal parameter range to form a process library, and corrects the simulation parameters of the digital twin model based on actual production data to improve the predictive accuracy of the compensation strategy.

[0037] The three-level feedback enables the optimization of data acquisition and preprocessing strategies. Based on the accuracy of defect identification and the results of data quality assessment, the sensor configuration scheme, sampling parameters and preprocessing strategies are dynamically adjusted. For example, the sensor combination and preprocessing parameter thresholds are optimized for newly introduced workpiece materials to improve the system's adaptability to new working conditions.

[0038] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0039] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online identification and closed-loop compensation control of hot melt adhesive bonding defects, characterized in that, It includes the multi-source acquisition stage, collaborative preprocessing stage, knowledge fusion stage, meta-learning recognition stage, twin compensation stage, edge execution stage, and closed-loop optimization stage; Multi-source acquisition stage: Construct an adaptive multi-source sensing acquisition system. Based on the process knowledge base, dynamically match the optimal sensor combination and acquisition parameters according to the workpiece material, specifications and bonding process. Integrate interference factor data and achieve multi-dimensional data alignment through spatiotemporal synchronous calibration to generate a structured dataset with associated process parameters and interference annotations. Collaborative preprocessing stage: Collaborative preprocessing is performed on the collected multimodal data, and the noise reduction and calibration strategies are optimized in conjunction with cross-modal correlation information. Effective information is screened and invalid interference is eliminated through multi-level signal quality assessment to form standardized data. Knowledge fusion stage: Construct a feature extraction network that integrates a knowledge graph of the hot melt adhesive bonding field with deep learning, constrain the multimodal data feature extraction process based on the knowledge graph, and adaptively allocate the feature weights of each modality through a dynamic attention mechanism to generate fused features; Meta-learning recognition stage: Based on fused features, an interpretable defect recognition and hierarchical model is constructed, and the feature transfer capability is integrated to achieve rapid adaptation of new defects in small samples, and the recognition results include defect location, type, level and hazard assessment. Twin compensation stage: Establish a two-layer compensation strategy system that combines the correspondence between defects and control parameters with digital twin pre-simulation, iteratively optimize the combination of process parameters based on multi-objective optimization algorithm, dynamically adjust the optimization objectives according to the severity of defects, and generate compensation instructions; Edge execution phase: Based on the industrial bus, a real-time collaborative control architecture is built at the edge, which transforms compensation commands into multi-actuator linkage actions and integrates equipment fault prediction and real-time feedback mechanism for compensation effect; Closed-loop optimization phase: Construct a full-link multi-level closed-loop iterative mechanism for data acquisition, model optimization, and knowledge updating, and optimize the identification model, compensation strategy, and acquisition preprocessing strategy through three levels of feedback.

2. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 1, characterized in that, The multi-source acquisition stage constructs a multi-source sensing dynamic reconstruction acquisition system, integrating a high frame rate industrial camera, ultrasonic sensor, infrared temperature sensor, pressure sensor, servo motor encoder, and sensor type adaptive selection module; the adaptive selection module has a built-in process knowledge base retrieval engine, which can automatically match the optimal sensor combination and acquisition parameters according to workpiece information and bonding process requirements. The system uses a filtering algorithm to fuse data on interference factors, including ambient temperature, humidity, and airflow speed, and dynamically adjusts the sensor sampling frequency and trigger timing. It employs a millisecond-level spatiotemporal synchronization calibration mechanism to achieve spatiotemporal alignment of multi-dimensional data through clock synchronization and matching of physical feature points in the image. An integrated sensor health status monitoring module can diagnose sensor accuracy degradation in real time and trigger adaptive calibration algorithms or redundancy switching.

3. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 2, characterized in that, The collaborative preprocessing stage is based on a cross-modal collaborative preprocessing framework designed with a multi-dimensional structured dataset; an adaptive denoising and enhancement strategy combining temperature signals with illumination interference is adopted for image signals; and a wavelet threshold adaptive denoising combined with a temperature correlation calibration scheme is adopted for ultrasonic signals. For time-series signals including temperature, pressure, and speed, a sliding window smoothing strategy combined with cross-modal redundancy verification is adopted, and abnormal fluctuations are eliminated through dynamic thresholds. A multi-level signal quality assessment model is constructed to judge the validity of signals based on indicators including signal-to-noise ratio and data integrity, identify non-interference inherent signal fluctuations, and retain valid information. An adaptive optimization module for preprocessing parameters is integrated to iteratively update the preprocessing strategy parameters according to the detection effect of different workpiece types, forming a parameter optimization library.

4. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 3, characterized in that, The knowledge fusion stage constructs a collaborative feature fusion network that combines a knowledge graph in the hot melt adhesive bonding field with deep learning; Based on the rheological properties of hot melt adhesives, the physical mechanism of bonding, the thermodynamic properties of materials, and historical defect data, an ontological modeling approach is used to build a multimodal knowledge graph containing the full-link relationship of materials, processes, defects, features, and mechanisms. An improved MobileNetV2 network is used for image signals, and prior constraints on defect morphology in the knowledge graph are introduced to enhance the feature extraction of defect regions. For signals including ultrasound and temperature, after time-domain and frequency-domain features are obtained, the knowledge graph is used for feature selection and preliminary weight allocation. A dynamic attention fusion mechanism is designed. Based on the relation weights of the knowledge graph and the preliminary judgment results of defects, the weights of each modality feature are adaptively allocated to generate a fused feature vector.

5. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 4, characterized in that, The meta-learning recognition stage uses fused feature vectors as input to construct a meta-learning-enhanced interpretability defect recognition hierarchical model. The main body of the model adopts an improved Centernet network to improve the accuracy of defect localization, integrates a meta-feature transfer module, integrates general feature transfer capabilities based on model-independent meta-learning algorithms, and fine-tunes model parameters based on a small number of new defect samples to achieve rapid adaptation. The improved Grad-CAM visualization technology is used to generate a full-chain interpretation report of defects, key features, physical mechanisms and impact levels by combining domain knowledge graphs. A dynamic matching model of defect classification and hazard level is built, and the safety threshold of the application scenario is preset. The classification threshold is dynamically adjusted according to the workpiece usage scenario to classify defects into three levels: mild, moderate and severe, and output defect identification results.

6. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 5, characterized in that, The twin compensation stage establishes a two-layer compensation strategy system based on defect identification results; and constructs a preliminary correspondence between defects and control parameters by combining the hot melt adhesive rheological property prediction model, historical compensation cases and process parameter constraints. A digital twin model of the hot melt adhesive bonding process is built, integrating a multiphysics simulation module to recreate the entire process of adhesive spraying, spreading, and curing. Preliminary compensation parameters are input into the model for virtual verification, simulating the flow state, temperature distribution, bonding strength, and defect elimination effect of the adhesive layer under different parameter combinations, and outputting a simulation error analysis report. An improved NSGAⅢ multi-objective optimization algorithm is adopted, with defect elimination, optimal bonding strength, minimum energy consumption, and minimum equipment loss as the core objectives. The combination of compensation parameters is iteratively optimized by combining simulation results and production cycle requirements. A priority mechanism for compensation strategy is constructed to dynamically adjust the weight of optimization objectives according to the severity of defects and production cycle requirements, and to generate compensation instructions.

7. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 6, characterized in that, The edge execution phase is based on the EtherCAT industrial bus to build a real-time collaborative control architecture at the edge. The compensation parameter optimization algorithm is deployed on the edge computing node, and the synchronous linkage adjustment of nozzle position, glue application pressure and workpiece conveying speed is realized through the servo drive system. The pressure control adopts a composite control algorithm combining PID and reinforcement learning; the temperature control module adopts a segmented temperature control strategy, dynamically adjusting the heating power based on infrared temperature feedback data; and an integrated actuator health monitoring module is used to establish a fault prediction model to identify fault precursors, trigger warnings, and adjust compensation parameters. A real-time feedback channel for the compensation effect is constructed, and the compensation effect is dynamically verified by synchronously collected image and sensor data.

8. The method for online identification and closed-loop compensation control of hot melt adhesive bonding defects according to claim 7, characterized in that, The closed-loop optimization stage combines compensation effect feedback data and full-process production data to construct a full-link multi-level closed-loop iteration mechanism; the first-level feedback uses an incremental learning algorithm to input compensation effect data and new defect samples back into the model to optimize the identification model and update the model parameters and feature weight rules in the knowledge graph. Secondary feedback statistics are used to collect multiple compensation data for defects of the same type and level, extract the optimal parameter range to form a process library, and at the same time correct the simulation parameters of the digital twin model. The three-level feedback dynamically adjusts the sensor configuration scheme, sampling parameters, and preprocessing strategy based on the defect identification accuracy and data quality assessment results.

Images