An umbrella canopy heat sealing splicing process parameter adaptive optimization control method

Abstract

This invention provides an adaptive optimization control method for the heat-sealing splicing process parameters of umbrella canopies, comprising: collecting raw data such as heat conduction, deformation, and energy consumption using a multi-source heterogeneous sensor array; obtaining high-precision multi-dimensional process data through time-series calibration, noise removal, and data alignment; decoupling key thermodynamic features based on physical mechanisms to construct a searchable process semantic fingerprint and establishing a mapping library between states and optimal heat-sealing parameters; combining a lightweight meta-policy network, achieving parameter migration and online fine-tuning through minimal calibration learning to meet the rapid adaptation requirements of new materials or abrupt changes in working conditions; and integrating quality feedback throughout the process to achieve parameter optimization and closed-loop process control. This invention can effectively improve the quality and production efficiency of heat-sealing splicing, ensuring no warping at splicing edges, continuous adhesive lines, and controllable thermal damage.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology for umbrella manufacturing process parameters, and in particular to an adaptive optimization control method for the heat-sealing and splicing process parameters of umbrella surfaces. Background Technology

[0002] With the continuous upgrading of umbrella manufacturing technology, the heat-sealing and splicing process of the umbrella surface, as a key link determining the consistency of the finished product's appearance and durability, is evolving from traditional experience-based parameter setting to intelligent and automated control. Currently, most industrial applications use semi-automatic or fully automatic heat-sealing equipment, which uses closed-loop digital control of process parameters such as pressure head temperature, pressure, and action time to adapt to the production needs of different materials and splicing processes. In terms of achieving dynamic adjustment and optimization of process parameters, mainstream solutions generally rely on data-driven methods. This involves using online / offline process monitoring systems to collect a large amount of historical sensor data during the heat-sealing process, forming a training set through data annotation, and using machine learning models such as deep neural networks and support vector regression to automatically learn the mapping relationship between inputs (such as raw material properties, environmental variables, and process history) and outputs (process parameter optimization suggestions and adjustment values). This is then combined with feedback control strategies to drive the field equipment. In terms of industry development trends, some high-end manufacturing enterprises have successively deployed adaptive process optimization systems based on multi-source data fusion. These systems emphasize real-time acquisition and fusion analysis of dynamic signals during the heat-sealing process, thereby achieving closed-loop intelligent adjustment of parameters; simultaneously, they continuously expand their ability to identify material types and environmental disturbances, pursuing process precision, green energy saving, and flexible production. However, current technologies generally rely on large-scale labeled data-driven models and experience mapping relationships trained on single materials or specific working conditions. Although some manufacturers have attempted to introduce lightweight neural networks or meta-learning mechanisms, supplemented by some physical modeling, their fundamental approach has not yet broken through the optimization logic for "data-rich / fixed-scenario processes." Representative features of existing technologies include: a large number of intelligent heat-sealing systems based on full-process perception and end-to-end regression network optimization, emphasizing the completeness of data acquisition and the accuracy of closed-loop feedback. Their specific application scope covers common umbrella materials such as polyester, nylon, and EVA coatings, as well as relatively stable production environments. However, for novel polymer films, composite fiber materials, or manufacturing scenarios with significant temperature and humidity fluctuations and large batch-to-batch dispersion of raw materials, the adaptive adjustment and generalization capabilities of existing systems remain significantly insufficient. Furthermore, existing methods are mostly black-box models, with weak transferability and physical interpretation of process failure boundaries, which is detrimental to flexible adjustment and agile troubleshooting when switching between materials or extreme environments. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an adaptive optimization control method for the heat-sealing process parameters of umbrella canopies.

[0004] The technical solution of this invention is implemented as follows: an adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy, comprising: S1: Acquire three types of raw signals in the umbrella heat-sealing pressure head area: the heat conduction response curve, the interface deformation time sequence, and the local heat dissipation rate, which are simultaneously collected by the micro temperature gradient sensor, the micro-strain film, and the infrared heat flux probe. This forms a multi-source heterogeneous process sensing dataset. S2: Perform physical-driven feature decoupling processing on the multi-source heterogeneous process sensing dataset, extract three indicators with clear thermodynamic meaning from the heat conduction response curve: the thermal front advancement slope, the steady-state thermal resistance jump point, and the cooling hysteresis width, and generate a physical decoupling feature vector of the thermal process. S3: Construct a process semantic fingerprint based on the combination encoding of the physical decoupling feature vectors of the heat sealing process, and map each process semantic fingerprint to a typical heat sealing state among the fabric fiber softening critical state, adhesive layer melting and spreading state, or interface stress relaxation completion state, and establish an initial fingerprint-parameter mapping library; S4: Using calibration experimental data of a small amount of standard materials and the initial fingerprint-parameter mapping library, construct a lightweight meta-policy network containing a two-layer fully connected structure and a gated memory unit. Set the training objective of the network to learn to generate a feasible subset of parameters that meet the heat sealing quality constraints based on the current process semantic fingerprint and historical adjustment trajectory. S5: For new materials or sudden environmental scenarios, collect the original signal of short-term heat sealing test and decouple it in real time to generate a new semantic fingerprint. By calculating the similarity between the new semantic fingerprint and the known fingerprints in the initial fingerprint-parameter mapping library, retrieve the three nearest known fingerprints as the migration adaptation benchmark. S6: Based on the local weight parameters of the meta-policy network corresponding to the three known fingerprints, perform weighted fusion, execute zero-sample start-up and three rounds of fine-tuning convergence operations, and generate a dynamic heat-sealing control parameter set that adapts to the current complex process environment. S7: Input the dynamic heat sealing control parameter group into the heat sealing actuator to control the pressure head temperature, pressure and action time, complete the heat sealing and splicing operation of the umbrella surface and output the spliced ​​finished product; S8: Monitor the edge curling, glue line continuity and heat damage area of ​​the finished collage. If the detection results do not meet the heat sealing quality constraints, trigger the feedback adjustment mechanism to update the state of the gated memory unit of the lightweight meta-policy network. Otherwise, maintain the current dynamic heat sealing control parameter set for subsequent production.

[0005] The present invention provides an adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy, which has the following beneficial effects: (1) This invention proposes a lightweight knowledge representation mechanism based on “process semantic fingerprint” by reconstructing the physical essence of parameter learning, which significantly improves the system’s ability to quickly adapt to unknown working conditions. Traditional methods often use black-box data-driven models to directly fit the input-output relationship, which makes the model prone to failure under material property deviations or environmental disturbances. However, this scheme abandons the end-to-end fitting path of the original signal and instead extracts key features with clear thermodynamic meaning from physical signals such as temperature gradient, interface deformation and heat dissipation—such as “thermal front advancement slope”, “steady-state thermal resistance jump point” and “cooling hysteresis width”—and encodes them into interpretable and reusable “semantic fingerprints”, with each fingerprint corresponding to an identifiable thermo-physical state. This design makes the system no longer dependent on the assumption of global data distribution, but based on a structured understanding of the inherent laws of the process, thereby effectively overcoming the model degradation problem caused by changes in material type or external interference, and achieving high robustness transfer across material types. (2) This invention introduces a dynamic invocation mechanism for the meta-policy network driven by semantic fingerprint retrieval, constructing an adaptive control architecture that can achieve rapid convergence without large-scale training, significantly shortening the new material introduction cycle and reducing computational resource consumption. Unlike traditional schemes that rely on complex optimization algorithms or cloud-based iterative updates, the meta-policy network in this invention contains only two fully connected layers and a gated memory unit. Its function is not to predict a single optimal parameter, but to generate a subset of feasible parameters that meet quality constraints based on the most similar known patterns in the historical experience database matched with the current semantic fingerprint, combined with local weights, and automatically excluding obviously invalid combinations. When facing new materials, the system only needs to perform 3–5 short-term trials to extract new fingerprints, and calls the policy fragments corresponding to neighboring fingerprints through similarity retrieval, completing a lightweight adaptation process of "zero-sample start-up + three rounds of fine-tuning convergence" within the edge controller. The entire process does not rely on image recognition, digital twin simulation, or external knowledge base support. All calculations can be completed on local low-power devices with a response latency of less than 80ms and a communication bandwidth of less than 12KB / s required for parameter updates. This significantly reduces the deployment threshold and maintenance costs, making it particularly suitable for industrial scenarios where production lines frequently switch materials, lack stable network connections, or have stringent real-time requirements. Attached Figure Description

[0006] Figure 1 This is a flowchart of an adaptive optimization control method for heat-sealing splicing process parameters of an umbrella canopy according to the present invention; Figure 2 This is a sub-flowchart of an adaptive optimization control method for heat-sealing splicing process parameters of an umbrella canopy according to the present invention; Figure 3This is another sub-flowchart of the adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. like Figure 1 As shown, this invention provides an adaptive optimization control method for the process parameters of heat-sealing splicing of umbrella canopies, specifically including: S1: Acquire three types of raw signals in the umbrella heat-sealing pressure head area: the heat conduction response curve, the interface deformation time sequence, and the local heat dissipation rate, which are simultaneously collected by the micro temperature gradient sensor, the micro-strain film, and the infrared heat flux probe. This forms a multi-source heterogeneous process sensing dataset. S2: Perform physical-driven feature decoupling processing on the multi-source heterogeneous process sensing dataset, extract three indicators with clear thermodynamic meaning from the heat conduction response curve: the thermal front advancement slope, the steady-state thermal resistance jump point, and the cooling hysteresis width, and generate a physical decoupling feature vector of the thermal process. S3: Construct a process semantic fingerprint based on the combination encoding of the physical decoupling feature vectors of the heat sealing process, and map each process semantic fingerprint to a typical heat sealing state among the fabric fiber softening critical state, adhesive layer melting and spreading state, or interface stress relaxation completion state, and establish an initial fingerprint-parameter mapping library; S4: Using calibration experimental data of a small amount of standard materials and the initial fingerprint-parameter mapping library, construct a lightweight meta-policy network containing a two-layer fully connected structure and a gated memory unit. Set the training objective of the network to learn to generate a feasible subset of parameters that meet the heat sealing quality constraints based on the current process semantic fingerprint and historical adjustment trajectory. S5: For new materials or sudden environmental scenarios, collect the original signal of short-term heat sealing test and decouple it in real time to generate a new semantic fingerprint. By calculating the similarity between the new semantic fingerprint and the known fingerprints in the initial fingerprint-parameter mapping library, retrieve the three nearest known fingerprints as the migration adaptation benchmark. S6: Based on the local weight parameters of the meta-policy network corresponding to the three known fingerprints, perform weighted fusion, execute zero-sample start-up and three rounds of fine-tuning convergence operations, and generate a dynamic heat-sealing control parameter set that adapts to the current complex process environment. S7: Input the dynamic heat sealing control parameter group into the heat sealing actuator to control the pressure head temperature, pressure and action time, complete the heat sealing and splicing operation of the umbrella surface and output the spliced ​​finished product; S8: Monitor the edge curling, glue line continuity and heat damage area of ​​the finished collage. If the detection results do not meet the heat sealing quality constraints, trigger the feedback adjustment mechanism to update the state of the gated memory unit of the lightweight meta-policy network. Otherwise, maintain the current dynamic heat sealing control parameter set for subsequent production.

[0009] Step S1: Acquire three types of raw signals—thermal conduction response curves, interface deformation time series, and local heat dissipation rate—simultaneously collected by a micro-temperature gradient sensor, a micro-strain thin film, and an infrared heat flux probe deployed in the umbrella heat-sealing pressure head region, forming a multi-source heterogeneous process sensing dataset. Specifically, this includes: S1.1: Plan the layout of a micro-sensor array for the contact interface of the umbrella heat-sealing pressure head. Based on the heat conduction path and stress distribution characteristics, determine the spatial installation coordinates of the micro temperature gradient sensor, micro-strain film and infrared heat flux probe, and generate a sensor spatial topology configuration scheme. For the contact interface of the umbrella heat-sealing pressure head, the required input objects for the layout planning are the installation positions and quantities of micro temperature gradient sensors, micro-strain films and infrared heat flux probes. It is necessary to obtain the spatial distribution model of the heat conduction path and the two-dimensional or three-dimensional numerical field of the interface stress distribution. Finite element meshing was performed on the spatial distribution model of the heat conduction path. The steady-state and unsteady-state heat flux density gradient fields from the surface of the pressure head to the interior of the umbrella were calculated. The node regions with significant temperature change rates were extracted as candidate installation points for temperature gradient sensors. Principal stress direction decomposition is performed on the numerical field of interface stress distribution to identify normal stress concentration areas and tangential stress concentration areas. Normal stress concentration areas are preferentially selected as candidate mounting points for micro-strain thin films to ensure high sensitivity of stress response. By combining the field-of-view characteristics and response bandwidth of the infrared heat flux probe, the heat dissipation rate distribution map is analyzed pixel by pixel to locate the heat dissipation peak region and calculate the projection distance to the center point of the pressure head, ensuring that the probe field of view covers the high heat dissipation region and avoids edge occlusion. Multi-objective optimization layout calculations are performed based on the spatial coordinates of candidate installation points. The optimization objectives are to maximize the complementarity of sensor signals and minimize spatial interference. A spatial topology configuration matrix is ​​constructed to establish a one-to-one mapping relationship between sensor categories and their installation coordinates. Through the above layout planning process, the heat conduction path and stress distribution characteristics of the previous step are transformed into a spatial topology configuration scheme, thereby achieving the accuracy and efficiency of sensor deployment. For example, on the interface of a heat-sealed pressure head for an umbrella canopy with a diameter of 850 mm, the pressure head surface and the umbrella fiber layer were divided into 4800 nodes using a finite element model of heat conduction. The temperature gradient field at each node was calculated, and a temperature gradient threshold was set. Temperature gradient sensor candidate points were obtained by filtering at ℃ / mm. Corresponding stress distribution model, the normal stress threshold was set to... MPa was used to screen 18 candidate sites for micro-strain thin films. For the infrared thermal flux probe, the field of view was set to... °, response bandwidth is Hz, locate 6 peak heat dissipation regions, and ensure the probe is installed at a distance from the center of the pressure head. Within mm. A multi-objective optimization algorithm was used to calculate the layout of the above candidate points, resulting in a spatial topology configuration matrix. In the matrix, the installation coordinates of the temperature gradient sensor, micro-strain film, and infrared heat flux probe are concentrated in the high-temperature region, normal stress concentration region, and high dissipation rate region of the indenter, respectively. Experimental results show that under this layout scheme, the complementarity of various sensor signals is enhanced, the correlation of acquired data is significantly improved, and the spatial interference between different sensors is reduced to a negligible level, providing an optimized physical basis for subsequent synchronous acquisition of multi-source heterogeneous signals. S1.2: Based on the aforementioned sensor spatial topology configuration scheme, a high-precision clock synchronization protocol is used to perform trigger timing calibration on the micro temperature gradient sensor, micro-strain thin film, and infrared heat flux probe to eliminate time jitter error in multi-channel signal acquisition and generate an original multi-channel sensing signal stream with a unified time reference. The input conditions are: the spatial installation coordinates of the miniature temperature gradient sensor, micro-strain film and infrared heat flux probe in the sensor spatial topology configuration scheme are clear, and each sensor hardware has an external clock synchronization interface function; Based on the aforementioned sensor spatial topology configuration scheme, for each micro temperature gradient sensor, micro strain film, and infrared heat flux probe, the initialization command of the high-precision clock synchronization protocol is invoked to establish a mapping relationship between a unified clock source and the internal sampling counter of each sensor. Perform clock deviation measurement operation by recording the timestamp of the first trigger time of each sensor under a uniform clock source period signal, calculating the deviation value relative to the reference sensor, and expressing it with nanosecond resolution; A timing calibration algorithm based on deviation values ​​is adopted to add delay compensation or advance trigger adjustment to the sampling trigger events of each sensor to ensure that the sampling events of all sensors occur synchronously under a unified time reference. Construct a time index table for multi-channel signal streams, and arrange the time-calibrated sampled data in chronological order under a unified sampling period to form a multi-channel original signal sequence with consistent timestamps. By employing a high-precision clock synchronization protocol and timing calibration processing method, the spatial topology configuration result from the previous step is transformed into a raw multi-channel sensor signal stream with a unified time reference, thereby achieving the elimination of time jitter errors in multi-source sampling and the effect of cross-sensor synchronous acquisition. For example, nine miniature temperature gradient sensors, six micro-strain films, and four infrared heat flux probes are deployed in the heat-sealing pressure head area of ​​the umbrella surface. The IEEE 1588 precision time protocol is used as the high-precision clock synchronization protocol, with a unified clock source frequency set to 10MHz. The initial deviation of the sampling counters inside the sensors is within... Microsecond interval. Calculated using the deviation measurement formula. in, For the reference sensor's first trigger timestamp, The timestamp of the first trigger moment of the sensor to be calibrated. This is the deviation value. The deviation value is converted into the number of sampling periods, through... Calculate the required compensation sampling period, where To unify the clock source frequency. For For a sensor with a time of 0.8 microseconds, when At 10MHz, After precisely adjusting the sampling event trigger delay by 8 cycles, all channels are sampled synchronously under a unified time reference. Verification shows that the maximum time jitter of the multi-channel signal stream within a 1-second sampling window is reduced to ±50 nanoseconds, significantly improving the timing consistency of multi-source heterogeneous process sensor data and providing a high-precision time reference for subsequent analog-to-digital conversion and multi-source data fusion. S1.3: Perform analog-to-digital conversion and noise filtering on the original multi-channel sensing signal stream with a unified time reference, and use an adaptive sliding window algorithm to remove electromagnetic interference spikes and environmental background noise to generate a denoised heat conduction response curve sequence, interface deformation time sequence and local heat dissipation rate sequence. S1.4: Based on the denoised heat conduction response curve sequence, interface deformation time sequence and local heat dissipation rate sequence, perform multi-source data frame-level fusion operation to map heterogeneous signals with different sampling frequencies to the same discrete time axis and generate a spatiotemporally aligned multidimensional process state data matrix. S1.5: Normalize and encapsulate the spatiotemporally aligned multidimensional process state data matrix, add timestamp tags and sensor identification labels according to the preset data communication protocol, and finally generate a standardized multi-source heterogeneous process sensing dataset.

[0010] Step S2: Perform physical-driven feature decoupling processing on the multi-source heterogeneous process sensing dataset, extracting three indicators with clear thermodynamic meanings from the thermal conduction response curve: the thermal front advancement slope, the steady-state thermal resistance jump point, and the cooling hysteresis width, to generate a physical decoupling feature vector for the thermal process. Specifically, this includes: S2.1: The thermal conduction response curve is subjected to baseline drift correction and high-frequency noise filtering to eliminate random disturbances introduced by sensor zero-point drift and environmental electromagnetic interference, and to obtain a standardized thermal conduction timing signal. S2.2: Perform sliding window first-order differential operation based on the standardized heat conduction time sequence signal to calculate the rate of temperature change per unit time and obtain the instantaneous heat flux density gradient sequence; S2.3: The instantaneous heat flux density gradient sequence is used to identify the linear rising segment and fit its slope parameter in order to quantify the initial propulsion efficiency of heat energy transfer into the fabric and obtain the thermal front propulsion slope. Based on the instantaneous heat flux density gradient sequence obtained in the previous steps, a detection window range is established to identify the linear rising section of the heat conduction response curve. The start and end time points of the detection window are set according to the continuous positive value interval of the temperature change rate. Local steady-state determination is performed on the instantaneous heat flux density gradient sequence within the detection window, and data segments containing abnormal peaks or reverse fluctuations are removed to ensure the accuracy of subsequent slope fitting. The least squares method is used to fit the linear rising segment data points after the rejection process to establish a linear model of heat flux density changing with time. The slope of the line is calculated as the original measure of the thermal front advancement slope. This slope quantifies the efficiency of heat energy transfer to the fabric per unit time. The difference between the fitted thermal front advancement slope and the preset thermal conduction physical benchmark value is corrected to correct the systematic deviation introduced by sensor response delay and sampling quantization error, and the corrected thermal front advancement slope index is obtained. Through the aforementioned processing method, the instantaneous heat flux density gradient sequence of the previous step is transformed into a quantitative physical index characterizing the initial propulsion efficiency of thermal energy, thereby achieving accurate state extraction of the early heat conduction stage of the heat-sealing process. For example, in a heat-sealing test of polyester fabric, the collected instantaneous heat flux density gradient sequence showed a stable positive value within the range of 0.8 seconds to 1.4 seconds, and the detection window was set to 0.8 seconds to 1.4 seconds. Through anomaly removal, the peak data point located at 1.05 seconds was removed, and the remaining data meeting the steady-state condition were retained. The least squares method was used to fit the retained data points, yielding the slope calculation formula for the linear model: in, For the i-th sampling time, This represents the instantaneous heat flux density value at that moment. The mean of the time series. The average heat flux density is used. The original slope value calculated using this formula is 35.6 W / (m²·s). This is then corrected for by comparing it with the baseline heat conduction advance value of 37.0 W / (m²·s). Considering a compensation factor of 0.05 seconds for sensor delay, the final thermal front advancement slope index after correction is 36.8 W / (m²·s). This index demonstrates high efficiency in heat transfer during the early stages of heat sealing of polyester fabrics. It is subsequently encapsulated into the physical decoupling feature vector of the heat sealing process, driving the generation of the process semantic fingerprint. Furthermore, it is verified that this index maintains high consistency in repeated experiments under the same temperature and humidity conditions, significantly improving the stability and reliability of the dynamic parameter adjustment strategy. S2.4: Detect the inflection point position of the transition from unsteady state to steady state based on the standardized heat conduction time sequence signal and calculate the temperature rise difference before and after, so as to characterize the phase change thermal resistance change characteristics generated at the moment of adhesive layer melting and obtain the steady-state thermal resistance jump point. S2.5: Based on the standardized heat conduction timing signal, extract the cooling trajectory after the heating is completed and measure the time span of a specific temperature difference range to evaluate the intensity of the hysteresis effect in the interface heat dissipation process, obtain the cooling hysteresis width, and encapsulate the thermal front advancement slope, steady-state thermal resistance jump point and cooling hysteresis width into a physical decoupling feature vector of the heat-sealing process.

[0011] like Figure 2 As shown, step S3 involves constructing a process semantic fingerprint based on the combination encoding of the physical decoupling feature vectors of the heat sealing process. Each process semantic fingerprint is mapped to a typical heat sealing state among the following: the critical state of fabric fiber softening, the state of adhesive layer melting and spreading, or the state of complete interface stress relaxation. An initial fingerprint-parameter mapping library is then established. Specifically, this includes: S3.1: Obtain the physical decoupling feature vector of the heat-sealing process generated in the previous steps, and perform weighted fusion processing on the three components of the thermal front advancement slope, steady-state thermal resistance jump point and cooling hysteresis width based on the preset normalized weight coefficients, so as to eliminate the difference in dimensions and highlight the dominant role of key thermodynamic indicators, and generate a standardized combination of heat-sealing state features. S3.2: Using the standardized heat-sealing state feature combination, perform a clustering operation based on physical thresholds. Based on three types of prior knowledge—the critical temperature range for fabric fiber softening, the range of adhesive layer melting viscosity, and the time constant of interface stress relaxation—discrete the continuous feature space into three typical heat-sealing state labels: the critical state for fabric fiber softening, the melting and spreading state of adhesive layer, or the completed state of interface stress relaxation. S3.3: Based on the typical heat-sealing state label and the corresponding standardized heat-sealing state feature combination, a unique process semantic fingerprint is constructed using a hash coding algorithm. The high-dimensional continuous features are mapped into a low-dimensional discrete coding sequence to form a process semantic fingerprint identifier with searchability and interpretability, which serves as a digital summary representing the core physical state of the current heat-sealing stage. The input conditions are the obtained typical heat-sealing state labels and the corresponding standardized heat-sealing state feature combinations, and the execution object is the multi-dimensional numerical feature vector in the combination. A feature dimension index mapping is established for the standardized combination of thermal state features. Each thermodynamic index is arranged in a fixed order and forms a high-dimensional continuous feature vector to ensure that the input data has a deterministic structure before encoding processing. The high-dimensional continuous feature vector is discretized into intervals, each continuous feature is divided into finite intervals according to a preset quantization precision, and each interval is assigned a unique integer label so that the subsequent hash encoding process can process the discretized data. The hash base value is constructed by performing a multinomial factorization operation on the discretized feature label sequence. The operation process is as follows: in, For hash base value, For the feature dimension, The preset weight coefficients for the i-th feature are: Let be the discrete label value of the i-th feature; The hash base value is subjected to a modulo operation to map it to a fixed-length encoding space to form a low-dimensional discrete codeword. The operation process is as follows: in, For encoded sequence values, This represents the maximum value of the encoding space. The encoded sequence value is combined with a typical heat-sealing state label to generate a unique process semantic fingerprint identifier, and retrieval index metadata is added to it, so that the identifier can be used as a key value for fast retrieval and also has physical interpretability. By forming a process semantic fingerprint identifier through hash encoding and state combination, the standardized heat-sealing state feature combination of the previous step is transformed into a low-dimensional, searchable digital summary with physical interpretation meaning, thereby realizing the input standardization for the subsequent fingerprint mapping library construction. For example, in the heat sealing process of polyester fabric, the input conditions are: thermal front advancement slope of 0.85, steady-state thermal resistance jump point of 5.2, and cooling hysteresis width of 1.35. The feature vector dimension is 3, and it is discretized into intervals with a preset quantization precision of 0.01, which are mapped to label values ​​of 85, 520, and 135 respectively. Let the weight coefficients w be 7, 11, and 13 respectively. The hash base value calculation process is as follows: Calculated The value is 8425. Let the maximum value of the encoding space be... The result of the modulo operation is 1000. =425. The encoded value 425 is combined with the status label "adhesive layer melt-spreading state" to generate the fingerprint identifier "425-GM", and index metadata is added. This fingerprint identifier can be quickly searched in the fingerprint mapping library, and physically corresponds to the typical state of the adhesive layer in melt-spreading. Verification results show that this identifier significantly shortens the average response time in the subsequent parameter retrieval process, improving the system's retrieval efficiency and control strategy generation speed; S3.4: For a small number of standard materials such as polyester, nylon and coated fabrics, calibration experiments are carried out to obtain the optimal pressure head temperature, pressure and action time parameter sets under each typical heat sealing state. The process semantic fingerprint identifier is stored in key-value association with the optimal pressure head temperature, pressure and action time parameter sets, and an initial fingerprint parameter mapping library containing state-parameter correspondence is initially constructed. The calibration experimental data input conditions required for associating and mapping typical heat-sealing state labels and process semantic fingerprint identifiers for polyester, nylon and coated fabrics are: a unique process semantic fingerprint encoding sequence and the corresponding standardized heat-sealing state feature combination generated in the previous step S3.3; Multiple heat-sealing trials were performed for each standard material. Under the critical state of fabric fiber softening, the state of adhesive layer melting and spreading, and the state of complete interface stress relaxation, the control parameter set that met the process consistency requirements was recorded by precisely adjusting the indenter temperature, contact pressure and action time. The ambient temperature and humidity and substrate moisture content during the process were stored as auxiliary annotation variables to ensure the contextual integrity of the state-parameter mapping. The optimal parameter sets for each typical heat-sealing condition were normalized, and the indenter temperature was mapped to a dimensionless temperature ratio according to material type and thickness distribution. ,in The current optimal pressure head temperature, Reference temperature for the corresponding material; Contact pressure is converted into a uniformly distributed stress value based on the clamping area. ,in For the total force, The surface area under stress is the interface; the duration of action is proportional to the integral ratio of heat flux. Dimensionless transformation is performed, where The integral of the actual heat flux. The reference heat flux is used to eliminate the influence of differences in material specific heat and thermal conductivity; When establishing key-value association mapping, the process semantic fingerprint identifier is used as the unique key value. The normalized head temperature ratio, normalized uniform stress value and normalized heat flux integral ratio are combined into parameter value tuples and stored in the form of hash table to ensure the efficiency and uniqueness of retrieval. By normalizing and key-value association of parameters, the process semantic fingerprint generated in the previous step is transformed into state-parameter correspondence data that can be reused under different standard material conditions, thereby realizing the construction of the initial fingerprint parameter mapping library and providing standardized input for subsequent meta-policy network calls. For example, in the molten and spread state of the adhesive layer on a polyester substrate, the calibration experiment yielded the optimal indenter temperature of 185℃, contact pressure of 150N, contact time of 3.5s, ambient temperature of 23℃, humidity of 50%, and moisture content of 3.2%. When mapping and normalizing the parameters, a reference temperature was set. =190℃, the normalized temperature ratio is calculated as follows: =0.9737; the interface stress area is 50cm², and the uniformly distributed stress value is normalized to =3.0 N / cm²; heat flux is based on the measured value of 15.8 kJ, with reference heat flux. =16.0kJ, normalization ratio is =0.9875. The semantic fingerprint of this process is associated with the parameter tuple (0.9737, 3.0, 0.9875) and stored as a key. The retrieval efficiency is verified in the standard library to be less than 1ms on average. This ensures that the corresponding heat sealing control parameter group can be generated quickly during the meta-policy network call process, and the output finished product has continuous glue lines and significantly improved interface flatness. S3.5: Perform consistency verification and redundancy removal on the initial fingerprint parameter mapping library. By calculating the Euclidean distance between adjacent process semantic fingerprint identifiers in the feature space, merge duplicate entries with similarity higher than a preset threshold and correct abnormal parameter records. Finally, generate a standardized initial fingerprint parameter mapping library with a compact structure and covering typical working conditions for subsequent meta-policy network calls.

[0012] like Figure 3 As shown, step S4 involves using calibration experimental data from a small amount of standard materials and the initial fingerprint-parameter mapping library to construct a lightweight meta-policy network containing a two-layer fully connected structure and a gated memory unit. The training objective of this network is set as learning to generate a subset of feasible parameters that satisfy the heat-sealing quality constraints based on the current process semantic fingerprint and historical adjustment trajectory. Specifically, this includes: S4.1: Based on the process semantic fingerprint vector in the initial fingerprint-parameter mapping library and the corresponding standard heat sealing control parameter group, construct a four-layer neural network topology structure including an input layer, a first fully connected hidden layer, a second fully connected hidden layer and an output layer, and embed a gated memory unit between the second fully connected hidden layer and the output layer to form the initial architecture of a lightweight meta-policy network, ensuring that the network has the ability to remember the timing adjustment trajectory; Based on the process semantic fingerprint vectors in the initial fingerprint-parameter mapping library and the corresponding standard heat sealing control parameter set, the input feature set is extracted and the input layer node mapping relationship is established. The components of each process semantic fingerprint vector are used as independent input nodes to form a complete input feature vector. The latitude scaling factor is adjusted on the input feature vector, and the rescaling coefficient matrix is ​​set according to the weight of each component in the heat-sealing quality constraint. The input data is subjected to linear weight transformation to ensure that features of different dimensions are delivered to the network at the same scale. When constructing the first fully connected hidden layer, the number of nodes is set to be an integer multiple of the number of nodes in the input layer, the weight matrix is ​​initialized to uniformly distributed random values, and an activation function is added to each node to enhance the nonlinear mapping capability. A second fully connected hidden layer is established and fully connected to the first fully connected hidden layer. The weight matrix is ​​initialized according to the variance balance principle to reduce the risk of gradient vanishing and to establish a stable data channel for the input of subsequent memory units. A gated memory unit is embedded between the second fully connected hidden layer and the output layer to map the hidden layer output into a two-component structure containing state vectors and control vectors. At the same time, initial weights and biases, including input gate, forget gate and output gate, are set for the gated variables to ensure that the network has the ability to remember the time-series adjustment trajectory. Through the above structural construction method, the standardized fingerprint and parameter mapping results of the previous step are transformed into a lightweight meta-policy network initial architecture with input, nonlinear mapping, temporal memory and control output capabilities, realizing the network's trainability and generalization foundation in a small sample environment. For example, in a heat-sealing production line for umbrella canopies, the initial fingerprint-parameter mapping library contains standard process data for polyester and nylon fabrics. The polyester fingerprint vector is [0.62, 1.15, 0.87], and the nylon fingerprint vector is [0.68, 1.12, 0.91]. The corresponding pressure head temperature, pressure, and action time parameter sets are 180℃, 45N, 3.2s and 175℃, 48N, 3.0s, respectively. The number of input layer nodes is set to 3, corresponding to the three components of the fingerprint vector. The rescalarization coefficient matrix is ​​set to diagonal elements [1.0, 0.8, 1.2]. After linear weight transformation, the polyester input vector is [0.62, 0.92, 1.044]. The number of nodes in the first fully connected hidden layer is set to 9, the weight initialization range is [-0.05, 0.05], and the activation function is ReLU. The second fully connected hidden layer has 6 nodes, and the weights are initialized using the variance equalization Glorot scheme. The input gate weight matrix of the gated memory unit is initialized in the range of [-0.03, 0.03], the forget gate weight matrix in the range of [-0.02, 0.02], and the output gate weight matrix in the range of [-0.04, 0.04], with all bias values ​​set to zero. After the structure is built, the network can receive the fingerprint vector of polyester or nylon and generate a fine feasible range of indenter temperature, pressure, and action time through two fully connected layers and the memory unit. The verification results show that the parameter combination generated by this architecture in a short training iteration can significantly improve the continuity of the glue line in the collage and maintain a stable indenter temperature curve under environmental temperature and humidity fluctuations. S4.2: Perform weight initialization processing on the initial architecture of the lightweight meta-policy network, use the physical decoupling feature vector of the heat-sealing process in the standard material calibration experimental data as input samples, generate a predicted heat-sealing control parameter set through forward propagation, compare the predicted heat-sealing control parameter set with the standard heat-sealing control parameter set, calculate the initial parameter deviation loss value, and provide an error benchmark for subsequent network training; The initial architecture of the lightweight meta-policy network is loaded with the decoupled feature vector of the heat-sealing process from the standard material calibration experimental data as input samples. The weight matrix of the input layer nodes is set to a random small-amplitude perturbation state to avoid the initial training falling into symmetry failure. The bias vector is set to zero to ensure the neutral baseline of the input feature components. The physical decoupled feature vector of the heat-sealing process is linearly transformed by the weight matrix and bias vector of the first fully connected hidden layer, and the unit node output is gated by a nonlinear activation function such as ReLU to generate the first hidden layer feature response matrix, ensuring that the feature space is initially mapped to the deep representation domain. The first hidden layer feature response matrix is ​​processed by the second fully connected hidden layer with similar weights, and the output is fed into the embedded gated memory unit. The input gate, forget gate and output gate are used to adjust the weights of the state vector of the time-series adjustment trajectory to achieve memory enhancement of the dynamic process sequence. The state vector processed by the gated memory unit is sent to the output layer, and the final weight product and bias summation operation is performed to generate a set of predicted heat sealing control parameters, including the predicted values ​​of pressure head temperature, contact pressure, and action time. A one-to-one correspondence is established between the predicted heat sealing control parameter set and the standard heat sealing control parameter set, and the initial parameter deviation loss value is calculated using the mean square error formula: in, This represents the initial parameter deviation loss value. To predict the components of the heat-sealing control parameters, These are the corresponding standard heat-sealing control parameter components. The number of parameter components; The mean squared error calculated above is used as the initial loss benchmark for training this lightweight meta-policy network, realizing a closed-loop derivation chain from the input feature vector to the loss value. By initializing weights and constructing a forward propagation loss benchmark, the results of the previous step are transformed into quantifiable network performance metrics, enabling gradient updates and generalization capabilities to be improved in subsequent training phases. For example, in the calibration experiment of polyester material, the thermal front advancement slope is set to 0.83 K / ms, the steady-state thermal resistance jump point temperature difference is set to 4.1 K, and the cooling hysteresis width is set to 1.9 s, forming an input sample vector [0.83, 4.1, 1.9]. After processing by the first hidden layer (weight matrix range ±0.05, bias 0), the hidden layer output [0.021, 0.164, 0.076] is obtained. Then, after the state is corrected by the second hidden layer and the gated memory unit, the state vector [0.013, 0.097, 0.052] is obtained. The output layer calculates the predicted pressure head temperature of 353.6 K, pressure of 81.3 N, and action time of 5.2 s. Substituting the predicted value and the standard value (357.0 K, 80.0 N, 5.0 s) into the mean square error formula, the loss is calculated as follows: The verification results show that the loss value is within a reasonable range during the initialization stage and can be significantly reduced in subsequent gradient descent training, ensuring the network's ability to quickly adapt to and control the state of polyester heat sealing under small sample conditions. S4.3: Execute the backpropagation algorithm based on the initial parameter deviation loss value to update the connection weight matrix of the first fully connected hidden layer and the second fully connected hidden layer in the lightweight meta-policy network. At the same time, adjust the internal state update threshold of the gated memory unit to minimize the difference between the predicted heat-sealing control parameter set and the standard heat-sealing control parameter set, and generate a pre-trained lightweight meta-policy network with basic generalization ability. Using the initial parameter deviation loss value, a hierarchical backpropagation operation is performed on the lightweight meta-policy network. First, the deviation gradient of the output layer is used as input to calculate the weight update amount and bias update amount corresponding to the second fully connected hidden layer. Stochastic gradient descent or improved momentum method is used to accumulate correction values ​​on the weight matrix and bias vector. For the output of the second fully connected hidden layer, the error gradient is back-calculated to the cell state and output state inside the cell according to the opening ratio of the gating state through the gated memory unit. The internal state update threshold is adjusted according to the memory decay function so that it meets the preset heat sealing process response speed constraint at the long and short time memory balance point. The updated error gradient is then passed to the first fully connected hidden layer, and the weight matrix and bias of this layer are corrected to ensure that the multi-dimensional feature mapping paths of the three parameters of temperature, pressure and action time remain consistent during training. The weight matrix update process can be achieved using the following formula: in, For updating the network connection weight matrix, This is the network connection weight matrix. For learning rate, Let be the gradient vector of the loss function with respect to the weights. The loss function is defined as the mean square error between the predicted set of heat-sealing control parameters and the standard set of heat-sealing control parameters. in, To predict output parameters, For ideal parameter values, The number of samples; Through the aforementioned hierarchical backpropagation and dynamic adjustment of the gating state, the initial architecture of the network is trained into a pre-trained lightweight meta-policy network with basic generalization capabilities, thereby achieving continuous convergence of prediction parameter bias. For example, in the initial training scenario for polyester fabric, the number of input samples is set to 120, and the learning rate is... Using momentum coefficient An improved SGD optimizer. The initial mean square error of the predicted heat-sealing control parameter set is... After 120 iterations of training, the error was reduced to During training, the state update threshold of the gated memory unit decreases exponentially according to the iteration cycle. Down to This allows the network to retain historical adjustment trajectories while enhancing its adaptation to new input fingerprints. The update magnitude of the weight matrix stabilizes at its initial value during the final training cycle. Within a certain range, the network parameters are ensured not to be over-adjusted, thus preventing disruption of the heat-sealing process stability. The pre-trained lightweight meta-policy network output by this implementation significantly improves parameter generation accuracy when facing heat-sealing splicing tasks of similar polyester materials, and exhibits a substantial improvement in stability across the combined deviation indices of pressure, temperature, and action time. S4.4: Utilize the pre-trained lightweight meta-policy network to receive historical adjustment trajectory sequences and current process semantic fingerprints, perform multi-round iterative reasoning operations, generate a subset of candidate feasible parameters, and introduce heat sealing quality constraints such as no warping at the splicing edge, glue line continuity greater than 98%, and thermal damage area less than 0.3 square millimeters. Perform feasibility screening on the subset of candidate feasible parameters and eliminate invalid parameter combinations that do not meet the constraints. Based on the input interface of the pre-trained lightweight meta-policy network, the historical adjustment trajectory sequence and the current process semantic fingerprint vector are loaded, and the initial parameter generation operation is completed by calling the combined inference path of the first fully connected hidden layer and the gated memory unit inside the network. In each round of iterative inference, the state transition function of the gated memory unit is used to retain and control the forgetting of the temporal information of the historical adjustment trajectory, and the feature components of the current process semantic fingerprint are weighted and mapped through the second fully connected hidden layer to synthesize candidate parameter combinations of pressure head temperature, pressure and action time. After generating the subset of candidate feasible parameters, heat sealing quality constraints are introduced. The quantification rule for no edge warping at the splicing edge is that the warping height is less than a preset threshold. A continuity of more than 98% in the adhesive thread corresponds to a break point distribution ratio that is less than the threshold. A thermal damage area of ​​less than 0.3 square millimeters corresponds to a pixel area statistic of less than a threshold. ; The compliance of each candidate parameter combination is determined using the following quality constraint feasibility screening formula: in This is the result of the feasibility assessment. The value of the warped edge height, For the continuity ratio of the adhesive line, The parameter group is considered feasible when all three conditions are met, representing the thermal damage area. This judgment formula eliminates invalid parameter combinations that do not meet the quality constraints, and retains a subset of candidate feasible parameters that meet the conditions. By repeating the above reasoning and screening process using a chain iteration method, the network gradually optimizes the size and quality distribution of the candidate set in each iteration. By using a feasibility screening process, the predicted parameters from the previous step are transformed into a subset of highly reliable and feasible parameters that meet the heat sealing quality constraints, thereby optimizing the network output and eliminating invalid combinations. For example, in an umbrella production line, the historical adjustment trajectory sequence contains the pressure head temperature, pressure, and action time variation curves for nearly 50 production cycles. The current process semantic fingerprint vector is {0.78, 0.62, 0.35}, corresponding to the normalized thermal front advancement slope, steady-state thermal resistance jump point, and cooling hysteresis width, respectively. The pre-trained lightweight meta-policy network generates 30 sets of candidate parameter combinations in the first round of inference, including the warp height threshold. Set to 0.12mm, adhesive line continuity threshold. The thermal damage area threshold is set to 0.98. The value is set to 0.3 mm². The actual measured warpage height, adhesive line continuity, and thermal damage area for each candidate parameter group are compared with a threshold using a formula. The formula uses a fixed iterative variable. The index represents the candidate set, and the judgment result retains 14 parameter combinations that satisfy the three quality constraints. These 14 combinations are used as a highly reliable and feasible parameter subset input to the subsequent fine-tuning training module. In actual trial production, the parameter sets significantly improved the flatness of the splicing interface, maintained a high level of adhesive line continuity, and controlled thermal damage within a small range, thus verifying the significant improvement effect of this screening process on the quality of dynamic heat sealing control parameters under complex process environments. S4.5: Based on the selected subset of feasible parameters and the corresponding heat sealing quality evaluation results, construct the final training objective function, fine-tune the pre-trained lightweight meta-policy network, so that the network learns to dynamically output the optimal dynamic heat sealing control parameter set according to the input process semantic fingerprint, complete the construction and deployment of the lightweight meta-policy network, and enable it to have the ability to quickly adapt to small samples in new material scenarios.

[0013] Step S5: For novel materials or sudden environmental scenarios, the original signal from a short-term heat-sealing test is collected and decoupled in real time to generate a new semantic fingerprint. By calculating the similarity between the new semantic fingerprint and known fingerprints in the initial fingerprint-parameter mapping library, the three nearest known fingerprints are retrieved as the migration adaptation benchmark. Specifically, this includes: S5.1: Acquire three types of raw signals synchronously collected by the micro-sensor array in the heat-sealing pressure head area of ​​the umbrella surface under the new material or environmental change scenario: heat conduction response curve, interface deformation time sequence and local heat dissipation rate. Use the physical driving feature decoupling algorithm to preprocess the multi-source heterogeneous process sensing dataset, extract the heat-sealing process physical decoupling feature vector containing the thermal front advancement slope, steady-state thermal resistance jump point and cooling hysteresis width, so as to construct the characterization basis of the new material process state to be matched. S5.2: Based on the physical decoupling feature vector of the heat sealing process, perform process semantic fingerprint encoding operation to map the combination of indicators with clear thermodynamic meaning into new semantic fingerprints that characterize the critical state of fabric fiber softening, the state of adhesive layer melting and spreading, or the state of complete interface stress relaxation, so as to form standardized process state encoding data that can be used for similarity comparison. S5.3: Call the established initial fingerprint-parameter mapping library, read the set of known process semantic fingerprints corresponding to the standard material stored therein, and use the multidimensional Euclidean distance metric algorithm to calculate the feature space distance value between the new semantic fingerprint and each known process semantic fingerprint in the set of known process semantic fingerprints, so as to quantify the degree of difference between the new material process state and the historical standard state. Based on the new semantic fingerprint encoding data generated by S5.2 and the calling conditions of the known process semantic fingerprint set, the input objects are limited to standardized low-dimensional process state encoding sequences and historical fingerprint encoding sequences stored in the initial fingerprint-parameter mapping library; The new semantic fingerprint encoding and each known process semantic fingerprint in the set are used as computation objects in turn, and the Euclidean distance metric method in the multidimensional feature space is used to perform differential calculation. By constructing a unified feature vector representation, the numerical values ​​of each dimension are corresponding to three types of structured indicators: the thermal front advancement slope, the steady-state thermal resistance jump point, and the cooling hysteresis width. Quantitative evaluation is achieved using precise numerical calculation formulas. The following formula is used: in, For the new semantic fingerprint in the first eigenvalues ​​of dimension To correspond to the feature values ​​of the known process semantic fingerprint in this dimension, Let be the Euclidean distance between the two in the feature space; For each pair of new semantic fingerprints combined with known fingerprints, the above operation is performed to obtain a sequence of distance values. This sequence of distance values ​​is then appended to the corresponding known fingerprint identifier to form a complete differential quantization dataset. This dataset can be used to directly characterize the physical similarity between the process state of new materials and the historical standard state, thus meeting the nearest neighbor retrieval input requirements of S5.4. By calculating the multidimensional Euclidean distance metric, the new semantic fingerprint features from the previous step are transformed into difference evaluation data with quantitative similarity indicators, thereby achieving accurate matching results for transfer adaptation benchmark screening. For example, under the standard polyester material conditions, the thermal front advancement slope of the known process semantic fingerprint in the initial fingerprint-parameter mapping library is 15.2, the steady-state thermal resistance jump point is 2.8, and the cooling hysteresis width is 4.6. The corresponding index values ​​for the new nylon fabric are 14.9, 2.6, and 4.8, respectively. The two types of fingerprints are constructed as three-dimensional feature vectors, and the Euclidean distance calculation formula is input: The calculated distance value is 0.412. This distance value is recorded in the differential quantization dataset and associated with the polyester fingerprint identifier. The above process is repeated, comparing the new semantic fingerprint of nylon with all known fingerprints in the mapping library one by one to form a complete distance value sequence. This sequence is then used in S5.4 to select the three nearest neighbor fingerprints with the smallest distance values, sorted by distance. In this example, the three nearest neighbor fingerprints of nylon fabric are the standard polyester state, the softened coated fabric state, and the molten nylon state, respectively. Their corresponding distance values ​​are all significantly lower than other fingerprint combinations, verifying the high accuracy of the distance metric in feature space similarity evaluation. S5.4: Execute the nearest neighbor retrieval strategy according to the feature space distance value, sort all known process semantic fingerprints in the initial fingerprint-parameter mapping library in ascending order of distance, and select the three known process semantic fingerprints with the smallest feature space distance value as the nearest known fingerprint set, so as to determine the historical process reference benchmark that is most similar to the new semantic fingerprint topology. The feature space distance value matrix of the new semantic fingerprint and the known process semantic fingerprint is sorted. The sorting algorithm is called to couple each distance value in the matrix with its corresponding process semantic fingerprint identifier, and all entries are sorted in ascending order of distance value to form a distance sorting list. The distance sorting list is subjected to distance threshold filtering. Entries with distance values ​​greater than the preset process similarity lower limit threshold are removed, thereby retaining a set of candidate fingerprints that have a high degree of matching in physical state representation. The filtered candidate fingerprint set is subjected to index position pruning processing, which directly extracts the first three entries at the beginning of the set while maintaining their original distance value sorting order, so as to preserve the similarity gradient with the new semantic fingerprint in subsequent weight calculation; For the three nearest known process semantic fingerprints selected, the complete feature vectors and process parameter records corresponding to them in the initial fingerprint-parameter mapping library are extracted, and a four-element information group containing fingerprint identifier, feature vector, distance value and association parameter is established as a standardized data structure for the nearest known fingerprint set. The nearest known fingerprint set is paired with the new semantic fingerprint topological feature vector to verify the high consistency of the selected entries in the feature space topology, ensuring that they can serve as a historical process reference benchmark for the new semantic fingerprint migration adaptation, and achieving accurate positioning of subsequent migration weight loading. By employing the aforementioned nearest neighbor retrieval strategy and item filtering process, the feature space distance matrix from the previous step is transformed into a set of nearest neighbor process semantic fingerprints with topological similarity guarantees, thereby enabling the determination of a high-precision historical reference benchmark for the process state of new materials. For example, in a short-time heat-sealing test of a new coated nylon material, the decoupled semantic fingerprints generated form distance values ​​of 0.85, 1.12, and 0.97 for polyester, nylon, and coated fabric fingerprints in the feature space and the initial fingerprint-parameter mapping library, respectively, along with several other entries. After sorting all distance values ​​using a fast sorting algorithm to form a list, a distance threshold of 1.5 is set to remove inferior matching entries. The first three items of the sorted list are then extracted, resulting in three records: polyester fingerprint (0.85), coated fabric fingerprint (0.97), and nylon fingerprint (1.12). For each record, feature vectors and parameter sets are extracted; for example, the indenter temperature of the polyester fingerprint. ℃, pressure MPa, action time The system establishes a four-element information group. During distance value verification, normalized Euclidean distance is used for calculation. After verifying that it meets the high consistency condition, the nearest neighbor process semantic fingerprint set is output and used for subsequent migration adaptation benchmark data package construction. In this scenario, the output set significantly improves the adaptability of the initial inference parameters in subsequent policy weight loading and enhances the process stability of the new material under zero-sample start-up conditions. S5.5: Based on the set of nearest known fingerprints, extract the corresponding local weight parameter index information of the meta-policy network, and generate a migration adaptation benchmark data package containing the three nearest known fingerprint identifiers and their associated weight parameters, so as to serve as the input basis for subsequent zero-sample initiation and three rounds of fine-tuning convergence operations, thus completing the key connection from new material perception to control policy migration.

[0014] Step S6: Based on the local weight parameters of the meta-policy network corresponding to the three known fingerprints, weighted fusion is performed, and zero-sample start-up and three rounds of fine-tuning convergence operations are executed to generate a dynamic heat-sealing control parameter set adapted to the current complex process environment. Specifically, this includes: S6.1: Obtain the local weight parameter matrix of the meta-policy network associated with each of the three known fingerprints, calculate the normalized weight coefficient based on the similarity distance between the new semantic fingerprint and each known fingerprint, and perform linear weighted fusion processing on the local weight parameter matrix of the meta-policy network to generate an initial dynamic heat sealing control strategy weight set with cross-material generalization ability. S6.2: Load the initial dynamic heat-sealing control strategy weight set into the lightweight meta-policy network, input the new semantic fingerprint generated by the current short-time heat-sealing test decoupling and the historical adjustment trajectory data, perform forward inference operation, and output a feasible parameter subset containing the initial values ​​of pressure head temperature, pressure and action time of zero sample period; The initial dynamic heat-sealing control strategy weight set generated by the S6.1 weighted fusion process is loaded into all neural network layers and gated memory units of the lightweight meta-policy network to ensure the consistency of the mapping between the weight matrix and the controller memory address space. Normalized linear transformation is performed on the new semantic fingerprint data vector to compress the physical decoupled feature components into the effective domain of the network input layer within the numerical range, and time-series encoding of historical adjustment trajectories is added to form a complete input sample matrix; The input sample matrix is ​​fed into the input layer of the lightweight meta-policy network. The intermediate feature embedding representation is obtained by weighted summation and nonlinear activation operation in the first fully connected hidden layer. This embedding representation is further combined with the temporal memory state for dynamic adjustment in the second fully connected hidden layer and gated memory unit. Using a forward inference process, based on the time-enhanced feature vector output by the gated memory unit, the initial prediction vectors for the pressure head temperature, pressure, and application time are calculated through the weight vector and bias term of the output layer. The prediction formula is as follows: in, To predict the output vector, This is the output layer weight matrix. The feature vector of the previous layer, For bias terms; The predicted output vector is truncated and corrected according to the preset physical constraint range, and the temperature, pressure and time values ​​that exceed the process safety range are eliminated, while the set of feasible parameters that meet the heat sealing quality constraints is retained. Through the above reasoning and screening process, the initial weight set of the previous step is transformed into a subset of feasible parameters under zero sample period, so as to quickly generate executable heat-sealing parameters in the absence of historical data of this material. For example, in a short-time heat-sealing test scenario for a novel polyester coated fabric, the input new semantic fingerprint vector includes a thermal front advancement slope of 0.85, a steady-state thermal resistance jump point of 12.4, and a cooling hysteresis width of 3.6. The historical adjustment trajectory is a time-series compressed encoding of three rounds of tests. After normalizing these features to the [0,1] interval, they are input into a lightweight meta-policy network. The first fully connected layer outputs a 64-dimensional embedding vector, and the second fully connected layer, combined with a gated memory unit, outputs a 48-dimensional temporal enhancement feature. The formula is applied... ,set up It is a 3×48 matrix. It is a 48-dimensional feature vector. Using a 3D bias vector, the initial values ​​of the indenter temperature (93.7°C), pressure (1.42°C), and action time (2.85°C) were calculated. After constraint interval correction, the temperature was set to 92.0°C, the pressure to 1.40°C, and the action time to 2.80°C. The generated subset of feasible parameters drove the heat-sealing actuator for the first round of trial production. During the verification process, the splicing edges were smooth and the adhesive lines were continuous, demonstrating the ability to significantly improve the process parameters to approach the target under zero-sample-period conditions. S6.3: Based on the feasible parameter subset of the zero-sample period, drive the heat sealing actuator to complete the first round of trial production, collect the real-time heat conduction response curve and interface deformation time sequence during the first round of trial production, calculate the deviation gradient between the actual heat sealing state and the target heat sealing quality constraint, so as to construct the heat sealing process error feedback vector for the first round. S6.4: Update and correct the state of the gated memory unit of the lightweight meta-policy network according to the heat sealing process error feedback vector, readjust the initial dynamic heat sealing control strategy weight set using the corrected gated memory unit state, and perform the second and third rounds of iterative fine-tuning operations to generate the converged optimized dynamic heat sealing control strategy weight set. Obtain the values ​​of each feature component in the error feedback vector of the first round of heat sealing process, call the current state matrix of the gated memory unit stored in the lightweight meta-policy network, construct the corresponding state update factor and forgetting factor based on the error feedback vector, and form a data input structure that can be used for internal state correction. Element-wise operations are performed on the state update factor and forgetting factor. A proportional adjustment strategy is used to adjust the memory retention ratio and the intensity of new information injection for each state unit. The updated gated state matrix is ​​then calculated using the following formula: in, For the updated gated state matrix, Forgetting factor matrix, This is the current state matrix of the gated memory unit. To update the factor matrix, This is the error feedback vector for the heat sealing process; The modified gated memory unit state matrix is ​​loaded into the original structure of the lightweight meta-policy network, triggering the second round of weight matrix adjustment operation. The connection weights of the first fully connected hidden layer and the second fully connected hidden layer are redistributed according to the output signal of the new state matrix. Forward reasoning calculation is performed on the adjusted weight matrix to generate the weight set of the second round of optimized dynamic heat sealing control strategy. The feasible parameter subset of its output is extracted and the heat sealing actuator is driven to complete the second round of trial production. The corresponding real-time heat conduction response curve and interface deformation time sequence are collected, and the new heat sealing process error feedback vector is calculated. The third-round fine-tuning module is invoked to input the heat-sealing process error feedback vector from the second round back into the corrected gated memory unit state matrix. The calculation of the update factor and forgetting factor and the readjustment of the weight matrix are repeated, and the optimized dynamic heat-sealing control strategy weight set for the third round is output. Through the above iterative fine-tuning operation, the initial dynamic heat sealing control strategy weight set is transformed into the converged optimized dynamic heat sealing control strategy weight set, so as to achieve high-precision adaptation of heat sealing parameters under new material scenarios. For example, during the trial production of nylon-coated umbrella fabric, the initial dynamic heat-sealing control strategy was set with the weighted concentrated pressure head temperature at 180 degrees Celsius, the initial contact pressure at 320 N, and the initial action time at 3.8 seconds. After the first round of trial production, the heat-sealing process error feedback vector was measured as [+5℃, -10N, +0.2s]. Substituting this into the above formula, the forgetting factor matrix α was set as... Update the factor matrix β as The corrected gated state matrix output signal was calculated. After loading this matrix into the network, the pressure head temperature corresponding to the second round weight set was set to 183 degrees Celsius, the contact pressure to 310 N, and the action time to 3.7 seconds. The error feedback vector after the second round of trial production was [+2℃, -5N, +0.05s]. An update was performed again, and the third round output pressure head temperature was set to 185 degrees Celsius, the contact pressure to 305 N, and the action time to 3.6 seconds. The continuity of the adhesive lines in the thermoplastic composite was significantly improved, the splicing interface had no curling edges, and the heat damage area was controlled within 0.25 square millimeters. This verified that this sub-step achieved round-by-round convergence of dynamic weights and precise optimization of process parameters in iterative fine-tuning. S6.5: Based on the converged optimized dynamic heat sealing control strategy weight set and the current new semantic fingerprint features, perform final parameter mapping decoding processing to generate a dynamic heat sealing control parameter set that adapts to the current complex process environment, ensuring that there is no edge curling at the splicing edge, the continuity of the adhesive line is greater than 98%, and the thermal damage area is less than 0.3 square millimeters.

[0015] Step S7: Input the dynamic heat-sealing control parameter set into the heat-sealing actuator to control the pressure head temperature, pressure, and action time, complete the heat-sealing and splicing operation of the umbrella surface, and output the spliced ​​finished product. Specifically, this includes: S7.1: Obtain the target pressure head temperature setpoint, target contact pressure setpoint, and target action time setpoint from the dynamic heat sealing control parameter group. Use the industrial bus protocol to parse the instruction data packet and generate a standard control instruction sequence containing the temperature closed-loop setpoint, pressure closed-loop setpoint, and timing trigger signal to establish a mapping channel between the digital control domain and the physical execution domain. S7.2: Based on the temperature closed-loop setpoint in the standard control command sequence, the pulse width modulation duty cycle of the built-in heating unit of the heat-sealing head is adjusted by proportional-integral-derivative method. The deviation between the infrared temperature measurement feedback loop and the temperature closed-loop setpoint is compared in real time to generate a dynamic heating drive waveform with thermal inertia compensation characteristics, so as to obtain a stable pressure head surface temperature field that meets the transient response requirements. S7.3: Based on the pressure closed-loop setpoint in the standard control command sequence, the output thrust of the servo electric cylinder is subjected to force-position hybrid control calculation. Combined with the interface contact stress data collected in real time by the micro-strain film, a feedforward compensation algorithm is executed to generate a high-precision pressure loading curve that eliminates the influence of mechanical clearance and friction nonlinearity, so as to achieve a uniform and constant normal contact pressure distribution at the umbrella splicing interface. S7.4: Using the timing trigger signal in the standard control command sequence as a synchronization reference, the synergistic window of the stable pressure head surface temperature field and normal contact pressure distribution is subjected to millisecond-level timing logic gating. Within the target action time set value range, the coupling state of the heating drive waveform and the pressure loading curve is locked to generate a spatiotemporal synchronous heat sealing process trajectory that meets the requirements of thermodynamic phase change kinetics. S7.5: Follow the spatiotemporal synchronous heat sealing process trajectory to perform pressure head lifting and holding actions, control the umbrella surface substrate to complete polymer chain segment diffusion and interface rearrangement in the molten spreading state, and release the normal contact pressure distribution and separate the heat sealing pressure head after reaching the preset cooling threshold, outputting an umbrella surface splicing product with continuous adhesive line structure and no edge curling defects. The closed-loop set values ​​of the target pressure head temperature, pressure and action time in the dynamic heat sealing control parameter group, along with the configuration information of the spatiotemporal synchronous heat sealing process trajectory, are used as the execution reference and loaded into the multi-axis drive and heating control unit of the heat sealing actuator. The pressure head lifting and pressing execution axis and the pressure loading execution axis are coordinated and scheduled under the trigger of the trajectory signal. By combining the position control curve of the lifting actuator with the temperature-pressure synchronization window parameters determined in the trajectory, the descent rate and stroke termination position of the pressure head are corrected in real time, so that the pressure head can quickly enter a stable contact state within the molten spreading state range. The thrust closed-loop data of the pressure loading actuator shaft is combined with the interface stress feedback signal for analysis to generate a dynamic pressure holding curve. A constant normal contact pressure is applied by a servo driver to keep the substrate interface stable under the mechanical conditions required for polymer chain diffusion and interface rearrangement. The pressure head lifting logic and the pressure holding curve are coupled and synchronized in the controller, so that the temperature field and pressure field form a fixed synergistic window, and the thermodynamic phase change process of the material is fully maintained within this window until the set cooling threshold temperature is reached. The cooling threshold detection signal is bound to the decompression command of the controller. At the instant when the temperature field reaches the threshold, the normal contact pressure distribution is released and the pressure head actuator shaft is driven to rise to a safe position, thereby achieving physical separation. Through the chain processing method of pressing head lifting-holding-decompression-separation, the spatiotemporal synchronous heat sealing process trajectory generated in the previous step is transformed into a physical execution action sequence, so as to realize the output of umbrella canopy splicing finished products with continuous glue line structure and no edge curling defects. For example, during the heat-sealing process of polyester umbrella material, the dynamic heat-sealing control parameter set sets the target pressure head temperature to 185℃, the contact pressure to 2.4kN, the action time to 4.8s, and the cooling threshold temperature to 85℃. After this parameter set is applied to the actuator, the pressure head descent rate is set to 30mm per second, and the pressure closed loop is triggered at 0.2mm from the substrate surface. The pressure closed loop is maintained at 2.4kN, and the interface stress feedback fluctuation amplitude does not exceed 0.05kN. The pressure holding curve is precisely applied by the servo driver, so that the interface stress and temperature field work together to complete the diffusion of polymer chain segments and interface rearrangement within the 4.8s action time. The cooling threshold detection uses an embedded infrared temperature probe. When the pressure head surface temperature drops to 85℃, the controller issues a decompression command to release the normal pressure and drive the pressure head to rise. The separation process takes less than 0.5s. The finished product testing results show that the continuity of the adhesive line structure is significantly improved, the warping height of the splicing interface is less than 0.15mm, and the heat damage area is controlled within 0.25mm², meeting the requirements of high-precision process quality.

[0016] Step S8: Monitor the edge curling, glue line continuity, and heat damage area of ​​the finished collage. If the detection results do not meet the heat sealing quality constraints, trigger the feedback adjustment mechanism to update the gated memory unit state of the lightweight meta-policy network; otherwise, maintain the current dynamic heat sealing control parameter set for subsequent production. Specifically, this includes: S8.1: Acquire high-resolution surface image data and three-dimensional contour scan data of the collage finished product, and use machine vision edge detection algorithm and laser triangulation principle to perform multimodal fusion processing on the data to extract the collage edge warping height value, glue line break point coordinate distribution set and thermal damage area pixel area statistics, and generate a collage quality quantitative feature vector containing three key indicators. S8.2: Based on the quantification feature vector of the collage quality, call the preset heat sealing quality constraint judgment logic module to perform threshold comparison calculation, and verify the collage edge warping height value with the maximum allowable warping threshold, the glue line continuity ratio derived from the glue line break point coordinate distribution set with the minimum continuity threshold, and the heat damage area pixel area statistics with the maximum damage area threshold one by one, and generate a Boolean flag bit representing whether the current process state is qualified for heat sealing quality compliance; S8.3: When the Boolean flag indicating compliance with heat sealing quality is unqualified, read the dynamic heat sealing control parameter group of the current cycle and the corresponding quantification feature vector of patchwork quality, use the error backpropagation mechanism to calculate the deviation gradient between the actual quality index and the ideal heat sealing quality constraint, and construct a feedback adjustment training sample triplet containing input fingerprint features, execution parameters and quality deviation information. S8.4: Based on the feedback-adjusted training sample triplet, locate the internal state register of the gated memory unit in the lightweight meta-policy network, perform gate signal reset and cell state incremental update operations, encode the quality deviation information into a gated forgetting factor and an input update factor, correct the weighted memory mode of the gated memory unit for the historical adjustment trajectory, and generate an updated gated memory unit state matrix with error compensation capability. S8.5: Replace the original gated memory unit state parameters in the lightweight meta-policy network with the updated gated memory unit state matrix to complete the online fine-tuning iteration of the model, and mark the current dynamic heat sealing control parameter group as parameters to be eliminated. Lock the next production cycle to call the corrected lightweight meta-policy network to regenerate the dynamic heat sealing control parameter group adapted to the current complex process environment, and realize the closed-loop feedback adjustment process.

[0017] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0018] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive optimization control of process parameters in the heat-sealing and splicing process of an umbrella canopy, characterized in that, Includes the following steps: S1: Obtain the multi-source heterogeneous process sensing dataset of the umbrella heat-sealing pressure head region; S2: Decouple the features of the multi-source heterogeneous process sensing dataset, extract the leading edge slope, steady-state thermal resistance jump point and cooling hysteresis width, and generate a physical decoupling feature vector of the thermal process. S3: Construct a process semantic fingerprint based on the combination encoding of the physical decoupling feature vectors of the heat sealing process, map each process semantic fingerprint to a typical heat sealing state, and establish an initial fingerprint-parameter mapping library; S4: Using a small amount of standard material calibration experimental data and the initial fingerprint-parameter mapping library, construct a meta-policy network, and set the training objective of the meta-policy network as learning to generate a feasible parameter subset based on the current process semantic fingerprint and historical adjustment trajectory; S5: For novel materials or sudden environmental scenarios, collect the original signals of short-term heat sealing tests and decouple them in real time to generate new semantic fingerprints. By calculating the similarity between the new semantic fingerprint and the known fingerprints in the initial fingerprint-parameter mapping library, retrieve the three nearest known fingerprints. S6: Based on the local weight parameters of the meta-policy network corresponding to the three known fingerprints, perform weighted fusion, execute zero-sample start-up and three rounds of fine-tuning convergence operations, and generate a dynamic heat-sealing control parameter set. S7: Input the dynamic heat sealing control parameter group into the heat sealing actuator to control the pressure head temperature, pressure and action time, complete the heat sealing and splicing operation of the umbrella surface and output the spliced ​​finished product.

2. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, Following S7, the following also includes: S8: Monitor the edge curling, glue line continuity and heat damage area of ​​the finished collage. If the detection results do not meet the heat sealing quality constraints, trigger the feedback adjustment mechanism to update the state of the gated memory unit of the lightweight meta-policy network. Otherwise, maintain the current dynamic heat sealing control parameter set for subsequent production.

3. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, The multi-source heterogeneous process sensing dataset includes three types of raw signals: thermal conduction response curves, interface deformation time series, and local heat dissipation rate.

4. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, The specific decoupling of the features involves: performing baseline drift correction and high-frequency noise filtering on the thermal conduction response curve to obtain a standardized thermal conduction time series signal; extracting the thermal front advancement slope using the sliding window difference and least squares method based on the standardized thermal conduction time series signal; extracting the steady-state thermal resistance jump point and cooling hysteresis width by using the temperature rise difference at the steady-state transition point and the cooling interval duration, respectively; and combining and encapsulating the thermal front advancement slope, the steady-state thermal resistance jump point, and the cooling hysteresis width into a physical decoupling feature vector of the thermal process.

5. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, S3 specifically includes: Based on the preset normalized weighting coefficients, the three components of the thermal front advancement slope, steady-state thermal resistance jump point and cooling hysteresis width in the physical decoupling feature vector of the heat-sealing process are weighted and fused to generate a standardized combination of heat-sealing state features. Using the standardized combination of heat-sealing state features, a clustering operation based on physical thresholds is performed. Based on three types of prior knowledge—the critical temperature range for fabric fiber softening, the range of adhesive layer melting viscosity, and the time constant of interfacial stress relaxation—the continuous feature space is discretized into three typical heat-sealing state labels: the critical state for fabric fiber softening, the state for adhesive layer melting and spreading, and the state for complete interfacial stress relaxation. Based on the typical heat-sealing state label and the corresponding standardized heat-sealing state feature combination, a unique process semantic fingerprint is constructed, which maps high-dimensional continuous features into low-dimensional discrete coding sequences to form a process semantic fingerprint identifier. For a small number of standard materials, calibration experiments are conducted to obtain the optimal indenter temperature, pressure and action time parameter sets under various typical heat sealing conditions. The process semantic fingerprint identifier is then stored as a key-value association with the optimal indenter temperature, pressure and action time parameter sets to construct an initial fingerprint parameter mapping library.

6. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 5, characterized in that, S3 further includes: The initial fingerprint parameter mapping library is subjected to consistency verification and redundancy removal. By calculating the Euclidean distance between adjacent process semantic fingerprint identifiers in the feature space, duplicate entries with similarity higher than a preset threshold are merged and abnormal parameter records are corrected to generate a standardized initial fingerprint parameter mapping library.

7. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, S4 specifically includes: Based on the process semantic fingerprint vectors in the initial fingerprint-parameter mapping library and the corresponding standard heat sealing control parameter group, a four-layer neural network topology is constructed, including an input layer, a first fully connected hidden layer, a second fully connected hidden layer and an output layer. A gated memory unit is embedded between the second fully connected hidden layer and the output layer to form the initial architecture of the meta-policy network. The initial architecture of the meta-policy network is initialized with weights. The physical decoupling feature vector of the heat sealing process in the standard material calibration experimental data is used as the input sample. The predicted heat sealing control parameter set is generated by forward propagation. The predicted heat sealing control parameter set is compared with the standard heat sealing control parameter set to calculate the initial parameter deviation loss value. Based on the initial parameter deviation loss value, the backpropagation algorithm is executed to update the connection weight matrix of the first fully connected hidden layer and the second fully connected hidden layer in the meta-policy network. At the same time, the internal state update threshold of the gated memory unit is adjusted to generate a pre-trained meta-policy network. The pre-trained meta-policy network is used to receive historical adjustment trajectory sequences and current process semantic fingerprints, perform multiple rounds of iterative reasoning operations, generate a subset of candidate feasible parameters, introduce heat sealing quality constraints, perform feasibility screening on the subset of candidate feasible parameters, eliminate invalid parameter combinations that do not meet the constraints, and obtain the filtered subset of feasible parameters.

8. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 7, characterized in that, S4 further includes: Based on the selected subset of feasible parameters and the corresponding heat-sealing quality evaluation results, the final training objective function is constructed, and the pre-trained meta-policy network is fine-tuned to complete the construction and deployment of the meta-policy network.

9. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 1, characterized in that, S5 specifically includes: The original signals synchronously collected by the micro-sensor array in the heat-sealing pressure head area of ​​the umbrella surface under the new material or environmental change scenario are obtained, and the original signals are preprocessed to extract the physical decoupling feature vector of the heat-sealing process. Based on the physical decoupling feature vector of the heat sealing process, the process semantic fingerprint encoding operation is performed to map the combination of indicators with clear thermodynamic meaning into a new semantic fingerprint characterizing the critical state of fabric fiber softening, the state of adhesive layer melting and spreading, or the state of complete interface stress relaxation. Call the established initial fingerprint-parameter mapping library, read the set of known process semantic fingerprints corresponding to standard materials stored therein, and calculate the feature space distance value between the new semantic fingerprint and each known process semantic fingerprint in the set of known process semantic fingerprints; Based on the feature space distance value, a nearest neighbor retrieval strategy is executed. All known process semantic fingerprints in the initial fingerprint-parameter mapping library are sorted in ascending order of distance, and the three known process semantic fingerprints with the smallest feature space distance value are selected as the nearest known fingerprint set.

10. The adaptive optimization control method for the heat-sealing splicing process parameters of an umbrella canopy according to claim 9, characterized in that, The S5 also includes: Based on the set of nearest known fingerprints, the corresponding local weight parameter index information of the meta-policy network is extracted to generate a migration adaptation baseline data package containing the three nearest known fingerprint identifiers and their associated weight parameters.

Images