A wing folding system and method based on sanitary napkin products

Abstract

The present application relates to sanitary napkin production equipment technical field, disclose a kind of wing folding system and method based on sanitary napkin product.The system includes the real-time deformation observation stream of wing captured by sensing acquisition module, behavior modeling module constructs dynamic behavior atlas accordingly, virtual simulation module simulates complete folding process and extracts folding interference characteristics according to atlas, behavior mapping and control module maps characteristics into action instruction sequence, and mechanical execution module executes instruction to complete interference elimination and form correction.The method predicts physical interference by virtual simulation, and generates adaptive control instructions based on the prediction results, realizes the active and accurate control of wing folding process, solves the technical problems of unstable product quality and high scrap rate caused by the inability to adapt to material dynamic deformation in traditional rigid folding mode.

Description

Technical Field

[0001] This invention relates to the field of sanitary napkin production equipment technology, specifically to a wing folding system and method based on sanitary napkin products. Background Technology

[0002] On automated sanitary napkin production lines, the precise folding of the wings is a critical process affecting the product's appearance and reliability. Current mainstream technologies rely on pre-set mechanical molds and fixed trajectory control. Mechanical actuators perform stamping or bending operations on the moving wings according to pre-programmed, fixed motion paths. This open-loop control method treats the wing material as a rigid or simply flexible geometric shape, and its motion logic is based on idealized position signals.

[0003] Existing technical solutions have shortcomings. The preset fixed trajectory cannot adapt to the dynamic behavior of the material itself. Sanitary napkin wing materials are viscoelastic and non-uniform; in actual high-speed production, their movement is affected by factors such as tension fluctuations, material elastic recovery, and surface friction, resulting in random angular deviations, localized wrinkles, and unpredictable deformations. The fixed folding method interferes with these dynamic deformations, leading to skewed fold lines, excessive material stretching, or even tearing, creating stress concentration points. These problems cannot be detected and compensated for in real time under fixed program control; defective products can only be removed through subsequent manual sampling, resulting in poor product consistency, high scrap rates, and difficulty in adapting the equipment to the varying characteristics of different batches of material.

[0004] The core problem that this invention needs to solve is: how to enable the folding system to break free from its dependence on a fixed trajectory, so that it can understand and predict the dynamic behavior of the protective wing material in a real physical environment in real time, and generate adaptive control commands that can actively avoid or eliminate physical interference, thereby achieving precise and reliable folding. Summary of the Invention

[0005] The purpose of this invention is to provide a wing folding system and method based on sanitary napkin products to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a wing folding system based on a sanitary napkin product, the system comprising:

[0007] The sensing and acquisition module is used to capture the real-time deformation observation stream of the sanitary napkin wings in motion through multiple synchronous sensing channels. The real-time deformation observation stream includes angular offset waveform, tension pulsation spectrum and surface undulation point cloud.

[0008] The behavior modeling module is used to perform cross-scale spatiotemporal fusion of the real-time deformation observation stream to construct a dynamic behavior map of the wing material. The dynamic behavior map records the response mode and structural evolution path of the material under different stress phases.

[0009] The virtual simulation module is used to drive the virtual folding process based on the dynamic behavior map. In the virtual process, it simulates the complete deformation sequence of the wing from the deployed state to the folded state, records key interference events and stress concentration points, and then extracts the folding interference features including material interference depth, folding line offset trajectory and stress peak distribution map.

[0010] The behavior mapping and control module is used to input the folding interference features into the behavior mapping network and map the interference features in the virtual process into the action command sequence of the actual mechanical actuator.

[0011] The mechanical execution module is used to receive the sequence of action instructions and coordinate multiple execution units to execute the action instructions in order to eliminate interference and correct the shape of the protective wings.

[0012] Preferably, the step of performing cross-scale spatiotemporal fusion of the real-time deformation observation stream to construct a dynamic behavior map of the wing material includes:

[0013] Low-frequency trend components and high-frequency jitter components are decomposed from the angular offset waveform. The low-frequency trend component reflects the bending intention of the main body of the wing, and the high-frequency jitter component reflects the local unstable vibration of the material.

[0014] By performing cross-correlation analysis between the tension pulsation spectrum and the high-frequency jitter component, the tension source causing material vibration and its excitation frequency can be identified.

[0015] The surface undulation point cloud is divided into grids, and each grid cell is associated with the low-frequency trend component value at its corresponding position and the excitation frequency to form a grid attribute cell.

[0016] Based on the historical evolution of each grid attribute unit, predict its deformation tendency in the next time slice, which includes continued bending, springback, or arching;

[0017] The predicted deformation tendencies of all grid attribute units are collected and clustered according to spatial adjacency to generate multiple material behavior regions with similar behavior patterns. Each material behavior region and its corresponding evolution path together constitute the dynamic behavior map.

[0018] Preferably, driving the virtual folding process based on the dynamic behavior graph includes:

[0019] A virtual agent is assigned to each material behavior region, and each virtual agent performs a predetermined folding action in the virtual space according to the evolution path of its region.

[0020] A constitutive model of the protective wing material is established in virtual space. The constitutive model defines the stress-strain relationship of the material under combined tension, bending, and compression states.

[0021] When the folding actions of different virtual agents overlap in the virtual space, interference detection is triggered. The interference detection calculates the material stress distribution in the overlapping area based on the constitutive relation model.

[0022] If the calculated stress in the overlapping region exceeds the material yield threshold, a critical interference event is determined to have occurred, and the type, location, and amount of stress exceeding the critical interference event are recorded.

[0023] The virtual folding process is continuously run until the preset folding endpoint posture is completed. All key interference events and the location and time of the global stress peak during the entire process are summarized to form the key interference events and stress concentration point record.

[0024] Preferably, the extraction of folding interference features, including material interference depth, folding line offset trajectory, and stress peak distribution map, includes:

[0025] For each key interference event, the deformation gradient field of the material in the interference direction is calculated by expanding outward from the event location in the virtual space.

[0026] Integrating along the direction of the deformation gradient field yields the maximum depth from the interference surface to the yield point inside the material, which is taken as the material interference depth of the key interference event.

[0027] Connect the center points of all key interference events in chronological order to form a broken line describing the migration of the interference position, which serves as the offset trajectory of the broken line;

[0028] On the global timeline of the virtual folding process, the stress field is sampled at fixed time intervals, and the stress peak at each sampling moment is marked on the corresponding spatial position of the protective wing, finally generating a stress peak distribution map marked with the magnitude and time of the stress peak.

[0029] The material interference depth, fold line offset trajectory, and stress peak distribution map together constitute a set of interference features used to guide actual action adjustments.

[0030] Preferably, the step of inputting the folded interference features into a behavior mapping network to map the interference features in the virtual process into a sequence of action commands for the actual mechanical actuator includes:

[0031] The behavior mapping network includes a feature interpretation layer and an instruction synthesis layer, wherein the feature interpretation layer receives the folded interference features;

[0032] The feature interpretation layer converts the material interference depth into initial requirements for the travel and holding time of the pressing module, converts the fold line offset trajectory into tracking requirements for the position sequence of the guide module, and converts the stress peak distribution map into modulation requirements for the output curve of the tension application module.

[0033] The instruction synthesis layer performs instruction feasibility verification and conflict resolution based on the initial requirements, tracking requirements, and modulation requirements, combined with the physical limits and kinematic constraints of each mechanical actuator;

[0034] After the conflict is resolved, the command synthesis layer generates a pressing command package containing stroke nodes and time nodes for the pressing module, a guidance command package containing position sequence and velocity curve for the guidance module, and a tension modulation command package containing the tension value changing over time for the tension application module.

[0035] The pressing instruction package, the guidance instruction package, and the tension modulation instruction package are synchronized and packaged according to the process timeline to form a sequence of action instructions that drive the entire folding station to work together.

[0036] Preferably, the instruction synthesis layer, based on the initial requirements, tracking requirements, and modulation requirements, and in conjunction with the physical limits and kinematic constraints of each mechanical actuator, performs instruction feasibility verification and conflict resolution, including:

[0037] Verify whether the initial stroke requirement of the pressing module exceeds its maximum stroke range. If it does, scale the stroke and holding time proportionally to generate adjusted pressing parameters that meet physical limits.

[0038] The position sequence tracking requirements of the verification guidance module are checked to see if they exceed the boundary of its motion space. If they do, the position sequence is smoothed by interpolation and boundary clipping to generate an reachable guidance path.

[0039] Verify whether the tension modulation requirements of the tension application module exceed its maximum load or minimum response frequency. If they do, limit and filter the tension change curve to generate an executable tension control curve.

[0040] Check whether there is spatial interference or dynamic contradiction in the actions of the adjusted pressing parameters, the reachable guide path, and the executable tension control curve at the same point in time. If so, introduce time lag or adjust the execution order to eliminate the contradiction.

[0041] The final output is a set of coordination instructions that are conflict-free in terms of time, space, and capability, and can best satisfy the virtual process's adjustment intentions.

[0042] Preferably, the system further includes closed-loop verification processing performed after the action instruction sequence is generated:

[0043] While the mechanical actuator operates the actual protective wing according to the sequence of action commands, the response deformation flow of the actual protective wing is collected;

[0044] The actual interference features are extracted from the actual wing's response deformation flow. These actual interference features include the actual material interference amount, the actual folding line path, and the actual stress response point.

[0045] In the feature comparison space, the actual interference features are compared with the folded interference features extracted from the virtual process item by item, and the feature matching degree and deviation vector are calculated.

[0046] If the feature matching degree is lower than the preset standard, the mapping weight parameters in the behavior mapping network are adjusted in reverse according to the deviation vector;

[0047] By using the adjusted behavior mapping network, the folding interference characteristics of subsequent wings in the same batch are remapped to generate a corrected sequence of action commands, thereby achieving closed-loop optimization of folding control based on actual response.

[0048] Preferably, the step of comparing the actual interference features with the folded interference features extracted from the virtual process item by item in the feature comparison space includes:

[0049] Establish an interference comparison diagram with material interference as the horizontal axis and spatial position as the vertical axis. Plot the predicted material interference depth curve and the actual material interference curve in the diagram. Calculate the area enclosed by the two curves as the interference deviation.

[0050] On the spatial location-time plane, the predicted fold line offset trajectory and the actual fold line path are plotted respectively, and the average Euclidean distance between the two trajectories is calculated as the path following deviation.

[0051] On the stress peak distribution map, the predicted stress peak points are paired with the actual stress response points, and the weighted sum of the stress value difference between the paired points and the spatial distance is calculated as the stress response deviation.

[0052] The interference deviation, path following deviation, and stress response deviation are normalized and weighted to obtain a comprehensive feature fit score.

[0053] The deviation vector is a multidimensional vector, with each dimension component corresponding to the magnitude and direction of a local deviation.

[0054] Preferably, the system further includes a process for initializing and training the behavior mapping network:

[0055] A training sample library was constructed by collecting morphological data of the wings of different models of sanitary napkins during historical production processes and their corresponding successful folding mechanical parameters.

[0056] The morphological data set is input into the processing flow of cross-scale spatiotemporal fusion and virtual folding process simulation to obtain the corresponding predicted folding interference features;

[0057] The behavior mapping network is trained in a supervised manner by taking the predicted folding interference features as input and the corresponding set of successful folding mechanical parameters as the expected output.

[0058] During training, the connection weights of the nodes inside the behavior mapping network are adjusted using the gradient descent algorithm to minimize the difference between the action command sequence output by the network and the set of mechanical parameters for successful folding.

[0059] When the difference is lower than the convergence threshold, training is stopped, and the network weight parameters at this point are fixed as the initial parameters of the behavior mapping network.

[0060] Preferably, the present invention also includes a method for folding the wings of a sanitary napkin product, the method comprising all the modules and process flow of the wing folding system for sanitary napkin products as described above.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By capturing multi-dimensional deformation data such as angular offset, tension pulsation, and surface undulation of the protective wing in real time, and fusing them to construct a dynamic behavior map, the system can accurately record the material's actual response mode and structural evolution path under different stresses. This elevates the understanding of the material state from a static geometric position to a dynamic physical behavior level, providing key inputs reflecting instantaneous mechanical properties for subsequent control decisions and overcoming the insufficient control basis caused by traditional methods neglecting material properties.

[0063] Based on dynamic behavior maps, the complete folding deformation sequence can be pre-simulated in a virtual environment, accurately recording and quantifying interference events such as abnormal hooking and wrinkle formation, as well as areas of stress concentration during the simulation process. This process identifies potential offset trajectories and interference depths of the folding line in advance, moving the discovery and location of folding defects from post-physical detection to the simulation prediction stage before physical occurrence, thus achieving proactive anticipation of risks during the folding process.

[0064] The folding interference features extracted from virtual simulation are input into a behavior mapping network, which maps these multidimensional features characterizing specific problems into a coordinated sequence of mechanical action commands. This process achieves an autonomous transformation from "physical problem feature description" to "corrective action generation." The mechanical actuator operates according to this real-time generated command sequence, and its actions are adaptively adjusted to the currently predicted specific interference and deformation, thereby achieving proactive and precise intervention in the folding process, replacing the passive execution logic of traditional fixed programs. Attached Figure Description

[0065] Figure 1 This is a timing diagram of the wing folding system based on sanitary napkin products described in this invention;

[0066] Figure 2 A flowchart for constructing a dynamic behavior map of wing protection materials;

[0067] Figure 3 A flowchart for extracting folded interference features;

[0068] Figure 4 Comparison images before and after adjustment of the pressing module stroke;

[0069] Figure 5 The convergence curve of the training loss function for the behavior mapping network. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Please see Figure 1 This invention provides a folding system for sanitary napkin wings. The system includes: a sensing and acquisition module that synchronously captures real-time deformation observation streams of sanitary napkin wings during movement via multiple synchronous sensing channels deployed at key locations on the production line. These real-time deformation observation streams include angle offset waveforms acquired by an angle sensor, tension pulsation spectra acquired by a tension sensor, and surface undulation point clouds acquired by a 3D vision sensor. A behavior modeling module receives these real-time deformation observation streams and performs cross-scale spatiotemporal fusion processing to construct a dynamic behavior map reflecting the dynamic response law of the wing material. A virtual simulation module, based on this dynamic behavior map, drives a virtual folding process synchronized with the physical world in a digital environment, simulating the complete deformation sequence of the wing from an unfolded state to a folded state. During this process, it records information on all key interference events and stress concentration points, ultimately extracting folding interference features including material interference depth, folding line offset trajectory, and stress peak distribution map. A behavior mapping and control module has a pre-set behavior mapping network that receives the folding interference features as input and converts them into a sequence of action commands suitable for actual mechanical actuators through internal mapping relationships. The mechanical execution module receives the sequence of action commands and coordinates its internal multiple execution units, such as the pressing unit and the guiding unit, to precisely execute the commands, thereby performing interference elimination and shape correction operations on the actual protective wing.

[0072] Example 1: See Figure 2 The low-frequency trend component and high-frequency jitter component are decomposed from the angular offset waveform using digital filtering technology. The low-frequency trend component reflects the macroscopic bending intention of the wing body, while the high-frequency jitter component reflects the local microscopic instability and vibration of the material. Cross-correlation analysis of the tension pulsation spectrum and the high-frequency jitter component identifies the tension source causing the material vibration and its corresponding excitation frequency. Next, the surface undulation point cloud is gridded, dividing the wing surface into numerous grid cells. Each grid cell is associated with its corresponding spatial location's low-frequency trend component value and excitation frequency, forming grid attribute cells with multi-dimensional attributes. Based on the historical state evolution of each grid attribute cell within consecutive time slices, its deformation tendency in the next time slice is predicted. This deformation tendency is classified as continued bending, springback, or arching. The predicted deformation tendencies of all grid attribute cells are collected and clustered according to spatial adjacency relationships to generate multiple material behavior regions with similar behavioral patterns. Each material behavior region and its corresponding evolution path throughout the observation time constitute the dynamic behavior map.

[0073] In practical implementation, the wing-folding system based on sanitary napkin products performs cross-scale spatiotemporal fusion through a behavior modeling module to construct a dynamic behavior map. The real-time deformation observation stream captured by the sensing and acquisition module is the starting point for processing. The real-time deformation observation stream includes angle offset waveforms, tension pulsation spectra, and surface undulation point clouds. The angle offset waveforms are continuously acquired by a high-precision rotary encoder installed on the wing-grabbing mechanism, the tension pulsation spectrum is acquired by a miniature tension sensor embedded in the guide path, and the surface undulation point cloud is synchronously generated by a 3D line laser scanner mounted above the production line. In some embodiments, the decomposition of low-frequency trend components and high-frequency jitter components from the angle offset waveform is achieved using a Butterworth filter in digital signal processing. A low-pass filter with a cutoff frequency of 5Hz is used to extract the low-frequency trend component reflecting the bending intention of the wing body, while a high-pass filter with a cutoff frequency of 5Hz is used to extract the high-frequency jitter component reflecting local unstable vibrations of the material. The cross-correlation analysis between the tension pulsation spectrum and the high-frequency jitter component was performed in the frequency domain. The cross-correlation coefficient between the power spectral density of the tension pulsation spectrum and the power spectral density of the high-frequency jitter component was calculated. The frequency corresponding to the peak value of the cross-correlation coefficient was identified as the excitation frequency of the tension source that caused the material vibration.

[0074] It is understandable that the Delaunay triangulation algorithm is used when meshing the surface undulation point cloud, transforming the continuous point cloud into a surface model composed of triangular mesh cells. Each triangular mesh cell is associated with the low-frequency trend component value and excitation frequency of the timestamp corresponding to its geometric center point. The low-frequency trend component value is obtained from the waveform data through spatial interpolation, while the excitation frequency is directly associated, thus forming a mesh attribute cell containing three basic attributes: position, low-frequency trend component value, and excitation frequency. The deformation tendency is predicted based on the historical evolution of the mesh attribute cells through a state transition model. The model input is the attribute sequence of the mesh attribute cells in three consecutive time slices, and the output is the classification probability of the deformation tendency in the next time slice. The deformation tendency includes three categories: continued bending, springback, or arching. The prediction process can be expressed by the following relationship:

[0075]

[0076] in: Indicates that the grid attribute cells tend to be categorical. The probability vector, This represents the softmax activation function. and These are the model weights and bias parameters. These are the grid attribute cell attribute vectors for the previous two time slices and the current time slice, respectively.

[0077] Optionally, after aggregating the predicted deformation tendencies of all grid attribute units, clustering is performed according to spatial adjacency using a density-based clustering algorithm. The algorithm uses the physical distance between grid units and the Euclidean distance of the deformation tendency probability vector as a joint metric. Grid attribute units with similar behavior patterns are grouped into the same material behavior region. The deformation tendency evolution sequence of each material behavior region within the entire observation time window constitutes the evolution path of that region. All material behavior regions and their corresponding evolution paths together constitute a dynamic behavior map, which is stored in the form of graph-structured data. Nodes in the graph represent material behavior regions, and node attributes include the spatial contour and initial deformation tendency of the region. Edges represent the adjacency relationships between regions, and the weight of the edges is determined by the correlation coefficient of the deformation tendency sequence.

[0078] Example 2: See Figure 3An independent virtual agent is assigned to each material behavior region in the dynamic behavior map. Each virtual agent performs a predetermined folding action in virtual space according to the evolution path of its region. A constitutive model of the wing material is synchronously established in virtual space, defining the stress-strain relationship of the material under combined tension, bending, and compression states. When the folding actions performed by different virtual agents spatially overlap in virtual space, an interference detection program is triggered. This interference detection calculates the material stress distribution in the overlapping region based on the constitutive model. If the calculated stress value in the overlapping region exceeds a preset material yield threshold, a critical interference event is identified, and the type, location, and stress excess of the critical interference event are recorded. The virtual folding process continues until the preset folding endpoint posture is achieved. All critical interference events and the location and time of the global stress peak throughout the entire process are summarized to form a complete record.

[0079] The process of extracting folding interference features from the virtual folding process includes the following steps: For each key interference event recorded, the deformation gradient field of the material in the direction of interference is calculated by expanding outward from the event location in the virtual space. Integration is performed along the direction of this deformation gradient field to obtain the maximum depth from the interference surface to the yield point inside the material; this depth is defined as the material interference depth corresponding to the key interference event. The center points of all key interference events are connected in chronological order to form a spatial polygonal line describing the migration of the interference position over time; this polygonal line is the folding line offset trajectory. On the global timeline of the virtual folding process, the stress field of the entire wing model is sampled at fixed time intervals, and the stress peak value at each sampling moment is marked at the corresponding spatial location of the wing, ultimately generating a stress peak distribution map marked with the magnitude and time information of the stress peak. The material interference depth, the folding line offset trajectory, and the stress peak distribution map together constitute a set of interference features used to guide actual action adjustments.

[0080] In practical implementation, the virtual simulation module drives the virtual folding process based on a dynamic behavior map, provided by the behavior modeling module, which includes multiple material behavior regions and their evolution paths. An independent virtual agent is assigned to each material behavior region in the dynamic behavior map. Each virtual agent is a logical entity in the virtual space. Each virtual agent executes a predetermined folding action in the virtual space according to the evolution path of its assigned material behavior region. The trajectory and velocity of the folding action are calculated from the deformation tendency sequence recorded in the evolution path. In some embodiments, a constitutive model of the wing material is established in the virtual space. This constitutive model is defined using a hyperelastic body model, and the model parameters are set based on actual mechanical test data of the wing composite material. The constitutive model defines the nonlinear stress-strain relationship of the material under combined tension, bending, and compression states. When the folding actions executed by different virtual agents spatially overlap in the virtual space, an interference detection based on the finite element method is triggered. The interference detection program calls the constitutive model to calculate the stress distribution of the mesh nodes in the overlapping region. If the calculated equivalent stress value at any node in the overlapping area exceeds the preset material yield threshold (set based on the yield strength of the nonwoven fabric and film of the wing), a critical interference event is identified. The type of critical interference event is recorded as tensile overload, compressive buckling, or shear slip. The location of the critical interference event is recorded as three-dimensional coordinates, and the stress excess is recorded as the difference between the actual stress value and the yield threshold. The virtual folding process continues until the wing's three-dimensional model reaches the preset folding endpoint posture, defined by the product design three-dimensional drawings. All critical interference events and the location and time of the global stress peak throughout the process are summarized to form a record of critical interference events and stress concentration points.

[0081] Optionally, the operation of extracting folding interference features from the virtual folding process is performed on key interference events and stress concentration points. For each key interference event in the record, a cubic analysis region with a side length of 5 mm is extended outward from the three-dimensional coordinates of the key interference event location in the virtual space. The deformation gradient field of the material in the principal interference direction within the analysis region is calculated. Numerical integration is performed along the direction of the deformation gradient field. The integration path starts from the surface node where the interference occurs and extends into the material until the equivalent stress of the node on the path first falls below the material yield threshold. The length of this integration path is defined as the maximum depth from the interference surface to the yield point inside the material. This depth value is the material interference depth corresponding to the key interference event. The center points of all key interference events are connected in chronological order to form a spatial polygonal line describing the migration of the interference position over time. This spatial polygonal line is the folding line offset trajectory, and linear interpolation is used between the trajectory points. On the global timeline of the virtual folding process, the Mises equivalent stress field of the entire wing model is sampled at fixed time intervals of 10 milliseconds. The global stress peak value and its three-dimensional spatial location at each sampling moment are recorded as a data point. All data points together constitute a stress peak distribution map, which is a spatiotemporal data set. The material interference depth, folding line offset trajectory, and stress peak distribution map together constitute a set of interference features used to guide actual action adjustments.

[0082] It is understandable that the integral of the deformation gradient field involved in calculating the material interference depth can be expressed by the following relationship:

[0083]

[0084] in: The depth of material interference is represented by the integration path starting from the surface point where the interference occurs. From the beginning, to the point within the material where the yield condition is met. Finish, Indicates the path along the integral. The modulus of the deformation displacement gradient tensor is calculated automatically in the finite element post-processing program of the virtual simulation module.

[0085] Example 3: The process of mapping folding interference features into a sequence of action commands by the behavior mapping and control module is achieved through its internal behavior mapping network. This behavior mapping network includes a feature interpretation layer and a command synthesis layer. The feature interpretation layer receives the folding interference features. The feature interpretation layer converts the material interference depth into initial requirements for the stroke and holding time of the pressing module in the mechanical execution module, converts the folding line offset trajectory into tracking requirements for the position sequence of the guiding module, and converts the stress peak distribution map into modulation requirements for the output curve of the tension application module. Based on these initial requirements, tracking requirements, and modulation requirements, and combined with the physical limits and kinematic constraints of each mechanical execution mechanism, the command synthesis layer performs command feasibility verification and conflict resolution.

[0086] The system verifies whether the initial stroke requirement of the pressing module exceeds its maximum stroke range. If so, it scales the stroke and corresponding holding time proportionally to generate adjusted pressing parameters that meet physical limits. It also verifies whether the position sequence tracking requirement of the guidance module exceeds its motion space boundary. If so, it performs smooth interpolation and boundary trimming on the position sequence to generate a reachable guidance path. Finally, it verifies whether the tension modulation requirement of the tension application module exceeds its maximum load capacity or minimum response frequency. If so, it limits and filters the tension change curve to generate an executable tension control curve. The system checks whether there is spatial interference or dynamic contradiction in the actions of the adjusted pressing parameters, the reachable guidance path, and the executable tension control curve at the same time point. If so, it introduces time lag or adjusts the execution order to eliminate the contradiction. The final output is a set of coordinated instruction elements that are conflict-free in time, space, and mechanism capabilities, and that best satisfy the virtual process adjustment intent. Based on these coordinating instruction elements, the instruction synthesis layer generates a pressing instruction package containing stroke nodes and time nodes for the pressing module, a guidance instruction package containing position sequences and speed curves for the guidance module, and a tension modulation instruction package containing the tension value's variation over time for the tension application module. The pressing instruction package, guidance instruction package, and tension modulation instruction package are then synchronized, aligned, and packaged according to the process timeline to form a sequence of action instructions driving the entire folding station to work collaboratively.

[0087] In practical implementation, the behavior mapping and control module maps folded interference features into a sequence of action commands through its internal behavior mapping network. The folded interference features are provided by the virtual simulation module and include the material interference depth, fold line offset trajectory, and stress peak distribution map. The behavior mapping network consists of a feature interpretation layer and a command synthesis layer. The feature interpretation layer receives the set of folded interference features as input and is composed of multiple parallel feature processors.

[0088] In some embodiments, the feature interpretation layer converts the material interference depth into initial requirements for the stroke and holding time of the pressing module in the mechanical actuation module. The conversion process uses a linear mapping relationship, where the material interference depth value is proportionally mapped to the target stroke of the linear motor of the pressing module, and the stress excess associated with the same interference depth value is mapped to the holding time of the pressing action at the target stroke point. The feature interpretation layer converts the fold line offset trajectory into tracking requirements for the position sequence of the guidance module. The tracking requirements include a set of three-dimensional coordinate points ordered by time and the expected transition time between the coordinate points. The feature interpretation layer converts the stress peak distribution map into modulation requirements for the output curve of the tension application module. The modulation requirements are a discrete curve of the tension setpoint changing over time, and the stress peak magnitude at each sampling moment in the stress peak distribution map is mapped to the expected output force of the tension application module at that moment.

[0089] It is understandable that the instruction synthesis layer, based on the initial requirements, tracking requirements, and modulation requirements output by the feature interpretation layer, and combined with the physical limits and kinematic constraints of each mechanical actuator, performs instruction feasibility verification and conflict resolution. It verifies whether the initial stroke requirement of the pressing module exceeds its maximum stroke range, which is 50 mm. If the target stroke in the initial requirement exceeds 50 mm, the target stroke and corresponding holding time are scaled proportionally. The scaling ratio is the ratio of the maximum stroke range to the initial target stroke, generating adjusted pressing parameters that meet the physical limits. It also verifies whether the position sequence tracking requirements of the guidance module exceed its motion space boundary. The motion space boundary of the guidance module is defined by a cuboid region. If any coordinate point in the position sequence exceeds the cuboid region, the position sequence undergoes smooth interpolation and boundary clipping. Smooth interpolation uses a cubic spline interpolation algorithm, and boundary clipping projects the coordinate points exceeding the boundary onto the nearest boundary surface, generating a reachable guidance path.

[0090] Optionally, the tension modulation requirement of the tension application module is checked to see if it exceeds its maximum load or minimum response frequency. The maximum load of the tension application module is 10 Newtons, and the minimum response frequency is 100 Hz. If the tension value at any point in the modulation requirement tension curve exceeds 10 Newtons, the curve is limited, and the excess portion is set to 10 Newtons. If the frequency component of the modulation requirement tension curve exceeds 100 Hz, the curve is low-pass filtered with a cutoff frequency of 100 Hz to generate an executable tension control curve. The actions of the adjusted pressing parameters, the reachable guide path, and the executable tension control curve at the same point in time are checked for spatial interference or dynamic contradictions. Spatial interference refers to the moving parts of different actuators occupying the same spatial position at the same time. Dynamic contradictions refer to the same actuator being required to achieve mutually exclusive motion states at the same time. If contradictions exist, time lag is introduced or the execution sequence is adjusted to eliminate the contradictions. Time lag is achieved by fine-tuning the start time of the action command, and the execution sequence is adjusted by rearranging the timing of the actions of different mechanisms. The final output is a set of coordination instructions that are conflict-free in terms of time, space, and capability, and can best satisfy the virtual process's adjustment intentions.

[0091] In some embodiments, the instruction synthesis layer generates a pressing instruction package for the pressing module, containing stroke nodes and time nodes, based on the coordination instruction elements. The pressing instruction package is a list, where each element contains a stroke position value and an absolute timestamp indicating arrival at that position. The instruction synthesis layer generates a guidance instruction package for the guidance module, containing a position sequence and a velocity curve. The guidance instruction package contains an array of three-dimensional coordinate points and a corresponding desired velocity array. The instruction synthesis layer generates a tension modulation instruction package for the tension application module, containing the time-varying pattern of tension values. The tension modulation instruction package is a set of (time, tension) data pairs. The pressing instruction package, guidance instruction package, and tension modulation instruction package are synchronized, aligned, and encapsulated according to a unified process timeline. The process timeline uses the start time of the folding process as zero. The alignment operation ensures that the instruction logic for the same timestamp in different instruction packages is consistent. The encapsulation format uses a binary data structure, ultimately forming a sequence of action instructions that drives the entire folding station to work collaboratively.

[0092] It is understandable that when performing smooth interpolation on the position sequence of the guidance module, an interpolation algorithm based on B-spline curves is used. The goal of this algorithm is to achieve smooth interpolation given a set of initial path points. In this case, a smooth trajectory is generated. Trajectory Depend on Control points and B-spline basis functions definition:

[0093]

[0094] Where: parameters In node vectors Changes within the defined interval Control points By solving a problem with an initial path point The least squares fitting problem with constraints is obtained, thus ensuring the generated trajectory. Smoothly and precisely traverse or approximate the original path point within the allowed boundaries. .

[0095] See Figure 4 This diagram visualizes the core data of the behavior mapping and control module in the sanitary napkin wing folding system, corresponding to the feasibility verification and adjustment of the pressing module's stroke by the instruction synthesis layer. The horizontal axis represents time, and the vertical axis represents the pressing module's stroke, containing three curves: the dashed line represents the initial stroke, the initial action requirement generated by the feature interpretation layer based on the material interference depth; the solid line represents the adjusted stroke, the executable stroke corrected by the instruction synthesis layer according to the equipment's physical limits; and the dashed line represents the equipment's maximum stroke threshold. This diagram intuitively presents the system's transformation process from virtual interference features to actual executable instructions, demonstrating the behavior mapping network's ability to coordinate virtual requirements with equipment capabilities. It is a visual representation of a crucial link in ensuring the safety and compliance of the wing folding action.

[0096] Example 4: After generating the action command sequence, the system performs closed-loop verification processing. While the mechanical actuator operates the actual wing according to the action command sequence, the system acquires the actual wing's response deformation flow through a sensing and acquisition module. Actual interference features are extracted from the actual wing's response deformation flow, including the actual material interference amount, the actual folding line path, and the actual stress response point. In a preset feature comparison space, the actual interference features are compared item by item with the folding interference features extracted from the virtual process, and the feature fit and deviation vector are calculated.

[0097] An interference comparison diagram is established with material interference as the horizontal axis and spatial position as the vertical axis. The predicted material interference depth curve and the actual material interference curve are plotted on the diagram, and the area enclosed by the two curves is calculated as the interference deviation. On the spatial position-time plane, the predicted folding line offset trajectory and the actual folding line path are plotted, and the average Euclidean distance between the two trajectories is calculated as the path following deviation. On the stress peak distribution map, the predicted stress peak points are paired with the actual stress response points, and the weighted sum of the stress value difference between the paired points and the spatial distance is calculated as the stress response deviation. The interference deviation, path following deviation, and stress response deviation are normalized and weighted to obtain a comprehensive feature fit score. The deviation vector is a multi-dimensional vector, where each dimension corresponds to the magnitude and direction of a local deviation. If the feature fit score is lower than a preset standard, the mapping weight parameters in the behavior mapping network are adjusted in reverse according to the deviation vector. Using the adjusted behavior mapping network, the folding interference features of subsequent wings in the same batch are remapped to generate a corrected action command sequence.

[0098] In practical implementation, the system performs closed-loop verification after generating the action command sequence. The action command sequence is generated by the behavior mapping and control module and sent to the mechanical execution module. While the mechanical execution mechanism performs the folding operation on the actual protective wing on the production line according to the action command sequence, the sensing and acquisition module collects the response deformation flow of the actual protective wing through multiple synchronous sensing channels. The format of the response deformation flow is consistent with the real-time deformation observation flow, including angle offset waveform, tension pulsation spectrum, and surface undulation point cloud.

[0099] In some embodiments, actual interference features are extracted from the actual wing's response deformation flow. The extraction process is logically the same as that used in the virtual process for extracting folding interference features. The actual interference features include the actual material interference amount, the actual folding line path, and the actual stress response points. The actual material interference amount is calculated by analyzing local deformation anomaly regions caused by mechanical interference in the response deformation flow. The actual folding line path is obtained by tracing the ridge lines of the surface undulation point cloud in the response deformation flow. The actual stress response points are obtained by mapping the peak position of the tension pulsation spectrum to the corresponding spatial coordinates of the surface undulation point cloud. In a preset feature comparison space, the actual interference features are compared item by item with the folding interference features extracted from the virtual process, and the feature fit score and deviation vector are calculated.

[0100] It is understandable that an interference comparison diagram is established with material interference quantity as the horizontal axis and spatial position as the vertical axis for comparison. The predicted material interference depth curve and the actual material interference quantity curve are plotted on the interference comparison diagram. The horizontal axis of both curves represents the standardized position along the length of the wing, and the vertical axis represents the normalized interference depth value. The area enclosed by the predicted material interference depth curve and the actual material interference quantity curve is calculated as the interference quantity deviation. A comparison is then performed on the spatial position-time plane, where the horizontal axis represents time and the vertical axis represents spatial position. The predicted folding line offset trajectory and the actual folding line path are plotted on the plane, and the average Euclidean distance between the corresponding time points of the two trajectories is calculated as the path following deviation. Finally, a comparison is performed on the stress peak distribution diagram. The predicted stress peak points and actual stress response points are paired according to time and spatial proximity. The weighted sum of the stress value difference and spatial distance between each pair of paired points is calculated, and the sum of all paired points is used as the stress response deviation.

[0101] Optionally, the interferometric deviation, path following deviation, and stress response deviation are normalized and weighted to obtain a comprehensive feature fit score. The normalization process uses a min-max normalization method, mapping each deviation to the 0-1 range. The weighting coefficients for the weighted fusion are pre-set based on the influence of each deviation on the folding quality. The deviation vector is a three-dimensional vector, where each dimension corresponds to the magnitude and direction of a local deviation. For example, the first component of the deviation vector records the value of the interferometric deviation and the offset direction of the actual curve relative to the predicted curve. Refer to Table 1, which shows the calculation of each deviation during a single alignment process.

[0102] Table 1: Comparison Table of Feature Deviation

[0103] Comparison Projects Source of predicted values Source of actual value Brief description of the calculation method Calculation results Interference deviation Virtual process material interference depth curve Actual response material interference curve Calculate the area enclosed between the two curves 0.15 (dimensionless) Path following deviation Virtual process folding line offset trajectory Actual response folding path Calculate the average Euclidean distance between trajectories 0.8 mm Stress response deviation Virtual process stress peak distribution diagram Actual response stress response point Calculate the stress difference between paired points and the distance-weighted sum. 2.3 (Stress-Distance Units)

[0104] In some embodiments, if the feature matching score is lower than a preset standard (set to 0.85), the mapping weight parameters in the behavior mapping network are adjusted in reverse according to the deviation vector. The adjustment process is based on the gradient descent principle; the deviation vector indicates the error direction between the network output and the actual expectation. The adjustment amount of the internal connection weights between the feature interpretation layer and the instruction synthesis layer of the behavior mapping network is calculated based on the error direction. Using the adjusted behavior mapping network, the folding interference features of subsequent wings in the same batch are remapped to generate a corrected action command sequence. This corrected action command sequence is used to drive the mechanical execution module to fold the next wing, thereby achieving closed-loop optimization of folding control based on actual response.

[0105] It is understandable that the feature fit score Calculated using the following formula:

[0106]

[0107] in: This represents the deviation of the normalized interference quantity. This represents the normalized path following deviation. This represents the normalized stress response deviation. , , These are the weighting coefficients assigned to each deviation, and they satisfy... When rating When the value falls below the threshold of 0.85, the network weight adjustment process is triggered.

[0108] Example 5: Before initial operation, the system initializes and trains the behavior mapping network. A training sample library is constructed by collecting morphological data sets of sanitary napkin wings of different models from historical production processes, along with their corresponding successful folding mechanical parameter sets. The morphological data sets are input into the cross-scale spatiotemporal fusion and virtual folding process simulation processing flow to obtain corresponding predicted folding interference features. The behavior mapping network is then trained under supervision using the predicted folding interference features as input and the corresponding successful folding mechanical parameter sets as the desired output. During training, the connection weights of the nodes within the behavior mapping network are adjusted using a gradient descent algorithm to minimize the difference between the network's output action command sequence and the successful folding mechanical parameter sets. When this difference falls below a preset convergence threshold, training stops, and the network weight parameters at this point are fixed as the initial parameters of the behavior mapping network.

[0109] In practical implementation, the system initializes and trains the behavior mapping network during the initial deployment phase. The behavior mapping network is the core component of the behavior mapping and control module. A training sample library is constructed by collecting morphological data sets of sanitary napkin wings of different models from historical production processes, along with their corresponding successful folding mechanical parameter sets. The morphological data set originates from the angle offset waveforms, tension pulsation spectra, and surface undulation point clouds recorded in the sensing and acquisition module's history. The successful folding mechanical parameter set comes from the pressing instruction packages, guidance instruction packages, and tension modulation instruction packages recorded in the mechanical execution module's history and manually verified as high-quality folding results.

[0110] In some embodiments, the morphological data set is input into a cross-scale spatiotemporal fusion and virtual folding process simulation processing flow. The cross-scale spatiotemporal fusion processing is completed by the behavior modeling module, and the virtual folding process simulation is completed by the virtual simulation module. For each set of morphological data in the training sample library, a corresponding dynamic behavior map is constructed through the cross-scale spatiotemporal fusion processing of the behavior modeling module. Then, the virtual simulation module drives the virtual folding process based on the dynamic behavior map and extracts folding interference features, thereby obtaining the predicted folding interference features corresponding to each set of historical morphological data. The predicted folding interference features are used as the input to the behavior mapping network, and the corresponding set of successful folding mechanical parameters is used as the expected output. The behavior mapping network is trained in a supervised manner, with the goal of making the action command sequence output by the behavior mapping network infinitely close to the set of successful mechanical parameters in history.

[0111] It is understandable that during training, the gradient descent algorithm is used to adjust the connection weights of nodes within the behavior mapping network. The behavior mapping network contains a feature interpretation layer and an instruction synthesis layer, both of which have adjustable connection weight parameters. The gradient descent algorithm calculates the difference between the network's actual output sequence of action instructions and the set of parameters for successful folding. This difference is quantified by a loss function, and the loss function's output value is distributed to the network's weight parameters through backpropagation. This process then adjusts the weight parameters along the gradient's inverse direction to reduce the difference. Each complete training iteration includes three steps: forward propagation to calculate the output, loss calculation, and backpropagation to update the weights. Repeating these training iterations minimizes the difference between the network's output sequence of action instructions and the set of parameters for successful folding.

[0112] Optionally, the training process continues, and the loss function value is recorded and monitored after each training iteration. When the loss function value falls below a preset convergence threshold, the training iteration stops. This threshold is set based on the accuracy requirements of the instruction sequence in actual production. At this point, all connection weights in the feature interpretation layer and instruction synthesis layer of the behavior mapping network are fixed as the initial parameters of the behavior mapping network. This fixed-weight network is deployed in the behavior mapping and control module for online mapping tasks. The loss function during training... Defined as the mean square error between the network output and the expected output, its relationship is as follows:

[0113]

[0114] in: This represents the value of the loss function. This indicates the number of samples in a training batch. This represents the total number of data points contained in each action instruction sequence. The behavior mapping network is represented as the first... The instruction sequence generated by the nth sample Predicted values ​​for each data point Indicates the corresponding number The successful folding mechanical parameter set of the sample is the first one. The true value of each data point. The network weights are continuously adjusted using the gradient descent algorithm to minimize... The value of .

[0115] See Figure 5 This graph corresponds to the initial training phase of the behavior mapping network and is a core verification graph for the effectiveness of network training. The horizontal axis represents the number of training iterations, the vertical axis represents the loss function value, the black curve represents the difference between the network output and the successfully folded mechanical parameters during training, and the dashed line represents the preset training convergence threshold. This graph visually demonstrates the convergence of the training process and is a key criterion for determining the completion of network initialization. Only when the loss is below the threshold can the network be fixed with the initial parameters, ensuring that during subsequent online mapping, virtual interference features can be accurately converted into actual action commands, providing a reliable algorithmic foundation for the adaptive control of the wing folding system.

[0116] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wing folding system based on sanitary napkin products, characterized by, The method comprises the following modules: a perception acquisition module, configured to capture a real-time deformation observation stream of a sanitary napkin wing in motion through a plurality of synchronous sensing channels, the real-time deformation observation stream comprising an angle offset waveform, a tension pulsation spectrum, and a surface fluctuation point cloud; a behavior modeling module, configured to perform cross-scale spatiotemporal fusion on the real-time deformation observation stream to construct a dynamic behavior atlas of the wing material, the dynamic behavior atlas recording response modes and structural evolution paths of the material under different stress phases; a virtual simulation module, configured to drive a virtual folding process according to the dynamic behavior atlas, simulate a complete deformation sequence of the wing from an unfolded state to a folded state in the virtual process, and record key interference events and stress concentration points, and further extract folding interference characteristics including material interference depth, folding line offset trajectory, and stress peak distribution map; a behavior mapping and control module, configured to input the folding interference characteristics into a behavior mapping network, and map the interference characteristics in the virtual process to a sequence of action instructions of an actual mechanical execution mechanism; a mechanical execution module, configured to receive the sequence of action instructions, and coordinate a plurality of execution units to execute the action instructions to interfere with and correct the wing; the cross-scale spatiotemporal fusion on the real-time deformation observation stream to construct the dynamic behavior atlas of the wing material comprises: decomposing a low-frequency trend component and a high-frequency jitter component from the angle offset waveform, the low-frequency trend component reflecting a bending intention of the wing body, and the high-frequency jitter component reflecting unstable vibration of the material locally; performing cross-correlation analysis on the tension pulsation spectrum and the high-frequency jitter component to identify a tension source causing the material vibration and an excitation frequency thereof; gridizing the surface fluctuation point cloud, and associating the low-frequency trend component and the excitation frequency of the corresponding position with each grid cell to form a grid attribute cell; predicting a deformation tendency of each grid attribute cell in a next time slice according to a historical evolution of the grid attribute cell, the deformation tendency including continuous bending, rebound, or arching; collecting the predicted deformation tendencies of all grid attribute cells, and clustering them according to spatial adjacency to generate a plurality of material behavior regions with similar behavior modes, each material behavior region and the corresponding evolution path thereof jointly constituting the dynamic behavior atlas; the driving of the virtual folding process according to the dynamic behavior atlas comprises: assigning a virtual agent to each material behavior region, each virtual agent performing a predetermined folding action in a virtual space according to the evolution path of the region to which the virtual agent belongs; establishing a constitutive relation model of the wing material in the virtual space, the constitutive relation model defining a stress-strain relationship of the material under a combined state of tension, bending, and compression; when folding actions of different virtual agents spatially overlap in the virtual space, triggering interference detection, the interference detection calculating a material stress distribution of the overlapping region based on the constitutive relation model; if the calculated stress of the overlapping region exceeds a material yield threshold, determining that a key interference event occurs, and recording a type, a location, and a stress excess amount of the key interference event. The virtual folding process is continuously run until a preset folding end posture is reached, and all key interference events and the positions and times of global stress peak values in the entire process are summarized to form a record of the key interference events and stress concentration points.

2. The sanitary napkin product based wing folding system of claim 1, wherein, The folding interference characteristics extracted include material interference depth, folding line offset trajectory, and stress peak value distribution diagram. For each key interference event, the event position is taken as the center to expand outward in the virtual space, and the deformation gradient field of the material in the interference direction is calculated. The maximum depth from the interference surface to the yield point inside the material is obtained by integrating along the direction of the deformation gradient field, serving as the material interference depth of the key interference event. The center points of all key interference events are connected in chronological order to form a broken line describing the migration of the interference position, serving as the folding line offset trajectory. The stress field is sampled at fixed time intervals on the global timeline of the virtual folding process, and the stress peak values at each sampling time are marked on the corresponding spatial positions of the wing, finally generating a stress peak value distribution diagram marked with stress peak value size and time. The material interference depth, folding line offset trajectory, and stress peak value distribution diagram together constitute an interference characteristic set for guiding actual action adjustment.

3. The panty liner product based wing folding system of claim 1, wherein, The folding interference characteristics are input into the behavior mapping network to map the interference characteristics in the virtual process to the action instruction sequence of the actual mechanical execution mechanism, including: The behavior mapping network includes a feature interpretation layer and an instruction synthesis layer, and the feature interpretation layer receives the folding interference characteristics. The feature interpretation layer converts the material interference depth into initial requirements for the pressing module stroke and holding time, converts the folding line offset trajectory into tracking requirements for the guide module position sequence, and converts the stress peak value distribution diagram into modulation requirements for the tension applying module output curve. The instruction synthesis layer performs instruction feasibility checking and conflict resolution according to the initial requirements, tracking requirements, and modulation requirements, combined with the physical limits and kinematic constraints of each mechanical execution mechanism. After conflict resolution, the instruction synthesis layer generates a pressing instruction package containing stroke nodes and time nodes for the pressing module, a guide instruction package containing a position sequence and a speed curve for the guide module, and a tension modulation instruction package containing a tension value change rule over time for the tension applying module. The pressing instruction package, guide instruction package, and tension modulation instruction package are synchronized and aligned according to the process time axis to form an action instruction sequence that drives the entire folding workstation to work cooperatively.

4. The panty liner product based wing folding system of claim 3, wherein, The instruction synthesis layer performs instruction feasibility checking and conflict resolution according to the initial requirements, tracking requirements, and modulation requirements, combined with the physical limits and kinematic constraints of each mechanical execution mechanism, including: Check whether the initial requirements for the stroke of the pressing module exceed its maximum stroke range, and if so, scale the stroke and holding time to generate adjusted pressing parameters that meet the physical limits. Check whether the position sequence tracking requirements of the guide module exceed its motion space boundary, and if so, perform smooth interpolation and boundary clipping on the position sequence to generate an achievable guide path. The tension modulation requirement of the tension applying module is checked to see if it exceeds its maximum load or minimum response frequency, and if it does, the tension change curve is limited and filtered to generate an executable tension control curve; The actions of the adjusted compression parameters, the reachable guide path and the executable tension control curve at the same time point are checked to see if there is spatial interference or dynamic contradiction, and if there is, time lag or execution order is adjusted to eliminate the contradiction; Finally, a set of coordinated instruction elements that are free of conflicts in time, space and ability and can best meet the adjustment intention of the virtual process are output.

5. The sanitary napkin product based wing folding system of claim 2, wherein, The system also includes a closed-loop verification process performed after the generation of the action instruction sequence: While the actual wing is being operated by the mechanical execution mechanism according to the action instruction sequence, the response deformation flow of the actual wing is collected; The actual interference features are extracted from the response deformation flow of the actual wing, including the actual material interference amount, the actual folding line path and the actual stress response point; In the feature comparison space, the actual interference features are compared with the folding interference features extracted from the virtual process one by one, and the feature fitting degree and deviation vector are calculated; If the feature fitting degree is lower than the preset standard, the mapping weight parameters in the behavior mapping network are adjusted in the opposite direction according to the deviation vector; Using the adjusted behavior mapping network, the folding interference features of the subsequent wings in the same batch are remapped to generate a corrected action instruction sequence, thereby realizing closed-loop optimization of folding control based on actual response.

6. The sanitary napkin product based wing folding system of claim 5, wherein, The comparison of the actual interference features with the folding interference features extracted from the virtual process in the feature comparison space includes: An interference comparison graph is established with the material interference amount as the horizontal axis and the spatial position as the vertical axis, and the predicted material interference depth curve and the actual material interference amount curve are plotted in the graph, and the area of the region surrounded by the two curves is calculated as the interference amount deviation degree; In the spatial position-time plane, the predicted folding line offset trajectory and the actual folding line path are plotted respectively, and the average Euclidean distance between the two trajectories is calculated as the path following deviation degree; In the stress peak distribution graph, the predicted stress peak points and the actual stress response points are paired, and the weighted sum of the stress value difference and the spatial distance between the paired points is calculated as the stress response deviation degree; The interference amount deviation degree, the path following deviation degree and the stress response deviation degree are normalized and weighted to obtain a comprehensive feature fitting degree score; The deviation vector is a multi-dimensional vector, and each dimension component corresponds to the amplitude and direction of a local deviation.

7. The sanitary napkin product based wing folding system of claim 1, wherein, The system also includes a process for initializing and training the behavior mapping network: A set of morphological data of different types of sanitary napkin wings in the historical production process and their corresponding successful folding mechanical parameter sets are collected to form a training sample library; The morphological data set is input into the cross-scale spatio-temporal fusion and virtual folding process simulation process to obtain the predicted folding interference features; The predicted folding interference features are used as input, and the corresponding successful folding mechanical parameter set is used as expected output to supervise the training of the behavior mapping network. During the training process, the connection weights of the internal nodes of the behavior mapping network are adjusted by a gradient descent algorithm, so that the difference between the action instruction sequence output by the network and the successful folding mechanical parameter set is minimized. When the difference is lower than the convergence threshold, the training is stopped, and the network weight parameters at this time are fixed as the initial parameters of the behavior mapping network.

Images