A corrugated box design optimization system and method based on a machine learning model

Abstract

The application relates to the technical field of machine learning, and particularly discloses a corrugated box design optimization system and method based on a machine learning model, which comprises the following steps: constructing a kinematics simulation model of a to-be-packaged article in a virtual empty box cavity; obtaining a simulation calibration coefficient and calibrating the kinematics simulation model; running the calibrated kinematics simulation model to generate a contact event sequence; analyzing the contact event sequence, extracting statistical characteristics of contact point spatial distribution on each inner surface of the virtual empty box cavity and corresponding impact force vectors, and generating an impact hotspot distribution map; mapping the impact hotspot distribution map to a three-dimensional grid model of the virtual empty box cavity, triggering an adaptive structure reinforcement generation algorithm, and generating local enhancement three-dimensional data at corresponding positions of the three-dimensional grid model; and outputting a corrugated box production file fused with the local enhancement three-dimensional data. The application improves the precision, consistency and engineering application value of corrugated box structure design.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and specifically to a system and method for optimizing the design of corrugated cardboard boxes based on machine learning models. Background Technology

[0002] In logistics and transportation, corrugated cardboard boxes are the most commonly used form of outer packaging. Their structural design usually relies on empirical rules or simple size matching, making it difficult to fully consider the actual force and movement behavior of different items under complex transportation scenarios. In actual transportation, items will experience uncertain displacement and collisions inside the cardboard box under conditions such as vibration, drop, and tilting. The contact position, impact frequency, and impact intensity have obvious randomness and spatial non-uniformity.

[0003] However, existing corrugated box design methods generally lack systematic analysis tools for the aforementioned dynamic contact behaviors, making it difficult to accurately identify the critical areas on the inner wall of the box that truly bear high impact loads. This leads to blind selection of reinforcement locations for the box structure, easily resulting in problems such as local over-design or insufficient protection of critical parts. At the same time, different items have significantly different physical properties and transportation scenarios, making it difficult for traditional design methods to achieve adaptive optimization for specific items and transportation conditions, resulting in unstable packaging protection effects and low material utilization efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a corrugated box design optimization system and method based on a machine learning model, thereby solving the above-mentioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A corrugated cardboard box design optimization system and method based on a machine learning model includes the following steps:

[0007] A kinematic simulation model of the items to be packaged within a virtual empty box cavity is constructed based on the physical parameters of the items to be packaged.

[0008] The pre-trained simulation calibration model processes the preset transportation scenario parameters and the physical parameters to obtain a set of simulation calibration coefficients, and then calibrates the kinematic simulation model based on the simulation calibration coefficients.

[0009] The calibrated kinematic simulation model is run under preset transportation scenario parameters. The positions, times, and impact force vectors of the items to be packaged coming into contact with the inner surfaces of the virtual empty box cavity during the simulation are recorded to generate a contact event sequence.

[0010] Analyze the contact event sequence, extract the statistical characteristics of the spatial distribution of contact points and corresponding impact force vectors on each inner surface of the virtual empty box cavity, and generate an impact hot spot distribution map;

[0011] The impact hotspot distribution map is mapped onto the 3D mesh model of the virtual empty box cavity, triggering an adaptive structural reinforcement generation algorithm to generate local enhanced 3D data at the corresponding positions in the 3D mesh model;

[0012] Output corrugated cardboard box production files that integrate locally enhanced 3D data.

[0013] Preferably, the process of constructing the kinematic simulation model is as follows:

[0014] The physical parameters include the external dimensions, mass, center of gravity, and surface friction coefficient of the item to be packaged. The length, width, and height values ​​of the external dimensions are obtained, and a predetermined buffer margin is added to each of the length, width, and height values ​​to generate the internal length, internal width, and internal height dimensions of the virtual empty box cavity.

[0015] The item to be packaged is simplified into a rigid body model. The circumscribed cuboid size of the rigid body model is equal to its external size. The mass attribute value of the rigid body model is equal to its mass, and the position of the center of mass of the rigid body model coincides with the position of the center of gravity.

[0016] Define the contact mechanics relationship between the rigid body model and the inner surface of the virtual empty box cavity. The contact mechanics relationship includes the normal contact force model and the tangential friction force model. The normal contact force model is defined based on the linear spring damping model, and the tangential friction force model is defined based on Coulomb's friction law. The static friction coefficient and the dynamic friction coefficient are equal to the surface friction coefficient. Define the six inner surfaces of the virtual empty box cavity as infinitely large rigid plane constraints. The defined rigid body model, contact mechanics relationship and rigid plane constraints together constitute the kinematic simulation model.

[0017] Preferably, the pre-training process of the simulation calibration model is as follows:

[0018] Input features for training samples are constructed based on the physical parameters of historical items and the corresponding transportation scenario parameters.

[0019] The sequence of real contact events is obtained from the physical transport test records of historical items; the sequence of simulated contact events is obtained from the initial kinematic simulation model with default parameters.

[0020] A set of simulation calibration coefficients is derived from the real contact event sequence and the simulated contact event sequence. These simulation calibration coefficients are used to calibrate the contact mechanical relationship.

[0021] All training samples form a training set. A neural network model is trained using the training set. The trained neural network model is the pre-trained simulation calibration model.

[0022] Preferably, the process of generating the impact hotspot distribution map is as follows:

[0023] The six inner surfaces of the virtual empty box cavity are unfolded into six two-dimensional analysis planes, and each contact event in the contact event sequence is assigned to the corresponding two-dimensional analysis plane.

[0024] On a single two-dimensional analysis plane, the contact locations of all contact events assigned thereto form a point set. The impact intensity corresponding to each contact location in the point set is calculated. Based on the impact intensity value of each contact location, an initial influence radius is calculated.

[0025] Using a two-dimensional Gaussian kernel function, a local impact density distribution field is generated at each contact point, with its initial influence radius as the smoothing range. The impact hotspot region is obtained based on the entire local impact density distribution field.

[0026] The boundary polygon coordinates, average impact density, and typical impact force direction of each impact hotspot region constitute an impact hotspot descriptor. The set of all impact hotspot descriptors on an inner surface constitutes the impact hotspot distribution map of that inner surface.

[0027] Preferably, the execution process of the adaptive structure reinforcement generation algorithm is as follows:

[0028] Input a 3D mesh model of the virtual empty box cavity and the impact hotspot distribution map of each inner surface; for an inner surface and its impact hotspot distribution map, traverse each impact hotspot descriptor in the distribution map; based on the boundary polygon coordinates in the impact hotspot descriptor, determine all mesh patches covered by the polygon on the inner surface mesh corresponding to the 3D mesh model; these covered mesh patches together constitute a region to be strengthened.

[0029] Based on the average impact density value and typical impact force direction in the impact hotspot descriptor, query the reinforced structure template library; the reinforced structure template library stores a variety of preset reinforced structure 3D templates and their selection conditions; the selection conditions include requirements for the range of average impact density values ​​and the range of the angle between the typical impact force direction and the inner surface normal; select a reinforced structure 3D template from the library that meets all selection conditions.

[0030] Based on the area covered by the boundary polygon coordinates in the impact hotspot descriptor, the size of the selected reinforcement structure 3D template is scaled proportionally; based on the typical impact force direction in the impact hotspot descriptor, the orientation of the selected reinforcement structure 3D template is rotated; the geometry of the scaled and rotated reinforcement structure 3D template is Boolean-merged with the original mesh portion containing all mesh patches constituting the region to be reinforced, so that the reinforcement structure geometry is merged into part of the original mesh; the mesh of the region to be reinforced and its surrounding merging region is smoothed and re-divided to eliminate sharp edges; all impact hotspot descriptors on all inner surfaces are processed iteratively, and finally all local reinforcement structures are generated on the 3D mesh model to obtain the merged 3D mesh model; the local reinforcement 3D data refers to the mesh geometry data representing the reinforcement structure in the merged 3D mesh model.

[0031] Preferably, the process of obtaining simulation calibration coefficients based on the pre-trained simulation calibration model is as follows:

[0032] Based on the preset transportation scenario parameters and the physical parameters, a set of input features is constructed. The input features are then input into the pre-trained simulation calibration model, which outputs a set of simulation calibration coefficients A. The simulation calibration coefficients A calibrate the contact mechanics relationship in the kinematic simulation model.

[0033] A corrugated cardboard box design optimization system based on a machine learning model includes:

[0034] Simulation module: Constructs a kinematic simulation model of the item to be packaged within a virtual empty box cavity based on the physical parameters of the item to be packaged;

[0035] Model calibration module: Based on the pre-trained simulation calibration model, the preset transportation scenario parameters and the physical parameters are processed to obtain a set of simulation calibration coefficients, and the kinematic simulation model is calibrated based on the simulation calibration coefficients;

[0036] Impact recording module: The calibrated kinematic simulation model is run under preset transportation scenario parameters to record the position, time and impact force vector of the item to be packaged contacting the inner surface of the virtual empty box cavity during the simulation process, and generate a contact event sequence.

[0037] Analyze the contact event sequence, extract the statistical characteristics of the spatial distribution of contact points and corresponding impact force vectors on each inner surface of the virtual empty box cavity, and generate an impact hot spot distribution map;

[0038] Optimization module: Maps the impact hotspot distribution map onto the 3D mesh model of the virtual empty box cavity, triggers the adaptive structural reinforcement generation algorithm, and generates local enhanced 3D data at the corresponding positions in the 3D mesh model;

[0039] Output corrugated cardboard box production files that integrate locally enhanced 3D data.

[0040] The beneficial effects of this invention compared to the prior art are as follows:

[0041] This invention enables more targeted optimization of the structural design of corrugated cardboard boxes for different items to be packaged and transportation conditions, thereby improving the rationality and consistency of packaging protection. Through a systematic analysis of the contact behavior between the items and the inner wall of the cardboard box during transportation, this invention clearly depicts the differences in stress distribution in different areas inside the box, facilitating accurate identification of areas with concentrated impact risks and avoiding insufficient protection or structural redundancy caused by relying on experience-based judgments. Optimizing the cardboard box structure based on the above analysis results allows the structural reinforcement locations to match the actual impact characteristics, improving the overall stability and reliability of the cardboard box under complex transportation conditions. Simultaneously, this invention can adapt to changes in the physical characteristics of different items and transportation scenarios, reducing the reliance on single empirical models in the design scheme and improving the universality and repeatability of the design results. Furthermore, by directly reflecting the optimization results into design data that can be used for production, it helps to shorten the conversion process between design and manufacturing, improving the efficiency and feasibility of corrugated cardboard box design optimization, thereby ensuring the effectiveness of packaging protection while improving the rationality of material utilization. Attached Figure Description

[0042] The invention will now be further described with reference to the accompanying drawings.

[0043] Figure 1 This is a flowchart illustrating a corrugated box design optimization method based on a machine learning model according to the present invention.

[0044] Figure 2 This is a schematic diagram of the process of executing the adaptive structure reinforcement generation algorithm of the present invention. Detailed Implementation

[0045] 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.

[0046] Please see Figures 1-2 As shown, this invention is a method for optimizing the design of corrugated cardboard boxes based on a machine learning model, comprising the following steps:

[0047] A kinematic simulation model of the item to be packaged within a virtual empty box cavity is constructed based on the physical parameters of the item.

[0048] In a preferred embodiment of the present invention, the process of constructing a kinematic simulation model is as follows:

[0049] The input consists of the physical parameters of the item to be packaged. These parameters include at least the external dimensions, mass, center of gravity, and surface friction coefficient. The external dimensions are given as length, width, and height values, while the center of gravity is given as its three-dimensional position in the item's coordinate system. During processing, the length, width, and height values ​​are read, and a predetermined buffer margin is added to each. This buffer margin, pre-set by the packaging design, is used to reserve space within the virtual empty box cavity for posture changes and contact compression. The output is the internal length, width, and height dimensions of the virtual empty box cavity. The virtual empty box cavity is described using a Cartesian coordinate system, with six internal surfaces corresponding to the top, bottom, left, right, front, and rear surfaces. Each internal surface is assigned a unique contact surface identifier for subsequent contact event recording.

[0050] In the simulation, the item to be packaged is simplified into a rigid body model. The geometry of the rigid body model is represented by a circumscribed cuboid, where the three sides are equal to the length, width, and height of the outer dimensions, respectively. The mass attribute of the rigid body model is equal to its mass. The center of mass of the rigid body model coincides with its center of gravity. The pose of the rigid body model is characterized by both position and attitude and is updated over time. A contact mechanics relationship is defined between the rigid body model and the inner surface of the virtual empty cavity. This contact mechanics relationship includes a normal contact force model and a tangential friction force model. The normal contact force model is based on a linear spring-damped model. When the circumscribed cuboid geometrically penetrates a certain inner surface, an elastic restoring component is formed based on the penetration depth, and a damping component is formed based on the normal relative velocity. These components are combined to form a normal contact force vector pointing towards the inner surface. The tangential friction model is based on Coulomb's law of friction. The direction of the friction force lies within the tangential plane of the inner surface and is opposite to the relative slippage tendency. The static friction state is used for tangential constraint when the relative tangential velocity approaches zero, and the dynamic friction state is used for resistance constraint when slippage occurs. The values ​​of both the static and dynamic friction coefficients are equal to the surface friction coefficient. The six inner surfaces of the virtual empty box cavity are defined as infinitely large rigid plane constraints. Each plane is determined by its spatial position and normal. The plane position is determined by the internal length, internal width, and internal height dimensions and is fixed to the cavity coordinate system. During simulation, geometric contact detection is performed at each time step to determine the contact relationship between the circumscribed cuboid and each rigid plane. If the contact condition is met, the impact force vector for that time step is output from the contact mechanics relationship. At the same time, the contact surface identifier, contact position, and contact occurrence time are also output. The contact position is represented by the position of the contact point in the corresponding inner surface coordinate system. The rigid body model, contact mechanics relationship, and rigid plane constraints together constitute the kinematic simulation model.

[0051] Abstracting the items to be packaged as rigid bodies with defined geometric shapes, masses, and center of mass positions, and abstracting the inner walls of the virtual empty box cavity as fixed rigid planes, allows the complex transportation process to be transformed into a calculable and repeatable dynamic problem while maintaining the main motion and contact characteristics. The displacement, tumbling, and collisions of the items within the box during transportation are essentially determined by gravity, inertial forces, and the interaction forces generated when in contact with the inner walls. The direction, magnitude, and timing of these forces directly depend on the object's geometric dimensions, mass distribution, and relative motion with the contact surfaces. Using a circumscribed cuboid to describe the item's geometry allows contact determination to be stably obtained through geometric distance and penetration relationships, thus continuously tracking the contact state between the item and each inner surface. The normal contact force is described in the form of linear spring damping, which simultaneously characterizes the reaction tendency caused by geometric compression and the energy dissipation caused by relative velocity at the time of contact, making the collision process smoothly change over time and avoiding the numerical instability caused by ideal rigid collisions. The tangential direction is described using Coulomb's law of friction, allowing tangential constraints and sliding behavior during contact to be naturally triggered by the normal contact state, thus reflecting the tendency of the object to remain stationary, slide, or flip due to friction within the container. The inner wall of the container is treated as an infinitely large rigid plane, ensuring the contact direction and constraint conditions remain stable during the simulation and preventing interference from boundary dimension changes in contact determination. Under this abstraction, the motion state of the object can be continuously updated over time; the surface, location, and time of contact, as well as the corresponding impact force vector, can be clearly identified and recorded, resulting in a kinematic simulation result that reflects the actual contact behavior during transportation.

[0052] It should be noted that when constructing the kinematic simulation model of the item to be packaged within a virtual empty box cavity, this invention defines the six inner surfaces of the virtual empty box cavity as infinitely large rigid planar constraints. This is an engineering abstraction used to describe contact events between the item and the inner wall of the box. The purpose of this abstraction is to stably and repeatably characterize contact event elements such as contact surface markings, contact positions, occurrence times, and impact force vectors, and thereby obtain contact event sequences and their statistical characteristics. Simultaneously, this invention calibrates the default stiffness parameters, default damping parameters, default static friction coefficients, and default dynamic friction coefficients in the contact mechanics relationship through simulation calibration coefficients. This improves the consistency between the contact event sequences generated under preset transportation scenario parameters and the contact behavior reflected in actual transportation test records. Therefore, the aforementioned rigid planar constraints, as the modeling basis for contact event analysis, do not affect the effectiveness of subsequent impact hotspot distribution map generation and structural reinforcement location determination.

[0053] The pre-trained simulation calibration model processes the preset transportation scenario parameters and the physical parameters to obtain a set of simulation calibration coefficients, and then calibrates the kinematic simulation model based on these coefficients.

[0054] In a preferred embodiment of the present invention, the pre-training process of the simulation calibration model is as follows:

[0055] Obtain a physical transport test record of a historical item. The test record includes the historical physical parameters of the historical item, the historical transport scenario parameters, and the real pressure data stream collected by a sensor array deployed on the inner wall of the test carton.

[0056] Process the real pressure data stream, identify and aggregate the events in which the historical items came into contact with the inner wall of the test carton, and generate a real contact event sequence; use historical physical parameters and historical transportation scenario parameters to run an initial kinematic simulation model based on default parameters to obtain a simulated contact event sequence;

[0057] Event matching is performed between real contact event sequences and simulated contact event sequences. The matching conditions are that the contact surface identifiers of the two events are the same and the difference in the timestamps of the events is less than the preset time tolerance. Based on the successfully matched event pairs, a set of simulation calibration coefficients is derived as the output label of the training sample. The simulation calibration coefficients include Class I coefficients and Class II coefficients. Class I coefficients include stiffness coefficients and damping coefficients, and Class II coefficients include static friction coefficients and dynamic friction coefficients.

[0058] The reverse calculation process is as follows: set the stiffness coefficient, damping coefficient, static friction coefficient, and dynamic friction coefficient as variables to be optimized, and perform iterative calculations. In each iteration, the default stiffness parameter of the initial kinematic simulation model is multiplied by the stiffness coefficient of the current iteration to obtain the temporary stiffness parameter; the default damping parameter of the initial kinematic simulation model is multiplied by the damping coefficient of the current iteration to obtain the temporary damping parameter; the default static friction coefficient of the initial kinematic simulation model is multiplied by the static friction coefficient of the current iteration to obtain the temporary static friction coefficient; and the default dynamic friction coefficient of the initial kinematic simulation model is multiplied by the dynamic friction coefficient of the current iteration to obtain the temporary dynamic friction coefficient.

[0059] The contact mechanics relationships of the initial kinematic simulation model are updated using temporary stiffness parameters, temporary damping parameters, temporary static friction coefficients, and temporary dynamic friction coefficients. The updated initial kinematic simulation model is run to generate a new sequence of simulated contact events. The error scalar of each pair of matching events between the new simulated contact event sequence and the real contact event sequence is calculated. The error scalars are summed to obtain the error metric for this iteration.

[0060] The values ​​of stiffness coefficient, damping coefficient, static friction coefficient, and dynamic friction coefficient are adjusted according to the direction of change of the error metric, and the next iteration is performed. When the error metric meets the preset convergence condition, the iteration stops, and the set of stiffness coefficient, damping coefficient, static friction coefficient, and dynamic friction coefficient at this time is used as the simulation calibration coefficient obtained by back-reasoning.

[0061] Historical physical parameters and historical transportation scenario parameters are combined as input features for training samples; multiple training samples are collected to form a training sample set; a neural network model is trained using the training sample set, and the trained neural network model is the pre-trained simulation calibration model.

[0062] The calculation error scalar includes:

[0063] Calculate the absolute value of the relative error between the two impact force vectors in magnitude and the radian value of the angle between the two vector directions. Then, perform a weighted summation of the absolute value of the relative error and the radian value to obtain the error scalar.

[0064] It should also be noted that the simulation calibration coefficients obtained in reverse engineering in this invention are used to calibrate the default parameters of the initial kinematic simulation model. Essentially, they are equivalent calibration coefficients that make the simulated contact event sequence statistically closer to the real contact event sequence. Since different parameter combinations may produce similar contact event matching results under certain conditions, this invention uses the satisfaction of a preset convergence condition as the stopping criterion to obtain a set of simulation calibration coefficients that can make the sum of matching event errors converge. The goal of this process is to improve the fitting degree of the contact event sequence in dimensions such as contact surface identification, occurrence timestamp, and impact force vector, rather than requiring the coefficients to be physically unique or equivalent to actual material parameters.

[0065] In this invention, the pre-trained simulation calibration model takes as input the physical parameters and preset transportation scenario parameters, and outputs simulation calibration coefficients to calibrate the contact mechanics relationships in the kinematic simulation model. Specifically, the calibration coefficients are multiplied by the default stiffness parameters, default damping parameters, default static friction coefficient, and default kinetic friction coefficient of the initial kinematic simulation model to obtain temporary stiffness parameters, temporary damping parameters, temporary static friction coefficient, and temporary kinetic friction coefficient for updating the contact mechanics relationships. Through this multiplicative calibration method, the contact event sequence recorded by the calibrated kinematic simulation model under the preset transportation scenario parameters more closely matches the contact behavior characteristics reflected in actual transportation test records.

[0066] After gathering a training sample set from multiple historical artifacts and transportation scenarios, the neural network model is constructed with regression as the objective. The input receives a feature vector composed of historical physical parameters, historical transportation scenario parameters, and error feature vectors obtained from event matching. The output provides a simulation calibration coefficient vector, where each component corresponds to the stiffness coefficient, damping coefficient, static friction coefficient, and dynamic friction coefficient, respectively. This allows the output to be directly used to scale and correct the contact mechanics parameters of the initial kinematic simulation model. The model architecture employs a feedforward neural network, consisting of an input layer, a feature transformation layer, and an output layer connected sequentially to form a layered structure. The input layer encodes various input features with a unified dimension and performs numerical scale normalization. The normalization process is determined based on statistics from the training sample set during the training phase and continues during the inference phase using the same normalization rule to ensure consistent input distribution. The feature transformation layer consists of multiple stacked fully connected networks. Nonlinear mapping is established between layers through nonlinear activation functions, enabling the network to express the coupled influence of physical parameters and transportation scenario parameters on simulation deviations. The depth of the feature transformation layer and the number of neurons in each layer are determined based on the training sample size and the convergence of the fitting error, maintaining a balance between expressive power and overfitting risk. To enhance the structured utilization of input features from different categories, a feature branch structure can be set before the feature transformation layer. Physical parameter sub-vectors and transportation scenario parameter sub-vectors are input to their respective sub-network branches for preliminary feature extraction. In the subsequent fusion layer, the latent features output from each branch are concatenated and fused with the error feature vector. The fused features are then input to subsequent fully connected layers to complete the joint regression of calibration coefficients, thus forming an end-to-end mapping from input to output. The output layer uses a linear output form, directly providing continuous values ​​for each calibration coefficient. Boundary constraints are applied to the coefficient value range at the output end, or the sign and scale of the coefficients are restricted through activation functions, ensuring that the output results meet the physical requirements of the calibration coefficients as scaling factors. During training, the simulation calibration coefficients obtained from the parameter back-calculation process are used as supervision labels. A loss function based on the coefficient vector regression error is employed for optimization. Weights for different coefficient components can be introduced into the loss function to balance the different impacts of stiffness / damping coefficients and friction coefficients on simulation error. In training iterations, mini-batch samples are used to update the weight parameters, and a validation set is set for generalization error monitoring. When the validation error meets the stopping condition, the trained network weights are output. After training, the neural network model is solidified as a pre-trained simulation calibration model. In actual use, only new physical parameters and transportation scenario parameters need to be input to obtain the corresponding simulation calibration coefficients through forward inference, avoiding the need for iterative optimization in each iteration. This improves the usability and stability of subsequent simulation calibration processes while maintaining calibration consistency.

[0067] In a preferred embodiment, the process of obtaining simulation calibration coefficients based on a pre-trained simulation calibration model is as follows:

[0068] Based on preset transportation scenario parameters and physical parameters, a set of input features is constructed. The input features are then input into a pre-trained simulation calibration model, which outputs a set of simulation calibration coefficients A. The simulation calibration coefficients A calibrate the contact mechanics relationship in the kinematic simulation model.

[0069] It is worth noting that during transportation, the contact behavior between the packaged item and the inner wall of the container is not only related to the item's dimensions, mass distribution, and surface friction characteristics, but also influenced by factors such as vibration intensity, direction changes, acceleration fluctuations, and impact patterns in the transportation environment. Items with the same physical parameters will exhibit significant differences in contact frequency, contact position distribution, and impact force variation under different transportation conditions. These differences are mainly reflected in the kinematic simulation as the influence of contact mechanics parameter values ​​on the simulation results. When preset transportation scenario parameters and physical parameters are used as input features, these features jointly describe the force environment and motion constraints of the item during transportation, enabling the model to distinguish the overall pattern of contact behavior under different working conditions. The pre-trained simulation calibration model is derived from the correspondence between a large amount of historical transportation test data and simulation results. Internally, it has learned which adjustments to contact mechanics parameters under different combinations of physical and transportation conditions can make the simulated contact events more closely resemble reality in time, space, and impact direction. Therefore, after inputting features corresponding to the current item and transportation conditions, the model can output a set of calibration coefficients to correct the contact mechanics relationship, so that the contact behavior generated by the simulation model under these conditions conforms to the laws reflected by the existing data.

[0070] Transportation scenario parameters describe the external environmental conditions of the packaged goods during transportation, characterizing the impact of the external environment on the overall motion of the goods. These parameters abstractly depict typical characteristics of the transportation process, including at least the vibration characteristics corresponding to the transportation mode, the types of impacts that may occur during transportation, the range of acceleration variations, and the attitude change trend. Vibration characteristics describe the periodic or random excitation characteristics experienced by the container in different directions during transportation; impact types describe the occurrence of possible transient strong excitation events; the acceleration variation range defines the amplitude variation interval of the external excitation in the time dimension; and the attitude change trend describes the overall law of the change of the gravity direction relative to the container coordinate system over time. These transportation scenario parameters do not directly limit the specific numerical calculation process but rather provide a unified conditional description for the generation and action of external excitations in subsequent kinematic simulations, enabling different transportation scenarios to be distinguished and expressed within the same simulation modeling framework.

[0071] The simulation calibration coefficient A is used to calibrate the contact mechanics relationships in the kinematic simulation model. Its function is manifested in the unified correction of the contact mechanics parameter values. During the initial construction of the kinematic simulation model, both the normal contact force model and the tangential friction force model use a set of preset default parameters. These default parameters characterize the contact characteristics between an object and the inner wall of a container under normal circumstances, but they are difficult to directly adapt to the differences caused by different object characteristics and transportation scenarios. When the simulation calibration coefficient A is applied to the contact mechanics relationships, it specifically multiplies the default stiffness parameter, default damping parameter, default static friction coefficient, and default dynamic friction coefficient by their corresponding coefficients in simulation calibration coefficient A, respectively, to obtain temporary stiffness parameters, temporary damping parameters, temporary static friction coefficients, and temporary dynamic friction coefficients used for the current simulation run. Subsequently, these temporary parameters replace the original default parameters, updating the parameter values ​​of the normal contact force model and the tangential friction force model in the contact mechanics relationships. This ensures that, during subsequent time progression, the calculated magnitude of the normal reaction force, damping attenuation characteristics, and tangential friction constraint strength at the time of contact all respond according to the calibrated parameters. Through this multiplicative calibration method, the kinematic simulation model can adapt parameters to different physical conditions and transportation scenarios without changing its structural form and computational logic. This allows the contact event sequence generated during the simulation to exhibit behavioral characteristics consistent with the corresponding transportation conditions in terms of contact time, contact position, and impact force vector characteristics.

[0072] The calibrated kinematic simulation model is run under preset transportation scenario parameters. The positions, times, and impact force vectors of the items to be packaged coming into contact with the inner surfaces of the virtual empty box cavity during the simulation are recorded, and a sequence of contact events is generated.

[0073] Specifically, the calibrated kinematic simulation model is run under preset transportation scenario parameter constraints. These parameters are given as a combination of vibration characteristics, impact types, acceleration variation ranges, and attitude change trends. Based on these parameters, a corresponding external excitation time series is generated. This external excitation time series is used to apply an equivalent inertial force to the rigid body model of the packaged item during the simulation. This includes: given the total simulation duration and sampling time step, establishing a box coordinate system, and assigning a timestamp to each time step. The acceleration variation range is used to provide the upper and lower limits of the three-axis acceleration; the vibration characteristics are used to provide the dominant frequency band, amplitude distribution, and randomness of the three-axis vibration; the attitude change trend is used to provide the time-varying law of the gravity direction relative to the box coordinate system; and the impact type is used to provide the morphology and triggering rules of the impact event. In the specific generation process, firstly, basic vibration components are constructed for each of the three axes. These basic vibration components are generated using a multi-sine superposition method. Several frequency points are selected on each axis, sampled from the main frequency band of that axis. Each frequency point is assigned an amplitude and an initial phase. The amplitude is sampled from the amplitude distribution of that axis, and the initial phase is sampled from a uniform distribution. The sinusoidal signals corresponding to each frequency point are superimposed to form the periodic vibration sequence of that axis. Next, random vibration components are superimposed on each of the three axes. These random vibration components are implemented using band-limited random sequences. First, a zero-mean random sequence is generated, then filtered using a bandpass filter consistent with the main frequency band, and finally scaled according to the amplitude distribution. Subsequently, impact components are generated based on the impact type. These impact components are superimposed onto the corresponding axes according to an event-based approach. The timing of the impact event is determined by the triggering rules of the impact type. The triggering rules can be implemented using one of the following methods: fixed interval, random arrival, or guaranteed occurrence within a specified time window. The waveform of the impact event is a finite-duration half-sine or triangular pulse. The pulse duration is given by the impact type. The pulse peak value is sampled from the acceleration variation range and scaled in conjunction with the intensity level of the impact type. The pulse direction is specified by the impact type as the positive or negative direction of a certain axis. The basic vibration component, random vibration component, and impact component are added step-by-step to obtain the three-axis external linear acceleration time series. At each time step, the three-axis acceleration is truncated to ensure it does not exceed the acceleration variation range. The attitude change trend is used to generate the external attitude or gravity direction time series. Specifically, at each time step, the rotational state of the box relative to the inertial space is given according to the attitude change trend. The rotational state can be represented by a slowly changing curve of roll, pitch, and yaw. The rate and amplitude of change of the curve are constrained by the attitude change trend. This rotational state is applied to the gravity vector to obtain the gravity component in the box coordinate system at that time step. The final output is the external excitation time series, which includes at least the box coordinate system linear acceleration vector and the box coordinate system gravity component for each time step. If necessary, the angular velocity or angular acceleration sequence can also be output simultaneously to drive the rigid body attitude update.

[0074] At the start of the simulation, the rigid body model of the item to be packaged is placed in a predefined initial position inside the virtual empty box cavity (e.g., near the geometric center of the box), and given a predetermined initial orientation (e.g., each face is parallel to the inner wall of the box). The system sets a fixed simulation time step (e.g., 0.001 seconds) and advances the simulation clock step by step according to this step.

[0075] At each time step, the system performs the following ordered operations to update the state of the rigid body model:

[0076] Based on the current simulation timestamp, the corresponding data is read from the preset "external excitation time series". This series predefines two main types of data for each time step: one is the "equivalent linear acceleration vector" acting on the box coordinate system, which simulates the vibration or impact of the transportation vehicle at that moment; the other is the "gravity component vector" expressed in the box coordinate system at that moment, which simulates the change in the direction of gravity caused by the tilting or rotation of the box.

[0077] The system calculates the resultant force acting on the rigid body. The resultant force consists of two main parts: the first part is the inertial force caused by external acceleration, obtained by multiplying the "equivalent linear acceleration vector" read in the previous step by the mass of the rigid body; the second part is the gravity component vector, obtained by directly multiplying the "gravity component vector" by the mass of the rigid body, resulting in gravity at the current orientation. These two force vectors are synthesized in the box coordinate system. Simultaneously, the system calculates the torque generated at the point of application of the resultant force relative to the center of mass of the rigid body, which affects the object's rotation.

[0078] Based on Newton's second law, the linear velocity of the rigid body is updated using the net force calculated in the previous step. Specifically, the net force is divided by the mass of the rigid body to obtain its linear acceleration at the current time step; then, this linear acceleration is multiplied by the simulation time step to obtain a velocity increment; finally, this velocity increment is added to the linear velocity of the rigid body in the previous time step to obtain the new current linear velocity. For rotational motion, the system uses a similar principle, calculating the angular acceleration using the calculated torque and the moment of inertia of the rigid body, and then updating the angular velocity of the rigid body.

[0079] Numerical integration is used to update the spatial position and attitude of the rigid body using newly calculated velocities. For position, the system multiplies the rigid body's current linear velocity by the simulation time step to obtain a displacement increment, which is then superimposed on the rigid body's position coordinates from the previous time step to determine its new position in the box coordinate system. For attitude (i.e., the object's orientation), the system calculates the rigid body's new attitude at the end of the current time step based on the current angular velocity using a mathematical attitude update algorithm (e.g., converting the angular velocity into a quaternion rate of change and integrating the quaternion).

[0080] After pose update, geometric contact detection is performed, using a rigid circumscribed cuboid and six infinitely large rigid planes as the objects. For each inner surface plane, the distance to the nearest point of the circumscribed cuboid along the plane's normal direction is calculated. If the nearest point distance is positive, no contact is considered; if the nearest point distance is negative, penetration is considered, and the penetration depth is defined as its absolute value. The contact point is determined as the projection of the point on the circumscribed cuboid closest to the plane along the plane's normal direction onto the plane. For planes where penetration occurs, the unit normal of that plane is taken as the contact normal, and the relative velocity of the rigid body at the contact point relative to the plane is calculated. Specifically, the linear velocity of the contact point is calculated from the linear velocity and angular velocity of the rigid body's center of mass, and then the plane velocity is subtracted to obtain the relative velocity, where the plane velocity is zero. The relative velocity is decomposed into normal and tangential components. The normal component is the projection of the relative velocity onto the contact normal direction, and the tangential component is the remaining vector after removing the normal component from the relative velocity. The normal contact force is calculated using a linear spring-damped model. The spring component, determined by the penetration depth and temporary stiffness parameters, pushes the rigid body out of the plane along the normal direction. The damping component, determined by the normal relative velocity and temporary damping parameters, suppresses the penetration velocity along the normal direction. The normal contact force vector is obtained by superimposing the spring and damping components along the normal direction. The tangential friction force is calculated using Coulomb's law of friction. First, the direction of the tangential component is taken as the opposite direction of the friction force. If the tangential component is close to zero, the direction of the friction force is maintained as the tangential direction of the previous time step or a zero vector is taken. Then, the maximum static friction limit is calculated, its magnitude being obtained by multiplying the normal contact force magnitude by the temporary static friction coefficient. Finally, the kinetic friction value is calculated, its magnitude being obtained by multiplying the normal contact force magnitude by the temporary kinetic friction coefficient. The friction state is determined based on whether the tangential relative velocity is zero and whether slippage occurs. When the tangential relative velocity is not zero, dynamic friction is directly applied, and the tangential friction force vector is the dynamic friction value along the friction direction. When the tangential relative velocity is zero or close to zero, static friction constraint is applied. First, the magnitude of the tangential constraint force required to counteract the tangential motion tendency is calculated. If this magnitude does not exceed the maximum static friction limit, the constraint force magnitude is taken and a static friction force vector is formed along the friction direction. If it exceeds the maximum static friction limit, slippage occurs, and the tangential friction force vector is formed using the dynamic friction value. The normal contact force vector and the tangential friction force vector are uniformly represented as a three-dimensional force vector in the box coordinate system. The components of both are added one by one along the three coordinate axes to obtain the composite contact force vector for that time step. This composite vector is defined as the impact force vector for that time step and is recorded along with the contact surface identifier, contact point position, and current time step timestamp. If multiple contacts occur on different inner surfaces at the same time step, the impact force vector of each contact surface can be output separately and multiple records can be generated, or the records can be archived separately according to the contact surface identifier during the event aggregation stage to maintain consistency with subsequent hotspot analysis.

[0081] Simultaneously, the inner surface markers where contact occurs are determined, the position coordinates of the contact point in the corresponding local coordinate system of the inner surface are calculated, and the current simulation time is recorded as the moment of contact occurrence. Contact records that occur continuously on the same inner surface over multiple consecutive time steps are merged, and continuous contact is regarded as a single contact event. For each contact event, the contact surface marker, contact position, moment of occurrence, and corresponding impact force vector are recorded. All contact events are then aggregated in chronological order to form a contact event sequence.

[0082] In this embodiment, a contact event is composed of contact determination results that occur consecutively over time on the same inner surface. When contact between the item to be packaged and the same inner surface is detected in multiple consecutive simulation time steps, the contact determination results within that consecutive time period are merged into a single contact event. The simulation time corresponding to the time step in which contact is first detected is taken as the start time of the contact event, and the simulation time corresponding to the time step in which contact ceases to occur is taken as the end time of the contact event. The duration of the contact event is defined as the time difference between the end time and the start time.

[0083] The aggregation and duration determination of contact events are achieved through the following steps: After obtaining the original contact data recorded according to the simulation time steps, the data is first grouped according to the contact surface identifier. For data on the same inner surface, scanning is performed in timestamp order. The system presets a maximum allowable time discontinuity value (e.g., the time length equivalent to several times the simulation step size). When the scan finds two records belonging to the same contact surface that are temporally adjacent, and the difference between their timestamps is less than or equal to this maximum allowable discontinuity value, these two contacts are considered to be continuous and belong to the same contact event. The timestamp of the latter record is then updated to the current termination time of the event. If the difference between the timestamps of adjacent records exceeds the maximum allowable discontinuity value, the previous contact event is considered to have ended, and a new contact event is started with the current record, whose start time is the timestamp of the current record. After traversal, the duration of each aggregated contact event is the difference between its termination time and start time. The contact event sequence during the training sample acquisition process is also obtained using the above method.

[0084] By analyzing the sequence of contact events, the statistical characteristics of the spatial distribution of contact points and the corresponding impact force vectors on each inner surface of the virtual empty box cavity are extracted, and an impact hotspot distribution map is generated.

[0085] In a preferred embodiment of the present invention, the process of generating an impact hotspot distribution map is as follows:

[0086] The six inner surfaces of the virtual empty box cavity are unfolded into six independent two-dimensional analysis planes. Each contact event in the contact event sequence is traversed, and the contact event is assigned to the corresponding two-dimensional analysis plane according to the contact surface identifier in the contact event.

[0087] On a single two-dimensional analysis plane, the contact positions of all contact events assigned to this plane form a point set. The impact intensity corresponding to each contact position in the point set is calculated by multiplying the magnitude of the impact force vector of the contact event corresponding to that point by the duration of the contact event.

[0088] Based on the impact intensity value at each contact point, an initial influence radius is calculated, and the impact intensity value is positively correlated with the initial influence radius value. Using a two-dimensional Gaussian kernel function, a local impact density distribution field is generated at each contact point as the center and its initial influence radius as the smoothing range. The local impact density distribution fields of all contact points on the two-dimensional analysis plane are superimposed to obtain the impact density cloud map of the inner surface.

[0089] The specific calculation steps for the initial influence radius are as follows: For the currently analyzed two-dimensional plane, find the maximum and minimum values ​​of the impact intensity at all contact points. Then, preset a minimum influence radius value and a maximum influence radius value as the upper and lower limits of the output range. For each contact point on the plane, its impact intensity value is linearly and proportionally mapped to the preset minimum and maximum influence radii according to the range from the minimum to the maximum intensity value across the entire plane. Through this mapping rule, the higher the impact intensity of a point, the larger its calculated initial influence radius, thus ensuring that it occupies a larger smooth range when generating the density field subsequently, and more accurately reflects the concentration of the impact.

[0090] Set a density threshold, extract continuous areas with density values ​​exceeding the density threshold from the impact density cloud map, and mark them as impact hotspot areas; record the boundary polygon coordinates of each impact hotspot area, the average impact density value within the area, and the average direction of all impact force vectors within the area as typical impact force directions.

[0091] The density threshold is set using a dynamic calculation method based on statistical characteristics. The specific process is as follows: After generating a complete impact density contour map of a given inner surface, the arithmetic mean of the density values ​​of all pixels or grid cells in the contour map is calculated, along with the standard deviation of these density values. Then, the density threshold is set to a certain multiple of this mean plus the standard deviation (this multiple is a configurable sensitivity parameter, such as 1.5 times). Finally, on the density contour map, all interconnected regions whose density values ​​exceed this calculated threshold are identified; these regions are designated as "impact hotspots." This method automatically adjusts the threshold based on the density field generated in each simulation, ensuring that statistically significant impact concentration areas can be effectively identified for different items and transportation scenarios.

[0092] The specific steps for calculating the typical impact force direction are as follows: For an identified impact hotspot area, firstly, collect the impact force vectors corresponding to all contact events falling within that area. Then, perform vector addition on all collected impact force vectors to obtain a resultant force vector. Next, convert this resultant force vector into a unit vector of length 1. The direction of this unit vector is defined as the "typical impact force direction" of the hotspot area. It represents the overall trend direction of the impact force experienced by the area. If the length of the calculated resultant force vector is very close to zero, it indicates that the impact force direction experienced by the area is very dispersed, with no obvious dominant direction. In this case, the system can mark the area as having insignificant directional characteristics, or instead use the direction of the single impact force vector with the largest magnitude recorded within the area as its typical direction.

[0093] The boundary coordinates, average impact density value, and typical impact force direction corresponding to each impact hotspot region constitute an impact hotspot descriptor. The set of all impact hotspot descriptors on an inner surface constitutes the impact hotspot distribution map of that inner surface.

[0094] The impact hotspot distribution map in this invention is derived from statistical analysis of contact event sequences. It characterizes the concentrated impact density areas formed by the spatial distribution of contact points on the inner surfaces of a virtual empty box and the corresponding statistical characteristics of the impact force vectors, under preset transportation scenario parameters. The impact hotspot distribution map reflects the concentration trend of contact events in both spatial and impact intensity dimensions, facilitating the mapping and determination of areas to be reinforced on a three-dimensional mesh model and the generation of local reinforcement structures. This allows the corrugated cardboard box structure design to be optimized for areas with more concentrated impact risks.

[0095] The impact hotspot distribution map is mapped onto the 3D mesh model of the virtual empty box cavity, triggering an adaptive structural reinforcement generation algorithm to generate local enhanced 3D data at the corresponding positions in the 3D mesh model;

[0096] In another preferred embodiment of the present invention, the execution process of the adaptive structure reinforcement generation algorithm is as follows:

[0097] The three-dimensional mesh model consists of vertex coordinates and mesh patch topological relationships. An indexable set of patches and internal surface normal information are established for each of the six inner surfaces. Each impact hotspot descriptor in the impact hotspot distribution map contains at least the boundary polygon coordinates, average impact density value, and typical impact force direction.

[0098] During processing, an inner surface is selected, and the coordinates of the two-dimensional boundary polygon corresponding to the inner surface are mapped to the inner surface region of the three-dimensional mesh model according to the correspondence between its local coordinate system and the parameterized coordinates of the inner surface of the three-dimensional mesh model. In the set of inner surface patches, the centroid point or its projection point of each patch is checked one by one to see if it falls within the coverage area of ​​the boundary polygon. All patches that fall within the coverage area are marked as covered mesh patches. The mesh subset corresponding to the covered mesh patches is used as the region to be strengthened, and its vertex set and patch set are output for subsequent geometric fusion.

[0099] The average impact density value and typical impact force direction of the impact hotspot descriptor are read. The angle between the typical impact force direction and the normal of the inner surface is calculated. The range of the average impact density value and the range of the angle are used as query conditions to access the reinforced structure template library. The reinforced structure template library stores a variety of preset reinforced structure 3D templates. Each 3D template is stored in the form of a 3D mesh or parametric geometry and is associated with selection conditions. The selection conditions include the applicable range of average impact density values ​​and the applicable range of angles. During the query, templates that simultaneously meet the density range and angle range are filtered out, and one of them is selected as the target reinforced structure 3D template. If multiple templates meet the conditions, a reporting process is performed.

[0100] The scaling ratio is determined based on the area covered by the boundary polygon on the inner surface. The 3D template of the target reinforcement structure is scaled proportionally to match the scale of the covered area, ensuring that the scaled template falls within or fits the boundary of the area to be reinforced within the planar projection range. Then, the scaled template is rotated according to the typical impact force direction, aligning the principal force direction with the typical impact force direction while maintaining the template's contact direction with the inner surface. The rotated template is positioned in the spatial location of the area to be reinforced by aligning the template's contact reference plane with a local plane of the inner surface and aligning the center of the template's planar projection with the geometric center of the boundary polygon, resulting in the placed reinforcement structure geometry. A Boolean union operation is performed on the placed reinforcement structure geometry and the corresponding original mesh portion of the area to be reinforced. The input to the Boolean union operation is a subset of the original mesh and the reinforcement structure geometry; the output is a merged mesh subset containing the reinforcement structure geometry and connected to the original mesh. The merged mesh subset replaces the original mesh subset and is written back to the 3D mesh model. The reinforced area and its surrounding fusion area undergo smooth re-meshing. During this process, local mesh refinement and vertex position smoothing are performed near the fusion boundary to ensure continuous surface transitions and eliminate sharp edges, while maintaining the overall shape of the inner surface from unexpected shifts. After processing one impact hotspot descriptor, the remaining impact hotspot descriptors on the same inner surface are traversed, and the mapping, model selection, scaling, rotation, Boolean merging, and re-meshing steps are repeated. Subsequently, the process is switched to the remaining inner surfaces and processed cyclically according to their respective impact hotspot distribution maps. Finally, all local reinforcement structures are generated on the 3D mesh model, resulting in a fused 3D mesh model. The mesh geometry representing the reinforcement structure in the fused 3D mesh model is output as local reinforcement 3D data and used for subsequent corrugated cardboard box production file generation.

[0101] The key points of the Boolean union operation are as follows: The purpose of this operation is to seamlessly merge the mesh geometry of the reinforced structure with the original mesh of the inner wall of the carton. During the operation, the system first accurately calculates the intersection line between the reinforced structure mesh and the original mesh of the area to be reinforced. Then, using this intersection line as the new boundary, the faces of the two meshes near the intersection are re-divided and triangulated to ensure that the merged surface is a continuous and complete mesh model, and to remove redundant parts of the original mesh that are obscured by the interior of the reinforced structure.

[0102] Key points for smooth re-partitioning: After Boolean union, the boundaries of the merged region may have sharp edges or uneven mesh. Smoothing first locates all vertices and edges at the merge junction. Then, using an iterative relaxation algorithm (such as Laplacian smoothing), the positions of these vertices are slightly adjusted to move closer to the average position of their neighboring vertices, thus smoothing out sharp seams. Simultaneously, the mesh quality is checked, and locally optimized triangles with poor shapes or uneven sizes are performed, for example, by splitting excessively long edges, merging excessively short edges, or adjusting diagonal connections, resulting in a more uniform and higher-quality mesh distribution in that region. This lays a good geometric foundation for subsequent carton unfolding and manufacturing document generation.

[0103] Output corrugated cardboard box production files that integrate locally enhanced 3D data.

[0104] The process of generating corrugated cardboard box production documents is as follows:

[0105] The process involves: inputting a 3D mesh model incorporating locally enhanced 3D data; converting the 3D mesh model into a solid 3D model of a cardboard box with defined wall thickness; performing geometric unfolding calculations on the solid cardboard box model, considering material extension and folding allowance, and generating a 2D planar unfolded diagram with adhesive edges; the planar unfolded diagram containing all crease lines, external contour cutting lines, and the projected contour lines of the locally enhanced structures on the 2D plane; identifying the regions defined by the projected contour lines of the locally enhanced structures in the planar unfolded diagram; adding a processing mark to each identified region in the associated data structure of the planar unfolded diagram, with the attribute fields of the processing mark containing at least an enhancement structure type identifier and a processing priority code; integrating the geometric data of the planar unfolded diagram, the attribute data of all processing marks, and the internal dimensions of the box derived from the final dimensions of the virtual empty box cavity, encoding and encapsulating them according to a predefined file format specification, and outputting a complete corrugated cardboard box production file; the corrugated cardboard box production file can be directly read and executed by downstream CNC cutting or laser engraving equipment.

[0106] A corrugated cardboard box design optimization system based on a machine learning model includes:

[0107] Simulation module: Constructs a kinematic simulation model of the item to be packaged within a virtual empty box cavity based on the physical parameters of the item to be packaged.

[0108] Model calibration module: Based on the pre-trained simulation calibration model, the preset transportation scenario parameters and the physical parameters are processed to obtain a set of simulation calibration coefficients, and the kinematic simulation model is calibrated based on the simulation calibration coefficients.

[0109] Impact recording module: The calibrated kinematic simulation model is run under preset transportation scenario parameters to record the position, time and impact force vector of the item to be packaged contacting the inner surface of the virtual empty box cavity during the simulation process, and generate a contact event sequence.

[0110] By analyzing the sequence of contact events, the statistical characteristics of the spatial distribution of contact points and the corresponding impact force vectors on each inner surface of the virtual empty box cavity are extracted, and an impact hotspot distribution map is generated.

[0111] Optimization module: Maps the impact hotspot distribution map onto the 3D mesh model of the virtual empty box cavity, triggers the adaptive structural reinforcement generation algorithm, and generates locally enhanced 3D data at the corresponding positions in the 3D mesh model.

[0112] Output corrugated cardboard box production files that integrate locally enhanced 3D data.

[0113] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for optimizing the design of corrugated cardboard boxes based on a machine learning model, characterized in that, Includes the following steps: A kinematic simulation model of the items to be packaged within a virtual empty box cavity is constructed based on the physical parameters of the items to be packaged. The pre-trained simulation calibration model processes the preset transportation scenario parameters and the physical parameters to obtain a set of simulation calibration coefficients, and then calibrates the kinematic simulation model based on the simulation calibration coefficients. The calibrated kinematic simulation model is run under preset transportation scenario parameters. The positions, times, and impact force vectors of the items to be packaged coming into contact with the inner surfaces of the virtual empty box cavity during the simulation are recorded to generate a contact event sequence. Analyze the contact event sequence, extract the statistical characteristics of the spatial distribution of contact points and corresponding impact force vectors on each inner surface of the virtual empty box cavity, and generate an impact hot spot distribution map; The impact hotspot distribution map is mapped onto the 3D mesh model of the virtual empty box cavity, triggering an adaptive structural reinforcement generation algorithm to generate local enhanced 3D data at the corresponding positions in the 3D mesh model; Output corrugated cardboard box production files that integrate locally enhanced 3D data.

2. The method for optimizing corrugated cardboard box design based on a machine learning model according to claim 1, characterized in that, The process of constructing a kinematic simulation model is as follows: The physical parameters include the external dimensions, mass, center of gravity, and surface friction coefficient of the item to be packaged, and a virtual empty box cavity is generated based on the external dimensions. Based on physical parameters, the item to be packaged is simplified into a rigid body model, and the contact mechanical relationship between the rigid body model and the inner surface of the virtual empty box cavity is defined. The six inner surfaces of the virtual empty box cavity are defined as infinitely large rigid plane constraints. The defined rigid body model, contact mechanics relationship and rigid plane constraints together constitute the kinematic simulation model.

3. The method for optimizing corrugated cardboard box design based on a machine learning model according to claim 2, characterized in that, The pre-training process of the simulation calibration model is as follows: Input features for training samples are constructed based on the physical parameters of historical items and the corresponding transportation scenario parameters. Obtain real contact event sequences based on physical transport test records of historical items; Based on the initial kinematic simulation model with default parameters, a sequence of simulated contact events is obtained. A set of simulation calibration coefficients is derived from the real contact event sequence and the simulated contact event sequence. These simulation calibration coefficients are used to calibrate the contact mechanical relationship. All training samples form a training set. A neural network model is trained using the training set. The trained neural network model is the pre-trained simulation calibration model.

4. The method for optimizing the design of corrugated cardboard boxes based on a machine learning model according to claim 3, characterized in that, The process of generating the impact hotspot distribution map is as follows: The six inner surfaces of the virtual empty box cavity are unfolded into six two-dimensional analysis planes, and each contact event in the contact event sequence is assigned to the corresponding two-dimensional analysis plane. On a single two-dimensional analysis plane, the contact locations of all contact events assigned thereto form a point set. The impact intensity corresponding to each contact location in the point set is calculated. Based on the impact intensity value of each contact location, an initial influence radius is calculated. Using a two-dimensional Gaussian kernel function, a local impact density distribution field is generated at each contact point, with its initial influence radius as the smoothing range. The impact hotspot region is obtained based on the entire local impact density distribution field. The boundary polygon coordinates, average impact density, and typical impact force direction of each impact hotspot region constitute an impact hotspot descriptor. The set of all impact hotspot descriptors on an inner surface constitutes the impact hotspot distribution map of that inner surface.

5. The method for optimizing corrugated cardboard box design based on a machine learning model according to claim 4, characterized in that, The execution process of the adaptive structure reinforcement generation algorithm is as follows: For an impact hotspot distribution map corresponding to an inner surface, a region to be strengthened is located in the three-dimensional mesh model based on the coordinates of the boundary polygon; Based on the impact hotspot descriptor, query the reinforced structure template library and select a reinforced structure 3D template from the library; After scaling, rotating, and performing Boolean union operations on the selected reinforced structure 3D template, the reinforced structure geometry is merged into part of the original mesh; All impact hotspot descriptors are processed, and all local enhancement structures are generated on the 3D mesh model to obtain the fused 3D mesh model; the local enhancement 3D data refers to the part of the mesh geometry data that represents the enhancement structure in the fused 3D mesh model.

6. The method for optimizing the design of corrugated cardboard boxes based on a machine learning model according to claim 5, characterized in that, The process of obtaining simulation calibration coefficients based on a pre-trained simulation calibration model is as follows: Based on the preset transportation scenario parameters and the physical parameters, a set of input features is constructed. The input features are then input into the pre-trained simulation calibration model, which outputs a set of simulation calibration coefficients A. The simulation calibration coefficients A calibrate the contact mechanics relationship in the kinematic simulation model.

7. A corrugated cardboard box design optimization system based on a machine learning model, characterized in that, include: Simulation module: Constructs a kinematic simulation model of the item to be packaged within a virtual empty box cavity based on the physical parameters of the item to be packaged; Model calibration module: Based on the pre-trained simulation calibration model, the preset transportation scenario parameters and the physical parameters are processed to obtain a set of simulation calibration coefficients, and the kinematic simulation model is calibrated based on the simulation calibration coefficients; Impact recording module: The calibrated kinematic simulation model is run under preset transportation scenario parameters to record the position, time and impact force vector of the item to be packaged contacting the inner surface of the virtual empty box cavity during the simulation process, and generate a contact event sequence. Analyze the contact event sequence, extract the statistical characteristics of the spatial distribution of contact points and corresponding impact force vectors on each inner surface of the virtual empty box cavity, and generate an impact hot spot distribution map; Optimization module: Maps the impact hotspot distribution map onto the 3D mesh model of the virtual empty box cavity, triggers the adaptive structural reinforcement generation algorithm, and generates local enhanced 3D data at the corresponding positions in the 3D mesh model; Output corrugated cardboard box production files that integrate locally enhanced 3D data.

Images