A golf club topology generation method based on reinforcement learning

Abstract

The application relates to the field of sports equipment manufacturing, in particular to a golf club topological structure generation method based on reinforcement learning, which comprises the following steps: constructing a historical hitting scatter point matrix and a corresponding group electromyogram inconsistency index set; performing impact simulation and obtaining physical simulation results, combining biomechanical characteristics with the physical simulation results, and calculating a deviation penalty term; and taking minimizing the deviation penalty term as a target, updating a strategy by using a reinforcement learning algorithm, and continuously iterating a club face thickness distribution and a club weight distribution. The scheme sets a deviation penalty term based on a non-perfect eccentric hitting, can specially perform topological inversion on the human-specific eccentric hitting error, generates an asymmetric intelligent thickness structure with automatic deviation correction physical characteristics, is higher in fitting degree with the actual situation, and can manufacture a club more in line with the actual use habit of human bodies.

Description

Technical Field

[0001] This invention relates to the field of sports equipment manufacturing, and in particular to a method for generating golf club topology based on reinforcement learning. Background Technology

[0002] With the continuous development of sports equipment manufacturing technology, golf club design has gradually shifted from the traditional trial-and-error method relying on experience to digital design based on computer-aided engineering (CAE). The topology of a golf club (especially the clubhead), namely the distribution of its internal materials and the form of its skeleton connections, directly determines the clubhead's weight, center of gravity (CG), moment of inertia (MOI), and energy transfer efficiency at impact. In pursuit of greater distance and higher forgiveness, designers typically need to perform topology optimization on the clubhead within strict weight constraints to achieve the ultimate improvement in mechanical performance.

[0003] The prior art, such as CN103357161A, discloses a method for selecting a golf club shaft, comprising the following steps: striking a golf ball with a golf club on which a sensor capable of measuring angular velocity about three axes is mounted on the grip and obtaining a measurement value from the sensor; determining the swing aim, swing apex, and impact point based on the measurement value; and selecting a shaft that matches the golfer using the following swing characteristic quantities (a) to (d) obtained from the measurement value: (a) the change in grip angular velocity in the flexion swing direction near the swing apex; (b) the average value of the grip angular velocity in the flexion swing direction from the swing apex until the flexion swing direction reaches its maximum during the downswing; (c) the average value of the grip angular velocity in the flexion swing direction from the moment the flexion swing direction reaches its maximum during the downswing until impact; and (d) the average value of the grip angular velocity in the flexion swing direction from the swing apex until impact.

[0004] Existing technologies generate topologies based on the assumption of a perfect center shot, pursuing clubface symmetry, which is significantly out of touch with actual usage. This invention was developed to address this common problem in the field. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of current methods by proposing a reinforcement learning-based method for generating golf club topology structures.

[0006] To overcome the shortcomings of the prior art, the present invention adopts the following technical solution:

[0007] A method for generating golf club topology based on reinforcement learning includes the following steps:

[0008] S1, Obtain historical shot data and biomechanical characteristics of the target player group, and construct a historical shot scatter matrix and the corresponding group electromyographic inconsistency index set; the shot data is the impact coordinates of the ball on the clubface;

[0009] S2, In the rigid body mechanics simulation engine, based on the current iteration of the rod surface thickness distribution and club weight distribution, the scatter point matrix is ​​subjected to impact simulation and physical simulation results are obtained. Combined with biomechanical characteristics and physical simulation results, the deviation penalty term is calculated.

[0010] S3, with the goal of minimizing the deviation penalty term, the strategy is updated using a reinforcement learning algorithm, and the clubface thickness distribution and club weight distribution are iterated continuously until the convergence condition is met, thereby autonomously generating an asymmetric and irregular intelligent clubface topology 3D model.

[0011] Furthermore, step S1 also includes the following sub-steps:

[0012] S11 acquires historical shot data of the target player group through high-speed cameras and shot monitoring radar, maps the impact coordinates of the ball on the clubface into a two-dimensional plane point set, and constructs a historical shot scatter matrix;

[0013] S12, synchronously collect the surface electromyography (EMG) signals of the player group during historical ball hits, and extract their corresponding EMG features;

[0014] S13, Spatial clustering is performed on the scatter point matrix, and the variance of electromyographic features within each cluster is normalized to calculate the population electromyographic inconsistency index corresponding to each scatter point in the matrix.

[0015] Furthermore, in S2, the impact simulation of the scattered point matrix and the resulting physical simulation include the following steps:

[0016] S21, calculate the local contact stiffness coefficient of each mesh node on the club face based on the current iteration of the club face thickness distribution, and calculate the rotational inertia tensor and center of mass position of the club head rigid body based on the current iteration of the club weight distribution;

[0017] S22, For the scattered points in the scattered point matrix, calculate the lever arm vector of the impact point relative to the center of mass of the club based on the impact coordinates corresponding to the scattered points and the corresponding club center of mass position, and calculate the impact impulse at the time of impact based on the local contact stiffness coefficient.

[0018] S23, calculate the torque generated by the collision, and solve the angular velocity change of the clubhead around the center of mass of the club by combining the moment of inertia tensor. The side spin angular velocity generated by the clubface is obtained by decomposing the angular velocity change.

[0019] S24: Obtain the collision reflection velocity based on the collision impulse, and calculate the golf ball's flight trajectory based on the collision reflection velocity and the side spin angular velocity, combined with the aerodynamic model, and calculate the geometric distance between the landing point and the preset target point.

[0020] Furthermore, in S3, the iterative process based on the deviation penalty term includes the following steps:

[0021] S31. Name the combined matrix of the current clubface thickness distribution matrix and the club weight distribution matrix as the state matrix. The combined matrix of thickness adjustment matrix and weight adjustment matrix for each position of the cue stick is named the motion matrix. The negative deviation penalty term is defined as a reward signal;

[0022] S32, Output the action matrix according to the current strategy. Applying to the rigid body mechanics simulation engine, the engine updates the state matrix as follows: And calculate the new based on step S2 Return reward signal ;

[0023] S33, with the goal of maximizing the expected cumulative reward signal, the policy network is updated, and the system updates the action matrix based on the policy network. ;

[0024] S34, Repeat the above steps until K consecutive generations The process terminates when the change is less than a preset threshold or when the maximum number of iterations is reached, outputting the finally converged asymmetric, irregular intelligent pole surface topology 3D model.

[0025] A golf club topology generation system based on full-scene coupling of reinforcement learning and rigid body dynamics, the system includes a data acquisition and processing module, a calculation module and a reinforcement learning iterative topology generation module;

[0026] The data acquisition and processing module includes a historical scatter matrix construction unit and a biomechanical feature extraction unit. The historical scatter matrix construction unit is used to acquire historical shot data of a specific player group, map it to a set of coordinate points on the clubface, and form a historical shot scatter matrix. The biomechanical feature extraction unit is used to simultaneously acquire electromyographic signals of the specific player group during historical imperfect shots, extract muscle exertion disorder features, and calculate the group electromyographic inconsistency index.

[0027] Furthermore, the calculation module includes a dynamic simulation execution unit and a coupling deviation penalty calculation unit;

[0028] The dynamics simulation execution unit includes a parametric physics mapping substructure, an eccentric collision calculation substructure, a dynamic response solving substructure, and a trajectory solving substructure. The parametric physics mapping substructure is used to calculate the local contact stiffness coefficient based on the current shaft thickness distribution and the rotational inertia tensor based on the club weight distribution. The eccentric collision calculation substructure is used to calculate the impact arm and impact impulse based on the coordinates of the scattering points. The dynamic response solving substructure is used to calculate the torque and solve for the output side spin angular velocity. The trajectory solving substructure calculates the flight trajectory and outputs the landing point deviation distance.

[0029] The coupling deviation penalty calculation unit is used to calculate the deviation penalty item based on various data.

[0030] Furthermore, the reinforcement learning iterative topology generation module includes a state-action matrix acquisition unit and a policy update and output unit;

[0031] The state-action space definition unit is used to generate action matrices and state matrices; the policy update and output unit is used to obtain reward signals, continuously update the policy network, and iteratively generate new pole face thickness and weight parameters until the deviation from the penalty term converges, and finally autonomously outputs an asymmetric and irregular intelligent pole face topology 3D model.

[0032] The beneficial effects achieved by this invention are as follows: 1. Existing technologies are based on the assumption of perfect center-of-shot impact and pursue clubface symmetry; this solution is based on the reality of imperfect eccentric impact and pursues reverse adaptive correction of human error distribution, actively generating asymmetric and irregular topology; by setting a deviation penalty term based on imperfect eccentric impact, it breaks through the shortcomings of traditional testing that pursues perfect center-of-shot impact, and can specifically perform topological inversion for the unique human eccentric impact error, generating an asymmetric intelligent thickness structure with automatic correction physical characteristics, which is more in line with reality and can manufacture clubs that are more in line with the actual human usage habits.

[0033] 2. Introducing the population electromyographic inconsistency index as an exponential term into the deviation penalty term, so that when the player population deviates from the target due to muscle exertion disorder (physiological error, exponent greater than or equal to 1) in a certain area, the penalty value is amplified exponentially, forcing the AI ​​to increase the thickness or shift the center of gravity in that area to provide extreme correction; while for accidental deviations when muscle exertion is stable (exponent less than 1), the penalty value is reduced logarithmically, avoiding over-optimization, and giving the topology a dynamic nonlinear fault tolerance capability for human physiological characteristics. Attached Figure Description

[0034] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate the same parts.

[0035] Figure 1 This is a flowchart of the process of the present invention.

[0036] Figure 2 This is a structural framework diagram of the present invention.

[0037] Figure 3 The figures show the comparative effects of the present invention and Comparative Examples 1 and 2.

[0038] Figure 4 This is a comparison diagram of the effects of the present invention and Comparative Example 3. Detailed Implementation

[0039] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0040] Example 1: This example provides a method for generating golf club topology based on reinforcement learning, including the following steps:

[0041] S1, Obtain historical shot data and biomechanical characteristics of the target player group, and construct a historical shot scatter matrix and the corresponding group electromyographic inconsistency index set; the shot data is the impact coordinates of the ball on the clubface;

[0042] S2, In the rigid body mechanics simulation engine, based on the current iteration of the rod surface thickness distribution and club weight distribution, the scatter point matrix is ​​subjected to impact simulation and physical simulation results are obtained. Combined with biomechanical characteristics and physical simulation results, the deviation penalty term is calculated.

[0043] S3, with the goal of minimizing the deviation penalty term, the strategy is updated using a reinforcement learning algorithm, and the clubface thickness distribution and club weight distribution are iterated continuously until the convergence condition is met, thereby autonomously generating an asymmetric and irregular intelligent clubface topology 3D model.

[0044] Furthermore, step S1 also includes the following sub-steps:

[0045] S11 acquires historical shot data of the target player group through high-speed cameras and shot monitoring radar, maps the impact coordinates of the ball on the clubface into a two-dimensional plane point set, and constructs a historical shot scatter matrix;

[0046] S12, synchronously collect the surface electromyography (EMG) signals of the player group during historical ball hits, and extract their corresponding EMG features;

[0047] S13, Spatial clustering is performed on the scatter point matrix, and the variance of electromyographic features within each cluster is normalized to calculate the population electromyographic inconsistency index corresponding to each scatter point in the matrix.

[0048] Specifically, the population electromyographic inconsistency index for the i-th scattering point is calculated using the following method:

[0049] Before performing spatial clustering on the scatter point matrix, the basic electromyographic disorder index for a single shot is first calculated according to the following formula:

[0050]

[0051] in, Let be the baseline electromyographic disorder level of the k-th player at the i-th scattering point of the ball. and The sum of and is 1, and both are preset weight coefficients. These weight coefficients are obtained by those skilled in the art through continuous training and optimization during model training. Let $\frac{k}{i}$ be the amplitude of forearm extensor muscle activation when the $k$ player hits the ball during the $i$ scattering phase. Let $\frac{k}{i}$ be the amplitude of flexor muscle activation when the $k$ player hits the ball at the $i$-th scattering point. Let the timing difference of flexor and extensor muscle activation of the k-th player at the i-th scattering point of the ball be the time difference. This is the preset maximum allowable timing difference;

[0052] Furthermore, spatial density clustering is then performed on the scatter point matrix to divide the matrix into D clusters, thus determining the cluster to which the i-th scatter point belongs. ;

[0053] Further calculations Basic electromyographic disorder variance The formula is as follows:

[0054]

[0055] in, Belongs to cluster The number of ball-hitting samples, This represents the mean baseline electromyographic disorder of all samples within the cluster.

[0056] Furthermore, the variance of the basic electromyographic disorder corresponding to all clusters is normalized using Min-Max and mapped to the interval (0, 2) to obtain the population electromyographic inconsistency index of the i-th scatter point. The formula is as follows:

[0057]

[0058] in, and These represent the maximum and minimum variances of the baseline electromyographic disorder for all clusters, respectively. To reach the minimum value, through the above calculations, Strictly confined to the interval (0, 2) and being a dimensionless exponent, when the flexor-extensor synergy and dispersion of the group hitting the ball in a certain area are poorer, the following applies: The closer it is to 2; the more stable and consistent the group's effort, the better. The closer it gets to 0.

[0059] Furthermore, in S2, the impact simulation of the scattered point matrix and the resulting physical simulation include the following steps:

[0060] S21, calculate the local contact stiffness coefficient of each mesh node on the club face based on the current iteration of the club face thickness distribution, and calculate the rotational inertia tensor and center of mass position of the club head rigid body based on the current iteration of the club weight distribution;

[0061] S22, For the scattered points in the scattered point matrix, calculate the lever arm vector of the impact point relative to the center of mass of the club based on the impact coordinates corresponding to the scattered points and the corresponding club center of mass position, and calculate the impact impulse at the time of impact based on the local contact stiffness coefficient.

[0062] S23, calculate the torque generated by the collision, and solve the angular velocity change of the clubhead around the center of mass of the club by combining the moment of inertia tensor. The side spin angular velocity generated by the clubface is obtained by decomposing the angular velocity change.

[0063] S24: Obtain the collision reflection velocity based on the collision impulse, and calculate the golf ball's flight trajectory based on the collision reflection velocity and the side spin angular velocity, combined with the aerodynamic model, and calculate the geometric distance between the landing point and the preset target point.

[0064] Specifically, the calculations from S21 to S24 all use existing methods. For example, the steps for calculating the local contact stiffness coefficient of each mesh node on the club face based on the current iteration's club face thickness distribution are as follows: Using contact algorithms in finite element analysis (FEA) (such as the contact stiffness models in ANSYS and Abaqus), inputting the material type, geometric thickness of the mesh nodes, and contact conditions automatically calculates the local contact stiffness. The steps for calculating the moment of inertia tensor and center of mass position of the club head rigid body based on the current iteration's club weight distribution are as follows: Using finite element software (such as ANSYS's CMOMENT or Abaqus's INERTIA), a club head geometric model must first be established. Inputting the weight distribution automatically yields the center of mass position, and then inputting the material density allows the software to calculate the moment of inertia tensor. The aerodynamic model mentioned is an existing model and will not be elaborated upon here.

[0065] Specifically, the deviation penalty term is calculated according to the following formula:

[0066]

[0067] The formula serves as the optimization objective for the reinforcement learning algorithm, comprehensively evaluating the degree of coupling deviation between the player's biomechanical characteristics and the physical shot result. It guides the algorithm to generate asymmetric topologies with automatic correction capabilities to address human-specific eccentricity errors. The deviation penalty term is represented by I, which is the total number of scatter points in the historical shot scatter point matrix. These are the weighting coefficients for the geometric distance term. These are the weighting coefficients for the spin term, which are obtained by those skilled in the art through continuous training and optimization during model training. Let be the geometric distance between the ball's landing point and the target point corresponding to the i-th scatter point. The system's preset geometric distance normalization reference value, Let i be the population electromyographic inconsistency index for the i-th scattering point. The Magnus rotation coupling coefficient reflects the efficiency of converting side spin into the actual yaw trajectory under specific environmental conditions and the microscopic features of the pole surface. This coefficient is obtained by the system by substituting the current environmental temperature and humidity sensor data and the pole surface micro-roughness matrix of the current iteration into the Magnus effect correction equation. The Magnus effect correction equation is existing technology and will not be elaborated upon here. Let be the sidespin angular velocity generated by the hit corresponding to the i-th scatter point. The preset normalized reference value for the side rotation angular velocity.

[0068] Furthermore, in S3, the iterative process based on the deviation penalty term includes the following steps:

[0069] S31. Name the combined matrix of the current clubface thickness distribution matrix and the club weight distribution matrix as the state matrix. The combined matrix of thickness adjustment matrix and weight adjustment matrix for each position of the cue stick is named the motion matrix. The negative deviation penalty term is defined as a reward signal;

[0070] Specifically, each position on the cue corresponds to an element in a matrix. By concatenating two matrices, a combined matrix can be obtained, representing the action. Each item in the table corresponds to a thickness adjustment or weight adjustment.

[0071] Specifically, the reward signal is , .

[0072] S32, Output the action matrix according to the current strategy. Applying to the rigid body mechanics simulation engine, the engine updates the state matrix as follows: And calculate the new based on step S2 Return reward signal ;

[0073] S33, with the goal of maximizing the expected cumulative reward signal, the policy network is updated, and the system updates the action matrix based on the policy network. ;

[0074] S34, Repeat the above steps until K consecutive generations When the change is less than a preset threshold, or when the maximum number of iterations is reached, the process terminates and outputs the finally converged asymmetric, irregular intelligent pole surface topology 3D model.

[0075] Specifically, the strategy network parameters include how much thickness and weight to add (or remove) at each position of the cue each time.

[0076] Specifically, each time the policy network is updated, the system will randomly update and record the action matrix corresponding to that update. And reward signals, and make the probability of updating methods that can increase reward signals increase in the next update.

[0077] Specifically, the K value is set by those skilled in the art based on actual needs and required precision.

[0078] A golf club topology generation system based on the coupling of reinforcement learning and rigid body dynamics across all scenarios is applied to a reinforcement learning-based golf club topology generation method. The system includes a data acquisition and processing module, a computation module, and a reinforcement learning iterative topology generation module.

[0079] The data acquisition and processing module includes a historical scatter matrix construction unit and a biomechanical feature extraction unit. The historical scatter matrix construction unit is used to acquire historical shot data of a specific player group, map it to a set of coordinate points on the clubface, and form a historical shot scatter matrix. The biomechanical feature extraction unit is used to simultaneously acquire electromyographic signals of the specific player group during historical imperfect shots, extract muscle exertion disorder features, and calculate the group electromyographic inconsistency index.

[0080] Furthermore, the calculation module includes a dynamic simulation execution unit and a coupling deviation penalty calculation unit;

[0081] The dynamics simulation execution unit includes a parametric physics mapping substructure, an eccentric collision calculation substructure, a dynamic response solving substructure, and a trajectory solving substructure. The parametric physics mapping substructure is used to calculate the local contact stiffness coefficient based on the current shaft thickness distribution and the rotational inertia tensor based on the club weight distribution. The eccentric collision calculation substructure is used to calculate the impact arm and impact impulse based on the coordinates of the scattering points. The dynamic response solving substructure is used to calculate the torque and solve for the output side spin angular velocity. The trajectory solving substructure calculates the flight trajectory and outputs the landing point deviation distance.

[0082] The coupling deviation penalty calculation unit is used to calculate the deviation penalty item based on various data.

[0083] Furthermore, the reinforcement learning iterative topology generation module includes a state-action matrix acquisition unit and a policy update and output unit;

[0084] The state-action space definition unit is used to generate action matrices and state matrices; the policy update and output unit is used to obtain reward signals, continuously update the policy network, and iteratively generate new pole face thickness and weight parameters until the deviation from the penalty term converges, and finally autonomously outputs an asymmetric and irregular intelligent pole face topology 3D model.

[0085] The beneficial effects of this solution are as follows: 1. Existing technologies are based on the assumption of a perfect center-of-shot strike and pursue clubface symmetry; this solution is based on the reality of imperfect eccentric strikes and pursues reverse adaptive correction of human error distribution, actively generating asymmetric and irregular topologies; by setting a deviation penalty term based on imperfect eccentric strikes, it breaks through the shortcomings of traditional testing that pursues a perfect center-of-shot impact, and can specifically perform topological inversion for the unique human eccentric strike error, generating an asymmetric intelligent thickness structure with automatic correction physical characteristics, which is more in line with reality and can manufacture clubs that are more in line with the actual human usage habits.

[0086] 2. Introducing the population electromyographic inconsistency index as an exponential term into the deviation penalty term, so that when the player population deviates from the target due to muscle exertion disorder (physiological error, exponent greater than or equal to 1) in a certain area, the penalty value is amplified exponentially, forcing the AI ​​to increase the thickness or shift the center of gravity in that area to provide extreme correction; while for accidental deviations when muscle exertion is stable (exponent less than 1), the penalty value is reduced logarithmically, avoiding over-optimization, and giving the topology a dynamic nonlinear fault tolerance capability for human physiological characteristics.

[0087] Example 2: This example should be understood as including all the features of any of the foregoing examples, and further improving upon them, and also includes a thickness distribution mapping method based on local Gaussian curvature deviation and material strain rate sensitivity index;

[0088] In Example 1, step S2 simply maps the clubface thickness distribution to a local contact stiffness coefficient, which is a prior art based on static linearity. However, in the actual physical scenario of a golf shot, the clubface impact time is extremely short (approximately 0.4-0.5 milliseconds), and clubface materials such as titanium alloys exhibit significant strain rate effects under high-speed impact. Furthermore, the local curvature of the clubface greatly affects the propagation of stress waves. The static mapping of prior art leads to distortion in the calculation of side spin under high-speed impact. To address this issue, this solution introduces a dynamic nonlinear mapping formula, enabling the simulation engine to approximate the real physical scenario.

[0089] The dynamic nonlinear mapping formula can be used to map the current iteration of the bar surface thickness distribution to the local contact stiffness coefficient of each grid node on the bar surface. Taking a certain grid node as an example, the formula is as follows:

[0090]

[0091] in, This represents the current local dynamic contact stiffness coefficient of the mesh node. Let $\frac{ ... The iteration thickness of the current grid node. The system's preset reference thickness is given by x, where x is the Gaussian curvature coupling weighting coefficient of the bar surface. The local Gaussian curvature deviation factor of the bar surface. It is a logarithmic function. Let be the normal velocity of the ball relative to the clubface at the moment of impact. The reference velocity is y (usually taken as 40 m / s), and y is the strain rate damping coefficient. It is an exponential function. The strain rate sensitivity index is the material of the bar surface. Let be the local volumetric average strain rate at the impact point of the rod surface.

[0092] Specifically, the local Gaussian curvature deviation factor (Unit is) The method for obtaining the impact point is as follows: In the 3D mesh model of the rod surface generated in the current iteration, locate the impact region (scatter point mapping position). Extract the mesh nodes at this location and its surrounding adjacent nodes, and fit a local quadratic surface using a differential geometry algorithm. Calculate the actual Gaussian curvature of the local quadratic surface of the rod surface. (i.e., the product of the two principal curvatures), local Gaussian curvature deviation factor According to the following formula:

[0093]

[0094] in, The target reference Gaussian curvature is preset for the system.

[0095] Specifically, the Gaussian curvature coupling weight coefficient x (unit: The parameters are empirical calibration parameters. They are obtained through orthogonal experimental design and physical prototype impact tests. Specifically, several rod head prototypes with different Gaussian curvatures (e.g., flat surfaces, convex surfaces with different arcs) but uniform thickness are fabricated. These prototypes are impacted at different speeds using a high-speed launcher (e.g., TrackMan radar or robotic swing device), and the actual impact contact stiffness is measured. Then, the same conditions are reproduced in the simulation engine, and the value of x is adjusted so that the calculated value is... The value of x is the calibration value when the error between the actual measured value and the physical value is minimized.

[0096] Specifically, the strain rate damping coefficient y reflects the saturation effect of the stiffness increase of the bar face material under extremely high strain rates. It is obtained by performing dynamic impact tests on a split Hopkinson bar (SHPB) to acquire the dynamic flow stress-strain curves of the bar head material within its strain rate range. The strain rate stiffening term in the dynamic nonlinear mapping formula is fitted using the experimental data, and the strain rate damping coefficient characterizing the material's saturation effect is extracted by reverse calculation.

[0097] Specifically, the strain rate sensitivity index of the bar surface material The value is used to characterize the severity of the response of the bar surface material to the relay rate. It is obtained by fitting constitutive curves at different strain rates through dynamic impact tests of Hopkinson bar (SHPB), and the unit is seconds.

[0098] Specifically, the local volume-average strain rate at the impact point The data is obtained as follows: during impact simulation in a dynamics simulation engine (such as a solver based on ANSYS LS-DYNA or Abaqus Explicit). When the ball travels at a velocity... Upon impact with the rod surface, the strain rate tensor of all Gaussian integration points of the finite element mesh within a certain volume range around the impact point (the scattering point mapping area) is extracted (usually a hemispherical influence zone with a radius of 5-10 mm centered at the impact point). The strain rate tensor is then integrated and averaged according to the volume weight of each mesh to obtain the local volume-averaged strain rate at the impact point. .

[0099] The beneficial effects of this embodiment are as follows: Compared to existing technologies, it takes into account the unique high-speed transient characteristics of golf shots. When an eccentric shot occurs, stress waves focus in areas of high curvature on the clubface, and the high strain rate causes the material to "harden" instantaneously. This mapping formula allows the simulation engine to realistically reflect the deflection effect of curvature on stress distribution and the nonlinear increase in stiffness caused by material strain rate hardening when calculating the impact impulse. This results in a reduction in the error of the calculated sidespin angular velocity and a more solid mechanical foundation for topology generation.

[0100] Example 3: This example should be understood as including all the features of any of the foregoing examples, and further improving upon them. It also includes more equipment details and implementation details, and provides a comparative example and comparative effect between the present invention and the prior art.

[0101] In this embodiment, 36 target players were selected as the target player group, including 12 beginner, 12 intermediate, and 12 advanced players. Each player used a test prototype with the same specification 7-iron or the same specification driver head to conduct the shot test, and each player completed 80 valid shots.

[0102] The ball-hitting monitoring device can be any one of the TrackMan 4 radar system, the Foresight GCQuad optical ball-hitting monitoring system, or the FlightScope X3 radar system; the high-speed camera device can be a Phantom VEO 710L high-speed camera or a Photron FASTCAM Mini AX200 high-speed camera, with a sampling frame rate set to 1000 fps to 3000 fps, preferably 2000 fps. The surface electromyography (EMG) acquisition device can be a Delsys Trigno Avanti wireless surface EMG system or a Noraxon Ultium EMG system, with a sampling frequency set to 2000 Hz. The EMG electrodes are respectively placed in the muscle belly regions of the forearm extensor and forearm flexor muscles, with an electrode spacing of 20 mm to 25 mm.

[0103] In one specific implementation, the clubface coordinate system is based on the geometric center of the clubface as the origin, with the toe direction of the clubface as the positive X-axis and the upper edge direction of the clubface as the positive Y-axis. The effective striking area of ​​the clubface is set as a two-dimensional region with a width of 95 mm and a height of 55 mm, and is divided into 19×11 grid nodes with a spacing of 5 mm between adjacent grid nodes.

[0104] For electromyographic signals, a bandpass filter from 20 Hz to 450 Hz is first applied, followed by a notch filter at 50 Hz or 60 Hz to remove power frequency noise. Then, within a time window from 150 ms before the ball is struck to 50 ms after the ball is struck, the activation amplitude of the forearm extensor muscles, the activation amplitude of the forearm flexor muscles, and the timing difference of the activation of the extensor and flexor muscles are extracted.

[0105] In this embodiment, spatial clustering can be performed using the DBSCAN clustering method, with the cluster radius parameter ε set to 6 mm and the minimum sample size MinPts set to 8. For each cluster, the variance of electromyographic disorder within the cluster is calculated and mapped to a dimensionless interval of 0 to 2. When the group electromyographic inconsistency index of a certain region is close to 2, it indicates that the eccentric hitting in that region mainly originates from group-wide instability in force exertion; when the index is close to 0, it indicates that the eccentric hitting in that region is mostly due to random deviation or individual differences among individual players.

[0106] In this embodiment, the total mass of the clubhead to be optimized is limited to 198 g to 205 g, preferably 202 g; the clubface material is Ti-6Al-4V titanium alloy with a material density of 4.43. The elastic modulus is 110 GPa, and the Poisson's ratio is 0.34. The initial thickness of the bar face is set to 2.40 mm, with an allowable optimization range of 1.80 mm to 3.20 mm. To avoid generating unmanufacturable extremely thin regions, the thickness variation gradient between two adjacent mesh nodes does not exceed 0.20 mm / mm.

[0107] In the rigid body dynamics simulation engine, the rod head is treated as a rigid body, and the local contact area on the rod surface is treated as a contact area with an equivalent elastic response. The simulation platform can use ANSYS LS-DYNA, Abaqus Explicit, Altair Radioss, or other explicit dynamics solvers. The time step is set to... to s, preferably The simulation time for a single impact was set to 1.5 ms, with the primary contact time between the ball and the clubface ranging from 0.35 ms to 0.55 ms. The initial incident velocity of the golf ball was set to 38 m / s to 48 m / s, preferably 42 m / s; the ball mass was set to 45.93 g; and the ball radius was set to 21.35 mm.

[0108] During flight trajectory calculation, the air density can be set to... The ambient temperature was set to 20℃, relative humidity to 50%, and wind speed to 0 m / s. For outdoor testing environments, these parameters can be updated in real time using temperature, humidity, and wind speed sensors. The trajectory calculation output includes the lateral deviation distance of the impact point, total flight distance, sidespin velocity, ballistic yaw angle, and target area hit probability.

[0109] In this embodiment, the current clubface thickness distribution matrix, club weight distribution coordinates, and moment of inertia tensor are collectively used as the reinforcement learning state; the thickness adjustment of each grid node and the club weight distribution adjustment are used as actions. In a single action, the adjustment of each thickness grid node is limited to -0.03 mm to 0.03 mm, and the single-step adjustment of the club weight distribution is limited to -0.10 mm to 0.10 mm. The reinforcement learning algorithm can employ policy gradient algorithms such as PPO, DDPG, or SAC, with PPO being preferred to ensure training stability in the continuous action space.

[0110] In one specific implementation, the policy network adopted a three-layer fully connected neural network with 256 hidden units per layer and the ReLU activation function; the learning rate was set to... The discount factor is set to 0.99, and the batch size is set to 128. The maximum number of iterations is set to 1500. Iteration terminates when the rate of change of the penalty term is less than 0.5% for 30 consecutive iterations, or when the maximum number of iterations is reached.

[0111] Based on the three-dimensional topological model of the rod face output by reinforcement learning iteration, prototypes can be prepared through five-axis CNC machining, precision casting followed by local CNC subtractive machining, selective laser melting of metal, or precision milling after forging. In specific implementation, Ti-6Al-4V rod face prototypes can be prepared using EOS M290 or SLM 280 metal additive manufacturing equipment, with a layer thickness set to 30 μm to 60 μm, preferably 40 μm; after forming, stress relief heat treatment is performed at 650℃ to 750℃, with a holding time of 2 h to 3 h.

[0112] After the prototype is manufactured, a coordinate measuring machine, industrial CT scanner, or laser scanner can be used to detect the thickness distribution of the clubface and the weight distribution of the club. The thickness detection error should be controlled within ±0.03 mm, and the weight distribution detection error should be controlled within ±0.10 mm. If the detection results exceed the error range, the local area should be further refined or reprocessed until it meets the requirements of the simulation output model.

[0113] The following is a comparison between this solution and existing technologies:

[0114] Comparative Example 1 uses the traditional center-hit assumption for symmetrical clubface thickness design, without introducing the historical hit scatter matrix or the population electromyography inconsistency index.

[0115] Comparative Example 2 uses a historical shot distribution matrix, but does not introduce a population electromyographic inconsistency index; it is optimized only based on the landing distance and sidespin angular velocity.

[0116] This scheme simultaneously incorporates historical shot distribution matrix, population electromyography inconsistency index, eccentric collision simulation, and reinforcement learning iterative topology generation.

[0117] The following table compares the effects of this solution with those of Comparative Example 1 and Comparative Example 2. Figure 3 Here is the corresponding effect image:

[0118]

[0119] As shown in the table above, compared to Comparative Example 1, the average landing point deviation distance in Example 3 decreased from 13.8 m to 7.9 m, a reduction of approximately 42.8%; the root mean square velocity of the sidespin angular velocity decreased from 742 rpm to 486 rpm, a reduction of approximately 34.5%; and the target area hit rate increased from 54.6% to 76.8%. This demonstrates that the asymmetric topology generation method based on the electromyographic inconsistency index can more effectively improve the landing point deviation and sidespin yaw caused by eccentric hitting.

[0120] The following table compares the effects of this scheme and Comparative Example 3. Figure 4 Here is the corresponding effect image:

[0121] As shown in the table above, under the same reinforcement learning iteration strategy and the same historical impact data, after adopting dynamic thickness-stiffness mapping, the collision impulse prediction error decreased from 11.7% to 6.2%, the sidespin velocity prediction error decreased from 15.9% to 8.1%, and the average landing point prediction error decreased from 4.8 m to 2.6 m. This indicates that local curvature and material strain rate correction can improve the ability of rigid body dynamics simulation to represent high-speed eccentric impacts, making it easier to reproduce the expected effects of the topology generated by reinforcement learning in physical prototypes.

[0122] The above-disclosed content is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the description and drawings of the present invention are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops. The above units are merely examples, and those skilled in the art can adopt corresponding units according to actual needs when implementing this solution.

Claims

1. A method for generating golf club topology based on reinforcement learning, characterized in that, Includes the following steps: S1, Obtain historical shot data and biomechanical characteristics of the target player group, and construct a historical shot scatter matrix and the corresponding group electromyographic inconsistency index set; the shot data is the impact coordinates of the ball on the clubface; S2, In the rigid body mechanics simulation engine, based on the current iteration of the rod surface thickness distribution and club weight distribution, the scatter point matrix is ​​subjected to impact simulation and physical simulation results are obtained. Combined with biomechanical characteristics and physical simulation results, the deviation penalty term is calculated. S3, with the goal of minimizing the deviation penalty term, the strategy is updated using a reinforcement learning algorithm, and the clubface thickness distribution and club weight distribution are iterated continuously until the convergence condition is met, thereby autonomously generating an asymmetric and irregular intelligent clubface topology 3D model.

2. The method for generating golf club topology based on reinforcement learning according to claim 1, characterized in that, Step S1 also includes the following sub-steps: S11 acquires historical shot data of the target player group through high-speed cameras and shot monitoring radar, maps the impact coordinates of the ball on the clubface into a two-dimensional plane point set, and constructs a historical shot scatter matrix; S12, synchronously collect the surface electromyography (EMG) signals of the player group during historical ball hits, and extract their corresponding EMG features; S13, Spatial clustering is performed on the scatter point matrix, and the variance of electromyographic features within each cluster is normalized to calculate the population electromyographic inconsistency index corresponding to each scatter point in the matrix.

3. The method for generating golf club topology based on reinforcement learning according to claim 1, characterized in that, In S2, the impact simulation of the scattered point matrix and the physical simulation results are obtained by the following steps: S21, calculate the local contact stiffness coefficient of each mesh node on the club face based on the current iteration of the club face thickness distribution, and calculate the rotational inertia tensor and center of mass position of the club head rigid body based on the current iteration of the club weight distribution; S22, For the scattered points in the scattered point matrix, calculate the lever arm vector of the impact point relative to the center of mass of the club based on the impact coordinates corresponding to the scattered points and the corresponding club center of mass position, and calculate the impact impulse at the time of impact based on the local contact stiffness coefficient. S23, calculate the torque generated by the collision, and solve the angular velocity change of the clubhead around the center of mass of the club by combining the moment of inertia tensor. The side spin angular velocity generated by the clubface is obtained by decomposing the angular velocity change. S24: Obtain the collision reflection velocity based on the collision impulse, and calculate the golf ball's flight trajectory based on the collision reflection velocity and the side spin angular velocity, combined with the aerodynamic model, and calculate the geometric distance between the landing point and the preset target point.

4. The method for generating golf club topology based on reinforcement learning according to claim 1, characterized in that, In S3, the iterative process based on the deviation penalty term includes the following steps: S31. Name the combined matrix of the current clubface thickness distribution matrix and the club weight distribution matrix as the state matrix. The combined matrix of thickness adjustment matrix and weight adjustment matrix for each position of the cue stick is named the motion matrix. The negative deviation penalty term is defined as a reward signal; S32, Output the action matrix according to the current strategy. Applying to the rigid body mechanics simulation engine, the engine updates the state matrix as follows: And calculate the new based on step S2 Return reward signal ; S33, with the goal of maximizing the expected cumulative reward signal, the policy network is updated, and the system updates the action matrix based on the policy network. ; S34, Repeat the above steps until K consecutive generations When the change is less than a preset threshold, or when the maximum number of iterations is reached, the process terminates and outputs the finally converged asymmetric, irregular intelligent pole surface topology 3D model.

5. A golf club topology generation system based on the coupling of reinforcement learning and rigid body dynamics across all scenarios, applied to the golf club topology generation method based on reinforcement learning as described in claim 1, characterized in that... The system includes a data acquisition and processing module, a computing module, and a reinforcement learning iterative topology generation module; The data acquisition and processing module includes a historical scatter matrix construction unit and a biomechanical feature extraction unit. The historical scatter matrix construction unit is used to acquire historical shot data of a specific player group, map it to a set of coordinate points on the clubface, and form a historical shot scatter matrix. The biomechanical feature extraction unit is used to simultaneously acquire electromyographic signals of the specific player group during historical imperfect shots, extract muscle exertion disorder features, and calculate the group electromyographic inconsistency index.

6. The golf club topology derivation and generation system based on full-scene coupling of reinforcement learning and rigid body dynamics as described in claim 5, characterized in that, The calculation module includes a dynamic simulation execution unit and a coupling deviation penalty calculation unit; The dynamics simulation execution unit includes a parametric physics mapping substructure, an eccentric collision calculation substructure, a dynamic response solving substructure, and a trajectory solving substructure. The parametric physics mapping substructure is used to calculate the local contact stiffness coefficient based on the current shaft thickness distribution and the rotational inertia tensor based on the club weight distribution. The eccentric collision calculation substructure is used to calculate the impact arm and impact impulse based on the coordinates of the scattering points. The dynamic response solving substructure is used to calculate the torque and solve for the output side spin angular velocity. The trajectory solving substructure calculates the flight trajectory and outputs the landing point deviation distance. The coupling deviation penalty calculation unit is used to calculate the deviation penalty item based on various data.

7. The golf club topology derivation and generation system based on full-scene coupling of reinforcement learning and rigid body dynamics as described in claim 5, characterized in that, The reinforcement learning iterative topology generation module includes a state-action matrix acquisition unit and a policy update and output unit. The state-action space definition unit is used to generate action matrices and state matrices; the policy update and output unit is used to obtain reward signals, continuously update the policy network, and iteratively generate new pole face thickness and weight parameters until the deviation from the penalty term converges, and finally autonomously outputs an asymmetric and irregular intelligent pole face topology 3D model.

Images