Machine Learning Model Construction and Practice for Personalized Training Programs in Athletics Training

Athletics is an important program in sports competition, which plays an important role in improving the physical quality of athletes. With the development of competitive sports, the competition in track and field is becoming increasingly fierce, and the training requirements for athletes are getting higher and higher. Therefore, the use of machine learning models to develop optimization and control measures for the technical training process aims to improve the athletic level of track and field athletes and provide coaches with more scientific and systematic training methods to promote the sustainable development of track and field [1-2].

Athletics training should contain unified training and individualized training, which continuously improves athletes’ basic skills and physical fitness in unified training sessions, and strengthens athletes’ individual special skills in individualized training [3]. And machine learning technology can provide sports data to help develop personalized training plans for athletes. Specifically, the machine learning model has a strong data analysis capability, which can analyze the athlete’s data, including step frequency, stride length, heart rate, etc. The analysis of the data can be used to understand the athlete’s strengths and room for improvement, as well as the direction of optimization [4-6]. It can identify the athlete’s current weaknesses and focuses, and guide the athlete which aspects need to be improved or strengthened, so as to develop targeted training methods, and allow the athlete to break through the training goals in a short period of time [7-9]. In addition, it can not only help coaches set goals and stage plans based on athletes’ abilities and goals, but also adjust and optimize training plans based on real-time monitoring data, specific performance, etc., to achieve the improvement of athletes’ technical abilities in track and field [10-13]. Although machine learning technology can provide accurate sports information for coaches’ reference, it is also necessary to focus on manual intervention when applying this technology, still focusing on the guidance of professional coaches, supplemented by intelligent devices, and do not put the cart before the horse [14-16].

Machine learning techniques play an important role in the field of sports. Dijkhuis, T. B. et al. trained different machine learning algorithms using collected data on participants’ daily steps, which not only predicted physical activity during the day, but also provided timely interventions that offered personalized feedback on progress towards individual step goals [17]. Wang, J. et al. created an AI coaching system based on a computer vision approach that extracts the trajectories of individual human instances through deep visual tracking, estimates human postures through a human joint relationship model, and generates visual suggestions to correct the athlete’s posture with the aim of providing a personalized sports training experience for sports athletes [18]. Stetter, B. J. et al. combined wearable sensors with an artificial neural network to model the correlation between the signals captured by an inertial measurement unit carried by the participant and the knee force time series to estimate the joint response in a sports application, which was used to observe biomechanical metrics of structural loading on the knee joint [19]. Den Hartigh, R. J. et al. elaborated on the concept of resilience in sport and physical education, showing that machine learning is crucial in improving mental or physical issues such as athlete resilience, and that the technique is capable of detecting early warning signals of loss of resilience in the context of psychological and physiological changes in the athlete and for providing personalized feedback about the athlete’s resilience [20]. Gámez Díaz, R. et al. introduced the field of digital twin technology in sports, pointing out that machine learning and coaching technologies are the focus of the field, i.e., analyzing the digital twin coaching features in terms of concepts such as interactivity, privacy, and security, and also depicting the ideal digital twin ecosystem for team sport athlete tracking [21]. Fister, I. et al. used athletes’ exercise heart rate etc. measured by sports watches as a source of data, which can help coaches to increase the experience of sports training and understand the athletes’ athletic ability in detail, and assist coaches to plan sports training by generating personalized training plans on digital computers through bat algorithms [22]. Jing, R. et al. constructed an evaluation model applicable to track and field training methods, introduced the sparrow search algorithm to improve the generalization ability of the model, and used the chicken flock algorithm to optimize the athlete’s training combinations, training cycles and training intensities, and experiments have shown that the model facilitates the risk management of the athletes and personalized training, which helps to improve the athletes’ performance [23].

The article collects track and field movement data by installing sensors in important movement parts of athletes, summarizing and filtering the data collected by the sensors. The real-time capture of track and field athletic data was accomplished by using microelectromechanical technology to optimize the collected movements and uploading the data of the movements to a computer for processing. Then, after giving a description of the personalized track and field training program generation problem, the general structural framework of the ACO algorithm was designed, and the algorithmic strategies such as the structure of the solution memory table, the evaluation of fitness considering constraints, the calculation of solution weights, and the sampling of baseline solutions in the ACO algorithm were elaborated. The construction of personalized training model on track and field based on ant colony algorithm is completed, which optimizes the design idea of training program for track and field athletes and improves their training efficiency side by side.

2

Machine Learning in Athletic Training

2.1

Characterization of machine learning in sports training

While artificial intelligence usually refers to the ability of a machine to perform a task in the way that a human would want it to, machine learning specifically denotes a type of computer application that is used to process data and learn from it, and can be defined as an endeavour to study how experience can be utilized to improve the performance of the system itself by means of computation. In computer systems, “experience” is usually in the form of “data”. The computer is trained on the data to obtain a “model” algorithm, i.e., a learning algorithm. It is then provided with empirical data and can generate models based on this data. When faced with a new situation, the model will provide us with the appropriate judgment. In short, machine learning is a “learning algorithm” that automatically acquires knowledge from experience and makes personalized predictions about new situations based on the knowledge model.

2.2

Advances in machine learning for sports training

The aim of sports training is to apply science and technology to the athlete’s current athletic performance in an attempt to carefully design a daily personalized training strategy to improve athletic performance and reduce athletic injuries, so that optimal athletic performance occurs at the right time. Thus, the implementation process of sports training can be simplified into four stages: planning (including competition schedules and training phases for specific objectives, etc.). Realization (including the warm-up phase of training equipment and equipment cattle psychological assessment, monitoring the intensity of the load of training sessions, tactical training in group projects, biological movement ability training, etc.). Control (including video analysis of motor skills and motor anatomy, intelligent calculation and analysis of technical and tactical skills in group events, and giving qualitative and quantitative feedback on sports performance). Evaluation (including evaluation of single training (short-term) loads, generally expressed as Banisters Training Impulse Volume (TRIMP). The assessment of the total training load (long-term) is generally measured in terms of the target performance of the competition and the total planned and actually completed training volume.

Currently, there is a consensus on leading athletic training with cutting-edge technology to personalize and efficiently enhance athletic performance. As a matter of fact, traditional sports training activities often require the full cooperation of professional coaching staff, spending a lot of manpower, material and financial resources, and adopting a “one-size-fits-all” training method to meet the demand due to the lack of high-quality coaching resources, and the phenomenon of suffering from sports injuries due to over-training or incorrect exercise patterns is very common, which seriously impedes the health and sustainability of athletic sports. This has seriously hindered the healthy and sustainable development of competitive sports. According to UNESCO, artificial intelligence technology can promote personalized learning. In view of this, there is an urgent need to build an artificial intelligence training system around the design and implementation of “sports training load” to create a precise personalized training mode suitable for different individuals to improve quality and efficiency. On the one hand, it can greatly reduce the amount of trivial work (training) done by coaches and athletes, freeing up more time to engage in creative work. On the other hand, the new model of combining man and machine is expected to improve the overall efficiency of sports training and reduce the risk of sports injuries.

3

Athletics Motion Capture

In sports teaching through the motion capture design for track and field training process of action capture, based on the actual situation can be analyzed comprehensively, to determine the problems that exist in the actual training process. The motion capture process will collect the athlete’s information for statistical analysis, and the standard database for comparison, on the basis of which the irregularities in track and field movements can be obtained, and parameterization of the athlete’s movement process, quantitative processing of individual movements, so that it becomes an understandable data pattern for athletes. In the actual exercise process, a personalized training model based on machine learning is used to match the more standardized training program for the athlete, and the problems are adjusted accordingly.

3.1

Design of the estimation function for track and field sports movements

The human body has a complex structure, and sensors are installed in various moving parts of the human body to collect track and field data, which is summarized and filtered. Then use microelectromechanical technology to optimize the track and field movement, upload it to the computer for processing and three-dimensional simulation display, to design the working principle of a real-time track and field data capture system.

Filtering is the most direct and useful processing step to avoid noise interference. The track and field data filtering operation uses Kalman filtering, which is a linear function that recursively processes all the input and output data from the sensors to complete optimal estimation. The biggest advantage of Kalman filtering is that there is no need for data format conversion and frequency conversion during the filtering process, and there is no need to store the historical filtering results, which saves system space and capture time.

In Kalman filtering [24], if you want to get the real-time state estimate of the current track and field sports data, you need to obtain two important data, including the last track and field sports data estimate and the current track and field sports data prediction Therefore, in the initial filtering situation, you need to arrange the filtering order of the track and field sports data first. Using x_k to represent the real-time state estimate of the knd track and field sports data, the estimation error of x_k is replaced by the symbol P_k, and the function of x_k and P_i is defined as: (1) ${\begin{array}{l} x_{k} = F_{k} x_{k - 1} + B_{k} u_{k} \\ P_{k} = F_{k}^{T} F_{k} P_{k - 1} + Q_{k} \end{array}$ Where: F_k is the sensor sampling interval. $F_{k}^{T}$ is the transpose matrix of F_k. Q_i is the bias derivative of F_k. B_k is the sensor sampling margin. u_k is the covariance of the sensor sampling margin.

The basis of MEMS is mechanical kinematics, and the important hardware devices to which it can be applied are accelerometers and magnetic field measuring instruments. Athletics cannot be separated from mechanical kinematics, so it is reasonable to use MEMS to optimize the Kalman filter estimates. Assuming, ideally (i.e., no gravity and magnetic field interference) the human body’s resting action matrix is $C_{n}^{n}$ and its functional expression is: (2) $C_{b}^{m} = [\begin{matrix} \cos α & \sin α \sin β & \sin α \cos β \\ 0 & \cos β & - \sin β \\ - \sin β & \cos α \sin β & \cos α \cos β \end{matrix}]$ Where: α and β are the pitch and roll angles of the human body. The designed real-time track and field sports data capture design in the human body to install accelerometers and magnetic field measurement instruments, the human body at rest in the acceleration coordinate system and magnetic field coordinate system is overlap, but in the movement state, the human body by the gravitational acceleration and the magnetic field of the influence caused by the separation of the two coordinate systems, at this time the human body stationary action matrix, in the acceleration coordinate system and the magnetic field coordinate system of the projected vector, respectively: (3) $[\begin{matrix} a_{x}^{b} \\ a_{y}^{b} \\ a_{z}^{b} \end{matrix}] = [\begin{matrix} \cos α & 0 & - \sin β \\ \sin α \cos β & \cos β & \cos α \sin β \\ \sin α \cos β & - \sin β & \cos α \cos β \end{matrix}] [\begin{matrix} 0 \\ 0 \\ g \end{matrix}]$ (4) $[\begin{matrix} h_{s}^{b} \\ h_{y}^{b} \\ h_{z}^{b} \end{matrix}] = [\begin{matrix} \cos α & 0 & - \sin β \\ \sin α \cos β & \cos β & \cos α \sin β \\ \sin α \cos β & - \sin β & \cos α \cos β \end{matrix}] [\begin{matrix} H_{s} \\ H_{s} \\ 0 \end{matrix}]$ where: g is the acceleration of gravity; H_x is the projection vector between the magnetic field coordinate system and the ground plane, H_y is the projection component of the magnetic field coordinate system perpendicular to the ground plane, and H_x and H_y serve to correct for the magnetic field bias of the track and field movement; $β = \arctan \frac{H_{x}}{H_{y}}$ .

Using Eq. (3), Eq. (4) to optimize the value of x_i in Eq. (1), the optimization process is to see x_i as $C_{b}^{n}$ , converted into matrix form, and the results obtained can be used to describe the optimal track and field sports data real-time estimates.

3.2

Design of real-time track and field data capture

The human body has 206 movable joints, but the angle of joint movement varies between individuals. If you want to accurately simulate all 206 joints is not achievable, the requirements of various industries for real-time capture of track and field data have not reached such a high level of accuracy, so as long as the important joints that can show the movement of track and field sports can be modeled.

The estimation function of track and field movement cannot describe the movement of specific joints of the human body, and the estimation result should be solved to obtain the joint movement angle, and then input into the human body model to realize real-time capture.

Choose a sensor that collects data that is not empty, set it as sensor 1, and the sensor connected to sensor 1 will inevitably generate rotation data. Taking sensor 1 as the resting point, the motion angle data of the neighboring sensor (set as sensor 2) relative to the resting point is obtained and labeled in the human body’s important motion joint model. Take sensor 2 as the resting point again and repeat the above steps until the motion data of all joints in the model is successfully filled.

When all the athletic movement data is solved successfully at once, the important movement joint model of the human body is output for real-time 3D virtual display. In order to ensure the real-time capture ability of the system, the bvh format file is used for the 3D virtual display.

4

Personalized training model for track and field based on ant colony algorithm

4.1

Problem description for the generation of personalized training programs in athletics

The training solution space (solution space Ω) optimization space can be defined as the time series shown in equation (5).

(5)

Ω = {X | X = < (V (I), A (I)), ..., (V (t), A (t)), ..., (V (l), A (l)) >}

$$\Omega = \{ \,X|\,X = \left\langle {(V(I),A(I)),...,(V(t),A(t)),...,(V(l),A(l))} \right\rangle \} $$

where V(t) and A(t) represent the speed and slope of the tth period, respectively, and t is the total number of periods. The value of training speed is taken from interval [V_min,V_max], which is a continuous variable.

Safe and effective heart rate interval (heart rate constraint) The safe and effective heart rate interval for athletes’ track and field training is defined as equation (6). HR_max is the maximum heart rate and HR_max = 220-age.

(6)

Safe and effective heart rate range = [0.64 H R_{\max}, 0.74 H R_{\max}]

Solution X Number of times the heart rate falls in a non-safe effective heart rate interval (optimization objective) The number of times the heart rate falls in a non-safe effective heart rate interval is defined as the function shown in equation (7) f(X).

(7)

f (X) = l - \sum_{t = 1}^{l} h (t)

(8)

h (t) = {\begin{array}{l} 1, H \hat{R} (t) \in [0.64 H R_{\max}, 0.74 H R_{\max}] \\ 0, H \hat{R} (t) \in [0.64 H R_{\max}, 0.74 H R_{\max}] \end{array}

Athletic training program generation for athletes in track and field can be described as a hybrid coded optimization problem with constraints [25], where the optimal solution X* satisfying Eq. (9) is searched for, subject to Eq. (8).

(9)

X^{*} = \max_{X \in Ω} f (X)

4.2

Generation of training scheme based on ant colony optimization algorithm

4.2.1

Structure of a mixed-encoding unmemory table

The ant colony optimization algorithm first constructs a solution memory table (pheromone model) based on mixed variables [26], the structure of which is shown in Fig. 1. The solution memory table contains k complete solution of the optimization problem {X₁,X₂,…,X_k}. The search path of the algorithm is to update the continuous variables first and then the discrete variables sequentially. Each solution consists of r continuous variables and d discrete variables (Paper r = d = l). Where t represents the time period.

For each iteration of the algorithm, m ants first sample m baseline solutions and guide the subsequent ants to randomly construct m new solutions, which are first stored in the solution memory table, and then the k + m solutions in the memory table are sorted according to the fitness values, removing the worst m solutions and retaining the best k , which results in the dynamic updating of the solution memory table in a positive-feedback manner, and the candidate solutions in the solution memory table are gradually optimized as a result.

4.2.2

Adaptation assessment considering constraints

The optimization objective for the generation of the optimal track and field training program is to minimize the value of the objective function, and the fitness is defined as in equation (10).

(10)

F (X_{i}) = R_{f} + R_{D}

This equation combines the objective function and the constraint violation degree penalty function, as well as introduces a comprehensive ranking mechanism. R_f, R_D denotes the ranking of the objective function and constraint violation degree penalty function from smallest to largest, respectively. The design of this adaptation degree mainly satisfies the following three principles:

1) $H \hat{R} (t)$ The fewer times it falls outside the interval [0.64HR_max,0.74HR_max], the better.

2) The closer $H \hat{R} (t)$ is to the boundary of the interval, the better the value of its fitness.

3) From the safety point of view, the fitness that $H \hat{R} (t)$ is smaller than the lower boundary is better than the fitness that $H \hat{R} (t)$ is larger than the upper boundary.

(11)

D (X_{i}) = \sum_{t = 1}^{l} g (t)

$$D\,({X_i}) = \mathop \sum \limits_{t = 1}^l g(t)$$

(12)

g (t) = {\begin{matrix} \frac{0.6 H R_{\max}}{H \hat{R} (t)} \times 0.8 & H \hat{R} (t) < 0.64 H R_{\max} \\ \frac{H \hat{R} (t)}{0.8 H R_{\max}} & H \hat{R} (t) > 0.74 H R_{\max} \\ 0 & o t h e r s \end{matrix}

4.2.3

Calculation of solution weights and sampling of benchmark solutions

The solutions in the memory table are ranked from smallest to largest according to the size of the fitness F(X_i) to get the ranking of each solution noted as rank(X_i), and then the weight of solution X_i is calculated ω(X_i). ω(X_i) is a Gaussian probability density function about the ranking rank(X_i), and the defining equation is as in equation (13).

(13)

ω (X_{i}) = \frac{1}{q k \sqrt{2 π}} e^{\frac{- {(r a n k (X_{i}) - 1)}^{2}}{2 q^{2} k^{2}}}

Eq. q is a coefficient that regulates the weights. After obtaining the weights and according to the probability defined in Eq. (14), m benchmark solutions are sampled (using the roulette method). The higher the ranking rank(X_i) the more attractive the solution is and the higher the probability of being sampled as a benchmark solution.

(14)

P (X_{i}) = \frac{ω (X_{i})}{\sum_{b = 1}^{k} ω (X_{b})}

4.2.4

Continuous variable update strategy ACOMV-V

In the baseline solution X_i, for each continuous variable $V_{i}^{j} (t)$ where 1≤j≤r , a Gaussian probability density function with mean μ and standard deviation σ is updated in the neighborhood of $V_{i}^{j} (t)$ to obtain a new velocity $V_{k + 1}^{j} (t)$ . It is defined as in equation (15).

(15)

g (x, μ, σ) = \frac{1}{σ \sqrt{2 π}} e^{\frac{- {(x - μ)}^{2}}{2 σ^{2}}}

where the mean is $μ_{i}^{j} = V_{i}^{j} (t)$ . The standard deviation $σ_{i}^{j}$ is jointly determined by the value of the j rd continuous variable for all solutions in the memory table, defined as in Eqs. (16) and (17).

(16)

σ_{i}^{j} = ε \sum_{a = 1, a \neq i}^{k} \frac{| V_{a}^{j} (t) \times P (X_{a}) - V_{i}^{j} (t) \times P (X_{i}) |}{k - 1}

(17)

\begin{array}{l} P (X_{a}) = \frac{1}{1 + F (X_{a})} \\ P (X_{i}) = \frac{1}{1 + F (X_{i})} \end{array}

This equation determines the standard deviation for each continuous variable, combined with how well each solution X_i fits. Since the data are Gaussian probability distributions, a large standard deviation has a more dispersed data distribution and a small standard deviation has a more concentrated data distribution. Scaling the data according to the fitness can reduce the perturbation of the worse solution for the better solution, and increase its standard deviation for the worse solution to make it more probable to produce a better solution.

Finally, if an illegal encoding is generated during the updating process (the resulting new velocity value is not in the desirable velocity interval [V_min,V_max]), the symmetry point is taken with symmetry axis V_suls when the new solution is smaller than the minimum velocity V_sub , and symmetry point is taken with symmetry axis V_max when the new solution is larger than the maximum velocity V_max , until the new solution is legal.

4.2.5

Discrete variable update strategy ACOMV-A

In the benchmark solution X_i, each discrete variable $A_{i}^{j} (t)$ , where 1≤j≤d , is updated using a continuous relaxation method. The method of successive relaxation used in the literature is to order the discrete variables and do the update operation directly on the indexes of the discrete variables instead of the variables themselves. The advantage of this method is that it can handle not only variables of numerical type, such as the set of integers {1, 2, 3, …, 10}, but also non-numeric variables such as the set {high, medium, low}.

Since the discrete variable type in this paper is the numerical type, in order to reduce the complexity of the optimization algorithm. The continuous relaxation method in this paper directly treats the discrete values as continuous values first, and its processing is the same as the update strategy of continuous variables, which is directly called ACO_MV-V. Then the result obtained from the update is compared with the discrete points in the optional set, and the discrete point closest to that continuous value is used as the updated $A_{k + 1}^{j} (t)$ .

5

Comparison of physical fitness test before and after the experiment

A total of 60 students (30 men and women each) from two classes with similar physical fitness test scores in track and field majors at a sports college were selected. The training program of the control group was carried out in accordance with the track and field training content, and the experimental group was carried out in accordance with the personalized training program of the personalized training model based on machine learning proposed in this paper. The test items include boys: 1000 meters, shot put, and standing long jump. Female students are participating in 800 meters, shot put, and standing long jump.

The physical fitness measurement data of the students are shown in Table 1, and the P-value of each test index data of the boys and girls in the control and experimental groups is greater than 0.05, with no significant difference. It indicates that there is no significant difference between the physical abilities of the two groups of students before conducting the experiment. The basis of the experiment is basically the same, and the grouping is reasonable. Note: (P>0.05 means there is no significant difference, 0.01<*P<0.05 means there is a significant difference, **P<0.01 means there is a very significant difference).

Table 1.

Shows the two groups of students’ physical test indicators compared

Test index	Control group (n= 30)	The experimental group (n= 30)	t	P
Height (male)	1.67±8.43	1.66±8.44	0.487	0.646
Weight (male)	53.25±11.23	56.02±12.24	-1.368	0.178
BMI (male)	19.15±3.57	20.11±3.45	-1.445	0.168
1000 meters (male)	244.58±32	243.57±28	0.522	0.626
Lead ball (male)	20.15±5.46	20.52±5.66	-1.445	0.156
Set the jump (male)	191.66±25.43	194.56±24.78	-1.323	0.179
Height (female)	156.45±5.63	157.13±5.47	-0.945	0.319
Weight (female)	48.56±12.43	49.27±11.89	-0.313	0.743
BMI (female)	20.54±3.69	19.99±3.49	0.545	0.824
800 meters (female)	227.56±30.51	225.98±29.87	0.255	0.196
Lead ball (female)	18.46±7.11	18.67±7.13	-1.287	0.218
To jump far (female)	163.77±17.87	164.29±17.18	-0.377	0.709

After the experiment, Table 2 displays the performance of students in track and field sports events in both groups. From the analysis of the results in Table 2, it can be seen that the experimental group, after 8 weeks of personalized training based on the machine learning model proposed in this paper, the track and field sports events such as 1000 m, shot put, 800 m, and standing long jump have a more significant difference compared with the control group, and have a larger effect size. It can be concluded that personalized training based on a machine learning model is effective in improving students’ performance in speed, endurance, and strength events. When each variable such as instructional design is certain, the use of personalized training program design tools to adjust the training of students can be of great help to the improvement of students’ performance in track and field events. Each student has a different physique, and the same intensity of training is often unable to meet the actual needs of all students at the same time, and even more unable to achieve the expected training goals. Therefore, in the actual sports training process, the application of personalized training plan based on machine learning model, according to the actual situation of different students, to develop specific training intensity, in order to improve the effect of track and field training, but also to avoid the occurrence of sports injuries due to excessive exercise intensity.

Table 2.

The two groups of students were compared in physical fitness test

Test index	Control group (n= 30)	The experimental group (n= 30)	t	P	Freedom degree	The effect is(cohen’sd)
1000 meters (male)/s	241.27±36.44	224.25±18.26	2.915	0.006	32	0.823
Lead ball (male)/m	21.58±6.45	25.64±6.47	-2.954	0.008	32	0.811
Set the jump (male) /cm	194.47±25.79	211.15±23.34	-2.788	0.007	32	0.758
800 meters (female) /s	223.44±25.460	207.41±21.43	2.784	0.007	24	0.817
Lead ball (female) /m	19.33±7.43	24.21±4.45	-2.726	0.009	24	0.779
To jump far (female) /cm	164.45±17.67	175.44±14.64	-2.934	0.008	24	0.789

Table 3 shows the comparison of physical fitness test indexes before and after the experiment of the students in the experimental group, and the analysis of the results in Table 3 shows that the students in the experimental group have a substantial increase in the indexes of various test items, especially the improvement of the performance of the shot put and the vertical jump project is very obvious, with a large effect size, the effect size of the boys’ shot put and the vertical jump project is 1.209 and 1.116, respectively, and that of the girls’ shot put and the vertical jump project is 1.45 and 1.209, respectively. The implementation of targeted and personalized physical training programs had a more significant effect on students’ speed and endurance.

Table 3.

The test index was compared with the physical test index

Test index	Preexperiment	After the experiment	t	P	Freedom degree	The effect is (cohen’sd)
1000 meters (male)/s	243.57±28	224.25±18.26	2.915	0.005	16	0.913
Lead ball (male)/m	20.52±5.66	25.64±6.47	-2.954	0.000	16	1.209
Set the jump (male) /cm	194.56±24.78	211.15±23.34	-2.788	0.000	16	1.116
800 meters (female) /s	225.98±29.87	207.41±21.43	2.784	0.000	13	0.923
Lead ball (female) /m	18.67±7.13	24.21±4.45	-2.726	0.000	13	1.45
To jump far (female) /cm	164.29±17.18	175.44±14.64	-2.934	0.003	13	1.299

Table 4 shows the comparison of physical fitness test indexes before and after the experiment of the control group students, as can be seen from Table 4, after 8 weeks of conventional training, the average scores of all test indexes of the students in the control group have been improved to different degrees, the effect size is small, and the largest effect size is only 0.241. It can be seen that conventional teaching training can also improve students’ physical fitness. But using conventional training to improve, it may take longer to achieve the desired effect.

Table 4.

The control of the control group was compared

Test index	Preexperiment	After the experiment	t	P	Freedom degree	The effect is (cohen’sd)
1000 meters (male)/s	244.58±32	241.27±36.44	2.904	0.006	15	0.108
Lead ball (male)/m	20.15±5.46	21.58±6.45	-6.975	0.006	15	0.027
Set the jump (male) /cm	191.66±25.43	194.47±25.79	-3.045	0.005	15	0.095
800 meters (female) /s	227.56±30.51	223.44±25.460	3.356	0.005	12	0.121
Lead ball (female) /m	18.46±7.11	19.33±7.43	-5.457	0.006	12	0.241
To jump far (female) /cm	163.77±17.87	164.45±17.67	-2.056	0.004	12	0.081

Table 5 shows the increase in the performance of track and field test items of the two groups, after 8 weeks of teaching, both the experimental group and the control group’s physical fitness have been improved, but the experimental group, through the development of personalized training, the performance of the experimental group has been improved to a greater extent, the reason is that the training process carried out through the personalized training program is more accurate and more in line with their own needs, the fastest increase in the performance of the girl’s shot put program, which is improved by 29.67%, but the control group The girls’ performance in the shot put event improved the fastest, by 29.67%, but the control group’s performance in the shot put event improved by only 4.71%. This shows that the personalized track and field training program can enable teachers to understand the actual situation of students more accurately and make appropriate adjustments to their training programs according to the personalized training model, which can help students achieve the best training effect to a large extent.

Table 5.

The comparison of test index growth before and after experiment

Test index	Control group (n= 30)	The experimental group (n= 30)
Test index	Increase/%	Increase/%
1000 meters (male)/s	1.35	7.9
Lead ball (male)/m	7.1	24.95
Set the jump (male) /cm	1.47	8.53
800 meters (female) /s	1.81	8.22
Lead ball (female) /m	4.71	29.67
To jump far (female) /cm	0.42	6.79

6

Aptitude enhancement tests for different categories

In order to verify and improve the personalized recommendation system designed in this research, an effectiveness experiment was conducted for the recommendation system, and the results of the experiment were analyzed and summarized.

The experimental group trained the training program recommended by the personalized recommendation system every day, and the control group trained the training program recommended by the teacher every day, and the daily training volume of each person was the same. At the end of the experimental task, the same satisfaction questionnaire was distributed to 60 experimental subjects. A total of 60 copies were distributed, 60 copies were recovered, and 60 valid questionnaires were distributed.

After 10 days of training, the training records of the relevant users were exported from the database and their scores were standardized. This experiment needs to analyze the competence of students in the training field in all aspects, as well as the overall development of their competence. The specific calculation rules are as follows: the score of the experimental subjects under a certain field is the arithmetic average of the scores of the track and field events under that field in which they trained, and the score of each group under a certain field is the arithmetic average of the scores of so many people in that group.

The overall score of the subjects was the arithmetic mean of the scores under each field, and the overall score of each group was the arithmetic mean of the scores of so many people in that group. The statistical results of the ten-day data of the students’ ability in each area are shown in Figure 2, which represents the change of the students’ ability in each area from the first day to the tenth day along the Y-axis from bottom to top, respectively, from the figure can be visualized to see that the experimental group of students’ ability in the area of speed category was improved from 0.55 to 0.95 on the tenth day, and the control group of students’ ability scores in the area of speed category was improved from 0.56 to 0.79 after the ten-day training, which is much lower than the experimental group’s ability score. 0.79, which is much lower than the degree of improvement of the students in the experimental group. In addition, students in the experimental group improved faster than those in the control group in the areas of bouncing, strength, and flexibility, indicating that the personalized training model proposed in this paper can provide students with more appropriate training programs, and has a more excellent effect on the improvement of students’ abilities in the areas of speed, bouncing, strength, and flexibility, which are closely related to the performance of the athletic events, and shows that the personalized recommender system proposed in this paper can provide students with more appropriate training programs. It shows that the personalized recommendation system proposed in this paper can improve the efficiency of users’ track and field training.

Figure 3 shows the overall ability scores of the two groups, from which it can be seen that the upward trend of the overall score of the experimental group is much higher than that of the control group, and the final overall score of the experimental group is 0.12 points higher than that of the control group. It shows that the personalized training recommendation system proposed in this paper can better combine the user’s own ability level to make recommendations, and at the same time further improve the user’s overall athletics athletic ability.

7

Conclusion

In the training of athletics events in competitive sports, the personalized training model based on machine learning provides students with a personalized training plan, which helps them to improve their motor skills, motor performance and motor level more efficiently.

1) The experimental group, through the formulation of personalized training, improved their performance in long-distance running, shot put, and standing long jump to a much greater extent than the control group, and the experimental group improved most significantly in the performance of the girls’ shot put category, whose performance was improved by 29.67%, which was 24.96% more than that of the control group. It shows that the personalized training program developed by the machine learning-based personalized training model is more accurate and more suitable for the different needs of each student, which can help students achieve the best track and field training effect to a greater extent.

2) On the tenth day of the experimental group, the students’ ability in the field of speed increased to 0.4 points, while the students’ ability in the field of speed in the control group only increased to 0.23 points after ten days of training, which is much lower than the improvement of the experimental group’s ability in the field of speed. In addition, the ability of the students in the control group in the areas of bouncing, strength, flexibility and the overall athletic ability were all improved to a lower degree than that of the experimental group, which indicates that the training program provided by the personalized training model based on machine learning can improve the user’s athletic ability in all aspects of track and field training and that it is more efficient, and that the personalized training model based on machine learning can be applied in the process of track and field training practice.

Idioma:: Inglés

Calendario de la edición:: 1 veces al año
Temas de la revista:: Ciencias de la vida, Ciencias de la vida, otros, Matemáticas, Matemáticas aplicadas, Matemáticas generales, Física, Física, otros

RSS Feed de revista

Machine Learning Model Construction and Practice for Personalized Training Programs in Athletics Training

Xi Chen

Yaosheng Zhang

Ruibin Jia

Yuquan Shi

Publicado en línea: 17 mar 2025

Recibido: 03 nov 2024

Aceptado: 11 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0286

Palabras claveAnt colony algorithm, Kalman filter, MEMS technology, Personalized training program

© 2025 Xi Chen et al., published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Palabras clave
Ant colony algorithm, Kalman filter, MEMS technology, Personalized training program