阅读说明：本技术 一种用户骑行能力分析方法 (A kind of user rides capability analysis method ) 是由 孔繁斌 于鉴 李珊 于 2019-07-08 设计创作，主要内容包括：一种用户骑行能力分析方法：对平台上的用户历史数据进行收集,筛选出满足条件的用户历史数据,所述用户历史数据包括：用户历史基础数据以及反映用户历史骑行能力的FTP值,构建FTP自动估测模型的样本数据集；利用反向传播神经网络,对所述样本数据集进行反复的训练迭代,得到FTP预测模型；根据所述当前用户在所述平台上输入的基础数据信息,基于所述FTP预测模型获得所述当前用户的FTP值。本发明应用神经网络智能算法进行机器学习训练,为大量低意愿进行FTP测试的用户提供高准确性的FTP估测,并基于预测得到的FTP值获得用户的运动能力等级,以及获得用户的运动能力类型,然后根据用户的需求、运动能力等级和运动能力类型为用户推荐训练课程。(A kind of user rides capability analysis method: being collected to the user's history data on platform, filter out the user's history data of the condition of satisfaction, the FTP value of ability that the user's history data include: user's history basic data and reflection user's history is ridden, constructs the sample data set of the automatic estimation models of FTP；Using reverse transmittance nerve network, training iteration repeatedly is carried out to the sample data set, obtains FTP prediction model；The basic data information inputted on the platform according to the active user obtains the FTP value of the active user based on the FTP prediction model.Present invention application neural network intelligent algorithm carries out machine learning training, the FTP estimation of high accuracy is provided for the user that a large amount of low wishes carry out FTP test, and the FTP value obtained based on prediction obtains the locomitivity grade of user, and obtain the locomitivity type of user, then according to the demand of user, locomitivity grade and locomitivity type be that user recommends training course.)

A kind of capability analysis method 1. user rides, it is characterised in that:

A: building sample data set is collected the user's history data on platform, filters out the user's history of the condition of satisfaction Data, the user's history data include: user's history basic data and reflection user's history the is ridden FTP value of ability, structure The sample data set of the automatic estimation models of FTP is built, wherein the basic data type N kind, N is integer, and N > 1；

B: building FTP prediction model carries out the sample data set anti-using reverse transmittance nerve network (BP neural network) Multiple training iteration obtains the FTP prediction model；

C: predicting the FTP value of active user, and the basic data information inputted on the platform according to the active user is based on The FTP prediction model obtains the FTP value of the active user.

2. analysis method according to claim 1, further comprises, the basic data information includes the weight letter of user Breath, according to the FTP value of the active user and the weight information of the active user, calculates the function of the active user Ride ability rating of the body than function body ratio=FTP/weight, according to function body than obtaining the active user.

3. analysis method according to claim 1, further comprises, the basic data information includes the weight letter of user Breath, grab active user described in database for a period of time in maximum 5 seconds, 1 minute, 5 minutes mean powers in motion recording, In conjunction with the FTP value and weight information of the active user, the power morphological data [MMP of the active user is formed_5s/ weight,MMP_1min/weight,MMP_5min/weight,FTP/weight]；Sentence according to active user's power morphological data Determine the locomitivity type of user.

4. analysis method according to claim 1, further comprises, according to the demand of the active user, locomitivity Grade and locomitivity type recommend targetedly drill program or training course to the active user.

5. analysis method according to claim 1, the basic data of user further comprises age, gender, height, body At least one of weight, sports level.

6. analysis method according to claim 1, the step a building sample data set further comprises:

A1 data collection: all user's history data in collecting platform the preceding paragraph time；

A2 data screening: the sample that user information is system default value is weeded out；Reject sluggish user's sample on platform； It rejects user information often to change, and maximum data value and minimum data value differ by more than the data of threshold value T, by screening, obtain To the sample data set.

7. analysis method according to claim 1, the step b building FTP prediction model further comprises:

B1: training set (X, Y) is input in BP neural network structure, determines that input layer number is equal in network by initialization Node in hidden layer l is arranged in the basic data type N, and output-index is the FTP of user, determines the number of nodes 1 of output layer；

Initialize the connection weight w between input layer, hidden layer and output layer neuron^ij,w^j, initialize hidden layer threshold value a^j, defeated Layer threshold value b out gives learning rate η and neuron activation functionsTarget error ε, the maximum times M of iteration are set；

B2: hidden layer output calculates, the output of j-th of node of hidden layer are as follows:

X=[x₁,x₂...x_i...x_N] it is input sample,It is activation primitive, wⁱJ is input layer and hidden layer The weight connected on different neurons, a^jFor the threshold value in hidden layer neuron, 1≤i < N, 1≤j≤l；

B3: the calculating of output layer exports the output of node layer are as follows:

The output of output layer neuron uses linear convergent rate, and neuron does not use activation primitive, H_jIt is the defeated of hidden layer neuron Out, w^jFor the weight that hidden layer is connect with output layer neuron, b is the weight of output layer neuron.

B4: error calculation, the prediction error E of network:

E=Y-O

Wherein Y indicates the true value of sample X, and O is the output valve of output layer neuron；

B5: judging whether iteration terminates, including two criterions:

B51: the prediction error of grid:

E≤ε

If so, it then terminates, retains current connection weight w^ij,w^j, hidden layer threshold value a^j, output layer threshold value b；If not, into The next judgement b52 of row.

B52: whether the number of iterations reaches maximum:

p≥M

If so, it then terminates, retains current connection weight w^ij,w^j, hidden layer threshold value a^j, output layer threshold value b；If not, more The new connection weight w^ij,w^j, hidden layer threshold value a^j, output layer threshold value b, return step b2.

8. analysis method according to claim 7, the update connection weight w^ij,w^jFurther comprise,

Network connection weight ω is updated according to neural network forecast error eⁱ _j,ω_j ^k

Right value update of the input layer to hidden layer are as follows:

w_ij=w_ij+ηH_j(1-H_j)x_iw_je

I=1,2,3j=1,2, l

Right value update of the hidden layer to output layer are as follows:

w_j=w_j+ηH_jE j=1,2, l

η is learning rate, H in formula_jIt is the output of hidden layer neuron, e is the error of true value and output valve.

9. analysis method according to claim 7, the update hidden layer threshold value a^j, output layer threshold value b further comprises, Network node threshold value a is updated according to neural network forecast error e^j, b.

The threshold value of input layer to hidden layer is updated to a^j；The threshold value of hidden layer to output layer is updated to b

a_j=a_j+ηew_jH_j(1-H_j) j=1,2, l

B=b+e

10. analysis method according to claim 2, the ability of riding of the user is divided into 4 grades.

11. according to the method described in claim 3, the locomitivity type includes spurt type, endurance type, chases hand and synthesis Type.

12. according to the method described in claim 4, user can customize and select the special training plan on platform.

Technical field

The present invention relates to locomitivity analysis field, especially a kind of user rides capability analysis method.

Background technique

With the increasingly raising of people's living standard, body-building has become essential part in daily life.Strong In the project of body, rides and attracted large quantities of fans to be added with its unique glamour.After riding vehicle, inevitable problem is just What the level for being me could how oneself constantly be promoted by performing physical exercise

For various puzzlements when riding, the scientific training ridden is suggested.That is for jockey, their energy Power needs to quantify, and then according to own situation, is targetedly trained.Because power can directly reflect user's output Strength, so power data enters everybody visual field.But the power data of a pile is faced, need a unified standard to comment These power datas of valence, allow them to play the value of oneself.

In current bicycle system, FTP (Functional Threshold Power, functional threshold power) is React one and its important index of driver's strength.FTP refers to jockey under the power currently ridden, the lactic acid that body generates The limiting value of body acid discharge is reached.Dr.Andrew Coggan is proposed: with the intensity of personal timing, being ridden 40 kilometers or 1 small When, the mean power obtained is FTP, this is the gold standard of FTP test.It is envisioned that if make every effort ride it is one small When or so, high-intensitive and high pressure is born to body bring.Most of jockey can test FTP to generate and retreat, and lead to function Rate training is difficult to carry out.

So a kind of it is proposed that method of intelligent estimation user FTP.Using user basic data for example the age, gender, Height, weight, sports level (including: that body is not poor, trained, general, medium, good, outstanding and outstanding) etc., carry out approximate estimation The FTP of user.This method application neural network intelligent algorithm carries out machine learning training, carries out FTP test for a large amount of low wishes User provide high accuracy FTP estimation.

User can carry out power training according to the FTP value estimated.User is allowed more to understand the direction of oneself training. It is for statistical analysis to the power data of riding of user, the advantage and disadvantage of user are found out using power data.User according to itself Situation carries out special training, strengthens advantage, weakens disadvantage.

Summary of the invention

The object of the invention is mainly: (1) user can be obtained by the FTP value of oneself without difficult arduous FTP test； (2) the FTP value obtained based on prediction obtains the locomitivity grade of user, and obtains the locomitivity type of user, then According to the demand of user, locomitivity grade and locomitivity type are that user recommends training course.

To reach above-mentioned technical purpose, technical solution of the present invention provides a kind of user and rides capability analysis method, comprising:

A: building sample data set is collected the user's history data on platform, filters out the user of the condition of satisfaction Historical data, the user's history data include: user's history basic data and reflection user's history the is ridden FTP of ability Value constructs the sample data set of the automatic estimation models of FTP, wherein the basic data type N kind, N is integer, and N > 1；

B: building FTP prediction model, using reverse transmittance nerve network (BP neural network), to the sample data set into The training iteration of row repeatedly, obtains the FTP prediction model；

C: predicting the FTP value of active user, the basic data information inputted on the platform according to the active user, The FTP value of the active user is obtained based on FTP prediction model.

According to the FTP value of the active user and the weight information of user, the function body of the active user is calculated Ability rating of riding than function body ratio=FTP/weight, according to function body than obtaining the active user.

Grab in database user for a period of time in maximum 5 seconds, 1 minute, 5 minutes mean powers, knot in motion recording Close the FTP value user of the active user and weight information, form the power morphological data of the active user

[MMP_5s/ weight, MMP_1min/ weight, MMP_5min/ weight, FTP/weight].According to the active user The locomitivity type of power morphological data judgement user (spurt type, chases hand, is comprehensive endurance type).

According to the demand of user, locomitivity grade and/or kinds of athletic capability class are that user recommends targetedly to instruct Practice plan or training course.

Detailed description of the invention

Fig. 1 predicts active user FTP value flow chart

Fig. 2 BP neural network structure chart

The training course of Fig. 3 user is distributed

Fig. 4 the method for the present invention test flow chart

The actual value of Fig. 5 FTP test set data is compared with predicted value

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that the described embodiment is only a part of the embodiment of the present invention, instead of all the embodiments.Based on this Embodiment in invention, every other reality obtained by those of ordinary skill in the art without making creative efforts Example is applied, shall fall within the protection scope of the present invention.

The present invention is collected the user data of (such as Onelap platform) on platform, constructs the FTP prediction mould of user Type.Then the FTP predicted value of active user is obtained using the FTP prediction model built according to the own situation of user.Its In predict active user FTP value as shown in Figure 1 the following steps are included:

A: building sample data set is collected the user's history data on platform, filters out the user of the condition of satisfaction Historical data, the user's history data include: user's history basic data and reflection user's history the is ridden FTP of ability Value constructs the sample data set of the automatic estimation models of FTP, wherein the basic data type N kind, N is integer, and N > 1.

Optional building sample data set includes: a1 data collection: collecting platform the preceding paragraph time all user's history numbers According to；A2 data screening: the sample that user information is system default value is weeded out；Reject sluggish user's sample on platform； It rejects user information often to change, and maximum data value and minimum data value differ by more than the data of threshold value T, by screening, obtain To sample data set, optional user information, which often changes, refers to that the average time that user information changes is less than certain threshold value, or uses The maximum time difference that family motion information changes is less than certain threshold value.

B: building FTP prediction model, using reverse transmittance nerve network (BP neural network), to the sample data set into The training iteration of row repeatedly, obtains the FTP prediction model.

Optionally, the FTP prediction model further comprises:

B1: training set (X, Y) is input in BP neural network structure, determines input layer number in network by initialization Equal to the basic data type N, node in hidden layer l is set, output-index is the FTP of user, determines the node of output layer Number 1；Figure (2) is to work as N=3, l=5, neural network structure figure when basic data includes age, height, weight.

Initialize the connection weight w between input layer, hidden layer and output layer neuron_ij,w_j, initialize hidden layer threshold value a_j, output layer threshold value b gives learning rate η and neuron activation functionsTarget error ε, the maximum times of iteration are set M。

B2: hidden layer output calculates, the output of j-th of node of hidden layer are as follows:

X=[x₁,x₂...x_i...x_N] be input sample, optionally, basic data include the age, height, weight, gender, At least one of sports level,It is activation primitive, w_ijTo be connected on input layer and hidden layer difference neuron Weight, a_jFor the threshold value in hidden layer neuron, 1≤i < N, 1≤j≤l.

B3: the calculating of output layer.Export the output of node layer are as follows:

The output of output layer neuron uses linear convergent rate, and neuron does not use activation primitive.H_jIt is hidden layer nerve The output of member, w_jFor the weight that hidden layer is connect with output layer neuron, b is the weight of output layer neuron.

B4: error calculation.The prediction error E of network:

E=Y-O

Wherein Y indicates the true value of sample X, and O is the output valve of output layer neuron.

B5: judging whether iteration terminates, including two criterions:

B51: the prediction error of grid:

E≤ε

If so, it then terminates, retains current connection weight w_ij,w_j, hidden layer threshold value a_j, output layer threshold value b；

If not, carry out next judgement b52.

B52: whether the number of iterations reaches maximum:

p≥M

If so, it then terminates, retains current connection weight w_ij,w_j, hidden layer threshold value a_j, output layer threshold value b；

If not, update connection weight w_ij,w_j, hidden layer threshold value a_j, output layer threshold value b, return step b2.

Optionally, the update connection weight w_ij,w_jFurther comprise,

Network connection weight ω is updated according to neural network forecast error e_ij,ω_jk

Right value update of the input layer to hidden layer are as follows:

w_ij=w_ij+ηH_j(1-H_j)x_iw_je

I=1,2,3j=1,2, l

Right value update of the hidden layer to output layer are as follows:

w_j=w_j+ηH_jE j=1,2, l

η is learning rate, H in formula_jIt is the output of hidden layer neuron.E is the error of true value and output valve.

Optionally, the hidden layer threshold value a is updated_j, output layer threshold value b further comprises, more according to neural network forecast error e New network node threshold value a_j, b.

The threshold value of input layer to hidden layer is updated to aj；The threshold value of hidden layer to output layer is updated to b

a_j=a_j+ηew_jH_j(1-H_j) j=1,2, l

B=b+e

C: predicting the FTP value of active user, the basic data information inputted on the platform according to the active user, The FTP value of the active user is obtained based on FTP prediction model.

Optionally, weight threshold updates the Decent Gradient Methods for using error back propagation.

Right value update is carried out using steepest descent method, activation primitive is

Hidden layer output, output layer output are as follows:

It is defined, is enabled according to steepest descent method

Then have:

w_j=w_j+ηH_je

w_ij=w_ij+ηH_j(1-H_j)x_jw_je

Wherein:

Similarly, the threshold value bj of the threshold value aj of hidden layer and output layer updates are as follows: is defined, is enabled by steepest descent methodThen have:

B=b+ η e

a_j=a_j+ηew_jH_j(1-H_j)

Optionally, according to the FTP value of the active user and the weight information of user, the active user is calculated Ride ability rating of the function body than function body ratio=FTP/weight, according to function body than obtaining the active user.Pass through function body Than user's rough ability of riding at present can be provided.The ability dividing condition of male user function body ratio is as shown in table 1.

The ability dividing condition of 1 male user function body ratio of table

Male	Occupation	Elite	It is outstanding	Commonly
					Function body ratio W/kg	>5	4~5	3.1~4	<3

Optionally, grab in database user for a period of time in maximum 5 seconds in motion recording, 1 minute, 5 minutes it is average Power, in conjunction with the active user the FTP value user and weight information, form the power form number of the active user According to [MMP_5s/weight,MMP_1min/weight,MMP_5min/weight,FTP/weight].According to active user's power shape The locomitivity type of state data judging user (spurt type, chases hand, is comprehensive endurance type).

After having the power form of user, the power come is sorted out in conjunction with Dr.AndrewCoggan and HunterAllen Form table further judges driver's type.(1) it after the power morphological data of user shows the curve of [one] type in table, says Bright user is comprehensive driver.To be showed in play, user should focus on promoting a certain of oneself Item ability, such user more have advantage in play.(2) when showed in power form table [] type curve after, illustrate to use Family is spurt type player.The Oxidation-free casting of the user is stronger, and quick muscle fiber is flourishing.User is aobvious in the sprint stage advantage of match It writes.(3) after the power morphological data of user shows [/] type curve in table, illustrate that user is endurance type driver.User's Lactic acid threshold value and aerobic capacity are relatively good, do well in FTP, ability to sprint is slightly worse.(4) when the power morphological data of user After showing [Λ] type curve in table, illustrate that user is to chase hand.The Oxidation-free casting of the user and maximum oxygen uptake capacity are all strong. It is being more suitable for the project of chasing during the games.

Optionally, figure (3) display according to the demand of user, locomitivity grade and locomitivity type to user recommend needle To the drill program or training course of property, user can customize and select the special training plan on platform.

Scheme (4) the method for the present invention test flow chart, optionally, collects all amounts for men of nearly 2 years register account numbers According to.The user information of collection include: height (cm), the age (year), on weight (kg) and current platform user FTP (W).It examines Consider in the data of acquisition, the accuracy of the information of user needs to screen data.By screening, 2464 numbers are obtained According to sample.For the superiority and inferiority of testing model, data are divided into two parts of training set and test set.Wherein training and test according to The ratio of 7:3 separates.Stepping is carried out to the data that data are concentrated first, according to the current strength situation of user, we are user It is divided into 7 grades.Data set is divided by different level crowds: body is poor, not training, it is medium, good, outstanding and outstanding. Data set becomes 7 Sub Data Sets at this time, and then this 7 small data sets are trained with stroke of collection and test set respectively Point.Test sample is randomly choosed according to the principle of " test number=0.3* sample number ", remaining sample is as training sample (note Meaning: test number here is the rounding of 0.3* sample number).The actual value of (5) FTP test set data is schemed compared with predicted value.