CN104200076B - A kind of Athletes damage risk method for early warning - Google Patents

Abstract

The invention discloses a kind of Athletes damage risk method for early warning, at the model proposed with reference to foreign scholar, multianalysis athlete injury gained in sports is caused injury on the basis of the factor, propose sports injury muscle injury Risk-warning and hinder the sub-dynamic chain model of reason, on the basis of this model, choose the corresponding factor at sportsman's Risk-warning database, utilize analytical hierarchy process to set up sports injury muscle injury Risk-warning and hinder the sub-dynamic chain quantitative model of reason.The SOM of utilization neural network discrete method of the present invention carries out sliding-model control to achievement data; Adopt and carry out Decision Table Reduction based on the clearly demarcated matrix method in rough set; RBF neural is built based on the decision table simplified; Training RBF Neural Network, finally obtains correct diagnostic result.This method can make the effective predicted motion person's injury gained in sports of early warning comparatively accurately risk class to injury gained in sports, is beneficial to treatment and the prevention of injury gained in sports.

Description

A kind of Athletes damage risk method for early warning

Technical field

The present invention relates to a kind of Athletes damage risk method for early warning based on Rough Sets Neural Networks.

Background technology

In track and field sports, the factor relevant to injury gained in sports is a lot, there are inner Risk Factors (age, sex, damage history etc.), outside Risk Factors (place, apparatus, weather etc.), bring out incentive condition (training burden, match density etc.), in so many factor, how to find out the relation of they and injury gained in sports risk, be a large order.And in these factors the existing index that can accurately measure as physiological and biochemical index, have again some qualitatively index as medical history, physical characteristic etc., and the index that each sportsman surveys in passing data is not quite similar again, therefore will carry out comprehensive assessment to injury gained in sports risk is a very difficult job.Radial basis function (RBF:Radial BasisFunction) Learning Algorithm, because of its good non-linear mapping capability and unique network structure, the every field such as pattern-recognition, data prediction, system identification, image procossing, speech understanding and Function Fitting are now widely used in.But, utilize neural network to carry out injury gained in sports risk assessment, there is the problems such as selected characteristic parameter is too much, scale of neural network is too huge, learning training time overlength, badly influence practical effect and the real-time of this model.This is the deficiencies in the prior art part.

Summary of the invention

The technical problem to be solved in the present invention is to provide a kind of Athletes damage risk method for early warning, and effective predicted motion person's injury gained in sports risk class, is beneficial to treatment and the prevention of injury gained in sports.

The present invention adopts following technological means to realize goal of the invention:

A kind of Athletes damage risk method for early warning, is characterized in that, comprise the steps:

(1) analysis of the injury gained in sports Risk-warning factor;

(2) propose sports injury muscle injury Risk-warning to cause injury factor dynamic chain model;

(3) sportsman's Risk-warning database is set up;

(4) use analytical hierarchy process, set up sports injury muscle injury Risk-warning and to cause injury factor dynamic chain quantitative model;

(5) choose sportsman and the Risk-warning factor according to sports events, utilize the corresponding data in exercise risk database, set up exercise risk early warning Sample Storehouse and the test sample book data of this project;

(6) judge the discreteness of the Risk-warning factor corresponding data in exercise risk early warning Sample Storehouse, if not discrete, call SOM neural network discrete method and sliding-model control is carried out to data;

(7) by the achievement data after utilizing SOM network discrete and the Data Synthesis decision table initially with regard to discrete valuation;

(8) according to the yojan process of RS theory, the extraction of attribute reduction Sum fanction is carried out to decision table, obtain one group of yojan result;

(9) RBF neural is built based on the decision table simplified;

(10) according to the decision table after simplification, the corresponding data in exercise risk early warning Sample Storehouse is chosen, Training RBF Neural Network;

(11) judge whether yojan result restrains, and whether meets accuracy requirement, if do not meet, modified RBF Neural Networks structure be connected weights, go to step (10);

(12) utilize test sample book data, whether the RBF neural that training of judgement is good obtains correct diagnostic result, if do not have, selects other yojan result, go to step (9) from step (8);

(13) preserve the RBF neural trained, utilize the new sportsman's risk assessment index detected of RBF neural assessment trained;

The SOM neural network discrete method of described step (6) comprises the steps:

(6.1) calculate the importance degree of each connection attribute in decision table and sort by order from big to small;

(6.2) each continuous print injury gained in sports is caused injury factor attribute a _i(i=1,2 ..., | C|) a corresponding SOM neural network, illustrate with the discrete example that turns to of an attribute below, the initial value of given SOM neural network output layer neuron number is 3, and input layer number is 1, by the connection weight w of network _jgive [Min (a _i), Max (a _i)] random value in interval, determine the initial value α (0) of learning rate α (t), α (t) slowly should reduce with the increase of t, total the study number of times T determined;

(6.3) by attribute a _iin numerical value x be input to SOM neural network;

(6.4) calculate input numerical value x and be connected weight w _jbetween distance;

D _j=(x-w _j) ², j=1,2,3 (formula 1)

(6.5) choose and make apart from minimum neuron k as winning node;

(6.6) the connection weights of winning node collection are upgraded

W _j(t+1)=w (t)-α (t) (x-w _j) (formula 2)

(6.7) return step (3), choose another data input network, until all mode in sample is supplied to network;

(6.8) renewal learning rate:

$\mathrm{\α}\left(t\right)=\mathrm{\α}\left(0\right)[1-\frac{t}{T}];$ (formula 3)

(6.9) make t=t+1, return step (6.2), repeat to t=T;

(6.10) calculate the incompatibility of decision table after discretize, if meet incompatibility threshold requirements, calculate end, otherwise turn next step;

(6.11) attribute of importance sorting first is selected, its SOM neural network output layer neuron number is adjusted to 4, the SOM neural network output layer neuron number that all the other attributes are corresponding is constant, again discretize decision table, and go to step (6.10), until the maximum cluster numbers of each SOM network is adjusted to 4, the maximum cluster numbers of each SOM network is defined as 4, if maximum its SOM neural network clustering number of attribute of importance degree reaches requirement, the attribute then selecting importance degree to take second place adjusts, by that analogy;

(6.12) final discretization operations is completed.

As the further restriction to the technical program, described step (8) adopts carries out Decision Table Reduction based on the clearly demarcated matrix method in rough set.

As the further restriction to the technical program, describedly to comprise the steps: based on the clearly demarcated matrix method in rough set

(8.1) exercise risk early warning Sample Storehouse data are read in, as decision table;

(8.2) the clearly demarcated Matrix C of this decision table is calculated _d;

(8.3) in clearly demarcated matrix, all values are non-zero, 1, element c _ij, set up corresponding logical expression L _ij, wherein

(formula 4)

(8.4) by all expression formula L that extracts _ijdo conjunction computing, obtain a conjunctive normal form L, namely

(formula 5)

(8.5) form above-mentioned conjunctive normal form being converted to disjunctive normal form is

(8.6) output attribute yojan result, the corresponding attribute reduction result of each conjunct in disjunctive normal form, often

The set of properties comprised in individual conjunct is into about the conditional attribute set after letter.

As the further restriction to the technical program, described step (9) comprises the steps:

(9.1) basis function is chosen

Select Gaussian function as the activation function of hidden layer unit, suppose that input vector is n dimension, have m output node, the number of hidden layer node is k, and the output of hidden node is expressed as:

$R_j\left(x\right)=e^{-\frac{{\left|\right|x-c_j\left|\right|}^2}{{2\mathrm{\σ}}_j^2}}$ (formula 6)

Wherein j={1,2 ..., k}, Rj (x) are for a jth hidden node is to output, and x is that the n of network ties up input vector, c _jbe the kernel function center vector of a jth hidden layer node, it has identical dimension with input vector x, σ _jfor the width of a jth hidden layer node Gaussian function, || x-c _j|| represent the distance measure between input vector and central point.Network exports and is defined as:

$y_i=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{R}_j\left(x\right)$ (formula 7)

Wherein i=1,2 ..., m; Yi is the output of i-th output layer node, w _ijfor a jth hidden layer node is to the connection weights of i-th output layer node.In this project, only need 1 export, i.e. i=1,

(9.2) hidden layer design

The hidden layer of design RBF neural, mainly determine hidden layer number of unit and their excitation function, adopt Gaussian function, Gaussian function is determined by cluster centre and cluster width, the classification of training mode is known, i.e. injury gained in sports Risk-warning grade, therefore a kind of simple mode can be adopted to determine, the i.e. corresponding Gaussian function totally three of each class, i.e. k=3, the width parameter of Gaussian function can get the average of three class samples and its central point distance, and the center of Gaussian function can get the central point of average as respective unit of three class samples respectively,

(9.3) radial basis function center, width and the renewal to the weights between output layer

Radial basis function center and other parameter all experience learning process, usually adopt error correction learning process, application gradient descent method, specific as follows:

Suppose have N number of sample to input, to all input amendment, definition error function:

$\mathrm{\ξ}=\frac{1}{2}\underset{q=1}{\overset{N}{\mathrm{\Σ}}}{e}_q^2$ (formula 8)

Wherein e _qfor error, be defined as follows:

$e_q=d_q-y\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\Σ}}}{w}_{1j}{R}_j\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\Σ}}}{w}_{1j}{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{{2\mathrm{\σ}}_j^2}}$ (formula 9)

Wherein d _qfor sample x _qthe value of required type, the d that the sample as risk class high in this project is corresponding _q=3.

As the further restriction to the technical program, described step (10) comprises the steps:

Step one: according to known class number determination hidden layer node number, k=3 in this project, asks for the center initial value c of mean value as such of all input vectors of this classification _j(1), using the width initial value σ of the mean value of all for every class input vectors and centre distance as such _j(1), each classification inputs a sample, asks for initial w _1j(1), setting allowable error ε, ε is an a small amount of setting learn efficiency η being greater than zero ₁=η ₂=η ₃=1/t, t are iterative loop variable, and setting loop variable t is initially 1, and maximal value is MaxT;

Step 2: input t training sample, obtains network actual output y (t);

Step 3: calculate the actual error ξ (t) exported between desired output, if error ξ (t) < is ε, then this sample does not need to adjust network parameter, jumps to step 6, otherwise enters next step;

Step 4: by formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;w}_{1j}\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}---\left(29\right)$

formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;c}_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^2}{e}^{-\frac{{\left|\right|x_q-c_j\left(n\right)\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}(x_q-c_j\left(n\right))---\left(31\right)$

formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;\mathrm{\&sigma;}}_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q(n\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^3}{\left|\right|x_q-c_j\left|\right|}^2{R}_j\left(x_q\right)---\left(33\right)$

Calculate $\frac{\&PartialD;\mathrm{\&xi;}\left(t\right)}{{\&PartialD;w}_{1j}\left(t\right)},\frac{\&PartialD;\mathrm{\&xi;}\left(t\right)}{{\&PartialD;c}_j\left(t\right)},\frac{\&PartialD;\mathrm{\&xi;}\left(t\right)}{{\&PartialD;\mathrm{\&sigma;}}_j\left(t\right)},$ And press formula 29

$w_{1j}(n+1)=w_{1j}\left(n\right)-{\mathrm{\&eta;}}_1\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;w}_{1j}}---\left(30\right)$

formula $c_j(n+1)=c_j\left(n\right)-{\mathrm{\&eta;}}_2\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;c}_j}---\left(32\right)$

formula ${\mathrm{\&sigma;}}_j(n+1)={\mathrm{\&sigma;}}_j\left(n\right)-{\mathrm{\&eta;}}_3\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;\mathrm{\&sigma;}}_j\left(n\right)}---\left(34\right)$ Calculate w _1j(t+1), c _j(t+1), σ _j(t+1);

Step 5: upgrade new network parameter w _1j(t+1), c _j(t+1), σ _j(t+1), t=t+1; If t>MaxT, display does not restrain, and enters step 6, otherwise forwards step 2 to;

Step 6: whole learning process terminates, preserves each parameter of current network.

Compared with prior art, advantage of the present invention and good effect are: SOM neural network discrete method of the present invention carries out sliding-model control to achievement data; Adopt and carry out Decision Table Reduction based on the clearly demarcated matrix method in rough set; RBF neural is built based on the decision table simplified; Training RBF Neural Network, finally obtains correct diagnostic result.This method can make the effective predicted motion person's injury gained in sports of early warning comparatively accurately risk class to injury gained in sports, is beneficial to treatment and the prevention of injury gained in sports.

Accompanying drawing explanation

Fig. 1 is the injury gained in sports cause of disease dynamic chain model that Meeuwisse (1994) sets up.

Fig. 2 is Sample Storehouse collecting flowchart figure of the present invention.

Fig. 3 is physiological and biochemical index information acquisition FB(flow block) of the present invention.

Fig. 4 is the potential factor information collecting flowchart block diagram of causing injury of discrete motion of the present invention.

Fig. 5 is present system master data fill flow path figure

Fig. 6 is present system day regular data fill flow path figure

Fig. 7 is sports injury muscle injury Risk-warning cause of disease dynamic chain model of the present invention.

Fig. 8 is that athlete of the present invention causes injury factor rating figure.

Fig. 9 is the process flow diagram of the preferred embodiment of the present invention.

Figure 10 is neural network topology structure figure of the present invention.

Figure 11 is that track and field team of the present invention injury gained in sports is caused injury risk facior data acquisition tables.

Figure 12 is sports injury muscle injury Risk Factors risk assessment table of the present invention.

Embodiment

Below in conjunction with accompanying drawing and preferred embodiment, the present invention is further described in detail.

One, the analysis of injury gained in sports Risk-warning correlation factor

A. cause injury the factor in inside

(1) age

Table-1 athlete's Age Indices damages risk assessment of causing injury

(2) sex

Table-2 track and field saltaroria sports events sex index damages risk assessment of causing injury

(3) menstrual cycle

Table-3 athlete's menstrual cycle index damages risk assessment of causing injury

(4) physical training amount (sport rank)

Table-4 athlete's sport ranks (physical training amount) are caused injury risk assessment

(5) history is damaged

Table-5 athletes damage history index and damage risk assessment of causing injury

(6) injury recovery situation

Table-6 athlete's injury recovery status index damages risk assessment of causing injury

(7) body shape (BMI index)

Table-7 athlete BMI index damages risk assessment of causing injury

(8) range of motion (ROM)

Table-8 athlete's range of motion (ROM) index damages risk assessment of causing injury

(9) stability of joint (arthrochalasis degree)

Table-9 athlete's stabilities of joint (arthrochalasis degree) index damages risk assessment of causing injury

(10) muscle strength and harmony

The harmonious index of table-10 athlete's muscle strength damages risk assessment of causing injury

(11) balance ability

Table-11 athlete's balance capacity index injurys gained in sports are caused injury risk assessment

(12) physical anatomy textural anomaly

Table-12 athlete's physical anatomy textural anomaly index injurys gained in sports are caused injury risk assessment

B. cause injury the factor in outside

(1) training, race conditions

Show-13 training athletes, risk assessment is caused injury in the injury gained in sports of race conditions index

(2) technical merit

Table-14 athlete's technical merit index injurys gained in sports are caused injury risk assessment

(3) safeguard measure

The injury gained in sports of table-15 athlete's safeguard measures (injury protection utensil) is caused injury risk assessment

(4) sport footwear

The damage of table-16 track and field running type games project sport footwears is caused injury risk assessment

(5) athletic ground

Table-17 athlete athletic ground index injurys gained in sports are caused injury risk assessment

(6) weather conditions (climate condition)

Table-18 athlete's weather conditions (climate condition) injurys gained in sports are caused injury the assessment of risk

C. the Induced by Stimulation factor

(1) error of performance

Table-19 athlete's error of performance index injurys gained in sports are caused injury risk assessment

(2) train, competition schedule is unreasonable

Table-20 athlete's exercise load figureofmerit injurys gained in sports are caused injury risk assessment

Table-21 athlete's training and competition plan index injurys gained in sports are caused injury risk assessment

(3) physical function is in poor shape

Table-22 athlete's physical function status index injurys gained in sports are caused injury risk assessment

(4) psychological factor

Table-23 athlete's psychological factor index injurys gained in sports are caused injury risk assessment

(5) with other people or apparatus accident collision

In athletic training and match, due to sportsman mutually between fierce contact-impact or cause injured possibility comparatively rare with the accidental cause such as apparatus collision, due to sportsman, in order to fight for vantage point, mutual crowded touching causes damage when competing the start of a race or intermediate run or making a spurt may to occur in middle-distance race, marathon etc., also may occur in hurdler and cause with the contact-impact of hurdle frame in training and competition.In a word, this kind of accidental cause is comparatively rare in track and field event, but once occur, just probably causing injured, therefore, is the risk factors causing athlete to damage with other people or apparatus accident collision, and concrete evaluation can reference table-24.

Table-24 athletes and other people or the injury gained in sports of apparatus accident collision factor are caused injury risk assessment

Two, the foundation of injury gained in sports Risk-warning cause of disease dynamic chain model

A. track and field event spy sends out breakdown diagnosis

By inquiry lot of documents, collect the data about track and field projects common injury, be summarized as follows by projects category induction.

(1) short in class

1) dash class

Table-25 sprint events spy sends out breakdown diagnosis

2) to hurdle class

Table-26 is hurdled, and project is special sends out breakdown diagnosis

(2) middle-distance race intermediate item

Table-27 middle-distance race intermediate item spy sends out breakdown diagnosis

(3) saltaroria project

Table-28 high jump project spy sends out breakdown diagnosis

Table-29 long-jump project spy sends out breakdown diagnosis

(4) intermediate item is thrown

1) javelin

Table-30 javelin project spy sends out breakdown diagnosis

2) discus

Table-31 discus project spy sends out breakdown diagnosis

3) hammer

Table-32 hammer project spy sends out breakdown diagnosis

4) shot

Table-33 shot project spy sends out breakdown diagnosis

B. potential sick and wounded generation sign is analyzed

Find by analyzing the reason causing various different motion to damage, athletes ' functions ability declines, and generation exercise induced fatigue and fatigue fail to recover in time, are the key factors of bringing out damage.Xu Yong etc. find in China long jumper injury gained in sports epidemiological study, long jumper when training different somatosensories and injury gained in sports have direct relation, when having 40.3% damage to occur, somatosensory is " fatigue ".If that is there is fatigue or recover not good general meaning that having a big risk of damage increases greatly in body.Therefore, a sign that exercise induced fatigue or the appearance recovering bad can be occurred as latent lesion.In addition, early stage in some sick and wounded appearance, have some specific signs, if energy Timeliness coverage, carry out early warning, take corresponding reduction to cause injury measure that the factor occurs, just can avoid sick and wounded generation or further develop, thus avoids or reduce sick and wounded to athletes ' body and the adverse effect of normally training.

(1) exercise induced fatigue and recover bad diagnostic factro

β 1 commonly uses physiological fatigue diagnostic factro

The conventional physiological fatigue diagnostic factro A of table-34

The conventional physiological fatigue diagnostic factro B of table-35

β 2 commonly uses mental fatigue diagnostic factro:

Motility mental fatigue is that sportsman is when tackling endogenous pressure and exogenous pressure, when Psychological Resources and physiology resource are not supplemented in time by constantly consuming the mental function that occurs can not maintain the phenomenon that original psychological activity level and mental function decline, be in particular in the change of emotional dimension, cognitive dimension, power dimension, behavior dimension and physical dimension.Motility mental fatigue has passive impact to athletic mental health state, physiological fatigue and damage etc., therefore monitors motility mental fatigue and just seems especially important.At present, following methods is mainly contained to the monitoring of motility mental fatigue:

The conventional mental fatigue diagnostic factro of table-36

(2) potential sick and wounded omen sign (feature)

Table-37 potential sick and wounded omen signs (feature)

(3) sportsman's self-monitoring

Sportsman passes through self-monitoring, can understand the change of some body abnormality in time, and doctor, trainer coordinate, timely adjusting training and scheme, contribute to prevention over training and traumatic incidents, sportsman also can be made to form note the custom observing training philosophy and personal hygiene.The content that oneself detects comprises subjective sensation and objective material.Subjective sensation refers to general sensation, sleep, appetite, sport mood etc.Objective material refers to pulse, body weight, vital capacity, grip, back strength, perspiration situation etc.These materials are wanted simultaneously and training content, match or performance of the test combine and analyze.

It is energetic that sportsman feels normally, happy, and training enthusiasm is high, and after exercise, the slightly tired and tired sense of muscle acid, recovers after rest very soon.As felt lassitude, unable, sleepy, dizzy, passionnate, the aching and limp numbness of local articular muscle, even pain, chest tightness, breathes hard, abdominal distension, and stomachache etc. are all abnormal occurrencies.Sleep is physiological requirements, and people can be made fully to be had a rest and regain one's strength.Frequent insomnia, is easily formed overtired.Appetite is all good under normal training, but in over training, overstretched or can reduce when there is some morbid state.At once the appetite decline phenomenon occurred after a hard training or after the match of anxiety is temporary transient, can recover normal very soon.Sport mood refers to that sportsman participates in the hope of training and competition.When training good, in high spirits, excitability is higher, full of physical strength, and sports achievement is generally also better.When physical function is not good, can occur training cold, even situation is weary of etc. to place, apparatus.Pulse generally records from radial artery, also can record by arteria carotis, should note pulse frequency, rhythmicity, the change such as strong and weak of beating.Time quiet, pulse may be slack-off with the raising of training degree, as there is suddenly acceleration or excessively slow, then will find reason.Sometimes can there is irregularity of pulse with breathing in pulse, be normal phenomenon; As obviously irregular in occurred, should doctor be notified at once.Body weight is general in group-up player keeps constant, slightly can alleviate, but may go up soon when just participating in training.As continued weight declines, should note whether there is certain deeline or seriously overtired.After a hard training, body weight may decline 1 ~ 4 kilogram, but recovers again normal after 1 ~ 2 day.Grip, arm strength all increase when training good, and continuous decrease is abnormal.Perspire how much because temperature, amount of drinking water, the level of training are different and different with personal touch.Sportswoman also will record menstruation.

Physical function is easy to the increase causing damage risk when declining, and this tells about above, therefore by athletic self-supervision, timely and effectively athletic health is fed back, and will contribute to the generation reducing sports wound.At this, sportsman's self-monitoring (mainly feeling based on sportsman) is classified as evaluation damage and causes injury risk factors, concrete evaluation can reference table-38.

Table-38 athlete's self-supervision index damages risk assessment of causing injury

(4) trainer's monitoring

As trainer, in the process arranging athletic training and competition, not only to understand the implementation status of training athlete and match, also should be able to judge athletic physical function situation.Can not only correct in time when technical movements mistake appears in sportsman, experience also by self judges whether sportsman occurs the degree of fatigue and fatigue in the athletic words and deeds of observation, so just can better grasp athletic body building situation and avoid sportsman to occur sick and wounded.In addition, outstanding trainer also should be able to understand or grasp the body function monitor indexs such as some customary physiological biochemistry, such could in conjunction with self observe and objective indicator reasonably adjust sport training plan and amount of exercise, can reach and improve locomitivity and achievement, the object avoiding exercise injury and disease can be played again.

The monitoring of trainer can comprise to be observed athlete's life and the body action situation in training, exchanges with athletic and judgement etc. to conventional physiological and biochemical index, therefrom show whether self may occur sick and woundedly making corresponding judgement to sportsman, and show that sportsman damages the height (reference table-39) of risk of causing injury.

Table-39 track and field monitoring sportsman damages risk assessment of causing injury

C. model is set up

On the multifactor dynamic chain basis of the Meeuwisse injury gained in sports cause of disease, we attempt by injury gained in sports risk feedback fusion wherein, establish the injury gained in sports Risk-warning cause of disease dynamic chain model for athlete, as Fig. 7.

In this dynamic chain, more to the right, the possibility that damage occurs is larger, and more to the left, the possibility that damage occurs is less.When the category-A factor (inside cause injury the factor) occurs, sportsman becomes " having damage tendency crowd "; Now, when the different B class factor (outside cause injury the factor) occurs, sportsman becomes " damage easily sends out crowd ", and the risk that certain class specific Damage occurs for the sportsman of track and field disparity items will increase greatly.Now should carrying out first time early warning, by reducing the measure of damage risk, the state of sportsman in dynamic chain being moved to the left, return to " having damage tendency crowd " stage, its danger occurring to damage is reduced.If sportsman belongs to " damage easily sends out crowd ", the appearance of the C class factor (bringing out stimulating factor), increases further by the risk making sportsman face damage.Now sportsman often may show sick and wounded sign, the feature that will occur, and now should carry out second time early warning, removes and causes sick and wounded all kinds of risk factors of causing injury, the position of sportsman in dynamic chain is moved to left, avoids the generation that it damages; Otherwise sportsman will there will be sick and wounded.

Three, sportsman's Risk-warning database is set up

A. injury gained in sports risk data gathers and pre-service

After the analysis of injury gained in sports Risk factors completes, can set about carrying out the foundation of injury gained in sports risk warning model, modeling method both can adopt dynamic chain model qualitatively, also can adopt quantitative model.Traditional linear model can be chosen in quantitative model, also can adopt non-linear modeling method based on this, as rough set or neural network etc.No matter adopt which kind of modeling method, first all will carry out the data collection effort of sample database.The collection of injury gained in sports risk data is mainly comprised the potential injury gained in sports obtained by injury gained in sports analysis and to cause injury the factor, and the value according to the factor can be divided into two classes: 1) the continuous data factor, is namely used for monitoring the conventional physiological and biochemical index of player's fatigue; 2) the discrete data factor, mainly comprises the information such as sports events, sport rank, stability of joint.The flow process of data acquisition, as shown in figure-2.

Data acquisition and pretreated concrete steps as follows:

Step one: the measurement of physiological and biochemical index;

Step 2: the process of physiological and biochemical index data;

Step 3: sports injury muscle injury is caused injury the establishment of risk facior data acquisition tables and data acquisition;

Step 4: sports injury muscle injury is caused injury the data statistics of risk facior data acquisition tables;

A.1. forecasting fatigue relevant physiological biochemical indicator gathers and pre-service

(1) measurement of forecasting fatigue relevant physiological biochemical indicator

1) routine blood test:

Routine blood test detects and comprises the contents such as RBC number, haemoglobin, hematocrit, leucocyte.

Play test morning on every Mondays, extracting vein blood.

Testing tool is that German BAYER Automatic Blood Cell Analyzer is tested.

2) serum testosterone, cortisol, serum testosterone (T)/serum cortisol (C) ratio:

Serum testosterone, cortisol are according to getting a venous blood test project characteristic every 2-4 week, and it is 1-2 week that pre-games generally increases test frequency.Serum testosterone (T)/serum cortisol (C) ratio generates index for it.

Testing tool is that German BAYER automatic clinical chemistry analyzer carries out.

3) creatine kinase and urea nitrogen:

Rise morning on every Mondays and get peripheral blood after emphasis training session.

Testing tool is German EPPENDORF semi-automatic biochemical analyzer.

(2) process of physiological and biochemical index data

After having gathered the physiological and biochemical index value that track and field team measures weekly, for the ease of final assessment Risk-warning grade, by according to the standard in table-40 by physiological and biochemical index discretize, then comprehensively become the fatigue monitoring factor for dynamic chain model.

Table-40 physiological and biochemical index grade scales

In physiological and biochemical index discretization process, larger difference is there is between the physiological and biochemical index due to Different Individual, can not go by unified standard to weigh Different Individual, conveniently the physiological and biochemical index of computerized algorithm to Different Individual is compared, and therefrom excavate and incidence relation between injury gained in sports risk, therefore must complete normalized to physiological and biochemical index when submitting sample data to.Concrete grammar be using sportsman cycle of training Different periods measure Physiological And Biochemical Values computation of mean values as its individual base value, by single measurement data in contrast, using ratio as this achievement data input value.

A.2. potential injury gained in sports is caused injury Qualitative factor data acquisition and pre-service

(1) sports injury muscle injury is caused injury the establishment of risk facior data acquisition tables

For setting up injury gained in sports Risk-warning cause of disease dynamic chain quantitative model, and the sick and wounded early warning system of athlete researched and developed based on data mining technology and mobile calculation technique, consulting a large amount of domestic and foreign literature, and with the expert of medicine in field of sports medicine and track and field team of Shandong Province special trainer, sportsman carries out interview, understand the incidence about track and field event injury gained in sports and common traumatogenic reason, and the factor of causing injury is conducted in-depth analysis and system sort out basis on, we draft " sports injury muscle injury cause injury risk facior data acquisition tables ", for the collection of cause injury factor information and sick and wounded library information.This data acquisition table is within cycle of training, and every circumferential sportsman provides once, and sportsman fills according to own actual situation.

(2) sports injury muscle injury is caused injury the data statistics of risk facior data acquisition tables

Option corresponding for this factor out, and is carried out assignment according to it to the size of injury gained in sports risk contribute by Factor Selection that first will be relevant to injury gained in sports risk in " sports injury muscle injury cause injury risk facior data acquisition tables ".For 3 option factors, the selection level assignment 1 minimum to injury gained in sports risk contribute, the assignment 3 of maximum option.In some specific factor, such as, damage history, if certain sportsman not previously damage, then this factor assignment 0.On this basis, with criteria table, the data acquisition table that sportsman fills in is added up.

B. the sick and wounded early warning system data inputting of athlete

The data general data that the sick and wounded early warning system of athlete is relevant to injury gained in sports early warning is distributed in four databases, comprise sportsman's basic information database, training supervision ADMINISTRATION SUBSYSTEM database, injury gained in sports Risk-warning subsystem database and the sick and wounded database of individual, according to the stability of data, filling in of whole database data is divided into fill in each cycle of training, fills in weekly, every day fills in and irregularly upgrade four classes.Wherein filling in of first two data is carried out according to sports injury muscle injury Warning System master data fill flow path block diagram, filling in of latter two data is carried out according to sports injury muscle injury Warning System daily workout data fill flow path block diagram, refers to Fig. 5, Fig. 6.In each database, contained data are as shown in table-41.

Table-41 athlete sick and wounded early warning system sportsman data inputting database

Three, use analytical hierarchy process, set up sports injury muscle injury Risk-warning and hinder the sub-dynamic chain quantitative model of reason

Although qualitative dynamic chain model can prevent the generation of injury gained in sports to a certain extent, the factor affecting early warning judgement is too much, and standard is comparatively fuzzy, practical operation difficulty.Therefore this project has carried out the structure of dynamic chain quantitative model on this basis.In documents and materials previously, have no the contrast of importance between different Risk Factors, the foundation of this specified rate dynamic chain model brings difficulty.The U.S. plans strategies for analytical hierarchy process (the Analyt ic Hierarchy Process that scholar A.L.saaty proposes in 20 century 70s, be called for short AHP method), be the multi index system of scheme is analyzed a kind of stratification, structured decision method, it is by policy-making thought process modelling, the quantification of decision maker to complication system.Should in this way, decision maker, by challenge is decomposed into some levels and some questions, simply compares and calculates between each factor, just can draw the weight of Different factor, for the selection of preferred plan provides foundation.The method can effectively solve in quantitative dynamic chain model process of establishing, the problem of different motion damage Risk Factors weight.

Because injury gained in sports risks and assumptions of causing injury is a lot, for the ease of analyzing, the Factor Weight problem analysis of injury gained in sports being caused injury of this project is divided into three levels.The superiors are destination layer, are injury gained in sports Risk-warning; Middle layer is indicator layer, refers to by the interview of sportsman, trainer and sports medical science expert and survey, three classifications injury gained in sports Risk Factors divided; Orlop is that object layer is concrete injury gained in sports Risk Factors.In order to set up indicator layer, namely classify to injury gained in sports Risk Factors, this project has made " athlete's injury gained in sports factor risk factors assessment table ".By being distributed to 15 senior above track and fields and team doctor fills in, result statistics following (for middle-distance runner).

α class is caused injury the factor (the most easily cause damage factor): comprise damage history, injury recovery situation, technical movements are wrong, stability of joint is poor, training and competition load is misorient, psychology is not in good state in training and competition.

β class is caused injury the factor (more easily cause damage factor): comprise muscle strength and harmony, balance ability, joint motion amplitude (pliability), with or without safeguard measure, physical anatomy structure have without exception, and other people or apparatus have collisionless etc.

γ class is caused injury the factor (factor that injurious effect is less): comprise age, sex, menstrual cycle, sport rank, athletic ground, weather conditions, sport footwear, training or competition status etc.

Cause injury after the factor determining three classes, we adopt analytical hierarchy process to carry out weight analysis to it, to show that this three class of α, β, γ is caused injury the weight to injurious effect of the factor, and adopt each sub-factor used the same method to three classes are caused injury in the factor and carry out weight analysis, draw the weight size of each sub-factor respectively.Circular is as follows.

1) hierarchy Model is set up.The foundation of hierarchical structure can problem methodization, stratification, builds three classes and to cause injury hierarchical model, and encode to the individual factor of causing injury causing the factor damaged.

2) Judgement Matricies A, compare between two for each element at the same level in hierarchy Model, its criterion is as table 42.

$A=\left[\begin{array}{cccc}{a}_{11}& a_{12}& ...& a_{1n}\\ a_{21}& a_{22}& ...& a_{2n}\\ ...& ...& ...& ...\\ a_{n1}& a_{n2}& ...& a_{\mathrm{nn}}\end{array}\right]$ (formula 10)

Table 42 judgment matrix scale and implication thereof

3) calculate each judgment matrix weight, and do consistency check.

The first step, by each column element normalization of judgment matrix, obtains the matrix after normalization

${\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}=\frac{a_{\mathrm{ij}}}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ij}}}(i,j=\mathrm{1,2},...,n)$ (formula 11)

                

Second step, is added the judgment matrix after normalization by row, obtains row vector

${\stackrel{\&OverBar;}{W}}_i=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{a}}_{\mathrm{ij}}(i=\mathrm{1,2},...,n)$ (formula 12)

3rd step, by row vector normalization, obtains characteristic vector W, is the weight of the factor of respectively causing injury

$W_i=\frac{{\stackrel{\&OverBar;}{W}}_i}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\stackrel{\&OverBar;}{W}}_j}(i=\mathrm{1,2},...,n)$ (formula 13)

4th step, calculates the Maximum characteristic root of judgment matrix

${\mathrm{\&lambda;}}_{\max }=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\frac{\left({\mathrm{AW}}_i\right)}{W_i}$ (formula 14)

5th step, carry out consistency check:

$\mathrm{CI}=\frac{{\mathrm{\&lambda;}}_{\max }-n}{n-1}$ (formula 15)

Look into the judgment matrix Aver-age Random Consistency Index RI that exponent number is n, in table 43.

Table 43RI value table

And calculate CR=CI/RI, when CR≤0.1, it is generally acknowledged that the consistance of judgment matrix is acceptable; Otherwise, then defective.

Concrete assessment result reference table 44.

Table 44 α, β, γ tri-class cause injury the factor total weight assessment

As shown in Table 44, cause injury in three class factors of risk affecting athlete's injury gained in sports, the α class factor is the most remarkable, and next is respectively β class and γ class and causes injury the factor, and α, β, γ tri-weight proportion shared by the class factor is respectively 72.4%, 19.3%, 8.3%.This shows, α class causes injury the various factors of the factor once occur, to causing sportsman to damage having the greatest impact of generation, to be also sportsman and trainer be in peacetime training and match, need the major risk factors evading and reduce for this.

Table-45 α classes are caused injury the right assessment of factor various factors inside

From table-45, in the weight analysis assessment result of the inner each Risk Factors of the factor that α class is caused injury, damage history (34.1%) causes injury risk the most obviously to sportsman, and next is respectively error of performance (19.8%), training and competition misorient (15.3%), injury recovery situation (12.5%), psychological factor (5.3%) and stability of joint (4.1%) etc.This is with document before and investigate basically identical, has its ratio again damaged of sportsman of damage history apparently higher than the sportsman not being subject to hinder; Technical movements mistake, action distortion easily causes the stressed improper generation causing damage; Training and competition are misorient refers to that training and competition arrangement does not meet athletes ' body function rule more, cause exercise intensity and amount of exercise excessive, cause athletes ' body overtired, thus cause sick and wounded generation; Injury recovery situation for most important sportsman, sick and wounded more not or participate in training and competition when not recovering completely and be easy to cause again injured and that other positions are injured generation; Athletic psychological factor also can have influence on the performance of training and competition, is overexcited, nervous, anxiety or psychological regulation ability etc. may cause athletes ' body muscle rigidity, inharmonious etc., thus increases sick and wounded danger; Stability of joint mainly refers to or without loose-jointed situation, and stability of joint declines and also makes easily to cause one of sick and wounded reason.Therefore, sportsman, in peacetime training and match, fully should understand the health of oneself, have not damaged and injury recovery situation, whether technical movements are wrong, whether training and competition amount is excessive, and psychological regulation training should be strengthened, to reduce and to reduce the generation of damage.

Table 46 β class is caused injury the right assessment of factor various factors inside

As shown in Table 46; in the weight analysis assessment of the inner each Risk Factors of the factor that β class is caused injury; with other people or apparatus collide (33.5%) be the factor the most easily causing damaging, next is respectively physical anatomy textural anomaly (26.0%), safeguard measure (13.9%), muscle strength and harmony (12.5%), range of motion (6.2%), balance ability (4.6%) and technical merit (3.2%).Collide comparatively rare in track and field event with other people or apparatus, generally see the sportsmen such as middle-distance race off or other squeeze mutually when fighting for position and touch or the accident collision of athletes ' body and hurdle frame in hurdling, once occur, be easy to cause injured, therefore, sportsman should focus one's attention in training and competition, avoids the generation of this type of accident; Whether physical anatomy textural anomaly mainly refers to that sportsman's lower limb length is whether consistent, have in knee to turn up and whether the shape of inside of the foot has abnormal situation, if sportsman has Leg length to differ or other abnormal conditions, all can increase the risk of damage; Safeguard measure, for most important pre-antisitic defect, especially had the position of old wound, answered more protection, to reduce the generation of damage; Muscle strength is a basic fitness, play most important for athletic achievement, when development force, if do not noted, the balanced development of front and back muscle group and left and right limb muscle strength is then easy to the generation causing damaging to sportsman, especially more obvious at lower limb, therefore, sportsman, while development muscle strength, must take into account the balanced development of strength; Athletes ' body major joint mobility (pliability), balance ability and technical merit also can have influence on sick and wounded generation, and therefore, sportsman wants to reduce damage, just needs the Comprehensive Quality and the improvement opportunity level that improve health.

Table 47 γ class is caused injury the right assessment of factor various factors inside

As shown in Table 47, in the weight analysis assessment of the inner each Risk Factors of the factor that γ class is caused injury, training or match (24.0%) cause injury risk significantly to motion, and next is respectively athletic ground (19.7%), menstrual cycle (14.1%), physical training amount (9.7%), body shape (BMI) (8.2%) and sport footwear (8.2%), weather conditions (7.5%), age (5.0%) and sex (3.6%) etc.For sportsman, the training and competition of ginseng more strength can increase injured possibility undoubtedly, its exercise intensity is larger, competition Shaoxing opera is strong, injured chance is more, not too can control for this type of factor sportsman undoubtedly, can only try one's best and reasonably arrange training and match to reach the object reducing damage; The change on athletic ground also can affect concerning sportsman, especially sportsman carries out training or competing at hard place (as road etc.), injured possibility can be increased undoubtedly, and injured possibility equally also can be increased in wet and slippery place or out-of-flatness place, therefore, sportsman should reduce the harmful effect brought by place as far as possible; For sportswoman, menstrual period participates in training or match may bring bad impact, especially the situation of menstrual cycle irregular or amenorrhoea, the probability of player injuries can be increased, therefore, sportswoman should understand self physical condition in time, adopts adequate measures with the harmful effect avoiding the menstrual period to cause; Athletic physical training amount mainly refers to the height of sportsman's grade, and research shows, sport rank level is higher, and the rate of injury is also larger, this and high performance athlete's exercise load more greatly, more more relevant; And athletes ' body form (BMI), sport footwear, weather conditions, age, sex etc. also respectively with certain correlativity that has of injury gained in sports, sportsman should be appreciated that the harmful effect that various Risk Factors may be brought in peacetime training and match, evade in time and reduce various risk of causing injury, sick and wounded to reach minimizing, the object improved results.

Cause injury factor information statistical form in table 48 inside

Cause injury factor information statistical form in table 49 outside

Table 50 incitant Information Statistics table

According to the weight of above biological factors, the different motion of table shown in 50-52 damages factor score of causing injury, each athletic PTS can be calculated, normal distribution based on sportsman's damage risk is supposed, data within the scope of next standard deviation in all sportsman's score averages are defined as medium risk grade, on be high-risk grade, under be low risk level.Based on the injury gained in sports Risk-warning result of dynamic chain quantitative model specifically in table 51, in table, 1,2,3 represent three Risk-warning grades from low to high respectively.

Table 51 is based on the injury gained in sports Risk-warning result early warning result table rear with correction of dynamic chain quantitative model

In sum, injury gained in sports risk factors of causing injury are factors that sportsman likely faces in each stage and link of training and competition, but the probability of happening of various risk factors of causing injury and the impact caused are not quite similar, therefore, by carrying out weight analysis assessment based on analytical hierarchy process to all kinds of factor, contribute to us to understand various factor pair injury gained in sports of causing injury and to cause injury the size of risk, for the factor that risk of causing injury is high, as α class is caused injury the factor, once occur, should adjust in time, to reduce injured possibility; And the β class factor of causing injury equally also can not be ignored, sportsman fully should understand own bodies situation in peacetime training and match, improves overall physical quality and technical merit, and reducing damage morbidity may; γ class cause injury the factor for cause sportsman damage impact less, sportsman should strengthen self-monitoring in peacetime training, evades and reduces all kinds of risk of causing injury.To cause injury the size impacted by calculating and assess all kinds of factor pair sportsman that causes injury, establish damage risk early warning cause of disease dynamic chain quantitative model.

Five, choose sportsman and the Risk-warning factor according to sports events, utilize the corresponding data in exercise risk database, set up exercise risk early warning Sample Storehouse and the test sample book data of this project;

The factor of injury gained in sports may be caused in track and field sports a lot, and different sports eventss is different again, therefore to set up one infeasible to the general injury gained in sports risk warning model of all sports eventss.Set up an Athletes damage risk assessment data storehouse like this, for different sports eventss when doing risk assessment, it will be optimal selection that selected part sportsman and part index number set up risk warning model targetedly.It is shown in table 28 that database should comprise variable, when wherein sportsman's essential information different motion project model is set up, and the screening of sportsman's sample.After sportsman has screened, according to the characteristic of disparity items, according to the conclusion of injury gained in sports early warning related-factors analysis, cause injury the factor and the Induced by Stimulation factor and choose the variable set up needed for injury gained in sports early warning Sample Storehouse in cause injury from the inside factor, outside, with the injury gained in sports risk class composition Risk-warning Sample Storehouse utilizing actual conditions+dynamic chain quantitative model+expert evaluation composite algorism to obtain, the factor in Risk-warning Sample Storehouse is stored in different databases in the sick and wounded early warning system of athlete, can call respectively when setting up Risk-warning Sample Storehouse.For long-distance runner, the concrete factor of Risk-warning Sample Storehouse is as follows:

1) sportsman's essential information: name, project, sportsman's grade, BMI;

2) injury gained in sports Risk-warning information: whether the state at heart in training load amount, error of performance, joint safeguard measure, Subjective fatigue situation, training and competition, training plan change, muscle strength contrast before and after training site weather condition, stability of joint, joint flexibility, thigh, the contrast of thigh circumference degree, balance ability, the shape of inside of the foot, technical merit, sport footwear resume training situation, menstrual cycle after changing the frequency, main training site, damage history, damage;

3) training supervision information: RBC number, haemoglobin, hematocrit, SCK, plasma wrea, serum testosterone, serum cortisol, serum testosterone (T)/serum cortisol (C), leukocyte count.

Six, judge the discreteness of the achievement data that risk warning indexes is corresponding, if not discrete, call SOM neural network discrete method and sliding-model control is carried out to achievement data

Can the factor of continuous coverage numerical value as training condition (physiological and biochemical index), carry out sliding-model control.The requirement of discretize is: (1) terseness, and the breakpoint number that discretize adopts is more few better; (2) consistance, does not change the compatibility of decision table after discretize; (3) accuracy, the precision of prediction of discretize result is more high better.

Discrete processes adopts SOM algorithm, SOM is a kind of without supervision self-organizing competing learning type feedforward neural network, utilize a large amount of training sample data to adjust the weights of network by Ad hoc mode, classification results can reflect the actual distribution situation of data comparatively objectively.When application SOM neural network carries out discrete processes to connection attribute, it is very crucial for correctly choosing clusters number, if because clusters number is very few, then may obtain inconsistent decision system, causes cannot making a policy according to Rule of judgment during practical application; If clusters number is too much.Then there will be discrete case, add the complicacy of calculating.In actual analysis, first using the incompatibility of decision table after discretize and the incompatibility of each attribute as the foundation of discrete performances evaluation.So just consider the incompatibility of decision table and the distribution characteristics of data itself after Data Discretization, both decreased the number of breakpoint as much as possible, in turn ensure that the compatibility of decision table.

The importance degree of definition 1: connection attribute a ∈ C (conditional attribute collection) is defined as follows:

1) calculate U/D, namely according to decision attribute, domain is divided into n class.Wherein in U and Sample Storehouse, sportsman is all; D is decision attribute, namely final injury gained in sports Risk-warning grade.

2) the mean value AVG (ai) of connection attribute a in each class (exercise risk damages the factor of causing injury) is calculated, i=1,2, n, and the maximal value Max (AVG (ai)) and the minimum M in (AVG (ai)) that find out mean value.

3) the standard deviation std (a) of all object connection attribute a is calculated.

Then importance degree SGF (a) of attribute a is:

$\mathrm{SGF}\left(a\right)=\frac{\mathrm{Max}\left(\mathrm{AVG}\left(a_i\right)\right)-\mathrm{Min}\left(\mathrm{AVG}\left(a_i\right)\right)}{\mathrm{std}\left(a\right)}$ (formula 16)

Definition 2: a given infosystem S=(U, C ∪ D), if incompatible object has m after discretize, concrete calculating sees formula 21, then the incompatibility α defining decision table is:

$\mathrm{\&alpha;}=\frac{m}{\left|U\right|}$ (formula 17)

Wherein | U| is the sum of object in decision table.In practical application, only need the incompatibility threshold values that setting one is less.In the applied research of current most of rough set, the neuronic number of SOM neural network output layer is that subjectivity is determined, easily occurs deviation.For this reason, according to above two definition, propose a kind of SOM neural network discretization method being applicable to rough set, step is as follows:

1) calculate the importance degree of each connection attribute in decision table and sort by order from big to small.

2) each continuous print injury gained in sports is caused injury factor attribute a _i(i=1,2 ... | C|) a corresponding SOM neural network, illustrate with the discrete example that turns to of an attribute below, the initial value of given SOM neural network output layer neuron number is 3, input layer number is 1, gives [Min (a by the connection weights wj of network _i), Max (a _i)] random value in interval, determine the initial value α (0) of learning rate α (t), α (t) slowly should reduce with the increase of t, total the study number of times T determined;

3) by attribute a _iin numerical value x be input to som network;

4) input numerical value x and the distance be connected between weights wj is calculated;

D _j=(x-w _j) ², j=1,2,3 (formula 18)

5) choose and make apart from minimum neuron k as winning node;

6) to winning node and connection weights upgrade

W _j(t+1)=w (t)-α (t) (x-w _j) (formula 19)

7) return step (3), choose another data input network, until all mode in sample is supplied to network;

8) renewal learning rate:

$\mathrm{\&alpha;}\left(t\right)=\mathrm{\&alpha;}\left(0\right)[1-\frac{t}{T}];$ (formula 20)

9) make t=t+1, return step 2), repeat to t=T.

10) calculate the incompatibility of decision table after discretize, if meet incompatibility threshold requirements, calculate end, otherwise turn next step.

11) attribute of importance sorting first is selected, its SOM neural network output layer neuron number is adjusted to 4, the SOM neural network output layer neuron number that all the other attributes are corresponding is constant, again discretize decision table, and go to step (10), until the maximum cluster numbers of each SOM network is adjusted to 4.The maximum cluster numbers of each SOM network is defined as 4, if maximum its SOM neural network clustering number of attribute of importance degree reaches requirement, then the attribute selecting importance degree to take second place adjusts, by that analogy.

12) final discretization operations is completed.

Because rough set theory is not high to the granularity requirements of knowledge, therefore maximum cluster numbers is defined as 4 and just substantially can meets requirement to incompatibility, if meet special circumstances, suitably increases that SOM network output layer is neuronic maximumly to limit the number.

Seven, by the achievement data after utilizing SOM network discrete and the Data Synthesis decision table initially with regard to discrete valuation

Eight, according to the yojan process of RS theory, the extraction of attribute reduction Sum fanction is carried out to decision table

By the data after utilizing SOM network discrete with initial with regard to after the Data Synthesis decision table of discrete valuation, then according to the yojan process of RS theory, the extraction of attribute reduction Sum fanction is carried out to decision table.The basic thought of RS uses relation of equivalence to be classified by the element in set, generates certain division of set, corresponding with relation of equivalence.According to the theory of relation of equivalence, the element in same grouping equivalence class can not be differentiated, and can carry out, can reach the object simplified information thus to the process of information in the granularity of equivalence class.It is a kind of uncertain information disposal route based on set theory.Can treatment and analysis out of true, incomplete various sizing effectively, quantitatively or the incomplete information of mixed type, therefrom find tacit knowledge, disclose potential rule.

Attribute reduction has diverse ways, and this project discernible matrix carries out yojan.A.Skowron proposed the method for the clearly demarcated matrix representation knowledge of a kind of use in 1991, this method for expressing is conducive to us to be explained and calculates data core and yojan.

Definition 3: establish S=(U, A, V, f) to be a decision table.Wherein U is domain, is the entirety of sportsman's Sample Storehouse in this application; A is community set, and A=CUD, C are conditional attribute collection, and namely all potential injurys gained in sports are caused injury the set of factor attribute, and D is decision kind set, only comprises injury gained in sports Risk-warning level attributed in this example; represent the set of all properties value, be the set of each attribute value of all sportsmen in Sample Storehouse;

F:U × A → V is information function, and it determines the value rule of each attribute of sportsman in Sample Storehouse.Definition a (x) is the value of sample x on attribute a, the then clearly demarcated Matrix C of this decision table _dcan be expressed as:

$c_{\mathrm{ij}}=\left\{\begin{array}{cc}\left\{a\right|a\&Element;A,a\left(x_i\right)\&NotEqual;a\left(x_j\right)\}& D\left(x_i\right)\&NotEqual;D\left(x_j\right)\\ 0& D\left(x_i\right)=D\left(x_j\right)\\ 1& D\left(x_i\right)\&NotEqual;D\left(x_j\right),a\left(x_i\right)=a\left(x_j\right)\end{array}\right.$ (formula 21)

Matrix is a matrix along principal diagonal symmetry, when analyzing, only considers triangle or lower triangular portions on it.When the decision attribute value of two samples is identical, the differential matrix element value corresponding to them is 0; When the decision attribute of two samples is different and when can be distinguished by the value difference of some conditional attribute, the value of the differential matrix element corresponding to them is the conditional attribute set that on these two sample attributes, value is different, namely can distinguish the conditional attribute set of these two samples; When two samples clash, when the identical and value of decision attribute of namely all conditional attribute values is different, the element value of the differential matrix corresponding to them is 1.Whether comprise 1 in differential matrix element and whether can comprise the foundation of inconsistent information and the foundation of inconsistent decision table as determination decisions table system.

Attribute reduction based on differential matrix can be divided into following six steps.

1) injury gained in sports Risk-warning Sample Storehouse data are read in, as decision table;

2) the clearly demarcated Matrix C of this decision table is calculated _d;

3) in clearly demarcated matrix, all values are non-zero, 1, element c _ij, set up corresponding logical expression L _ij, wherein

(formula 22)

4) by all expression formula L that extracts _ijdo conjunction computing, obtain a conjunctive normal form L, namely

(formula 23)

5) form above-mentioned conjunctive normal form being converted to disjunctive normal form is

6) output attribute yojan result.The result of the corresponding attribute reduction of each conjunct in disjunctive normal form, the set of properties comprised in each conjunct is into about the conditional attribute set after letter.These different yojan results available, Training RBF Neural Network, chooses the final yojan result of the best conduct of classifying quality respectively.

Specific algorithm is as follows:

Nine, RBF neural is built based on the decision table simplified

RBF neural structure and feedforward neural network similar, it is a kind of three-layer forward networks.Input layer is made up of signal source node.The second layer is hidden layer, the needs of the visual described problem of nodes and determining.Third layer is for exporting, and it makes response to the effect of input pattern.Conversion from the input space to hidden layer space is nonlinear, and is linear from hidden layer space to the conversion in output layer space.The transforming function transformation function of hidden layer unit is radial basis function, and it is a kind of non-negative nonlinear function to the decay of central point radial symmetry of local distribution.

RBF neural topological structure is as shown in table 8.RBF neural is nonlinear from the input space to the conversion in implicit space, is then linear from hidden layer space to the conversion of output region.This is a kind of topological structure of feedforward network.

Approach nonlinear system by RBF neural, after given learning sample data, its algorithm mainly solves two problems:

1) Neural Network Structure Design problem, comprises the determination of network node in hidden layer, the determination of the cluster centre C of RBF;

2) modified weight problem, namely adjusts the connection weight matrix w of hidden layer space to output layer space.

This project mainly uses RBF neural, according to input data separation athletic injury gained in sports Risk-warning grade be high, in or low, namely sportsman is divided three classes according to injury gained in sports risk size.For the purpose of simple, represent this three classifications by 1,2,3 of RBF neural output valve.Specific as follows:

Table-52RBF exports and the risk class table of comparisons

Namely the object of this project constructs a sorter for using RBF neural.Mainly comprise that basis function is chosen, hidden layer design and radial basis function center, weights and width change.This project neural network hidden layer center, width and the weights between output layer adopt gradient descent method to determine.

(1) the choosing of basis function

Select Gaussian function as the activation function of hidden layer unit, suppose that input vector is n dimension, have m output node, the number of hidden node is k, and the output of hidden node is expressed as:

$R_j\left(x\right)=e^{-\frac{{\left|\right|x-c_j\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}$ (formula 24)

Wherein j={1,2 ..., k}, Rj (x) are for a jth hidden layer node is to output, and x is that the n of network ties up input vector, c _jbe the kernel function center vector of a jth hidden layer node, it has identical dimension with input vector x, σ _jfor the width of a jth hidden layer node Gaussian function, || x-c _j|| represent the distance measure between input vector and central point.Network exports and is defined as:

$y_i=\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{R}_j\left(x\right)$ (formula 25)

Wherein i=1,2 ..., m; Yi is the output of i-th output layer node, w _ijfor a jth hidden layer node is to the connection weights of i-th output layer node.In this project, only need 1 export, i.e. i=1.

(2) hidden layer design

The hidden layer of design RBF neural, mainly determines hidden layer number of unit and their excitation function.This project adopts Gaussian function, and Gaussian function is determined by cluster centre and cluster width.In this project, the classification of training mode is known, i.e. injury gained in sports Risk-warning grade, a kind of simple mode therefore can be adopted to determine, i.e. the corresponding Gaussian function totally three of each class, i.e. k=3.The width parameter of Gaussian function can get the average of three class samples and its central point distance, and the center of Gaussian function can get the central point of average as respective unit of three class samples respectively.Although the network of this method construct seems somewhat coarse relatively, by effective learning algorithm and adopt the strategy of error correction effectively can improve the performance of network.In actual classification application, the network constructed by such mode is succinctly efficient, can not affect classification accuracy simultaneously.

(3) radial basis function center, width and the renewal to the weights between output layer

Radial basis function center and other parameter all experience learning process.Usual employing error correction learning process, application gradient descent method, specific as follows:

Suppose have N number of sample to input, to all input amendment, definition error function:

$\mathrm{\&xi;}=\frac{1}{2}\underset{q=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_q^2$ (formula 26)

Wherein e _qfor error, be defined as follows:

$e_q=d_q-y\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{1j}{R}_j\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{1j}{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}$ (formula 27)

Wherein d _qfor sample x _qthe value of required type.The d that sample as risk class high in this project is corresponding _q=3.

The iterative process of each free parameter is as follows:

1) weights of output unit

$\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;w}_{1j}\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}$ (formula 28)

$w_{1j}(n+1)=w_{1j}\left(n\right)-{\mathrm{\&eta;}}_1\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;w}_{1j}}$ (formula 29)

Wherein n represents current variable-value, and n+1 represents the value after iterated revision.

2) hidden unit center

$\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;c}_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^2}{e}^{-\frac{{\left|\right|x_q-c_j\left(n\right)\left|\right|}^2}{{2\mathrm{\&sigma;}}_j^2}}(x_q-c_j\left(n\right))$ (formula 30)

$c_j(n+1)=c_j\left(n\right)-{\mathrm{\&eta;}}_2\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;c}_j}$ (formula 31)

3) function widths

$\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;\mathrm{\&sigma;}}_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q(n\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^3}{\left|\right|x_q-c_j\left|\right|}^2{R}_j\left(x_q\right)$ (formula 32)

${\mathrm{\&sigma;}}_j(n+1)={\mathrm{\&sigma;}}_j\left(n\right)-{\mathrm{\&eta;}}_3\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{{\&PartialD;\mathrm{\&sigma;}}_j\left(n\right)}$ (formula 33)

Wherein η ₁, η ₂, η ₃for learning efficiency also can be variable for constant.

Ten, according to the decision table after simplification, choose the corresponding data in exercise risk early warning Sample Storehouse, Training RBF Neural Network, learning method detailed process is as follows:

Step one: according to known class number determination hidden node number, k=3 in this project, asks for the center initial value c of mean value as such of all input vectors of this classification _j(1), using the width initial value σ of the mean value of all for every class input vectors and centre distance as such _j(1), each classification inputs a sample, asks for initial w _1j(1), setting allowable error ε, ε is an a small amount of being greater than zero.Setting learn efficiency η ₁=η ₂=η ₃=1/t, t are iterative loop variable.Setting loop variable t is initially 1, and maximal value is MaxT.

Step 2: input t training sample, obtains network actual output y (t).

Step 3: calculate the actual error ξ (t) exported between desired output, if error ξ (t) < is ε, then this sample does not need to adjust network parameter, jumps to step 6, otherwise enters next step.

Step 4: by formula 28,30,32 calculate and press formula 29,31,33 calculate w _1j(t+1), c _j(t+1), σ _j(t+1).

Step 5: upgrade new network parameter w _1j(t+1), c _j(t+1), σ _j(t+1), t=t+1; If t>MaxT, display does not restrain, and enters step 6, otherwise forwards step 2 to.

Step 6: whole learning process terminates, preserves each parameter of current network;

Step 7: utilize the network parameter preserved to test new athletic damage alarming grade.

11, judge whether yojan result restrains, and whether meets accuracy requirement, if do not meet, modified RBF Neural Networks structure be connected weights, go to step ten;

12, utilize test sample book data, whether the RBF neural that training of judgement is good obtains correct diagnostic result, if do not have, selects other yojan result, go to step nine from step 8;

13, preserve the RBF neural trained, utilize the new sportsman's risk assessment index detected of RBF neural assessment trained.

Step 12 specifically comprises following process:

(1) test sample book data are set up

When setting up exercise risk early warning Sample Storehouse Sample Storehouse, leaving and taking a part of data as test sample book, this ensures that thering the consistance of injury gained in sports Risk-warning grade discrimination, can effectively verify the accuracy of model prediction.Because athlete crowd more than domestic one-level is less, the present embodiment gathers 148 player information altogether, and 131 people's information are used for setting up training sample database, for Training RBF Neural Network; Choose 17 athletes as test sample book, the accuracy of the sick and wounded early warning system warning algorithm of athlete is tested.

(2) data importing

The cause injury typing of the factor of the sick and wounded Risk-warning of Athletes is mainly divided into four parts: individual essential information directly imports when building storehouse, and information can personal information page amendment in movable information management; Physiological and biochemical index adds in monitor message management, amendment; Mostly discrete damage is caused injury, and the sick and wounded risk information page of the individual of factor information in sick and wounded information management adds, amendment; The part of often change in discrete element, the sick and wounded early warning page in injury gained in sports early warning adds and directly submission obtains early warning result.

(3) Risk-warning diagnostic result

The injury gained in sports advanced warning grade of this project to test sample book utilizes actual conditions+quantitative model, actual conditions+quantitative model+expert evaluation and the sick and wounded intelligent early-warning system of athlete to predict respectively, and result is as shown in table 53.Rough Set-Neural Network algorithm is adopted to predict damage risk in the sick and wounded intelligent early-warning system of athlete, the sick and wounded composite algorism determination injury gained in sports risk class that a situation arises, train expertise and the quantification dynamic chain model based on analytical hierarchy process of actual motion in combining information collection in Sample Storehouse construction, therefore after in final table 53, the risk class of two row is completely the same, illustrates that this method reaches designing requirement.

Actual motion is sick and wounded, and a situation arises, coach's expertise and the composite algorism concrete steps based on the linear dynamic chain quantitative model of analytical hierarchy process as follows:

(1) as having injury gained in sports to occur in actual training process, then the data that the damage of this Athletes gathers before occurring are high-risk grade;

(2) for the situation not moving damage, then use according to image data and dynamically connect linear model calculation risk grade, and this athletic injury gained in sports risks and assumptions scoring event is supplied to 15 senior above traineres, team doctor and scientific research personnel simultaneously and carries out risk class marking.

(3) if coach, team doctor and scientific research personnel and marking result are same risk class more than 80%, then using this result as this athletic risk rating result; Otherwise using dynamic chain quantitative model result as risk rating result.

The comparison of table 53 distinct methods exercise risk early warning result

Known by testing above, use rough set & RBF network algorithm to achieve ideal result to injury gained in sports Risk-warning, therefore this algorithm is used for athletic injury gained in sports Risk-warning is feasible.

Certainly, above-mentioned explanation is not limitation of the present invention, and the present invention is also not limited only to above-mentioned citing, and the change that those skilled in the art make in essential scope of the present invention, remodeling, interpolation or replacement, also belong to protection scope of the present invention.

Claims (5)

1. an Athletes damage risk method for early warning, is characterized in that, comprises the steps:

(1) analysis of the injury gained in sports Risk-warning factor;

(2) propose sports injury muscle injury Risk-warning to cause injury factor dynamic chain model;

(3) sportsman's Risk-warning database is set up;

(4) use analytical hierarchy process, set up sports injury muscle injury Risk-warning and to cause injury factor dynamic chain quantitative model;

(5) choose sportsman and the Risk-warning factor according to sports events, utilize the corresponding data in exercise risk database, set up exercise risk early warning Sample Storehouse and the test sample book data of this project;

(6) judge the discreteness of the Risk-warning factor corresponding data in exercise risk early warning Sample Storehouse, if not discrete, call SOM neural network discrete method and sliding-model control is carried out to data;

(7) by the achievement data after utilizing SOM neural network discrete and the Data Synthesis decision table initially with regard to discrete valuation;

(8) according to the yojan process of RS theory, the extraction of attribute reduction Sum fanction is carried out to decision table, obtain one group of yojan result;

(9) RBF neural is built based on one of them decision table simplified selected;

(10) according to the decision table after simplification, the corresponding data in exercise risk early warning Sample Storehouse is chosen, Training RBF Neural Network;

(11) judge whether yojan result restrains, and whether meets accuracy requirement, if do not meet, modified RBF Neural Networks structure be connected weights, go to step (10);

(12) utilize test sample book data, whether the RBF neural that training of judgement is good obtains correct diagnostic result, if do not have, selects other yojan result, go to step (9) from step (8);

(13) preserve the RBF neural trained, utilize the new sportsman's risk assessment index detected of RBF neural assessment trained;

The SOM neural network discrete method of described step (6) comprises the steps:

(6.1) calculate the importance degree of each connection attribute in decision table and sort by order from big to small;

(6.2) each continuous print injury gained in sports is caused injury factor attribute a _i(i=1,2 ..., | C|) a corresponding SOM neural network, the initial value of given SOM neural network output layer neuron number is 3, and input layer number is 1, by the connection weight w of network _jgive [Min (a _i), Max (a _i)] random value in interval, determine the initial value α (0) of learning rate α (t), α (t) slowly should reduce with the increase of t, t representative study number of times, and initial value is 1, total the study number of times T determined;

(6.3) by attribute a _iin numerical value x be input to SOM neural network;

(6.4) calculate input numerical value x and be connected weight w _jbetween distance;

D _j=(x-w _j) ², j=1,2,3 (formula 1)

(6.5) choose and make apart from minimum neuron k as winning node;

(6.6) the connection weights of winning node collection are upgraded

W _j(t+1)=w (t)-α (t) (x-w _j) (formula 2)

(6.7) return step (6.3), choose another data input network, until all mode in sample is supplied to network;

(6.8) renewal learning rate:

$\mathrm{\&alpha;}\left(t\right)=\mathrm{\&alpha;}\left(0\right)[1-\frac{t}{T}];$ (formula 3)

(6.9) make t=t+1, return step (6.2), repeat to t=T;

(6.10) calculate the incompatibility of decision table after discretize, if meet incompatibility threshold requirements, calculate end, otherwise turn next step;

(6.11) attribute of importance sorting first is selected, its SOM neural network output layer neuron number is adjusted to 4, the SOM neural network output layer neuron number that all the other attributes are corresponding is constant, again discretize decision table, and go to step (6.10), until the maximum cluster numbers of each SOM neural network is adjusted to 4, the maximum cluster numbers of each SOM neural network is defined as 4, if maximum its SOM neural network clustering number of attribute of importance degree reaches requirement, the attribute then selecting importance degree to take second place adjusts, by that analogy;

(6.12) final discretization operations is completed.

2. Athletes damage risk method for early warning according to claim 1, is characterized in that, described step (8) adopts carries out Decision Table Reduction based on the clearly demarcated matrix method in rough set.

3. Athletes damage risk method for early warning according to claim 2, is characterized in that, describedly comprises the steps: based on the clearly demarcated matrix method in rough set

(8.1) exercise risk early warning Sample Storehouse data are read in, as decision table;

(8.2) the clearly demarcated Matrix C of this decision table is calculated _d;

(8.3) in clearly demarcated matrix, all values are non-zero, 1, element c _ij, set up corresponding logical expression L _ij, wherein

(formula 4)

I, j are clearly demarcated entry of a matrix element subscript;

(8.4) by all expression formula L that extracts _ijdo conjunction computing, obtain a conjunctive normal form L, namely

(formula 5)

(8.5) form above-mentioned conjunctive normal form being converted to disjunctive normal form is the final yojan result number obtained of i representative;

(8.6) output attribute yojan result, the corresponding attribute reduction result of each conjunct in disjunctive normal form, the set of properties comprised in each conjunct is into about the conditional attribute set after letter.

4. Athletes damage risk method for early warning according to claim 1, it is characterized in that, described step (9) comprises the steps:

(9.1) basis function is chosen

Select Gaussian function as the activation function of hidden layer unit, suppose that input vector is n dimension, have m output node, the number of hidden layer node is k, and the output of hidden layer node is expressed as:

$R_j\left(x\right)=e^{-\frac{{\left|\right|x-c_j\left|\right|}^2}{2{\mathrm{\&sigma;}}_j^2}}$ (formula 6)

Wherein j={1,2 ..., k}, R _jx (), for a jth hidden node is to output, x is that the n of network ties up input vector, c _jbe the kernel function center vector of a jth hidden layer node, it has identical dimension with input vector x, σ _jfor the width of a jth hidden layer node Gaussian function, || x-c _j|| represent the distance measure between input vector and central point, network exports and is defined as:

$y_i=\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{R}_j\left(x\right)$ (formula 7)

Wherein i=1,2 ..., m; y _ibe the output of i-th output layer node, w _ijfor a jth hidden layer node is to the connection weights of i-th output layer node;

(9.2) hidden layer design

The hidden layer of design RBF neural, mainly determine hidden layer number of unit and their excitation function, adopt Gaussian function, Gaussian function is determined by cluster centre and cluster width, the classification of training mode is known, i.e. injury gained in sports Risk-warning grade, therefore a kind of simple mode can be adopted to determine, the i.e. corresponding Gaussian function totally three of each class, i.e. k=3, the width parameter of Gaussian function can get the average of three class samples and its central point distance, and the center of Gaussian function can get the central point of average as respective unit of three class samples respectively,

(9.3) radial basis function center, width and the renewal to the weights between output layer

Radial basis function center and other parameter all experience learning process, usually adopt error correction learning process, application gradient descent method, specific as follows:

Suppose have N number of sample to input, to all input amendment, definition error function:

$\mathrm{\&xi;}=\frac{1}{2}\underset{q=1}{\overset{N}{\mathrm{\&Sigma;}}}{e}_q^2$ (formula 8)

Wherein e _qfor error, be defined as follows:

$e_q=d_q-y\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{1j}{R}_j\left(x_q\right)=d_q-\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{w}_{1j}{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{2{\mathrm{\&sigma;}}_j^2}}$ (formula 9)

Wherein d _qfor sample x _qthe value of required type, the d that the sample of high risk class is corresponding in this project _q=3.

5. Athletes damage risk method for early warning according to claim 1, it is characterized in that, described step (10) comprises the steps:

Step one: according to known class number determination hidden layer node number, k=3 in this project, asks for the center initial value c of mean value as such of all input vectors of this classification _j(1), using the width initial value σ of the mean value of all for every class input vectors and centre distance as such _j(1), each classification inputs a sample, asks for initial w _1j(1), setting allowable error ε, ε is an a small amount of setting learn efficiency η being greater than zero ₁=η ₂=η ₃=1/t, t are iterative loop variable, and setting loop variable t is initially 1, and maximal value is MaxT;

Step 2: input t training sample, obtains network actual output y (t);

Step 3: calculate the actual error ξ (t) exported between desired output, if error ξ (t) < is ε, then this sample does not need to adjust network parameter, jumps to step 6, otherwise enters next step;

Step 4: by formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;w_{1j}\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q{e}^{-\frac{{\left|\right|x_q-c_j\left|\right|}^2}{2{\mathrm{\&sigma;}}_j^2}}---\left(29\right)$

formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;c_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)R_j\left(x_q\right)=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^2}{e}^{-\frac{{\left|\right|x_q-c_j\left(n\right)\left|\right|}^2}{2{\mathrm{\&sigma;}}_j^2}}(x_q-c_j\left(n\right))---\left(31\right)$

formula $\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;{\mathrm{\&sigma;}}_j\left(n\right)}=-\underset{q=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}_q\left(n\right)\frac{w_{1j}\left(n\right)}{{\mathrm{\&sigma;}}_j^3\left(n\right)}{\left|\right|x_q-c_j\left|\right|}^2{R}_j\left(x_q\right)---\left(33\right)$

Calculate and press formula

$w_{1j}(n+1)=w_{1j}\left(n\right)-{\mathrm{\&eta;}}_1\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;w_{1j}}---\left(30\right)$

formula $c_j(n+1)=c_j\left(n\right)-{\mathrm{\&eta;}}_2\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;c_j}---\left(32\right)$

formula ${\mathrm{\&sigma;}}_j(n+1)={\mathrm{\&sigma;}}_j\left(n\right)-{\mathrm{\&eta;}}_3\frac{\&PartialD;\mathrm{\&xi;}\left(n\right)}{\&PartialD;{\mathrm{\&sigma;}}_j\left(n\right)}---\left(34\right)$ Calculate w _1j(t+1), c _j(t+1), σ _j(t+1);

Step 5: upgrade new network parameter w _1j(t+1), c _j(t+1), σ _j(t+1), t=t+1; If t>MaxT, display does not restrain, and enters step 6, otherwise forwards step 2 to;

Step 6: whole learning process terminates, preserves each parameter of current network.