CN106355034A - Sub-health prediction method and prediction system based on incremental neural network model - Google Patents

Abstract

The invention discloses a sub-health prediction method based on an incremental neural network model. The sub-health prediction method comprises the following steps that a sub-health daily data database is established; the neural network model is trained; daily life data is acquired and sent to a server; current-day data is extracted from a daily data record sheet of a user to form n-dimensional vectors, normalization processing is performed, and then the vectors are input into the neural network model of sub-health pathology to perform sub-health danger level possibility prediction; an intelligent household sub-health nursing device judges whether a sub-health danger level value W is greater than 03 or not; when the user receives warning of a warning indicator, the user goes to a hospital by himself/herself for examination and transmits an examination result to the server, and the server judges whether the examination result is correct or not; when the examination result is wrong, an incremental algorithm is executed, and dynamic correction is conducted on the neural network model. The sub-health prediction method is accurate in prediction, and the neural network model is customized for each user.

Description

A kind of subhealth state Forecasting Methodology based on increment type neural network model and prognoses system

Technical field

The invention belongs to field of medical technology, more particularly to a kind of subhealth state based on increment type neural network model is pre- Survey method and prognoses system.

Background technology

Currently domestic each health management system arranged be respectively provided with subhealth state prediction and evaluation, its use prediction mode be data Join.Its principle is that by system matches fixed data and then personal lifestyle data entry system is shown ill probability.But due to people The complexity of body and disease, unpredictability, in the form of expression of bio signal and information, Changing Pattern (Self-variation with Change after medical intervention) on, it is detected and signal representation, the data of acquisition and the analysis of information, decision-making etc. are all multi-party All there is extremely complex non-linear relationship in face.So using traditional Data Matching can only be blindness data examination it is impossible to Judge the logic association between data and data and variable, the codomain deviation obtaining is big, causes the specificity ten of system prediction Point poor, domestic health management system arranged effectively Accurate Prediction cannot be carried out to personal subhealth state so current.

Most of before this is all using bp neural network model to subhealth state prediction, but when new detection data generation When it is necessary to train neural network model again, operation efficiency is extremely low.And after system user scale increases, server Will be unable to complete in time training mission.

Content of the invention

The purpose of the present invention is that and overcomes the deficiencies in the prior art, there is provided one kind is based on increment type neural network model Subhealth state Forecasting Methodology and prognoses system, the present invention by neural network model train predict a large amount of patient in hospital pathology numbers According to, find subhealth state pathology and subhealth state earlier life variations in detail, clinical symptoms, examination criteria value, high-risk group's feature, this Logic association between several causes of disease and variable, ultimately form the subhealth state pathology nerve to subhealth state illness probability Accurate Prediction Network model, the present invention passes through to gather user's daily life data, the periodicity of its data of active analysis, regular eventually through Subhealth state pathology Neural Network model predictive user suffers from subhealth state probability, immediately reminds user in the way of visual effect Doctor, constantly revises neural network model when Neural Network model predictive is inaccurate by increasable algorithm, to set for each Standby user sets up the neural network model training for this user, with the increase of use time, to set up to this customer volume Body neural network model customized, accuracy rate is greatly improved.

To achieve these goals, the invention provides a kind of subhealth state prediction side based on increment type neural network model Method, comprises the steps:

Step (1), acquisition hospital's subhealth state etiology and pathology data source and the daily monitoring data of patient, thus set up subhealth state Daily data database；

Step (2), the daily data database of subhealth state set up according to step (1) are off-line manner to neutral net mould Type is trained, to obtain the subhealth state pathology neural network model training；

Step (3), by intelligent monitoring device, the daily life data of user is acquired, and will collection daily life Live data sends to server, and server preserves the daily life data of user to the daily data logger of user；

Step (4), from the daily data logger of user, extract same day data, form n-dimensional vector, and n-dimensional vector is done Carry out subhealth state degree of danger general in the subhealth state pathology neural network model training in input step (2) after normalized Rate is predicted, obtains subhealth state probability results array p, and server is by corresponding for maximum probability in array p subhealth state degree of danger value w Send wired home subhealth state care appliances to；

After step (5), subhealth state degree of danger value w of wired home subhealth state care appliances the reception server transmission, sentence Whether disconnected subhealth state degree of danger value w is more than or equal to 3, and if greater than equal to 3, then attention device warns to remind user, if little In 3, then attention device does not warn；

Step (6), when user receive attention device warning when, user voluntarily removes examination in hospital, and inspection result is passed through Wired home subhealth state care appliances send back server, and server judges whether inspection result is correct, if inspection result is wrong By mistake, then explanation subhealth state pathology Neural Network model predictive is inaccurate, if inspection result is correct, subhealth state pathology god is described Through network model's prediction accurately；

Step (7), when inspection result mistake, from the daily data logger of user extract m days in record preserve to In incremental data table, when the record quantity in incremental data table is more than h bar, execute increasable algorithm, to subhealth state pathology god Carry out dynamic corrections through network model；

Step (8), repeat step (3)～(7).

Further, the input layer of neural network model is n node, and hidden layer number is n*2+1, and output layer is 1 Node, extracts k bar record from subhealth state daily data database table and is trained, every record is a n-dimensional vector, owns Data before use first through normalized so as to numerical value is interval in [0,1], then execution following steps are to neural network model It is trained:

1) one n-dimensional vector of input, to neural network model, calculates all of weight vector in neural network model defeated to this Enter the distance of n-dimensional vector, closest neuron is as won neuron, its computing formula is as follows:

$\left|\right|x_i-w_k\left|\right|=\underset{j}{\min }\left(\right||x_i-w_j|\left|\right);$

Wherein: w_kIt is the weight vector of triumph neuron, | | ... | | for Euclidean distance；

2) weight vector of the neuron in adjustment triumph neuron and triumph neuron field, formula is as follows:

Wherein: w_jT () is neuron；w_j(t+1) weight vector before being adjustment and after adjustment；J belongs to triumph neuron neck Domain；α (t) is learning rate, and it is as the function that the increase of iterationses is gradually successively decreased, and span is [0 1], through multiple It is 0.71 that Optimal learning efficiency is chosen in experiment；d_jIt is the distance of neuron j and triumph neuron；σ (t) is as the letter that the time successively decreases Number；Iteration all input n-dimensional vectors is input in neural network model and is trained each time, when the iteration reaching regulation After number of times, neural network model training terminates.

Further, inspection result is sent back the form of the object information of server by wired home subhealth state care appliances For: { the subhealth state degree of danger value of the actual judgement of doctor }, whether server, after receiving object information, judges inspection result Correctly.

Further, the increasable algorithm carrying out dynamic corrections to subhealth state pathology neural network model is:

Every vector v { v in incremental data table₁,v₂,…,v_n, it is sent in neural network model learning function Row study, learning procedure is as follows:

1) first to output layer, each weight vector is assigned little random number and is done normalized, then utilizes input mode vector v Meansigma methodss avg (v), be initialized as the weights of unique neuron in the 0th layer of neural network model, and be set to win nerve Unit, calculates its quantization error qe；

2) expand out 2 × 2 structures som from the 0th layer of neuron, and its level identities layer is set to 1；

3) for each 2 × 2 structure som subnet expanded out in layer layer, initialize the power of this 4 neurons Value；The input vector set ci of i-th neuron is set to sky, main label is set to null, the main label ratio r of neuron i_i It is set to 0；The abnormity early warning data vector v of new som inherits the triumph input vector set vx of his father's neuron；

4) select vector v x from vx_iDo following judgement:

If vx_iFor the data of not tape label, then calculate its Euclidean distance with each neuron, chosen distance is the shortest Neuron is as triumph neuron；

If vx_iFor the data of tape label, then select main label and vx_iLabel is identical and r_iThe maximum neuron of value is made For triumph neuron, update this triumph neuron main label；

If can not find main label and vx_iLabel identical neuron, then find and vx_iClosest neuron i makees For triumph neuron；

5) weights of neuron in triumph neuron and its neighborhood are adjusted, update the vectorial set w=w ∪ that wins {vx_i, calculate main label, the main label ratio r of triumph neuron_iWith comentropy e_iIf. not up to predetermined frequency of training, Go to step 4)；

6) quantization error qe of each neuron in this neural network model after calculating is adjusted_i, neuronal messages entropy e_i With the average quantization error mqe of subnet, formula is as follows:

${\mathrm{qe}}_i=\underset{x_j\&element;c_i}{\σ}\left\{\right||w_i-x_j|\left|\right\};$

Wherein: w_iFor the weight vector of neuron i, c_iThe set constituting for all input vectors being mapped to neuron i；

$e_i={\σ}_{i\&element;t}(-lb\frac{n_i}{m})\×\frac{n_i}{m};$

Wherein: n_iRepresent to fall that label is the number of samples of i on neuron, m represents to fall label data on neuron Sum, t represents to fall the sample label species set on neuron；

Then judge:

If mqe > qe × threshold value q of father node, wherein q=0.71, then insert a line neuron in this som, turn step Rapid 4)；

If e_i> father node e_i× threshold value p, wherein p=0.42, then grow one layer of new subnet from this neuron, will The subnet newly growing increases in the subnet queue of layer+1 layer；

If being not inserted into new neuron in som also do not grow new subnet, illustrate that the training of this subnet completes；

7) for all 2 × 2 structures som of the layer+1 layer newly expanded out, iteration operating procedure 3)～5) to it again It is trained, until neural network model no longer produces new neuron and new layering, whole training terminates.

Further, in step (4), subhealth state probability results array p obtaining is one 6 dimension variable, its 6 dimension variable It is respectively 6 subhealth state degree of danger probability, the corresponding subhealth state of maximum probability in 6 subhealth state degree of danger probability is dangerous Degree value w sends wired home subhealth state care appliances to；In step (5), represent when subhealth state degree of danger value w=1 Ideal, represents normal during w=2, represent slight subhealth state during w=3, represents moderate subhealth state during w=4, represents dangerous during w=5, Represent abnormally dangerous during w=6.

Further, if user includes health check-up by other means and checks oneself, learn that oneself has suffered from subhealth state, and intelligence The attention device of family's subhealth state care appliances do not warn then it represents that wired home subhealth state care appliances judge inaccurate, this When execution step (6)～(7), wired home subhealth state care appliances are sent to object information on server.

Present invention also offers a kind of prognoses system of described subhealth state Forecasting Methodology, including intelligent monitoring device, intelligence Device data acquisition device, server and wired home subhealth state care appliances, described intelligent monitoring device and described smart machine Data acquisition unit is connected, and described smart machine data acquisition unit passes through communication device one and described server network communication, institute State wired home subhealth state care appliances and pass through communication device two and described server network communication.

Further, described wired home subhealth state care appliances are provided with attention device.

Further, described intelligent monitoring device includes Intelligent worn device, Intelligent water cup, Intelligent weight claim, intelligent horse Bucket and Intelligent light sensing equipment.

Beneficial effects of the present invention:

1st, the present invention by neural network model train predict a large amount of patient in hospital pathological data, find subhealth state pathology with Subhealth state earlier life variations in detail, clinical symptoms, examination criteria value, high-risk group's feature, the logic between this several causes of disease Association and variable, ultimately form the subhealth state pathology neural network model to subhealth state illness probability Accurate Prediction, and the present invention is led to Cross collection user's daily life data, the periodicity of its data of active analysis, regularity are eventually through subhealth state pathology nerve net Network model prediction user suffers from subhealth state degree of danger probability, reminds user's instant hospitalizing and prevention in the way of visual effect.

2nd, when Neural Network model predictive is inaccurate, neural network model is constantly revised, to be directed to by increasable algorithm Each equipment user sets up the neural network model training for this user, with the increase of use time, to set up to this The neural network model that user makes to measure, accuracy rate is greatly improved.

Brief description

In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing Have technology description in required use accompanying drawing be briefly described it should be apparent that, drawings in the following description be only this Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, acceptable Other accompanying drawings are obtained according to these accompanying drawings.

Fig. 1 is the flow chart of the embodiment of the present invention.

Specific embodiment

Below in conjunction with the accompanying drawings invention is further illustrated, but be not limited to the scope of the present invention.

Embodiment

As shown in figure 1, a kind of subhealth state Forecasting Methodology based on increment type neural network model that the present invention provides, including Following steps:

Step (1), acquisition hospital's subhealth state etiology and pathology data source and the daily monitoring data of patient, thus set up subhealth state Daily data database；

Wherein daily monitoring data is 21 item data, and its 21 item data is the age, sex, heart rate, body fat, and systolic pressure is relaxed Open and press, smoking capacity (daily), the amount of drinking (daily), the amount of drinking water and frequency, body weight, total urination time, urine color, times of defecation, Whether constipation, bmi index, body temperature, pursue an occupation, the length of one's sleep and quality, 21 item data such as travel distance (daily), the present invention Set up 21 dimensional vectors with 21 item data；

Step (2), the daily data database of subhealth state set up according to step (1) are off-line manner to neutral net mould Type is trained, to obtain the subhealth state pathology neural network model training；

Step (3), by intelligent monitoring device, the daily life data of user is acquired, and will collection daily life Live data sends to server, and server preserves the daily life data of user to the daily data logger of user；

Step (4), from the daily data logger of user, extract same day data, form n-dimensional vector, and n-dimensional vector is done Carry out subhealth state degree of danger general in the subhealth state pathology neural network model training in input step (2) after normalized Rate is predicted, obtains subhealth state probability results array p, and server is by corresponding for maximum probability in array p subhealth state degree of danger value w Send wired home subhealth state care appliances to；

Wherein in step (4), subhealth state probability results array p obtaining is one 6 dimension variable, and its 6 dimension variable is respectively For 6 subhealth state degree of danger probability, maximum probability corresponding subhealth state degree of danger in 6 subhealth state degree of danger probability Value w sends wired home subhealth state care appliances to.

After step (5), subhealth state degree of danger value w of wired home subhealth state care appliances the reception server transmission, sentence Whether disconnected subhealth state degree of danger value w is more than or equal to 3, and if greater than equal to 3, then attention device warns to remind user, if little In 3, then attention device does not warn；

Wherein in step (5), represent preferable when subhealth state degree of danger value w=1, during w=2, represent normal, w=3 When represent slight subhealth state, represent moderate subhealth state during w=4, represent dangerous during w=5, represent abnormally dangerous during w=6.I.e. originally Subhealth state degree of danger is divided into 6 grades by invention, more accurate and visual.

Step (6), when user receive attention device warning when, user voluntarily removes examination in hospital, and inspection result is passed through Wired home subhealth state care appliances send back server, and server judges whether inspection result is correct, if inspection result is wrong By mistake, then explanation subhealth state pathology Neural Network model predictive is inaccurate, if inspection result is correct, subhealth state pathology god is described Through network model's prediction accurately；

The form that wherein inspection result is sent back the object information of server by wired home subhealth state care appliances is: { the subhealth state degree of danger value of the actual judgement of doctor }, just whether server, after receiving object information, judge inspection result Really.

Step (7), when inspection result mistake, from the daily data logger of user extract 7 days in record preserve to In incremental data table, when the record quantity in incremental data table is more than 100, execute increasable algorithm, to subhealth state pathology Neural network model carries out dynamic corrections；

Step (8), repeat step (3)～(7).

The input layer of the neural network model of the present invention is 21 nodes, and hidden layer number is 43, and output layer is 6 nodes (i.e. 6 subhealth state degree of danger probability), extracts 400000 records from subhealth state daily data database table and is trained, Every record is 21 dimensional vectors, all data before use first through normalized so as to numerical value is interval in [0,1], so Execution following steps are trained to neural network model afterwards: 1) one 21 dimensional vector of input, to neural network model, calculate god Arrive through weight vector all of in network model

The distance of this input 21 dimensional vector, closest neuron is as won neuron, and its computing formula is as follows:

$\left|\right|x_i-w_k\left|\right|=\underset{j}{\min }\left(\right||x_i-w_j|\left|\right);$

Wherein: w_kIt is the weight vector of triumph neuron, | | ... | | for Euclidean distance；

2) weight vector of the neuron in adjustment triumph neuron and triumph neuron field, formula is as follows:

Wherein: w_jT () is neuron；w_j(t+1) weight vector before being adjustment and after adjustment；J belongs to triumph neuron neck Domain；α (t) is learning rate, and it is as the function that the increase of iterationses is gradually successively decreased, and span is [0 1], through multiple It is 0.71 that Optimal learning efficiency is chosen in experiment；d_jIt is the distance of neuron j and triumph neuron；σ (t) is as the letter that the time successively decreases Number；Iteration all input n-dimensional vectors is input in neural network model and is trained each time, when the iteration reaching regulation After number of times, neural network model training terminates.

The increasable algorithm carrying out dynamic corrections to subhealth state pathology neural network model of the present invention is:

Every vector v { v in incremental data table₁,v₂,…,v_n, it is sent in neural network model learning function Row study, learning procedure is as follows:

1) first to output layer, each weight vector is assigned little random number and is done normalized, then utilizes input mode vector v Meansigma methodss avg (v), be initialized as the weights of unique neuron in the 0th layer of neural network model, and be set to win nerve Unit, calculates its quantization error qe；

2) expand out 2 × 2 structures som from the 0th layer of neuron, and its level identities layer is set to 1；

3) for each 2 × 2 structure som subnet expanded out in layer layer, initialize the power of this 4 neurons Value；The input vector set ci of i-th neuron is set to sky, main label is set to null, the main label ratio r of neuron i_i It is set to 0；The abnormity early warning data vector v of new som inherits the triumph input vector set vx of his father's neuron；

4) select vector v x from vx_iDo following judgement:

If vx_iFor the data of not tape label, then calculate its Euclidean distance with each neuron, chosen distance is the shortest Neuron is as triumph neuron；

If vx_iFor the data of tape label, then select main label and vx_iLabel is identical and r_iThe maximum neuron of value is made For triumph neuron, update this triumph neuron main label；

If can not find main label and vx_iLabel identical neuron, then find and vx_iClosest neuron i makees For triumph neuron；

5) weights of neuron in triumph neuron and its neighborhood are adjusted, update the vectorial set w=w ∪ that wins {vx_i, calculate main label, the main label ratio r of triumph neuron_iWith comentropy e_iIf. not up to predetermined frequency of training, Go to step 4)；

6) quantization error qe of each neuron in this neural network model after calculating is adjusted_i, neuronal messages entropy e_i With the average quantization error mqe of subnet, formula is as follows:

${\mathrm{qe}}_i=\underset{x_j\&element;c_i}{\σ}\left\{\right||w_i-x_j|\left|\right\};$

Wherein: w_iFor the weight vector of neuron i, c_iThe set constituting for all input vectors being mapped to neuron i；

$e_i={\σ}_{i\&element;t}(-lb\frac{n_i}{m})\×\frac{n_i}{m};$

Wherein: n_iRepresent to fall that label is the number of samples of i on neuron, m represents to fall label data on neuron Sum, t represents to fall the sample label species set on neuron；

Then judge:

If mqe > qe × threshold value q of father node, wherein q=0.71, then insert a line neuron in this som, turn step Rapid 4)；

If e_i> father node e_i× threshold value p, wherein p=0.42, then grow one layer of new subnet from this neuron, will The subnet newly growing increases in the subnet queue of layer+1 layer；

If being not inserted into new neuron in som also do not grow new subnet, illustrate that the training of this subnet completes；

7) for all 2 × 2 structures som of the layer+1 layer newly expanded out, iteration operating procedure 3)～5) to it again It is trained, until neural network model no longer produces new neuron and new layering, whole training terminates.

If user includes health check-up by other means and checks oneself, learn that oneself has suffered from subhealth state, and wired home is sub- strong The attention device of health care appliances does not warn then it represents that wired home subhealth state care appliances judge inaccurate, now executes step Suddenly (6)～(7), wired home subhealth state care appliances are sent to object information on server.

Present invention also offers a kind of prognoses system of described subhealth state Forecasting Methodology, including intelligent monitoring device, intelligence Device data acquisition device, server and wired home subhealth state care appliances, described intelligent monitoring device and described smart machine Data acquisition unit is connected, and described smart machine data acquisition unit passes through communication device one and described server network communication, institute State wired home subhealth state care appliances and pass through communication device two and described server network communication.

It is provided with attention device on the described wired home subhealth state care appliances of the present invention.

The described intelligent monitoring device of the present invention includes Intelligent worn device, Intelligent water cup, Intelligent weight claim, intelligent closestool With Intelligent light sensing equipment etc..

The present invention is trained by neural network model and predicts a large amount of patient in hospital pathological data, finds subhealth state pathology and Asia Healthy earlier life variations in detail, clinical symptoms, examination criteria value, high-risk group's feature, the logic between this several causes of disease is closed Connection and variable, ultimately form the subhealth state pathology neural network model to subhealth state illness probability Accurate Prediction, and the present invention passes through Collection user's daily life data, the periodicity of its data of active analysis, regularity are eventually through subhealth state pathology neutral net Model prediction user suffers from subhealth state probability, reminds user's instant hospitalizing and prevention in the way of visual effect.

All data of the present invention preserve to server, can significantly save calculating cost, hardware configuration is low, thus selling Valency is also low.

The present invention carries communication device one and communication device two, by wifi from the Internet that is dynamically connected, and can protect for a long time Hold online.Various intelligent monitoring devices can easily access present device by modes such as network or bluetooths, sets in acquisition The daily life data of the monitoring of intelligent monitoring device, the data that therefore present device obtains can automatically be uploaded after standby mandate It is real-time, accurate, polynary.

Because everyone physical trait is different, the data characteristicses being shown during subhealth state morbidity also can be different.Cause This is not conventional high by the method accuracy rate of neural network prediction subhealth state.The present invention is directed to each equipment user and sets up training Go out the neural network model for this user, running after a period of time, by producing to measure neutral net is being made to this user Forecast model, accuracy rate is greatly improved.

When neural network model is judged by accident, error message be will be feedbacked to server by wired home subhealth state care appliances, For this user's dynamic corrections neural network model, when similar characteristics data in this user next, will not judge by accident again.Cause This, with the increase of use time, the judgement of the wired home subhealth state care appliances of the present invention will be more and more accurate.

Ultimate principle, principal character and the advantages of the present invention of the present invention have been shown and described above.The technology of the industry , it should be appreciated that the present invention is not restricted to the described embodiments, the simply explanation described in above-described embodiment and description is originally for personnel Invention principle, without departing from the spirit and scope of the present invention the present invention also have various changes and modifications, these change Change and improvement both falls within scope of the claimed invention.Claimed scope by appending claims and its Equivalent defines.

Claims (9)

1. a kind of subhealth state Forecasting Methodology based on increment type neural network model is it is characterised in that comprise the steps:

Step (1), obtain hospital subhealth state and cure the disease etiology and pathology data source and the daily monitoring data of patient, thus setting up subhealth state Daily data database；

Step (2), off-line manner neural network model is entered according to the daily data database of subhealth state that step (1) is set up Row training, to obtain the subhealth state pathology neural network model training；

Step (3), by intelligent monitoring device, the daily life data of user is acquired, and will collection daily life number According to sending to server, server preserves the daily life data of user to the daily data logger of user；

Step (4), from the daily data logger of user, extract same day data, form n-dimensional vector, and normalizing is done to n-dimensional vector Carry out subhealth state degree of danger probability pre- in the subhealth state pathology neural network model training in input step (2) after change process Survey, obtain subhealth state probability results array p, corresponding for maximum probability in array p subhealth state degree of danger value w is transmitted by server To wired home subhealth state care appliances；

After step (5), subhealth state degree of danger value w of wired home subhealth state care appliances the reception server transmission, judge sub- Whether health risk degree value w is more than or equal to 3, and if greater than equal to 3, then attention device warns to remind user, if less than 3, Then attention device does not warn；

Step (6), when user receive attention device warning when, user voluntarily removes examination in hospital, and by inspection result pass through intelligence Family's subhealth state care appliances send back server, and server judges whether inspection result is correct, if inspection result mistake, Illustrate that subhealth state pathology Neural Network model predictive is inaccurate, if inspection result is correct, subhealth state pathology nerve net is described Network model prediction is accurate；

Step (7), when inspection result mistake, from the daily data logger of user extract m days in record preserve to increment In tables of data, when the record quantity in incremental data table is more than h bar, execute increasable algorithm, to subhealth state pathology nerve net Network model carries out dynamic corrections；

Step (8), repeat step (3)～(7).

2. a kind of subhealth state Forecasting Methodology based on increment type neural network model according to claim 1, its feature exists In the input layer of neural network model is n node, and hidden layer number is n*2+1, and output layer is 1 node, from subhealth state day Extract k bar record in regular data database table to be trained, every record is a n-dimensional vector, all data are first before use Through normalized so as to numerical value is interval in [0,1], then execution following steps are trained to neural network model:

1) one n-dimensional vector of input, to neural network model, calculates all of weight vector in neural network model and ties up to this input n The distance of vector, closest neuron is as won neuron, and its computing formula is as follows:

Wherein: w_kIt is the weight vector of triumph neuron, | | ... | | for Euclidean distance；

2) weight vector of the neuron in adjustment triumph neuron and triumph neuron field, formula is as follows:

Wherein: w_jT () is neuron；w_j(t+1) weight vector before being adjustment and after adjustment；J belongs to triumph neuron field；α T () is learning rate, it is as the function that the increase of iterationses is gradually successively decreased, and span is [0 1], through many experiments Choosing Optimal learning efficiency is 0.71；d_jIt is the distance of neuron j and triumph neuron；σ (t) is as the function that the time successively decreases； Iteration all input n-dimensional vectors is input in neural network model and is trained each time, when the iteration time reaching regulation After number, neural network model training terminates.

3. a kind of subhealth state Forecasting Methodology based on increment type neural network model according to claim 1, its feature exists In the form that inspection result is sent back the object information of server by wired home subhealth state care appliances is: { doctor is actual to be sentenced Disconnected subhealth state degree of danger value }, server, after receiving object information, judges whether inspection result is correct.

4. a kind of subhealth state Forecasting Methodology based on increment type neural network model according to claim 1, its feature exists In the increasable algorithm carrying out dynamic corrections to subhealth state pathology neural network model is:

Every vector v { v in incremental data table₁,v₂,…,v_n, it is sent in neural network model learning function and learned Practise, learning procedure is as follows:

1) first to output layer, each weight vector is assigned little random number and is done normalized, then utilizes the flat of input mode vector v Average avg (v), is initialized as the weights of unique neuron in the 0th layer of neural network model, and is set to triumph neuron, meter Calculate its quantization error qe；

2) expand out 2 × 2 structures som from the 0th layer of neuron, and its level identities layer is set to 1；

3) for each 2 × 2 structure som subnet expanded out in layer layer, the weights of this 4 neurons are initialized；Will The input vector set ci of i-th neuron is set to sky, and main label is set to null, the main label ratio r of neuron i_iIt is set to 0；The abnormity early warning data vector v of new som inherits the triumph input vector set vx of his father's neuron；

4) select vector v x from vx_iDo following judgement:

If vx_iFor the data of not tape label, then calculate its Euclidean distance with each neuron, chosen distance nerve the shortest Unit is as triumph neuron；

If vx_iFor the data of tape label, then select main label and vx_iLabel is identical and r_iThe maximum neuron of value is as obtaining Victory neuron, updates this triumph neuron main label；

If can not find main label and vx_iLabel identical neuron, then find and vx_iClosest neuron i is as obtaining Victory neuron；

5) weights of neuron in triumph neuron and its neighborhood are adjusted, update the vectorial set w=w ∪ { vx that wins_i, Calculate main label, the main label ratio r of triumph neuron_iWith comentropy e_iIf. not up to predetermined frequency of training, go to step 4)；

6) quantization error qe of each neuron in this neural network model after calculating is adjusted_i, neuronal messages entropy e_iAnd son The average quantization error mqe of net, formula is as follows:

Wherein: w_iFor the weight vector of neuron i, c_iThe set constituting for all input vectors being mapped to neuron i；

Wherein: n_iRepresent to fall that label is the number of samples of i on neuron, m represents to fall the total of label data on neuron Number, t represents to fall the sample label species set on neuron；

Then judge:

If mqe > qe × threshold value q of father node, wherein q=0.71, then insert a line neuron in this som, go to step 4)；

If e_i> father node e_i× threshold value p, wherein p=0.42, then grow one layer of new subnet from this neuron, will be new The subnet growing increases in the subnet queue of layer+1 layer；

If being not inserted into new neuron in som also do not grow new subnet, illustrate that the training of this subnet completes；

7) for all 2 × 2 structures som of the layer+1 layer newly expanded out, iteration operating procedure 3)～5) it is re-started Training, until neural network model no longer produces new neuron and new layering, whole training terminates.

5. a kind of subhealth state Forecasting Methodology based on increment type neural network model according to claim 1, its feature exists In, in step (4), subhealth state probability results array p obtaining is one 6 dimension variable, and its 6 dimension variable is respectively 6 Asias and is good for Health degree of danger probability, sends maximum probability corresponding subhealth state degree of danger value w in 6 subhealth state degree of danger probability to Wired home subhealth state care appliances；In step (5), represent preferable when subhealth state degree of danger value w=1, table during w=2 Show normal, represent slight subhealth state during w=3, during w=4, represent moderate subhealth state, represent dangerous during w=5, represent non-during w=6 Often dangerous.

6. a kind of subhealth state Forecasting Methodology based on increment type neural network model according to claim 1, its feature exists In, if user includes health check-up by other means and checks oneself, learn that oneself has suffered from subhealth state, and the nursing of wired home subhealth state The attention device of equipment do not warn then it represents that wired home subhealth state care appliances judge inaccurate, now execution step (6)～ (7), wired home subhealth state care appliances are sent to object information on server.

7. a kind of prognoses system of subhealth state Forecasting Methodology described in employing claim 1～7 is supervised it is characterised in that including intelligence Control equipment, smart machine data acquisition unit, server and wired home subhealth state care appliances, described intelligent monitoring device and institute State smart machine data acquisition unit to be connected, described smart machine data acquisition unit passes through communication device one and described server net Network communicates, and described wired home subhealth state care appliances pass through communication device two and described server network communication.

8. according to claim 8 the prognoses system of subhealth state Forecasting Methodology it is characterised in that described wired home subhealth state Attention device is provided with care appliances.

9. according to claim 8 the prognoses system of subhealth state Forecasting Methodology it is characterised in that described intelligent monitoring device bag Include Intelligent worn device, Intelligent water cup, Intelligent weight claim, intelligent closestool and Intelligent light sensing equipment.