CN106779084B - Machine learning system and method - Google Patents
Machine learning system and method Download PDFInfo
- Publication number
- CN106779084B CN106779084B CN201610881058.0A CN201610881058A CN106779084B CN 106779084 B CN106779084 B CN 106779084B CN 201610881058 A CN201610881058 A CN 201610881058A CN 106779084 B CN106779084 B CN 106779084B
- Authority
- CN
- China
- Prior art keywords
- data
- neural network
- model
- processor
- scheme
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Medical Treatment And Welfare Office Work (AREA)
- Electrically Operated Instructional Devices (AREA)
Abstract
This patent application discloses machine learning systems, including being used to acquire the cloud server to the client of rehabilitation data and with client by network connection;The client includes the first processor for being provided with improved intelligent algorithm;The cloud server includes the second processor for being provided with BP neural network model, wherein for second processor by the initial parameter value of improved intelligent algorithm Optimized BP Neural Network while waiting for rehabilitation data by BP neural network model treatment, the initial parameter value includes initial weight and initial threshold.The machine learning system of present patent application has faster pace of learning and study precision.
Description
Technical field
The present invention relates to health and fitness information interaction fields, and in particular to a kind of machine learning system.
Background technique
Machine learning (Machine Learning, ML) is a multi-field cross discipline, be related to probability theory, statistics,
The multiple subjects such as Approximation Theory, convextiry analysis, algorithm complexity theory.Specialize in the study that the mankind were simulated or realized to computer how
Behavior reorganizes the existing structure of knowledge and is allowed to constantly improve the performance of itself to obtain new knowledge or skills.Engineering
Habit is the core of artificial intelligence, is the fundamental way for making computer have intelligence, and application spreads the every field of artificial intelligence.
China human mortality aging in recent years accelerates, and as the fast-developing people's lives rhythm in epoch constantly adds
Fastly, life stress is continuously increased, and consequent is the growth of various chronic disease disease incidence and the appearance of New Type of Diseases, this
Serious negative effect is brought to the physical and mental health of people, the stabilization of society and development.In order to facilitate people rapidly
It obtains the health and fitness information situation of itself and can be improved in daily life for the health and fitness information situation, newly developed now
Some information exchange platforms for health control are gone out.Because intelligently to judge the health status of each different people and go back
It needs to identify the New Type of Diseases that judgement continuously emerges, is both provided in these information exchange platforms for making information exchange platform more
Add intelligent machine learning system.
BP neural network is the important component of machine learning system.BP (Back Propagation) neural network,
That is the learning process of error-duration model error backpropagation algorithm, by the forward-propagating of information and two processes of backpropagation of error
Composition.Each neuron of input layer is responsible for receiving from extraneous input information, and passes to each neuron of middle layer;Middle layer is
Internal information process layer, be responsible for information transformation, according to the demand of information change ability, middle layer can be designed as single hidden layer or
More hidden layer configurations;The last one hidden layer is transmitted to the information of each neuron of output layer, after further treatment after, complete primary study
Forward-propagating treatment process, by output layer outwardly output information processing result.When reality output and desired output are not inconsistent,
Into the back-propagation phase of error.Error corrects each layer weight in the way of error gradient decline by output layer, to hidden
Layer, the layer-by-layer anti-pass of input layer.Information forward-propagating and error back propagation process in cycles are that each layer weight constantly adjusts
Process and neural network learning training process, this process be performed until network output error be reduced to and can connect
Until the degree received or preset study number.
The initial weight and threshold value of existing BP neural network are typically all to use random initializtion for the area [- 0.5,0.5]
Between random number, in fact and accurately can not obtain or predict specific initial weight and initial threshold.And BP neural network
Initial weight and initial threshold are very big on network training influence, can directly affect the convergence and convergence speed of BP neural network
Degree, and then influence study speed and the study accuracy of entire machine learning system.This makes the machine learning system built
It unites unpredictable, there is very big security risk.
It is badly in need of developing a kind of machine learning system that can obtain determining initiation parameter value now.
Summary of the invention
It is not true to solve existing machine learning system initiation parameter value the invention is intended to provide a kind of machine learning system
Fixed problem.
Machine learning system in this programme, including being used to acquire to the client of rehabilitation data and passing through net with client
The cloud server of network connection;The client includes the first processor for being provided with improved intelligent algorithm;The cloud clothes
Business device includes the second processor for being provided with BP neural network model, and wherein second processor is by BP neural network model
Pass through the initial parameter value of improved intelligent algorithm Optimized BP Neural Network, the initial parameter value while managing to rehabilitation data
Including initial weight and initial threshold.
Explanation of nouns:
Improved intelligent algorithm: refer to and calculated in such as artificial neural network, genetic algorithm, simulated annealing, TABU search, ant colony
The basis of the intelligent algorithms such as method, particle flow algorithm carries out parameter optimization, more accurately can simulate or disclose certain natures
The symptom or process on boundary's (living nature) and establish be solve challenge thinking and means, thought and content are related to counting
, physics, biological evolution, artificial intelligence, neural network and statistical mechanics etc..
To rehabilitation data: the set for all data that the people of non-rehabilitation inputs to machine learning system.
Initial parameter value: refer to the initial parameter setting value of decision model scale and arithmetic speed.
Principle and the utility model has the advantages that
In use, user will be input to client to rehabilitation data, client passes through the improved intelligence in first processor
Can algorithm according to rehabilitation data optimization is determined to the initial parameter value of the BP neural network model in second processor
Obtain the initial weight and initial threshold of BP neural network model.Client will be to rehabilitation data and initial weight and initial threshold
Value is sent to cloud server together, and the BP neural network model in second processor is according to the specific initial weight received
And initial threshold, become the specific model that a convergence rate determines.In BP neural network model after optimization determines, BP
Neural network model will calculate corresponding BP calculated result after iterative calculation as input data after rehabilitation data.
BP neural network model calculates optimal BP and calculates knot in which can be convenient as classical machine learning model
Fruit.Server carries out this calculating process beyond the clouds, can greatly reduce the burden of client, and exist calculating resulting data
It is saved in cloud server, relevant information can be obtained from cloud server by being conducive to multiple client.At by second
Before managing device progress BP neural network model calculating, by intelligent algorithm improved in first processor first to BP neural network model
It optimizes, this makes the different BP neural network models that it is specifically calculated to rehabilitation data, and there are in iteration speed
Difference.In other words, optimized BP neural network model, can be according to specifically to rehabilitation data point reuse at being most suitable for this
The iteration speed calculated a bit to rehabilitation data improves calculating speed on the basis of not influencing and calculating accuracy as far as possible.This
So that the pace of learning of entire machine learning system is become faster, save because iteration speed is inappropriate and caused by over head time.
The present invention is optimized by initial parameter value of the improved intelligent algorithm to BP neural network, makes initial parameter value
Become a specific computable numerical value rather than a random value, calculates optimal result to shorten BP neural network model
Time.Calculating speed by improving BP neural network model makes entire machine learning system calculating speed faster, makes to reach
Time to the accurate machine learning system of result becomes faster.
Further, the cloud server further includes the database for real-time update user evaluation questionnaire and its answer,
The database is connect with second processor.It treats rehabilitation data by way of evaluation questionnaire to be acquired, database will
User data and its relevant treatment data by second processor carry out real-time update.
Further, the cloud server further includes the identity information for inputting to user and evaluation questionnaire answer progress
The checking assembly of verifying and for storing the second memory of push scheme, the checking assembly includes being provided with plan model
Authentication module, the authentication module are connect with second memory, and the checking assembly is connect with second processor.
Authentication module using store in the database to rehabilitation data be used as the input data of plan model, pass through planning mould
Type calculates the first propelling data to match with second memory and these first propelling datas is sent to second processor,
Push scheme corresponding in second memory is transferred to client according to the first propelling data by second processor.
Evaluation questionnaire answer be to rehabilitation data, and because having multiple problems in evaluation questionnaire, each problem reaction
Characterizing situation also can be different, therefore can be generally divided into user's symptom for characterizing the specific symptomatic condition of user to rehabilitation data
Data and for characterize user whether health and health status user health data.And authentication module use to rehabilitation number
According to mainly user health data.
Push scheme includes one of video, picture and text and audio or diversified forms, and each scheme have it is corresponding
Protocol Numbers carry out unique identification;Push scheme can be regarded as machine learning system and obtain to what user inputted to rehabilitation data
The result finally fed back.
Because client will be by cloud database after the database that rehabilitation data are transmitted in cloud server
Authentication module rehabilitation data, which be calculated the first propelling data, to be waited for these, and by first as healthy judging result
Propelling data is directly passed to second processor without passing through any other pilot process.This to input from client to health
After complex data, the response message of cloud server can be obtained as soon as possible.And cloud server is because of its wide memory space,
Compared to client, relevant calculation is directly carried out on server beyond the clouds, speed faster, there is entire machine learning system faster
Reaction speed.
Further, the second processor includes being provided with the screening module of decision-tree model and being provided with support vector machines
The correction module of model;Block is touched in the screening and correction module is connect with database and second memory.
Screening module according in database real-time update user is inputted to rehabilitation data questionnaire survey answer (i.e. to
Rehabilitation data itself) recurrence tree classification is carried out, make screening mould obtain judging data.Screening module judges Decision Classfication result
Data pass to the BP neural network model being similarly disposed in second processor.BP neural network model is according to judging that data determine
Whether treat rehabilitation data calmly to be calculated, if judging, data meet the design conditions of BP neural network model, BP nerve net
Network model can calculate BP calculated result, and (BP calculated result directly can supply the with pushing scheme in second memory and match
Two processors are to client push corresponding scheme, therefore BP calculated result also can be regarded as another propelling data, can claim
Be the second propelling data), it is on the contrary then not will do it calculating.In this way, only meeting can just carrying out to rehabilitation data for condition
Subsequent operation, it is on the contrary then will not.Different from plan model, BP neural network model is as input to be mainly to rehabilitation data
User's symptom data.So plan model obtains in authentication module the first propelling data and BP neural network model obtain
Two propelling datas, corresponding scheme are also different.The corresponding scheme of first propelling data be whether health and health status
Illustrate scheme, and the corresponding scheme of the second propelling data is the operation scheme for helping user to alleviate symptom.Because of each scheme
There are the Protocol Numbers of unique identification, so the first propelling data and the second propelling data also can be regarded as the collection of Protocol Numbers
It closes.
After the judgement of screening module, BP neural network model directly in second memory push scheme carry out
Match, and push scheme is directly sent to by cloud server by client by second processor.When user is receiving push
Scheme and after used a period of time, user are used to push the feedback coefficient that scheme is evaluated to last time by client input
According to.Feedback data is equally sent to database holding.Correction module is used as by the feedback data provided in database and is supported
The input data of vector machine model obtains amendment data after being handled, and according to amendment data to the push in second memory
Scheme push parameter corresponding with propelling data is modified, and revised push parameter is updated into database and is protected
It deposits.Here push parameter refers to, by the push scheme and the one-to-one relationship limit of propelling data progress in second memory
Fixed parameter setting, original push parameter are setting value.
Second processor can be allow directly by input after the judgement of rehabilitation data by screening module, according to BP
The second propelling data that neural network model obtains obtains pushing video to second memory.Input user to health from client
Push scheme can be quickly obtained after complex data.Decision-tree model is to use the more skilled and simple model of structure, can be fast
Speed completes the setting of decision-tree model in screening module.By correction module, can be convenient to push scheme accuracy into
Row monitoring, and by push parameter amendment come be continuously improved propelling data and push scheme corresponding relationship, make according to
The push scheme that rehabilitation data obtain is more and more accurate, and continuous the destination of study of machine learning system is reached with this.
Further, checking assembly includes study module for carrying out study early period and for carrying out the later period of later period study
Study module, and BP neural network model is provided in study module and later period study module;In the later period study module
The interstitial content of the BP neural network model of setting is more than the number of nodes in the BP neural network model in the study module.
The node for the BP neural network model being arranged in later period study module is than the BP neural network in study module
Model is more, illustrates that the calculating of the BP neural network model in later period study module is more accurate.Study module and later period learn mould
Block is directed to the machine learning in early period and later period respectively to be run.It is numerous after counting accuracy and calculating process first letter, it is suitble to each
From the machine learning stage.In early period, the corresponding related data of overall user is less, with the simple some BP nerve nets of calculating
Network model is relatively suitble to.In the later period, the corresponding related data of overall user run up to it is a certain amount of, can be by more multiple
Miscellaneous BP neural network model improves the accuracy of entire machine learning system.Pass through the multinode in later period study module
The calculating of BP neural network model can more optimize the relevant push parameter of push scheme, make the scheme of machine learning system
It pushes more intelligent.
Further, machine learning method, comprising the following steps:
Step 1: client is collected to rehabilitation data by way of questionnaire survey and reaches the database of cloud server
Middle storage;
Step 2: according in database real-time update to rehabilitation data using decision-tree model be directed to questionnaire survey answer
Recurrence tree classification is carried out, and the judgement data that classification obtains are directly transmitted into back client and BP neural network model;
Step 3: the initial weight and initial threshold of BP neural network are optimized using improved intelligent algorithm;
Step 4: BP neural network receives the judgement data that decision-tree model transmitting comes, and BP neural network is by determining
When the judgement of plan tree-model needs to push scheme, BP neural network is by obtaining Protocol Numbers collection after treating the calculating of rehabilitation data
It closes, second processor combines the scheme in second memory with the one-to-one scheme of Protocol Numbers set and is transferred to client
End, while these are waited for that the push parameter list that rehabilitation data and Protocol Numbers set are correspondingly formed is transferred to data by second processor
Library;
Step 5: the evaluation questionnaire input feedback number that user is provided after the scheme combination pushed by client
According to;
Step 6: amendment data will be obtained after handling using supporting vector machine model feedback data, will correct number
According to database and second memory is transferred to simultaneously, second processor is joined according to push corresponding when correcting data to push scheme
Number table is modified;
Step 7: by storage, obtain after real-time update in the database includes to rehabilitation number to BP neural network model
According to the user data online updating BP neural network parameter including, Protocol Numbers set and push parameter list;When user passes through visitor
Family end is inputted again to rehabilitation data, BP neural network model be added during calculating BP neural network parameter obtain by
Revised Protocol Numbers set;Protocol Numbers set is transferred to second processor and database simultaneously.
BP neural network parameter refers to determine all parameters of BP neural network topological structure.
User is answered by client for the evaluation questionnaire in client, and the evaluation questionnaire that user answers is answered
Case is input in the first processor in client as to rehabilitation data by input module.Passing through BP neural network model
Calculating after, a certain amount of original user data is collected into database (including to rehabilitation data, Protocol Numbers set and push
Parameter list, original push parameter list are pre-set).
Screening module in second processor, according in database real-time update to rehabilitation data (user's symptom data
With user health data) recurrence tree classification is carried out to the questionnaire survey answer (i.e. to rehabilitation data) inputted from input module, make
Screening module draws the judgement data for deciding whether push scheme according to the input of user's symptom data.
Study module in checking assembly receives the judgement data that the expression that screening module transmission comes needs to push scheme,
Study module will store the user data obtained after real-time update in the database online more by BP neural network model
New BP neural network parameter.(before input is to rehabilitation data really when user carries out suggested design customization by input module
The fixed operation for whether needing to push scheme), study module will join after the process of numerical procedure number set according to BP neural network
Number resets the topological structure and scale of BP neural network model, changes the corresponding Protocol Numbers of the scheme of push,
Optimization push scheme combination, for use next time when can obtain more optimized result.Including simultaneously will be comprising current input
BP neural network parameter save to database.Second processor passes through obtained scheme combination and pushes away to client progress scheme
It send.
After the scheme that user is pushed combines and operates according to the content that scheme combines, user passes through defeated
Enter the feedback data of scheme combination of the component input for push, wherein user is by filling in evaluation questionnaire input feedback data.
Correction module in second processor is repaired using feedback data as the input data of supporting vector machine model and being obtained by calculation
Correction data, supporting vector machine model are modified push parameter list according to amendment data, and by revised push parameter list
It updates in database and saves.
The present invention at least undergoes the period of scheme push twice, and the push of first time scheme is mainly to deposit according to setting first
Healthy symptom scales video parameter table in reservoir directly carries out scheme by plan model as original push parameter list
Push, original user data is effectively had accumulated by this process, is followed by machine learning and to constantly improve entire machine
Device learning system provides premise.
The present invention is acquired the various input information of user by way of questionnaire survey, has to the input of user
There is guiding performance, allow users to categorical data required for machine learning system is provided, improves effective benefit that user inputs information
With rate.
It, can be automatically by by the optimization module being arranged in first processor after the completion of the push of first time scheme
Improved intelligent algorithm pair is optimized with BP neural network model initial weight and initial threshold, makes BP neural network model
The more preferable convergence rate of convergence faster, makes machine learning system have faster pace of learning.
When second of scheme pushes, database has had collected a certain amount of original user data.In collection
Passed sequentially through on the basis of user data regression tree model, BP neural network model and supporting vector machine model to push parameter list into
Row is adjusted and improved, and the scheme of entire machine learning system push is made gradually to become more accurate.It can be according to the product of user data
Tired, the push for combining scheme is more humanized, intelligent, improves the precision combined to rehabilitation data and scheme and matching
Degree.
Further, the supporting vector machine model uses two category support vector machines models, after user's operational version
The vector set that feedback data is formed: { feeling of pain after receiving treatment, function improve degree } carries out as known training set
It calculates.
Further, adaptive operator is added during improved intelligent algorithm Optimized BP Neural Network model
Change mutation operator in the iterative process of BP neural network model, by way of gradually reducing, population is made to exist
Early period keeps diversity of individuals, avoids precocity, retains excellent information in the later period, avoids optimal solution from being destroyed, increase searches
The probability of globally optimal solution.
Further, the intersection that random range is added during improved intelligent algorithm Optimized BP Neural Network model is calculated
Son:
The adjustment of the operator considers possible random variation in difference vector amplification, helps to keep in search process
Population diversity.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of the embodiment of the present invention.
Fig. 2 is the operation schematic diagram of the embodiment of the present invention.
Fig. 3 is the BP neural network model flow figure that the embodiment of the present invention uses the optimization of improved intelligent algorithm.
Specific embodiment
Below by specific embodiment, the present invention is described in further detail:
Appended drawing reference in Figure of description includes: checking assembly 1, second memory 2, database 3, outside port 4, nothing
Line communication component 5, second processor 6, input module 7, video resolution component 8, display screen 9, first memory 10, network transmitting-receiving
Component 11, first processor 12.
Embodiment:
As shown in Figure 1, machine learning system, including cloud server and with cloud server by being connected to the network at least
Three clients.
Client includes first processor 12, input module 7, video resolution component 8, display screen 9,10 and of first memory
Network transmitting-receiving subassembly 11, input module 7, video resolution component 8, display screen 9, first memory 10 and network transmitting-receiving subassembly 11 are logical
It crosses data/address bus to connect with first processor 12, first processor 12 controls the operation of each component of client;Network transmitting-receiving subassembly
11 establish the data link of client and cloud server by least one network protocol;Input module 7 is used for user
Input the information such as authentication information and evaluation questionnaire answer;The storage of first memory 10 connects via aforementioned network transmitting-receiving subassembly 11
The scheme from cloud server push received;Video resolution component 8 is received for parsing via aforementioned network transmitting-receiving subassembly 11
From cloud server push video type scheme;Display screen 9 is for playing the video scheme after parsing, audio side
Case or display picture and text scheme.
Cloud server is including being stored with the second memory 2 of the various schemes including video, audio and picture and text, using
Carry out real-time update user data (to join including what is inputted as evaluation questionnaire answer to rehabilitation data, Protocol Numbers set and push
Number table, wherein including that can be used to characterize user's symptom data of user's symptomatic condition and be good for for characterizing user to rehabilitation data
The user health data of health situation) database, verify for the identity information inputted to user and evaluation questionnaire answer
Checking assembly 1, wireless communication components 5, outside port 4 and second processor 6;Outside port 4, second memory 2, data
Library, checking assembly 1, wireless communication components 5 are connect with second processor 6;Outside port 4 is suitable for the second of cloud server
Memory 2 and customer data base direct or through network and are indirectly coupled to other equipment;Checking assembly 1 is for verifying via nothing
Subscriber identity information that line communication component 5 receives simultaneously sends verification result signal to second processor 6;6 sound of second processor
Should in checking assembly 1 send be verified signal, control database in evaluation questionnaire be sent to via wireless communication components 5
Aforementioned client.
User will be input at first by input module 7 for the answer (i.e. to rehabilitation data) of the problems in evaluation questionnaire
It manages in device 12, these are waited for that rehabilitation data are sent to cloud server by network transmitting-receiving subassembly 11 by first processor 12.Cloud
The wireless communication components 5 of server receive user data and by these wait for rehabilitation data be transferred to second processor 6 or directly
It is transferred to database 3.Second processor 6 will be transmitted these data while these user data deliveries to customer data base 3
To checking assembly 1, checking assembly 1 sends the first propelling data to second processor 6 and second memory after passing through Validation Answer Key
2, so that second processor 6 is transferred corresponding scheme from second memory 2, and by these schemes component 5 by wireless communication
It is transferred to client.
The verifying for contacting the corresponding Protocol Numbers of each scheme and concrete scheme is provided in checking assembly
Module;Plan model is provided in authentication module.Medical expert is provided in authentication module for each in questionnaire
Each option in factor is the healthy symptom scales reference table of suggested design setting;Pass through healthy symptom scales reference table, rule
Drawing model can direct obtaining and matched first propelling data of scheme in second memory to rehabilitation data according to input.
Correspondence scheme in second memory directly can be transmitted to client according to the first propelling data by second processor.Generally
Be generally picture and text scheme by the scheme that plan model is calculated, mainly judge user whether health and health status such as
What.
The optimization that Automatic Optimal is carried out for the scheme push to entire machine learning system is additionally provided in first processor
Module;It is provided with improved intelligent algorithm in optimization module, is equipped with the doctor worked out according to a variety of common symptons in optimization module
Instruction data, these judgment criterias of physician guidance opinion data as improved intelligent algorithm, enable optimization module
Continue to optimize push parameter list.Optimization module is calculated corresponding Protocol Numbers by physician guidance opinion data and is compiled with scheme
Number adjustment amount be the combination of variable calculated scheme and the total effect of optimal combination that is obtained according to physician guidance opinion data it
Absolute value of the difference is objective function, carries out successive iteration using improved intelligent algorithm and obtains multiple groups comprehensive parameters adjustment amount, root
The distribution of factor option is carried out to adjustment amount according to weight, the corresponding Protocol Numbers table of suggested design is adjusted optimization.
Screening module is equipped in second processor;Screening module is by the user's symptom data and use of real-time update in database
Family health data is used respectively as the input data and categorical data of the decision-tree model being arranged in screening module with these
The row classification of the problems in family symptom data and the corresponding evaluation questionnaire of user health data;By the total Options in each problem
Regard 0 or 1 bi-values as, and these bi-values are linked to be one group of binary string, is converted binary system using system conversion method
For decimal number;Feature value for single choice problem is discrete type { 21,22,…,2n-1, n is option number corresponding to the problem
Mesh };For multiple-choice question, then feature value is continuous type { 1~2n-1- 1, n are number of options corresponding to the problem }), it uses
CART algorithm constructs decision-tree model;Metric processing is converted by the answer progress binary system of evaluation questionnaire and forms feature
Collection Feature1, Feature2, Feature3, Feature4, Feature5, Feature6, Feature7, Feature8,
Feature9};Feature set is corresponding with feature value set, as judging data;For defeated from input module in screening module
What is entered judges data to first processor transmitting according to categorical data to rehabilitation data, judges that data illustrate whether push rehabilitation
The judging result of video is (because be mostly health by the scheme that the second propelling data that BP neural network model obtains mainly pushes
Multiple video scheme, thus directly say here whether pushing video scheme, but be also not excluded for the second propelling data also and have and is corresponding non-
Video scheme).
The study module being arranged in checking assembly passes through all data that BP neural network model will be transmitted in database
It is trained and learns, obtain corresponding BP neural network parameter and save in the database.Rehabilitation view is received in checking assembly
The judging result of frequency combination, study module according to input module input to rehabilitation data and BP neural network parameter at second
Reason device exports the video number that video is recommended in corresponding push;BP mind before background program updates this data at the same time
Through network mould parameter;BP neural network parameter can determine the topological structure and scale of BP neural network model, after convenient
The calculating in face more optimizes accurately.
Correction module is additionally provided in second processor;Supporting vector machine model, correction module root are provided in correction module
Amendment data are obtained according to the feedback data that user inputs, and according to amendment data to the push scheme and push in second memory
The corresponding push parameter of data is modified, and revised push parameter is updated into database and is saved.I.e. user is defeated
It is constantly corrected between the video of the symptom and push that enter;And revised associated data is transmitted to database and is saved.
In the present solution, it is preferably rehabilitation training video, rehabilitation guide picture and text and health that good push scheme, which is previously set,
Refreshment, which is practiced, explains one of audio or a variety of;Each independent push scheme by professional person it is artificial impart comprehensive ginseng
Number.Comprehensive parameters make Protocol Numbers are corresponding with caused by corresponding scheme to close as the auxiliary Optimal Parameters to Protocol Numbers
System's (i.e. push parameter list) is more accurate.
Comprehensive parameters include but is not limited to: single push scheme for it is multinomial to rehabilitation data efficacy parameter respectively,
The priority parameters etc. of single push scheme;
Such as: the corresponding effective treatment ability of medicine in illness B of push option A is rAB(rAB∈ [0,1]), the size of value
The power of the effective treatment ability of medicine of push scheme can be measured.
Priority parameters are that push scheme carries out precedence sequence, first filter out push associated with scheme to be pushed
The push scheme of scheme nothing to do with connection, the push alternative priority for setting onrelevant are classified as 0, and associated push scheme is then pressed excellent
First sequence grade setting 1,2,3 ... ..., and 1 <, 2 < ... < n (there are multiple videos to be in same priority), for same
The information of dominance relation will carry out number limitation output according to the effective treatment ability of medicine height of synthesis.
Meanwhile restriction parameter has also been previously set in the combination of push scheme, the preferably single push scheme of the restriction parameter
Combine the quantity for the single push scheme for being included;
According to the restriction parameter of comprehensive parameters and push scheme combination to rehabilitation data, each push scheme, calculate
The push scheme combination optimized out is sent to user.
If after user carries out the i-th wheel rehabilitation training the diagnosis of new round questionnaire will be participated in, when diagnostic result shows the user still
It need to receive next round training, system is then according to next round diagnostic data, the rehabilitation training and corresponding training of recommending the wheel to carry out
Number of days, the feedback in training process carries out i+1 wheel rehabilitation training push, until medical health can be fully achieved in user
Until standard.This large amount of and unduplicated specific aim push scheme push are able to achieve the real time monitoring to user health situation,
Be conducive to the gradually rehabilitation of user's body.
As shown in Fig. 2, in the early period of machine learning system operation, there is no any user data in machine learning system,
It is constantly collected in use by user including (being wherein most importantly to rehabilitation to the user data including rehabilitation data
Data and corresponding propelling data)
Firstly, evaluation questionnaire is sent to client by cloud server, user reads evaluation questionnaire by display screen 9 and leads to
Crossing input module 7 will be input in first processor 12 for the answer (i.e. to rehabilitation data) of evaluation questionnaire.
First processor 12 lead to judging result after tentatively judging to the Consideration by the judgment module being arranged
It crosses display screen 9 to be presented to the user, so that user carries out subsequent operation selection.
It is carried out in first processor 12 using all problems input that AHP model is directed to all evaluation questionnaires in advance
AHP assigns power.Solve the relative weighting of each factor using two-phase method: using (0,1,2) three scale method to each reference factor into
Row build comparator matrix more afterwards two-by-two, by by user for evaluation questionnaire input answer be compared judgement after obtain by
Processing to rehabilitation data.Client is transmitted to what is generated by evaluation questionnaire in cloud server to rehabilitation data, finally quilt
It is stored in database 3.Second processor 6 is directly passed to checking assembly 1 what is come to transmission after rehabilitation data.It tests
The authentication module in component 1 is demonstrate,proved using user health data as the input data of plan model.
Medical expert is previously provided with to each evaluation questionnaire problematic factor option point using SPA model in advance in authentication module
Three following healthy symptom scales reference tables are established for essential body position after analysis.The material elements of suggested design are in total
There are nine, respectively to corresponding various Considerations.
Table 1-1
Table 1-2
Painful area (neck) | Serious symptom | Symptom is general | Health | Explanation |
Consideration 1 | 1,3 | 2 | 4 | Number represents option code name |
Consideration 2 | 4,5,6 | 2,3 | 1 | Number represents option code name |
Consideration 3 | 3 | 2 | 1 | Number represents option code name |
Consideration 4 | 3,4 | 2 | 1 | Number represents option code name |
Consideration 5 | 4 | 2,3 | 1 | Number represents option code name |
Consideration 6 | 1 | 2 | 2 | Number represents option code name |
Consideration 7 | 1 | 2 | 2 | Number represents option code name |
Consideration 8 | 1 | 2 | 2 | Number represents option code name |
Consideration 9 | 4,5 | 2,3 | 1 | Number represents option number |
Table 1-3
Painful area (shoulder) | Serious symptom | Symptom is general | Health | Explanation |
Consideration 1 | 1,3 | 2 | 4 | Number represents option code name |
Consideration 2 | 4,5,6 | 2,3 | 1 | Number represents option code name |
Consideration 3 | 3 | 2 | 1 | Number represents option code name |
Consideration 4 | 3,4 | 2 | 1 | Number represents option code name |
Consideration 5 | 4 | 2,3 | 1 | Number represents option code name |
Consideration 6 | 1 | 2 | 2 | Number represents option code name |
Consideration 7 | 1 | 2 | 2 | Number represents option code name |
Consideration 8 | 1 | 2 | 2 | Number represents option code name |
Consideration 9 | 4,5,6 | 2,3,4 | 1 | Number represents option number |
Table 1-4
Number | 1 | 2 | 3 | 4 | 5 | 6 |
Option | A | B | C | D | E | F |
Into table 1-3, number represents option code name and refers to that the number of choosing is exactly the code name i.e. rehabilitation video group of option table 1-1
The number of conjunction.Number represents option number, refers to that number limits the number of option, for example selected AB/AC/BD, is all 2,
ABC/BCD is then 3.The conversion relation of each of assessment table answer choice is write in table 1-4 exactly.Wherein, Consideration
It can be the factors relevant to symptom checking such as pain cause, pain duration and pain grade, and these factors all can
In the problem of being imbedded in evaluation questionnaire.User will include these factors in the form of to rehabilitation data by answering evaluation questionnaire
Data input come in.
Three-level is divided into user health situation according to healthy symptom scales reference table, therefore the connection member in SPA is ternary.Root
Meter is weighted to same, different, counter in the assessment system to rehabilitation data and healthy symptom scales reference table according to what user submitted
It calculates, finally obtains the first propelling data as Health Category data, determined and corresponded to according to the size relation of the first propelling data
Healthy symptom.Serious symptom grade is corresponded to coefficient, different coefficient corresponds to the general grade of symptom, and reciprocal coefficient corresponds to the good type of symptom
(health type), the maximum is the Health Category of user in calculated each coefficient of first propelling data.First processor 12 from
The Health Category data are transferred in checking assembly 1 and corresponding picture and text in its second memory 2 are illustrated that scheme is transmitted to
On display screen 9, the health status for being intuitive to see oneself is enabled users to.
As shown in healthy symptom scales reference table above, in addition to the judgment part of Health Category, pass through medical expert's needle
To each problem of evaluation questionnaire extract corresponding Consideration and it is corresponding represent numerical value, directly by each video for a certain specific
The therapeutic effect for the symptom (such as pain degree 0-3 point) being likely to occur have it is much, establish video and influence video recommendations it is each because
The table of corresponding relationship between plain option.Using dominance relation of the video between stochastic variable, each video or incidence relation and specially
Family gives model constraint condition of the standard combination number of video scheme as plan model, is obtained most by maximizing treatment degree
The stochastic variable that whole value is 1, the corresponding video of these variables are final rehabilitation training scheme combination.
Checking assembly 1 obtains the as rehabilitation training scheme data splitting by plan model for user's symptom data
One propelling data, and the data are transferred to second processor 6, second processor 6 is transferred according to the data is stored in advance in
Rehabilitation video combination in two memories 2, and the data that the rehabilitation video combines are sent to client.Client passes through access
The combination of the cloud server rehabilitation video, can see specific rehabilitation video by display screen 9.Second can also directly be passed through
Processor 6 will be combined and be transmitted directly to the first propelling data, that is, corresponding video of video number set in second memory 2
Client.
In the present embodiment, by client receive user based on evaluation questionnaire reply about user's current physical condition
Description information;To rehabilitation data it can be appreciated that the symptom description information extracted, the i.e. selected choosing of problem and user
?.
Specific questionnaire form, including multiple problems include each multiple options for problem, and each option is specially a kind of disease
Shape description, I are the number of problem, JiFor the option sum of i-th of problem;
For each option, then with the significance level of weight video the parameter characterization option pair and affiliated problem;
For each problem, then the selection combined with another weight video parameter characterization problem pair with rehabilitation video
Significance level;
The comprehensive parameters of rehabilitation training video include value be 0-100% multiple efficacy parameters, value 0,1,2,
The priority parameters (M indicates greatest priority number) of 3 ... M, and indicate the stage parameter in video affiliated stage, value 1,
2,3 ... K (K indicates greatest priority number);
The restriction parameter of rehabilitation video combination is preferably single rehabilitation video and combines included rehabilitation training video
Quantity;
These parameters can correspond to table by the parameter being preset in storage equipment and be inquired.
The following are the parameter to plan model in the present embodiment, setting is specific as shown in table 2:
Table 2
Then, the select permeability for pushing result is converted into plan model;Specifically, i.e. in the case where meeting certain condition
It maximizes and calculates
Max g (x)=Kernel (Kernel (R, Handamard (λ, U)), x)
Hadamard symbol indicates the Hadamard product of matrix/vector in formula;Kernel indicates vector dot product
Maximization calculating needs to meet following constraint condition:
The quantity of video is N;
In the rehabilitation stage locating for user, the video for not meeting the stage-training must not be exported
It only include the video in the stage that belongs to, the stage parameter choosing in rehabilitation stage and each video according to corresponding to user
It takes;
In the case where same satisfaction treatment, the high video (being determined by the priority parameters of video) of selection priority;
Specific maximize is calculated as solving:
It maximizes the purpose calculated to be to solve for meeting the X of constraint condition, in the X solved, be chosen according to its result corresponding
The combination of rehabilitation video is pushed to user, and maximized numerical value is then the integrated effect parameter of video combination at this time, is remembered below
Forg。
Physician guidance opinion, which is further used, in order to further optimize update, in this implementation cooperates video parameter table
The amount of being updated calculates.
Doctor suggests that the source of data is, in advance, expert works out tens common class illnesss, and every one kind illness includes multiple
Symptom is equal to and has selected multiple symptom options in questionnaire;If there is P illness, the symptom option of each illness can be used
One answer vector UpTo indicate;
Then, it holds charrette and provides objective and accurate push scheme combination to each illness.
Further parameter setting is as shown in table 3:
Table 3
Then, minimization problem below is converted into for the update of comprehensive parameters:
In general, authentication module mainly pushes or picture and text scheme, because main video scheme is still by learning
Neural network model in module is pushed.
As shown in Fig. 2, firstly, the answer of evaluation questionnaire is passed through input module to rehabilitation data by client by user
7 are input in first processor 12, these are waited for that rehabilitation data are transmitted to cloud by network transmitting-receiving subassembly 11 by first processor 12
It holds in server.Wireless communication components 5 in cloud server receive to rehabilitation data and will be transferred to second to rehabilitation data
Processor 6.
With the decision-tree model in screening module in second processor 6, by customer data base 3 in plan model operation
User's symptom data for being collected into and user health data respectively as decision-tree model input data and categorical data, to commenting
Estimate the row classification of the problems in questionnaire, is decimal number by the data processing of these questionnaire problems, i.e., it will be all in each problem
Option all regards 0 or 1 bi-values as, when option is selected, then takes 1, otherwise takes 0;Then by these values be linked to be one group two into
System string, use system conversion method by binary system be converted into decimal number (for single choice problem feature value for discrete type
{21,22,…,2n-1, n is number of options corresponding to the problem };For multiple-choice question, then feature value is continuous type { 1~2n -1- 1, n be the problem corresponding to number of options), using CART algorithm construct decision-tree model, by input to rehabilitation
Whether the calculating of data can need the judging result of pushing video scheme to BP neural network Model Transfer.Specific construction step
It is as follows:
Step 1: original decision is calculated according to classification situation of the original decision tree to training dataset and beta pruning data set
Set TmaxImportant coefficient I (Tmax), by TmaxAs T0It is put into this season i=1 in optimal tree candidate collection;Wherein, definition is determined
Plan tree T=(v1,v2,v3), wherein vector v1v2v3Respectively represent nicety of grading a (T) classification stability s (T) decision of decision tree
Set scale d (T) three Xiang Zhibiao;
The significance level of decision tree T defines decision tree important coefficientWherein WjIt is j-th point
The weighted value of vector, Wj>=0 and meet condition
Step 2: bottom-up that the internal node that each left and right is leaf node is successively attempted by cut operator compares
The beta pruning tree T obtained after node is removed every timecutImportant coefficient I (Tcut), retain I (Tcut) the maximum beta pruning tree T of valuecutMake
For TiIt is put into optimal subtree candidate collection and enables i=i+1;
Step 3: in subtree Ti-1On the basis of repeat step 2 to carry out the construction that candidate optimal beta pruning tree is gathered straight
Proceed to original decision tree T to cut operatormaxRoot node until;
Step 4: in optimal subtree candidate collection { T0,T1,T2,T3,...,TiIn more each beta pruning tree TiImportance
Coefficient I (Ti) retain I (Ti) value maximum one as final optimum decision tree;
Step 5: the optimum decision tree that is obtained with step 4 is handled to rehabilitation data, and by optimum decision tree processing result
It is transferred directly to the checking assembly connecting with second processor;Convolutional neural networks are provided in the checking assembly.
The nicety of grading of decision treeWherein Nl is all leaf nodes in decision tree T;
Define the prior probability that the example that beta pruning is concentrated reaches node tWherein N ' is the test in beta pruning sample set
Example sum, n ' (t) are that the example number for entering the node t of decision tree is concentrated in beta pruning;Wherein e ' (t)
The class label example sum for reaching node t and belonging to node t is concentrated for beta pruning.
The classification stability of decision treeWherein s (t)=min { a ' (t)/a (t), a
(t)/a′(t)}。
Decision tree scaleWherein, m is the leaf node number of decision tree T, and h is decision tree T
Depth capacity.
By the problems in questionnaire survey data by pre-processing, regard the total Options in each problem the two of 0 or 1 as
Member value, when option is selected, then takes 1, otherwise takes 0;Then these values are linked to be a binary string, using system conversion side
Binary system is converted decimal number by method;Feature set is formed, feature set can specifically indicate are as follows: Feature1, Feature2,
Feature3, Feature4, Feature5, Feature6, Feature7, Feature8, Feature9 }
It is calculated by conversion it is known that the specific value condition of each feature set is as shown in table 4 below:
Table 4
It is arranged according to decision-tree model, classification collection is { Level1, Level2, Level3 }, and corresponding value may be configured as
{ 1,2,3 }.
Classified by what decision-tree model inputted user from client to rehabilitation data, and by as classification knot
The judgement data of fruit judge whether to need to carry out video scheme push to user.It then proves to need if it is Level 1 or Level 2
Rehabilitation Training in Treating is carried out, otherwise need not recommend (assessment table is please referred to, specifies that the first second level illustrates that symptom is obvious,
The third level illustrates that body is healthier, without progress rehabilitation training).
If it is determined that the result is that do not need to push, then can be unsatisfactory for conditional information from cloud server to client transmission
Picture and text scheme: you very it is healthy does not need push Regimen Chemotherapy.If it is judged that need to push scheme, then decision-tree model
The judgement data of output can be transmitted directly in the BP neural network model being arranged in checking assembly 1.
User accesses cloud server by client, can see on display screen 9 and judge by decision-tree model
The health status of oneself.And then following operation is carried out according to the prompt on display screen 9.
Data collected by operation phase early period are trained and are learned on backstage by the BP neural network in checking assembly 1
It practises, obtains corresponding BP neural network parameter and be stored in database 3.When there is user to be customized scheme, it is stored in data
BP neural network parameter in library 3 constitutes suitable BP neural network model, by according to user submit to rehabilitation data meter
It calculates and exports corresponding rehabilitation video combination;BP neural network model before background program updates this data at the same time is joined
Number.During entire model training, network inputs are user's symptom data (binary-coded character string), and desired output is doctor
It is raw to carry out the view after treatment video data obtained by plan model is modified to by video parameter table (i.e. push parameter list)
Frequency output data.Data (binary decision vector form) after being modified to the obtained video scheme of preliminary stage, network
It is set as two hidden-layer structure, and is learnt using gradient descent algorithm.
After the judgement of screening module, BP neural network model directly in second memory push scheme carry out
Match, and push scheme is directly sent to by cloud server by client by second processor.When user is receiving push
Scheme and after used a period of time, user are used to push the feedback coefficient that scheme is evaluated to last time by client input
According to.Feedback data is equally sent to database holding.Correction module is used as by the feedback data provided in database and is supported
The input data of vector machine model obtains amendment data after being handled, and according to amendment data to the push in second memory
Scheme push parameter corresponding with propelling data is modified, and revised push parameter is updated into database and is protected
It deposits.
Improved intelligent algorithm model is provided in the optimization module being arranged in first processor 12, to entire machine learning
The model of system optimizes.Physician guidance opinion data is provided in first processor 12, which is
Medical expert works out tens common classes and waits for rehabilitation data, and holds medical expert's seminar and provide objective standard to each symptom
True therapeutic scheme combination.Wait for that rehabilitation data calculate the corresponding comprehensive parameters of video according to this tens class, with the tune of comprehensive parameters
Whole amount is variable, and the calculated video combination of model and the absolute value of the difference of the maximum therapy degree of expert are objective function, is adopted
Successive iteration is carried out with improved intelligent algorithm and obtains multiple groups comprehensive parameters adjustment amount, and factor choosing is carried out to adjustment amount according to weight
Item distribution and entropy processing, are finally adjusted optimization for the corresponding entire video parameter table of video.
In the present embodiment, intelligent algorithm Optimized BP Neural Network is broadly divided into: neural network structure determines that intelligent algorithm is excellent
Change weight and threshold value, BP neural network training and prediction.
Wherein the topological structure of BP neural network is determined according to the input/output parameters number of sample, it is possible thereby to determine
The number of genetic algorithm optimization parameter determines the code length of population at individual.Because intelligent algorithm Optimal Parameters are BP nerve nets
The initial weight and threshold value of network, as long as network structure it is known that there is known as long as the number of weight and threshold value.The weight of neural network and
Threshold value generally passes through the random number that random initializtion is [- 0.5,0.5] section, which influences very network training
Greatly, it but can not accurately obtain, for identical initial weight and threshold value, network training introduces intelligence and calculates the result is that the same
Method is exactly for the optimal initial weight of optimization and threshold value.The comprehensive parameters update generated with plan model, BP nerve
The comprehensive parameters update that network generates finally also transitions between the calculated result for seeking BP neural network and sample calculated result
The problem of minimizing the error.
It is solved using the minimization problem that improved intelligent algorithm is converted into the update of comprehensive parameters;Herein for
Selection uses, genetic annealing algorithms/improvement particle algorithm/differential evolution algorithm adaptively adjusted, three kinds of algorithms.
In genetic annealing algorithms, after specifically using binary coding, population primary to generate, according to the survival of the fittest and winning
The bad principle eliminated develops by generation and produces best xpApproximate solution;And in order to accelerate search speed and enhance the part of algorithm
Search capability is scanned for by defining the energy assessment function and initial temperature of single individual using simulated annealing.
In improved particle algorithm, the calculation of new group's extreme value and individual extreme value has been used.
For each individual extreme value Pbi, from remaining with for the extreme value for randomly selecting an individual in other individual extreme values
It is denoted as Pbj, newly generated individual extreme value is denoted as PbiThen:
Pbi(t)=r*Pbi(t)+(1-r)*Pbj(t)
Wherein t expression runs to t generation, and i indicates that i-th of particle, j indicate j-th of particle, and r is [0,1] being randomly generated
Between number.
It is as follows to the improvement of a group's extreme value Gb:
All individual extreme values are ranked up, then from wherein picking out K best individual Pb1′,Pb2′,,,
Pbk', indicate Gb ' with the weighted average of this K individual, then:
Wherein t expression runs to t generation, and i indicates i-th of the particle picked out, aiMeet:
Adaptive operator is added during improved intelligent algorithm Optimized BP Neural Network model:
In formula: F0For mutation operator;GmIt is current evolutionary generation for maximum evolutionary generation: G.
Crossover operator:
Change mutation operator in an iterative process, by way of gradually reducing, population is made to keep individual multiplicity in early period
Property, precocity is avoided, retains excellent information in the later period, optimal solution is avoided to be destroyed, increases the probability for searching globally optimal solution.
The adjustment of crossover operator considers possible random variation in difference vector amplification, helps to protect in search process
Hold population diversity.
By experimental verification: two operators are added later can be more for the function of complex (such as multidimensional function)
Optimal solution is searched out fastly, and result is accurate compared with more.
By one of above method, x is finally solvedpBest fit approximation solution, and obtain according to weight final micro-
Adjust matrix Yp, it may be assumed that
Yp=Kernel (xp,πp);
Y is individually calculated to all illnesssp;
Finally, being updated using the fine tuning matrix to rehabilitation video and the incidence matrix R of symptom description information, it may be assumed that
R ' is then updated incidence matrix.Incidence matrix herein is video parameter table, can equally be extended to institute
Some schemes push parameter list.
Wherein, the specific calculating of BP neural network model is as follows:
Step 1: initialization network: by the user data passed over by decision-tree model and supporting vector machine model
These input datas are processed into binary data, the input and output of neural network constitute one by the input for forming neural network
A sequence, is expressed as (x, y), and according to this sequence, we are capable of determining that the input layer, hidden layer and output node layer of network
Number, respectively n node, l node, m node.Then, connection weight and threshold value are initialized, if input layer, hidden layer are refreshing
It is w through the connection weight between memberij, the connection weight between hidden layer and output layer neuron is wjk, hidden layer and output layer
Threshold value be respectively a and b.Finally, the neuron excitation function f (X) and learning rate η of neural network are provided.
Step 2: the output of hidden layer is calculated: the connection according to the input variable X of network, between input layer and hidden layer
Weight wij, and the threshold value a of hidden layer, it can be concluded that the output H of hidden layer:
F is the excitation function of hidden layer in above formula;L is the node number of hidden layer.Excitation function are as follows:
Step 3: it calculates the output of output layer: calculating the output of hidden layer, connection hidden layer and output according to step 2
The connection weight w of layerjkAnd output layer threshold value b, it can be concluded that BP neural network prediction output 0:
Step 4: it calculates error: O is exported according to the neural network forecast that the desired output Y and step 3 of network are obtained, it can
Calculate the prediction error e of network:
ek=Yk-OkK=1,2 ..., m
Step 5: weight is updated: according to the prediction error e of neural network to the connection weight w of networkijAnd wjkIt carries out
It updates:
wjk=wjk+ηHjekJ=1,2 ..., l;K=1,2 ..., m
Wherein, η is the learning rate of neural network.
Step 6: update threshold value: according to the prediction error e of network again to threshold value a, b of network hidden layer and output layer into
Row updates:
bk=bk+ekK=1,2 ..., m
Step 7: judging whether to meet algorithm iteration termination condition, if being unsatisfactory for, then return step two.
The BP calculated result that BP neural network model obtains is the second propelling data, the second propelling data and the second storage
Video scheme in device 2 is corresponding, and video scheme is sent to client or will store the position of video scheme by second processor 6
Data are sent to client, are obtained by client.
In order to improve the convergence rate of BP neural network, second order gradient method is used after step 5:
Wherein:
Although second order gradient method has relatively good convergence, need to calculate E to the second dervative of w, calculation amount is very big.
It does not directly adopt second order gradient method generally, and uses variable-metric method or conjugate gradient method, there is second order gradient method such as to restrain for they
Fast advantage, and without directly calculating second order gradient.
Wherein, variable metric algorithm:
W (t+1)=w (t)+μ H (t) D (t)
Δ w (t)=w (t)-w (t-1)
Δ D (t)=D (t)-D (t-1)
The supporting vector machine model being arranged in second processor 6 uses two common category support vector machines models, will
Client uses the feedback data set after video: { feeling of pain after receiving treatment, function improve degree } is used as known training
Collection.
Step 1: known training set is set:
T={ (x1,y1),…,(xl,yl)}∈(X×Y)l
Wherein, xi∈X∈Rn, yi∈ Y={ 1, -1 } (i=1,2 ..., l);xiFor feature vector.
Step 2: the outstanding K of kernel function appropriate (x, x ') and parameter C appropriate are selected, constructs and solves optimization problem:
Obtain optimal solution:
Step 3: α is chosen*A positive componentAnd threshold value is calculated accordingly:
Step 4: construction decision function:
It is 1 or -1 according to f (x), it is resolved that its classification ownership.
Indicate that user's each section function is recovered normal when f (x) is 1, by neural network, model is pushed to user
Video be it is accurate useful, the rehabilitation training for still needing to participate in next stage is indicated when f (x) is -1, at this time by the { pain of user feedback
Pain: mean value, variance, function improve degree: mean value, variance } it is input in supporting vector machine model, then again by BP nerve net
Network model exports video data to client again, i.e., optimizes to the calculating of BP neural network model.
In the continuous feedback of supporting vector machine model, can according to the feedback data of user by symptom that user input with push away
Constantly corrected between the video sent, make below neural network model for user's rehabilitation video combination when it is more accurate.
Meanwhile correcting data and being tracked for user when carrying out video rehabilitation training, it is able to use family and realizes more rapid rehabilitation
Effect (the symptomatic reaction adjustment later period after capable of receiving rehabilitation training according to user pushes scheme, reaches the customization of hommization).
Deep learning model is mainly established in the later period operation of machine learning system, increases the machine learning early period stage
Neural network hidden node number, and data will be corrected as the interim foundation treated, the larger nerve of direct construction
Network model all before replacing.
As shown in figure 3, improved intelligent algorithm optimizes the initial weight and initial threshold of BP neural network model
Specific step is as follows:
Step 1: according to BP neural network parameter in database 3 is stored in, determine neural network model topological structure and
Type.
Step 2: weight and threshold value to BP neural network model all in database 3 are encoded to obtain initial kind
Group.
Step 3: it determines and specific weight and threshold value is obtained by decoding after initial population.
Step 4: by obtained weight and threshold value to newly-established BP neural network model carry out using.
Step 5: it obtains training network using pre-set training sample, keeps entire BP neural network model constantly excellent
Change.
Step 6: pre-set test sample test network is used.
Step 7: it compares test result and official result that test network obtains to obtain test error.
Step 8: the test error of individual each in population is analyzed and processed and calculates fitness.
Step 9: the high individual, that is, chromosome of selection fitness is replicated.
Step 10: intersected according to conventional algorithm.
Step 11: it makes a variation according to conventional algorithm.
Step 12: obtaining new group, judges whether to meet termination according to pre-set physician guidance opinion data
Condition;If met, directly it is decoded to obtain optimal weight and threshold value;If not meeting termination condition, step is returned
Rapid three.
Improved intelligent algorithm used in the present embodiment is differential evolution algorithm.The specific steps of which are as follows:
Step 1: initialization.Using differential evolution algorithm using NP dimension is the real number value parameter vector of D as each
The population in generation, each individual indicate are as follows: Xi,G(i=1,2 ..., NP)
In formula: i expression-sequence of the individual in population;G indicates evolutionary generation;NP indicates population scale, is minimizing
NP is remained unchanged in the process.
It is assumed that meeting non-uniform probability distribution to all random initializtion populations.The boundary of setting parameter variable isThen:
In formula, uniform random number that rand [0,1] is generated between [0,1].
Step 2: variation.To each object vector Xi,G(i=1,2 ..., NP), the variation of basic differential evolution algorithm
Vector generates as follows:
Wherein, randomly selected serial number r1 r2And r3It is different, and r1 r2And r3Should not with object vector serial number i yet
Together, so must satisfy NP >=4.Mutation operator F ∈ [0,2] is a real constant factor, controls the amplification of variation variable.
Because mutation operator takes real constant, the more difficult determination of mutation operator in implementation, aberration rate is too big, makes algorithm search inefficiency, asks
It is low to obtain globally optimal solution precision;Aberration rate is too small, and population diversity reduces, and " precocity " phenomenon easily occurs.Therefore it is added as follows
Adaptive mutation rate λ, according to algorithm search progress, adaptive mutation rate design is as follows:
In formula: F0For mutation operator;GmIt is current evolutionary generation for maximum evolutionary generation: G.
When algorithm starts, adaptive mutation rate is F0~2F0, there is the larger value, keep diversity of individuals in the early stage, keep away
Exempt from precocity;As algorithm progress mutation operator gradually reduces, to later period aberration rate close to F0, retain excellent information, avoid optimal
Solution is destroyed, and the probability for searching globally optimal solution is increased.
Step 3: intersect.In order to increase the diversity of interference parameter vector, crossover operation is introduced.Then trial vector becomes:
ui,G+1=(u1i,G+1,u2i,G+1,...,uDi,G+1)
(i=1,2 ... NP;J=1,3 ..., D)
In formula: randb (j) generates j-th of estimated value of randomizer between [0,1];Rnb (i) ∈ 1,2 ..., D
It is a randomly selected sequence, ensures u with iti,G+1At least from Vi,G+1Obtain a parameter;CR is crossover operator, value
Range is [0,1].
In order to keep the diversity of group, the crossover operator of following random range is devised:
Step 4: selection.To determine trial vector ui,G+1Whether member in the next generation can be become, it will according to greedy criterion
Object vector X in trial vector and current populationi,GIt is compared.If objective function will be minimized, have smaller
The vector of target function value will win a position on the ground in next-generation population.All individuals in the next generation are all than current population
Corresponding individual is more preferably or at least equally good.
Step 5: the processing of boundary condition.In having the problem of boundary constraint, it is ensured that the parameter value for generating new individual is located at
Be necessary in the feasible zone of problem, a straightforward procedure be will not meet boundary constraint new individual be used in it is random in feasible zone
The parameter vector of generation replaces.
That is: ifOrSo:
Test example:
Compare for convenience, selects be equipped with the present embodiment machine learning system (this example) and existing machine learning respectively
The health and fitness information interaction platform of system (comparative example) is tested, both health and fitness information interaction platforms are in addition to machine learning system
System is different outer, other various pieces are identical, and the scheme of the usage mode of two health and fitness information interaction platforms and push
Type is identical.
Firstly, choose two groups of ages, gender and conditions two groups of group members all the same, respectively simultaneously with being equipped with this reality
A machine learning system (this example) and existing machine learning system (comparative example) is applied every other day to carry out once in the identical period
It uses.In order to facilitate comparison, each age bracket and the identical user of conditions take two kinds of genders of male and female respectively
People compares use.Each group member is had recorded in table 5 after input is to rehabilitation data to the sound for receiving push scheme
Between seasonable.
Table 5
Unit in table 5 is the second.As can be seen from the above table, response time 20 of the people of each age bracket in the same period
Year and the response time of 30 years old people are most short, this skilled operation with the people at the two ages is related.And with same in a small group
The response time difference of the male and female of sample age bracket and identical conditions is little.
For in two groups of group members of the same same age bracket of conditions and same gender, using installation in every a line
There is the group member of the health and fitness information interaction platform of the machine learning system in the present embodiment, can contract with the increase of access times
Short response time has absolutely proved that machine learning system can with being continually entered in user's use process to rehabilitation data
Study optimization is constantly carried out, calculating speed is improved, the scheme of suitable user is rapidly pushed to user.And use is equipped with now
There is the health and fitness information interaction platform of machine learning system, although the response time is gradually shortened also with increasing for access times,
But the use time of correspondence period is longer than another group of group member.And increasing with access times, the variation of response time
Also below another group group member of speed.Illustrate that the machine learning system in the present embodiment has faster than existing machine learning system
Pace of learning.
Then, the rehabilitation training for carrying out every group of group member according to the 4th obtained push scheme in table 5 six months by a definite date,
And at the third moon, April, May and June, the rehabilitation efficacy of each group member is evaluated respectively.Evaluation criterion
For go identical hospital completion identical standard physical examination, table 6 is that the result carried out according to physical examination result counts, wherein 100 expressions
It fully recovers, 0 indicates do not have rehabilitation completely, i.e. push scheme does not play a role completely.
Table 6
As can be seen from the above table, the people of each age bracket is different in the rehabilitation speed of same period, wherein 10 years old, 20 years old
Very fast with the rehabilitation speed of 30 years old people, this is related with the physical fitness of the people at these three ages.And with year same in a small group
The rehabilitation speed difference of the male and female of age section and identical conditions is little.
For in two groups of group members of the same same age bracket of conditions and same gender, using this reality in every a line
The physical examination result that the machine learning system in example is better than another group of group member in each physical examination middle of the month physical examination result is applied, is sufficiently said
The scheme that the health and fitness information interaction platform push constituted using the machine learning system of the present embodiment is illustrated is more suitable user, this
The precision of machine learning system push scheme in embodiment is higher, and the effect of generation is more preferable, the engineering in the present embodiment
Learning system is more preferable by the identical learning effect obtained to rehabilitation data.
What has been described above is only an embodiment of the present invention, and the common sense such as well known specific structure and characteristic are not made herein in scheme
Excessive description, technical field that the present invention belongs to is all before one skilled in the art know the applying date or priority date
Ordinary technical knowledge can know the prior art all in the field, and have using routine experiment hand before the date
The ability of section, one skilled in the art can improve and be implemented in conjunction with self-ability under the enlightenment that the application provides
This programme, some typical known features or known method should not become one skilled in the art and implement the application
Obstacle.It should be pointed out that for those skilled in the art, without departing from the structure of the invention, can also make
Several modifications and improvements out, these also should be considered as protection scope of the present invention, these all will not influence the effect that the present invention is implemented
Fruit and patent practicability.The scope of protection required by this application should be based on the content of the claims, the tool in specification
The records such as body embodiment can be used for explaining the content of claim.
Claims (8)
1. machine learning system, including being used to acquire the cloud to the client of rehabilitation data and with client by network connection
Server;It is characterized by: the client includes the first processor for being provided with improved intelligent algorithm;The cloud service
Device includes the second processor for being provided with BP neural network model, and wherein second processor is passing through BP neural network model treatment
Pass through the initial parameter value of improved intelligent algorithm Optimized BP Neural Network, the initial parameter value packet while to rehabilitation data
Include initial weight and initial threshold;
Its machine learning method the following steps are included:
Step 1: client is collected to rehabilitation data and is reached by way of questionnaire survey and deposited in the database of cloud server
Storage;
Step 2: according in database real-time update to rehabilitation data using decision-tree model for questionnaire survey answer progress
Tree classification is returned, and the judgement data that classification obtains are directly transmitted into back client and BP neural network model;
Step 3: the initial weight and initial threshold of BP neural network are optimized using improved intelligent algorithm;
Step 4: BP neural network receives the judgement data that decision-tree model transmitting comes, and BP neural network is passing through decision tree
Model judgement is when needing to push scheme, and BP neural network is by obtaining Protocol Numbers set after treating the calculating of rehabilitation data, and the
Scheme combination in second memory is transferred to client by two processors, and the scheme and Protocol Numbers collection in scheme combination are unified
One is corresponding, while these are waited for that the push parameter list that rehabilitation data and Protocol Numbers set are correspondingly formed is transferred to by second processor
Database;
Step 5: the evaluation questionnaire input feedback data that user is provided after the scheme combination pushed by client;
Step 6: will obtain amendment data after handling using supporting vector machine model feedback data, same by data are corrected
When be transferred to database and second memory, second processor according to amendment data to push scheme when corresponding push parameter list
It is modified;
Step 7: BP neural network model according to the user data that is obtained after real-time update in the database of storage online more
New BP neural network parameter, user data include to rehabilitation data, Protocol Numbers set and push parameter list;When user passes through visitor
Family end is inputted again to rehabilitation data, BP neural network model be added during calculating BP neural network parameter obtain by
Revised Protocol Numbers set;Protocol Numbers set is transferred to second processor and database simultaneously.
2. machine learning system according to claim 1, it is characterised in that: the cloud server further includes for real-time
The database of user's evaluation questionnaire and its answer is updated, the database is connect with second processor.
3. machine learning system according to claim 1, it is characterised in that: the cloud server further include for
The checking assembly and stored for storing the second of push scheme that the identity information of family input and evaluation questionnaire answer are verified
Device, the checking assembly include the authentication module for being provided with plan model, and the authentication module is connect with second memory, described
Checking assembly is connect with second processor.
4. machine learning system according to claim 1, it is characterised in that: the second processor includes being provided with decision
The screening module of tree-model and the correction module for being provided with supporting vector machine model;It is described screening touch block and correction module with number
It is connected according to library with second memory.
5. machine learning system according to claim 3, it is characterised in that: checking assembly includes being used to carry out study early period
Study module and for carrying out the later period study module of later period study, and be provided in study module and later period study module
BP neural network model;The interstitial content for the BP neural network model being arranged in the later period study module is than the study module
In BP neural network model in number of nodes it is more.
6. machine learning system according to claim 1, it is characterised in that: the supporting vector machine model is using two classification
Supporting vector machine model, the vector set that the feedback data after user's operational version is formed: the feeling of pain after receiving treatment,
Function improves degree } it is calculated as known training set.
7. machine learning system according to claim 1, it is characterised in that: optimize BP nerve net in improved intelligent algorithm
Adaptive operator is added during network modelF=F0·2λ, in formula: F0For mutation operator;GmIt evolves for maximum
Algebra: G is current evolutionary generation.
8. machine learning system according to claim 1, it is characterised in that: optimize BP nerve net in improved intelligent algorithm
The crossover operator of random range is added during network model:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610881058.0A CN106779084B (en) | 2016-09-30 | 2016-09-30 | Machine learning system and method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610881058.0A CN106779084B (en) | 2016-09-30 | 2016-09-30 | Machine learning system and method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106779084A CN106779084A (en) | 2017-05-31 |
CN106779084B true CN106779084B (en) | 2019-07-09 |
Family
ID=58972351
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610881058.0A Active CN106779084B (en) | 2016-09-30 | 2016-09-30 | Machine learning system and method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106779084B (en) |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107273705A (en) * | 2017-07-25 | 2017-10-20 | 北京品驰医疗设备有限公司 | The determination of human body implantation type medical treatment device parameter, method to set up and equipment |
CN107659609B (en) * | 2017-07-26 | 2021-01-12 | 北京天云融创软件技术有限公司 | Deep learning support platform based on cloud computing and deep learning training method |
CN108247764B (en) * | 2017-12-14 | 2021-04-27 | 北京木业邦科技有限公司 | Wood board cutting method and device based on machine learning, electronic equipment and medium |
CN108154913A (en) * | 2017-12-29 | 2018-06-12 | 北京精密机电控制设备研究所 | A kind of rehabilitation training parameter determination method, device, electronic equipment and system |
CN109993391B (en) * | 2017-12-31 | 2021-03-26 | 中国移动通信集团山西有限公司 | Method, device, equipment and medium for dispatching network operation and maintenance task work order |
CN108897283B (en) * | 2018-05-31 | 2022-07-19 | 中策橡胶集团股份有限公司 | Rubber mixing production line data analysis processing method |
CN111316227B (en) * | 2018-08-20 | 2021-08-13 | 华为技术有限公司 | Method and equipment for debugging application program |
CN109800277A (en) * | 2018-12-18 | 2019-05-24 | 合肥天源迪科信息技术有限公司 | A kind of machine learning platform and the data model optimization method based on the platform |
CN110164524A (en) * | 2019-04-29 | 2019-08-23 | 北京国润健康医学投资有限公司 | A kind of hemiplegia patient training mission adaptive matching method and its system |
CN111913743B (en) * | 2019-05-09 | 2023-04-14 | 杭州海康威视数字技术股份有限公司 | Data processing method and device |
CN111008322B (en) * | 2019-11-27 | 2020-10-30 | 广州快决测信息科技有限公司 | Method and system for automatically identifying effective data acquisition module |
CN111523646B (en) * | 2020-04-23 | 2023-06-23 | 国家开放大学 | Intelligent perception network and management method for remote education learning center based on Internet of things |
CN111816281B (en) * | 2020-06-23 | 2024-05-14 | 无锡祥生医疗科技股份有限公司 | Ultrasonic image inquiry device |
CN112420215A (en) * | 2020-10-16 | 2021-02-26 | 童心堂健康科技(北京)有限公司 | Physiological data judgment method and system based on artificial intelligence |
CN112905591B (en) * | 2021-02-04 | 2022-08-26 | 成都信息工程大学 | Data table connection sequence selection method based on machine learning |
CN113017585A (en) * | 2021-03-18 | 2021-06-25 | 深圳市雅士长华智能科技有限公司 | Health management system based on intelligent analysis |
CN115577225B (en) * | 2022-12-07 | 2023-05-12 | 通号万全信号设备有限公司 | Two-out-of-two voting processing method, device and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9022140B2 (en) * | 2012-10-31 | 2015-05-05 | Resource Energy Solutions Inc. | Methods and systems for improved drilling operations using real-time and historical drilling data |
CN104794355A (en) * | 2015-04-29 | 2015-07-22 | 孙卫唯 | Method for executing online rehabilitation guidance in computer system and rehabilitation guidance system |
CN105636512A (en) * | 2013-08-14 | 2016-06-01 | 南洋理工大学 | Systems and methods for revascularization assessment |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2009068072A1 (en) * | 2007-11-30 | 2009-06-04 | Kinkadee Systems Gmbh | Scalable associative text mining network and method |
-
2016
- 2016-09-30 CN CN201610881058.0A patent/CN106779084B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9022140B2 (en) * | 2012-10-31 | 2015-05-05 | Resource Energy Solutions Inc. | Methods and systems for improved drilling operations using real-time and historical drilling data |
CN105636512A (en) * | 2013-08-14 | 2016-06-01 | 南洋理工大学 | Systems and methods for revascularization assessment |
CN104794355A (en) * | 2015-04-29 | 2015-07-22 | 孙卫唯 | Method for executing online rehabilitation guidance in computer system and rehabilitation guidance system |
Non-Patent Citations (2)
Title |
---|
基于BP神经网络的脑卒中患者居家上肢康复动作识别技术;庞古山 等;《计算机工程与应用》;20121231;第48卷(第S2期);第96-99页 |
基于EJB的B_S模式神经网络智能诊断方法研究;冯瑶 等;《现代制造工程》;20051231(第5期);第117-119页 |
Also Published As
Publication number | Publication date |
---|---|
CN106779084A (en) | 2017-05-31 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106779084B (en) | Machine learning system and method | |
CN106682385B (en) | Health information interaction system | |
Han et al. | A survey on metaheuristic optimization for random single-hidden layer feedforward neural network | |
Neysiani et al. | Improve performance of association rule-based collaborative filtering recommendation systems using genetic algorithm | |
Wang et al. | Flower pollination algorithm with bee pollinator for cluster analysis | |
Li et al. | Bi-goal evolution for many-objective optimization problems | |
García-Hernández et al. | A novel hybrid evolutionary approach for capturing decision maker knowledge into the unequal area facility layout problem | |
Jiang et al. | Design of structural modular neural networks with genetic algorithm | |
CN103984721B (en) | Books individuation search method based on interactive evolutionary optimization | |
CN109214503B (en) | Power transmission and transformation project cost prediction method based on KPCA-LA-RBM | |
Ding et al. | Product color emotional design considering color layout | |
CN106407711A (en) | Recommendation method and recommendation system of pet feeding based on cloud data | |
CN106485069A (en) | The method and system of rehabilitation information pushing | |
CN106447117B (en) | Pet feeding method and system based on the daily data analysis of pet | |
Dolhopolov et al. | Use of artificial intelligence systems for determining the career guidance of future university student | |
CN110211684A (en) | The electrocardiogram classification method of BP neural network based on genetic algorithm optimization | |
Dell’Anna et al. | Evolving fuzzy logic systems for creative personalized socially assistive robots | |
Xie et al. | Prediction of transfers to tertiary care and hospital mortality: A gradient boosting decision tree approach | |
Orong et al. | A new crossover mechanism for genetic algorithm with rank-based selection method | |
Devi et al. | Monkey behavior based algorithms-a survey | |
Czajkowski et al. | Steering the interpretability of decision trees using lasso regression-an evolutionary perspective | |
Li et al. | Characterizing player’s playing styles based on player vectors for each playing position in the Chinese Football Super League | |
Yu et al. | Evaluation of sports training effect based on GABP neural network and artificial intelligence | |
Reddy et al. | A Beginner's Guide to Federated Learning | |
CN112070200B (en) | Harmonic group optimization method and application thereof |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |