CN106691440A - Controller fatigue detection method and system based on BP neural network - Google Patents
Controller fatigue detection method and system based on BP neural network Download PDFInfo
- Publication number
- CN106691440A CN106691440A CN201611115813.0A CN201611115813A CN106691440A CN 106691440 A CN106691440 A CN 106691440A CN 201611115813 A CN201611115813 A CN 201611115813A CN 106691440 A CN106691440 A CN 106691440A
- Authority
- CN
- China
- Prior art keywords
- ripples
- slow
- power ratio
- neural network
- perclos
- 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.)
- Pending
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B2503/00—Evaluating a particular growth phase or type of persons or animals
- A61B2503/20—Workers
Abstract
The invention relates to the field of fatigue detection, in particular to a controller fatigue detection method and system based on a BP neural network. The controller fatigue detection method based on the BP neural network includes the steps that brain waves of a controller are collected, and a slow alpha wave power percentage, a power ratio of alpha waves to beta waves and a power ratio of theta waves to slow alpha waves are obtained according to the brain waves; the slow alpha wave power percentage, the power ratio of the alpha waves to the beta waves and the power ratio of the theta waves to the slow alpha waves are input into a pre-trained BP neural network model, and a PERCLOS value simulation result is obtained; if the PERCLOS value simulation result is larger than a fatigue threshold value, the controller is judged to be in a fatigue state. According to the controller fatigue detection method and system based on the BP neural network, based on the BP neural network, the fatigue degree of the human body is detected in real time by detecting the brain waves, which makes the real-time fatigue detection simpler and reduces the detection cost.
Description
Technical field
The present invention relates to fatigue detecting technology field, and in particular to a kind of controller's fatigue detecting based on BP neural network
Method and system.
Background technology
Growing with air traffic, the live load of air traffic controller is increasing, its tired journey
Degree is to Air Traffic System level of security important.International Civil Aviation Organization has been that tired risk management is formulated
Doc9966 rules and regulations handbooks.European and American developed countries are also successively by for fatigue detecting system or the method extension of pilot
Onto controller's fatigue detecting application.CAAC innings is to instruct with International Civil Aviation Organization Doc9966, also in CCAR-121 files
In specify that the rule of tired risk management.
But, up to the present, although domestic and international researcher proposes various fatigue detectings and management method and system, but
It is that these methods mainly have three aspects not enough.One is subjective, and such as a large amount of questionnaire forms are used for fatigue judgement and predict
In, researcher can incorporate experience into according to the answer result of measured and be given a mark to determine degree of fatigue, can so receive unavoidably
To the influence of researcher's subjective judgement；Two is to be difficult to real-time detection, and it is logical to have quite a few method being currently in use
The performance of (such as continuous tens days) measured was crossed in the observation long period, so that tired trend prediction chart is set up, further according to figure
Table come judge certain a period of time in controller it is whether tired.Controller's current physical condition is thus directly have ignored, may
Testing result is affected；Three is that the current existing method suitable for real-time fatigue detecting is used to face mostly
Feature is acquired and knows method for distinguishing, and this method needs high accuracy video detecting device to shoot controller at any time, from cost
Angle analysis are provided no advantage against.
The content of the invention
For defect of the prior art, controller's fatigue detection method based on BP neural network that the present invention is provided and
System, based on BP neural network, by detecting brain wave come the degree of fatigue of real-time detection human body, becomes real-time fatigue detecting
It is more simple, and reduce testing cost.
In a first aspect, a kind of controller's fatigue detection method based on BP neural network that the present invention is provided, including：Collection
The brain wave of controller, power ratio, θ ripples and the slow α of slow α wave powers percentage, α ripples and β ripples are obtained according to the brain wave
The power ratio of ripple；By the power ratio of the power ratio, θ ripples and slow α ripples of the slow α wave powers percentage, the α ripples and β ripples
The power ratio of value is input into the good BP neural network model of training in advance, obtains PERCLOS value simulation results；If described
PERCLOS values simulation result is more than fatigue threshold, then judge that the controller is in fatigue state.
The controller's fatigue detection method based on BP neural network that the present invention is provided, only with by letter in real-time detection
Single economic mode detects the brain wave of controller, and the BP neural network model that brain wave input is built in advance just can be accurately
The current PERCLOS value simulation results of controller are estimated, so as to detect the fatigue state of controller.Therefore, the present invention is provided
Method real-time fatigue detecting is become more simple, and reduce testing cost.
Preferably, the training method of the BP neural network model includes：Set up BP neural network model and generate at random
The parameter of the BP neural network model, the BP neural network model includes input layer, intermediate layer, output layer, the input
Layer includes 3 nodes, and the intermediate layer includes multiple nodes, and the output layer includes 1 node, the input layer and it is described in
Full connection mode is used between interbed, and between the intermediate layer and the output layer；Gather controller brain wave with
And corresponding catacleisis data, power ratio, the θ ripples of slow α wave powers percentage, α ripples and β ripples are obtained according to the brain wave
With the power ratio of slow α ripples, according to the PERCLOS value measurement results that the catacleisis data are obtained, and multiple samples are generated,
The power ratio of each sample including slow α wave powers percentage, α ripples and β ripples, the power ratio of θ ripples and slow α ripples and corresponding
PERCLOS value measurement results；A sample is chosen from the sample of generation, by the slow α wave powers percentage in sample, α ripples and β
The power ratio of the power ratio, θ ripples and slow α ripples of ripple is input into the BP neural network model, obtains PERCLOS values and estimates knot
Really；According to the PERCLOS values estimation results and the error of the PERCLOS value measurement results of the sample chosen, the BP is updated
The parameter of neural network model；If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
Preferably, the PERCLOS value measurement results obtained according to the catacleisis data, including：From the eye
The upper palpebra inferior ultimate range under controller's waking state is obtained in eyelid closure data, the catacleisis data are eyelid
Closed amplitude changes with time, and the catacleisis amplitude is the distance between upper palpebra inferior；By the catacleisis data
Divided by the upper palpebra inferior ultimate range, catacleisis degree is obtained；According to the catacleisis degree, in the unit of account time
Closed-eye time；Closed-eye time is obtained into PERCLOS value measurement results divided by the unit interval.
Preferably, it is described according to the catacleisis degree, the closed-eye time in the unit of account time, including：In unit
In time, when the summation of corresponding time period of the catacleisis degree more than 70% or 80% is the eye closing in the unit interval
Between.
Preferably, if reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again, including：
PERCLOS value measurement results in the PERCLOS values estimation results and the sample chosen obtain global error, if described
Global error reaches default maximum times less than error threshold or frequency of training, then terminate training, and sample is otherwise chosen again
It is trained again.
Preferably, according to the PERCLOS values estimation results and the mistake of the PERCLOS value measurement results of the sample chosen
Difference, updates the parameter of the BP neural network model, including：Calculate the PERCLOS values estimation results and choose sample in
PERCLOS value measurement results output error；According to the output error relative to intermediate layer to each side right value of output layer
Partial derivative, updates the intermediate layer to each side right value of output layer；It is each to intermediate layer relative to input layer according to the output error
The partial derivative of side right value, updates the input layer to each side right value in intermediate layer；It is inclined relative to output layer according to the output error
The partial derivative put, updates the output layer biasing；According to the output error relative to the partial derivative that intermediate layer biases, institute is updated
State intermediate layer biasing.
Preferably, it is described according to the brain wave obtain the power ratio of slow α wave powers percentage, α ripples and β ripples, θ ripples and
The power ratio of slow α ripples, including：Gather the brain wave of controller；Wavelet Denoising Method is carried out to the brain wave；According to Wavelet Denoising Method
Later brain wave obtains the power ratio of the power ratio, θ ripples and slow α ripples of slow α wave powers percentage, α ripples and β ripples.
Second aspect, a kind of controller's fatigue detecting system based on BP neural network that the present invention is provided, including：Brain electricity
Ripple processing module, the brain wave for gathering controller obtains slow α wave powers percentage, α ripples and β ripples according to the brain wave
Power ratio, θ ripples and slow α ripples power ratio；Fatigue data output module, for by the slow α wave powers percentage, described
The power ratio of the power ratio of α ripples and β ripples, the θ ripples and slow α ripples is input into the good BP neural network model of training in advance, obtains
To PERCLOS value simulation results；Tired judge module, if being more than fatigue threshold for the PERCLOS values simulation result, sentences
The controller of breaking is in fatigue state.
The controller's fatigue detecting system based on BP neural network that the present invention is provided, passes through brain wave in real-time detection
Sensor detects the brain wave of controller, by brain wave input processor, by the BP neural network built in advance in processor
Model just can accurately estimate the current PERCLOS value simulation results of controller, so as to be examined according to PERCLOS values simulation result
The fatigue state of controller is surveyed, and alarm is sent by alarm, remind controller.Therefore, what the present invention was provided is refreshing based on BP
Through controller's fatigue detecting system of network, real-time fatigue detecting is set to become more simple, and reduce testing cost.
Preferably, also it is used for including BP neural network model instruction module：Set up BP neural network model and generate institute at random
The parameter of BP neural network model is stated, the BP neural network model includes input layer, intermediate layer, output layer；The input layer
Comprising 3 nodes, the intermediate layer includes multiple nodes, and the output layer includes 1 node；The input layer and the centre
Full connection mode is used between layer, and between the intermediate layer and the output layer；Gather controller brain wave and
Corresponding catacleisis data, according to the brain wave obtain the power ratio of slow α wave powers percentage, α ripples and β ripples, θ ripples and
The power ratio of slow α ripples, according to the PERCLOS value measurement results that the catacleisis data are obtained, and generates multiple samples, often
The power ratio of individual sample including slow α wave powers percentage, α ripples and β ripples, the power ratio of θ ripples and slow α ripples and corresponding
PERCLOS value measurement results；A sample is chosen from the sample of generation, by the slow α wave powers percentage in sample, α ripples and β
The power ratio of the power ratio, θ ripples and slow α ripples of ripple is input into the BP neural network model, obtains PERCLOS values and estimates knot
Really；According to the PERCLOS values estimation results and the error of the PERCLOS value measurement results of the sample chosen, the BP is updated
The parameter of neural network model；If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
Preferably, it is described to be obtained according to the catacleisis data in BP neural network model instruction module
PERCLOS value measurement results, including：The upper palpebra inferior under controller's waking state is obtained from the catacleisis data
Ultimate range, the catacleisis data change with time for catacleisis amplitude, and the catacleisis amplitude is upper and lower eye
The distance between eyelid；By the catacleisis data divided by the upper palpebra inferior ultimate range, catacleisis degree is obtained；According to
The catacleisis degree, the closed-eye time in the unit of account time；Closed-eye time is obtained divided by the unit interval
PERCLOS value measurement results.
Brief description of the drawings
Fig. 1 is the schematic diagram of PERCLOS measuring principles；
Three layers of structural representation of BP neural network model that Fig. 2 is used by the embodiment of the present invention；
The flow chart of the controller's fatigue detection method based on BP neural network that Fig. 3 is provided by the embodiment of the present invention；
The structural frames of the controller's fatigue detecting system based on BP neural network that Fig. 4 is provided by the embodiment of the present invention
Figure.
Specific embodiment
The embodiment of technical solution of the present invention is described in detail below in conjunction with accompanying drawing.Following examples are only used for
Technical scheme is clearly illustrated, therefore is intended only as example, and protection of the invention can not be limited with this
Scope.
It should be noted that unless otherwise indicated, technical term used in this application or scientific terminology should be this hair
The ordinary meaning that bright one of ordinary skill in the art are understood.
Brain wave is highly important Human Physiology index, can indirectly reflect the degree of fatigue of human body.PERCLOS values
Ratio shared by the eyes closed time in the unit time, is the value that admittedly can directly reflect degree of fatigue.It is of the invention real
Apply example offer the controller's fatigue detection method based on BP neural network, by BP neural network model obtain brain wave with
Relation between PERCLOS values.
The brain wave of controller is gathered by brain wave sensor, and by high-definition intelligent algorithm video camera to measured's face
Portion's feature carries out whole real-time video, obtains the catacleisis data synchronous on a timeline with brain wave.
Brain wave to collecting is processed, and therefrom extracts key parameter.Specific process step includes：To Noise
The brain wave of signal carries out wavelet transformation；Noise reduction process is carried out to the wavelet coefficient that conversion is obtained, with making an uproar that removal is wherein included
Sound；Wavelet inverse transformation is carried out to the wavelet coefficient after treatment.According to the waveform obtained after Wavelet Denoising Method, three ginsengs are calculated
Number：The power ratio of the power ratio, θ ripples and slow α ripples of slow α wave powers percentage, α ripples and β ripples.
Measured's face feature video to collecting carries out associated picture treatment, obtains catacleisis data.Specific side
Method includes lower three steps：
Step S10, carries out human eye positioning.The main process for carrying out human eye positioning is as follows：
Eye areas compared with peripheral region, with gray value is relatively low and the characteristics of larger rate of gray level.Therefore can base
Positioned in the half-tone information of eye image.It is divided into following two steps：
(1) eyes coarse localization.
After being accurately positioned face, it is distributed according to face organ, human eye can very simply determine one in the first half of face
Individual general area.Observation face picture, finds eye in the horizontal direction by skin, the left eye white of the eye, pupil of left eye, left eye eye
In vain, skin, the right eye white of the eye, pupil of right eye, the right eye white of the eye, skin, grey scale change are larger.Carried out at grey scale change mutation micro-
Point, high level will be produced, its absolute value is added up, then that bigger a line of grey scale change, accumulated value is bigger.Computing formula is as follows：
ΔhF (x, y)=f (x, y)-f (x-1, y)
F (x, y) is the gray level image of the human face region for obtaining, and is found through experiments that, the derivative changing value sum at eyes
Maximum absolute value, the online position of human eye can roughly be judged by the method.
(2) human eye is accurately positioned.Make discovery from observation C around eyesbValue is higher, and CrValue is relatively low, therefore according to following public affairs
Formula is calculated characteristic pattern, to protrude eye feature：
Wherein, EyeMap is eye feature figure, (Cb)2,(Cb/Cr) all normalize between [0,255],It is
Negated by Cr and obtained (255, Cr).After EyeMap figures are obtained, threshold values T is set, the value by EyeMap less than T is set to 0, this step
A simple filtering is can be considered to remove the interference of non-eye feature.
After obtaining EyeMap filtering figures, with reference to human eye coarse positioning result, from left to right search for, define in proportion relative to people
A certain size frame of face region, when frame enters it is EyeMap filtering map values and maximum when, as human eye.
Step S20, after completing positioning, eyes is tracked using the method for Deformable Template.Deformable Template process has
Body includes：If the position in the eye template upper left corner is (x, y), the hunting zone of next frame is along upper and lower, left and right 4 on original position
Individual direction respectively extends 10 pixels, and its formula is
In above formula, N is the number of picture rope in template；M is template；I is part to be matched in image.All can be more than
Coordinate corresponding to the maximum of the p of threshold value is the position for most matching, and the eye image obtained using this is used as next two field picture
Template.During tracking, the detection of eyes is come back to if the p for obtaining respectively less than threshold values or the line-spacing of two are excessive
Journey.
Step S30, according to the image that eyes are traced into from video, measurement obtains the distance between palpebra inferior, i.e. eye
Eyelid closed amplitude, catacleisis amplitude changes with time as catacleisis data；According to catacleisis number obtained above
According to the PERCLOS value measurement results for obtaining, specific implementation is comprised the following steps：Controller is obtained from catacleisis data
Upper palpebra inferior ultimate range under waking state；By catacleisis data divided by upper palpebra inferior ultimate range, catacleisis is obtained
Degree, catacleisis degree is as shown in Figure 1 with the relation of time；According to catacleisis degree, the eye closing in the unit of account time
Time；Closed-eye time is obtained into PERCLOS value measurement results divided by the unit interval.
Fatigue identification based on PERCLOS P80 (or P70) model, will catacleisis degree be more than 80% (or 70%)
Eye state be judged as closure state.With upper palpebra inferior ultimate range of initial time controller when clear-headed as standard, if with
The distance for obtaining afterwards is then judged as closure less than 80% (or 70%) of this distance.PERCLOS values measurement result is opened and closed by eyes
Scope and duration is short is determined, its measuring principle as shown in figure 1, once closing one's eyes-eye opening process as a example by, t1~t4When
Between section be unit time, time period of the correspondence catacleisis degree more than 20%；t2~t3Time period is closed-eye time, correspondence eye
Time period of the eyelid closure degree more than 80% (or 70%)；PERCLOS values can be calculated by following equation,
Wherein, f is the percentage of eyes closed time, is represented during once closing one's eyes-opening eyes, and f is bigger, and eyes connect
The time of nearly closure is more long, and the possibility of fatigue is bigger, and f values are the PERCLOS values for needing to solve.Actually used the method
When, according to the image that eyes are traced into from video, measurement obtains the distance between palpebra inferior；Upper and lower eye is obtained according to measurement
The distance between eyelid and the upper palpebra inferior ultimate range of the measured for obtaining in advance, obtain catacleisis degree；Gather many frame numbers
Catacleisis degree curve corresponding with number of frames (equivalent to the time) is can obtain according to rear, i.e., during with number of frames to represent
Between.
Above-mentioned be analyzed by the curve to catacleisis degree-time, using P80 (or P70) model measurement
PERCLOS values, the method can be accurately obtained PERCLOS values, but on condition that need to obtain accurate catacleisis degree-when
Between curve map, this is accomplished by accurately analyzing video collection.
In order to simplify the calculating process of PERCLOS values, another implementation of step S30 is：According to from video with
Track is measured and obtains the distance between palpebra inferior to the image of eyes；The distance between upper palpebra inferior and pre- is obtained according to measurement
The upper palpebra inferior ultimate range of the measured for first obtaining, obtains catacleisis degree, if catacleisis degree more than 80% (or
70%), then judge that the two field picture is catacleisis frame；By unit interval palpebra interna closure frame number and the ratio of the totalframes for the treatment of
As PERCLOS values.The method first determines whether that the eyes in single-frame images are closed or opened, and then counts eye closing frame number and exists
Whether the ratio accounted in totalframes judges controller in fatigue state, and required precision of the method to video acquisition be lower,
Processing speed is faster.Assuming that the frame per second of experiment video is 10fs-1, resolution ratio is 640 × 480, duration 60s, then regarded with every 6s
Frequency takes 1 frame and makees eyes detection as 1 detection unit, interval 0.33s.Count the shape of 18 two field pictures in each detection unit
State, obtains the totalframes SumFrame_Num of catacleisis frame number CloseFrame_Num and treatment, calculates corresponding according to formula
PERCLOS values
If gained PERCLOS values are more than 50%, judge that now controller has been in fatigue state, by warning system
Alerted.
Due to the input vector X=(x of BP neural network model used in the embodiment of the present invention1,x2,x3) and export to
The dimension of amount Y=(y) is relatively low, influences to calculate effect in real time to avoid BP neural network model excessively complicated, it is preferred to use
Three layers of BP neural network model are predicted, specific as shown in Fig. 2 BP neural network model includes input layer, intermediate layer, output
Layer.Input layer includes 3 nodes, and the power ratio of the slow α wave powers percentage, α ripples and β ripples of single test sample is corresponded to respectively
The power ratio of value, θ ripples and slow α ripples.Intermediate layer includes multiple nodes, middle layer node number not only with input layer and output layer
Nodes are relevant, more with the factor such as the characteristic of the complexity of the problem that need to be solved and the form of transfer function and sample data
It is relevant, in the embodiment of the present invention, the preferred value for obtaining middle layer node number is tested by network training in certain span
It is 8, intermediate layer sets 8 nodes, can guarantee that network performance, reduces the systematic error of network, while shortens net training time.
Output layer includes 1 node, the PERCLOS values of the single test sample of correspondence.Between input layer and intermediate layer, and intermediate layer and
Using each node of full connection mode, i.e. input layer, to intermediate layer, each node uses a line phase between output layer
Even, same connected mode is also adopted by between intermediate layer and output layer.
As shown in figure 3, setting input layer respectively i1、i2、i3, middle layer node is respectively h1、h2、……h8, output
Node layer is o1.Input layer is set to wi to the side right value of middle layer nodeij, 1≤i≤3,1≤j≤8, middle layer node arrives
The side right value for exporting node layer is set to woij, 1≤i≤8, j=1.
The mapping relations of the input in intermediate layer and the mapping relations of output, the input of output layer and output use S type functions,
I.e.
To middle layer node, its input form isOutput form is hok=f
(hik), 1≤k≤8, wherein, bikIt is bias.
To output node layer, its input form isOutput form is yo=f (yi), wherein
Bo is bias.
The training method of the BP neural network model used in the embodiment of the present invention includes：Gather controller brain wave with
And corresponding catacleisis data, the power ratio of slow α wave powers percentage, α ripples and β ripples, θ ripples and slow are obtained according to brain wave
The power ratio of α ripples, according to the PERCLOS value measurement results that catacleisis data are obtained, and generates multiple samples, each sample
The power ratio and corresponding PERCLOS of power ratio, θ ripples and slow α ripples including slow α wave powers percentage, α ripples and β ripples
Value measurement result；A sample is chosen from the sample of generation, by the work(of slow α wave powers percentage, α ripples and β ripples in sample
The power ratio input BP neural network model of rate ratio, θ ripples and slow α ripples, obtains PERCLOS value estimation results；According to
The error of the PERCLOS value measurement results in PERCLOS values estimation results and the sample chosen, updates BP neural network model
Parameter；If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
Wherein, if reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again, including：Root
Global error is obtained according to the PERCLOS value measurement results in PERCLOS values estimation results and the sample chosen, if global error is small
Default maximum times are reached in error threshold or frequency of training, then terminates training, sample otherwise chosen again and is trained again.
Wherein, the error of the PERCLOS value measurement results in PERCLOS values estimation results and the sample chosen, more
The parameter of new BP neural network model, including：The PERCLOS values calculated in the sample of PERCLOS values estimation results and selection are surveyed
Measure the output error of result；Partial derivative according to output error relative to intermediate layer to each side right value of output layer, updates intermediate layer
To each side right value of output layer；Partial derivative according to output error relative to input layer to each side right value in intermediate layer, updates input layer
To each side right value in intermediate layer；According to output error relative to the partial derivative that output layer is biased, output layer biasing is updated；According to output
Error updates intermediate layer biasing relative to the partial derivative that intermediate layer biases.
With reference to specific formula, the training method of BP neural network model is illustrated：
Step one：Netinit.It is one [- 1,1] to all side right values and excitation function bias random initializtion
Number on interval.Study frequency n=1 is set, i.e., is learnt by the 1st sample.
Step 2：Give a sampleWherein,N-th study is represented respectively (i.e.
N-th sample) in use slow α wave powers percentage, α ripples and β ripples power ratio, θ ripples and slow α ripples power ratio,
Represent the PERCLOS numerical value used in n-th study.The theoretical output yo of content calculating network first according to the 3rd trifle.
Step 3：Define the error of reality output and network theory output
Step 4：Calculate partial derivative of the output error relative to intermediate layer to each side right value of output layer
Therefore,
Step 5：Calculate partial derivative of the output error relative to input layer to each side right value in intermediate layer
Therefore,
Step 6：Output error is calculated relative to the partial derivative that output layer is biased, its derivation is similar with step 4.
Result is directly given herein
Step 7：Output error is calculated relative to the partial derivative that intermediate layer biases, its derivation is similar with step 5.
Result is directly given herein
Step 8：Update side right value and bias.
Step 9：Judge study end condition.Global error is calculated first
Wherein, yomNetwork theory output valve in learning at the m times is represented,Represent the reality output in learning at the m times
Value.Can be by yomThe PERCLOS value estimation results of network output in the m times study are interpreted as, andRepresent in the m times study
The PERCLOS value measurement results of middle use, i.e., the m-th PERCLOS value measurement result of sample.If E is default less than one
Value, or n reaches default maximum study number of times, then terminate study, provides the three-layer neural network structure for succeeding in school.Otherwise,
N=n+1 is made, two are gone to step, beginning learns next time.
Based on above-mentioned BP neural network, a kind of controller's fatigue inspection based on BP neural network is the embodiment of the invention provides
Survey method, as shown in figure 3, including：
Step S1, gathers the brain wave of controller, and the work(of slow α wave powers percentage, α ripples and β ripples is obtained according to brain wave
The power ratio of rate ratio, θ ripples and slow α ripples.
Step S2, the power ratio of the power ratio, θ ripples and slow α ripples of slow α wave powers percentage, α ripples and β ripples is input into
The good BP neural network model of training in advance, obtains PERCLOS value simulation results.
Step S3, if PERCLOS values are more than fatigue threshold simulation result, judges that controller is in fatigue state.
Wherein, fatigue threshold preferably 0.5.
Controller's fatigue detection method based on BP neural network provided in an embodiment of the present invention, only uses in real-time detection
The brain wave of controller is detected by way of simple economy, the BP neural network model that brain wave input is built in advance is with regard to energy
The current PERCLOS value simulation results of controller are accurately estimated, so as to detect the fatigue state of controller.Prior art is straight
Connected high definition camera detection face features, and obtained PERCLOS values to judge the fatigue state of controller, in order to obtain compared with
The requirement of accuracy of detection high to testing equipment is high, and this can greatly increase the cost of detection, and corresponding face features are calculated
Method is also complex, is unfavorable for real-time detection.And the present invention provide method detection be controller brain wave, compared to face
For identification, it is necessary to equipment and the algorithm that uses it is all relatively simple, to realize that real-time fatigue detecting provides favourable support,
And reduce testing cost.
In order that the brain wave for collecting can more accurately react the real fatigue state of controller, it is necessary to the brain of collection
Electric wave is pre-processed, and in step s 2, the brain wave according to controller obtains the work(of slow α wave powers percentage, α ripples and β ripples
The power ratio of rate ratio, θ ripples and slow α ripples, including：Gather the brain wave of controller；Wavelet Denoising Method is carried out to brain wave；According to
The later brain wave of Wavelet Denoising Method obtains the power ratio of the power ratio, θ ripples and slow α ripples of slow α wave powers percentage, α ripples and β ripples
Value.The brain wave of controller is surveyed by brain wave sensor and is acquired, and the brain wave to collecting is processed, and is therefrom extracted
Key parameter.Specific process step includes：Brain wave to noisy acoustical signal carries out wavelet transformation；The small echo obtained to conversion
Coefficient carries out noise reduction process, to remove the noise for wherein including；Wavelet inverse transformation is carried out to the wavelet coefficient after treatment.According to small
The waveform obtained after ripple denoising, is calculated three parameters：Power ratio, the θ ripples of slow α wave powers percentage, α ripples and β ripples
With the power ratio of slow α ripples.
Based on above-mentioned controller's fatigue detection method identical inventive concept based on BP neural network, the present invention implement
A kind of controller's fatigue detecting system based on BP neural network that example is provided, as shown in figure 4, including：Brain wave processing module
101, the brain wave for gathering controller obtains power ratio, the θ of slow α wave powers percentage, α ripples and β ripples according to brain wave
The power ratio of ripple and slow α ripples；Fatigue data output module 102, for by the power ratio of slow α wave powers percentage, α ripples and β ripples
The power ratio of value, θ ripples and slow α ripples is input into the good BP neural network model of training in advance, obtains PERCLOS value simulation results；
Tired judge module 103, if being more than fatigue threshold for PERCLOS values simulation result, judges that controller is in fatigue state.
Controller's fatigue detecting system based on BP neural network provided in an embodiment of the present invention, only uses in real-time detection
The brain wave of controller is detected by way of simple economy, the BP neural network model that brain wave input is built in advance is with regard to energy
The current PERCLOS value simulation results of controller are accurately estimated, so as to detect the fatigue state of controller.Prior art is straight
Connected high definition camera detection face features, and obtained PERCLOS values to judge the fatigue state of controller, in order to obtain compared with
The requirement of accuracy of detection high to testing equipment is high, and this can greatly increase the cost of detection, and corresponding face features are calculated
Method is also complex, is unfavorable for real-time detection.And the present invention provide method detection be controller brain wave, compared to face
For identification, it is necessary to equipment and the algorithm that uses it is all relatively simple, to realize that real-time fatigue detecting provides favourable support,
And reduce testing cost.
Wherein, the controller's fatigue detecting system based on BP neural network provided in an embodiment of the present invention also includes BP nerves
Network model instructs module, for setting up BP neural network model and the at random parameter of generation BP neural network model, BP nerve nets
Network model includes input layer, intermediate layer, output layer；Input layer includes 3 nodes, and intermediate layer includes multiple nodes, output layer bag
Containing 1 node；Full connection mode is used between input layer and intermediate layer, and between intermediate layer and output layer；Collection control
The brain wave and corresponding catacleisis data of member, the power of slow α wave powers percentage, α ripples and β ripples is obtained according to brain wave
The power ratio of ratio, θ ripples and slow α ripples, according to the PERCLOS value measurement results that catacleisis data are obtained, and generates multiple
Sample, the power ratio of each sample including slow α wave powers percentage, α ripples and β ripples, the power ratio of θ ripples and slow α ripples and
Corresponding PERCLOS values measurement result；From generation sample in choose a sample, by the slow α wave powers percentage in sample,
The power ratio input BP neural network model of the power ratio, θ ripples and slow α ripples of α ripples and β ripples, obtains PERCLOS values and estimates knot
Really；According to PERCLOS values estimation results and the error of the PERCLOS value measurement results of the sample chosen, BP neural network is updated
The parameter of model；If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
Wherein, in BP neural network model instruction module, knot is measured according to the PERCLOS values that catacleisis data are obtained
Really, including：The upper palpebra inferior ultimate range under controller's waking state is obtained from catacleisis data, catacleisis data are
Catacleisis amplitude changes with time, and catacleisis amplitude is the distance between upper palpebra inferior；By catacleisis data divided by
Upper palpebra inferior ultimate range, obtains catacleisis degree；According to catacleisis degree, the closed-eye time in the unit of account time；
Closed-eye time is obtained into PERCLOS value measurement results divided by the unit interval.
Finally it should be noted that：Various embodiments above is merely illustrative of the technical solution of the present invention, rather than its limitations；To the greatest extent
Pipe has been described in detail with reference to foregoing embodiments to the present invention, it will be understood by those within the art that：Its according to
The technical scheme described in foregoing embodiments can so be modified, or which part or all technical characteristic are entered
Row equivalent；And these modifications or replacement, the essence of appropriate technical solution is departed from various embodiments of the present invention technology
The scope of scheme, it all should cover in the middle of the scope of claim of the invention and specification.
Claims (10)
1. a kind of controller's fatigue detection method based on BP neural network, it is characterised in that including：
The brain wave of controller is gathered, power ratio, the θ of slow α wave powers percentage, α ripples and β ripples are obtained according to the brain wave
The power ratio of ripple and slow α ripples；
By the power ratio of the power ratio of the power ratio, θ ripples and slow α ripples of the slow α wave powers percentage, the α ripples and β ripples
The good BP neural network model of value input training in advance, obtains PERCLOS value simulation results；
If the PERCLOS values simulation result is more than fatigue threshold, judge that the controller is in fatigue state.
2. method according to claim 1, it is characterised in that the training method of the BP neural network model includes：
Set up BP neural network model and generate the parameter of the BP neural network model at random, the BP neural network model bag
Input layer, intermediate layer, output layer are included, the input layer includes 3 nodes, and the intermediate layer includes multiple nodes, the output
Layer includes 1 node, is used between the input layer and the intermediate layer, and between the intermediate layer and the output layer
Full connection mode；
The brain wave and corresponding catacleisis data of controller are gathered, slow α wave powers percentage is obtained according to the brain wave
Than, α ripples and the power ratio of the power ratio, θ ripples and slow α ripples of β ripples, according to the PERCLOS that the catacleisis data are obtained
Value measurement result, and multiple samples are generated, each sample includes power ratio, the θ ripples of slow α wave powers percentage, α ripples and β ripples
With the power ratio of slow α ripples and corresponding PERCLOS values measurement result；
A sample is chosen from the sample of generation, by power ratio, the θ of slow α wave powers percentage, α ripples and β ripples in sample
The power ratio of ripple and slow α ripples is input into the BP neural network model, obtains PERCLOS value estimation results；
According to the PERCLOS values estimation results and the error of the PERCLOS value measurement results of the sample chosen, the BP is updated
The parameter of neural network model；
If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
3. method according to claim 2, it is characterised in that described to be obtained according to the catacleisis data
PERCLOS value measurement results, including：
The upper palpebra inferior ultimate range under controller's waking state is obtained from the catacleisis data, the eyelid is closed
Data are closed for catacleisis amplitude changes with time, the catacleisis amplitude is the distance between upper palpebra inferior；
By the catacleisis data divided by the upper palpebra inferior ultimate range, catacleisis degree is obtained；
According to the catacleisis degree, the closed-eye time in the unit of account time；
Closed-eye time is obtained into PERCLOS value measurement results divided by the unit interval.
4. method according to claim 3, it is characterised in that described according to the catacleisis degree, during unit of account
Interior closed-eye time, including：Within the unit interval, catacleisis degree is more than the total of 70% or 80% corresponding time period
Be the closed-eye time in the unit interval.
5. method according to claim 2, it is characterised in that if reaching preset stopping condition, terminates training, otherwise weighs
New sample of choosing is trained again, including：According to the PERCLOS values in the PERCLOS values estimation results and the sample chosen
Measurement result obtains global error, if the global error reaches default maximum times less than error threshold or frequency of training,
Then terminate training, sample is otherwise chosen again and is trained again.
6. method according to claim 2, it is characterised in that according to the PERCLOS values estimation results and the sample chosen
The error of this PERCLOS value measurement results, updates the parameter of the BP neural network model, including：
Calculate the output error of the PERCLOS value measurement results in the sample of the PERCLOS values estimation results and selection；
Partial derivative according to the output error relative to intermediate layer to each side right value of output layer, updates the intermediate layer to output
Each side right value of layer；
Partial derivative according to the output error relative to input layer to each side right value in intermediate layer, updates the input layer to centre
Each side right value of layer；
According to the output error relative to the partial derivative that output layer is biased, the output layer biasing is updated；
According to the output error relative to the partial derivative that intermediate layer biases, the intermediate layer biasing is updated.
7. method according to claim 1 and 2, it is characterised in that described that slow α wave powers hundred are obtained according to the brain wave
Point than, α ripples and the power ratio of the power ratio, θ ripples and slow α ripples of β ripples, including：
Gather the brain wave of controller；
Wavelet Denoising Method is carried out to the brain wave；
Power ratio, θ ripples and the slow α ripples of slow α wave powers percentage, α ripples and β ripples are obtained according to the later brain wave of Wavelet Denoising Method
Power ratio.
8. a kind of controller's fatigue detecting system based on BP neural network, it is characterised in that including：
Brain wave processing module, the brain wave for gathering controller obtains slow α wave powers percentage, α according to the brain wave
The power ratio of the power ratio, θ ripples and slow α ripples of ripple and β ripples；
Fatigue data output module, for by the power ratio of the slow α wave powers percentage, the α ripples and β ripples, the θ ripples and
The power ratio of slow α ripples is input into the good BP neural network model of training in advance, obtains PERCLOS value simulation results；
Tired judge module, if being more than fatigue threshold for the PERCLOS values simulation result, judges that the controller is in
Fatigue state.
9. system according to claim 8, it is characterised in that be also used for including BP neural network model instruction module：
Set up BP neural network model and generate the parameter of the BP neural network model at random, the BP neural network model bag
Include input layer, intermediate layer, output layer；The input layer includes 3 nodes, and the intermediate layer includes multiple nodes, the output
Layer includes 1 node；Used between the input layer and the intermediate layer, and between the intermediate layer and the output layer
Full connection mode；
The brain wave and corresponding catacleisis data of controller are gathered, slow α wave powers percentage is obtained according to the brain wave
Than, α ripples and the power ratio of the power ratio, θ ripples and slow α ripples of β ripples, according to the PERCLOS that the catacleisis data are obtained
Value measurement result, and multiple samples are generated, each sample includes power ratio, the θ ripples of slow α wave powers percentage, α ripples and β ripples
With the power ratio of slow α ripples and corresponding PERCLOS values measurement result；
A sample is chosen from the sample of generation, by power ratio, the θ of slow α wave powers percentage, α ripples and β ripples in sample
The power ratio of ripple and slow α ripples is input into the BP neural network model, obtains PERCLOS value estimation results；
According to the PERCLOS values estimation results and the error of the PERCLOS value measurement results of the sample chosen, the BP is updated
The parameter of neural network model；
If reaching preset stopping condition, terminate training, sample is otherwise chosen again and is trained again.
10. system according to claim 9, it is characterised in that in BP neural network model instruction module, described
The PERCLOS value measurement results obtained according to the catacleisis data, including：
The upper palpebra inferior ultimate range under controller's waking state is obtained from the catacleisis data, the eyelid is closed
Data are closed for catacleisis amplitude changes with time, the catacleisis amplitude is the distance between upper palpebra inferior；
By the catacleisis data divided by the upper palpebra inferior ultimate range, catacleisis degree is obtained；
According to the catacleisis degree, the closed-eye time in the unit of account time；
Closed-eye time is obtained into PERCLOS value measurement results divided by the unit interval.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611115813.0A CN106691440A (en) | 2016-12-07 | 2016-12-07 | Controller fatigue detection method and system based on BP neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611115813.0A CN106691440A (en) | 2016-12-07 | 2016-12-07 | Controller fatigue detection method and system based on BP neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106691440A true CN106691440A (en) | 2017-05-24 |
Family
ID=58937437
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611115813.0A Pending CN106691440A (en) | 2016-12-07 | 2016-12-07 | Controller fatigue detection method and system based on BP neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106691440A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107316083A (en) * | 2017-07-04 | 2017-11-03 | 北京百度网讯科技有限公司 | Method and apparatus for updating deep learning model |
| CN107316532A (en) * | 2017-05-27 | 2017-11-03 | 西南交通大学 | The method of testing and system of dispatcher's inferential capability |
| CN107657868A (en) * | 2017-10-19 | 2018-02-02 | 重庆邮电大学 | A kind of teaching tracking accessory system based on brain wave |
| CN107890338A (en) * | 2017-09-22 | 2018-04-10 | 杭州爱上伊文化创意有限公司 | A kind of breast development data collecting system and its underwear |
| CN108090698A (en) * | 2018-01-08 | 2018-05-29 | 聚影汇（北京）影视文化有限公司 | A kind of film test and appraisal service system and method |
| CN109425669A (en) * | 2017-09-01 | 2019-03-05 | 中国民用航空局民用航空医学中心 | A kind of method that liquid chromatography-mass spectrometry screens degree of fatigue associated biomarkers in human body fluid |
| CN109425670A (en) * | 2017-09-01 | 2019-03-05 | 中国民用航空局民用航空医学中心 | A method of teams and groups' degree of fatigue is detected based on human urine |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101596101A (en) * | 2009-07-13 | 2009-12-09 | 北京工业大学 | Judge the method for fatigue state according to EEG signals |
| CN105139070A (en) * | 2015-08-27 | 2015-12-09 | 南京信息工程大学 | Fatigue driving evaluation method based on artificial nerve network and evidence theory |
| CN105286890A (en) * | 2015-09-22 | 2016-02-03 | 江西科技学院 | Driver sleepy state monitoring method based on electroencephalogram signal |
-
2016
- 2016-12-07 CN CN201611115813.0A patent/CN106691440A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101596101A (en) * | 2009-07-13 | 2009-12-09 | 北京工业大学 | Judge the method for fatigue state according to EEG signals |
| CN105139070A (en) * | 2015-08-27 | 2015-12-09 | 南京信息工程大学 | Fatigue driving evaluation method based on artificial nerve network and evidence theory |
| CN105286890A (en) * | 2015-09-22 | 2016-02-03 | 江西科技学院 | Driver sleepy state monitoring method based on electroencephalogram signal |
Non-Patent Citations (1)
| Title |
|---|
| GANG LI等: "Estimation of eye closure degree using EEG sensors and its application in drive drowsiness detection", 《SENSORS》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107316532A (en) * | 2017-05-27 | 2017-11-03 | 西南交通大学 | The method of testing and system of dispatcher's inferential capability |
| CN107316083A (en) * | 2017-07-04 | 2017-11-03 | 北京百度网讯科技有限公司 | Method and apparatus for updating deep learning model |
| CN107316083B (en) * | 2017-07-04 | 2021-05-25 | 北京百度网讯科技有限公司 | Method and apparatus for updating deep learning model |
| US11640550B2 (en) | 2017-07-04 | 2023-05-02 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for updating deep learning model |
| CN109425669A (en) * | 2017-09-01 | 2019-03-05 | 中国民用航空局民用航空医学中心 | A kind of method that liquid chromatography-mass spectrometry screens degree of fatigue associated biomarkers in human body fluid |
| CN109425670A (en) * | 2017-09-01 | 2019-03-05 | 中国民用航空局民用航空医学中心 | A method of teams and groups' degree of fatigue is detected based on human urine |
| CN109425670B (en) * | 2017-09-01 | 2022-09-16 | 中国民用航空局民用航空医学中心 | Method for detecting fatigue degree of team based on human urine |
| CN109425669B (en) * | 2017-09-01 | 2022-09-16 | 中国民用航空局民用航空医学中心 | Method for screening biomarkers related to fatigue degree in human body fluid by liquid chromatography-mass spectrometry |
| CN107890338A (en) * | 2017-09-22 | 2018-04-10 | 杭州爱上伊文化创意有限公司 | A kind of breast development data collecting system and its underwear |
| CN107657868A (en) * | 2017-10-19 | 2018-02-02 | 重庆邮电大学 | A kind of teaching tracking accessory system based on brain wave |
| CN108090698A (en) * | 2018-01-08 | 2018-05-29 | 聚影汇（北京）影视文化有限公司 | A kind of film test and appraisal service system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106691440A (en) | Controller fatigue detection method and system based on BP neural network | |
| CN106599821A (en) | Controller fatigue detection method and system based on BP neural network | |
| CN106539581B (en) | Controller's fatigue detection method and system based on probabilistic method | |
| CN101583313B (en) | Awake state judging model making device, awake state judging device, and warning device | |
| CN103824420B (en) | Fatigue driving identification system based on heart rate variability non-contact measurement | |
| CN104063719B (en) | Pedestrian detection method and device based on depth convolutional network | |
| CN110322423A (en) | A kind of multi-modality images object detection method based on image co-registration | |
| CN109635875A (en) | A kind of end-to-end network interface detection method based on deep learning | |
| CN109730637A (en) | A kind of face face-image quantified system analysis and method | |
| CN106886792A (en) | A kind of brain electricity emotion identification method that Multiple Classifiers Combination Model Based is built based on layering | |
| CN105488456B (en) | Method for detecting human face based on adaptive threshold adjustment rejection sub-space learning | |
| CN110197169A (en) | A kind of contactless learning state monitoring system and learning state detection method | |
| CN106128032A (en) | A kind of fatigue state monitoring and method for early warning and system thereof | |
| CN108664947A (en) | A kind of fatigue driving method for early warning based on Expression Recognition | |
| CN104103033B (en) | View synthesis method | |
| CN106650688A (en) | Eye feature detection method, device and recognition system based on convolutional neural network | |
| CN107169998A (en) | A kind of real-time tracking and quantitative analysis method based on hepatic ultrasound contrast enhancement image | |
| CN109389806A (en) | Fatigue driving detection method for early warning, system and medium based on multi-information fusion | |
| CN107832802A (en) | Quality of human face image evaluation method and device based on face alignment | |
| CN106909220A (en) | A kind of sight line exchange method suitable for touch-control | |
| CN110119672A (en) | A kind of embedded fatigue state detection system and method | |
| CN108596087B (en) | Driving fatigue degree detection regression model based on double-network result | |
| CN102902971A (en) | Method and system for conducting statistics on elevator visitor flow based on intelligent visual perception | |
| CN106570516A (en) | Obstacle recognition method using convolution neural network | |
| CN113743471B (en) | Driving evaluation method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170524 |
|
| RJ01 | Rejection of invention patent application after publication |