CN109919298A - A kind of airfield runway cutting automatic identification and measurement method based on long memory network and Naive Bayes Classifier in short-term - Google Patents
A kind of airfield runway cutting automatic identification and measurement method based on long memory network and Naive Bayes Classifier in short-term Download PDFInfo
- Publication number
- CN109919298A CN109919298A CN201910124537.1A CN201910124537A CN109919298A CN 109919298 A CN109919298 A CN 109919298A CN 201910124537 A CN201910124537 A CN 201910124537A CN 109919298 A CN109919298 A CN 109919298A
- Authority
- CN
- China
- Prior art keywords
- recess
- model
- groovenet
- short
- seam
- 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.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
- Length Measuring Devices By Optical Means (AREA)
Abstract
The present invention relates to a kind of airfield runway cutting automatic identifications and measurement method based on long memory network and Naive Bayes Classifier in short-term, acquire airport runway surface elevation profile information by vehicle-mounted laser profiler device.According to the relevant feature of acquisition data, design GrooveNet model, the model is used for the identification of airfield runway recess, and the model first traverses the starting point for determining each recess of airfield runway using a bounding box in whole segment data, and the size of recess is then calculated according to the position of starting point；The judgement for belonging to cutting or seam for the airfield runway recess recognized, classifies to it using Naive Bayes Classifier；The strategy of a comparison probability is formulated finally to improve the accuracy of recess classification.The present invention may be implemented to airfield runway cutting automatic identification and classification, and then carry out efficiently safely to road face, and objective appraisal improves the accuracy of measurement.
Description
Technical field
The present invention relates to road automatic measurement technique fields, especially a kind of to be based on long memory network in short-term and simple pattra leaves
The airfield runway cutting automatic identification and measurement method of this classifier.
Background technique
Currently, for the evaluation of airfield runway performance, usually there are three types of methods both at home and abroad: first is using simulation and emulation
Method come estimate vehicle occur neatly when critical speed；Second for smooth instrument trailer, swimming cloths test carriage and
Runway friction test vehicle etc. carries out field exploring coefficient of friction；Third is the size (width, depth and spacing) according to cutting
To determine the performance of airfield runway.The calculation method of groove dimensions need to only be obtained by starting point and deepest point.
For starting point determination method there are mainly two types of: first determines for the method based on filter, uses first
Filter obtains filtered outline data, in the case where given threshold, according to initial data and filtered outline data
Depth displacement determine whether to be herein cutting, if then positive negative slope joint is starting point, but this method is set due to threshold value
Fixed subjectivity is easy the shallower recess of missing inspection depth.Second is the method based on gradient to determine, resulting by that will calculate
The gradient value of adjacent two o'clock obtains two endpoints of recess compared with the threshold value of setting, but this method has spike in recess
In the case where the internal point of recess can be mistaken for endpoint.Third is the method based on cluster to determine, the method is customized
In the case where one threshold value, judge that data point belongs to the interior point or exterior point of cutting according to the slope of adjacent two o'clock, it is inside and outside
The junction of point is starting point, to calculate the size of cutting.The method is not suitable for the cutting of serious wear.These methods
Problem all not strong enough in the prevalence of generalization ability, for shallower recess, the recess of serious wear and irregular recessed
It falls into.Therefore a kind of method that urgently accuracy rate is high, generalization ability is strong determines the starting point of recess.
The antiskid performance evaluation of airfield runway after identifying and positioning recess, is also needed to recessed as unit of slab
It is trapped into capable classification, the position for correctly finding seam is a critical issue of antiskid Performance Evaluation.Currently used threshold value
Method is easy to appear classification error when encountering serious wear or irregular recess.
Airfield runway is usually to be safeguarded according to slab one by one, therefore the position determination of seam is also most important.It carves
Slot and the differentiating method of seam are usually the feature larger compared with depth and spacing according to the depth of seam, and given threshold connects to distinguish
Seam and cutting, the disadvantages of this method are according to said method to be easy to appear misjudgement when seam two sides serious wear.
In conclusion the method for existing airfield runway cutting identification and measurement has some limitations.Main problem
It is that the effect that cutting more serious for abrasion, that shape is irregular and depth is shallower identifies is not ideal enough, generalization ability is inadequate
By force.
Summary of the invention
In view of this, the purpose of the present invention is to propose to a kind of based on long memory network in short-term and Naive Bayes Classifier
Airfield runway cutting automatic identification and measurement method carry out data acquisition using laser contour equipment, and propose an energy certainly
The dynamic model for identifying and positioning recess, it is last to be classified again using Naive Bayes Classifier to recess；It can effectively improve
The recall rate of cutting reduces the quantity of cutting missing inspection and erroneous detection.
The present invention is realized using following scheme: a kind of airport based on long memory network and Naive Bayes Classifier in short-term
Runway cutting automatic identification and measurement method, comprising the following steps:
Step S1: laser displacement sensor is provided, and is acquired using the laser displacement sensor and is recessed on airfield runway
Altitude data；
Step S2: building GrooveNet model is simultaneously trained the GrooveNet model；
Step S3: GrooveNet model trained in step S2 is treated as into a mobile bounding box, in a whole segment data
On traversed with step-length 1；Output valve after traversal is built into a string of sequences being made of 0 and 1, the sequence that the output valve is constituted
0 and 1 position having a common boundary determines the starting point of recess in column；
Step S4: the position of recess described in aligning step S3；
Step S5: classify to the recess；
Step S6: sorted result in step S5 is corrected.
Further, the step S2 specifically includes the following steps:
Step S21: database is established: by carrying out probability statistical analysis to groove dimensions data, with specimen width for 50
A pixel value is that 50 can and be only capable of comprising a recess, the size the reason of constructing the database of recess, select 50
Selection be the key that the later period determine recess starting point.The database is divided into both positive and negative sample, wherein complete comprising one
The sample of whole recess is known as positive sample, on the contrary then be negative sample；
Step S22: cell factory size in long short-term memory artificial neural network network layers: the selection of cell factory size is selected
Should match with trained and test sample size, thus in long short-term memory artificial neural network network layers cell factory size selection
It is 50；
Step S23: select the GrooveNet prototype network number of plies: GrooveNet is a disaggregated model, the last layer setting
For full articulamentum, so that the output valve of model is single value i.e. 0 or 1；One layer long short-term memory artificial neural network is used first
Layer is mutually arranged in pairs or groups with the full articulamentum, meanwhile, then add one layer long short-term memory artificial neural network network layers more, it is quasi- to improve model
Close effect；Finally, model initial structure is determined as two layers long short-term memory artificial neural network network layers and a full articulamentum；
Step S24: avoid model from over-fitting occur: over-fitting is in test data, and model performance is good,
But in test data, effect is poor.In order to avoid this phenomenon, in the long short-term memory artificial neural network of each layer
All add one layer Dropout layers in layer end；Dropout layers can randomly choose several nodes and information is enabled to fail, and avoid model
The overfitting in training data；The model framework for finally determining GrooveNet at this time is long short-term memory artificial neural network
+ Dropout layers of layer+long+Dropout layers of short-term memory artificial neural network network layers+full articulamentum；
Step S25: the training GrooveNet model: GrooveNet model first passes through propagated forward and obtains a prediction
Value, predicted value obtains a penalty values compared with true value, then the power of the GrooveNet model is adjusted by backpropagation
Weight, to reduce penalty values；Propagated forward is combined with backpropagation, is constantly adjusted the weight of the GrooveNet model, is made
The GrooveNet model output value is close with predicted value；Finally save the weighted value of the GrooveNet model.
Further, determine that the formula such as (1) of recess starting point is shown in the step S3:
SP indicates that the start position of recess, EP indicate the final position of recess, xjIndicate in model output value 1 and 0 friendship
Boundary position, xiIndicate in model output value 0 and 1 boundary position.
Further, the particular content of the position of recess is corrected described in step S4 are as follows: originate point when front and back two is recessed
When setting in the presence of the part to overlap, two recess need to be merged into a recess, wherein the start position being newly recessed is the latter
The position of the starting point of recess, and the final position being newly recessed is the terminal of previous recess.
Further, step S5 specifically includes the following steps:
Step S51: characteristic of division is selected.Cutting and the difference of seam are: seam is between left and right away from meeting under normal circumstances
It is greater than cutting；Seam is deeper than groove depth, and width is wider.Therefore made herein with left spacing, right spacing, depth and width
For characteristic of division.
Step S52: the probability that the recess under single features belongs to cutting or seam, calculation formula such as (2) are calculated separately
(3) shown in；
Step 53: calculating four characteristic of division, that is, left spacing, right spacing, depth and width, merge lower recess and belong to cutting
Or the probability of seam, calculation formula is such as shown in (4) (5)；
Step 54: comparing the probability that recess belongs to seam and recess belongs to cutting, obtain preliminary classification result.
Further, the particular content of the step S6 are as follows: when seam continuously occurs in preliminary classification results, compare
On the contrary the two, which are recessed, belongs to the probability of seam, and the big person of probability is then used as true seam, then be cutting.
Compared with prior art, the invention has the following beneficial effects:
The present invention can effectively improve the recall rate of cutting, reduce the quantity of cutting missing inspection and erroneous detection, to be maintenance portion
Door assessment track performance provides reliable data and supports.
Detailed description of the invention
Fig. 1 is the LSTM architecture diagram of present example.
Fig. 2 is the architecture diagram of the GrooveNet of present example.
Fig. 3 is the positive sample exemplary diagram of present example.
Fig. 4 is the negative sample exemplary diagram of present example.
Specific embodiment
With reference to the accompanying drawing and example the present invention will be further described.
In this example, overall technological scheme is as follows:
(1) the airfield runway recess automatic identification based on long memory network in short-term
The characteristics of this example depends on front and back data dependence for the position of recess starting point, devises one by five layers
Neural network constitute model (as shown in Figure 1, 2), wherein comprising two layers grow in short-term memory network, two layers dropout layers and
One layer of full articulamentum, and the database (positive negative sample is as shown in Figure 3,4) of recess is constructed for 50 for model training with width.
In given new data, the model trained can return to an output valve (0 or 1).0 represents in data not comprising recess, 1 representative
Include recess in data.Project can realize the automatic identification of recess by this disaggregated model.Due to the multiplicity being recessed in training library
Property can preferably identify shallower recess, the recess of serious wear and irregular recessed so that model has higher recall rate
It falls into, generalization ability is stronger.
(2) airfield runway based on mobile bounding box, which is recessed, to be automatically positioned
Model after training is step-length with 1 by this example, in the whole segment data set of laser contour equipment acquisition back
Upper traversal, bounding box can obtain a succession of sequence being made of 0 and 1 after having traversed one piece of data；The position of 0 and 1 intersection
Set the key node for exactly calculating recess starting point.It has determined recess starting point and then recess has been corrected, corrected herein
Purpose be region is had coincidence recess merger be one recess.The output valve of model is combined and can be calculated with mobile bounding box
The starting point being recessed out, so that it is determined that the position of recess, and calculate the size of recess.
(3) the recess classification based on naive Bayesian
This example has chosen width, depth and between left and right away from this four incoherent features as Naive Bayes Classification
The characteristic of division of device, and trained library is constructed for disaggregated model.Model after training can calculate separately each recess and belong to quarter
Slot and the probability for belonging to seam, to obtain preliminary classification results.The recess adjacent with true seam because there is biggish spacing,
Therefore be easy to be mistaken for seam, for this phenomenon, project, which has been formulated a strategy and worked as, continuous seam in preliminary classification results
When appearance, the probability that the two recess belong to seam is compared, the high person of probability is as true seam.
Specifically, this example provide it is a kind of based on the long airfield runway of memory network and Naive Bayes Classifier in short-term
Cutting automatic identification and measurement method, comprising the following steps:
Step S1: laser displacement sensor is provided, and is acquired using the laser displacement sensor and is recessed on airfield runway
Altitude data；
Step S2: building GrooveNet model is simultaneously trained the GrooveNet model；
Step S21: database is established: by carrying out probability statistical analysis to groove dimensions data, with specimen width for 50
A pixel value is that 50 can and be only capable of comprising a recess, the size the reason of constructing the database of recess, select 50
Selection be the key that the later period determine recess starting point.The database is divided into both positive and negative sample, wherein complete comprising one
The sample of whole recess is known as positive sample, on the contrary then be negative sample；
Step S22: the selection of cell factory size in long short-term memory artificial neural network network layers: the choosing of cell factory size
Selecting should match with trained and test sample size, therefore the size of cell factory is selected in long short-term memory artificial neural network network layers
It is selected as 50.
The step S23:GrooveNet prototype network number of plies selects: GrooveNet is a disaggregated model, therefore the last layer
It is set as full articulamentum, so that the output valve of model is single value (i.e. 0 or 1).It is manually refreshing using one layer of long short-term memory first
It mutually arranges in pairs or groups through network layer and full articulamentum, due to depth deficiency, feature extraction not enough makes models fitting effect bad；Therefore it is more again
Add one layer long short-term memory artificial neural network network layers, models fitting effect is preferable at this time；In order to test model fitting effect whether
There can be the space continued to lift up, and be added to one layer long short-term memory artificial neural network network layers more, model accuracy is promoted at this time
Less, but parameter increase brings bigger calculating consumption, and model initial structure is determined as two layers of length in short-term after weighing the advantages and disadvantages
Remember artificial neural network network layers and a full articulamentum.
Step S24: avoid model from over-fitting occur: over-fitting is in test data, and model performance is good,
But in test data, effect is poor.In order to avoid this phenomenon, herein in the long short-term memory artificial neural network network layers of each layer
End is all added to one layer Dropout layers.Dropout layers can randomly choose several nodes and information is enabled to fail, and avoid model
The overfitting in training data.The model framework of GrooveNet is finally determined at this time are as follows: long short-term memory artificial neural network
+ Dropout layers of layer+long+Dropout layers of short-term memory artificial neural network network layers+full articulamentum.
Step S25: the training GrooveNet model: GrooveNet model first passes through propagated forward and obtains a prediction
Value, predicted value obtains a penalty values compared with true value, then the power of the GrooveNet model is adjusted by backpropagation
Weight, to reduce penalty values；Propagated forward is combined with backpropagation, is constantly adjusted the weight of the GrooveNet model, is made
The GrooveNet model output value is close with predicted value；Finally save the weighted value of the GrooveNet model.
Step S3: GrooveNet model trained in step S2 is treated as into a mobile bounding box, in a whole segment data
On with 1 for step-length traverse；Output valve after traversal is built into a string of sequences being made of 0 and 1, the sequence that the output valve is constituted
In 0 and 1 have a common boundary position determine recess starting point；
Step S4: the position of recess described in aligning step S3；
Step S5: classify to the recess；
Step S6: sorted result in step S5 is corrected.
In this example, specific embodiment is as follows:
(1) device parameter and its working principle
This example acquires the altitude data being recessed on airfield runway using laser contour equipment.The equipment is by laser displacement
Sensor composition, can accurately capture the feature of runway recess.The spot size of the sensor is 1 millimeter, measurement range
It is ± 200 millimeters, resolution ratio is 0.049 millimeter, uses rate for 32kHz.
(2) the airfield runway recess automatic identification based on long memory network in short-term
A) GrooveNet model
The model key core that this example proposes is LSTM layers.The key concept of LSTM is cell state and door knot
Structure.There are three types of the doors of type by LSTM: forgeing door, input gate and out gate.
Before step introduction, several important concepts in LSTM are introduced first.Input step-length: when data input model, it is
According to the one data input of a data, therefore whole process can include 50 input step-length (sizes that the present embodiment is selected
For 50)；Previous location mode: the advantages of LSTM, is that step-length 1 step-length t can be transmitted to the important information in step-length t-1
In, wherein previous location mode just refers to the information being retained in 1~t-1 of step-length；The discarding and reservation of information: it loses
It abandons and retains the significance level for referring to information, which is usually completed by sigmoid activation primitive, and discarding, which refers to, becomes 0 for information, is protected
Stay and point to information multiplied by remaining after 1, with should the output valve of sigmoid be 0.5 when, it is meant that current information is multiplied by 0.5
After be retained.New information: the input value of model when referring to t moment.
Step 1: determination needs how much information retained from previous location mode.This process is completed by forgetting door.Come
It is transmitted in sigmoid function simultaneously from the information of previous element state and information currently entered.Output one is between 0
And the value between 1, the value are to indicate that previous location mode needs are retained by how much percentage.Formula is such as shown in (1).
f(t)=σ (Wf[h(t-1),x(t)]+bf) (1)
Wherein, f(t)Indicate that original state needs the degree retained when t moment, σ indicates sigmoid activation primitive, WfIt indicates
Forget the weighted value of door, h(t-1)Indicate the location mode of previous moment, x(t)Indicate current input value, bfIt indicates to forget the inclined of door
Item is set, t indicates t moment.
Step 2: determining that new information needs are added in previous element state by how many significance level, the process is by defeated
Introduction is completed.Two steps can be divided into:
(1) it determines the significance level of new information: first passing the information of previous element state and information currently entered
It is delivered in sigmoid function.For output valve between 0 and 1, which indicates significance level.
(2) a candidate value vector is created, includes all information currently entered in candidate value vector.By preceding layer list
The status information and information currently entered of member are transmitted in tanh function, create a new candidate value vector.Formula is such as
(2), shown in (3).
i(t)=σ (Wi[h(t-1),x(t)]+bi) (2)
C (t)=tanh (Wa[h(t-1),x(t)]+ba) (3)
i(t)Indicate the significance level of new information when t moment, WiIndicate the weighted value of input gate, biIndicate input gate
Bias term, C(t)Indicate the candidate value vector of t moment, WaIndicate the weighted value of candidate value vector, baIndicate the inclined of candidate value vector
Set item.
Step 3: updating unit state.The part is added by two parts.
(1) determine how much preceding cells state needs reservation.It is previous when first by the location mode and t moment at t-1 moment
State needs the degree retained to be multiplied point by point.
(2) determine that current input value needs are added in previous element state by the importance of how many degree.When by t
The candidate value multiplication of vectors of the significance level and t moment of new information when quarter.
Finally the status information of the previous element remained is added to obtain current unit with newly added information
State.Formula is such as shown in (4).
C(t)=C(t-1)*f(t)+i(t)*C (t) (4)
Step 4: determining that the output valve of active cell, information can one information of corresponding output when having flowed through a unit
Value.This process is completed by out gate.Determine that candidate output valve, candidate output valve are determined by current location mode first
It is fixed, the significance level of candidate's output valve is determined later, is finally multiplied the two to obtain final unit output valve.Formula is such as
(5), shown in (6).
o(t)=σ (Wo[h(t-1),x(t)]+bo) (5)
h(t)=o(t)*tanh(C(t)) (6)
o(t)Indicate the significance level of candidate's output valve when t moment, WoIndicate weight information, boIndicate bias term.
B) the training of model
Step 1: the foundation of database.This example, for 50 database to construct recess, database is divided into just with width
Minus two kinds of samples, it is on the contrary then be negative sample wherein be known as positive sample comprising a sample being completely recessed.Sample instantiation is as schemed
Show.
Step 2: training pattern.Initial model first passes through propagated forward and obtains a predicted value, predicted value and true value phase
Compare to obtain a penalty values, then adjust the weight of model by backpropagation, to achieve the purpose that reduce penalty values.Forward direction passes
It broadcasts and is combined with backpropagation, constantly adjust the weight of model, keep model output value close with predicted value.Last preservation model
Weighted value.
Whether trained model in this example can automatically identify in sample comprising a complete recess.
(3) airfield runway based on mobile bounding box, which is recessed, to be automatically positioned
Trained model is treated as a mobile bounding box by this example, is traversed in a whole segment data with 1 for step-length.It is defeated
Value is built into a string of sequences being made of 0 and 1 out.
A. the starting point that is recessed primarily determines.
0 and 1 position having a common boundary is the key point for determining starting point in the sequence that output valve is constituted.Specific reduction formula is such as
(7) shown in.
SP=xj+1
EP=49+xi-1 (7)
SP indicates that the start position of recess, EP indicate the final position of recess, xjIndicate in model output value 1 and 0 friendship
Boundary position, xiIndicate in model output value 0 and 1 boundary position.
B. the position of recess is corrected.
When front and back two is recessed initial point position in the presence of the part to overlap, two recess need to be merged into a recess,
The start position being wherein newly recessed is the position of the starting point of the latter recess, and the final position being newly recessed is previous recess
Terminal.
This example orients the starting point of recess by mobile bounding box, and the minimum point of elevation is defined as most between starting point
Deep, the size of recess is calculated by these three points.Wherein calculation formula such as (8) is shown.
(4) the recess classification based on naive Bayesian
A. the preliminary classification being recessed
This example has the characteristics of biggish width, depth and spacing according to seam compared with cutting, has chosen four points
Category feature (left spacing, right spacing, width and depth), wherein in order to guarantee to be independent from each other between four features, herein
Spacing refers to the terminal of previous recess and the distance between the starting point of the latter recess.One data of project build later
Library, to obtain the mean value and variance of four characteristic of division of cutting and seam.
Step 1: calculating separately the probability that single features lower recess belongs to cutting or seam, calculation formula such as (8) (9) institute
Show.
Step 2: calculating four characteristic of division and merge the probability that lower recess belongs to cutting or seam, calculation formula such as (10)
(11) shown in.
Step 3: comparing the probability that recess belongs to seam and recess belongs to cutting, obtain preliminary classification result.
B. the correction of classification results
Since the recess adjacent with true seam also has biggish spacing, therefore this kind of recess is easy to be missed in preliminary classification
It is judged to seam, this example has formulated an improvement strategy thus.When seam continuously occurs in preliminary classification results, compare this
On the contrary two are recessed and belong to the probability of seam, and the big person of probability is then used as true seam, then be cutting.
Preferably, this example combine laser contour equipment, propose one kind can automate, efficiently, accurately identify and survey
The method for measuring cutting.This method can be recessed with automatic identification, determined the starting point of recess, calculated the size of recess, and to recess
Classify.Data processor based on full automation, this method can be applied in the Performance Evaluation of airfield runway.
Meanwhile this example has fully considered that shallow-layer recess, the recess of serious wear and irregular recess are difficult to be detected
The truth of survey.This example targetedly establishes the database of model training, and the diversity of database makes the extensive energy of model
Power increases.The average recall rate of recess reaches 97.51%.It is close with true seam when this example is classified also directed to recess
Recess be easy classification error the case where, formulated a probability comparison method to promote the accuracy of classification.
It is essential for regularly carrying out assessment to the non-skid property of airfield runway, and this example provides for maintenance department
One efficient, accurate method, substantial saved manpower and material resources cost.
Particularly, this example has built the model of an entitled GrooveNet, which is made of five layers of neural network,
Including two layers LSTM layers, two layers dropout layers and one layer of full articulamentum.The model, can be to being given after training
Data carry out two classification, judge to data whether include recess, thus achieve the purpose that identification recess.In tranining database
Model generalization ability comprising different recess, therefore after training is stronger.
This example assigns GrooveNet as a bounding box, traverses whole segment data with 1 for step-length, can export a string at this time
The sequence being made of 0 and 1.And the position of 0 and 1 intersection then may be used to determine the starting point of recess.This example also ties positioning
Fruit is corrected, and the recess that position is overlapped mutually is merged.
This example classifies to recess using Naive Bayes Classifier, according to the different place of seam and cutting, choosing
Four characteristic of division (left spacing, right spacing, depth and width) is taken.And it calculates separately each recess and belongs to cutting and belong to and connect
The probability of seam compares two probability later and obtains preliminary classification results.This example is it is contemplated that adjacent with true seam is recessed
The truth for being easy to happen classification error is fallen into, has formulated a probability comparison method to correct preliminary classification result.The foregoing is merely
Preferred embodiments of the invention, it is all according to equivalent changes and modifications within the scope of the patent application of the present invention, it should all belong to culvert of the invention
Lid range.
Claims (6)
1. a kind of based on the long airfield runway cutting automatic identification of memory network and Naive Bayes Classifier in short-term and measurement side
Method, it is characterised in that: the following steps are included:
Step S1: laser displacement sensor is provided, and utilizes the height being recessed on laser displacement sensor acquisition airfield runway
Number of passes evidence；
Step S2: building GrooveNet model is simultaneously trained the GrooveNet model；
Step S3: by GrooveNet model trained in step S2 treat as a mobile bounding box, in a whole segment data with
Step-length 1 is traversed；Output valve after traversal is built into a string of sequences being made of 0 and 1, in the sequence that the output valve is constituted
0 and 1 position having a common boundary determines the starting point of recess；
Step S4: the position of recess described in aligning step S3；
Step S5: classify to the recess；
Step S6: sorted result in step S5 is corrected.
2. according to claim 1 a kind of based on the long quarter of the airfield runway of memory network and Naive Bayes Classifier in short-term
Slot automatic identification and measurement method, it is characterised in that: the step S2 specifically includes the following steps:
Step S21: database is established: by carrying out probability statistical analysis to groove dimensions data, with specimen width for 50 pictures
Plain value is that 50 can and be only capable of being recessed comprising one, the choosing of the size the reason of constructing the database of recess, select 50
Select is the key that determine recess starting point the later period.The database is divided into both positive and negative sample, wherein complete recessed comprising one
Sunken sample is known as positive sample, on the contrary then be negative sample；
Step S22: select cell factory size in long short-term memory artificial neural network network layers: the selection of cell factory size should be with
Trained and test sample size matches, therefore cell factory is sized in long short-term memory artificial neural network network layers
50；
Step S23: select the GrooveNet prototype network number of plies: GrooveNet is a disaggregated model, and the last layer is set as complete
Articulamentum, so that the output valve of model is single value i.e. 0 or 1；First using one layer long short-term memory artificial neural network network layers with
The full articulamentum is mutually arranged in pairs or groups, meanwhile, then add one layer long short-term memory artificial neural network network layers more, to improve models fitting effect
Fruit；Finally, model initial structure is determined as two layers long short-term memory artificial neural network network layers and a full articulamentum；
Step S24: avoid model from over-fitting occur: over-fitting is in test data, and model performance is good, but
In test data, effect is poor.In order to avoid this phenomenon, at the long short-term memory artificial neural network network layers end of each layer
All add one layer Dropout layers in end；Dropout layers can randomly choose several nodes and information is enabled to fail, and model is avoided to instruct
Practice overfitting in data；At this time finally determine GrooveNet model framework be long short-term memory artificial neural network network layers+
Dropout layers+long+Dropout layers of short-term memory artificial neural network network layers+full articulamentum；
Step S25: the training GrooveNet model: GrooveNet model first passes through propagated forward and obtains a predicted value,
Predicted value obtains a penalty values compared with true value, then the weight of the GrooveNet model is adjusted by backpropagation,
To reduce penalty values；Propagated forward is combined with backpropagation, is constantly adjusted the weight of the GrooveNet model, is made institute
It is close with predicted value to state GrooveNet model output value；Finally save the weighted value of the GrooveNet model.
3. according to claim 1 a kind of based on the long quarter of the airfield runway of memory network and Naive Bayes Classifier in short-term
Slot automatic identification and measurement method, it is characterised in that: determine that the formula such as (1) of recess starting point is shown in the step S3:
SP indicates that the start position of recess, EP indicate the final position of recess, xjIndicate in model output value 1 and 0 boundary position
It sets, xiIndicate in model output value 0 and 1 boundary position.
4. according to claim 1 a kind of based on the long quarter of the airfield runway of memory network and Naive Bayes Classifier in short-term
Slot automatic identification and measurement method, it is characterised in that: the particular content of the position of recess is corrected described in step S4 are as follows: work as front and back
When two recess initial point positions have the part to overlap, two recess need to be merged into a recess, wherein what is be newly recessed rises
Point is set to the position of the starting point of the latter recess, and the final position being newly recessed is the terminal of previous recess.
5. according to claim 1 a kind of based on the long quarter of the airfield runway of memory network and Naive Bayes Classifier in short-term
Slot automatic identification and measurement method, it is characterised in that: step S5 specifically includes the following steps:
Step S51: characteristic of division is selected.Cutting and the difference of seam are: seam is between left and right away from meeting than carving under normal circumstances
Slot is greater；Seam is deeper than groove depth, and width is wider.Therefore herein using left spacing, right spacing, depth and width as divide
Category feature.
Step S52: the probability that the recess under single features belongs to cutting or seam, calculation formula such as (2) (3) institute are calculated separately
Show；
Step 53: calculating four characteristic of division, that is, left spacing, right spacing, depth and width, merge lower recess and belong to cutting or connect
The probability of seam, calculation formula is such as shown in (4) (5)；
Step 54: comparing the probability that recess belongs to seam and recess belongs to cutting, obtain preliminary classification result.
6. according to claim 5 a kind of based on the long quarter of the airfield runway of memory network and Naive Bayes Classifier in short-term
Slot automatic identification and measurement method, it is characterised in that: the particular content of the step S6 are as follows: when seam in preliminary classification results
Continuously when occurring, compare the probability that the two recess belong to seam, on the contrary the big person of probability is then used as true seam, then be cutting.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910124537.1A CN109919298B (en) | 2019-02-19 | 2019-02-19 | Airport runway grooving automatic identification and measurement method based on long-short-time memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910124537.1A CN109919298B (en) | 2019-02-19 | 2019-02-19 | Airport runway grooving automatic identification and measurement method based on long-short-time memory network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109919298A true CN109919298A (en) | 2019-06-21 |
| CN109919298B CN109919298B (en) | 2023-04-28 |
Family
ID=66961722
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910124537.1A Active CN109919298B (en) | 2019-02-19 | 2019-02-19 | Airport runway grooving automatic identification and measurement method based on long-short-time memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109919298B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751199A (en) * | 2019-10-15 | 2020-02-04 | 南京航空航天大学 | Satellite anomaly detection method based on Bayesian neural network |
| CN111368804A (en) * | 2020-03-31 | 2020-07-03 | 河北科技大学 | Lane line detection method, system and terminal equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107665279A (en) * | 2017-09-26 | 2018-02-06 | 福建农林大学 | Road face cutting automatic identifying method |
| US10037704B1 (en) * | 2017-02-01 | 2018-07-31 | David Myr | Automatic real-time air traffic control system and method for maximizing landings / takeoffs capacity of the airport and minimizing aircrafts landing times |
| CN108460744A (en) * | 2018-03-21 | 2018-08-28 | 武汉理工大学 | A kind of enhancing of cement cutting pavement image noise reduction and FRACTURE CHARACTERISTICS extracting method |
-
2019
- 2019-02-19 CN CN201910124537.1A patent/CN109919298B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10037704B1 (en) * | 2017-02-01 | 2018-07-31 | David Myr | Automatic real-time air traffic control system and method for maximizing landings / takeoffs capacity of the airport and minimizing aircrafts landing times |
| CN107665279A (en) * | 2017-09-26 | 2018-02-06 | 福建农林大学 | Road face cutting automatic identifying method |
| CN108460744A (en) * | 2018-03-21 | 2018-08-28 | 武汉理工大学 | A kind of enhancing of cement cutting pavement image noise reduction and FRACTURE CHARACTERISTICS extracting method |
Non-Patent Citations (2)
| Title |
|---|
| 叶珍: "基于ABAQUS的刻槽尺寸与路面磨损的相关性研究与应用", 《中国优秀硕士学位论文全文数据库》 * |
| 英红等: "基于BP神经网络的沥青路面裂缝识别方法研究", 《河南理工大学学报(自然科学版)》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110751199A (en) * | 2019-10-15 | 2020-02-04 | 南京航空航天大学 | Satellite anomaly detection method based on Bayesian neural network |
| CN111368804A (en) * | 2020-03-31 | 2020-07-03 | 河北科技大学 | Lane line detection method, system and terminal equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109919298B (en) | 2023-04-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111081016B (en) | Urban traffic abnormity identification method based on complex network theory | |
| CN103605362B (en) | Based on motor pattern study and the method for detecting abnormality of track of vehicle multiple features | |
| CN106997669B (en) | Method for judging traffic jam cause based on feature importance | |
| CN103473540B (en) | The modeling of intelligent transportation system track of vehicle increment type and online method for detecting abnormality | |
| CN105069779A (en) | Building ceramic surface pattern quality detection method | |
| CN102982559A (en) | Vehicle tracking method and system | |
| CN106650913A (en) | Deep convolution neural network-based traffic flow density estimation method | |
| CN102279929B (en) | Remote-sensing artificial ground object identifying method based on semantic tree model of object | |
| CN107255803A (en) | A kind of ship track logic start algorithm could | |
| Chiang et al. | Auto-control of pumping operations in sewerage systems by rule-based fuzzy neural networks | |
| CN106846816A (en) | A kind of discretization traffic state judging method based on deep learning | |
| CN107153846A (en) | A kind of road traffic state modeling method based on Fuzzy C-Means Cluster Algorithm | |
| CN105447504A (en) | Traffic mode behavior recognition method and corresponding recognition model construction method | |
| CN110135354A (en) | A kind of change detecting method based on outdoor scene threedimensional model | |
| CN109919298A (en) | A kind of airfield runway cutting automatic identification and measurement method based on long memory network and Naive Bayes Classifier in short-term | |
| CN113064450B (en) | Quantum particle swarm unmanned aerial vehicle path planning method based on annealing algorithm | |
| CN109472226A (en) | A kind of sleep behavioral value method based on deep learning | |
| CN110533103A (en) | A kind of lightweight wisp object detection method and system | |
| CN112365708A (en) | Scenic spot traffic volume prediction model establishing and predicting method based on multi-graph convolution network | |
| CN113989179B (en) | Train wheel set tread defect detection method and system based on target detection algorithm | |
| CN112819087B (en) | Method for detecting abnormality of BOD sensor of outlet water based on modularized neural network | |
| CN114297914A (en) | Deep neural network result credibility guaranteeing method for large power grid reliability evaluation | |
| CN109492816A (en) | A kind of coal and gas prominent dynamic prediction method based on hybrid intelligent | |
| CN105911553A (en) | Determination method and system for road feasible region | |
| CN110992375B (en) | Chord angle discrimination clustering single tree segmentation method for layer-by-layer LiDAR point cloud |
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 |