CN109919298B - Airport runway grooving automatic identification and measurement method based on long-short-time memory network - Google Patents
Airport runway grooving automatic identification and measurement method based on long-short-time memory network Download PDFInfo
- Publication number
- CN109919298B CN109919298B CN201910124537.1A CN201910124537A CN109919298B CN 109919298 B CN109919298 B CN 109919298B CN 201910124537 A CN201910124537 A CN 201910124537A CN 109919298 B CN109919298 B CN 109919298B
- Authority
- CN
- China
- Prior art keywords
- model
- layer
- long
- short
- depression
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Length Measuring Devices By Optical Means (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to an automatic identification and measurement method for airfield runway notch grooves based on a long-short-term memory network and a naive Bayesian classifier, which collects airfield runway surface elevation profile information through vehicle-mounted laser profiling equipment. According to the relevant characteristics of the collected data, a groonet model is designed, the model is used for identifying the pits of the airport runway, firstly, the model can utilize a bounding box to traverse and determine the starting points of all the pits of the airport runway on the whole section of data, and then the sizes of the pits are calculated according to the positions of the starting points; judging that the identified airfield runway depression belongs to the notch or seam, and classifying the identified airfield runway depression by adopting a naive Bayes classifier; finally, a strategy for comparing probabilities is formulated to improve the accuracy of the pit classification. The method can realize automatic identification and classification of the notch of the airport runway, further carry out efficient and objective evaluation on the safety of the runway surface, and improve the accuracy of measurement.
Description
Technical Field
The invention relates to the technical field of automatic road detection, in particular to an automatic identification and measurement method for airfield runway grooving based on a long-short-time memory network.
Background
At present, three methods are generally used for evaluating the performance of an airport runway at home and abroad: firstly, estimating critical speed of a vehicle when water skid occurs by adopting a simulation and emulation method; secondly, carrying out field investigation on friction coefficients by using a slipperiness instrument trailer, a surface friction testing vehicle, a runway friction testing vehicle and the like; third, the performance of the airfield runway is determined based on the size (width, depth, and pitch) of the score grooves. The calculation method for the groove size only needs to be obtained through the starting point and the deepest point.
There are two main methods for determining the starting point: the first method is to determine based on a filter, firstly, the filtered profile data is obtained by adopting the filter, and under the condition of setting a threshold value, whether the profile data is notched or not is judged according to the height Cheng Chalai of the original data and the filtered profile data, if yes, a positive slope intersection point and a negative slope intersection point are used as starting points, but due to subjectivity of the threshold value setting, the method is easy to miss a pit with shallow depth. The second is to determine by a gradient-based method, and the two end points of the recess are obtained by comparing the calculated gradient values of the two adjacent points with a set threshold value, but the method can misjudge the inner point of the recess as the end point under the condition that the peak exists in the recess. And thirdly, determining based on a clustering method, wherein the method judges whether the data points belong to inner points or outer points of the notch according to the slopes of two adjacent points under the condition of self-defining a threshold value, and the joint of the inner points and the outer points is a starting point, so that the size of the notch is calculated. This method is not suitable for grooving with severe wear. These methods have the general problem of insufficient generalization ability for shallower depressions, severely worn depressions, and irregular depressions. Therefore, a method with high accuracy and high generalization capability is needed to determine the starting point of the pit.
The anti-skid performance evaluation of the airfield runway takes slab as a unit, after the pit is identified and positioned, the pit is also required to be classified, and the accurate position finding of the joint is a key problem of the anti-skid performance evaluation. The currently used thresholding method is prone to classification errors when it encounters heavy or irregular pits.
Airport runways are usually maintained from slabs, and therefore seam location determination is also critical. The method for distinguishing the notch from the seam is generally to set a threshold value according to the characteristics of deeper depth and larger spacing of the seam, and has the defect that erroneous judgment is easy to occur when two sides of the seam are severely worn.
In summary, the existing method for identifying and measuring the notch of the airfield runway has certain limitations. The main problems are that the effect of identifying the notch with serious abrasion, irregular shape and shallow depth is not ideal enough, and the generalization capability is not strong enough.
Disclosure of Invention
In view of the above, the invention aims to provide an airport runway grooving automatic identification and measurement method based on a long-short-term memory network, adopts laser profiling equipment to collect data, provides a model capable of automatically identifying and positioning pits, and finally adopts a naive Bayesian classifier to classify the pits; the recall rate of grooving can be effectively improved, and the number of missed detection and false detection of grooving is reduced.
The invention is realized by adopting the following scheme: an automatic identifying and measuring method for airfield runway grooving based on a long-short-time memory network comprises the following steps:
step S1: providing a laser displacement sensor, and acquiring elevation data of the pit on the airfield runway by using the laser displacement sensor;
step S2: constructing a GrooveNet model and training the GrooveNet model;
step S3: taking the groovelet model trained in the step S2 as a mobile bounding box, and traversing the whole data with a step length of 1; constructing a series of 0 and 1 output values after traversing, wherein the positions of the 0 and 1 junctions in the series of the output values determine the starting point of the concave;
step S4: correcting the position of the recess in step S3;
step S5: classifying the depressions;
step S6: and correcting the result after classification in the step S5.
Further, the step S2 specifically includes the following steps:
step S21: establishing a database: by performing a probabilistic statistical analysis on the score size data, a database of depressions is constructed with a sample width of 50 pixel values, the choice 50 being made in that 50 can and can only contain one depression, the choice of size being the key to later determining the depression starting point. Dividing the database into positive and negative samples, wherein the sample containing one complete depression is called positive sample, and the negative sample is called negative sample;
step S22: selecting long-time memory of cell unit size in artificial neural network layer: the cell unit size is selected to match the size of the training and testing samples, so the size of the cell unit in the long-and-short-term memory artificial neural network layer is selected to be 50;
step S23: selecting the network layer number of the GrooveNet model: the GrooveNet is a classification model, and the last layer is set as a full-connection layer, so that the output value of the model is a single value, namely 0 or 1; firstly, matching a layer of long-short-time memory artificial neural network layer with the fully-connected layer, and simultaneously, adding a layer of long-short-time memory artificial neural network layer to improve the model fitting effect; finally, determining the model primary architecture as two layers of long-short-time memory artificial neural network layers and a full-connection layer;
step S24: the phenomenon of overfitting of the model is avoided: the overfitting phenomenon is that in the test data, the model performance is good, but in the test data, the effect is poor. In order to avoid the phenomenon, a Dropout layer is added at the tail end of each long-short-time memory artificial neural network layer; the Dropout layer can randomly select a plurality of nodes to disable information, so that the model is prevented from being excessively fitted in training data; at the moment, the model architecture of the GrooveNet is finally determined to be a long-short-time memory artificial neural network layer, a Dropout layer, a long-short-time memory artificial neural network layer, a Dropout layer and a full-connection layer;
step S25: training the groovent model: firstly, a predicted value is obtained through forward propagation of the GrooveNet model, a loss value is obtained through comparison of the predicted value and a true value, and then the weight of the GrooveNet model is adjusted through backward propagation so as to reduce the loss value; combining forward propagation and backward propagation, and continuously adjusting the weight of the GrooveNet model to enable the output value of the GrooveNet model to be similar to the predicted value; and finally, saving the weight value of the groovelet model.
Further, the formula for determining the starting point of the recess in step S3 is shown in (1):
SP represents the start position of the depression, EP represents the end position of the depression, x j Representing the boundary position of 1 and 0 in the model output value, x i Indicating the boundary position between 0 and 1 in the model output value.
Further, the specific contents of the correction recess position in step S4 are: when there is a portion where the start points of the front and rear recesses overlap each other, the two recesses are combined into one recess, wherein the start point of the new recess is the position of the start point of the subsequent recess, and the end point of the new recess is the end point of the previous recess.
Further, the step S5 specifically includes the following steps:
and S51, selecting classification features. The difference between the grooving and the joint is that: typically, the left-right spacing of the seam is greater than that of the score line; the seam is deeper than the score groove depth and wider. The left pitch, right pitch, depth, and width are therefore used herein as classification features.
Step S52: respectively calculating the probability that the concave part belongs to a notch or a seam under a single characteristic, wherein a calculation formula is shown as (2) and (3);
step 53: calculating four classification features, namely left spacing, right spacing, depth and width, combining the probability that the undercut belongs to a notch or a seam, wherein a calculation formula is shown as (4) (5);
step 54: and comparing the probability that the pit belongs to the joint and the probability that the pit belongs to the notch, and obtaining a preliminary classification result.
Further, the specific content of step S6 is as follows: when the seam continuously appears in the preliminary classification result, the probability that the two pits belong to the seam is compared, the probability is high, the seam is taken as a real seam, and otherwise, the seam is notched.
Compared with the prior art, the invention has the following beneficial effects:
the invention can effectively improve the recall rate of grooving and reduce the number of missed detection and false detection of grooving, thereby providing reliable data support for the maintenance department to evaluate the performance of the runway.
Drawings
Fig. 1 is an LSTM architecture diagram of an example of the present invention.
Fig. 2 is an architecture diagram of a groovent of an example of the invention.
Fig. 3 is a diagram of a positive example of an example of the present invention.
Fig. 4 is a negative example diagram of an example of the present invention.
Detailed Description
The invention is further illustrated by the following examples in conjunction with the accompanying drawings.
In this example, the overall technical solution is as follows:
(1) Airport runway depression automatic identification based on long-short-term memory network
The present example designs a model composed of five layers of neural networks (as shown in fig. 1 and 2) for the characteristic that the position of the starting point of the depression depends on the correlation of the front and back data, wherein the model comprises two long and short time memory networks, two dropout layers and one full connection layer, and a depression database (positive and negative samples are shown in fig. 3 and 4) is constructed with the width of 50 for model training. The trained model returns an output value (0 or 1) given the new data. 0 represents that no recess is included in the data, and 1 represents that recess is included in the data. The item can realize automatic identification of the dent through the classification model. Due to the diversity of the depressions in the training library, the model has higher recall rate, can better identify shallower depressions, serious-abrasion depressions and irregular depressions, and has stronger generalization capability.
(2) Airport runway pit automatic positioning based on movable bounding box
The embodiment takes 1 as step length of the trained model, traverses the whole section of data set acquired by the laser profiling equipment, and obtains a series of sequences consisting of 0 and 1 after the bounding box traverses the whole section of data; the position of the intersection of 0 and 1 is the key node for calculating the starting point of the depression. After determining the starting point of the recess, the recess is corrected again, where the purpose of the correction is to merge the recess with the region overlapping into one recess. The output of the model, combined with the moving bounding box, can calculate the starting point of the depression, thereby determining the position of the depression, and calculating the size of the depression.
(3) Naive Bayes-based concave classification
In the embodiment, four uncorrelated features of width, depth and left-right spacing are selected as classification features of a naive Bayesian classifier, and a training library is constructed for a classification model. The trained model calculates the probability that each depression belongs to the notch and the seam respectively, so that a preliminary classification result is obtained. The recesses adjacent to the actual joint are easily misjudged as joints because of larger spacing, and aiming at the phenomenon, a strategy is formulated by the project, namely when continuous joints appear in the primary classification result, the probability that the two recesses belong to the joints is compared, and the probability is high as the actual joint.
Specifically, the embodiment provides an airport runway grooving automatic identification and measurement method based on a long-short-time memory network, which comprises the following steps:
step S1: providing a laser displacement sensor, and acquiring elevation data of the pit on the airfield runway by using the laser displacement sensor;
step S2: constructing a GrooveNet model and training the GrooveNet model;
step S21, establishing a database: by performing a probabilistic statistical analysis on the score size data, a database of depressions is constructed with a sample width of 50 pixel values, the choice 50 being made in that 50 can and can only contain one depression, the choice of size being the key to later determining the depression starting point. Dividing the database into positive and negative samples, wherein the sample containing one complete depression is called positive sample, and the negative sample is called negative sample;
step S22, selecting cell unit sizes in the artificial neural network layer after long and short time storage: the cell size is chosen to match the size of the training and test samples, so the long and short term memory cell size in the artificial neural network layer is chosen to be 50.
Step S23, selecting the number of network layers of the GrooveNet model: the groovent is a classification model, so the last layer is set to be the fully connected layer, so that the output value of the model is a single value (i.e., 0 or 1). Firstly, a layer of long-short-term memory artificial neural network layer is matched with a full-connection layer, and the model fitting effect is poor due to insufficient depth and insufficient feature extraction; therefore, a layer of long-and-short-term memory artificial neural network layer is added, and the model fitting effect is better; in order to test whether the model fitting effect can have a space for continuously improving, a layer of long-short-term memory artificial neural network layer is additionally arranged, at the moment, the model accuracy is not improved greatly, but the parameter increase brings greater calculation consumption, and after weighing the advantages and disadvantages, the model primary framework is determined to be two layers of long-short-term memory artificial neural network layers and a full-connection layer.
Step S24: the phenomenon of overfitting of the model is avoided: the overfitting phenomenon is that in the test data, the model performance is good, but in the test data, the effect is poor. To avoid this, a Dropout layer is added at the end of each layer of long and short term memory artificial neural network. The Dropout layer can randomly select a plurality of nodes to disable information, and the model is prevented from being excessively fitted in training data. The model architecture of the groovent is finally determined at this time as follows: the long-time memory artificial neural network layer is connected with the Dropout layer, and the full connection layer is connected with the Dropout layer.
Step S25: training the groovent model: firstly, a predicted value is obtained through forward propagation of the GrooveNet model, a loss value is obtained through comparison of the predicted value and a true value, and then the weight of the GrooveNet model is adjusted through backward propagation so as to reduce the loss value; combining forward propagation and backward propagation, and continuously adjusting the weight of the GrooveNet model to enable the output value of the GrooveNet model to be similar to the predicted value; and finally, saving the weight value of the groovelet model.
Step S3: taking the groovelet model trained in the step S2 as a mobile bounding box, and traversing the whole data by taking 1 as a step length; constructing a series of 0 and 1 output values after traversing, wherein the positions of the 0 and 1 junctions in the series of the output values determine the starting point of the concave;
step S4: correcting the position of the recess in step S3;
step S5: classifying the depressions;
step S6: and correcting the result after classification in the step S5.
In this example, the specific embodiments are as follows:
(1) Device parameters and their working principle
The present example employs laser profiling equipment to collect elevation data for depressions on airport runways. The device is composed of a laser displacement sensor, and can capture the characteristics of the track depression with high precision. The spot size of the sensor is 1 millimeter, the measuring range is +/-200 millimeters, the resolution is 0.049 millimeter, and the adoption rate is 32kHz.
(2) Airport runway depression automatic identification based on long-short-term memory network
A) GrooveNet model
The key core of the model proposed by this example is the LSTM layer. The core concept of LSTM is in the cellular state and gate structure. LSTM has three types of gate structures, forget gate, input gate and output gate.
Before the step introduction, several important concepts in LSTM are first introduced. Inputting step length: in the data input model, the data is input one by one, so the whole process comprises 50 input steps (the selected size is 50); previous cell state: LSTM has the advantage of being able to transfer important information in step 1 to step t-1 into step t, where the previous cell state refers to the information that was retained in steps 1 to t-1; discarding and retaining of information: discarding and preserving refer to the importance of information, which is usually done by a sigmoid activation function, discarding refers to the information being changed to 0, preserving refers to the information being preserved after multiplying by 1, and similarly when the output value of sigmoid is 0.5, it means that the current information is preserved after multiplying by 0.5. New information: refers to the input value of the model at time t.
Step 1: it is determined how much information needs to be retained from the previous cell state. This process is done by a forgetful gate. Information from the previous cell state and the currently entered information are passed on to the sigmoid function at the same time. A value between 0 and 1 is output that indicates how much of the previous cell state needs to be preserved. The formula is shown as (1).
f (t) ＝σ(W f [h (t-1) ,x (t) ]+b f )(1)
Wherein f (t) Represents the degree to which the previous state needs to be preserved at time t, sigma represents the sigmoid activation function, W f Weight value indicating forgetting gate, h (t-1) Representing the state of the cell at the previous time, x (t) Representing the current input value, b f The bias term of the forgetting gate is shown, and t is the time t.
Step 2: determining how much importance new information needs to be added to the previous cell state is done by the input gate. The method can be divided into two steps:
(1) Determining the importance degree of the new information: the information of the previous cell state and the information of the current input are firstly transferred to the sigmoid function. The output value is between 0 and 1, which represents the degree of importance.
(2) Creating a candidate value vector which contains all information of the current input. The state information of the previous layer unit and the current input information are transferred to the tanh function to create a new candidate value vector. The formulas are shown as (2) and (3).
i (t) ＝σ(W i [h (t-1) ,x (t) ]+b i )(2)
C (t) ＝tanh(W a [h (t-1) ,x (t) ]+b a )(3)
i (t) Represents the importance degree of new information at time t, W i Representing the weight value of the input gate, b i Representing the bias term representing the input gate, C (t) A candidate value vector representing time t, W a Weight value representing candidate value vector, b a Representing the bias of candidate value vectorsSetting items.
Step 3: updating the cell state. The part is formed by adding two parts.
(1) It is determined how much of the previous cell state needs to be preserved. The state of the cell at time t-1 is first multiplied point by the degree to which the previous state needs to be preserved at time t.
(2) It is determined to what degree of importance the current input value needs to be added to the previous cell state. The importance of the new information at time t is multiplied by the candidate value vector at time t.
And finally, adding the state information of the previous cell which is reserved with the newly added information to obtain the current cell state. The formula is shown as (4).
C (t) ＝C (t-1) *f (t) +i (t) *C (t) (4)
Step 4: and determining the output value of the current unit, and correspondingly outputting an information value when the information flows through one unit. This is done by the output gate. Firstly, determining a candidate output value, wherein the candidate output value is determined by the current unit state, then determining the importance degree of the candidate output value, and finally multiplying the candidate output value and the importance degree to obtain a final unit output value. The formulas are shown as (5) and (6).
o (t) ＝σ(W o [h (t-1) ,x (t) ]+b o )(5)
h (t) ＝o (t) *tanh(C (t) )(6)
o (t) Indicating the importance level of the candidate output value at time t, W o Representing weight information, b o Representing the bias term.
B) Training of models
Step 1: and (3) establishing a database. The present example builds a concave database with a width of 50, and divides the database into positive and negative samples, wherein the sample containing one complete concave is called positive sample, and the negative sample is the negative sample. Sample examples are shown.
Step 2: and training a model. The initial model firstly obtains a predicted value through forward propagation, the predicted value is compared with a true value to obtain a loss value, and then the weight of the model is adjusted through backward propagation, so that the purpose of reducing the loss value is achieved. The forward propagation and the backward propagation are combined, and the weight of the model is continuously adjusted to enable the model output value to be similar to the predicted value. And finally, saving the weight value of the model.
The trained model in this example can automatically identify whether the sample contains a complete depression.
(3) Airport runway pit automatic positioning based on movable bounding box
The present example uses the trained model as a moving bounding box, traversing 1 step over a whole piece of data. The output values are constructed as a series of 0 and 1 sequences.
a. Preliminary determination of the starting point of the depression.
The position of the boundary between 0 and 1 in the sequence of output values is the key point for determining the starting point. The specific conversion formula is shown in (7).
SP＝x j +1
EP＝49+x i -1 (7)
SP represents the start position of the depression, EP represents the end position of the depression, x j Representing the boundary position of 1 and 0 in the model output value, x i Indicating the boundary position between 0 and 1 in the model output value.
b. Correcting the position of the recess.
When there is a portion where the start points of the front and rear recesses overlap each other, the two recesses are combined into one recess, wherein the start point of the new recess is the position of the start point of the subsequent recess, and the end point of the new recess is the end point of the previous recess.
The present example locates the starting points of the depressions by moving the bounding box, the lowest elevation points between the starting points being defined as the deepest points, and the dimensions of the depressions being calculated from these three points. Wherein the calculation formula is shown as (8).
(4) Naive Bayes-based concave classification
a. Preliminary classification of depressions
In this example, four classification features (left pitch, right pitch, width and depth) are selected based on the characteristics of the seam having a larger width, depth and pitch compared to the score groove, where the pitch refers to the distance between the end of the previous recess and the start of the next recess in order to ensure that the four features are independent of each other. The project then builds a database to obtain the mean and variance of the four classification features of the score and seam.
Step 1: the probability that the single feature undercut belongs to the notch or seam is calculated respectively, and the calculation formula is shown as (8) and (9).
Step 2: the probability that the four classification feature combined undercut belongs to a notch or seam is calculated, and a calculation formula is shown as (10) and (11).
Step 3: and comparing the probability that the pit belongs to the joint and the probability that the pit belongs to the notch, and obtaining a preliminary classification result.
b. Correction of classification results
Since the depressions adjacent to the actual seam are also widely spaced, such depressions are easily misidentified as seams in the preliminary classification, for which this example creates an improved strategy. When the seam continuously appears in the preliminary classification result, the probability that the two pits belong to the seam is compared, the probability is high, the seam is taken as a real seam, and otherwise, the seam is notched.
Preferably, the embodiment combines laser profiling equipment to provide a method capable of automatically, efficiently and accurately identifying and measuring the notch. The method can automatically identify the depressions, determine the starting points of the depressions, calculate the dimensions of the depressions and classify the depressions. The method can be applied in the performance assessment of airport runways based on fully automated data processing procedures.
Meanwhile, the present example fully considers the facts that shallow depressions, depressions with serious wear, and irregular depressions are difficult to detect. The embodiment has a database for targeted model training establishment, and the diversity of the database improves the generalization capability of the model. The average recall rate of the pits reached 97.51%. In the embodiment, a probability comparison method is formulated to improve classification accuracy aiming at the situation that the pits close to the real joints are easy to classify errors during pit classification.
The anti-skid performance of the airport runway is required to be evaluated regularly, and the method provides an efficient and accurate method for maintenance departments, so that the cost of manpower and material resources is greatly saved.
In particular, this example builds a model called a groovent, which consists of five layers of neural network, including two LSTM layers, two dropout layers, and a fully connected layer. After training, the model can conduct two-class classification on the given data and judge whether the given data contains the pits, so that the purpose of identifying the pits is achieved. The training database comprises pits with different shapes, so that the generalization capability of the trained model is strong.
The present example uses the GrooveNet as a bounding box, traverses the entire piece of data in 1 step, and outputs a series of 0 and 1 sequences. The position of the intersection of 0 and 1 can be used to determine the starting point of the depression. The item also corrects the positioning result, merging the recesses that overlap in position with each other.
The embodiment adopts a naive Bayes classifier to classify the concave, and selects four classification features (left interval, right interval, depth and width) according to the difference between the joint and the notch. And respectively calculating the probability that each pit belongs to the notch and the probability that each pit belongs to the joint, and then comparing the two probabilities to obtain a preliminary classification result. In this example, a probability comparison method is formulated to correct the preliminary classification result, taking into account the fact that the recesses adjacent to the true seams are prone to classification errors.
The foregoing is only illustrative of the present invention and all changes and modifications that come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein.
Claims (2)
1. The automatic identifying and measuring method for the airfield runway grooving based on the long-short-term memory network is characterized by comprising the following steps of: the method comprises the following steps:
step S1: providing a laser displacement sensor, and acquiring elevation data of the pit on the airfield runway by using the laser displacement sensor;
step S2: constructing a GrooveNet model and training the GrooveNet model;
step S3: taking the groovelet model trained in the step S2 as a mobile bounding box, and traversing the whole data with a step length of 1; constructing a series of 0 and 1 output values after traversing, wherein the positions of the 0 and 1 junctions in the series of the output values determine the starting point of the concave;
step S4: correcting the position of the recess in step S3;
step S5: classifying the depressions;
step S6: correcting the classified result in the step S5;
the step S2 specifically includes the following steps:
step S21: establishing a database: carrying out probability statistical analysis on the notch size data, and constructing a concave database with a sample width of 50 pixel values; dividing the database into positive and negative samples, wherein the sample containing one complete depression is called positive sample, and the negative sample is called negative sample;
step S22: selecting long-time memory of cell unit size in artificial neural network layer: the size of the cell unit is matched with the size of the training and testing sample, and the size of the cell unit in the artificial neural network layer is 50 for long and short time;
step S23: selecting the network layer number of the GrooveNet model: the GrooveNet is a classification model, and the last layer is set as a full-connection layer, so that the output value of the model is a single value, namely 0 or 1; firstly, matching a layer of long-short-time memory artificial neural network layer with the fully-connected layer, and simultaneously adding one more layer of long-short-time memory artificial neural network layer; finally, determining the model primary architecture as two layers of long-short-time memory artificial neural network layers and a full-connection layer;
step S24: the phenomenon of overfitting of the model is avoided: adding a Dropout layer at the tail end of each layer of long-short-term memory artificial neural network layer; at the moment, the model architecture of the GrooveNet is finally determined to be a long-short-time memory artificial neural network layer, a Dropout layer, a long-short-time memory artificial neural network layer, a Dropout layer and a full-connection layer;
step S25: training the groovent model: firstly, a predicted value is obtained through forward propagation of the GrooveNet model, a loss value is obtained through comparison of the predicted value and a true value, and then the weight of the GrooveNet model is adjusted through backward propagation so as to reduce the loss value; combining forward propagation and backward propagation, and continuously adjusting the weight of the GrooveNet model to enable the output value of the GrooveNet model to be similar to the predicted value; finally, saving the weight value of the groovelet model;
the specific contents of the position of the recess in the correcting step S3 in the step S4 are: when the initial point positions of the front and rear depressions are overlapped with each other, the two depressions are combined into one depression, wherein the initial point position of the new depression is the position of the initial point of the next depression, and the final point position of the new depression is the final point of the previous depression;
the step S5 specifically comprises the following steps:
step S51, sorting classification features: taking left spacing, right spacing, depth and width as classification features;
step S52: respectively calculating the probability that the concave part belongs to a notch or a seam under a single characteristic, wherein a calculation formula is shown as (2) and (3);
step 53: calculating four classification features, namely left spacing, right spacing, depth and width, combining the probability that the undercut belongs to a notch or a seam, wherein a calculation formula is shown as (4) (5);
step 54: comparing the probability that the pit belongs to the joint and the probability that the pit belongs to the notch, and obtaining a preliminary classification result;
the specific content of the step S6 is as follows: when the seam continuously appears in the preliminary classification result, the probability that the two pits belong to the seam is compared, the probability is high, the seam is taken as a real seam, and otherwise, the seam is notched.
2. The automatic identification and measurement method for airfield runway score based on long and short time memory network of claim 1, wherein the method comprises the following steps: the formula for determining the starting point of the recess in step S3 is shown in (1):
SP represents the start position of the depression, EP represents the end position of the depression, x j Representing the boundary position of 1 and 0 in the model output value, x i Indicating the boundary position between 0 and 1 in the model output value.
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 CN109919298A (en) | 2019-06-21 |
| CN109919298B true 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) |
Families Citing this family (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 |
|---|
| 基于ABAQUS的刻槽尺寸与路面磨损的相关性研究与应用;叶珍;《中国优秀硕士学位论文全文数据库》;20190315;全文 * |
| 基于BP神经网络的沥青路面裂缝识别方法研究;英红等;《河南理工大学学报(自然科学版)》;20180611(第04期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109919298A (en) | 2019-06-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109658387B (en) | Method for detecting defects of pantograph carbon slide plate of electric train | |
| CN111444767B (en) | Pedestrian detection and tracking method based on laser radar | |
| CN109919298B (en) | Airport runway grooving automatic identification and measurement method based on long-short-time memory network | |
| CN113158560B (en) | Intelligent driving vehicle autonomous capability test method based on scene opposition | |
| CN101976504B (en) | Multi-vehicle video tracking method based on color space information | |
| Kortmann et al. | Detecting various road damage types in global countries utilizing faster r-cnn | |
| CN109948683B (en) | Method and device for dividing difficulty degree of point cloud data and related equipment thereof | |
| KR102185225B1 (en) | Method for detecting sinkhole using deep learning and data association and sinkhole detecting system using it | |
| CN108052886A (en) | A kind of puccinia striiformis uredospore programming count method of counting | |
| Guo et al. | Classification of airborne laser scanning data using JointBoost | |
| CN113609895A (en) | Road traffic information acquisition method based on improved Yolov3 | |
| CN116168356A (en) | Vehicle damage judging method based on computer vision | |
| CN114609609A (en) | Speed estimation method for extracting static point cloud by FMCW laser radar random sampling | |
| CN117809458A (en) | Real-time assessment method and system for traffic accident risk | |
| CN113989179A (en) | Train wheel set tread defect detection method and system based on target detection algorithm | |
| CN115239733B (en) | Crack detection method and apparatus, terminal device and storage medium | |
| CN116452826A (en) | Coal gangue contour estimation method based on machine vision under shielding condition | |
| CN115993365A (en) | Belt defect detection method and system based on deep learning | |
| CN112598621B (en) | Intelligent determination method for surface quality of cold-rolled strip steel | |
| CN115564792A (en) | Welding path extraction method based on octree region growth | |
| CN108037488A (en) | Method of Sample Selection when a kind of non-homogeneous environment is empty | |
| Yao et al. | Automated detection and identification of area-based distress in concrete pavements | |
| Zhan et al. | Crack detection method of shadow asphalt pavement based on cuckoo search | |
| Yuan et al. | Real-time long-range road estimation in unknown environments | |
| Marianingsih et al. | Machine vision for the various road surface type classification based on texture feature |
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 |