CN114202540B - Intelligent detection method for split pin defect of high-speed rail contact network - Google Patents
Intelligent detection method for split pin defect of high-speed rail contact network Download PDFInfo
- Publication number
- CN114202540B CN114202540B CN202210144493.0A CN202210144493A CN114202540B CN 114202540 B CN114202540 B CN 114202540B CN 202210144493 A CN202210144493 A CN 202210144493A CN 114202540 B CN114202540 B CN 114202540B
- Authority
- CN
- China
- Prior art keywords
- layer
- dense
- speed rail
- module
- contact network
- 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
- 230000007547 defect Effects 0.000 title claims abstract description 75
- 238000001514 detection method Methods 0.000 title claims abstract description 31
- 238000012549 training Methods 0.000 claims abstract description 53
- 238000012360 testing method Methods 0.000 claims abstract description 34
- 238000013135 deep learning Methods 0.000 claims abstract description 7
- 238000000034 method Methods 0.000 claims description 21
- 230000002159 abnormal effect Effects 0.000 claims description 12
- 239000011159 matrix material Substances 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 11
- 230000007704 transition Effects 0.000 claims description 11
- 238000001914 filtration Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 4
- 239000000126 substance Substances 0.000 claims description 4
- 210000000707 wrist Anatomy 0.000 claims description 4
- 230000008569 process Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000007429 general method Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 239000000725 suspension Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
Abstract
The invention discloses an intelligent detection method and system for a defect of a cotter pin of a railway contact network, wherein a data set consisting of a sample containing the defect of the cotter pin of the railway contact network and a sample with a normal cotter pin of the railway contact network is obtained at first, and the data set is divided into a training set and a testing set according to a proportion; then, building an identification network model by using a Pythrch deep learning frame; training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the split pin defect of the high-speed rail contact network; and finally, taking a real-time image of the high-speed rail contact network, capturing a subimage containing the oblique cantilever and a structure below the oblique cantilever, identifying through an identification network model, obtaining an identification result, and further capturing the image of the high-speed rail contact network containing the cotter pin defect.
Description
Technical Field
The invention belongs to the technical field of high-speed rails, relates to a technology for identifying a defect of a cotter pin of a high-speed rail contact network, and particularly relates to an intelligent detection method and system for the defect of the cotter pin of the high-speed rail contact network.
Background
The operation mileage of high-speed railways in China is increased year by year, the high-speed railways are supplied with power through a contact network in real time in the operation process, so the contact network can also receive impact and vibration brought by a high-speed rail train, the contact network is possibly damaged due to the impact vibration, the damage of the contact network is timely discovered, the high-speed railways can be guaranteed to continuously operate for a long time only by timely repairing and perfecting, the high-speed railways are extremely long, the damaged parts in the contact network are timely discovered to be difficult work, and therefore, the key problem is solved from various angles through various ways.
A high-speed rail power supply safety detection monitoring system (6C system) is constructed in China, and the detection of the states of parts of a contact network suspension system by using non-contact detection equipment is an important component.
The method has the advantages that the defects possibly occurring on different parts of the contact net are correspondingly different detection schemes, and different treatment schemes are provided for the defects such as missing or damage of split pins and the like easily occurring on the contact net, but the accuracy rate of the existing treatment scheme is low, and the treatment effect is still to be improved.
The inventor researches an intelligent detection method for the split pin defect of the high-speed rail contact network on the basis of a deep learning technology, and aims to design a method and a system capable of solving the problems.
Disclosure of Invention
In order to overcome the problems, the inventor of the invention carries out intensive research and provides an intelligent detection method and system for the high-speed rail contact network cotter pin defect, wherein a data set consisting of a sample containing the high-speed rail contact network cotter pin defect and a sample with a normal high-speed rail contact network cotter pin is obtained at first, and the data set is divided into a training set and a testing set according to a proportion; then, building an identification network model by using a Pythrch deep learning frame; training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the split pin defect of the high-speed rail contact network; and finally, taking a real-time image of the high-speed rail overhead line system, intercepting a subimage containing the inclined cantilever and a structure below the inclined cantilever, identifying through an identification network model, obtaining an identification result, and further capturing the image of the high-speed rail overhead line system containing the split pin defect, thereby completing the invention.
The invention aims to provide an intelligent detection method for the defect of a cotter pin of a high-speed rail contact network, which comprises the following steps:
step 1, acquiring a data set consisting of a sample containing a defect of a cotter pin of a high-speed rail contact network and a sample with a normal cotter pin of the high-speed rail contact network, and dividing the data set into a training set and a test set according to a proportion;
step 2, building an identification network model by using a Pythrch deep learning framework;
step 3, training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the cotter pin defect of the high-speed rail contact network;
and 4, taking a real-time high-speed rail contact network image, capturing sub-images containing the inclined cantilever and the structure below the inclined cantilever, identifying through an identification network model, obtaining an identification result, and further capturing the high-speed rail contact network image containing the cotter pin defect.
Wherein, step 1 comprises the following substeps:
the method comprises the following steps that 1-1, a plurality of high-definition images including a high-speed rail contact network are obtained, wherein the high-definition images are shot by cameras and come from at least two cameras with different shooting angles;
the substep 1-2, intercepting the acquired data image into a sub-image, wherein only the inclined wrist arm and the lower structure of the original image are reserved in the sub-image;
in the substep 1-3, dividing subimages into a sample containing the defect of the cotter pin of the high-speed rail contact network and a sample with the normal cotter pin of the high-speed rail contact network, and respectively marking the samples;
1-4, randomly dividing the marked sample into a training data set and a testing data set according to a proportion, wherein in the training data set, the proportion of a normal sample of a cotter pin of the high-speed rail contact network to a sample containing the defect of the cotter pin of the high-speed rail contact network is 50: 1; in the test data set, the proportion of the normal sample of the cotter pin of the high-speed rail contact network to the sample containing the defect of the cotter pin of the high-speed rail contact network is 100: 1.
The defects of the split pin of the high-speed rail contact network comprise the defect of the split pin, abnormal split pin breaking angle and abnormal split pin appearance state.
In step 2, the identification network model sequentially comprises a convolution layer, a down-sampling layer, an alternate distribution layer and a classification layer;
wherein the convolution kernel size of the convolutional layer is 5x5, and the step distance is 2;
the downsampling layer is a maximum pooling downsampling layer with the step distance of 2;
the alternating arrangement of layers means that 9 dense modules and 8 transition layers are arranged alternately.
Wherein the dense module is formed by stacking a plurality of dense layers,
preferably, the 9 dense modules comprise a first dense module, a second dense module, a third dense module, a fifth dense module, a sixth dense module, a seventh dense module, an eighth dense module and a ninth dense module;
the first dense module comprises a 3-layer dense layer;
the second dense module comprises 4 dense layers;
the dense module III comprises 5 dense layers;
the dense module IV comprises 7 dense layers;
the dense module V comprises 7 dense layers;
the dense module six comprises 8 dense layers;
the dense module seventh comprises 6 dense layers;
the dense module eight comprises 4 dense layers;
the dense module nine comprises a 3-layer dense layer;
preferably, the dense layer is formed by sequentially stacking a BN layer, a ReLU layer, a point-by-point convolution layer, a BN layer, a Sigmod layer, a first channel-by-channel convolution layer, a BN layer, a second channel-by-channel convolution layer and a point-by-point convolution layer;
wherein, the point-by-point Convolution layer is Pointwise Convolution, the Convolution kernel size is 2x2, and the step pitch is 1;
the convolution kernel size of the first channel-by-channel convolution layer is 2x2, and the step pitch is 1;
the convolution kernel size of the second channel-by-channel convolution layer is 3x3, and the step pitch is 1.
Wherein, in the dense module, the
The output feature matrix formula for the layer is as follows:
wherein the content of the first and second substances,
Wherein the transition layer comprises a BN layer, a ReLU layer, a convolution layer with convolution kernel size of 1x1 and step size of 1, and an average pooled downsampling layer with filter size of 3x3 and step size of 2; adjusting the width, height and depth of an output feature matrix by interposing the transition layer between a dense module and a dense module;
preferably, wherein the depth of the feature matrix is adjusted by setting the convolution kernel to 1 × 1 with a step size of 1;
the width and height of the output feature matrix are adjusted by setting the average pooled downsampled layer with the filter size of 3x3 and step size of 2.
Wherein, in the step S2, the classification layer includes a BN layer, an average pooled downsampled layer with a filter size of 6x6 and a step size of 1, and a full connection layer.
In step 3, the model training system comprises a data preprocessing module, a loss function module, a training module and a training log storage module;
in the preprocessing module, removing noise of images in a training set and noise of images in a testing set by adopting a median filtering method; scaling the image to 256 pixels wide and 256 pixels high; finally, converting the read image data into a tensor format in a Pythrch;
and evaluating the consistency between the predicted output of the recognition network model and the real label of the input image through the loss function module, and considering that the recognition network model obtains the capacity of recognizing the cotter pin defect of the high-speed rail overhead line system when the consistency reaches more than 99 percent.
The invention also provides an intelligent detection system for the high-speed rail contact network cotter pin defect, which is used for implementing the intelligent detection method for the high-speed rail contact network cotter pin defect.
The intelligent detection method and system for the high-speed rail contact net cotter pin defect provided by the invention have the following beneficial effects:
(1) according to the intelligent detection method and system for the cotter pin defect of the high-speed rail contact network, the improved dense module is adopted to multiplex the characteristic data, so that the identification accuracy is improved, and the method and the system can effectively identify the cotter pin defect.
(2) According to the intelligent detection method and system for the cotter pin defect of the high-speed rail contact network, provided by the invention, the cotter pin defect of the high-speed rail contact network can be detected and captured in real time, so that real-time detection and discovery are realized, the defect position can be repaired in time, and larger loss is avoided.
Drawings
Fig. 1 shows an overall logic diagram of an intelligent detection method for a split pin defect of a catenary of a high-speed rail in a preferred embodiment of the invention;
FIG. 2 shows a schematic view of a high-speed rail catenary cotter pin in the absence;
fig. 3 shows a schematic diagram of a split pin of a high-speed rail contact net when the split pin is not broken.
Detailed Description
The invention is explained in more detail below with reference to the figures and examples. The features and advantages of the present invention will become more apparent from the description.
The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.
According to the intelligent detection method for the defect of the split pin of the high-speed rail contact network, as shown in fig. 1, the method comprises the following steps:
step 1, acquiring a data set consisting of a sample containing a defect of a cotter pin of a high-speed rail contact network and a sample with a normal cotter pin of the high-speed rail contact network, and dividing the data set into a training set and a test set according to a proportion;
preferably, step 1 comprises the sub-steps of:
the method comprises the following steps that 1-1, a plurality of high-definition images including a high-speed rail contact network are obtained, wherein the high-definition images are shot by cameras and come from at least two cameras with different shooting angles; a general method is to arrange at least two cameras on a mobile device, which can run along a high-speed rail line, and can be a common high-speed rail train or a special device, so as to be able to photograph a catenary in real time, preferably, the two cameras are oriented to different directions, so that they can photograph the catenary from two angles, preferably two opposite directions; in this substep 1-1, at least ten thousand images need to be obtained;
when the mobile equipment moves along the high-speed rail, the speed is kept consistent, the focal length of the camera is adjusted as soon as possible and kept unchanged as much as possible, so that the definition and the size of the obtained image are consistent, and the subsequent processing is facilitated; in addition, in order to make the data sufficient, it is necessary to take a picture of samples on at least two high-speed railway lines, respectively.
And a substep 1-2, intercepting the acquired data image into a sub-image, wherein only the inclined wrist arm and the lower structure of the original image are reserved in the sub-image. The applicant finds that at least 3 bolt connecting structures are arranged below each inclined cantilever on a high-speed rail contact network, and are provided with cotter pins, the identification technology of the inclined cantilevers is simple and mature enough, the inclined cantilevers can be identified quickly and accurately, on the basis, the inclined cantilevers in images of each contact network can be known quickly and accurately, so that the images containing the inclined cantilevers can be intercepted, square sub-images containing complete inclined cantilevers can be intercepted, the images can contain the inclined cantilevers and the structures below the inclined cantilevers and also contain the cotter pins required to be detected, the workload of subsequent image processing can be greatly reduced through the sub-step, and a database for cotter pin defect detection can be formed quickly.
In the substep 1-3, dividing subimages into a sample containing the defect of the cotter pin of the high-speed rail contact network and a sample with the normal cotter pin of the high-speed rail contact network, and respectively marking the samples;
the defects of the split pin of the high-speed rail contact network comprise that the split pin is missing as shown in fig. 2, the opening angle of the split pin is abnormal, and the appearance state of the split pin is abnormal; the normal condition is that the opening angle reaches about 120 degrees, the abnormal condition is that the opening angle is too large or too small, the contact pin is not opened, and the abnormal condition is also that the opening angle is abnormal, as shown in fig. 3; the abnormal appearance state means that the two wires of the cotter are different in length, and generally, the two wires are partially damaged or cut.
In the marking process of the sub-step, the positions of the cotters in the image are observed one by one manually, whether the defects such as cotter missing, abnormal opening angle or abnormal cotter appearance state exist or not is judged, and the image is marked.
1-4, randomly dividing the marked sample into a training data set and a testing data set according to a proportion, wherein in the training data set, the proportion of a normal sample of a cotter pin of the high-speed rail contact network to a sample containing the defect of the cotter pin of the high-speed rail contact network is 50: 1; in the test data set, the proportion of the normal sample of the cotter pin of the high-speed rail contact network to the sample containing the defect of the cotter pin of the high-speed rail contact network is 100: 1.
Step 2, building an identification network model by using a Pythrch deep learning framework;
preferably, in step 2, the identification network model sequentially comprises a convolutional layer, a downsampling layer, an alternating distribution layer and a classification layer;
wherein the convolution kernel size of the convolutional layer is 5x5, and the step distance is 2;
the downsampling layer is a maximum pooling downsampling layer with the step distance of 2;
the alternating arrangement of layers means that 9 dense modules and 8 transition layers are arranged alternately.
In the dense module, the dense layer is connected with all the previous layers as input, so that characteristic reuse can be realized, and the efficiency is improved.
In a preferred embodiment, the dense module is formed from a plurality of dense layers stacked,
specifically, the 9 dense modules comprise a dense module I, a dense module II, a dense module III, a dense module IV, a dense module V, a dense module VI, a dense module VII, a dense module VIII and a dense module IX;
the first dense module comprises a 3-layer dense layer;
the second dense module comprises 4 dense layers;
the dense module III comprises 5 dense layers;
the dense module IV comprises 7 dense layers;
the dense module V comprises 7 dense layers;
the dense module six comprises 8 dense layers;
the dense module seventh comprises 6 dense layers;
the dense module eight comprises 4 dense layers;
the dense module nine comprises a 3-layer dense layer. Preferably, the dense layer numbers can be freely combined, and the combination is selected by comparing different number combinations, so that the identification precision and the operation efficiency have better comprehensive performance;
preferably, the dense layer is formed by sequentially stacking a BN layer, a ReLU layer, a point-by-point convolution layer, a BN layer, a Sigmod layer, a first channel-by-channel convolution layer, a BN layer, a second channel-by-channel convolution layer and a point-by-point convolution layer;
wherein, the point-by-point Convolution layer is Pointwise Convolution, the Convolution kernel size is 2x2, and the step pitch is 1;
the convolution kernel size of the first channel-by-channel convolution layer is 2x2, and the step pitch is 1;
the convolution kernel size of the second channel-by-channel convolution layer is 3x3, and the step pitch is 1.
The point-by-point convolution layer and channel-by-channel convolution layer are combined, so that the amount of operation and the number of parameters can be greatly reduced, the algorithm efficiency is improved, and a normal distribution with the mean value of 0 and the variance of 1 is input by using a BN layer, so that the training can be accelerated; the Sigmod layer is a Sigmod function layer, and two classifications are performed by the Sigmod layer.
More preferably, in the channel-by-channel convolution layer, one convolution kernel is responsible for one channel, one channel is convolved by only one convolution kernel, and the number of feature maps after convolution is the same as the number of channels of the input layer, so that the operation amount and the parameter number can be reduced.
The BN is Batch Normalization, the characteristic diagram meets the distribution rule that the mean value is 0 and the variance is 1 through the BN, and in the scheme of the application, the convergence of the network can be accelerated and the accuracy rate can be improved through setting the BN layer.
The ReLU is proposed by Xavier Glorot et al in 2009 paper Deep spark Rectifier Neural Networks, and a trained Neural network has certain sparsity and higher training speed, so that the training period can be shortened;
the ReLU activation function is:
in a preferred embodiment, in the dense module, the first
The output feature matrix formula for the layer is as follows:
wherein the content of the first and second substances,
is shown as
Outputting the layer;
In a preferred embodiment, the transition layers include a BN layer, a ReLU layer, a convolutional layer with a convolutional kernel size of 1x1 and a stride of 1, and an average pooled downsampled layer with a filter size of 3x3 and a stride of 2; adjusting the width, height and depth of an output feature matrix by disposing the transition layer between a dense module and a dense module;
preferably, the depth of the feature matrix is adjusted by setting the convolution kernel to be 1 × 1 and the step pitch to be 1, so as to prevent the number of output channels from being too large;
by setting the average pooling downsampling layer with the filter size of 3x3 and the step size of 2, the width and the height of an output feature matrix are adjusted, features are compressed, the network complexity is simplified, the calculation amount is reduced, and the memory consumption is reduced.
In a preferred embodiment, in step S2, the classification layer includes sequentially connected BN layers, an average pooled downsampling layer with a filter size of 6x6 and a step size of 1, and a full connection layer, and the distributed feature representation extracted by the pre-network is mapped to a sample label space;
the fully-connected layer refers to a layer in which each node in the layer will connect all nodes of its next layer.
Step 3, training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the cotter pin defect of the high-speed rail contact network;
preferably, in step 3, the model training system includes a data preprocessing module, a loss function module, a training module, and a training log storage module;
in the preprocessing module, removing noise of images in a training set and noise of images in a testing set by adopting a median filtering method; scaling the image to 256 pixels wide and 256 pixels high; finally, converting the read image data into a tensor format in a Pythrch; the window size in the median filtering method is 3x 3;
evaluating the consistency between the predicted output of the recognition network model and the real label of the input image through the loss function module, and considering the recognition network model to obtain the capacity of recognizing the cotter pin defect of the high-speed rail contact network when the consistency reaches more than 99%;
the loss function is:
wherein the content of the first and second substances,
the predicted value is represented by a value of the prediction,
representing the actual output of the neuron;
training in the training module in an end-to-end training mode, wherein a training optimizer is an RMSProp optimizer; the end-to-end training is a training mode of inputting a sub-image output recognition result; the learning rate of the RMSProp optimizer is set to 10-5；
The training log storage module is used for recording information such as loss values, accuracy rates and weight file storage paths in the training process;
in a specific training and testing process, a tensor format is input into an identification network model, an output value can be obtained in a forward propagation process, an error value can be obtained by comparing the output value with an expected output value, an error gradient of each node can be obtained by calculating a partial derivative of each node, and an obtained loss value is reversely applied to the loss gradient, so that the error is reversely propagated;
the identification precision detection process for detecting the trained identification network model to act on the test set through the model test system comprises the steps of inputting a sample of the test data set into the identification network model, obtaining a predicted result, comparing the predicted result with a true value of the sample to obtain identification precision, and storing the training weight when the precision is greater than the weight of the network model trained at the last time; and when the test recognition accuracy is not improved any more than 30 times, finishing the training process.
And 4, taking a real-time high-speed rail contact network image, capturing sub-images containing the inclined cantilever and the structure below the inclined cantilever, identifying through an identification network model, obtaining an identification result, and further capturing the high-speed rail contact network image containing the cotter pin defect.
Wherein the sub-image is a square image comprising a complete oblique carpal arm structure.
Preferably, after the high-speed rail overhead line system image is obtained, the high-speed rail overhead line system image is firstly input into the oblique cantilever identification module, the oblique cantilever in the overhead line system image is identified and locked, so that the length of the oblique cantilever is obtained, the oblique cantilever is placed in the middle of the square selection frame, and finally the square selection frame is intercepted, so that a sub-image containing the oblique cantilever and a structure below the oblique cantilever is obtained.
The invention also provides an intelligent detection system for the defect of the cotter pin of the high-speed rail contact network, which is used for implementing the intelligent detection method for the defect of the cotter pin of the high-speed rail contact network.
The recognition network model can recognize and mark defects of cotter pins in the subimages. The identification network model comprises a convolution layer, a down-sampling layer, an alternate distribution layer and a classification layer in sequence;
wherein the convolution kernel size of the convolutional layer is 5x5, and the step distance is 2;
the downsampling layer is a maximum pooled downsampling layer with the step pitch of 2;
the alternating arrangement layer means that 9 dense modules and 8 transition layers are alternately arranged;
the dense module is formed from a plurality of dense layers stacked,
preferably, the 9 dense modules comprise first dense module, second dense module, third dense module, fifth dense module, sixth dense module, seventh dense module, eighth dense module and ninth dense module;
the first dense module comprises a 3-layer dense layer;
the second dense module comprises a 4-layer dense layer;
the dense module III comprises 5 dense layers;
the dense module IV comprises 7 dense layers;
the dense module V comprises 7 dense layers;
the dense module six comprises 8 dense layers;
the dense module seventh comprises 6 dense layers;
the dense module eight comprises 4 dense layers;
the dense module nine comprises a 3-layer dense layer;
preferably, the dense layer is formed by sequentially stacking a BN layer, a ReLU layer, a point-by-point convolution layer, a BN layer, a Sigmod layer, a first channel-by-channel convolution layer, a BN layer, a second channel-by-channel convolution layer and a point-by-point convolution layer;
wherein, the point-by-point Convolution layer is Pointwise Convolution, the Convolution kernel size is 2x2, and the step pitch is 1;
the convolution kernel size of the first channel-by-channel convolution layer is 2x2, and the step pitch is 1;
the second channel-by-channel convolutional layer has a convolutional kernel size of 3x3 and a step size of 1.
Examples
Step 1, taking 1.2 million pictures of contact networks along the Jinghush high-speed rail, and selecting 110 pictures containing the split pin defect of the high-speed rail contact network from the pictures; intercepting the acquired data image into a square sub-image, wherein only the inclined wrist arm and the lower structure of the original image are reserved in the sub-image; dividing the subimages into a sample containing the defect of the cotter pin of the high-speed rail contact network and a sample with the normal cotter pin of the high-speed rail contact network, and setting a training data set and a test data set according to the samples; in the training data set, the proportion of a normal sample of the cotter pin of the high-speed rail contact network to a sample containing the defect of the cotter pin of the high-speed rail contact network is 50: 1; in the test data set, the proportion of the normal sample of the cotter pin of the high-speed rail contact network to the sample containing the defect of the cotter pin of the high-speed rail contact network is 100: 1.
Step 2, building an identification network model by using a Pythrch deep learning framework;
the identification network model sequentially comprises a convolution layer, a down-sampling layer, an alternate distribution layer and a classification layer;
wherein the convolution kernel size of the convolutional layer is 5x5, and the step distance is 2;
the downsampling layer is a maximum pooling downsampling layer with the step distance of 2;
the alternating arrangement layer means that 9 dense modules and 8 transition layers are alternately arranged;
the 9 dense modules comprise a first dense module, a second dense module, a third dense module, a fifth dense module, a sixth dense module, a seventh dense module, an eighth dense module and a ninth dense module;
the first dense module comprises a 3-layer dense layer;
the second dense module comprises 4 dense layers;
the dense module III comprises 5 dense layers;
the dense module IV comprises 7 dense layers;
the dense module V comprises 7 dense layers;
the dense module six comprises 8 dense layers;
the dense module seventh comprises 6 dense layers;
the dense module eight comprises 4 dense layers;
the dense module nine comprises a 3-layer dense layer;
the dense layer is formed by sequentially stacking a BN layer, a ReLU layer, a point-by-point convolution layer, a BN layer, a Sigmod layer, a first channel-by-channel convolution layer, a BN layer, a second channel-by-channel convolution layer and a point-by-point convolution layer;
the convolution kernel size of the point-by-point convolution layer is 2x2, and the step pitch is 1;
the convolution kernel size of the first channel-by-channel convolution layer is 2x2, and the step pitch is 1;
the convolution kernel size of the second channel-by-channel convolution layer is 3x3, and the step pitch is 1.
Step 3, training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the cotter pin defect of the high-speed rail contact network;
removing noise of images in a training set and noise of images in a testing set by adopting a median filtering method; scaling the image to 256 pixels wide and 256 pixels high; finally, converting the read image data into a tensor format in a Pythrch; the window size in the median filtering method is 3x 3;
evaluating the consistency between the predicted output of the recognition network model and the real label of the input image through the loss function module, and considering the recognition network model to obtain the capacity of recognizing the cotter pin defect of the high-speed rail contact network when the consistency reaches more than 99%;
the loss function is:
the training optimizer is a RMSProp optimizer.
And 4, taking 50 real-time high-speed rail contact network images, intercepting subimages containing the inclined cantilever and the structure below the inclined cantilever, identifying through an identification network model, obtaining an identification result, and further capturing 3 high-speed rail contact network images with split pin defects contained in the subimages.
Step 5, enabling experts to watch the 50 high-speed rail overhead contact system images called in the step 4 one by one, finding that the number of the high-speed rail overhead contact system images with the cotter pin defects is 3, and enabling the number of the high-speed rail overhead contact system images to be consistent with the result given in the step 4;
and 6, repeating the step 4 and the step 5 for 10 times to obtain that the accuracy rate of identifying the cotter pin defect of the high-speed rail contact network in the step 4 is more than 99.9%.
The present invention has been described above in connection with preferred embodiments, but these embodiments are merely exemplary and merely illustrative. On the basis of the above, the invention can be subjected to various substitutions and modifications, and the substitutions and the modifications are all within the protection scope of the invention.
Claims (8)
1. The intelligent detection method for the defect of the split pin of the high-speed rail contact network is characterized by comprising the following steps of:
the method comprises the following steps of 1, acquiring a data set consisting of a sample containing the defect of the cotter pin of the high-speed rail contact network and a sample with the normal cotter pin of the high-speed rail contact network, and dividing the data set into a training set and a testing set according to a proportion;
step 2, building an identification network model by using a Pythrch deep learning framework;
step 3, training the recognition network model through a model training system, and testing the trained recognition network model through a model testing system, so that the recognition network model can recognize the cotter pin defect of the high-speed rail contact network;
step 4, taking a real-time high-speed rail contact network image, intercepting subimages containing the inclined cantilever and the structure below the inclined cantilever, identifying through an identification network model, obtaining an identification result, and further capturing the high-speed rail contact network image containing the cotter pin defect;
in step 2, the identification network model sequentially comprises a convolution layer, a down-sampling layer, an alternate distribution layer and a classification layer;
wherein the convolution kernel size of the convolutional layer is 5x5, and the step distance is 2;
the downsampling layer is a maximum pooling downsampling layer with the step distance of 2;
the alternating arrangement layer means that 9 dense modules and 8 transition layers are alternately arranged;
the transition layer comprises a BN layer, a ReLU layer, a convolution layer with convolution kernel size of 1x1 and step size of 1, and an average pooling downsampling layer with a filter size of 3x3 and step size of 2; adjusting the width, height and depth of an output feature matrix by interposing the transition layer between a dense module and a dense module;
wherein the depth of the feature matrix is adjusted by setting the convolution kernel to be 1x1 and the convolution layer with the step pitch of 1;
the width and height of the output feature matrix are adjusted by setting the average pooled downsampled layer with the filter size of 3x3 and step size of 2.
2. The intelligent detection method for the defect of the cotter pin of the high-speed rail contact network according to claim 1, characterized in that,
step 1 comprises the following substeps:
the method comprises the following steps that 1-1, a plurality of high-definition images including a high-speed rail contact network are obtained, wherein the high-definition images are shot by cameras and come from at least two cameras with different shooting angles;
the substep 1-2, intercepting the acquired data image into a sub-image, wherein only the inclined wrist arm and the lower structure of the original image are reserved in the sub-image;
in the substep 1-3, dividing the subimages into a sample containing the defect of the cotter pin of the high-speed rail contact network and a sample with the normal cotter pin of the high-speed rail contact network, and respectively marking;
1-4, randomly dividing the marked sample into a training data set and a testing data set according to a proportion, wherein in the training data set, the proportion of a normal sample of a cotter pin of the high-speed rail contact network to a sample containing the defect of the cotter pin of the high-speed rail contact network is 50: 1; in the test data set, the proportion of the normal sample of the cotter pin of the high-speed rail contact network to the sample containing the defect of the cotter pin of the high-speed rail contact network is 100: 1.
3. The intelligent detection method for the cotter pin defect of the high-speed rail contact network according to claim 1, characterized in that in substeps 1-3,
the defects of the split pin of the high-speed rail contact network comprise the defect of the split pin, abnormal split angle of the split pin and abnormal appearance state of the split pin.
4. The intelligent detection method for the defect of the cotter pin of the high-speed rail contact network according to claim 1, characterized in that,
the dense module is stacked from a plurality of dense layers,
the 9 dense modules comprise a first dense module, a second dense module, a third dense module, a fifth dense module, a sixth dense module, a seventh dense module, an eighth dense module and a ninth dense module;
the first dense module comprises a 3-layer dense layer;
the second dense module comprises 4 dense layers;
the dense module III comprises 5 dense layers;
the dense module IV comprises 7 dense layers;
the dense module V comprises 7 dense layers;
the dense module six comprises 8 dense layers;
the dense module seventh comprises 6 dense layers;
the dense module eight comprises 4 dense layers;
the dense module nine comprises a 3-layer dense layer;
the dense layer is formed by sequentially stacking a BN layer, a ReLU layer, a point-by-point convolution layer, a BN layer, a Sigmod layer, a first channel-by-channel convolution layer, a BN layer, a second channel-by-channel convolution layer and a point-by-point convolution layer;
wherein, the point-by-point Convolution layer is Pointwise Convolution, the Convolution kernel size is 2x2, and the step pitch is 1;
the convolution kernel size of the first channel-by-channel convolution layer is 2x2, and the step pitch is 1;
the convolution kernel size of the second channel-by-channel convolution layer is 3x3, and the step pitch is 1.
5. The intelligent detection method for the defect of the cotter pin of the high-speed rail contact network according to claim 4, characterized in that,
in the dense module, the first
The output feature matrix formula for the layer is as follows:
wherein the content of the first and second substances,
is shown as
Outputting the layer;
6. The intelligent detection method for the defect of the cotter pin of the high-speed rail contact network according to claim 1, characterized in that,
in the step 2, the classification layer comprises a BN layer, an average pooled downsampling layer with a filter size of 6x6 and a step size of 1, and a full connection layer.
7. The intelligent detection method for the defect of the cotter pin of the high-speed rail contact network according to claim 1, characterized in that,
in step 3, the model training system comprises a data preprocessing module, a loss function module, a training module and a training log storage module;
in the preprocessing module, removing noise of images in a training set and noise of images in a testing set by adopting a median filtering method; scaling the image to 256 pixels wide and 256 pixels high; finally, converting the read image data into a tensor format in a Pyorch;
and evaluating the consistency between the predicted output of the recognition network model and the real label of the input image through the loss function module, and considering that the recognition network model obtains the capacity of recognizing the cotter pin defect of the high-speed rail overhead line system when the consistency reaches more than 99 percent.
8. An intelligent detection system for the split pin defect of the high-speed rail contact network, which is used for implementing the intelligent detection method for the split pin defect of the high-speed rail contact network according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210144493.0A CN114202540B (en) | 2022-02-17 | 2022-02-17 | Intelligent detection method for split pin defect of high-speed rail contact network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210144493.0A CN114202540B (en) | 2022-02-17 | 2022-02-17 | Intelligent detection method for split pin defect of high-speed rail contact network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114202540A CN114202540A (en) | 2022-03-18 |
| CN114202540B true CN114202540B (en) | 2022-05-13 |
Family
ID=80645581
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210144493.0A Active CN114202540B (en) | 2022-02-17 | 2022-02-17 | Intelligent detection method for split pin defect of high-speed rail contact network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114202540B (en) |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115170883B (en) * | 2022-07-19 | 2023-03-14 | 哈尔滨市科佳通用机电股份有限公司 | Brake cylinder piston push rod opening pin loss fault detection method |
| CN115228690B (en) * | 2022-08-08 | 2023-08-08 | 深圳市深赛尔股份有限公司 | Paint surface repairing device and method based on environment-friendly water-based coil steel coating |
| CN115359307B (en) * | 2022-10-24 | 2023-01-03 | 成都诺比侃科技有限公司 | Big data-based overhead contact system wear-checking defect data management method and system |
| CN115830012B (en) * | 2023-02-08 | 2023-05-02 | 诺比侃人工智能科技(成都)股份有限公司 | Method for detecting and analyzing clue damage data of overhead contact system |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507972A (en) * | 2020-04-20 | 2020-08-07 | 南京航空航天大学 | Tunnel surface defect detection method combining convolutional neural network and support vector machine |
| CN111667455B (en) * | 2020-04-28 | 2023-10-24 | 广东三三智能科技有限公司 | AI detection method for brushing multiple defects |
| CN112258496A (en) * | 2020-11-02 | 2021-01-22 | 郑州大学 | Underground drainage pipeline disease segmentation method based on full convolution neural network |
-
2022
- 2022-02-17 CN CN202210144493.0A patent/CN114202540B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN114202540A (en) | 2022-03-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114202540B (en) | Intelligent detection method for split pin defect of high-speed rail contact network | |
| Faghih-Roohi et al. | Deep convolutional neural networks for detection of rail surface defects | |
| CN107316007B (en) | Monitoring image multi-class object detection and identification method based on deep learning | |
| CN110246132A (en) | Rail vehicle bolt looseness detection method and system | |
| CN103745224B (en) | Image-based railway contact net bird-nest abnormal condition detection method | |
| CN105260744B (en) | The automatic on-line diagnostic method and system of a kind of goods train coupler yoke key position failure | |
| CN108711148B (en) | Tire defect intelligent detection method based on deep learning | |
| CN104504377B (en) | A kind of passenger on public transport degree of crowding identifying system and method | |
| CN108932523B (en) | Image classification and classified data creation system and method, and storage medium | |
| CN108073898B (en) | Method, device and equipment for identifying human head area | |
| CN113947731B (en) | Foreign matter identification method and system based on contact net safety inspection | |
| CN111489339A (en) | Method for detecting defects of bolt spare nuts of high-speed railway positioner | |
| CN114399672A (en) | Railway wagon brake shoe fault detection method based on deep learning | |
| CN114743119B (en) | High-speed rail contact net hanger nut defect detection method based on unmanned aerial vehicle | |
| CN111311567A (en) | Method for identifying fastener and steel rail diseases of track line image | |
| CN109801265A (en) | A kind of real-time transmission facility foreign matter detection system based on convolutional neural networks | |
| CN112464846B (en) | Automatic recognition method for abnormal faults of freight train carriage of station | |
| CN104268588A (en) | Automatic detection method for brake shoe borer loss fault of railway wagon | |
| CN108508023A (en) | The defect detecting system of end puller bolt is contacted in a kind of railway contact line | |
| CN112801037A (en) | Face tampering detection method based on continuous inter-frame difference | |
| CN115861210A (en) | Transformer substation equipment abnormity detection method and system based on twin network | |
| CN116883801A (en) | YOLOv8 target detection method based on attention mechanism and multi-scale feature fusion | |
| CN112070712A (en) | Printing defect detection method based on self-encoder network | |
| CN114387261A (en) | Automatic detection method suitable for railway steel bridge bolt diseases | |
| CN109389170A (en) | A kind of gradation type operating condition method for early warning based on 3D convolutional neural networks |
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 |