CN112308135A - Railway motor car sand spreading pipe loosening fault detection method based on deep learning - Google Patents

Railway motor car sand spreading pipe loosening fault detection method based on deep learning Download PDF

Info

Publication number
CN112308135A
CN112308135A CN202011183656.3A CN202011183656A CN112308135A CN 112308135 A CN112308135 A CN 112308135A CN 202011183656 A CN202011183656 A CN 202011183656A CN 112308135 A CN112308135 A CN 112308135A
Authority
CN
China
Prior art keywords
network
target
residual error
deep learning
detection method
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202011183656.3A
Other languages
Chinese (zh)
Inventor
付德敏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Kejia General Mechanical and Electrical Co Ltd
Original Assignee
Harbin Kejia General Mechanical and Electrical Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Harbin Kejia General Mechanical and Electrical Co Ltd filed Critical Harbin Kejia General Mechanical and Electrical Co Ltd
Priority to CN202011183656.3A priority Critical patent/CN112308135A/en
Publication of CN112308135A publication Critical patent/CN112308135A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

A railway motor car sanding pipe loosening fault detection method based on deep learning belongs to the technical field of image detection. The method aims to solve the problems of low detection accuracy and low detection efficiency in a manual image checking mode and the problem that the detection accuracy of the existing deep learning network is to be improved. Firstly, acquiring an image of a sanding pipe to be detected, and detecting by using an CASCADERCNN network model; CASCADERCNN the network includes a deep residual network, an RPN network and a prediction network; and the sand scattering pipe image to be detected firstly passes through a depth residual error network, then passes through an RPN network and finally passes through a prediction network, and the fault category and the corresponding position are output. The method is mainly used for detecting the loosening fault of the sand spraying pipe of the railway motor car.

Description

Railway motor car sand spreading pipe loosening fault detection method based on deep learning
Technical Field
The invention belongs to the technical field of image detection, and particularly relates to a loose fault detection method for a sanding pipe.
Background
The loose fault of the sand spraying pipe of the bullet train is a fault which endangers the traffic safety of the railway bullet train, and in the fault detection of the loose fault of the sand spraying pipe, the fault detection is basically carried out by adopting a mode of manually checking images in the prior art. The fault detection method for manually checking the images has low efficiency, needs a large amount of manpower, and can cause missed detection and false detection due to the fact that human factors such as fatigue and omission easily occur in the working process of vehicle detection personnel, so that the driving safety is influenced.
Therefore, an automatic detection method is needed to replace a fault detection method for manually checking images, and the fault detection accuracy and stability can be improved by automatically identifying faults according to image information. In recent years, with the continuous development of deep learning and artificial intelligence, the technology is mature, and the deep learning network has achieved a few effects in image recognition. But the existing deep learning network is utilized to identify the loosening fault of the sanding pipe, so that the problems still exist. For example, although the fastercnnn detection network is widely used in target detection, the problems of inaccurate detection frames and multiple classes of false detection can occur in the detection process.
Disclosure of Invention
The invention aims to solve the problems of low detection accuracy and low detection efficiency in a manual image checking mode and the problem that the detection accuracy of the conventional deep learning network is to be improved.
The railway motor car sanding pipe loosening fault detection method based on deep learning comprises the following steps:
acquiring an image of the sanding pipe to be detected, detecting by using an CASCADERCNN network model, and outputting fault types and corresponding positions;
CASCADERCNN the loss function of the network during training is as follows:
Figure BDA0002750863690000011
wherein N isclsTo classify the total number of samples, piIs the classification probability of the different classes; n is a radical ofregNumber of regression frames, ti＝{tx,ty,tw,thIs a vector, tx,ty,tw,thRespectively representing the offset of the x coordinate, the y coordinate, the width and the height of the candidate frame; the parameter λ is used to weigh the ratio of classification loss to regression loss, when the sample is positive
Figure BDA0002750863690000012
Is 1, negative sample
Figure BDA0002750863690000013
Is 0;
Figure BDA0002750863690000014
is and tiVectors with the same dimension represent the offset of the candidate frame to the mark frame; l iscls(pi) A classification loss function that is predicted for the target,
Figure BDA0002750863690000015
a position loss function that is a regression prediction;
classification loss function for target prediction:
Figure BDA0002750863690000021
wherein gamma is a positive number, alphaiThe method is used for adjusting the proportion of positive and negative samples, h represents the probability value of the current prediction class, and t represents the class value of the real label.
Has the advantages that:
1. the invention can realize the loosening fault detection of the sand spraying pipe of the railway motor car, and the image automatic identification mode is used for replacing manual detection, thereby improving the detection efficiency and accuracy.
2. According to the method, a deep learning algorithm is applied to automatic identification of the sand scattering pipe loosening fault, and the robustness and the precision of the overall algorithm are improved.
3. The invention improves a residual error module and introduces the improved module into an CASCADERCNN model, multiplies multi-scale information by respective weight and then fuses the information together, learns the importance of different input characteristics by introducing the weight, increases the adaptation of the network to the scale and enables the output fusion characteristics to be closer to real characteristics, adopts CASCADERCNN network to train a high-quality detector under the condition of ensuring no reduction of the number of samples by using continuously improved threshold values, and achieves the purpose of predicting the result by cascading detection networks. The accuracy of the target frame is improved, and the accuracy of detection is improved.
4. For small targets in detection, the method improves focalloss loss functions and improves detection accuracy.
Drawings
FIG. 1 is a flow chart of fault identification;
FIG. 2 is a flow chart of weight coefficient determination during training;
FIG. 3 is a schematic view of the CASCADERCNN training process;
FIG. 4 is a diagram of a depth residual block;
FIG. 5 is a depth residual network diagram;
FIG. 6 is a flow chart of an embodiment.
Detailed Description
The first embodiment is as follows:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and as shown in fig. 1, the method comprises the following steps:
acquiring an image of the sanding pipe to be detected, detecting by using an CASCADERCNN network model, and outputting fault types and corresponding positions;
CASCADERCNN the loss function of the network during training is as follows:
Figure BDA0002750863690000031
wherein N isclsTo classify the total number of samples, piIs the classification probability of the different classes; n is a radical ofregNumber of regression frames, ti＝{tx,ty,tw,thIs a vector, tx,ty,tw,thRespectively representing the offset of the x coordinate, the y coordinate, the width and the height of the candidate frame; the parameter λ is used to weigh the ratio of classification loss to regression loss, when the sample is positive
Figure BDA0002750863690000032
Is 1, negative sample
Figure BDA0002750863690000033
Is 0;
Figure BDA0002750863690000034
is and tiVectors with the same dimension represent the offset of the candidate frame to the mark frame; l iscls(pi) A classification loss function that is predicted for the target,
Figure BDA0002750863690000035
a position loss function that is a regression prediction;
classification loss function for target prediction:
Figure BDA0002750863690000036
wherein gamma is a positive number, alphaiThe method is used for adjusting the proportion of positive and negative samples, h represents the probability value of the current prediction class, and t represents the class value of the real label.
The second embodiment is as follows:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and the method comprises the following steps:
acquiring an image of the sanding pipe to be detected, detecting by using an CASCADERCNN network model, and outputting fault types and corresponding positions;
CASCADERCNN the loss function of the network during training is as follows:
Figure BDA0002750863690000037
wherein N isclsTo classify the total number of samples, piIs the classification probability of the different classes; n is a radical ofregNumber of regression frames, ti＝{tx,ty,tw,thIs a vector, tx,ty,tw,thRespectively representing the offset of the x coordinate, the y coordinate, the width and the height of the candidate frame; the parameter λ is used to weigh the ratio of classification loss to regression loss, when the sample is positive
Figure BDA0002750863690000038
Is 1, negative sample
Figure BDA0002750863690000039
Is 0;
Figure BDA00027508636900000310
is and tiVectors with the same dimension represent the offset of the candidate frame to the mark frame; l iscls(pi) A classification loss function that is predicted for the target,
Figure BDA00027508636900000311
a position loss function that is a regression prediction;
classification loss function for target prediction:
Figure BDA00027508636900000312
wherein gamma is a positive number, alphaiAnd h represents a predicted value and t represents a real label value, wherein the proportional parameter is a proportion parameter of positive and negative samples.
The regression predicted position loss function is as follows:
Figure BDA0002750863690000041
wherein the content of the first and second substances,
Figure BDA0002750863690000042
the third concrete implementation mode:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and the method comprises the following steps:
acquiring an image of the sanding pipe to be detected, detecting by using an CASCADERCNN network model, and outputting fault types and corresponding positions;
CASCADERCNN the network comprises a deep residual error network, an RPN network and a prediction network, and the concrete structure is as follows:
the deep residual error network, the RPN and the prediction network are connected in sequence, the output of the deep residual error network is input into the RPN, and the RPN enters and exits the foreground candidate frame and is input into the prediction network.
Other steps and parameters are the same as in the first or second embodiment.
The fourth concrete implementation mode:
the implementation mode is a railway motor car sanding pipe loosening fault detection method based on deep learning, and the deep residual error network structure in the implementation mode is a conv layer, a first pool layer, a plurality of residual error modules, a second pool layer and an FC layer.
Other steps and model structures are the same as those in the third embodiment.
The fifth concrete implementation mode:
the embodiment is a railway motor car sand pipe loosening fault detection method based on deep learning, and the residual error module structure in the embodiment is as follows:
the input passes through 1 × 1 conv, then is divided into two branches, one branch passes through 3 × 3 conv, then passes through 1 × 1 conv, the other branch passes through 1 × 1 conv, and then the outputs of the two branches are merged to be the final output.
The other steps and model structure are the same as those in the fourth embodiment.
The sixth specific implementation mode:
the implementation mode is a railway motor car sanding pipe loosening fault detection method based on deep learning, and in the deep residual error network structure of the implementation mode, a conv layer is 7 × 7 conv; the first pool layer was 2 x 2 pools.
The other steps and model structures are the same as those of the fourth or fifth embodiment.
The seventh embodiment:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and in the deep residual error network structure of the embodiment, the second pool layer is an avgpool layer.
The other steps and model structures are the same as in the fourth, fifth or sixth embodiment.
The specific implementation mode is eight:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and in the deep residual error network structure of the embodiment, the number of residual error modules is 6.
The other steps and model structures are the same as in embodiments four, five, six or seven.
The specific implementation method nine:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, and the training process of the CASCADERCNN network in the embodiment comprises the following steps:
inputting a target image into a depth residual error network to perform image feature extraction to obtain a feature map;
inputting the feature map output by the depth residual error network into an RPN (resilient packet network), generating a target region candidate frame, determining the type of a foreground or a background of the region according to the classification of the candidate frame, and performing regression adjustment on the position of the candidate frame;
inputting the foreground candidate frame obtained by the RPN into the prediction network to obtain each target classification and the position of the target frame, then bringing the obtained target frame into the prediction network again, repeating iteration for 3 times, and setting the iou of 3 times as 0.5,0.6 and 0.7 respectively to obtain the final each target classification and the position of the target frame.
The other steps and the model structure are the same as in one of the first to eighth embodiments.
The detailed implementation mode is ten:
the embodiment is a railway motor car sanding pipe loosening fault detection method based on deep learning, in the embodiment, a feature map output by a deep residual error network is input into an RPN network, and the process of generating a target area candidate frame comprises the following steps:
sliding on the low-dimensional feature map by using a sliding window, mapping on a target image corresponding to the sanding pipe component area in the center of the sliding window, when the IOU of the area mapped on the target image and the target position of the mark in the mark file is more than 0.7, the candidate frame area is a positive sample, when the IOU of the area mapped on the target image and the target position of the mark in the mark file is less than 0.3, the candidate frame area is a negative sample, and then, taking the positive sample as 1: training an RPN (resilient packet network), wherein the final output classification regression task of training is to randomly extract 64 regions with the IOU (input output) more than 0.5 from the real target mark position as a foreground and extract the regions with the IOU more than 0.1 and less than 0.5 as a background.
The other steps and model structure are the same as in the ninth embodiment.
Examples
As shown in fig. 6, the method for detecting loosening fault of a railway bullet train sanding pipe based on deep learning in the embodiment includes the following steps:
1. establishing a sample data set
And (4) building imaging equipment on two sides of the railway track, and acquiring high-definition images after the motor car passes through the equipment. The image is a sharp grayscale image. The motor train parts can be influenced by natural conditions such as rainwater, mud stains, oil stains, black paint and the like or artificial conditions. Also, there may be differences in the images taken at different sites. Therefore, the images of the sanding pipe components vary widely. Therefore, in the process of collecting the image data of the sanding pipe, the diversity is ensured, and the sanding pipe images under various conditions are collected as completely as possible.
The shape of the sandpipe sections may vary among different types of trucks. But some of the less common truck types have sanding pipe components that are difficult to collect due to the large frequency differences that occur between the different types. Accordingly, all types of sandpipe components are collectively referred to as a class, and the sample data set is established all by class.
The sample data set includes: a set of grayscale images and a set of markup files.
The grayscale image set is a high-definition grayscale image shot by the device. The marking file set is a file set which stores the names of the gray level images, the types of the fault forms and the coordinates of the upper left corner and the lower right corner of the target area in the xml file, and the sanding pipe is loosened and has six fault forms, so that the images are marked as six types and are acquired in a manual marking mode, wherein the six types of fault forms comprise: the shape of the sand spreading pipe joint is broken, the shape of the two sand spreading pipe joints is broken, the shape of the sand spreading pipe with metal luster after being loosened (showing falling), the shape of the sand spreading pipe bound with a white adhesive tape (simulating fault), the shape of the sand spreading pipe bound with a black adhesive tape (simulating fault) and the normal shape. There is a one-to-one correspondence between the grayscale image set and the marker file set, i.e., each grayscale image corresponds to one marker file.
Although the creation of the sample data set includes images under various conditions, data amplification of the sample data set is still required to improve the stability of the algorithm. The amplification form comprises operations of rotation, translation, zooming, mirror image and the like of the image, and each operation is performed under random conditions, so that the diversity and applicability of the sample can be ensured to the greatest extent.
2. Initial positioning of sand spraying pipe
According to the hardware equipment, the wheel base information, the relevant position and other prior knowledge, the area image of the sanding pipe component can be preliminarily intercepted from the gray level image acquired by the camera.
3. Improving image contrast
Because the angle distances of the imaging devices of all stations are different, the brightness degrees of the collected images are different, and some images are too dark to clearly observe the loosening area of the sanding pipe, the contrast of the images is improved in a self-adaptive mode before entering a deep learning network, and the contrast of the images can be improved in a self-adaptive mode by calculating the average pixel of the images first and improving the contrast of the pixels with the pixel number smaller than 40. Taking the image with the improved contrast as a target image;
4. multi-target detection training
1) Initializing CASCADERCNN network parameters by using model parameters of the coco data set; CASCADERCNN the network includes a deep residual network, an RPN network and a prediction network (prediction part of the existing CASCADERCNN network);
as shown in fig. 5, the depth residual error network is an improved depth residual error network, and the network structure is as follows: conv layer + pool layer + a plurality of residual modules + pool layer + FC layer; preferably, the total number of the layers is 7 × 7conv +2 × 2pool +6 residual modules + avgpool + FC, the preferred network structure is specially set for the sanding pipe and the image characteristics, and the preferred network structure has a very good effect when the sanding pipe is detected by using the preferred network structure;
the residual module is specifically formed as shown in fig. 4, the input first passes through 1 × 1 conv, then is divided into two branches, one branch passes through 3 × 3 conv, then passes through 1 × 1 conv, the other branch passes through 1 × 1 conv, and then the outputs of the two branches are fused to be the final output;
Figure BDA0002750863690000071
wherein, IiRepresenting the output before fusion of residual modules, WiAnd (3) representing the weights of the feature maps with different scales, and O representing the final output of the feature map with different scales.
As shown in fig. 3, fig. 3 is a schematic diagram of a processing flow corresponding to the CASCADERCNN training process, where I denotes an input image, conv + pool denotes a backbone convolutional network, i.e., a deep residual network, H denotes a portion corresponding to the RPN and a prediction network, and a prediction is made using the previously extracted features, B0 denotes a candidate box, B1, B2, and B3 denote position regression on a current target, and C1, C2, and C3 denote classes of the target, respectively.
2) Inputting the target image into a depth residual error network to extract the characteristics of the image to obtain a characteristic diagram;
3) and inputting the feature map output by the depth residual error network into an RPN (resilient packet network), generating a target region candidate frame, determining the type of the region as a foreground or a background according to the classification of the candidate frame, and performing regression adjustment on the position of the candidate frame. The method specifically comprises the steps of sliding on a low-dimensional feature map by using a sliding window, mapping the center of the sliding window to a target image corresponding to a sanding pipe component area, when the IOU of the area mapped to the target image and the target position of a mark in a mark file is larger than 0.7, determining that the candidate frame area is a positive sample, when the IOU of the area mapped to the sanding pipe component on the target image and the target position of the mark in the mark file is smaller than 0.3, determining that the candidate frame area is a negative sample, and then taking the positive sample as 1: training an RPN (resilient packet network), wherein the final output classification regression task of training is to randomly extract 64 regions with the IOU (input output) more than 0.5 from the real target mark position as a foreground and extract the regions with the IOU more than 0.1 and less than 0.5 as a background.
4) Inputting the obtained foreground candidate frame into a prediction network to obtain each target classification and a target frame position, then bringing the obtained target frame into the prediction network again, repeating iteration for 3 times, and setting the iou of 3 times as 0.5,0.6 and 0.7 respectively to obtain the final each target classification and the target frame position.
During training, the loss function is as follows:
Figure BDA0002750863690000072
wherein N isclsTo classify the total number of samples, piIs the classification probability of the different classes; n is a radical ofregNumber of regression frames, ti＝{tx,ty,tw,thIs a vector, tx,ty,tw,thRespectively representing the offset of the x coordinate, the offset of the y coordinate, the offset of the width and the offset of the height of the candidate frame; the parameter λ is used to weigh the ratio of classification loss to regression loss, when the sample is positive
Figure BDA0002750863690000081
Is 1, negative sample
Figure BDA0002750863690000082
Is 0;
Figure BDA0002750863690000083
is and tiAnd vectors with the same dimension represent the offset of the candidate frame to the marked frame.
The formula (1) is divided into two parts, wherein the first part is a classification loss function of target prediction, and the second part is a position loss function of regression prediction:
classification loss function for target prediction:
Figure BDA0002750863690000084
wherein gamma is a positive number, alphaiFor adjusting the ratio of positive and negative samples, alphai∈[0，1]H represents the predicted current category probability value, t represents the real label category value, when the category of the prediction frame is the same as that of the label frame, the value of t is 1, and when the category of the prediction frame is different from that of the label frame, the value of t is 0.
Figure BDA0002750863690000085
Representing the similarity of the predicted frame class to the real frame class. For example, when the prediction box class is the same as the tag box class, and the probability is 1,
Figure BDA0002750863690000086
the loss function is-1, so that the influence caused by the uneven category in the cascadercnn is solved, and the accuracy is improved.
Regression predicted position loss function:
Figure BDA0002750863690000087
Figure BDA0002750863690000088
Figure BDA0002750863690000089
a flow chart of weight coefficient determination during training is shown in fig. 2. And obtaining a trained overall prediction network model by using the sample data set, and outputting the fault category and the corresponding position by using the overall prediction network model.
5. And (3) detecting the sanding pipe: and acquiring an image of the sanding pipe to be detected, and detecting by using the trained integral prediction network model to obtain the fault category and the corresponding position.
The present invention is capable of other embodiments and its several details are capable of modifications in various obvious respects, all without departing from the spirit and scope of the present invention.

Claims (10)

1. The railway motor car sanding pipe loosening fault detection method based on deep learning is characterized by comprising the following steps of:
acquiring an image of the sanding pipe to be detected, detecting by using an CASCADERCNN network model, and outputting fault types and corresponding positions;
CASCADERCNN the loss function of the network during training is as follows:
Figure FDA0002750863680000011
wherein N isclsTo classify the total number of samples, piIs the classification probability of the different classes; n is a radical ofregNumber of regression frames, ti＝{tx,ty,tw,thIs a vector, tx,ty,tw,thRespectively representing the offset of the x coordinate, the y coordinate, the width and the height of the candidate frame; the parameter λ is used to weigh the ratio of classification loss to regression loss, when the sample is positive
Figure FDA0002750863680000012
Is 1, negative sample
Figure FDA0002750863680000013
Is 0;
Figure FDA0002750863680000014
is and tiVectors with the same dimension represent the offset of the candidate frame to the mark frame; l iscls(pi) A classification loss function that is predicted for the target,
Figure FDA0002750863680000015
a position loss function that is a regression prediction;
classification loss function for target prediction:
Figure FDA0002750863680000016
wherein gamma is a positive number, alphaiThe method is used for adjusting the proportion of positive and negative samples, h represents the probability value of the current prediction class, and t represents the class value of the real label.
2. The deep learning-based railway car sanding pipe loosening fault detection method according to claim 1, wherein the regression predicted position loss function is as follows:
Figure FDA0002750863680000017
wherein the content of the first and second substances,
Figure FDA0002750863680000018
3. the railway bullet train sanding pipe loosening fault detection method based on deep learning of claim 2, wherein the CASCADERCNN network comprises a deep residual error network, an RPN network and a prediction network, and the specific structure is as follows:
the deep residual error network, the RPN and the prediction network are connected in sequence, the output of the deep residual error network is input into the RPN, and the RPN enters and exits the foreground candidate frame and is input into the prediction network.
4. The railway vehicle sanding pipe loosening fault detection method based on deep learning of claim 3, wherein the deep residual error network structure is a conv layer, a first pool layer, a plurality of residual error modules, a second pool layer and an FC layer.
5. The railway motor car sanding pipe loosening fault detection method based on deep learning of claim 4, wherein the residual error module structure is as follows:
the input passes through 1 × 1 conv, then is divided into two branches, one branch passes through 3 × 3 conv, then passes through 1 × 1 conv, the other branch passes through 1 × 1 conv, and then the outputs of the two branches are merged to be the final output.
6. The railway motor car sanding pipe loosening fault detection method based on deep learning of claim 5, wherein in the deep residual error network structure, the conv layer is 7 x 7 conv; the first pool layer was 2 x 2 pools.
7. The method for detecting loosening faults of sanding pipes of railway motor cars based on deep learning of claim 6, wherein in the deep residual error network structure, the second pool layer is avgpool.
8. The railway motor car sanding pipe loosening fault detection method based on deep learning of claim 4, 5, 6 or 7, wherein in the deep residual error network structure, the number of residual error modules is 6.
9. The railway bullet sanding pipe loosening fault detection method based on deep learning of claim 8, wherein the training process of the CASCADERCNN network comprises the following steps:
inputting a target image into a depth residual error network to perform image feature extraction to obtain a feature map;
inputting the feature map output by the depth residual error network into an RPN (resilient packet network), generating a target region candidate frame, determining the type of a foreground or a background of the region according to the classification of the candidate frame, and performing regression adjustment on the position of the candidate frame;
inputting the foreground candidate frame obtained by the RPN into the prediction network to obtain each target classification and the position of the target frame, then bringing the obtained target frame into the prediction network again, repeating iteration for 3 times, and setting the iou of 3 times as 0.5,0.6 and 0.7 respectively to obtain the final each target classification and the position of the target frame.
10. The railway bullet train sanding pipe loosening fault detection method based on deep learning of claim 9, wherein the feature map output by the deep residual error network is input into the RPN network, and the process of generating the target area candidate frame comprises the following steps:
sliding on the low-dimensional feature map by using a sliding window, mapping on a target image corresponding to the sanding pipe component area in the center of the sliding window, when the IOU of the area mapped on the target image and the target position of the mark in the mark file is more than 0.7, the candidate frame area is a positive sample, when the IOU of the area mapped on the target image and the target position of the mark in the mark file is less than 0.3, the candidate frame area is a negative sample, and then, taking the positive sample as 1: training an RPN (resilient packet network), wherein the final output classification regression task of training is to randomly extract 64 regions with the IOU (input output) more than 0.5 from the real target mark position as a foreground and extract the regions with the IOU more than 0.1 and less than 0.5 as a background.
CN202011183656.3A 2020-10-29 2020-10-29 Railway motor car sand spreading pipe loosening fault detection method based on deep learning Pending CN112308135A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011183656.3A CN112308135A (en) 2020-10-29 2020-10-29 Railway motor car sand spreading pipe loosening fault detection method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011183656.3A CN112308135A (en) 2020-10-29 2020-10-29 Railway motor car sand spreading pipe loosening fault detection method based on deep learning

Publications (1)

Publication Number Publication Date
CN112308135A true CN112308135A (en) 2021-02-02

Family

ID=74332160

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011183656.3A Pending CN112308135A (en) 2020-10-29 2020-10-29 Railway motor car sand spreading pipe loosening fault detection method based on deep learning

Country Status (1)

Country Link
CN (1) CN112308135A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170882A (en) * 2022-07-19 2022-10-11 哈尔滨市科佳通用机电股份有限公司 Optimization method of rail wagon part detection network and guardrail breaking fault identification method

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110414394A (en) * 2019-07-16 2019-11-05 公安部第一研究所 A kind of face blocks face image method and the model for face occlusion detection
CN111079747A (en) * 2019-12-12 2020-04-28 哈尔滨市科佳通用机电股份有限公司 Railway wagon bogie side frame fracture fault image identification method
CN111402226A (en) * 2020-03-13 2020-07-10 浙江工业大学 Surface defect detection method based on cascade convolution neural network
EP3686772A1 (en) * 2019-01-25 2020-07-29 Tata Consultancy Services Limited On-device classification of fingertip motion patterns into gestures in real-time
CN111832398A (en) * 2020-06-02 2020-10-27 国网浙江嘉善县供电有限公司 Unmanned aerial vehicle image distribution line pole tower ground wire broken strand image detection method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3686772A1 (en) * 2019-01-25 2020-07-29 Tata Consultancy Services Limited On-device classification of fingertip motion patterns into gestures in real-time
CN110414394A (en) * 2019-07-16 2019-11-05 公安部第一研究所 A kind of face blocks face image method and the model for face occlusion detection
CN111079747A (en) * 2019-12-12 2020-04-28 哈尔滨市科佳通用机电股份有限公司 Railway wagon bogie side frame fracture fault image identification method
CN111402226A (en) * 2020-03-13 2020-07-10 浙江工业大学 Surface defect detection method based on cascade convolution neural network
CN111832398A (en) * 2020-06-02 2020-10-27 国网浙江嘉善县供电有限公司 Unmanned aerial vehicle image distribution line pole tower ground wire broken strand image detection method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
KAIMING HE: ""Focal Loss for Dense Object Detection"", 《ARXIV》 *
REN S: ""Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks"", 《ARXIV》 *
ZHAOWEI CAI: ""Cascade R-CNN: Delving into High Quality Object Detection"", 《AIXIV》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115170882A (en) * 2022-07-19 2022-10-11 哈尔滨市科佳通用机电股份有限公司 Optimization method of rail wagon part detection network and guardrail breaking fault identification method

Similar Documents

Publication Publication Date Title
CN109977812B (en) Vehicle-mounted video target detection method based on deep learning
CN107316064B (en) Asphalt pavement crack classification and identification method based on convolutional neural network
CN111079747B (en) Railway wagon bogie side frame fracture fault image identification method
Chen et al. A self organizing map optimization based image recognition and processing model for bridge crack inspection
CN113160192B (en) Visual sense-based snow pressing vehicle appearance defect detection method and device under complex background
CN112434695B (en) Upper pull rod fault detection method based on deep learning
CN111899288B (en) Tunnel leakage water area detection and identification method based on infrared and visible light image fusion
CN111091544B (en) Method for detecting breakage fault of side integrated framework of railway wagon bogie
CN111080611A (en) Railway wagon bolster spring fracture fault image identification method
CN114998852A (en) Intelligent detection method for road pavement diseases based on deep learning
CN112330593A (en) Building surface crack detection method based on deep learning network
CN107798293A (en) A kind of crack on road detection means
CN111080621B (en) Method for identifying railway wagon floor damage fault image
CN111652295A (en) Railway wagon coupler yoke key joist falling fault identification method
CN113436157A (en) Vehicle-mounted image identification method for pantograph fault
CN108648210B (en) Rapid multi-target detection method and device under static complex scene
CN112101138A (en) Bridge inhaul cable surface defect real-time identification system and method based on deep learning
CN114723709A (en) Tunnel disease detection method and device and electronic equipment
CN115527170A (en) Method and system for identifying closing fault of door stopper handle of automatic freight car derailing brake device
CN111667473A (en) Insulator hydrophobicity grade judging method based on improved Canny algorithm
CN115995056A (en) Automatic bridge disease identification method based on deep learning
CN112329858B (en) Image recognition method for breakage fault of anti-loosening iron wire of railway motor car
Guo et al. Surface defect detection of civil structures using images: Review from data perspective
CN112308135A (en) Railway motor car sand spreading pipe loosening fault detection method based on deep learning
CN113762247A (en) Road crack automatic detection method based on significant instance segmentation algorithm

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
RJ01 Rejection of invention patent application after publication

Application publication date: 20210202

RJ01 Rejection of invention patent application after publication