CN116310649A - Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster - Google Patents

Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster Download PDF

Info

Publication number
CN116310649A
CN116310649A CN202310299281.4A CN202310299281A CN116310649A CN 116310649 A CN116310649 A CN 116310649A CN 202310299281 A CN202310299281 A CN 202310299281A CN 116310649 A CN116310649 A CN 116310649A
Authority
CN
China
Prior art keywords
round pin
unit
cotter
image
label
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
CN202310299281.4A
Other languages
Chinese (zh)
Inventor
孙晶
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 CN202310299281.4A priority Critical patent/CN116310649A/en
Publication of CN116310649A publication Critical patent/CN116310649A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The method for detecting the loss of the round pin and the round pin cotter pin of the adjusting screw of the brake adjuster solves the problem that the accuracy of the loss detection of the round pin and the round pin cotter pin of the adjusting screw of the existing brake adjuster is low, and belongs to the technical field of railway wagon fault detection. The invention comprises the following steps: s1, acquiring an image of the bottom of a railway wagon passing through a vehicle, intercepting images of round pins and round pin cotter pins of a brake adjuster control rod, S2, sending the intercepted images into a fault detection model, and outputting a detection result by the fault detection model; the fault detection model is a Mask R-CNN network model, and a backbone network of the Mask R-CNN network model is ResNet50; the weight of the pre-training ResNet50 is trained by adopting a comparison self-supervision method, and then a fault detection model is trained by adopting an example segmentation labeled sample data set; s3, if the round pin and the round pin cotter pin of the brake adjuster control rod are not detected in the detection result, the image is a fault image, and fault alarm is given, otherwise, no fault exists.

Description

Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster
Technical Field
The invention relates to a method for detecting loss of a round pin and a round pin cotter pin of an adjusting screw of a brake adjuster, and belongs to the technical field of railway wagon fault detection.
Background
Compared with the traditional railway wagon fault detection method adopting manual image checking, the automatic fault detection method based on deep learning can remarkably improve the detection efficiency, reduce the cost and avoid the phenomena of missing detection and false detection caused by fatigue of the car checking staff. However, as the brake adjuster adjusts the screw rod round pin and the round pin cotter pin to be lost and located at the bottom of the train, the round pin image target is small and has ambiguity, the shape of the cotter pin target has diversity, and the accuracy of fault detection by adopting the traditional Mask R-CNN deep learning segmentation network is low.
Disclosure of Invention
Aiming at the problem that the accuracy of the loss detection of the round pin and the round pin cotter of the adjusting screw of the brake adjuster is low in the traditional Mask R-CNN deep learning segmentation network, the invention provides a method for detecting the loss of the round pin and the round pin cotter of the adjusting screw of the brake adjuster.
The invention relates to a method for detecting loss of a round pin and a round pin cotter pin of an adjusting screw of a brake adjuster, which comprises the following steps:
s1, acquiring an image of the bottom of a railway wagon passing through a vehicle, intercepting the images of round pins and round pin cotter pins of a brake adjuster control rod,
s2, sending the intercepted image into a fault detection model, and outputting a detection result by the fault detection model;
the fault detection model is a Mask R-CNN network model, and a backbone network of the Mask R-CNN network model is ResNet50;
the fault detection model training method comprises the following steps:
s21, building images of different parts, different textures and different scales of a label-free railway wagon as a pre-training set, and carrying out data enhancement on the pre-training set to enable each image in the pre-training set to generate a positive sample pair (x) i ,x j ) And negative sample z k ；
S22, x i And x j Respectively inputting the two features into two backbone network encoders, respectively extracting the features from the two backbone network encoders, respectively inputting the features into one MLP network, and outputting z by the two MLP networks i And z j The method comprises the steps of carrying out a first treatment on the surface of the Two backbone network encoders have the same architecture, with the first backbone network Encoder employing back propagation to update parameters and the second backbone network EncThe oder updates parameters by adopting a momentum method according to the parameters of the first backbone network Encoder; both backbone networks Encoder are ResNet50;
s23, z output according to two MLP networks i And z j Negative sample z k Calculating a loss function, and adjusting the weight of a first backbone network Encoder;
s24, marking the round pin and the round pin cotter pin parts of the brake adjuster adjusting screw, and carrying out data amplification to obtain an example-segmented label sample data set;
s25, after the pre-training of the ResNet50 in the fault detection model is completed by adopting S21 to S23, taking the backbone network Encoder as the ResNet50 in the fault detection model, and training the fault detection model by adopting an example segmentation label sample data set;
s3, if the round pin and the round pin cotter pin of the brake adjuster control rod are not detected in the detection result, the image is a fault image, and fault alarm is given, otherwise, no fault exists.
Preferably, each MLP network is composed of a Linear layer Linear, a batch normalized BN, an activation function ReLU, a Linear layer Linear connected in sequence.
Preferably, the method for performing data amplification in S24 includes:
firstly, carrying out Laplace gradient calculation on a round pin cotter image, and carrying out Gaussian blur on the image meeting gradient requirements by adopting a probability method;
the method 2, copying round pin cottage pin images to a new scene;
and 3, randomly adjusting the brightness of the image by using the images obtained by the method 1 and the method 2, and carrying out self-adaptive histogram equalization conceptualization and noise adding operation.
Preferably, S24 further comprises screening the data amplified instances segmented with the tag sample data set:
if the confidence of the prediction frame of the segmentation target of a certain sample in the sample data set segmented with the label in the example is less than 0.5, deleting the sample;
if the confidence of the prediction frame of the segmentation target of a certain sample in the sample data set of the sample segmentation label is not less than 0.5 and the true value file is not available, adding the label information;
if the IOU of the division target prediction frame and the truth frame of a certain sample in the sample data set divided by the label is larger than a set threshold value, but the categories are inconsistent, the label information is corrected;
if a certain sample in the sample data set with the label is segmented, the target is not segmented, but the label information exists, whether the label information is reasonable or not is judged, and if not, the sample is deleted.
Preferably, the method of segmented data amplification is adopted, during the training process, the label sample data set is segmented by using the amplified examples in the 1 st to 18 th rounds, and the label sample data set is segmented by using the amplified examples in the 19 th to 20 th rounds.
In S21, acquiring a passing image, and collecting images of different parts, different textures and different dimensions of a railway wagon to form a pre-training set according to wheelbase information and prior information of the parts, wherein the different parts comprise springs, bolts, nuts, chains, cylinders, pull rods, pipe bodies, round pins and cotters; the different scales include 1024×1024, 1024×512, 512×512, 256×512, 512×256, 256×256.
Preferably, the feature pyramid of the Mask R-CNN network model is a PAN-FPN network, the ResNet50 comprises 4 groups of residual units which are sequentially connected, the intercepted image is sent to the 1 st group of residual units, and a C2 feature map, a C3 feature map and a C4 feature map which are output by the last 3 groups of residual units are input to the PAN-FPN network;
the PAN-FPN network comprises a No. 1 up-sampling module, a No. 1 connecting unit C, a No. 1C 4f unit, a No. 2 up-sampling unit, a No. 2 connecting unit C, a No. 2C 4f unit, a No. 1 CBS unit, a No. 3 connecting unit C, a No. 3C 4f unit, a No. 2 CBS unit, a No. 4 connecting unit C and a No. 4C 4f unit;
the C4 feature map is simultaneously input to an up-sampling unit No. 1 and a connecting unit No. 4; the output of the No. 1 up-sampling unit and the C3 feature map are input to a No. 1 connecting unit for connection, the output of the No. 1 connecting unit is input to a No. 1C 4f unit, the output of the No. 1C 4f unit is input to a No. 2 up-sampling unit and a No. 3 connecting unit simultaneously, the output of the No. 2 up-sampling unit and the C2 feature map are input to a No. 2 connecting unit for connection, the output of the No. 2 connecting unit is input to a No. 2C 4f unit, the C4f unit outputs the feature P2, the feature P2 is input to a No. 1 CBS unit, the output of the No. 1 CBS unit is input to a No. 3 connecting unit, the output of the No. 3 CBS unit is input to a No. 4 connecting unit, the output of the No. 4 connecting unit is input to a No. 4C 4f unit, and the C4f unit outputs the feature P4;
the C4f unit divides an input characteristic diagram into 4 channels according to the channels, the 1 st channel is subjected to a Bottleneck operation, the 2 nd channel is respectively subjected to 1 Bottleneck operation and 2 Bottleneck operation, the 3 rd channel is respectively subjected to 1 Bottleneck operation, 2 Bottleneck operation and 3 Bottleneck operation, all channels after the Bottleneck operation are connected with the 4 th channel, and then the output of the C4f unit is obtained through a CBS operation;
the CBS units are Conv-BN-SiLU combination operations.
Preferably, the optimization of the Mask R-CNN network model is AdamW, adamW is obtained by adding the Adam optimization with regularization, wherein the initial learning rate is 0.001, the weight attenuation coefficient is 0.005, training of the learning rate is updated by adopting a segmented Cosine LR method, the fixed learning rate is adopted in the first 60%, and the learning rate is updated by adopting a Cosine annealing mode in the latter 40%.
The method has the beneficial effects that the pre-training weight is trained by adopting the specific component data of the trucks and the comparison self-supervision method, so that the Mask R-CNN backbone network has stronger feature extraction capability for truck components with different scales aiming at different textures, and the method is more beneficial to improving the precision of subsequent downstream segmentation tasks. According to the invention, the FPN in the Mask R-CNN network model is improved, the traditional FPN is replaced by the improved PAN-FPN, and the separation channel operation is added, so that the gradient reflux is increased, and the network performance is further improved. The invention adopts the image amplification optimization method aiming at the cotter trucks, and because the copy examples and some conditions after Gaussian blur amplification deviate from real prediction, the invention adopts the segmentation amplification method, and has a certain new performance improvement compared with the whole-course copy examples and Gaussian blur. The data amplification is combined with the label sample data set optimization strategy to be more beneficial to performance improvement. The optimizer in the Mask R-CNN network model is replaced by AdamW, and the training of the learning rate is updated by adopting a segmented Cosine LR method, so that the accuracy is improved.
Drawings
FIG. 1 is a flow chart of the detection method of the invention;
FIG. 2 is a diagram of self-supervising and supervised relationships;
FIG. 3 is a flow chart of the self-supervised training pre-training weights of the present invention;
FIG. 4 is a schematic diagram of a PAN-FPN map of the present invention;
FIG. 5 is a schematic diagram of a c4f unit;
fig. 6 is a schematic diagram of the principle of the Bottleneck operation.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.
The method for detecting loss of the round pin and the round pin cotter of the damper adjusting screw according to the present embodiment comprises:
step 1, after a railway motor car passes through high-definition imaging equipment erected around, acquiring an image of the bottom of a railway wagon passing through the railway motor car, and intercepting the images of round pins and round pin cotter pins of a brake adjuster control rod;
step 2, sending the intercepted image into a fault detection model, and outputting a detection result by the fault detection model;
and step 3, if the round pin and the round pin cotter pin of the brake adjuster control rod are not detected in the detection result, the image is a fault image, a fault message is uploaded, and a vehicle inspection person carries out next processing according to the fault message. Otherwise, the next image is continuously detected.
The fault detection model of the embodiment is a Mask R-CNN network model, and a backbone network of the Mask R-CNN network model is ResNet50;
the Mask R-CNN example segmentation method is an example segmentation Mask map which can effectively detect targets and output high quality, mask R-CNN is an extension of fast R-CNN, and parallel prediction segmentation branches are performed at the same time of boundary frame detection, in order to extract deeper information, resNet-50 is used as a backbone network, and self-supervision and non-label truck data training pre-training weights are used, self-supervision learning is to learn features on a large number of non-label samples to obtain the pre-training weights, and fine adjustment is performed on a small number of label methods to obtain a better model, as shown in fig. 2. In the pre-training weight process, a large number of unlabeled TFDS truck linear array images are adopted as training data sets of a self-supervision method, and the method is easier to learn characteristic information of the linear array truck images than a common pre-training method using an ImageNet data set, and can obtain higher recognition rate on a downstream target detection task when the method is adopted as the pre-training weight.
The present embodiment adopts a self-supervision method to train a pre-training model, so that it is hoped to learn a general feature for downstream detection and segmentation tasks, the goal of contrast learning is to make the model learn to distinguish between positive and negative samples, make the similarity between positive samples as high as possible, and make the similarity between negative samples as low as possible, and the specific flow is shown in fig. 3.
The fault detection model training method comprises the following steps:
step 21, establishing images of different parts, different textures and different scales of the label-free railway wagon as a pre-training set, and carrying out data enhancement on the pre-training set to enable each image in the pre-training set to generate positive valuesSample pair (x) i ,x j ) And negative sample z k ；
And 21, when data are collected, high-definition imaging equipment is set up around a railway motor car track, after a motor car passes, a car passing image is obtained, sub-image information of different parts is collected according to wheelbase information and prior information of the parts, and area images of the same part are placed in a folder to serve as a class, and the fact that the area images added with the parts are compared with the images of the whole car to serve as a class has better effect. In order to extract the texture features and the features of different scales of different trucks, the collected components comprise springs, bolts, nuts, chains, cylinders, pull rods, pipe bodies, round pins, cotters and the like, and the sizes of the components are 1024 x 1024, 1024 x 512, 512 x 512, 256 x 512, 512 x 256, 256 x 256 and the like, and the prepared truck component data without labels are used as a self-supervision label-free sample data set. For a truck image x, a positive sample pair (x i ,x j ) Taking other images in batch as negative samples z k 。
In the step 21, when data enhancement is performed, after random transformation such as clipping, overturning, color dithering and the like is performed on a truck image, a combined data amplification method such as image stretching, color dithering, brightness adjustment, gaussian blur and the like is found for railway image target detection, and the amplification method in the embodiment is more suitable for the characteristics of railway data such as image stretching, different illumination, different shooting time and complex background.
Step 22, X i And x j Respectively inputting the two features into two backbone network encoders, respectively extracting the features from the two backbone network encoders, respectively inputting the features into one MLP network, and outputting z by the two MLP networks i And z j The method comprises the steps of carrying out a first treatment on the surface of the The two backbone network encoders have the same architecture, wherein the first backbone network Encoder adopts back propagation to update parameters, and the second backbone network Encoder adopts a momentum method to update parameters according to the parameters of the first backbone network Encoder; both backbone networks Encoder are ResNet50;
step 23, z according to two MLP network outputs i And z j Negative sample z k Calculating a loss function, and adjusting the weight of a first backbone network Encoder;
step 24, marking the round pin and the round pin cotter pin parts of the brake adjuster adjusting screw, and carrying out data amplification to obtain an example-segmented label sample data set;
and 25, after the pre-training of the ResNet50 in the fault detection model is completed by adopting the steps 21 to 23, taking the backbone network Encoder as the ResNet50 in the fault detection model, and training the fault detection model by adopting the sample data set with the label by example segmentation.
The MLP of the embodiment is a nonlinear projection layer, the MLP is a nonlinear projection layer, the embodiment improves the original MLP of Linear-ReLU-Linear into Linear layer Linear, batch normalization BN, activation function ReLU, linear layer Linear, batch normalization BN, activation function ReLU and Linear layer Linear which are connected in sequence in self-supervision training, firstly, the effect of increasing BN is to provide better initialization scale and compensate incorrect initialization, secondly, a plurality of Linear-BN-ReLU combinations are added, the nonlinear learning capacity can be increased, the learning quality can be greatly improved, and the effect on the subsequent downstream target segmentation task is better.
Step 24 of the embodiment establishes high-definition imaging equipment around a railway motor car track, acquires a car passing image after a motor car passes, and intercepts a brake adjuster control lever round pin and a round pin cotter pin. Taking the images with the brake adjuster adjusting screw round pins and the round pin cotter pins as positive samples, taking the images with the brake adjuster adjusting screw round pins and the round pin cotter pins which are lost as negative samples, ensuring that the number of the positive samples is basically consistent with that of the negative samples, marking the parts of the brake adjuster adjusting screw round pins and the round pin cotter pins by marking software, completing the manufacture of a data set, and dividing a training set, a verification set and a test set for the data set. And the following three special data amplification methods are adopted for the characteristics of the components:
in the method 1, because round pins and round pin cotter pin parts are smaller at the middle part of a truck body, and the images shot at the positions of cameras of different truck types are blurred, the Laplace gradient calculation is firstly carried out on the round pin cotter pin images aiming at the situation, and the Gaussian blur is carried out on the images meeting the gradient requirement by adopting a probability method;
the method 2 is that aiming at the condition that the round pin cotter pin has stretching on the image, the round pin cotter pin image is copied to a new scene, and more complex new data are created:
firstly, selecting two truck images, firstly carrying out random scale dithering on round pins and cotters on the two truck images, wherein the random scale dithering is an operation combination of zooming, cutting, overturning, stretching, filling and the like, then copying the round pin instance and the cotter instance on one of the two truck images into the other truck image, and if the other truck image also has the same instance object, ensuring that the two instance objects are not overlapped, and creatively extracting complex new data of more scenes by copying the instance objects at the pixel level.
And 3, randomly performing image brightness adjustment, self-adaptive histogram equalization conceptualization and noise adding operation on the images obtained by the method 1 and the method 2 to finish amplification of a data set, wherein the data amplification operation can enhance the generalization capability of a subsequent detection network and reduce the probability of network overfitting.
The embodiment also comprises a method for optimizing the sample data set with the label divided for the example, wherein the established brake adjuster adjusts the image data of the round pin and the round pin cotter of the screw rod, and the method is formed by manual labeling. Therefore, there must be many false marks and false leaks in the component data set, both of which can negatively impact the effectiveness of model training. In order to overcome the defect, the patent designs a method for screening a data set for multiple times based on an optimized data updating algorithm to improve the accuracy of model prediction, wherein the specifically screened data is shown in a chart 1. The present embodiment screens the sample data set of the label-segmented data after the data amplification:
(1) Checking data that does not meet the image quality: if the confidence of a segmentation target prediction frame of a certain sample in the sample data set with the labels in the example segmentation is smaller than 0.5, deleting the sample, namely considering that the quality of the current image cannot be accurately segmented, and the camera possibly has faults, so that the quality of the shot image does not reach the standard, and further the segmentation task cannot be performed.
(2) Checking missing mark data: if the confidence of the prediction frame of the segmentation target of a certain sample in the sample data set of the sample segmentation label is not less than 0.5 and the true value file is not available, adding the label information;
(3) Checking for false labels: if the IOU of the division target prediction frame and the truth frame of a certain sample in the sample data set divided by the label is larger than a set threshold value, but the categories are inconsistent, the label information is corrected;
(4) Checking false mark data: if a certain sample in the sample data set with the label is segmented, the target is not segmented, but the label information exists, whether the label information is reasonable or not is judged, and if not, the sample is deleted.
The Mask R-CNN network model of the embodiment adopts an improved feature pyramid PAN-FPN to replace the traditional FPN, the FPN is from top to bottom, the high-level strong semantic features are transferred, the whole pyramid is enhanced, only semantic information is enhanced, and no positioning information is transferred. PAN is aimed at this point by adding a bottom-up pyramid behind the FPN, supplementing the FPN, transferring the strong locating features of the lower layers up, also called "double tower tactics", while in order to increase the gradient reflux of the PAN pyramid, channel separation is used, increasing the gradient reflux.
The ResNet50 comprises 4 groups of residual units which are sequentially connected, the intercepted image is sent to the 1 st group of residual units, and a C2 characteristic diagram, a C3 characteristic diagram and a C4 characteristic diagram which are output by the last 3 groups of residual units are input to a PAN-FPN network;
the PAN-FPN network comprises a No. 1 up-sampling module, a No. 1 connecting unit C, a No. 1C 4f unit, a No. 2 up-sampling unit, a No. 2 connecting unit C, a No. 2C 4f unit, a No. 1 CBS unit, a No. 3 connecting unit C, a No. 3C 4f unit, a No. 2 CBS unit, a No. 4 connecting unit C and a No. 4C 4f unit;
the C4 feature map is simultaneously input to an up-sampling unit No. 1 and a connecting unit No. 4; the output of the No. 1 up-sampling unit and the C3 feature map are input to a No. 1 connecting unit for connection, the output of the No. 1 connecting unit is input to a No. 1C 4f unit, the output of the No. 1C 4f unit is input to a No. 2 up-sampling unit and a No. 3 connecting unit simultaneously, the output of the No. 2 up-sampling unit and the C2 feature map are input to a No. 2 connecting unit for connection, the output of the No. 2 connecting unit is input to a No. 2C 4f unit, the C4f unit outputs the feature P2, the feature P2 is input to a No. 1 CBS unit, the output of the No. 1 CBS unit is input to a No. 3 connecting unit, the output of the No. 3 CBS unit is input to a No. 4 connecting unit, the output of the No. 4 connecting unit is input to a No. 4C 4f unit, and the C4f unit outputs the feature P4;
the C4f unit divides an input characteristic diagram into 4 channels according to the channels, the 1 st channel is subjected to a Bottleneck operation, the 2 nd channel is respectively subjected to 1 Bottleneck operation and 2 Bottleneck operation, the 3 rd channel is respectively subjected to 1 Bottleneck operation, 2 Bottleneck operation and 3 Bottleneck operation, all channels after the Bottleneck operation are connected with the 4 th channel, and then the output of the C4f unit is obtained through a CBS operation;
the CBS units are Conv-BN-SiLU combination operations.
As shown in FIG. 4, the specific PAN-FPN is a backbone network based on ResNet50, and comprises 4 groups of residual units, wherein the output characteristic diagrams of each group of residual units are C1, C2, C3 and C4 respectively, and because round pin cotter pin components are smaller, the embodiment adopts C2, C3 and C4 as PAN pyramids, the pyramids mainly comprise three branches P2, P3 and P4, and the C4f unit is a gradient reflux for enriching the gradient reflux; the Bottleneck operation is an operation in which two CBS operations and one shortcut are added. The improved Mask R-CNN network model performs subsequent classification, regression and masking operations on the PAN pyramid features with more reflow gradient information.
In the embodiment, the pre-training weight is trained by adopting the data of the specific components of the truck and the contrast self-supervision method, so that the Mask R-CNN backbone network has stronger feature extraction capability of truck components with different scales aiming at different textures, is more beneficial to improving the precision of the subsequent downstream segmentation task, and is improved by about 0.9% by map.
In the embodiment, the improved Mask R-CNN is adopted to replace the traditional FPN in a network structure by adopting PAN-FPN, and a separation channel is added for operation, so that gradient reflux is increased, the performance of the network is further improved, and the map is improved by about 1.1%.
In the embodiment, an image amplification optimization method for a cotter truck is adopted, and because the copy example and some cases after Gaussian blur amplification deviate from real prediction, a segmentation data amplification method is adopted, the 1 st to 18 th round of use examples after amplification are divided into label sample data sets, the last 2 round of use are not used, and compared with the whole-course copy example and Gaussian blur, certain performance is improved. The data amplification is more beneficial to performance improvement by combining the label sample data set optimization strategy, and the map improvement is about 0.8%.
The optimizer of the Mask R-CNN network model adopted in this embodiment is AdamW, where AdamW is an Adam optimizer plus regularization, to limit parameter values not too large, where the initial learning rate is 0.001, the weight attenuation coefficient is 0.005, and training of the learning rate is updated by using a segmented Cosine LR method, where the fixed learning rate is adopted in the first 60%, and the learning rate is updated by using a Cosine annealing mode in the latter 40%, and the optimizer policy adjusts, so that the map is improved by about 0.7%.
The four optimization schemes are integrated, the overall experimental results of the embodiment are compared with those shown in the following table 4, and the table shows that the four optimization schemes can improve the overall performance of the Mask R-CNN (ResNet 50) algorithm, the four optimization schemes respectively improve the mAP value by 0.9%, 1.1%, 0.8% and 0.7%, and the overall improvement of the combination of the four optimization schemes is 2.4%, so that the overall optimization is helpful to the recognition rate.
Table 1 optimization scheme
Figure BDA0004144393110000091
Table 2 data screening results
Figure BDA0004144393110000092
TABLE 3 super parameters
Figure BDA0004144393110000093
TABLE 4 optimization results
Figure BDA0004144393110000101
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other embodiments.

Claims (10)

1. A method for detecting loss of a round pin and a round pin cotter of a damper adjusting screw, the method comprising:
s1, acquiring an image of the bottom of a railway wagon passing through a vehicle, and intercepting the images of round pins and round pin cotter pins of a brake adjuster control rod;
s2, sending the intercepted image into a fault detection model, and outputting a detection result by the fault detection model;
the fault detection model is a Mask R-CNN network model, and a backbone network of the Mask R-CNN network model is ResNet50;
the fault detection model training method comprises the following steps:
s21, establishing images of different parts, different textures and different scales of the label-free railway wagon as a pre-training set, carrying out data enhancement on the pre-training set,each image in the pre-training set is made to generate a positive sample pair (x i ,x j ) And negative sample z k ；
S22, x i And x j Respectively inputting the two features into two backbone network encoders, respectively extracting the features from the two backbone network encoders, respectively inputting the features into one MLP network, and outputting z by the two MLP networks i And z j The method comprises the steps of carrying out a first treatment on the surface of the The two backbone network encoders have the same architecture, wherein the first backbone network Encoder adopts back propagation to update parameters, and the second backbone network Encoder adopts a momentum method to update parameters according to the parameters of the first backbone network Encoder; both backbone networks Encoder are ResNet50;
s23, z output according to two MLP networks i And z j Negative sample z k Calculating a loss function, and adjusting the weight of a first backbone network Encoder;
s24, marking the round pin and the round pin cotter pin parts of the brake adjuster adjusting screw, and carrying out data amplification to obtain an example-segmented label sample data set;
s25, after the pre-training of the ResNet50 in the fault detection model is completed by adopting S21 to S23, taking the backbone network Encoder as the ResNet50 in the fault detection model, and training the fault detection model by adopting an example segmentation label sample data set;
s3, if the round pin and the round pin cotter pin of the brake adjuster control rod are not detected in the detection result, the image is a fault image, and fault alarm is given, otherwise, no fault exists.
2. The method for detecting loss of round pin and round pin cotter of a brake adjuster adjusting screw according to claim 1, wherein each MLP network is composed of Linear layer Linear, batch normalized BN, activation function ReLU, linear layer Linear connected in sequence.
3. The method for detecting loss of round pin and round pin cotter of damper adjusting screw according to claim 1, wherein the method for performing data amplification in S24 comprises:
firstly, carrying out Laplace gradient calculation on a round pin cotter image, and carrying out Gaussian blur on the image meeting gradient requirements by adopting a probability method;
the method 2, copying round pin cottage pin images to a new scene;
and 3, randomly adjusting the brightness of the image by using the images obtained by the method 1 and the method 2, and carrying out self-adaptive histogram equalization conceptualization and noise adding operation.
4. The method for detecting loss of round pin and round pin cotter for damper adjustment screw according to claim 3, wherein S24 further comprises screening the amplified instance of the data segmented with tag sample data set:
if the confidence of the prediction frame of the segmentation target of a certain sample in the sample data set segmented with the label in the example is less than 0.5, deleting the sample;
if the confidence of the prediction frame of the segmentation target of a certain sample in the sample data set of the sample segmentation label is not less than 0.5 and the true value file is not available, adding the label information;
if the IOU of the division target prediction frame and the truth frame of a certain sample in the sample data set divided by the label is larger than a set threshold value, but the categories are inconsistent, the label information is corrected;
if a certain sample in the sample data set with the label is segmented, the target is not segmented, but the label information exists, whether the label information is reasonable or not is judged, and if not, the sample is deleted.
5. The method for detecting loss of round pin and round pin cotter of damper adjustment screw according to claim 4, characterized in that, in the training process, the 1 st to 18 th round of use amplified examples are segmented with the label sample data set, the 19 th to 20 th round of use amplified examples are segmented with the label sample data set.
6. The method for detecting the loss of the round pin and the round pin cotter of the brake adjuster adjusting screw according to claim 1, wherein in the step S21, a passing image is acquired, images of different components, different textures and different dimensions of a railway wagon are collected to form a pre-training set according to wheelbase information and prior information of the components, and the different components comprise springs, bolts, nuts, chains, cylinders, pull rods, pipe bodies, round pins and cotters; the different scales include 1024×1024, 1024×512, 512×512, 256×512, 512×256, 256×256.
7. The method for detecting the loss of the round pin and the round pin cotter of the brake adjuster adjusting screw according to claim 1 is characterized in that a characteristic pyramid of a Mask R-CNN network model is a PAN-FPN network, resNet50 comprises 4 groups of residual units which are sequentially connected, the intercepted image is sent to the 1 st group of residual units, and a C2 characteristic graph, a C3 characteristic graph and a C4 characteristic graph which are output by the last 3 groups of residual units are input to the PAN-FPN network;
the PAN-FPN network comprises a No. 1 up-sampling module, a No. 1 connecting unit C, a No. 1C 4f unit, a No. 2 up-sampling unit, a No. 2 connecting unit C, a No. 2C 4f unit, a No. 1 CBS unit, a No. 3 connecting unit C, a No. 3C 4f unit, a No. 2 CBS unit, a No. 4 connecting unit C and a No. 4C 4f unit;
the C4 feature map is simultaneously input to an up-sampling unit No. 1 and a connecting unit No. 4; the output of the No. 1 up-sampling unit and the C3 feature map are input to a No. 1 connecting unit for connection, the output of the No. 1 connecting unit is input to a No. 1C 4f unit, the output of the No. 1C 4f unit is input to a No. 2 up-sampling unit and a No. 3 connecting unit simultaneously, the output of the No. 2 up-sampling unit and the C2 feature map are input to a No. 2 connecting unit for connection, the output of the No. 2 connecting unit is input to a No. 2C 4f unit, the C4f unit outputs the feature P2, the feature P2 is input to a No. 1 CBS unit, the output of the No. 1 CBS unit is input to a No. 3 connecting unit, the output of the No. 3 CBS unit is input to a No. 4 connecting unit, the output of the No. 4 connecting unit is input to a No. 4C 4f unit, and the C4f unit outputs the feature P4;
the C4f unit divides an input characteristic diagram into 4 channels according to the channels, the 1 st channel is subjected to a Bottleneck operation, the 2 nd channel is respectively subjected to 1 Bottleneck operation and 2 Bottleneck operation, the 3 rd channel is respectively subjected to 1 Bottleneck operation, 2 Bottleneck operation and 3 Bottleneck operation, all channels after the Bottleneck operation are connected with the 4 th channel, and then the output of the C4f unit is obtained through a CBS operation;
the CBS units are Conv-BN-SiLU combination operations.
8. The method for detecting loss of round pin and round pin cotter of a brake adjuster adjusting screw according to claim 1, wherein an optimizer of a Mask R-CNN network model is AdamW, adamW is obtained by adding a regular to an AdamW optimizer, wherein an initial learning rate is 0.001, a weight attenuation coefficient is 0.005, training of a learning rate is updated by a segmented Cosine LR method, a fixed learning rate is used in the first 60%, and a Cosine annealing mode is used in the last 40%.
9. A computer-readable storage device storing a computer program, wherein the computer program when executed implements the gatherer adjusting screw round pin and round pin cotter loss detection method according to any one of claims 1 to 8.
10. A brake lever round pin and round pin cotter loss detection apparatus comprising a storage device, a processor and a computer program stored in the storage device and executable on the processor, wherein execution of the computer program by the processor implements the brake adjustment screw round pin and round pin cotter loss detection method according to any one of claims 1 to 8.
CN202310299281.4A 2023-03-24 2023-03-24 Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster Pending CN116310649A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310299281.4A CN116310649A (en) 2023-03-24 2023-03-24 Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310299281.4A CN116310649A (en) 2023-03-24 2023-03-24 Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster

Publications (1)

Publication Number Publication Date
CN116310649A true CN116310649A (en) 2023-06-23

Family

ID=86792435

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310299281.4A Pending CN116310649A (en) 2023-03-24 2023-03-24 Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster

Country Status (1)

Country Link
CN (1) CN116310649A (en)

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111339882A (en) * 2020-02-19 2020-06-26 山东大学 Power transmission line hidden danger detection method based on example segmentation
US10713794B1 (en) * 2017-03-16 2020-07-14 Facebook, Inc. Method and system for using machine-learning for object instance segmentation
CN114821278A (en) * 2022-04-22 2022-07-29 南通大学 Power transmission line part identification method based on improved YOLOv5
US20220309674A1 (en) * 2021-03-26 2022-09-29 Nanjing University Of Posts And Telecommunications Medical image segmentation method based on u-net
CN115359366A (en) * 2022-08-19 2022-11-18 中国人民解放军军事科学院系统工程研究院 Remote sensing image target detection method based on parameter optimization
CN115526874A (en) * 2022-10-08 2022-12-27 哈尔滨市科佳通用机电股份有限公司 Round pin of brake adjuster control rod and round pin split pin loss detection method
CN115546664A (en) * 2022-09-30 2022-12-30 湖北省电力勘测设计院有限公司 Cascaded network-based insulator self-explosion detection method and system
CN115731164A (en) * 2022-09-14 2023-03-03 常州大学 Insulator defect detection method based on improved YOLOv7
CN115829999A (en) * 2022-12-22 2023-03-21 国网新疆电力有限公司信息通信公司 Insulator defect detection model generation method, device, equipment and storage medium

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10713794B1 (en) * 2017-03-16 2020-07-14 Facebook, Inc. Method and system for using machine-learning for object instance segmentation
CN111339882A (en) * 2020-02-19 2020-06-26 山东大学 Power transmission line hidden danger detection method based on example segmentation
US20220309674A1 (en) * 2021-03-26 2022-09-29 Nanjing University Of Posts And Telecommunications Medical image segmentation method based on u-net
CN114821278A (en) * 2022-04-22 2022-07-29 南通大学 Power transmission line part identification method based on improved YOLOv5
CN115359366A (en) * 2022-08-19 2022-11-18 中国人民解放军军事科学院系统工程研究院 Remote sensing image target detection method based on parameter optimization
CN115731164A (en) * 2022-09-14 2023-03-03 常州大学 Insulator defect detection method based on improved YOLOv7
CN115546664A (en) * 2022-09-30 2022-12-30 湖北省电力勘测设计院有限公司 Cascaded network-based insulator self-explosion detection method and system
CN115526874A (en) * 2022-10-08 2022-12-27 哈尔滨市科佳通用机电股份有限公司 Round pin of brake adjuster control rod and round pin split pin loss detection method
CN115829999A (en) * 2022-12-22 2023-03-21 国网新疆电力有限公司信息通信公司 Insulator defect detection model generation method, device, equipment and storage medium

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
小书童: "YOLO系列又双叒更新！详细解读YOLOv8的改进模块", pages 1 - 12, Retrieved from the Internet <URL:https://zhuanlan.zhihu.com/p/598189585> *
王昕钰 等: "基于三级级联架构的接触网定位管开口销缺陷检测", 仪器仪表学报, vol. 40, no. 10, 15 October 2019 (2019-10-15), pages 1 - 10 *

Similar Documents

Publication Publication Date Title
WO2020244261A1 (en) Scene recognition system for high-resolution remote sensing image, and model generation method
CN111583229B (en) Road surface fault detection method based on convolutional neural network
CN110210608B (en) Low-illumination image enhancement method based on attention mechanism and multi-level feature fusion
CN109559302A (en) Pipe video defect inspection method based on convolutional neural networks
CN112232351B (en) License plate recognition system based on deep neural network
CN112800838A (en) Channel ship detection and identification method based on deep learning
CN108734717B (en) Single-frame star map background dark and weak target extraction method based on deep learning
CN111008608B (en) Night vehicle detection method based on deep learning
CN111242955B (en) Road surface crack image segmentation method based on full convolution neural network
CN115526874B (en) Method for detecting loss of round pin and round pin cotter pin of brake adjuster control rod
CN115035418A (en) Remote sensing image semantic segmentation method and system based on improved deep LabV3+ network
CN112215907A (en) Automatic extraction method for weld defects
CN115393698A (en) Digital image tampering detection method based on improved DPN network
CN115019340A (en) Night pedestrian detection algorithm based on deep learning
CN114724052A (en) Electric power image skyline segmentation method based on deep learning network model
CN113033371A (en) CSP model-based multi-level feature fusion pedestrian detection method
CN116310649A (en) Method for detecting loss of round pin and round pin cotter of adjusting screw of brake adjuster
CN116309545A (en) Single-stage cell nucleus instance segmentation method for medical microscopic image
CN113205078B (en) Crowd counting method based on multi-branch progressive attention-strengthening
CN113239865B (en) Deep learning-based lane line detection method
CN112991257B (en) Heterogeneous remote sensing image change rapid detection method based on semi-supervised twin network
CN117391177B (en) Construction method and application of driver behavior detection model
CN116778346B (en) Pipeline identification method and system based on improved self-attention mechanism
CN110472528B (en) Subway environment target training set generation method and system
CN114627005B (en) Rain density classification guided double-stage single image rain removing method

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