CN108960320B - Real-time detection method for fault image of angle cock of train - Google Patents
Real-time detection method for fault image of angle cock of train Download PDFInfo
- Publication number
- CN108960320B CN108960320B CN201810706143.2A CN201810706143A CN108960320B CN 108960320 B CN108960320 B CN 108960320B CN 201810706143 A CN201810706143 A CN 201810706143A CN 108960320 B CN108960320 B CN 108960320B
- Authority
- CN
- China
- Prior art keywords
- image
- fault
- train
- angle cock
- normalization processing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Train Traffic Observation, Control, And Security (AREA)
Abstract
The invention relates to the field of image processing and fault identification, and discloses a real-time detection method for fault images of a folding angle cock of a train, which comprises the following steps: and using an image manufacturing database to extract multi-scale combined features, carrying out normalization processing, further training a cascade detector, then collecting an image to be detected of the angle cock, extracting the multi-scale combined features, carrying out normalization processing, then sending the features into the trained cascade detector, positioning a target area, calculating confidence coefficient, comparing the confidence coefficient with a set threshold value, and detecting whether the angle cock is in fault. The method for detecting the fault image of the angle cock of the train in real time has high detection efficiency and accuracy, strong real-time performance and can find the fault of the angle cock in time.
Description
Technical Field
The invention relates to the field of image processing and fault identification, in particular to a real-time detection method for a fault image of a folding angle cock of a train.
Background
In recent years, the technical inspection (train inspection for short) of trains in China mainly depends on the manual inspection mode of train inspectors. The mode is influenced by human subjective factors such as eyesight and fatigue degree and external environments such as climate and illumination, so that the detection efficiency and accuracy are influenced by different degrees. High-speed rails and electric power businesses in China are developed vigorously nowadays, the logistics and transportation industry changes with the place covered, railway transportation is used as an important component of the logistics and transportation industry and is in the position of a medium current pillar in national economy, and higher requirements are put forward for the following railway transportation safety problems. Then, TFDS (train operation failure moving picture Detection System) is gradually applied to a portion of the railway line in China. The system applies computer, network communication, automatic control and image acquisition processing technology and introduces scientific management method and systematized development method, provides dynamic collection, storage, transmission and early warning service of fault picture information for railway train operation fault detection, improves train inspection operation quality and efficiency and vehicle safety precaution level, and strengthens man-machine system for fault basic information collection and management in train application. The system can automatically capture all pictures of the bottom and the lower part of the train, mainly comprises devices such as train coupler buffering accessories, an underframe, a bogie, the lower part of the side of the train body, a folding angle cock and the like, analyzes the captured contents in a man-machine combination mode, and can judge whether the devices of the train have faults such as defects, breakage, loss and the like, thereby realizing the crossing from manual detection to man-machine combination detection.
The angle cock is a key part of a train air braking system, and a train transmits compressed air to each carriage through the angle cock on a main pipeline and performs braking by using the compressed air. Only when the angle cock is opened, the compressed air can be transmitted to each carriage, and the smoothness of the main pipe of the train is ensured. This can cause a significant accident if the angle cock is closed during the running of the train. Although many researchers in China carry out many relevant researches on the detection of the TFDS fault, the folding angle cock is too small in area relative to a train and difficult to locate, and the complexity of the background of the folding angle cock also increases the difficulty of the fault detection of the folding angle cock to a certain extent. In recent years, related scholars propose different algorithms, such as Suzhou university, the shape representation and matching algorithm based on target contour is proposed for detecting and identifying train bogie faults, Qina of southwest university of transportation aims at the problems of train bogie fault signal feature extraction, key component performance degradation estimation, multi-feature fusion and dimension reduction, the bogie fault signal feature extraction and analysis framework is proposed, and a data driving mode is used for solving the problems of bogie fault diagnosis and component form estimation, so that a research idea is provided.
Disclosure of Invention
The invention aims to overcome the defects of the technology, provides a real-time detection method for a fault image of a train angle cock, has high detection efficiency and accuracy and strong real-time performance, and can find the fault of the angle cock in time.
In order to achieve the purpose, the invention discloses a method for detecting fault images of a folding angle cock of a train in real time, which comprises the following steps:
A) preprocessing an original image acquired by a dynamic image detection system of a railway train running fault, eliminating the influence of a shooting environment on the original image, labeling the denoised image, and making an image database with labels;
B) extracting image features from the image database manufactured in the step A), extracting multi-scale combined features of the image by using a rapid feature pyramid, and performing feature normalization processing;
C) training a cascade detector based on SVM-Adaboost by using the image database after the characteristic normalization processing in the step B);
D) acquiring an image to be detected of the angle cock by a railway train operation fault dynamic image detection system, sliding on each layer of an image pyramid to be detected by using a sliding window in different step lengths to generate a series of small windows to be detected, extracting multi-scale combined features of each small window to be detected by using a rapid feature pyramid, and performing normalization processing;
E) sending the features after normalization processing in the step D) to the cascade detector trained in the step C), locating a target area and calculating a confidence coefficient;
F) comparing the confidence coefficient calculated in the step E) with a set threshold, if the confidence coefficient is larger than the threshold, reserving the target area positioning frame positioned in the step E), and if the folded angle cock is not in fault, deleting the target area positioning frame, and enabling the folded angle cock to be in fault.
Preferably, in the step a), the original image is a dynamic image of a brake device passing through the bottom of the train, which is shot by an outdoor image acquisition device of the train fault rail edge image detection system in real time, and is transmitted through an optical fiber network.
Preferably, in the step a), the image target area of the positive sample is labeled, and then the positive sample and the negative sample are respectively made into a folding image database with labels, where the positive sample is a non-failure image, the negative sample is a failure image, and the folding image database is made by separately storing the failure image and the non-failure image.
Preferably, in the step B), the multi-scale joint feature includes 8 channels including 1 gradient amplitude, 6 directional gradient histograms and 1 rotation invariant LBP, and the step of extracting the image multi-scale joint feature by using the fast feature pyramid and performing the feature normalization process includes:
a) accurately calculating the combined features of one scale in every eight scales by using a quick feature pyramid;
b) calculating the characteristics of the images of other scales in the eight scales by using the characteristics;
c) and carrying out normalization processing on the combined features extracted by the eight scales.
Preferably, in the step C), the cascade detector is formed by collecting a plurality of strong classifiers, the strong classifiers are formed by a group of SVM weak classifiers, when the cascade detector is trained, the inner loop trains each strong classifier by using an AdaBoost iterative algorithm, and the outer loop trains the cascade detector formed by all the strong classifiers.
Preferably, in the step F), the confidence threshold is 0.1-0.4.
Compared with the prior art, the invention has the following advantages: the cascade detector is trained by establishing the image database, extracting the multi-scale joint features and carrying out normalization processing, and then the multi-scale joint features after the normalization processing of the image to be detected are detected by the cascade detector, so that whether the folding angle cock is in failure or not can be detected in time, and the detection efficiency and the accuracy are high.
Drawings
FIG. 1 is a schematic flow chart of a method for detecting fault images of a folding angle cock of a train in real time according to the invention;
FIG. 2 is a schematic view of a non-fault image of a dog-ear door according to the present invention;
FIG. 3 is a schematic view of a fault image of a folding angle plug door according to the present invention;
FIG. 4 is a schematic view of a multi-scale combined feature extraction process of the angle cock of the present invention;
FIG. 5 is a schematic diagram of a detection flow based on an SVM-Adaboost cascade detector in the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments.
A method for detecting fault images of a folding angle cock of a train in real time is shown in figure 1 and comprises the following steps:
A) preprocessing an original image acquired by a railway train running fault dynamic image detection system, eliminating the influence of a shooting environment on the original image, wherein the original image is a dynamic image of a brake device passing through the bottom of a train and shot in real time by an outdoor image acquisition device of the train fault rail edge image detection system, and is transmitted through an optical fiber network, wherein an image target area of a positive sample is labeled, positive and negative samples are respectively manufactured into a folding cock image database with labels, the positive sample is a non-fault image, as shown in fig. 2, the negative sample is a fault image, as shown in fig. 3, the fault image and the non-fault image are separately stored by manufacturing the folding cock image database;
B) extracting image features from the image database manufactured in the step a), as shown in fig. 4, extracting multi-scale combined features of the image by using a fast feature pyramid, and performing feature normalization processing, in this embodiment, the multi-scale combined features include 8 channels including 1 gradient amplitude, 6 directional gradient histograms, and 1 rotation invariant LBP, the extracting image multi-scale combined features by using the fast feature pyramid, and the performing feature normalization processing includes:
a) accurately calculating the combined features (reference features) of one scale in every eight scales by using a quick feature pyramid;
b) calculating the characteristics of the images of other scales in the eight scales by using the characteristics;
c) and carrying out normalization processing on the combined features extracted by the eight scales.
Specifically, in this embodiment, since the acquired image of the folding angle cock is a gray scale image, the image is selected with the features of 8 channels including 1 gradient amplitude, 6 directional gradient histograms and 1 rotation invariant LBP, and the fast feature pyramid can greatly accelerate the feature extraction speed without losing key information. The fast feature pyramid is implemented as follows:
Isrepresenting the image I at the scale s, R (I, s) representing the image I to be s-resampled, and defining a function omega with translation invariance to extract s-resampled channel characteristics CsΩ (R (I, s)), and C is usedsPerforming calculation layer by layer, and calculating C for increasing speed and simplifying calculationsThe following approximate calculation can be performed:
where C ═ Ω (I) is a channel characteristic of the image I, and λΩIs a characteristic parameter, λΩCan be obtained by fitting according to the specific channel characteristic distribution.
And one feature pyramid is used for representing the image I in multiple scales and extracting the image features corresponding to each scale. The scale s is sampled at equal intervals in a log space from 1, generally, 4-12 scales are selected from an eight-scale (octave), and each eight-scale interval is half of the previous eight-scale interval. Establishing a quick characteristic pyramid, and firstly calculating the characteristic C of each eight-scale images′=Ω(R(I,s′)),Then calculating the characteristics of the image between eight scales by using a formulaBecome intoWhere s is the closest dimension to s',the fast feature pyramid only calculates the features of one scale in each eight scale, and then uses the features to calculate the features of images of other scales in the eight scales.
After 8 scale features of an image are extracted, the dimension of the features of 8 scales are inconsistent, which may cause that the training of a detector later cannot converge to a global optimal solution, so that the features of each scale are normalized by the following calculation method:
wherein |2Represents a 2-norm; epsilon is a small constant.
C) Training an SVM-Adaboost-based cascade detector by utilizing the image database after the characteristic normalization processing in the step B), wherein the cascade detector is formed by integrating a plurality of strong classifiers, each strong classifier is formed by a group of SVM weak classifiers, an AdaBoost iterative algorithm is used for training each strong classifier in an inner loop during the training of the cascade detector, and an outer loop is used for training the cascade detector formed by all the strong classifiers, and in the embodiment, the training process is as follows:
step 2, internal circulation: training a strong classifier stage by using an AdaBoost algorithm, readjusting the added threshold of the next SVM weak classifier when adding one SVM weak classifier in the cycle, and ensuring that the detection rate of the strong classifier stage is not less than dminAnd re-evaluating the strong classifier to determine the false positive rate f of the strong classifieriWhether or not f is lower thanmaxIf so, finishing the training of the strong classifier at the stage, otherwise, continuously adding the SVM weak classifier until the termination condition of the internal circulation is met;
step 3, external circulation: judging the whole false positive rate F of the current targetiWhether the overall false positive rate F is lower than the initialization targettargetIf so, terminating the training, otherwise, readjusting the negative samples and putting all the samples with detection errors into the database N, and then performing internal circulation until the current target overall false positive rate FiLower than the overall false positive rate F of the initialization targettargetUntil the end;
D) acquiring an image to be detected of the angle cock by a railway train operation fault dynamic image detection system, sliding on each layer of an image pyramid to be detected by using a sliding window in different step lengths, generating a series of small windows to be detected, referring to step B), extracting multi-scale joint features of each small window to be detected by using a rapid feature pyramid, and carrying out normalization processing;
E) sending the features normalized in step D) to the cascade detector trained in step C), as shown in FIG. 5, locating the target area and calculating the confidence level;
F) and E) comparing the confidence coefficient calculated in the step E) with a set threshold, wherein the confidence coefficient threshold is 0.1-0.4, in the embodiment, the threshold is 0.25, if the confidence coefficient is larger than the threshold, the target area positioning frame positioned in the step E) is reserved, the folded angle cock is not in fault, otherwise, the target area positioning frame is deleted, and the folded angle cock is in fault.
Claims (6)
1. A method for detecting fault images of a folding angle cock of a train in real time is characterized by comprising the following steps: the method comprises the following steps:
A) preprocessing an original image acquired by a dynamic image detection system of a railway train running fault, eliminating the influence of a shooting environment on the original image, labeling the denoised image, and making an image database with labels;
B) extracting image features from the image database manufactured in the step A), extracting multi-scale combined features of the image by using a rapid feature pyramid, and performing feature normalization processing;
C) training a cascade detector based on SVM-Adaboost by using the image database after the characteristic normalization processing in the step B);
D) acquiring an image to be detected of the angle cock by a railway train operation fault dynamic image detection system, sliding on each layer of an image pyramid to be detected by using a sliding window in different step lengths to generate a series of small windows to be detected, extracting multi-scale combined features of each small window to be detected by using a rapid feature pyramid, and performing normalization processing;
E) sending the features after normalization processing in the step D) to the cascade detector trained in the step C), locating a target area and calculating a confidence coefficient;
F) comparing the confidence coefficient calculated in the step E) with a set threshold, if the confidence coefficient is larger than the threshold, reserving the target area positioning frame positioned in the step E), and if the folded angle cock is not in fault, deleting the target area positioning frame, and enabling the folded angle cock to be in fault.
2. The method for detecting the fault image of the angle cock of the train in real time according to claim 1, wherein the method comprises the following steps: in the step A), the original image is a dynamic image of a brake device passing through the bottom of the train and shot in real time by an outdoor image acquisition device of the train fault rail edge image detection system, and the dynamic image is transmitted through an optical fiber network.
3. The method for detecting the fault image of the angle cock of the train in real time according to claim 1, wherein the method comprises the following steps: in the step A), the image target area of the positive sample is labeled, then the positive sample and the negative sample are respectively made into a folding cock image database with labels, the positive sample is a non-fault image, the negative sample is a fault image, and the fault image and the non-fault image are separately stored by making the folding cock image database.
4. The method for detecting the fault image of the angle cock of the train in real time according to claim 1, wherein the method comprises the following steps: in the step B), the multi-scale combined features include 8 channels including 1 gradient amplitude, 6 directional gradient histograms and 1 rotation invariant LBP, and the step of extracting the image multi-scale combined features by using the fast feature pyramid and performing the feature normalization processing includes:
a) accurately calculating the combined features of one scale in every eight scales by using a quick feature pyramid;
b) calculating the characteristics of the images of other scales in the eight scales by using the characteristics;
c) and carrying out normalization processing on the combined features extracted by the eight scales.
5. The method for detecting the fault image of the angle cock of the train in real time according to claim 1, wherein the method comprises the following steps: in the step C), the cascade detector is formed by collecting a plurality of strong classifiers, the strong classifiers are formed by a group of SVM weak classifiers, when the cascade detector is trained, the internal loop trains each strong classifier by using an AdaBoost iterative algorithm, and the external loop trains the cascade detector formed by all the strong classifiers.
6. The method for detecting the fault image of the angle cock of the train in real time according to claim 1, wherein the method comprises the following steps: in the step F), the confidence threshold is 0.1-0.4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810706143.2A CN108960320B (en) | 2018-06-26 | 2018-06-26 | Real-time detection method for fault image of angle cock of train |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810706143.2A CN108960320B (en) | 2018-06-26 | 2018-06-26 | Real-time detection method for fault image of angle cock of train |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108960320A CN108960320A (en) | 2018-12-07 |
CN108960320B true CN108960320B (en) | 2021-04-27 |
Family
ID=64484846
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810706143.2A Active CN108960320B (en) | 2018-06-26 | 2018-06-26 | Real-time detection method for fault image of angle cock of train |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108960320B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112733706B (en) * | 2021-01-07 | 2022-09-16 | 河南理工大学 | Motor fault diagnosis method based on bilinear LBP, electronic equipment and medium |
CN113830136B (en) * | 2021-10-20 | 2022-04-19 | 哈尔滨市科佳通用机电股份有限公司 | Method for identifying malposition fault of angle cock handle of railway wagon |
CN115170828B (en) * | 2022-07-15 | 2023-03-14 | 哈尔滨市科佳通用机电股份有限公司 | Angle cock clip loss fault detection method based on deep learning |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105260744A (en) * | 2015-10-08 | 2016-01-20 | 北京航空航天大学 | Automatic on-line diagnosis method for freight train coupler tail cotter position faults and system |
CN105260749A (en) * | 2015-11-02 | 2016-01-20 | 中国电子科技集团公司第二十八研究所 | Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM |
CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
CN106485694A (en) * | 2016-09-11 | 2017-03-08 | 西南交通大学 | A kind of high ferro contact net double-jacket tube connector six-sided nut based on cascade classifier comes off defective mode detection method |
CN106778833A (en) * | 2016-11-28 | 2017-05-31 | 北京航天控制仪器研究所 | Small object loses the automatic identifying method of failure under a kind of complex background |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CA2979201A1 (en) * | 2015-03-11 | 2016-09-15 | Siemens Industry, Inc. | Prediction in building automation |
-
2018
- 2018-06-26 CN CN201810706143.2A patent/CN108960320B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105260744A (en) * | 2015-10-08 | 2016-01-20 | 北京航空航天大学 | Automatic on-line diagnosis method for freight train coupler tail cotter position faults and system |
CN105260749A (en) * | 2015-11-02 | 2016-01-20 | 中国电子科技集团公司第二十八研究所 | Real-time target detection method based on oriented gradient two-value mode and soft cascade SVM |
CN105447511A (en) * | 2015-11-15 | 2016-03-30 | 安徽清新互联信息科技有限公司 | SVM object detection method based on Adaboost Haar-Like features |
CN106485694A (en) * | 2016-09-11 | 2017-03-08 | 西南交通大学 | A kind of high ferro contact net double-jacket tube connector six-sided nut based on cascade classifier comes off defective mode detection method |
CN106778833A (en) * | 2016-11-28 | 2017-05-31 | 北京航天控制仪器研究所 | Small object loses the automatic identifying method of failure under a kind of complex background |
Non-Patent Citations (2)
Title |
---|
《Automated visual inspection of angle cocks during train operation》;Zhou Fuqiang, et al;《Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit 》;20130709;794-806 * |
《基于多特征融合和BP-AdaBoost算法的列车关键零件故障自动识别》;孙国栋 等31;《中国机械工程》;20171130;2587-2594 * |
Also Published As
Publication number | Publication date |
---|---|
CN108960320A (en) | 2018-12-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108960320B (en) | Real-time detection method for fault image of angle cock of train | |
CN106226050B (en) | A kind of TFDS fault picture automatic identifying method based on convolutional neural networks | |
CN105260744B (en) | The automatic on-line diagnostic method and system of a kind of goods train coupler yoke key position failure | |
WO2019127273A1 (en) | Multi-person face detection method, apparatus, server, system, and storage medium | |
CN105424330B (en) | A kind of lorry bolts at axle end fault detection method and device | |
CN108898137A (en) | A kind of natural image character identifying method and system based on deep neural network | |
WO2021238826A1 (en) | Method and apparatus for training instance segmentation model, and instance segmentation method | |
CN105227907B (en) | Unsupervised anomalous event real-time detection method based on video | |
CN110264440B (en) | Large-scale train displacement fault detection method and system based on deep learning | |
CN105701448B (en) | Three-dimensional face point cloud nose detection method and the data processing equipment for applying it | |
CN103745224A (en) | Image-based railway contact net bird-nest abnormal condition detection method | |
CN103295027A (en) | Freight wagon blocking key missing fault identification method based on support vector machine | |
CN113947731B (en) | Foreign matter identification method and system based on contact net safety inspection | |
CN111080612B (en) | Truck bearing damage detection method | |
CN103778435A (en) | Pedestrian fast detection method based on videos | |
CN112215822A (en) | Face image quality evaluation method based on lightweight regression network | |
CN108009472A (en) | A kind of finger back arthrosis line recognition methods based on convolutional neural networks and Bayes classifier | |
CN104615986A (en) | Method for utilizing multiple detectors to conduct pedestrian detection on video images of scene change | |
CN110858276A (en) | Pedestrian re-identification method combining identification model and verification model | |
CN109858456A (en) | A kind of rolling stock status fault analysis system | |
CN115272830A (en) | Pantograph foreign matter detection method based on deep learning | |
CN115424128A (en) | Fault image detection method and system for lower link of freight car bogie | |
CN113723188A (en) | Dress uniform person identity verification method combining face and gait features | |
CN115631488A (en) | Jetson Nano-based fruit maturity nondestructive testing method and system | |
CN111325076A (en) | Aviation ground building extraction method based on U-net and Seg-net network fusion |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |