CN112907534A

CN112907534A - Fault detection method and device based on door closing part position image

Info

Publication number: CN112907534A
Application number: CN202110187677.0A
Authority: CN
Inventors: 金佳鑫
Original assignee: Harbin Kejia General Mechanical and Electrical Co Ltd
Current assignee: Harbin Kejia General Mechanical and Electrical Co Ltd
Priority date: 2021-02-18
Filing date: 2021-02-18
Publication date: 2021-06-04

Abstract

The invention discloses a fault detection method and a fault detection device based on a door closing part position image, relates to a fault detection method and a fault detection device of a door closing vehicle in a train, and aims to solve the problems that the existing door closing fault detection mode needs manual detection and is easy to generate fault and omission, and the method comprises the following steps: step 1, obtaining an image to be detected; the image to be detected comprises a door closing part position image of a real-time passing vehicle; step 2, calculating characteristic points of the image to be detected; step 3, clustering the characteristic points of the image to be detected, and extracting the characteristic vector of the image to be detected; step 4, inputting the feature vector of the image to be detected into a Support Vector Machine (SVM) classifier, and obtaining the classification result of the image to be detected by introducing a finally trained SVM feature model: and the SVM classifier classifies the images to be detected through the SVM feature model.

Description

Fault detection method and device based on door closing part position image

Technical Field

The invention relates to a method and a device for detecting a door closing fault in a train.

Background

The door closing failure means that a cut-off cock handle at the bottom of the car body is in a closed state. In a railway train, a vehicle which must stop a brake action due to a loaded cargo is called a "door-closing vehicle", and a vehicle in which a brake temporarily fails is permitted to close a shutoff valve to stop the brake action of the vehicle, and simultaneously pulls a release valve to discharge air in an auxiliary reservoir so as not to be braked again due to air leakage from a branch pipe. The cut-off cock has to be opened when braking, and when the brake breaks down, the cut-off cock is closed in order to cut off the compressed air supply passage of the brake main pipe. If the cock handle is closed, the brake pipeline is not communicated with the main pipeline, and at the moment, the truck loses the brake capability, cannot play a role in braking, and is easy to cause serious safety faults.

The train door closing brake cars are strictly regulated, the door closing cars cannot be woven in the rear three cars of the locomotive according to the regulation, the number of the door closing cars in the train cannot exceed 6 percent of the number of the current cars because the number of the door closing cars cannot exceed two continuous woven cars, the tail cannot be closed, the second and third cars at the tail cannot be closed continuously, and therefore detection of the failure of the door closing cars is a very important link for ensuring the running safety of the railway freight car.

At present, each railway bureau still adopts a traditional truck fault detection mode, needs to manually check the passing images of each train one by one, and marks and uploads a vehicle inspection platform when a fault is found.

Disclosure of Invention

The invention aims to solve the problems that the existing door closing failure detection mode needs manual detection and failure and detection omission easily occur, and provides a failure detection method and device based on a door closing component position image.

The invention discloses a fault detection method based on a door closing vehicle part position image, which comprises the following specific steps:

step 1, obtaining an image to be detected; the image to be detected comprises a door closing part position image of a real-time passing vehicle;

the category of the door closing part position image includes a truncated door part image, a door handle image, and other images; the other images include images other than a truncated stopper part image and a stopper handle image;

step 2, calculating characteristic points of the image to be detected;

step 3, clustering the characteristic points of the image to be detected, and extracting the characteristic vector of the image to be detected;

step 4, inputting the feature vector of the image to be detected into a Support Vector Machine (SVM) classifier, and obtaining the classification result of the image to be detected by introducing a finally trained SVM feature model: the SVM classifier classifies the images to be detected through an SVM feature model;

if the type of the image to be detected is an intercepted door closing part image or other images, judging that the door closing vehicle corresponding to the image to be detected has no fault;

and if the type of the image to be detected is the door handle image, judging that the door closing vehicle corresponding to the image to be detected has a fault.

The step of obtaining the support vector machine SVM feature model trained in the step 4 is as follows:

step 41, acquiring pre-acquired door closing component position images and category labels corresponding to the door closing component position images, and forming a sample training set;

step 42, acquiring a training set characteristic vector set corresponding to each door closing vehicle component position image in a sample training set;

and 43, training the SVM feature model through the training set feature vector set and the class label corresponding to the door closing vehicle component position image to obtain the finally trained SVM feature model.

Wherein, step 42 is specifically as follows:

step 421, calculating characteristic points of all door closing component position images in the sample training set, and combining the characteristic points of all door closing component position images into a training set characteristic point set;

step 422, classifying all feature points in the feature point set of the training set into K categories by adopting a K-means clustering algorithm, and obtaining corresponding K clustering centers; k is a positive integer;

and 423, classifying the feature points of each door closing component position image in the training set according to K clustering centers to obtain a feature vector of each door closing component position image, and combining the feature vectors of each door closing component position image into a training set feature vector set.

Wherein, step 422 is as follows;

4221, finding K feature points from a training set feature point set to serve as K clustering centers, wherein the K clustering centers correspond to K categories;

4222, calculating distances between K clustering centers in the training set feature point set and other feature points respectively, and dividing each other feature point into a category in which the clustering center with the minimum distance to the other feature points is located; the other characteristic points are characteristic points except the clustering center;

4223, calculating the mean value of all the feature points in each category, and taking the mean value as a new clustering center;

and step 4224, replacing the original clustering center with the new clustering center, returning to execute the step 4222 and the step 4223 until the clustering center is not changed any more, and outputting the final K clustering centers and the corresponding categories.

Wherein, step 41 further comprises: acquiring a pre-acquired door closing vehicle component position image and a category label corresponding to the door closing vehicle component position image, and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

step 43 is specifically as follows:

431, creating a Support Vector Machine (SVM) classifier, and setting training parameters of an SVM feature model;

step 432, inputting the training set feature vector set and the class labels corresponding to the door closing vehicle component position images into a Support Vector Machine (SVM) classifier, and training a Support Vector Machine (SVM) feature model to obtain a preliminarily trained SVM feature model;

step 433, testing the preliminarily trained SVM feature model by using a test set;

if the classification accuracy does not reach 100%, go back to step 431 and reset the training parameters,

and if the classification accuracy reaches 100%, finishing the training of the SVM feature model to obtain the finally trained SVM feature model.

The invention relates to a fault detection device based on a door closing vehicle part position image, which comprises:

the image acquisition module to be detected is used for acquiring an image to be detected; the image to be detected comprises a door closing part position image of a real-time passing vehicle;

the image feature point calculation module to be detected is connected with the image acquisition module to be detected and is used for calculating the feature points of the image to be detected;

the image feature vector extraction module to be detected is connected with the image feature point calculation module to be detected and is used for clustering the feature points of the image to be detected and extracting the feature vector of the image to be detected;

the classification judgment module of the images to be detected is connected with the clustering module of the characteristic points of the images to be detected and is used for inputting the characteristic vectors of the images to be detected into the SVM classifier, and the classification result of the images to be detected is obtained by leading in the SVM characteristic model of the final training: the SVM classifier classifies the images to be detected through an SVM feature model;

The device for acquiring the trained SVM (support vector machine) feature model in the classification judgment module of the image to be detected comprises the following steps:

the system comprises a sample building module, a sample training set and a data processing module, wherein the sample building module is used for obtaining a pre-collected door closing vehicle component position image and a category label corresponding to the door closing vehicle component position image and forming a sample training set;

the training set feature vector set acquisition module is connected with the sample construction module and is used for acquiring a training set feature vector set corresponding to each door closing vehicle component position image in the sample training set;

and the support vector machine SVM feature model training module is connected with the training set feature vector set calculating module and used for training the support vector machine SVM feature model through the training set feature vector set and the category label corresponding to the door closing vehicle component position image to obtain a finally trained support vector machine SVM feature model.

Wherein, the training set feature vector set calculation module comprises:

the training set characteristic point calculation module is used for calculating characteristic points of all door closing component position images in the sample training set and combining the characteristic points of all door closing component position images into a training set characteristic point set;

the training set characteristic point clustering module is connected with the training set characteristic point calculating module and used for classifying all the characteristic points in the training set characteristic point set into K categories by adopting a K-means clustering algorithm and obtaining corresponding K clustering centers; k is a positive integer;

and the feature vector calculation module is connected with the training set feature point clustering module and is used for classifying the feature points of each door closing vehicle component position image in the training set according to K clustering centers to obtain the feature vector of each door closing vehicle component position image and combining the feature vectors of each door closing vehicle component position image into a training set feature vector set.

Wherein, training set feature point clustering module includes:

the system comprises a clustering center generation module, a training set feature point collection module and a training set feature point selection module, wherein the clustering center generation module is used for finding K feature points from the training set feature point collection as K clustering centers, and the K clustering centers correspond to K categories;

the class division module is connected with the clustering center generation module and is used for respectively calculating the distances between the K clustering centers in the training set feature point set and other feature points and dividing each other feature point into the class of the clustering center with the minimum distance from the other feature points; the other characteristic points are characteristic points except the clustering center;

the cluster center adjusting module is connected with the category dividing module and used for respectively calculating the mean value of all the feature points in each category and taking the mean value as a new cluster center;

and the clustering center determining module is simultaneously connected with the clustering center adjusting module and the classification dividing module and used for replacing the original clustering center with a new clustering center and returning to the classification dividing module and the clustering center adjusting module until the clustering center is not changed any more, and outputting the final K clustering centers and the corresponding classifications.

The system comprises a sample building module, a test set building module and a data processing module, wherein the sample building module further comprises a test set building module, and the test set building module is used for acquiring a door closing vehicle component position image acquired in advance and a category label corresponding to the door closing vehicle component position image and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

the SVM feature model training module comprises:

the classifier creating module is used for creating a Support Vector Machine (SVM) classifier and setting training parameters of an SVM feature model;

the feature model preliminary training module is connected with the classifier creating module and used for inputting the feature vector set of the training set and the class labels corresponding to the door closing vehicle component position images into a SVM classifier of a Support Vector Machine (SVM), and training the feature model of the SVM to obtain a preliminarily trained SVM feature model;

the characteristic model verification module is simultaneously connected with the sample building module and the characteristic model preliminary training module and is used for testing the preliminarily trained support vector machine SVM characteristic model by using a test set;

if the classification accuracy rate does not reach 100%, returning to the classifier creating module and resetting the training parameters,

The invention has the beneficial effects that:

therefore, in order to overcome the defect of manual detection of faults, a mode of replacing manual detection with machine detection is adopted, and the faults of the parts of the door closing car are detected by adopting an image processing technology and a machine vision technology which are mature in recent years, so that the operation efficiency can be improved, the personnel cost is reduced, the running safety of the truck is strongly ensured, the railway safe running detection system is promoted to be more perfect, and the method has a wide application prospect and has the following effects:

1. the classification algorithm mainly based on the SVM is adopted to classify the component images, the algorithm is mature, the classification effect is good, and compared with deep learning image classification, the classification accuracy is high, the classification speed is higher, and the real-time requirement of truck fault detection can be met.

2. The method adopts an ORB algorithm to extract the key point characteristics of the image, converts the image characteristics into digital characteristics to describe one image, can quickly extract the key points in the image, has high speed, and is not influenced by noise points and image transformation.

3. The improved K-means clustering algorithm is provided, the risk that the algorithm converges to a local minimum value is reduced by defining an initialized clustering center, the convergence speed is improved, the clustering of the component characteristic points is more accurate, and the interference of non-component characteristic points is reduced.

The multi-scale judgment is carried out on the SVM classification result of the door closing part position image, the fault category can be further accurately judged, the false alarm rate of fault judgment is reduced, and the fault detection accuracy is improved.

Drawings

FIG. 1 is an image of a cut-off door component in an image of a door closing vehicle component location;

FIG. 2 is a representation of a door handle in an image of a door closing vehicle component position;

FIG. 3 is a further image of the door closing vehicle component position image;

fig. 4 is a flowchart of a fault detection method based on a door closing component position image according to the present invention.

Detailed Description

In a first specific embodiment, a method for detecting a fault based on a door closing vehicle component position image in the present embodiment includes the following specific steps:

step 2, calculating characteristic points of the image to be detected;

step 4, inputting the feature vector of the image to be detected into a Support Vector Machine (SVM) classifier, and obtaining the classification result of the image to be detected by introducing a finally trained SVM feature model:

Further, the step of obtaining the support vector machine SVM feature model trained in step 4 is specifically as follows:

the category of the door closing part position image includes an intercepted door part image, a door handle image, and other images except the intercepted door part image and the door handle image;

step 42, acquiring a training set characteristic vector set of each door closing vehicle component position image in a sample training set;

Further, step 42 is specifically as follows:

Further, step 422 is specifically as follows;

Further, step 41 further comprises: acquiring a pre-acquired door closing vehicle component position image and a category label corresponding to the door closing vehicle component position image, and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

step 43 is specifically as follows:

Specifically, a truck fault rail edge image detection system (TFDS-3) adopts a high-speed camera to shoot an image of a traveling truck, reads a database and the image of the truck, intercepts a position image of a door closing truck part, classifies the image through an SVM (support vector machine) machine learning technology and an image processing technology, and judges whether the door closing truck fault exists.

Firstly, an image of the position of a door closing vehicle part is collected and intercepted, and the image has three conditions: intercepting the images of the cock component, the handle image and other images; the three types of door closing vehicle component position images are combined into a training set and a test set. And then, extracting ORB (feature extraction algorithm) features of each image in the training set by adopting an ORB algorithm, and combining feature points of all the images into an ORB feature point set of the training set. Similar feature points in the training set ORB feature point set are classified into one class through an improved K-means clustering algorithm, and K clustering centers are finally obtained to represent the ORB feature point set. Classifying ORB feature points of each training set image to K clustering centers according to the K clustering centers, and obtaining K-dimensional feature point description of each image after classification, namely called feature vectors of the image; and finally, combining the feature vectors of each image into a training set feature vector set to finish the extraction of the feature vectors of the training set.

Inputting the training set feature vectors into an SVM classifier to train a feature model of a sample set image, verifying the classification accuracy of the feature model through a test set after the feature model is trained until the feature model capable of accurately classifying the test set is trained, wherein the model can be applied to classification of the position image of the door closing vehicle component; the model is called through an SVM classifier, the fact that the position image of the door closing part of the truck passes through the model is judged, if the classification result is a handle image, the door closing fault is considered, an alarm is given immediately, fault information is timely fed back to a manual truck inspection worker, and safe operation of the truck is ensured through manual further verification.

The specific steps of the embodiment are as follows: collecting a data set, calculating ORB feature points, clustering K-means feature points, extracting feature vectors, training an SVM feature model, predicting and classifying SVM images, judging faults and outputting.

Firstly, collecting a door closing vehicle part image data set:

the method comprises the steps of collecting vehicle passing images through a TFDS-3 truck rail edge image detection system, reading a database, intercepting vehicle closing part position images, and collecting the vehicle closing part position images of different vehicle types in different time periods and different outdoor environments as far as possible. The position images of the door closing car parts are divided into three types by adopting a manual mode: intercepting the images of the cock component, the handle image and other images; finally, the collected image is divided into a sample set and a test set for use in the subsequent steps. The training set is used for constructing a feature model, and the testing set is used for testing the classification effect of the feature model.

Secondly, extracting ORB features of the image:

the orb (organized Fast and Rotated brief) algorithm is a method for extracting and describing features, and is divided into two parts, feature point extraction and feature point description. The algorithm mainly extracts image features through a FAST (features from segmented segment test) feature detection algorithm, and then performs feature description on the image by using a BRIEF algorithm. The algorithm is selected to have the advantages of high calculation speed and capability of meeting real-time application. The steps of extracting features using the ORB algorithm are:

(1) initializing an ORB detector;

(2) searching image characteristic key points, traversing pixel points of each sample image, and quickly selecting image key points through a Fast algorithm;

(3) and (3) calculating a feature descriptor, and creating a vector for each key point in the image by adopting a BRIEF algorithm, namely converting the image key points obtained in the step (2) into feature vectors, and using the feature vectors as the feature descriptor of the image.

(4) And combining the feature vectors of each image in the training set to form an ORB feature point set of the training set.

Thirdly, clustering image feature points by K-means

And secondly, obtaining an ORB feature point set of the sample training set, clustering the feature points by adopting a K-means clustering algorithm for improving an initial clustering center, and clustering similar feature points into one class to form a feature point statistical table containing K classes, wherein the K classes are the feature point clustering centers of the training set. The improved K-means clustering algorithm comprises the following specific steps:

(1) finding out K-150 characteristic points from an ORB characteristic point set of a training set as an initial clustering center of a K-means algorithm; the selection principle of the 150 characteristic points is that every two points are not adjacent, if the characteristic points are less than 150, 150 characteristic points are randomly selected from the characteristic point set to serve as initial clustering centers;

(2) calculating the distances from other feature points in the feature point set to the 150 initial clustering centers, and dividing each feature point into initial categories with the minimum distance from the feature point set;

(3) calculating the mean value of all the characteristic points divided into each category, and taking the mean value as a new clustering center of each category;

(4) repeating the processes of (2) and (3) until the cluster center is not changing;

(5) and outputting the final clustering center and the category to which each feature point corresponds.

And finally, obtaining a clustering center of the training set ORB characteristic point set and the category of the clustering center, and storing a data file of the clustering center for the characteristic vector extraction of the next image.

Fourthly, generating a characteristic vector set

Quantizing the image data features into digital features, generating a feature vector set to describe a sample training set, and specifically comprising the following steps of:

(1) traversing each image in the sample training set again, and calculating an ORB characteristic point set of each image according to the step 2;

(2) calculating the distances between the ORB characteristic points of each image and the K clustering centers obtained in the step 3, and sequentially classifying all the characteristic points into the category with the minimum distance;

(3) counting the number of the ORB characteristic points in K categories of each image after classification, and further representing the image into a K-dimensional numerical vector;

(4) and (3) repeating the steps (1) to (3) on each image in the sample training set, respectively calculating K-dimensional numerical vectors of the images, and combining the K-dimensional numerical vectors into a training set characteristic vector set.

The characteristic vector set obtained in the steps is the characteristic vector set of the sample training set, namely, the image characteristics are converted into digital characteristics to represent the image training sample set, and the digital characteristics are input into an SVM algorithm to train a door closing vehicle component characteristic model in the next step.

Fifthly, training SVM characteristic model

Since the class labels of the sample training set are known, the training of the feature model can be started only by inputting the feature vector set and the corresponding labels in the previous step into the SVM classifier. The method comprises the following specific steps:

(1) creating an SVM classifier, and setting training parameters of an SVM model;

(2) inputting a sample training set feature vector set and corresponding labels, and starting training an SVM feature model;

(3) storing the trained feature model;

(4) and (4) testing the classification effect of the feature model in the step (3) by using the verification set, returning to the step (1) to retrain the feature model if the classification accuracy rate does not reach 100%, and ending the training of the SVM feature model until the classification accuracy rate reaches 100%.

Sixthly, classification prediction and fault judgment of vehicle passing image SVM (support vector machine)

The method comprises the following steps of obtaining a car passing image in real time through a TFDS-3 truck image detection system, obtaining a car closing component position image, using an SVM feature model trained in step 5, and starting to predict a car closing component fault, wherein the method specifically comprises the following steps:

(1) acquiring a door closing part position image;

(2) calculating ORB characteristic points of the image;

(3) and (4) clustering the ORB feature points according to the method in the step (3) to extract the image feature vector.

(4) Inputting the feature vectors into an SVM classifier, and importing the trained SVM feature models in the step 5 to obtain an image classification prediction result;

(5) if the classification result is 'cut-off door component image class' or 'other image class', the closed door vehicle is considered to have no fault, and the detection of the next vehicle is continued;

(6) and if the classification result is 'handle image type', the door closing fault is considered, at the moment, the fault information is uploaded to a vehicle detection platform, and the door closing fault is confirmed manually.

In a second embodiment, the device for detecting a failure based on a door closing member position image according to the present embodiment includes:

the classification judgment module of the images to be detected is connected with the clustering module of the characteristic points of the images to be detected and is used for inputting the characteristic vectors of the images to be detected into the SVM classifier, and the classification result of the images to be detected is obtained by leading in the SVM characteristic model of the final training:

Further, the device for acquiring the trained SVM feature model in the classification judgment module of the image to be detected comprises:

the training set feature vector set acquisition module is connected with the sample construction module and is used for acquiring a training set feature vector set of each door closing vehicle component position image in the sample training set;

Further, the training set feature vector set calculating module comprises:

Further, the training set feature point clustering module comprises:

Further, the sample building module further comprises a test set building module, which is used for obtaining the pre-collected door closing vehicle component position image and the category label corresponding to the door closing vehicle component position image, and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

the SVM feature model training module comprises:

It should be noted that the present application also includes other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and the substance of the present application, but these corresponding changes and modifications should fall within the scope of the appended claims of the present application.

Claims

1. The fault detection method based on the door closing part position image is characterized by comprising the following specific steps:

the category of the door closing part position image comprises a cut-off door part image, a door handle image and other images; the other images include images other than a truncated stopper part image and a stopper handle image;

step 2, calculating the characteristic points of the image to be detected;

step 4, inputting the feature vector of the image to be detected into a Support Vector Machine (SVM) classifier, and obtaining a classification result of the image to be detected by introducing a finally trained SVM feature model: the SVM classifier classifies the images to be detected through an SVM feature model;

if the type of the image to be detected is an intercepted door plug part image or other images, judging that the door closing vehicle corresponding to the image to be detected has no fault;

and if the category of the image to be detected is the door handle image, judging that the door closing vehicle corresponding to the image to be detected has a fault.

2. The method for detecting the fault based on the door closing vehicle component position image according to claim 1, wherein the step of obtaining the SVM feature model trained in the step 4 is as follows:

step 41, acquiring a pre-acquired door closing component position image and a category label corresponding to the door closing component position image, and forming a sample training set;

3. The method for detecting faults based on the door closing car component position image as claimed in claim 2, wherein the step 42 is specifically as follows:

4. The method of claim 3, wherein step 422 is as follows;

5. The method of claim 4, wherein the fault detection based on the image of the location of the door closing component is performed by a computer,

step 41 further comprises: acquiring a pre-acquired door closing vehicle component position image and a category label corresponding to the door closing vehicle component position image, and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

step 43 is specifically as follows:

6. Fault detection device based on door closing car part position image, its characterized in that includes:

the classification judgment module of the images to be detected is connected with the clustering module of the characteristic points of the images to be detected and is used for inputting the characteristic vectors of the images to be detected into the SVM classifier, and the classification result of the images to be detected is obtained by leading in the finally trained SVM characteristic model: the SVM classifier classifies the images to be detected through an SVM feature model;

7. The device for detecting the fault based on the door closing vehicle component position image as claimed in claim 6, wherein the device for obtaining the trained SVM feature model in the image classification judgment module to be detected comprises:

8. The device of claim 7, wherein the training set feature vector set calculating module comprises:

9. The device of claim 8, wherein the training set feature point clustering module comprises:

the system comprises a clustering center generation module, a training set feature point collection module and a training set feature point collection module, wherein the clustering center generation module is used for finding K feature points from the training set feature point collection as K clustering centers, and the K clustering centers correspond to K categories;

10. The device of claim 9, wherein the image of the location of the door closing device is based on a location of the door closing device,

the sample building module also comprises a test set building module which is used for acquiring a pre-acquired door closing vehicle component position image and a category label corresponding to the door closing vehicle component position image and forming a test set; the door closing vehicle component position images in the test set are different from the door closing vehicle component position images in the sample training set; and

the SVM feature model training module comprises: