CN115661776A - Method and system for identifying railway wagon brake beam safety chain falling fault image - Google Patents
Method and system for identifying railway wagon brake beam safety chain falling fault image Download PDFInfo
- Publication number
- CN115661776A CN115661776A CN202211314124.8A CN202211314124A CN115661776A CN 115661776 A CN115661776 A CN 115661776A CN 202211314124 A CN202211314124 A CN 202211314124A CN 115661776 A CN115661776 A CN 115661776A
- Authority
- CN
- China
- Prior art keywords
- image
- brake beam
- railway wagon
- fault
- box
- 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
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a method and a system for identifying a safe chain falling fault image of a brake beam of a railway wagon, and relates to a method and a system for identifying a safe chain falling fault image of a brake beam of a wagon. The invention aims to solve the problem that the detection precision is low due to the fact that a railway wagon brake beam safety chain is positioned behind a brake beam and is obviously shielded and overlapped in the existing target detection method. The process is as follows: establishing an original sample image data set; training the network model based on the original sample image data set until the model converges to obtain a trained network model; the specific process is as follows: the network structure is an SSD network; in the training process, an improved repetition Loss function and a DIoU-NMS non-maximum value inhibition method are adopted to remove a repeated frame; and carrying out fault identification on the image to be detected based on the obtained trained network model. The invention is used for the field of fault image identification.
Description
Technical Field
The invention relates to a truck brake beam safety chain falling fault image identification method and a truck brake beam safety chain falling fault image identification system.
Background
The train braking system is an important component of a train, and the performance and the braking capacity of the train braking system are directly related to whether the train can run safely or not. The brake beam is one of the most important parts on the railway vehicle, and the damage and the failure of the brake beam have great influence on the traffic safety. In the running process of a train, a brake beam bears large alternating load and impact force, and bears braking force and reaction force of wheels on brake shoes during braking, so that the stress condition is severe. The connection between the brake beam safety chain and the beam body is fastened together by an integral safety chain clip and an eye bolt. When the brake beam safety chain breaks down, the fixing effect on the brake beam body is lost, and the local abrasion of the beam body is caused by the direct friction between the safety chain and the beam body, so that potential safety hazards are brought to the running of the truck. The existing car inspection operation mode of manually looking at pictures one by one has the problems of influence of personnel quality and responsibility, error and omission detection, difficulty in ensuring the operation quality, high labor cost, low efficiency and the like.
Therefore, the method has important significance for automatic detection of the falling fault of the brake beam safety chain. By combining image processing and deep learning technologies, automatic fault identification and alarm are realized, and the quality and efficiency of vehicle inspection operation are effectively improved.
Disclosure of Invention
The invention aims to solve the problem that the detection precision is low due to obvious shielding and overlapping of a railway wagon brake beam safety chain behind a brake beam in the existing target detection method, and provides a railway wagon brake beam safety chain falling fault image identification method and system.
The method for identifying the falling fault image of the safety chain of the brake beam of the railway wagon comprises the following specific processes:
establishing an original sample image data set;
the image is a railway wagon brake beam safety chain image;
training the network model based on the original sample image data set until the model converges to obtain a trained network model; the specific process is as follows:
the network structure is an SSD network;
in the training process, an improved repetition Loss function and a DIoU-NMS non-maximum value inhibition method are adopted to remove a repeated frame;
the improved reporting Loss function is specifically as follows:
the reporting loss is composed of three loss items;
L＝L Attr +αL RepGT +βL RepBox
first item L Attr The loss values generated for the prediction frame and the real target frame;
second term L RepGT Is the loss value generated by the prediction frame and the surrounding real target frame;
third term L RepBox Is the loss value generated by the prediction box and the prediction box of other real targets;
alpha and beta represent weights;
and improving the surrounding real target frame in the second item, wherein the improved surrounding real target frame selection mode is represented by the following formula:
wherein D represents a default box in which the user can see,
a real target frame around the display area is shown,
representing a real target box matching the default box, G representing a real target box,
representing other real target frames except the real frame of the target to be regressed;
and carrying out fault identification on the image to be detected based on the obtained trained network model.
Preferably, an original sample image dataset is established; the specific process is as follows:
1. acquiring a linear array image of the railway wagon;
2. acquiring a rough positioning image of a safety chain component of a brake beam of the railway wagon based on the linear array image of the railway wagon;
3. carrying out data set amplification on the rough positioning image of the railway wagon brake beam safety chain component;
4. based on the amplified dataset, an original sample image dataset is created.
Preferably, acquiring a linear array image of a railway wagon in the first step; the specific process is as follows:
high definition equipment is set up respectively at freight train track bottom, shoots the freight train that passes through, acquires the image of freight train bottom.
Preferably, in the second step, a railway wagon brake beam safety chain component rough positioning image is obtained based on the railway wagon linear array image; the specific process is as follows:
and carrying out rough positioning on the position of the safety chain of the brake beam in the image of the bottom of the truck according to the truck wheel base information and the bogie type information to obtain a component rough positioning image.
Preferably, in the third step, data set amplification is carried out on the railway wagon brake beam safety chain component rough positioning image; the specific process is as follows:
the amplification form comprises rotation, translation, scaling and mirror image operation of the image, and each operation is performed under random conditions.
Preferably, an original sample image dataset is established based on the amplified dataset; the specific process is as follows:
the original sample image dataset comprises: a component coarse positioning image set and a marking information set;
there is a one-to-one correspondence between the set of part coarse positioning images and the set of marker information data, i.e., each part coarse positioning image corresponds to one marker data.
Preferably, the specific process of the DIoU-NMS non-maximum value inhibition method is as follows:
assuming that a default box set of the SSD network model is B, the category confidence score corresponding to the default box set B is s, and for the prediction box M with the highest score, the classification score of the ith default box is updated:
wherein s is i Represents a classification score, s i E is s; ε represents the NMS threshold, R DIOU (M,B i ) Represents the distance between the center point of the predicted frame with the highest score and the center point of the ith default frame, B i Represents the ith default box and the IOU represents the degree of overlap between the default box and the highest scoring predicted box.
Preferably, fault recognition is carried out on the image to be detected based on the obtained trained network model; the specific process is as follows:
step six: coarse positioning of the part:
taking out a subregion image containing the component according to the bogie type information and the prior knowledge of the region where the component is located;
step six and two: and (3) fault judgment:
inputting the sub-region image into the trained model, completing the detection of the brake beam safety chain component in the sub-region image, judging whether a fault exists, and if so, executing a sixth step and a third step; if not, continuously detecting the next image;
step six and three: uploading and alarming:
and generating a message according to the fault information and the fault position and fault category, and uploading the message to an alarm platform.
Preferably, the weights α and β both take values of 0.5.
The wagon brake beam safety chain falling fault image recognition system is used for executing a wagon brake beam safety chain falling fault image recognition method based on deep learning.
The invention has the beneficial effects that:
compared with the target detection of other truck parts, the brake beam safety chain is positioned behind the brake beam, and obvious shielding exists. There are two cases of occlusion: one is the situation that the safety chain display is incomplete due to the shielding of the brake beam, so that the deep learning detection model cannot learn complete features and false detection is caused; another is due to the flexible nature of the chain, which can cause different sections of chain to stack together, and the detection model cannot distinguish, resulting in missed detection. From the angle of the Loss function, the method adopts an improved replication Loss function and a DIoU-NMS non-maximum value inhibition method to remove the repeated frame, and solves the problem of low detection precision of the brake beam safety chain caused by shielding and overlapping conditions.
The invention selects the surrounding real target frames based on the real target frames matched with the default frame. The IoU of the target frame and other surrounding real target frames is calculated by taking the real target frame as a standard, and the selected target frame with the largest IoU is the surrounding real target frame.
The invention adopts a DIoU-NMS non-maximum value inhibition method, not only analyzes the overlapping area in the inhibition criterion, but also calculates the distance of the central point between the two detection frames, effectively avoids the problem of low target detection precision under the shielding and overlapping conditions, and can effectively improve the target detection precision under the shielding and overlapping conditions.
The high-definition imaging device at the bottom of the rail of the truck is used for shooting the truck moving at high speed to obtain a high-definition image of the bottom of the truck. And acquiring a coarse positioning image containing the component according to the wheel base information, the bogie type and other prior knowledge. And collecting, sorting and amplifying data of the images to obtain a training image sample set. And establishing a proper deep neural network according to the fault type, and training for multiple times until the model converges to obtain corresponding parameters. And in the identification stage, loading parameters, inputting the shot images into a network to obtain a prediction result, judging whether the images are in failure or not according to the prediction result, and alarming a failure area if the images are in failure.
According to the invention, an automatic identification technology is introduced into truck fault detection, so that automatic fault identification and alarm are realized, and only the alarm result needs to be confirmed manually, so that the labor cost is effectively saved, and the operation quality and the operation efficiency are improved. The deep learning algorithm is applied to automatic identification of the falling fault of the safety chain of the brake beam, and compared with the traditional machine vision detection method, the method has higher accuracy and stability.
Drawings
FIG. 1 is a flow chart of model training according to the present invention;
FIG. 2 is a diagram of an SSD network architecture;
fig. 3 is a failure determination flowchart.
Detailed Description
It should be noted that, in the case of conflict, the various embodiments disclosed in the present application may be combined with each other.
The first specific implementation way is as follows: the embodiment is described with reference to fig. 1 and 3, and the method for identifying the breakage fault of the connection pull rod of the railway wagon comprises the following specific processes:
establishing an original sample image data set;
the image is a railway wagon brake beam safety chain image;
training the network model based on the original sample image data set until the model converges to obtain a trained network model; the specific process is as follows:
the network structure is an SSD network;
in the training process, an improved repetition Loss function and a DIoU-NMS non-maximum value inhibition method are adopted to remove repeated frames;
the improved reporting Loss function is specifically as follows:
the reporting loss is composed of three loss items;
L＝L Attr +αL RepGT +βL RepBox
first item L Attr The loss values generated for the prediction box and the real target box;
second term L RepGT Is the loss generated by the predicted frame and the surrounding real target frame (i.e. the real target frame with the largest IoU in addition to the matched real target frame)A value;
third term L RepBox Is the loss value generated by the prediction box and the prediction box of other real targets;
alpha and beta represent weights;
and improving the surrounding real target frame in the second item, wherein the improved surrounding real target frame selection mode is represented by the following formula:
wherein D represents a default box in which the user can see,
a real target frame around the object is shown,
representing a real target box matching the default box, G representing a real target box,
representing other real target frames except the real frame of the target to be regressed;
and carrying out fault identification on the image to be detected based on the obtained trained network model.
After the training data set is established, a deep learning target detection algorithm is selected to detect the brake beam safety chain component, and the model training process is shown in figure 1.
Compared with the target detection of other truck parts, the brake beam safety chain is positioned behind the brake beam, and obvious shielding exists. There are two cases of occlusion: one is the situation that the safety chain display is incomplete due to the shielding of the brake beam, so that the deep learning detection model cannot learn complete features and false detection is caused; another is due to the flexible nature of the chain, which can cause different sections of chain to stack together, and the detection model cannot distinguish, resulting in missed detection. According to the method, from the angle of a Loss function, a repeated frame is removed by adopting an improved repetition Loss function and a DIoU-NMS non-maximum value inhibition method, and the problem of poor detection effect of the safety chain of the brake beam caused by the shielding condition is solved.
SSD target detection algorithm
Currently, commonly used target detection algorithms include Faster R-CNN and YOLO, wherein the Faster R-CNN discards the previous method for generating a candidate region by target detection, and the RPN generates the candidate region for detection, so that the detection precision is high, but the speed is low. And YOLO directly obtains the target position through regression and judges the type of the target position, so that the detection task is converted into a regression task. The generation process of the candidate area is omitted, the target frame position and category prediction can be directly obtained in one step, the detection speed is greatly increased, and the detection precision is relatively poor.
The invention adopts the SSD (solid State disk) target detection algorithm, which integrates the advantages of YOLO and fast R-CNN, so that the algorithm can keep high precision and the speed can meet the requirement of real-time detection.
Fig. 2 is a network structure of an SSD, which mainly includes three parts: a basic network, a feature extraction network and a detection network;
(1) Basic network: the basic network in the original text is built by using the VGG16, and an author replaces fc6 and fc7 full-connection layers in the original VGG16 by convolution layers with convolution kernels of 3 × 3 × 1024 and 1 × 1 × 1024 respectively, because the full-connection structure can destroy the position information of the features;
(2) A feature extraction network: after the conv7_2 feature maps in the modified VGG16 are convolved, 4 feature maps, conv8_2, conv9_2, conv10_2 and conv11_2, are generated through convolution combination of 1 × 1 and 3 × 3, respectively, and the conv4_3 and conv7 feature maps form a feature pyramid of the SSD, and the network generating the feature maps is a feature extraction network. The general detection algorithm uses one feature map for detection, and the SSD uses 6 feature maps.
(3) Detecting a network: the feature extraction network generates 6 feature maps in total, and the detection network convolutes each feature map by respectively adopting two parallel 3 x 3 convolution kernels to generate a coordinate bias map of which each feature point represents a coordinate value and a category confidence map of which each feature point represents a prediction score of a category.
Principle of reporting Loss design
The solution idea of reporting Loss is to design a new Loss function, which includes two parts: one is an attraction term that brings the prediction box closer to the real target box, and the other is a repulsion term that brings the prediction box as far away as possible from other real target boxes and other prediction boxes around. This arrangement enables the prediction box to be gradually approached to the matched real target during training, and avoids being filtered out due to non-extreme inhibition caused by entering the wrong target area.
Repulsion Loss function definition
The reporting loss is composed of three loss items;
L＝L Attr +αL RepGT +βL RepBox
(1)Attr Loss
first item L Attr The penalty for the predicted frame and the actual target frame is greater as the predicted frame and the actual target frame are farther apart, which brings the predicted frame and the matching actual target frame as close as possible.
Where P is the default box, P + For all of the sets of positive samples,
is the true target box with the largest IoU with the default box, B P Is a prediction box obtained from P regression, and Attr Loss adopts Smooth L1 The distance is calculated.
(2)RepGT Loss
Second term L RepGT Is the predicted frame and the surrounding real target frame (i.e. the predicted frame and the default frame except the matched real target frame)The actual target box with the largest frame IoU) the predicted box is as close as possible to the surrounding actual target box, which makes the predicted box as far away as possible from the surrounding actual target box.
Wherein
The default frame P is the real target frame with the maximum IoU except the matched real target frame, and the RepGT Loss adopts Smooth ln The distance is calculated.
(3)RepBox Loss
Third term L RepBox Is the loss value generated by the prediction box and the prediction boxes of other real targets, and if the prediction box is closer to the prediction boxes of other real targets, the loss value is larger, so that the prediction boxes are far away from each other as possible.
Wherein the prediction box
And prediction blocks for other objects
The larger the IoU is, the larger the loss is, so that two prediction frames can be effectively prevented from being filtered out by non-extreme inhibition due to close proximity, and the missing detection is reduced.
Implementation details of Repulsion Loss on SSD and improvements thereof
Unlike the method of classifying and regressing through candidate regions commonly used in the R-CNN series algorithm, the SSD classifies and regresses in a default frame manner, so there are many differences in the implementation details of replication Loss.
The replication Loss mainly increases two losses, i.e. the RepGT Loss and the RepBox Loss, compared with the original multitask Loss function of the SSD. The existence of the RepGT Loss enables the prediction frame to be far away from the real target frame near the matching target as far as possible, because the realization of the Repulsion Loss in the original text is based on the default frame, the selected real target frame is the surrounding real target frame with the largest IoU except the real target frame matched with the default frame, and the specific selection method is shown in the formula.
Where P is the default box, G is the true target box,
is the default box P, except for the matching real target box, the real target box with the largest IoU.
Is the true target box with the largest IoU with the default box P.
But for SSDs there is no default block, instead a default block. The default box is a series of fixed-position boxes selected based on the feature map, which are selected as positive samples if matched with the real target box, and then classified and regressed based on the positive sample default boxes. However, this scheme also has a problem in that, because the weight of the network is initialized randomly at the beginning of the training, the initial deviation of the prediction block may be relatively large, which makes the next positioning of the prediction block difficult to predict, and further affects the convergence and detection performance of the training.
In order to solve the problem, an improved scheme provided by the patent is to select a surrounding real target frame based on a real target frame matched with a default frame. The method has the advantages that the real target frame is used as a standard, the IoU of the real target frame and the IoU of other real target frames around the real target frame are calculated, the target frame with the largest IoU is selected as the real target frame around the real target frame, and the strategy which is selected without depending on a prediction frame does not influence training convergence even if the prediction frame has a large deviation, so that the algorithm is more robust. The improved selection manner of the real surrounding target frame can be represented by the following formula.
Wherein D represents a default box in which the user can see,
a real target frame around the display area is shown,
representing a real target box matched with the default box, and G representing the real target box;
for RepBox Loss, its role is to keep the prediction boxes that predict different targets away from each other. However, in implementation, if two prediction boxes are too close to each other, the design mechanism of RepBox Loss may cause the two prediction boxes to be out of range of the true target box of each other, which may cause the positioning of the prediction boxes to be deviated. To solve this problem, we set an overlap threshold, i.e. allow some intersection between the predicted box and the surrounding target real box, and only overlap exceeding this threshold will calculate RepBox Loss.
DIoU-NMS non-maximum suppression method
NMS is a generic term for methods of suppressing non-maximum elements, and is used to remove adjacent duplicate test frames, leaving the most accurate test frame. In the conventional NMS, a defined IoU is used for inhibiting repeated detection frames, but because only overlapping areas are analyzed, when the detection targets have shielding and overlapping conditions, unexpected inhibition effect is easy to occur, and the targets which need to be detected originally are filtered out, so that the condition that the targets are missed to be detected is caused. In order to solve the problem, the invention adopts a DIoU-NMS non-maximum value inhibition method, analyzes the overlapping area in the inhibition criterion, and calculates the central point distance between two detection frames, thereby effectively avoiding the problem of target missing detection under the conditions of shielding and overlapping.
Assuming that a default box set of the model is B and the corresponding category confidence set is s, for the prediction box M with the highest score, the classification score is updated:
wherein s is i Is the classification score,. Epsilon.is the NMS threshold, R DIOU (M,B i ) Is the center point distance between the predicted box with the highest score and the ith default box, B i Is the ith default box;
for two rectangular frames with a farther center point, it is possible to locate different objects and therefore not filter out. The detection frame removal is carried out by analyzing the distance between the IoU of the two rectangular frames and the central point at the same time, so that the target detection precision under the shielding and overlapping conditions can be effectively improved.
The second embodiment is as follows: the difference between the present embodiment and the first embodiment is that an original sample image dataset is established; the specific process is as follows:
1. acquiring a linear array image of the railway wagon;
2. acquiring a rough positioning image of a safety chain component of a brake beam of the railway wagon based on the linear array image of the railway wagon;
3. carrying out data set amplification on the rough positioning image of the safety chain component of the railway wagon brake beam;
4. based on the amplified dataset, an original sample image dataset is created.
Other steps and parameters are the same as those in the first embodiment.
The third concrete implementation mode: the difference between the first embodiment and the second embodiment is that the first embodiment obtains a linear array image of the rail wagon; the specific process is as follows:
high definition equipment is set up respectively at freight train track bottom, shoots the freight train that passes through, acquires the image of freight train bottom.
By adopting line scanning, seamless splicing of images can be realized, and a two-dimensional image with a large visual field and high precision is generated.
Other steps and parameters are the same as those in the first embodiment.
The fourth concrete implementation mode is as follows: the difference between the first embodiment and the third embodiment is that in the second embodiment, a railway wagon brake beam safety chain component rough positioning image is obtained based on a railway wagon linear array image; the specific process is as follows:
and carrying out rough positioning on the position of the safety chain of the brake beam in the image of the bottom of the truck according to the truck wheel base information and the bogie type information to obtain a component rough positioning image.
The local area image containing the part is intercepted from the large bottom image, so that the time required by fault identification can be effectively reduced, and the identification accuracy rate is improved.
Other steps and parameters are the same as those in one of the first to third embodiments.
The fifth concrete implementation mode is as follows: the difference between the embodiment and one of the first to the fourth embodiments is that in the third embodiment, data set amplification is carried out on the railway wagon brake beam safety chain component rough positioning image; the specific process is as follows:
although the creation of the sample data set includes images under various conditions, data amplification of the sample data set is still required to improve the stability of the algorithm.
The amplification form comprises rotation, translation, scaling and mirror image operation of the image, and each operation is carried out under random conditions, so that the diversity and applicability of the sample can be ensured to the maximum extent.
The truck parts can be influenced by natural conditions such as rainwater, mud, oil, black paint and the like or artificial conditions. Also, there may be differences in the images taken at different sites.
In the process of collecting the training image data set, the diversity is ensured, and images of different sites under various conditions are collected as much as possible.
Other steps and parameters are the same as in one of the first to fourth embodiments.
The sixth specific implementation mode: the difference between the first embodiment and the fifth embodiment is that, in the fourth embodiment, based on the amplified data set, an original sample image data set is established; the specific process is as follows:
the original sample image dataset comprises: a component coarse positioning image set and a marking information set; marking information fault positions and fault categories;
the original image set is a rough positioning image which is shot by equipment and contains a brake beam safety chain component;
the marking information set is information of a rectangular subarea containing a component and is obtained in a manual marking mode;
there is a one-to-one correspondence between the component coarse positioning image set and the marker information data set, i.e., each component coarse positioning image corresponds to one marker data.
Other steps and parameters are the same as in one of the first to fifth embodiments.
The seventh concrete implementation mode: the difference between this embodiment and one of the first to sixth embodiments is that the specific process of the DIoU-NMS non-maximum suppression method is as follows:
assuming that a default box set of the SSD network model is B, the category confidence score corresponding to the default box set B is s, and for the prediction box M with the highest score, the classification score of the ith default box is updated:
wherein s is i Represents a classification score, s i E is s; ε represents the NMS threshold, R DIOU (M,B i ) Represents the distance between the center point of the predicted frame with the highest score and the center point of the ith default frame, B i Represents the ith default box and the IOU represents the degree of overlap between the default box and the highest scoring predicted box.
Other steps and parameters are the same as those in one of the first to sixth embodiments.
The specific implementation mode is eight: the difference between the first embodiment and the seventh embodiment is that the fault recognition is carried out on the image to be detected based on the obtained trained network model; the specific process is as follows:
step six: coarse positioning of the parts:
taking out a subregion image containing the component according to the bogie type information and the prior knowledge of the region where the component is located;
step six two: and (3) fault judgment:
inputting the sub-region image into the trained model, completing the detection of the brake beam safety chain component in the sub-region image, carrying out logic analysis on the position and the shape of the component, judging whether a fault exists, and if so, executing a sixth step and a third step; if not, continuously detecting the next image;
step six and three: uploading an alarm:
and generating a message according to the fault information and the fault position and fault category, and uploading the message to an alarm platform.
Other steps and parameters are the same as those in one of the first to seventh embodiments.
The specific implementation method nine: this embodiment is different from the first to eighth embodiments in that the values of the weights α and β are both 0.5.
Other steps and parameters are the same as those in one to eight of the embodiments.
The specific implementation mode is ten: the railway wagon brake beam safety chain falling fault image recognition system is used for executing a railway wagon connecting pull rod breaking fault recognition method based on deep learning.
The above-described calculation examples of the present invention are merely to explain the calculation model and the calculation flow of the present invention in detail, and are not intended to limit the embodiments of the present invention. It will be apparent to those skilled in the art that other variations and modifications of the present invention can be made based on the above description, and it is not intended to be exhaustive or to limit the invention to the precise form disclosed, and all such modifications and variations are possible and contemplated as falling within the scope of the invention.
Claims (10)
1. The method for identifying the falling fault image of the safety chain of the brake beam of the railway wagon is characterized by comprising the following steps of: the method comprises the following specific processes:
establishing an original sample image data set;
the image is a railway wagon brake beam safety chain image;
training the network model based on the original sample image data set until the model converges to obtain a trained network model; the specific process is as follows:
the network structure is an SSD network;
in the training process, an improved repetition Loss function and a DIoU-NMS non-maximum value inhibition method are adopted to remove a repeated frame;
the improved reporting Loss function is specifically as follows:
the reconstruction loss consists of three loss terms;
L＝L Attr +αL RepGT +βL RepBox
first item L Attr The loss values generated for the prediction box and the real target box;
second term L RepGT Is the loss value generated by the prediction frame and the surrounding real target frame;
third item L RepBox Is the loss value generated by the prediction box and the prediction box of other real targets;
alpha and beta represent weights;
and improving the surrounding real target frame in the second item, wherein the improved surrounding real target frame selection mode is represented by the following formula:
wherein D represents a default box in which the user can select,
a real target frame around the object is shown,
representing a real target box matching the default box, G representing a real target box,
representing other real target frames except the real frame of the target to be regressed;
and carrying out fault identification on the image to be detected based on the obtained trained network model.
2. The method for identifying a railway wagon brake beam safety chain disengagement fault image as claimed in claim 1, wherein the establishing a raw sample image dataset; the specific process is as follows:
1. acquiring a linear array image of a railway wagon;
2. acquiring a railway wagon brake beam safety chain component coarse positioning image based on the railway wagon linear array image;
3. carrying out data set amplification on the rough positioning image of the safety chain component of the railway wagon brake beam;
4. based on the amplified dataset, an original sample image dataset is created.
3. The method for identifying the safe chain falling fault image of the brake beam of the railway wagon as claimed in claim 2, wherein the first step is to obtain a linear array image of the railway wagon; the specific process is as follows:
high definition equipment is set up respectively at freight train track bottom, shoots the freight train that passes through, acquires the image of freight train bottom.
4. The method for identifying the falling fault image of the safety chain of the brake beam of the railway wagon as claimed in claim 3, wherein in the second step, a rough positioning image of the safety chain component of the brake beam of the railway wagon is obtained based on a linear array image of the railway wagon; the specific process is as follows:
and carrying out rough positioning on the position of the safety chain of the brake beam in the image of the bottom of the truck according to the truck wheel base information and the bogie type information to obtain a component rough positioning image.
5. The method for identifying the image of the railway wagon brake beam safety chain dropping fault according to the claim 4, wherein the third step is to perform data set amplification on the railway wagon brake beam safety chain component coarse positioning image; the specific process is as follows:
the amplification form comprises rotation, translation, zooming and mirror image operation of the image, and each operation is performed under random conditions.
6. The method for identifying the safe chain falling fault image of the railway wagon brake beam as claimed in claim 5, wherein the four data sets are based on an amplified data set, and an original sample image data set is established; the specific process is as follows:
the original sample image dataset comprises: a component coarse positioning image set and a marking information set;
there is a one-to-one correspondence between the component coarse positioning image set and the marker information data set, i.e., each component coarse positioning image corresponds to one marker data.
7. The method for identifying the safe chain dropping fault image of the brake beam of the railway wagon according to claim 6, wherein the DIoU-NMS non-maximum value inhibition method comprises the following specific processes:
assuming that a default box set of the SSD network model is B, the category confidence score corresponding to the default box set B is s, and for the prediction box M with the highest score, the classification score of the ith default box is updated:
wherein s is i Represents a classification score, s i Belongs to s; ε represents the NMS threshold, R DIOU (M,B i ) Represents the distance between the center point of the predicted frame with the highest score and the center point of the ith default frame, B i Denotes the ith default box, IOU denotes defaultThe degree of overlap between the bounding box and the prediction box with the highest score.
8. The method for identifying the image of the falling fault of the safety chain of the brake beam of the railway wagon according to claim 7, wherein the image to be detected is subjected to fault identification based on the obtained trained network model; the specific process is as follows:
step six: coarse positioning of the parts:
taking out a subregion image containing the component according to the bogie type information and the prior knowledge of the region where the component is located;
step six and two: and (3) fault judgment:
inputting the sub-region image into the trained model, completing the detection of the brake beam safety chain component in the sub-region image, judging whether a fault exists, and if so, executing a sixth step and a third step; if not, continuously detecting the next image;
step six and three: uploading and alarming:
and generating a message according to the fault information and the fault position and fault category, and uploading the message to an alarm platform.
9. The method for identifying the safe chain dropping fault of the brake beam of the railway wagon as claimed in claim 8, wherein the values of the weights α and β are both 0.5.
10. A railway wagon brake beam safety chain falling fault image identification system, which is characterized by being used for executing the deep learning-based railway wagon brake beam safety chain falling fault image identification method as claimed in one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211314124.8A CN115661776A (en) | 2022-10-25 | 2022-10-25 | Method and system for identifying railway wagon brake beam safety chain falling fault image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211314124.8A CN115661776A (en) | 2022-10-25 | 2022-10-25 | Method and system for identifying railway wagon brake beam safety chain falling fault image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115661776A true CN115661776A (en) | 2023-01-31 |
Family
ID=84991939
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211314124.8A Pending CN115661776A (en) | 2022-10-25 | 2022-10-25 | Method and system for identifying railway wagon brake beam safety chain falling fault image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115661776A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111080608A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for recognizing closing fault image of automatic brake valve plug handle of railway wagon in derailment |
| CN111091547A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon brake beam strut fracture fault image identification method |
| CN114743099A (en) * | 2022-03-29 | 2022-07-12 | 北京京天威科技发展有限公司 | Method and device for identifying faults of train key components |
-
2022
- 2022-10-25 CN CN202211314124.8A patent/CN115661776A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111080608A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Method for recognizing closing fault image of automatic brake valve plug handle of railway wagon in derailment |
| CN111091547A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon brake beam strut fracture fault image identification method |
| CN114743099A (en) * | 2022-03-29 | 2022-07-12 | 北京京天威科技发展有限公司 | Method and device for identifying faults of train key components |
Non-Patent Citations (2)
| Title |
|---|
| XINLONG WANG: "Repulsion_Loss_Detecting_Pedestrians_in_a_Crowd" * |
| ZHAOHUI ZHENG: "Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression" * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Liu et al. | A review of applications of visual inspection technology based on image processing in the railway industry | |
| Li et al. | Automatic defect detection of metro tunnel surfaces using a vision-based inspection system | |
| James et al. | Tracknet-a deep learning based fault detection for railway track inspection | |
| CN111079819B (en) | Method for judging state of coupler knuckle pin of railway wagon based on image recognition and deep learning | |
| CN111080608A (en) | Method for recognizing closing fault image of automatic brake valve plug handle of railway wagon in derailment | |
| CN111079746B (en) | Railway wagon axle box spring fault image identification method | |
| CN104599249A (en) | Cable rope bearing bridge deck vehicle load distribution real-time detection method | |
| CN111311567A (en) | Method for identifying fastener and steel rail diseases of track line image | |
| Yella et al. | Condition monitoring of wooden railway sleepers | |
| Katsamenis et al. | Simultaneous Precise Localization and Classification of metal rust defects for robotic-driven maintenance and prefabrication using residual attention U-Net | |
| KR102185225B1 (en) | Method for detecting sinkhole using deep learning and data association and sinkhole detecting system using it | |
| CN111080601A (en) | Method for identifying fault image of pull ring grinding shaft of derailment brake device of railway wagon | |
| CN113221839B (en) | Automatic truck image identification method and system | |
| Ye et al. | Application of lightweight railway transit object detector | |
| EP3385884A1 (en) | Method for recognising an oject of a mobile unit | |
| CN115527170A (en) | Method and system for identifying closing fault of door stopper handle of automatic freight car derailing brake device | |
| Koh et al. | Autonomous road potholes detection on video | |
| Aydin et al. | Defect classification of railway fasteners using image preprocessing and alightweight convolutional neural network | |
| Wang et al. | Anodfdnet: A deep feature difference network for anomaly detection | |
| CN115082850A (en) | Template support safety risk identification method based on computer vision | |
| CN114067296A (en) | Method and device for identifying surface defects of steel rail | |
| Jenefa et al. | Real-Time Rail Safety: A Deep Convolutional Neural Network Approach for Obstacle Detection on Tracks | |
| CN115661776A (en) | Method and system for identifying railway wagon brake beam safety chain falling fault image | |
| Chen | Migration learning-based bridge structure damage detection algorithm | |
| Kumari et al. | YOLOv8 Based Deep Learning Method for Potholes Detection |
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 |