CN116091870A - Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat - Google Patents
Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat Download PDFInfo
- Publication number
- CN116091870A CN116091870A CN202310184867.6A CN202310184867A CN116091870A CN 116091870 A CN116091870 A CN 116091870A CN 202310184867 A CN202310184867 A CN 202310184867A CN 116091870 A CN116091870 A CN 116091870A
- Authority
- CN
- China
- Prior art keywords
- shape
- training
- image
- loss function
- board seat
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Train Traffic Observation, Control, And Security (AREA)
Abstract
The utility model provides a follow-up plate seat damage fault recognition detection network training and detection method, system and medium, relates to image processing technology field, and is because there are various natural interference, such as illumination, rainwater, mud stain to current follow-up plate seat image, and then leads to the model to be unable to concentrate on the overall shape in the training process to influence the effect of training, further led to the problem that the model recognition accuracy rate that trains is low. The improved loss function can filter the influence of individual pixel points on the recognition, so that the model is more focused on the overall shape, a more accurate model can be obtained under the same training set and the same training times, and the accuracy of the recognition of the damaged faults of the slave plate seat is improved. The loss function after the application is improved can also solve the problem caused by unbalance of positive and negative samples in the training process to a certain extent.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a network training and detecting method, a system and a medium for identifying and detecting plate seat breakage faults.
Background
The slave plate seat is used for fixing a coupler buffer device, is used for realizing the coupling and traction between a locomotive and a vehicle or between the vehicle and the vehicle, and is required to bear traction force and impact force generated when the train runs or is in shunting operation. In order to ensure the stable and safe running of the train, the validity and the integrity of the slave plate seat are required to be identified and detected, and once faults such as breakage, rivet breakage and the like are found, the faults are required to be rapidly processed. At present, a mode of manually checking images is adopted to perform fault checking on the slave board seat, and the situations of fatigue, omission and the like are very easy to occur in the working process of car checking staff, so that the occurrence of missed checking and false checking is caused, the driving safety is influenced, the manual checking efficiency is low, and a large amount of time is required for fault checking.
With the development of deep learning technology, deep learning should be able to realize automated detection in various fields. For the damage faults of the slave plate seat, when the slave plate seat is detected by adopting a deep learning technology, the acquired slave plate seat image is abnormal in part of pixel points due to various natural interferences such as illumination, rainwater and mud stains, and the model cannot concentrate on the whole shape due to the pixel points, so that the training effect is influenced, and the trained model is low in recognition accuracy.
Disclosure of Invention
The purpose of the invention is that: the method aims at solving the problems that the model cannot concentrate on the whole shape in the training process due to various natural interferences such as illumination, rainwater and mud stains existing in the existing slave board seat image, so that the training effect is affected, and the recognition accuracy of the trained model is low.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a method of training a network for identifying and detecting a failure from a breakage of a board seat, comprising:
a step of acquiring a railway train slave board seat area image and constructing a training set using the slave board seat area image, and
training a Mask-RCNN network by using the training set;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing a Dice Loss function;
the Focal Loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter, wherein gamma is ≡1.
Further, the Dice Loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
Further, the steps of collecting the images of the slave board seat area of the railway train, constructing a training set by utilizing the images of the slave board seat area, and training the Mask-RCNN network by utilizing the training set are specifically as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: carrying out image amplification on the fault image, and determining a training set based on the amplified fault image and the slave board seat area image;
step four: and (5) training the Mask-RCNN network by using the training set.
And further, intercepting the image of the slave plate seat area in the second step, determining the start and stop positions of the slave plate seat area according to the wheelbase information, and intercepting the slave plate seat area according to the start and stop positions.
Further, the image amplification includes: panning, zooming, and sharpening.
A failure recognition and detection method for damage of a slave plate seat comprises the following steps:
acquiring a railway train slave board seat area image as an image to be detected, carrying out segmentation detection on the image to be detected by using a trained Mask-RCNN network, and judging whether faults exist according to detection results;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing a Dice Loss function;
the Focal Loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter, wherein gamma is ≡1.
Further, the Dice Loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
Further, the training steps of the Mask-RCNN network are as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: carrying out image amplification on the fault image, and determining a training set based on the amplified fault image and the slave board seat area image;
step four: and (5) training the Mask-RCNN network by using the training set.
A failure recognition detection system for breakage of a slave board seat, comprising: the device comprises an acquisition module, a training module and a detection module;
the acquisition module is used for acquiring an image of a slave plate seat area of the railway train and taking the acquired image of the slave plate seat area as an image to be detected;
the training module is used for training the Mask-RCNN network;
the detection module is used for carrying out segmentation detection on the image to be detected by utilizing the trained Mask-RCNN network, and judging whether faults exist or not according to the detection result;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing the Dice Loss function.
A computer-readable storage medium in which a computer-readable program is stored for performing any one of the methods described above.
The beneficial effects of the invention are as follows:
the improved loss function can filter the influence of individual pixel points on the recognition, so that the model is more focused on the overall shape, a more accurate model can be obtained under the same training set and the same training times, and the accuracy of the recognition of the damaged faults of the slave plate seat is improved. The problem that the unbalance of positive and negative samples caused in the training process can also be solved to a certain extent to the loss function after the application is improved, so that the application deep learning network identifies the slave plate seat state and realizes fault alarm, thereby improving the detection efficiency and the stability and the safety of train operation. The method replaces manual detection by utilizing the mode of automatic image identification, and reduces labor cost. The deep learning algorithm is applied to automatic identification of the fault of the slave board seat, so that the stability and the accuracy of the whole algorithm are improved, and the fault detection time is effectively shortened. In the model training process, a method of combining the shape correlation coefficient and the Focal Loss is adopted as a new Loss function, optimization calculation is carried out, the training effect is improved, and the accuracy of the trained model in identifying the slave plate seat area image is further improved.
Drawings
Fig. 1 is a flow chart of the present application.
Detailed Description
It should be noted in particular that, without conflict, the various embodiments disclosed herein may be combined with each other.
The first embodiment is as follows: the network training method for identifying and detecting the breakage fault of the plate seat according to the embodiment comprises the following steps:
a step of acquiring a railway train slave board seat area image and constructing a training set using the slave board seat area image, and
training a Mask-RCNN network by using the training set;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing the Dice Loss function.
The Focal Loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter which is used for respectively balancing the number of positive and negative samples and adjusting the imbalance problem of the difficult-to-separate and easy-to-separate samples, wherein gamma is more than or equal to 1.
However, rail wagons have their specificity, i.e. the probability of failure is low, resulting in a very large proportion of normal pictures in the acquired data set and a sparse number of negative samples of failure. For this feature, the loss function of the model is improved. In replacing and improving the Loss function, the Focal Loss function is sometimes used to solve the problem caused by sample imbalance during training.
The Focal Loss function can be expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma is an adjustable parameter which is used for respectively balancing the number of positive and negative samples and adjusting the imbalance problem of the easily separable samples, wherein gamma is more than or equal to 1.
For a scoring sample with high probability of occurrence, the probability of the model identifying the target as the sample class is higher, i.e., p t Closer to 1, which tends to be taken to be smaller alpha t The value reduces the influence of the value on the calculation of the overall loss function; at the same time, p t The larger the 1-p t Smaller, and (1-p) due to gamma.gtoreq.1 t ) γ The smaller the loss power function of the easily separable sample is, the two phases are combined, and the problem of unbalance of the positive and negative samples in the training process is solved to a certain extent.
The second embodiment is as follows: this embodiment is a further description of the first embodiment, and the difference between this embodiment and the first embodiment is that the Dice Loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
Since segmentation detection is performed, judgment of shape similarity is introduced into the Loss function on the basis of Focal Loss. Conventional methods, such as Dice Loss, tend to be designed for pixel-level shape similarity, but sometimes the target morphology is not completely fixed for the recognition of the failure from the board seat, and the image quality may be problematic, such as shadowing, shielding or noise, if the conventional shape similarity analysis method is used, the change of part of key pixels may affect the training result. Thus, a new judgment shape similarity loss function is proposed, which can be expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is hereinafter simply referred to as shape A; b represents model predictionA set of target pixels, i.e. the shape of the predicted target, hereafter simply referred to as shape B; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same;
From this equation, it can be seen that if shape A is similar to shape B overall, then
The value of (2) and->
Should be close, the loss function is close to 0; otherwise, if the shapes have larger difference, the difference value of the two is larger, the loss function is larger, the deviation between the predicted value of the characterization model and the actual target is overlarge, and the model is required to be further optimized or trained. It can also be seen that if only some pixels change, the overall impact is not very great, thereby reducing the impact of key pixels on the loss function calculation.
And a third specific embodiment: this embodiment is a further description of the second embodiment, and the difference between the second embodiment and the second embodiment is that the steps of acquiring an image of a slave board seat area of a railway train, and constructing a training set using the image of the slave board seat area, and the steps of training a Mask-RCNN network using the training set are specifically as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: amplifying the fault image, and determining a training set and a testing set based on the amplified fault image and the slave board seat area image;
step four: and training the Mask-RCNN network by using the training set and the testing set.
The specific embodiment IV is as follows: the third embodiment is further described, and the difference between the third embodiment and the third embodiment is that in the second step, the image of the slave board seat area is taken, the start-stop position of the slave board seat area is determined according to the wheelbase information, and the slave board seat area is taken according to the start-stop position.
Fifth embodiment: the present embodiment is further described with respect to the fourth embodiment, and the difference between the present embodiment and the fourth embodiment is that the image amplification includes: panning, zooming, and sharpening.
Specific embodiment six: the method for identifying and detecting the damage faults of the slave board seat according to the embodiment comprises the following steps:
acquiring a railway train slave board seat area image as an image to be detected, carrying out segmentation detection on the image to be detected by using a trained Mask-RCNN network, and judging whether faults exist according to detection results;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing the Dice Loss function.
The whole thought of this application is: high-definition linear array imaging devices are built on two sides of a railway vehicle, and when a train passes through, a sensor is triggered to collect images. And intercepting a target subgraph and amplifying a fault image by utilizing the wheelbase information and priori knowledge of the slave board seat position so as to construct a data set. Based on the obtained data set training deep learning network model, the acquired image is identified by using the model, and the fault form and the fault position are determined. And mapping the identification result into the original image according to the mapping relation, and uploading the alarm. And the staff performs corresponding processing according to the identification result to ensure the safe operation of the train. When in actual use, a complete picture of the side part of the train is acquired and a target image to be detected is cut out; and detecting through the network model to obtain a detection result, analyzing the detection result, judging whether a fault exists, and if so, alarming is needed.
Specifically, the step of collecting an image of a railway train from a slab seat area image to an image to be measured generally comprises the steps of collecting the image and intercepting the area:
the image acquisition step specifically comprises the following steps:
and (3) constructing high-speed imaging equipment, acquiring high-definition linear array gray level images of all parts of the truck, and splicing the high-definition linear array gray level images into complete pictures of the side part and the bottom part of the train. Images in different environments in different time periods are collected, samples are expanded, various natural interferences such as illumination, rainwater, mud stains and the like exist in the data samples, the robustness of the algorithm is enhanced, and the algorithm can be applied to different working conditions of a train.
The step of region interception specifically comprises:
the detected slave plate seat targets are only distributed at the side part of the train, and the distribution has certain characteristics, so that the possible starting and stopping positions of the slave plate seat targets are determined according to priori knowledge such as the wheel base information and the like, the target areas are intercepted according to the starting and stopping positions to obtain target images, the sizes of the images to be identified are reduced, the relative proportion of the targets in the identification images is increased, the interference is reduced, the training of a model is facilitated, and the identification rate can be improved.
In the algorithm model used in the deep learning, a loss function is needed for measuring the model effect, characterizing the error between the model predicted value and the true value, and adjusting the weight parameter according to the result obtained by the loss function calculation so as to obtain a more accurate model.
For fault recognition from the board seat, a Mask-RCNN segmentation algorithm is used, which is a Two Stage target detection algorithm, and commonly used loss functions are a cross entropy loss function, a Smoooh L1 loss function, and the like.
Wherein the cross entropy loss function can be expressed as:
wherein x is i Represents the i-th sample, p (x i ) Representing the true distribution of the samples, q (x i ) Representing the distribution of model predictions.
Whereas Smoooh L1 loss function can be expressed as:
the present application combines the Focal Loss with the proposed shape similarity Loss function, resulting in a new Loss function as follows:
L＝θL Focal +(1-θ)L shape
wherein θ is an adjustable parameter, and can be adjusted according to the actual application scene, and is usually between 0 and 1.
The improved loss function can solve the problem caused by unbalance of positive and negative samples in the training process to a certain extent, and can filter the influence of individual pixel points on recognition, so that the model is more focused on the whole shape, a more accurate segmentation model can be obtained under the same training set and the same training times, and the research and development efficiency and the accuracy of plate seat damage fault recognition are improved.
The FocalLoss loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter, wherein gamma is greater than or equal to 1;
seventh embodiment: this embodiment is a further description of the sixth embodiment, and the difference between this embodiment and the sixth embodiment is that the Dice Loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
eighth embodiment: this embodiment is a further description of embodiment seven, and the difference between this embodiment and embodiment seven is that the training steps of the Mask-RCNN network are as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: carrying out image amplification on the fault image, and determining a training set based on the amplified fault image and the slave board seat area image;
step four: and (5) training the Mask-RCNN network by using the training set.
Detailed description nine: the present embodiment provides a failure recognition and detection system for breakage of a slave board, comprising: the device comprises an acquisition module, a training module and a detection module;
the acquisition module is used for acquiring an image of a slave plate seat area of the railway train and taking the acquired image of the slave plate seat area as an image to be detected;
the training module is used for training the Mask-RCNN network;
the detection module is used for carrying out segmentation detection on the image to be detected by utilizing the trained Mask-RCNN network, and judging whether faults exist or not according to the detection result;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing the Dice Loss function.
Detailed description ten: a computer-readable storage medium of the present embodiment stores therein a computer-readable program for executing the method of any one of the first to sixth embodiments.
It should be noted that the detailed description is merely for explaining and describing the technical solution of the present invention, and the scope of protection of the claims should not be limited thereto. All changes which come within the meaning and range of equivalency of the claims and the specification are to be embraced within their scope.
Claims (10)
1. The network training method for identifying and detecting the breakage faults of the plate seat is characterized by comprising the following steps of:
a step of acquiring a railway train slave board seat area image and constructing a training set using the slave board seat area image, and
training a Mask-RCNN network by using the training set;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing a Dice Loss function;
the Focal Loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter, wherein gamma is ≡1.
2. The method for training a network for identifying and detecting a broken fault from a board seat according to claim 1, wherein the Dice Loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
3. The method for training the slave board seat damage fault recognition detection network according to claim 2, wherein the steps of acquiring the slave board seat area image of the railway train and constructing a training set by using the slave board seat area image, and training the Mask-RCNN network by using the training set are specifically as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: carrying out image amplification on the fault image, and determining a training set based on the amplified fault image and the slave board seat area image;
step four: and (5) training the Mask-RCNN network by using the training set.
4. A method for training a slave board seat breakage fault recognition and detection network according to claim 3, wherein in the second step, the slave board seat area image is intercepted, the start-stop position of the slave board seat area is determined according to the wheelbase information, and the slave board seat area is intercepted according to the start-stop position.
5. The method for training a network for identifying and detecting a failure from a board seat failure according to claim 4, wherein the image amplification comprises: panning, zooming, and sharpening.
6. The method for identifying and detecting the damage faults of the slave plate seat is characterized by comprising the following steps:
acquiring a railway train slave board seat area image as an image to be detected, carrying out segmentation detection on the image to be detected by using a trained Mask-RCNN network, and judging whether faults exist according to detection results;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing FocalLoss loss function, L shape Representing a DiceLoss loss function;
the FocalLoss loss function is expressed as:
FL(p t )＝-α t (1-p t ) γ ×log(p t )
wherein p is t Representing confidence level of model prediction of a certain class t, 0 < p t ＜1；α t And gamma represents an adjustable parameter, wherein gamma is ≡1.
7. The method for identifying and detecting a failure from a board seat breakage according to claim 6, wherein the DiceLoss loss function is expressed as:
wherein A represents a set of actual target pixels, namely the shape of an actual target, which is simply called shape A; b represents a set of model prediction target pixels, namely the shape of a prediction target, which is called shape B for short; x represents the coordinates of each pixel point within a given range;
representing the average value of the coordinates of each point on the shape A, namely the geometric center coordinate of the shape A; B-A represents the set of points in shape B that are not in shape A at the same time, i.e., B-A n B; i A-B I are the same; />
Representing the distance of each point in shape B that is not in shape a from the geometric center of shape a; />
And the same is done; the two are subjected to difference and square, and an improved shape loss function is obtained;
8. The method for identifying and detecting the breakage fault of the slave board seat according to claim 7, wherein the training step of the Mask-RCNN network is as follows:
step one: taking side and bottom images of the railroad train;
step two: intercepting a slave board seat area image in side and bottom images of the railway train, and simulating a fault image based on the slave board seat area image;
step three: carrying out image amplification on the fault image, and determining a training set based on the amplified fault image and the slave board seat area image;
step four: and (5) training the Mask-RCNN network by using the training set.
9. A failure recognition and detection system for breakage of a slave board seat, characterized by comprising: the device comprises an acquisition module, a training module and a detection module;
the acquisition module is used for acquiring an image of a slave plate seat area of the railway train and taking the acquired image of the slave plate seat area as an image to be detected;
the training module is used for training the Mask-RCNN network;
the detection module is used for carrying out segmentation detection on the image to be detected by utilizing the trained Mask-RCNN network, and judging whether faults exist or not according to detection results;
the loss function of the Mask-RCNN network is expressed as:
L＝θL Focal +(1-θ)L shape
wherein θ represents an adjustable parameter, θ is between 0 and 1, L Focal Representing the Focal Loss function, L shape Representing the Dice Loss function.
10. A computer readable storage medium, characterized in that the medium has stored therein a computer readable program for executing the method according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310184867.6A CN116091870B (en) | 2023-03-01 | 2023-03-01 | Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310184867.6A CN116091870B (en) | 2023-03-01 | 2023-03-01 | Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116091870A true CN116091870A (en) | 2023-05-09 |
| CN116091870B CN116091870B (en) | 2023-09-12 |
Family
ID=86204515
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310184867.6A Active CN116091870B (en) | 2023-03-01 | 2023-03-01 | Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116091870B (en) |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111080612A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Truck bearing damage detection method |
| 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 |
| CN112489031A (en) * | 2020-12-10 | 2021-03-12 | 哈尔滨市科佳通用机电股份有限公司 | Mask-rcnn-based oil leakage detection method for snake-shaped-resistant shock absorber |
| US20210090247A1 (en) * | 2019-09-23 | 2021-03-25 | Samsung Sds Co., Ltd. | Apparatus and method for medical diagnostic |
| CN112990085A (en) * | 2021-04-08 | 2021-06-18 | 海南长光卫星信息技术有限公司 | Method and device for detecting change of culture pond and computer readable storage medium |
| CN113240014A (en) * | 2021-05-18 | 2021-08-10 | 长春理工大学 | Class II segmentation loss function and construction method and application thereof |
| WO2021179692A1 (en) * | 2020-10-22 | 2021-09-16 | 平安科技（深圳）有限公司 | Head ct image segmentation method and apparatus, electronic device and storage medium |
| CN114092487A (en) * | 2021-10-13 | 2022-02-25 | 山东师范大学 | Target fruit instance segmentation method and system |
| CN114511850A (en) * | 2021-12-30 | 2022-05-17 | 广西慧云信息技术有限公司 | Method for identifying image of fruit size and granule of sunshine rose grape |
| CN115616674A (en) * | 2022-10-12 | 2023-01-17 | 广州肖宁道路工程技术研究事务所有限公司 | Ground penetrating radar training data set simulation amplification and road nondestructive testing method and system |
| WO2023284340A1 (en) * | 2021-07-15 | 2023-01-19 | 北京小蝇科技有限责任公司 | Method for classifying species and development stage of parasite and classifying image pixel from bottom to top |
-
2023
- 2023-03-01 CN CN202310184867.6A patent/CN116091870B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210090247A1 (en) * | 2019-09-23 | 2021-03-25 | Samsung Sds Co., Ltd. | Apparatus and method for medical diagnostic |
| CN111080612A (en) * | 2019-12-12 | 2020-04-28 | 哈尔滨市科佳通用机电股份有限公司 | Truck bearing damage detection method |
| 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 |
| WO2021179692A1 (en) * | 2020-10-22 | 2021-09-16 | 平安科技（深圳）有限公司 | Head ct image segmentation method and apparatus, electronic device and storage medium |
| CN112489031A (en) * | 2020-12-10 | 2021-03-12 | 哈尔滨市科佳通用机电股份有限公司 | Mask-rcnn-based oil leakage detection method for snake-shaped-resistant shock absorber |
| CN112990085A (en) * | 2021-04-08 | 2021-06-18 | 海南长光卫星信息技术有限公司 | Method and device for detecting change of culture pond and computer readable storage medium |
| CN113240014A (en) * | 2021-05-18 | 2021-08-10 | 长春理工大学 | Class II segmentation loss function and construction method and application thereof |
| WO2023284340A1 (en) * | 2021-07-15 | 2023-01-19 | 北京小蝇科技有限责任公司 | Method for classifying species and development stage of parasite and classifying image pixel from bottom to top |
| CN114092487A (en) * | 2021-10-13 | 2022-02-25 | 山东师范大学 | Target fruit instance segmentation method and system |
| CN114511850A (en) * | 2021-12-30 | 2022-05-17 | 广西慧云信息技术有限公司 | Method for identifying image of fruit size and granule of sunshine rose grape |
| CN115616674A (en) * | 2022-10-12 | 2023-01-17 | 广州肖宁道路工程技术研究事务所有限公司 | Ground penetrating radar training data set simulation amplification and road nondestructive testing method and system |
Non-Patent Citations (2)
| Title |
|---|
| RUPAL R. AGRAVAT等: "3D Semantic Segmentation of Brain Tumor for Overall Survival Prediction", 《RESEARCHGATE》, pages 1 - 11 * |
| 宋浩然: "高速铁路无砟轨道异物视觉智能识别算法研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅰ辑》, no. 1, pages 026 - 223 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116091870B (en) | 2023-09-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115272334B (en) | Method for detecting tiny defects on surface of steel rail under complex background | |
| CN109489724B (en) | Tunnel train safe operation environment comprehensive detection device and detection method | |
| CN111080598B (en) | Bolt and nut missing detection method for coupler yoke key safety crane | |
| US11827228B2 (en) | Road abnormality detection apparatus, road abnormality detection method and road abnormality detection program | |
| CN102759347B (en) | Online in-process quality control device and method for high-speed rail contact networks and composed high-speed rail contact network detection system thereof | |
| CN109238756B (en) | Dynamic image detection equipment and detection method for freight car operation fault | |
| CN111091546B (en) | Railway wagon coupler tail frame breaking fault identification method | |
| CN113112501B (en) | Vehicle-mounted track inspection device and method based on deep learning | |
| CN109242035B (en) | Vehicle bottom fault detection device and method | |
| CN113516629A (en) | Intelligent detection system for TFDS passing operation | |
| CN108508023B (en) | Defect detection system for contact end jacking bolt in railway contact network | |
| CN115331086A (en) | Brake shoe breaking and rivet losing fault detection method | |
| CN114972177A (en) | Road disease identification management method and device and intelligent terminal | |
| CN113639685B (en) | Displacement detection method, device, equipment and storage medium | |
| CN113487561B (en) | Pantograph foreign matter detection method and device based on gray gradient abnormal voting | |
| CN116091870B (en) | Network training and detecting method, system and medium for identifying and detecting damage faults of slave plate seat | |
| CN112435289B (en) | Method and device for detecting width of rail joint gap of steel rail | |
| CN113191239A (en) | Vehicle overall dimension dynamic detection system based on computer vision | |
| KR101560810B1 (en) | Space controled method and apparatus for using template image | |
| CN107194923B (en) | Ultraviolet image diagnosis method for defect inspection of contact network power equipment | |
| Li et al. | Road Damage Evaluation via Stereo Camera and Deep Learning Neural Network | |
| CN110189301B (en) | Foreign matter detection method for generator stator core steel sheet stacking platform | |
| CN109064417B (en) | Subway tunnel image processing method and device | |
| CN116129374B (en) | Multifunctional flaw detection cluster device beside rail and control method thereof | |
| CN113640307B (en) | Rail condition monitoring method adopting machine vision |
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 |