CN116453040A - Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault - Google Patents

Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault Download PDF

Info

Publication number
CN116453040A
CN116453040A CN202310222178.XA CN202310222178A CN116453040A CN 116453040 A CN116453040 A CN 116453040A CN 202310222178 A CN202310222178 A CN 202310222178A CN 116453040 A CN116453040 A CN 116453040A
Authority
CN
China
Prior art keywords
iron wire
loosening
train
feature
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310222178.XA
Other languages
Chinese (zh)
Inventor
张闽东
王增
王盼盼
耿雪霏
朱善玮
武慧杰
刘鹏飞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Gtv Technology Development Co ltd
Original Assignee
Beijing Gtv Technology Development Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Gtv Technology Development Co ltd filed Critical Beijing Gtv Technology Development Co ltd
Priority to CN202310222178.XA priority Critical patent/CN116453040A/en
Publication of CN116453040A publication Critical patent/CN116453040A/en
Pending legal-status Critical Current

Links

Classifications

    • 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
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G06V10/765—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects using rules for classification or partitioning the feature space
    • 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
    • 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
    • G06V10/778—Active pattern-learning, e.g. online learning of image or video features
    • 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
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • 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
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/809—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of classification results, e.g. where the classifiers operate on the same input data
    • 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
    • G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations

Landscapes

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

Abstract

The invention discloses an intelligent image analysis method for a broken iron wire type fault of a train, which comprises the following steps: step 1, acquiring image data containing anti-loosening iron wires on a train, marking the anti-loosening iron wires to be detected in the image, and dividing the whole data into a training set, a verification set and a test set; step 2, performing data enhancement operation, configuring super parameters such as learning rate, momentum gradient, weight attenuation and the like required by an algorithm, and inputting a training sample into an improved yolov7 model to perform feature learning; and 3, inputting the new driving image into the learned feature model for target recognition, and finally framing out the failed region and outputting corresponding alarm information. The technical scheme of the invention can balance the performance of model training and reasoning, so that the algorithm has real-time performance and stronger feature extraction capability, and the features of a small target are reserved through an attention mechanism, thereby meeting the capability of detecting the faults of the anti-loose iron wires.

Description

Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault
Technical Field
The invention relates to the technical field of machine vision and rail transit fault detection, in particular to an intelligent image analysis method for a train vehicle anti-loosening iron wire fracture fault.
Background
Rail transit is an important traffic mode, and the mileage of construction and operation is greatly increased. The life safety of passengers and the running safety of trains are guaranteed, and the safety of passengers and the running safety of trains are an indispensable part of the development of rail transit. Wherein, different automobile body structures are complicated, and a lot of important parts are fixed by bolted connection, in order to prevent that the bolt from losing and not hard up, fixed locking iron wire between the bolt. Along with the long-time running of the train, the conditions of breakage of the anti-loosening iron wires can occur due to vibration, impact and the like, so that the bolts fall off, and the safety of running is threatened.
At present, most of the detection to the anti-loosening iron wire faults is also through a manual detection method, so that not only is a great deal of labor cost required, but also the detection efficiency is low, visual fatigue can be brought to detection personnel through long-time detection, and false detection is more likely to occur when small targets are inspected. The image recognition technology based on deep learning realizes high-efficiency detection of the target by learning fault characteristic information, and helps to save a lot of time cost manually. The current algorithm has good fault detection effect only on larger targets, but still has the problems of poor detection and more missed detection on small targets such as anti-loose iron wires.
Disclosure of Invention
The invention aims to provide an intelligent image analysis method for a train anti-loosening iron wire fracture fault, which solves the problems in the prior art.
In order to achieve the purpose, the intelligent image analysis method for the broken type faults of the anti-loosening iron wires of the train is provided for solving the problem that intelligent identification detection is carried out on faults and missed detection of the anti-loosening iron wires. Specifically, the method comprises the following steps:
step 1, acquiring image data containing anti-loosening iron wires on a train, marking the anti-loosening iron wires to be detected in the acquired image data, and dividing the anti-loosening iron wire image data and corresponding marking files into a training set, a verification set and a test set;
step 2, performing data enhancement operation, configuring super parameters required by an algorithm, and inputting a training sample into an improved yolov7 model to perform feature learning;
and step 3, inputting a new passing image into the learned feature model for target recognition, framing out a failed region and outputting corresponding alarm information.
Further, the specific process for acquiring the image data of the anti-loosening iron wire contained on the train comprises the following steps: when passing by the track side equipment, the train triggers the sensor to receive signals, the high-speed linear array scanning equipment acquires images of the whole train in operation, the images of the whole train are divided into high-definition images with the same resolution through cutting, the high-definition images are stored in the server, and the images containing the anti-loosening iron wires are selected out to serve as data to be marked.
Further, the specific flow of marking the anti-loosening iron wire is as follows: manually labeling the cut high-definition image by means of a LabelImg tool, wherein labeling comprises the steps of drawing a 2D surrounding rectangular frame for an anti-loosening iron wire to be detected, acquiring position information of the anti-loosening iron wire in a current labeling image, defining category information to which the anti-loosening iron wire belongs, and generating an xml file, wherein the content of the xml file comprises category information, x-axis and y-axis coordinates of the upper left corner and x-axis and y-axis coordinates of the lower right corner of the rectangular frame, and the category information comprises normal and/or broken conditions.
Further, the overall data is divided into a training set, a verification set and a test set according to the proportion of 7:2:1 by randomly dividing the marked anti-loosening iron wire sample data set; the training set is used for learning fault and non-fault characteristics of the model, the verification set is used for adjusting and optimizing weight parameters of the network model, and the test set is used for checking generalization capability of the model.
Further, the specific flow of data enhancement includes: randomly extracting 1/3 of the images in the training sample to perform rotation, contrast enhancement, mosaics enhancement and color space enhancement; providing abundant image background and fault morphology for the sample by adopting a supervised image processing method, amplifying the number of targets in the image, and enhancing the characteristic information of the anti-loosening iron wire;
the configuration of the super-parameters comprises: input image size, pre-training model, learning rate, momentum gradient descent method, training times and batch processing picture number.
Further, the basic composition modules of the overall network framework of the improved yolov7 model include: an input layer, a feature extraction network, a path aggregation feature pyramid network and a prediction output layer; the feature extraction network includes a CBS module, an ELAN structure, and an MP structure layer.
Further, the path aggregation feature pyramid network structure is characterized in that the last three feature layers of the trunk feature extraction network are used for transmitting deep information into a shallow layer by introducing a top-down path, so that features of different layers are fused efficiently, and feature semantic information is enhanced; and then the feature information between the network layers is fused deeply by downsampling and deep features.
And the prediction output layer is used for adjusting the number of image channels of three characteristic layers with different scales of P3, P4 and P5 output by the characteristic pyramid network through a RepConv module structure, and finally predicting confidence, category and offset information of an anchor frame through 1X 1 convolution.
Further, the improved integral network framework of yolov7 replaces the trunk feature extraction network with a RepConv structure, a multi-branch structure is adopted during model training, different receptive fields are obtained through different convolution kernels, different receptive field information is added, the capability of extracting feature information is further enhanced, and parameters are reconstructed during reasoning by the model; based on the feature fusion network for optimization, a SimAM attention mechanism is embedded before the path aggregation feature pyramid network structure, the importance of each neuron is evaluated, important neurons are searched by measuring the linear separability among the neurons, the neurons with key features are given higher weight, and surrounding interference information is suppressed.
Further, the yolov7 improved integral network frame further comprises the steps of suppressing redundant predicted frames, replacing an original NMS with a Soft-NMS algorithm to filter candidate frames, performing frame suppression on the original confidence score by using function operation in the algorithm execution process according to the superposition degree between the score and the frames, and reducing rejection of effective fault detection frames by reducing the confidence score under the condition that a plurality of anti-loosening iron wires are aggregated.
The method of the invention has the following advantages:
the invention provides an intelligent image analysis method for a broken iron wire type fault of a train vehicle, which is improved based on a yolov7 model, and a RepVGG network is used as a trunk feature extraction network, so that the model is more flexible and easy to deploy, and meanwhile, the capability of extracting feature information is enhanced. The performance of the model is greatly improved through parameter reconstruction during model reasoning. By suppressing surrounding interference information before feature aggregation, the weight of targets with fewer features in the image is improved, and the detection capability of the model on small targets is realized. By the method for reducing the confidence of redundant prediction frames, the elimination of effective detection frames can be avoided.
Drawings
FIG. 1 is a flow chart of an intelligent identification method for breaking faults of anti-loose iron wires;
FIG. 2 is a diagram of a network architecture based on the yolov7 improvement of the present invention;
FIG. 3 is a network block diagram of RepConv in the backbone network of the present invention.
Detailed Description
The technical solution of the present invention will be clearly and completely described in conjunction with the specific embodiments, but it should be understood by those skilled in the art that the embodiments described below are only for illustrating the present invention and should not be construed as limiting the scope of the present invention. All other embodiments, which can be made by one of ordinary skill in the art without undue burden on the person of ordinary skill in the art based on the embodiments of the present invention, are within the scope of the present invention.
The experimental methods used in the following examples are conventional methods unless otherwise specified. Materials, reagents and the like used in the examples described below are commercially available unless otherwise specified.
As shown in fig. 1, the present invention provides a technical solution: the embodiment is an intelligent image analysis method aiming at the fault of breaking of an anti-loose iron wire of a train, and the specific steps of the integral operation include:
firstly, acquiring image data containing anti-loosening iron wires on a train, marking the anti-loosening iron wires to be detected in the image, and dividing the anti-loosening iron wire image and a corresponding marking file into a training set, a verification set and a test set.
And performing data enhancement operation, configuring super parameters required by an algorithm, and inputting training samples into an improved yolov7 model to perform feature learning. Where the super-parameters are parameters that are set to values before starting the learning process when training the network model in deep learning. In general, the super parameters need to be optimized, and a group of optimal super parameters need to be selected for the learning machine so as to improve the learning performance and efficiency.
And inputting the new passing image into the learned feature model for target recognition, and finally framing out the failed region and outputting corresponding alarm information.
The specific process for acquiring the image data of the anti-loosening iron wire comprises the following steps:
when the train passes by the trackside equipment, the sensor is triggered to receive signals, the high-speed linear array scanning equipment acquires images of the whole train in operation, the images of the whole train are divided into high-definition images with the same resolution through cutting, and the high-definition images are stored in the server. And then selecting the image containing the anti-loosening iron wires as data to be marked.
The specific flow of image marking and data dividing is as follows:
the method comprises the steps that a cut high-definition image is manually marked by means of a Labelimg tool, the marked content comprises that a 2D surrounding rectangular frame is drawn on an anti-loosening iron wire to be detected, position information of the anti-loosening iron wire in a current marked image is obtained, the category (normal/broken) to which the anti-loosening iron wire belongs is defined, an xml file is generated, and the file content comprises category information (normal/broken), x-axis coordinate and y-axis coordinate information of an upper left corner and a lower right corner of the rectangular frame.
Randomly dividing the marked anti-loosening iron wire sample data set into a training set, a verification set and a test set according to the proportion of 7:2:1. The training set is used for learning fault and non-fault characteristics of the model, the verification set is used for adjusting and optimizing weight parameters of the network model, and the test set is used for checking generalization capability of the model.
The configuration of the super parameters comprises the input image size, a pre-training model, a learning rate, a momentum gradient descent method, training times and batch processing picture numbers.
The specific method of the image enhancement strategy comprises the following steps:
1/3 of the images in the samples are randomly extracted for rotation, contrast enhancement, mosaics enhancement and color space enhancement. A supervised image processing method is adopted to provide abundant image background and fault morphology for the sample, the number of targets in the image is amplified, and the characteristic information of the anti-loosening iron wires is enhanced.
The basic composition modules of the network framework include:
the yolov7 comprises an input layer, a feature extraction network, a path aggregation feature pyramid network and a prediction output layer. The feature extraction network comprises a CBS module, an ELAN structure and an MP structure layer. The CBS module consists of a convolution layer, a feature normalization layer and a nonlinear activation layer, input features are extracted by 3×3 convolution, data distribution is adjusted through a BN layer, overfitting is prevented, and features are fitted through nonlinear mapping. The ELAN is composed of a plurality of CBS modules, so that the deep network can learn and converge effectively by controlling the number of characteristic channels. The MP structure is composed of a maximum pooling module and a CBS module, and feature dimensions are reduced by using two feature compression modes of 3×3 convolution of maximum pooling and stride=2, so that the calculation overhead of the whole network is reduced.
According to the path aggregation feature pyramid network structure, the last three feature layers of the backbone network are used for transmitting deep information into a shallow layer through introducing a top-down path, so that efficient fusion of features of different levels is realized, and the feature semantic information is enhanced. And then the feature information between the network layers is deeply fused through downsampling and deep feature fusion, so that the relevance between the features is enhanced, and the characterization capability of the network can be improved.
The prediction output module is used for adjusting the number of image channels of three characteristic layers with different scales, namely P3, P4 and P5, output by the characteristic pyramid network through a RepConv module structure, and finally predicting confidence, category and offset information of an anchor frame through 1X 1 convolution.
The Loss function comprises three parts of coordinate Loss, target confidence Loss and classification Loss, wherein the target confidence Loss and the classification Loss adopt binary cross entropy Loss, and the coordinate Loss adopts CIoU_Loss.
The binary cross entropy loss is expressed as:
wherein i represents sample data; y is a real label; a represents the predicted output; n represents the total sample size.
The coordinate loss function is expressed as:
the IOU is the ratio of the intersection and union of the prediction frame and the real frame; distance_2 is the Euclidean Distance between two center points of the prediction frame and the real frame; c is the minimum circumscribed rectangle of the intersection and union of the prediction frame and the real frame, and distance_C is the diagonal Distance of C; v is a parameter that measures aspect ratio uniformity.
As shown in fig. 2, the overall network framework based on yolov7 improvement is first optimized based on the backbone feature extraction network:
the feature extraction network in the backbone network is uniformly replaced by a RepConv structure, and the characteristics of the structure can reconstruct parameters. Referring to fig. 3, the network adopts a multi-branch structure in the training stage, and combines 3×3 convolution, 1×1 convolution branches and identity residual branches, and adopts different convolution kernels to obtain different receptive fields, so that information obtained by the different receptive fields is added, further, extraction of fault characteristic information is enhanced, and characteristic information extraction efficiency is improved. By stacking the structures multiple times, the network is enabled to learn enough characteristic information and strengthen the nonlinearity of the network. Converting into a one-way structure in an inference network, redistributing the parameters of the branches to the main branches, merging weights, and performing inference calculation by taking 3×3 main branch convolution, so that the model achieves high-performance inference efficiency, and reduces the memory occupancy rate. Where receptive field refers to a single-step calculated range of image values for the network model.
The optimization based on the feature fusion network is as follows:
the importance of each neuron is firstly evaluated, the neurons with rich information usually show different point-placing modes from the surrounding neurons, the important neurons are searched by measuring the linear separability among the neurons, the neurons with key characteristics are given higher weight, and the surrounding interference information is suppressed. The loss of effective features is reduced, so that feature weights are evaluated more comprehensively and efficiently, and the network retains position information with few features in the image.
Further improvements to redundant prediction block suppression are:
the method comprises the steps of replacing an original NMS with a Soft-NMS algorithm to filter candidate frames, performing frame inhibition on original confidence scores in the algorithm execution process according to the coincidence degree between the scores and the frames, and reducing rejection of effective fault detection frames by reducing the confidence scores under the condition that a plurality of anti-loosening iron wires are gathered. The algorithm comprises the following steps:
selecting the frame with the highest probability from the detection frames to add into the candidate set, and removing the detection frame from the detection frame set;
calculating IoU of the rest detection frames and selecting the detection frames in the first step, and if IoU of the detection frames and the detection frames are larger than a specified threshold value, updating the probability of the detection frames to be s (1-IoU); the probability of the kuze is still s. If the probability of the detection frame is smaller than a certain set threshold value, searching for the detection frame from the set and deleting the detection frame;
repeating the steps 1 and 2 until the detection frame set is empty.
The inhibition expression of Soft-NMS is:
where si represents the class probability corresponding to the first detection frame; m, a detection frame with the maximum current probability; nt—inhibition threshold; bi-ith detection box to be filtered.
And identifying the latest passing image, loading the image to the local, and calling an identification model by a starting program to automatically detect the image. In the detection process, when the model algorithm judges that the anti-loosening iron wire breaks down, corresponding alarm log information is generated.
While the invention has been described in detail in the foregoing general description and specific examples, it will be apparent to those skilled in the art that modifications and improvements can be made thereto. Accordingly, such modifications or improvements may be made without departing from the spirit of the invention and are intended to be within the scope of the invention as claimed.

Claims (10)

1. An intelligent image analysis method for train vehicle anti-loose iron wire fracture faults comprises the following steps:
step 1, acquiring image data containing anti-loosening iron wires on a train, marking the anti-loosening iron wires to be detected in the acquired image data, and dividing the anti-loosening iron wire image data and corresponding marking files into a training set, a verification set and a test set;
step 2, performing data enhancement operation, configuring super parameters required by an algorithm, and inputting a training sample into an improved yolov7 model to perform feature learning;
and step 3, inputting a new passing image into the learned feature model for target recognition, framing out a failed region and outputting corresponding alarm information.
2. The intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault-like faults according to claim 1, wherein the specific flow for acquiring the image data of the anti-loosening iron wire contained on the train is as follows: when passing by the track side equipment, the train triggers the sensor to receive signals, the high-speed linear array scanning equipment acquires images of the whole train in operation, the images of the whole train are divided into high-definition images with the same resolution through cutting, the high-definition images are stored in the server, and the images containing the anti-loosening iron wires are selected out to serve as data to be marked.
3. The intelligent image analysis method for the breakage of the anti-loose iron wires of the train vehicles according to claim 2, wherein the specific process for marking the anti-loose iron wires is as follows: manually labeling the cut high-definition image by means of a LabelImg tool, wherein labeling comprises the steps of drawing a 2D surrounding rectangular frame for an anti-loosening iron wire to be detected, acquiring position information of the anti-loosening iron wire in a current labeling image, defining category information to which the anti-loosening iron wire belongs, and generating an xml file, wherein the content of the xml file comprises category information, x-axis and y-axis coordinates of the upper left corner and x-axis and y-axis coordinates of the lower right corner of the rectangular frame, and the category information comprises normal and/or broken conditions.
4. The intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault class according to claim 3, wherein the overall data is divided into a training set, a verification set and a test set according to the proportion of 7:2:1 by randomly dividing the marked anti-loosening iron wire sample data set; the training set is used for learning fault and non-fault characteristics of the model, the verification set is used for adjusting and optimizing weight parameters of the network model, and the test set is used for checking generalization capability of the model.
5. The intelligent image analysis method for train vehicle anti-loose iron wire breakage fault-like faults according to claim 1, wherein the specific flow of data enhancement comprises the following steps: randomly extracting 1/3 of the images in the training sample to perform rotation, contrast enhancement, mosaics enhancement and color space enhancement; providing abundant image background and fault morphology for the sample by adopting a supervised image processing method, amplifying the number of targets in the image, and enhancing the characteristic information of the anti-loosening iron wire;
the configuration of the super-parameters comprises: input image size, pre-training model, learning rate, momentum gradient descent method, training times and batch processing picture number.
6. The intelligent image analysis method for train vehicle anti-loose iron wire breakage fault-like faults according to claim 1, wherein basic composition modules of an integral network framework of the improved yolov7 model comprise: an input layer, a feature extraction network, a path aggregation feature pyramid network and a prediction output layer; the feature extraction network includes a CBS module, an ELAN structure, and an MP structure layer.
7. The intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault class according to claim 6, wherein the path aggregation feature pyramid network structure is characterized in that the last three feature layers of a main feature extraction network are used for transmitting deep information into a shallow layer by introducing a top-down path, so that features of different layers are fused efficiently, and feature semantic information is enhanced; and then the feature information between the network layers is fused deeply by downsampling and deep features.
8. The intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault type faults according to claim 6 is characterized in that the prediction output layer is used for adjusting the number of image channels of three characteristic layers of different scales of P3, P4 and P5 output by a characteristic pyramid network through a RepConv module structure, and finally confidence, category and anchor frame offset information are predicted through 1X 1 convolution.
9. The intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault according to claim 7, wherein the yolov7 improved integral network framework is characterized in that the trunk feature extraction network is replaced by a RepConv structure, a multi-branch structure is adopted during model training, different receptive fields are obtained through different convolution kernels, different receptive field information is added, the capability of extracting feature information is further enhanced, and parameters are reconstructed during reasoning by the model; based on the feature fusion network for optimization, a SimAM attention mechanism is embedded before the path aggregation feature pyramid network structure, the importance of each neuron is evaluated, important neurons are searched by measuring the linear separability among the neurons, the neurons with key features are given higher weight, and surrounding interference information is suppressed.
10. The intelligent image analysis method for train vehicle anti-loose iron wire breakage fault type according to claim 9, wherein the yolov7 improved integral network frame further comprises the steps of suppressing redundant prediction frames, replacing original NMS with Soft-NMS algorithm to filter candidate frames, performing function operation on original confidence scores in the algorithm execution process, performing frame suppression according to the coincidence degree between the scores and the frames, and reducing rejection of effective fault detection frames by reducing confidence scores in the case that a plurality of anti-loose iron wires are clustered.
CN202310222178.XA 2023-03-09 2023-03-09 Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault Pending CN116453040A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310222178.XA CN116453040A (en) 2023-03-09 2023-03-09 Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310222178.XA CN116453040A (en) 2023-03-09 2023-03-09 Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault

Publications (1)

Publication Number Publication Date
CN116453040A true CN116453040A (en) 2023-07-18

Family

ID=87131073

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310222178.XA Pending CN116453040A (en) 2023-03-09 2023-03-09 Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault

Country Status (1)

Country Link
CN (1) CN116453040A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117671243A (en) * 2023-12-07 2024-03-08 百鸟数据科技(北京)有限责任公司 Small target detection method, device, computer equipment and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117671243A (en) * 2023-12-07 2024-03-08 百鸟数据科技(北京)有限责任公司 Small target detection method, device, computer equipment and storage medium

Similar Documents

Publication Publication Date Title
CN111815605B (en) Sleeper defect detection method based on step-by-step deep learning and storage medium
CN111680613B (en) Method for detecting falling behavior of escalator passengers in real time
CN111199257A (en) Fault diagnosis method and device for high-speed rail driving equipment
CN111046583A (en) Switch machine fault diagnosis method based on DTW algorithm and ResNet network
CN108764169A (en) A kind of driver's Emotion identification based on machine learning and display device and method
CN116453040A (en) Intelligent image analysis method for train vehicle anti-loosening iron wire fracture fault
CN106570444A (en) On-board smart prompting method and system based on behavior identification
CN112818871A (en) Target detection method of full-fusion neural network based on half-packet convolution
CN115272850A (en) Railway wagon BAB type brake adjuster pull rod head breaking fault identification method
Zhou et al. Fault detection of train height valve based on nanodet-resnet101
CN111563468A (en) Driver abnormal behavior detection method based on attention of neural network
CN114743099A (en) Method and device for identifying faults of train key components
CN118015562A (en) Method and system for extracting key frames of traffic accident monitoring video in severe weather
CN111027859B (en) Driving risk prevention method and system based on motor vehicle state monitoring data mining
CN113343926A (en) Driver fatigue detection method based on convolutional neural network
CN110502995A (en) Driver based on subtle facial action recognition yawns detection method
US11954955B2 (en) Method and system for collecting and monitoring vehicle status information
CN112633163B (en) Detection method for realizing illegal operation vehicle detection based on machine learning algorithm
CN114758326A (en) Real-time traffic post working behavior state detection system
CN114858455A (en) Rolling bearing fault diagnosis method and system based on improved GAN-OSNet
CN114495160A (en) Pedestrian detection method and system based on improved RFBNet algorithm
CN114372556A (en) Driving danger scene identification method based on lightweight multi-modal neural network
CN114244596A (en) Vehicle-mounted CAN (controller area network) anomaly detection method and system based on HTM (hyper text transport protocol)
CN112766174A (en) Railway train carriage group bottom plate loss fault detection method
TWI762375B (en) Semantic segmentation failure detection system

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