CN113591270B - MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set - Google Patents
MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set Download PDFInfo
- Publication number
- CN113591270B CN113591270B CN202110728637.2A CN202110728637A CN113591270B CN 113591270 B CN113591270 B CN 113591270B CN 202110728637 A CN202110728637 A CN 202110728637A CN 113591270 B CN113591270 B CN 113591270B
- Authority
- CN
- China
- Prior art keywords
- model
- crack
- index
- stage
- rail weld
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000012544 monitoring process Methods 0.000 title claims abstract description 41
- 229910000831 Steel Inorganic materials 0.000 title claims abstract description 38
- 239000010959 steel Substances 0.000 title claims abstract description 38
- 238000004458 analytical method Methods 0.000 title claims abstract description 16
- 230000006378 damage Effects 0.000 claims abstract description 68
- 238000003466 welding Methods 0.000 claims abstract description 32
- 235000019687 Lamb Nutrition 0.000 claims abstract description 26
- 238000012549 training Methods 0.000 claims abstract description 11
- 230000000694 effects Effects 0.000 claims description 40
- 238000004422 calculation algorithm Methods 0.000 claims description 18
- 238000004364 calculation method Methods 0.000 claims description 15
- 238000011156 evaluation Methods 0.000 claims description 15
- 238000013135 deep learning Methods 0.000 claims description 13
- 238000013461 design Methods 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 230000015556 catabolic process Effects 0.000 claims description 8
- 238000006731 degradation reaction Methods 0.000 claims description 8
- 230000003902 lesion Effects 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 230000004807 localization Effects 0.000 claims description 4
- 238000013526 transfer learning Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 239000013256 coordination polymer Substances 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000036541 health Effects 0.000 abstract description 9
- 238000000354 decomposition reaction Methods 0.000 description 21
- 238000001514 detection method Methods 0.000 description 10
- 238000012360 testing method Methods 0.000 description 10
- 238000009826 distribution Methods 0.000 description 8
- 238000005457 optimization Methods 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 5
- 238000013459 approach Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 238000003672 processing method Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 239000006247 magnetic powder Substances 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000035515 penetration Effects 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 230000006735 deficit Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000012014 frustrated Lewis pair Substances 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 229910052751 metal Inorganic materials 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 150000002739 metals Chemical class 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000009659 non-destructive testing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000000523 sample Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06N3/045—Combinations of networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
The invention discloses a monitoring method for performance analysis and comparison on a steel rail weld data set based on MDCD, which comprises the following steps: s1: analyzing Lamb wave structures in the steel rail welding seams; s2: monitoring Lamb wave data characteristics on rail weld crack damage; s3: building and training an MDCD model for monitoring the health of a steel rail weld joint structure; s4: the method and the device can effectively detect crack damage on the weld joint structure, simultaneously have lower propagation damage, and ensure that Lamb waves can monitor the weld joint structure within a certain distance and range.
Description
Technical Field
The invention relates to the technical field of nondestructive testing, in particular to a monitoring method for performance analysis and comparison on a steel rail weld data set based on MDCD.
Background
The rail weld joint is used as a component for bearing the continuity of two modules in the rail, and the health state of the rail weld joint is closely related to the safety of railway transportation. In the railway transportation production operation process, the shape of the manual welding part of the steel rail welding line is irregular, so that when a train passes through, the force generated by the wheel-rail relationship cannot be uniformly conducted and dispersed at the welding line, and the position stress of the steel rail welding line part is serious. Under the condition that a long-time train passes through the high-frequency train, vibration fatigue generated by the wheel-rail relationship can directly lead to fatigue crack damage at the stressed serious part of the steel rail weld joint, and the rail breakage phenomenon finally occurs along with deepening of the crack damage.
Because of the design characteristics of the railway steel rail, the rail web part is not stressed obviously, crack fatigue damage is not easy to occur, and the cracks of the steel rail welding seam in the current railway transportation network are mostly occurred at the welding root parts at the two ends of the welding seam, and are usually vertical cracks. Along with the increase of fatigue in the use process of the steel rail, cracks are in different lengths and depths under different conditions. According to the analysis of the structure and the stress condition of the steel rail, the crack occurrence part is generally the rail head bottom surface, the rail bottom top surface and the rail bottom position. The rail bottom is serious in stress condition, once small crack damage occurs, the rail bottom can rapidly evolve into damage degree which can be distinguished by naked eyes, and rail breakage accidents can be caused under severe conditions.
In the field of nondestructive inspection in modern industry, conventional nondestructive inspection methods such as Ultrasonic (UT), magnetic powder (MT), radiation (RT), penetration (PT) and the like are generally adopted for crack monitoring of a welded seam structure. In the selection of the rail weld flaw detection method, ultrasonic detection is mainly adopted in each country.
In order to ensure the detectability of the rail welding seam crack, the conventional nondestructive ultrasonic detection method needs to slide on the rail continuously by means of equipment such as a flaw detector car and a flaw detector to realize crack damage, so that for the rail welding seam with a fixed position and a smaller range, the crack detection can be realized by sliding on the rail welding seam continuously, the ultrasonic probe can not be solidified to the periphery of the rail welding seam by moving the detectability, so that the continuous structural health monitoring of the rail welding seam is realized, and each detection consumes manpower and material resources. Therefore, the real-time, accurate and efficient rail weld crack damage and structural health monitoring technology aiming at the rail weld and capable of detecting small cracks has important significance for guaranteeing railway transportation safety.
Structural health monitoring is a proposal provided for a specific structure, which can monitor structural integrity continuously and automatically for a long time, discover and locate defects, monitor the change condition of existing defects, and even estimate the service life of the structure. Lamb waves can realize long-distance detection by virtue of the characteristic of small energy loss when the Lamb waves propagate along the surface of a medium, and have high sensitivity to tiny defects in a structure, so that the Lamb waves are widely applied to structural health monitoring (Structure Health Monitoring, SHM) of metals. For structural health monitoring of steel rail welding seams, lamb waves generally have larger activation frequency and shorter wavelength, so that excellent damage detection resolution of the Lamb waves is given, crack damage on the welding seam structure can be effectively detected, meanwhile, lower propagation damage of the Lamb waves is guaranteed, the Lamb waves can monitor the welding seam structure within a certain distance and range, and the Lamb waves are proved to have important significance for structural health monitoring of the steel rail welding seams.
Disclosure of Invention
The present invention aims to solve the above technical problems to a certain extent.
In view of the above, the invention provides a monitoring method for performance analysis and comparison on a steel rail weld data set based on MDCD, which can meet the requirement for monitoring the steel rail weld crack to a certain extent under the condition of ensuring the running speed.
In order to solve the technical problems, the invention provides a monitoring method for analyzing and comparing performance of MDCD (MDCD) -based steel rail weld data set, which comprises the following steps: s1: analyzing Lamb wave structures in the steel rail welding seams; s2: monitoring Lamb wave data characteristics on rail weld crack damage; s3: the method comprises the steps of structurally designing a deep learning network model of a generator and a discriminator of an countermeasure generation network, wherein the first stage is based on a deep learning neural network which is complex in design structure and more in parameters and can effectively extract and process data characteristics so as to meet the requirement on deep characteristic extraction of input data, and the second stage is based on a teacher and student theory in transfer learning to design a network structure model with simpler design, so that the characteristic extraction capacity and the appointed data output capacity of the first stage are learned; s4: and evaluating the model training effect of the two stages, and comparing the performance degradation rate of the second stage model relative to the first stage model. The execution effect of the two-stage training model on different tasks is compared with the current advanced and popular deep learning algorithm and the traditional digital signal processing algorithm.
Further, the step S4 includes: s41: and defining the result evaluation indexes of the two-stage model after training, ensuring that the indexes can fully reflect the performance of the model on the related rail weld crack damage monitoring task, and displaying each index of the two-stage model.
Further, the step S4 further includes: s42: and (3) comparing the cost performance of two stage models referencing different Efficient model architectures to obtain the conclusion that the comprehensive cost performance of the MDDs 2-Eb3 model is highest in the process of executing rail weld Lamb wave modal decomposition and damage monitoring tasks.
Further, the step S4 further includes: s43: the two-stage models of different Efficient model architectures are compared and analyzed with the current advanced and popular deep learning and traditional digital signal processing methods, evaluation indexes of the models among different tasks are analyzed, and the effects that MDDs 1-Eb7 can reach advanced accuracy in each task and MDDs 2-Eb3 can approach to the advanced models in each task on the premise of ensuring the running efficiency are obtained.
Further, the step S4 further includes: s44: the mode generation effect of the model is visually displayed, the mode decomposition effect between the MDCDs2-Eb3 and the advanced method is visually compared, and the MDCDs2-Eb3 has the effect of the advanced model, but still has an optimization space in part of the phase intervals of the mode generation result.
Further, the step S4 further includes: s45: the visual display is carried out on the crack positioning effect of the model, the crack positioning effect between MDCD2-Eb3 and an advanced method is compared visually, meanwhile, fitting description is carried out on an approximate crack width curve after the crack is positioned, and the fitting result of MDCD2-Eb3 and the advanced model in crack width prediction is approximately displayed, so that the MDCD2-Eb3 can effectively complete the crack damage positioning task.
Further, the step S4 further includes: s46: and comparing the crack existence probability and the crack depth prediction result between the MDCDs2-Eb3 and the advanced method.
Further, the step S1 further includes: s14: the feasibility of the model in transfer learning between similar tasks is explored, and a method for compressing model quantity under the condition of ensuring model accuracy is discussed under the condition of building an initial complex model.
The invention has the technical effects that: the cost performance between two stage models of different EfficientNet model architectures is referred to for calculation comparison, and the conclusion that the comprehensive cost performance of the MDDs 2-Eb3 model is highest in the process of executing rail welding seam Lamb wave modal decomposition and damage monitoring tasks is obtained; comparing and analyzing two-stage models of different Efficient model architectures with the current advanced and popular deep learning and traditional digital signal processing methods, and analyzing evaluation indexes of the models among different tasks to obtain the effect that MDDs 1-Eb7 can reach advanced accuracy in each task and MDDs 2-Eb3 can approach to the advanced model in each task on the premise of ensuring the running efficiency; the mode generation effect of the model is visually displayed, the mode decomposition effect between the MDCDs2-Eb3 and the advanced method is visually compared, and the MDCDs2-Eb3 has the effect of the advanced model, but still has an optimization space in part of the phase intervals of the mode generation result; the visual display is carried out on the crack positioning effect of the model, the crack positioning effect between MDCD2-Eb3 and an advanced method is visually compared, meanwhile, fitting description is carried out on an approximate crack width curve after the crack is positioned, the fitting result of MDCD2-Eb3 and the advanced model in crack width prediction is approximately displayed, and the fact that MDCD2-Eb3 can effectively complete the crack damage positioning task is shown, but certain deviation still exists for the prediction of width distribution; compared with the advanced model pre-algorithm, the MDCDs2-Eb3 has excellent capability in the task of predicting the crack existence probability and the crack depth compared with the advanced model pre-algorithm by comparing the crack existence probability and the crack depth prediction result between the MDCDs2-Eb3 and the advanced method. The comparison result shows that the model has certain task execution capacity, can meet the requirement for monitoring the rail weld crack to a certain extent under the condition of ensuring the running speed, but still has certain error and optimization space under partial conditions.
Drawings
FIG. 1 is a flow chart of a MDCD-based monitoring method for performance analysis comparison on a rail weld data set in accordance with the present invention;
FIG. 2 is a model of running time at different GPUs according to the present invention;
FIG. 3 is the same parameter but different GPU runtime according to the present invention;
FIG. 4 is a graph of model trend of different indicators according to the present invention;
FIG. 5 is a parameter-index trend graph according to the present invention;
FIG. 6 is a graph comparing the results of the decomposition of the ASD and VMD modes with the MDDs 2-Eb3 signal modes according to the present invention;
FIG. 7 is a graph comparing the width predictions between MDCDs2-Eb3 and the executable crack locating task model and algorithm in accordance with the present invention;
FIG. 8 is a graph comparing depth predictions between MDCDs2-Eb3 and an executable crack locating task model and algorithm in accordance with the present invention.
Detailed Description
The present invention will be further described with reference to the accompanying drawings and specific examples, which are not intended to limit the invention, so that those skilled in the art may better understand the invention and practice it.
A monitoring method for performance analysis and comparison on a steel rail weld data set based on MDCD comprises the following steps: s1: analyzing Lamb wave structures in the steel rail welding seams; s2: monitoring Lamb wave data characteristics on rail weld crack damage; s3: the method comprises the steps of structurally designing a deep learning network model of a generator and a discriminator of an countermeasure generation network, wherein the first stage is based on a deep learning neural network which is complex in design structure and more in parameters and can effectively extract and process data characteristics so as to meet the requirement on deep characteristic extraction of input data, and the second stage is based on a teacher and student theory in transfer learning to design a network structure model with simpler design, so that the characteristic extraction capacity and the appointed data output capacity of the first stage are learned; s4: and evaluating the model training effect of the two stages, and comparing the performance degradation rate of the second stage model relative to the first stage model. The execution effect of the two-stage training model on different tasks is compared with the current advanced and popular deep learning algorithm and the traditional digital signal processing algorithm.
The step S4 includes: s41: defining the result evaluation indexes of the two-stage models after training, ensuring that the indexes can fully reflect the performance of the models on the related rail weld crack damage monitoring task, and displaying all indexes of the two-stage models; s42: the cost performance between two stage models referring to different Efficient model architectures is calculated and compared, and a conclusion that the comprehensive cost performance of the MDDs 2-Eb3 model is highest in the process of executing rail welding seam Lamb wave modal decomposition and damage monitoring tasks is obtained; s43: comparing and analyzing two-stage models of different Efficient model architectures with the current advanced and popular deep learning and traditional digital signal processing methods, and analyzing evaluation indexes of the models among different tasks to obtain the effect that MDDs 1-Eb7 can reach advanced accuracy in each task and MDDs 2-Eb3 can approach to the advanced model in each task on the premise of ensuring the running efficiency; s44: the mode generation effect of the model is visually displayed, the mode decomposition effect between the MDCDs2-Eb3 and the advanced method is visually compared, and the MDCDs2-Eb3 has the effect of the advanced model, but still has an optimization space in part of the phase intervals of the mode generation result; s45: the visual display is carried out on the crack positioning effect of the model, the crack positioning effect between MDCD2-Eb3 and an advanced method is visually compared, meanwhile, fitting description is carried out on an approximate crack width curve after the crack is positioned, the fitting result of MDCD2-Eb3 and the advanced model in crack width prediction is approximately displayed, and the MDCD2-Eb3 can effectively complete the crack damage positioning task; s46: and comparing the crack existence probability and the crack depth prediction result between the MDCDs2-Eb3 and the advanced method.
According to the specific embodiment of the invention, the performance comparison analysis of the MDCD model on the rail weld crack damage monitoring data and the performance comparison analysis are beneficial to exploring the execution capacity of the trained model in various tasks, and the performance and effect of the model are intuitively displayed by indexes. Meanwhile, the effect of the model and the current advanced algorithm on the same task is compared, so that the difference between the task execution capacity of the model and the front edge algorithm is facilitated to be judged, and a certain basis is provided for further development and improvement of the follow-up model.
The analysis of the rail weld crack damage model indexes of the first stage and the second stage is to evaluate and check whether the model can complete the designated rail weld crack damage monitoring task according to the expected, so the task completion degree evaluation indexes of the model are described as follows:
aiming at the evaluation index of the model Lamb wave modal decomposition result, the similarity between the model generation modal component and the actual modal component is checked;
aiming at the evaluation index of the rail weld crack positioning probability matrix label, the positioning accuracy of the model on the crack position on the rail weld with crack damage is checked, wherein the positioning accuracy comprises crack length accuracy, crack position accuracy and crack width accuracy;
aiming at the evaluation index of the occurrence probability of the steel rail weld cracks, checking the recall ratio of the number of the cracks and the existence of crack damages by the model under the condition that the cracks exist;
aiming at the evaluation index of the depth of the steel rail weld joint crack, the prediction accuracy of the model on the depth of the crack is checked under the condition that the crack exists.
There are therefore four criteria for model evaluation: model impairment classification accuracy index (Classification Accuracy Index, CAI), model depth prediction accuracy index (Depth Accuracy Index, DAI), localization accuracy index (Localization Accuracy Index, LAI), modal component accuracy index (IMFs Accuracy Index, IMFAI).
(1) Damage classification accuracy index
Aiming at the steel rail welding seam damage classification, the probability of the welding seam crack damage and the probability of the welding seam without the crack damage are generated, and the sum result between the probability and the probability is 1. Aiming at a steel rail weld crack monitoring task, the principle is that the fault detection is rather than the missed detection, so that the damage classification of the weld crack is realized, and the Recall rate is adopted to judge the performance effect of the model on the task.
The recall parameter represents the sensitivity of the model to rail weld cracks, and the higher the recall, the more sensitive the model to the weld cracks. The recall ratio is calculated as follows:
in the formula, phi is a trained rail weld crack monitoring model;
Φ TP ，Φ FN -phi, as shown in Table 5-1 below TP Representing the probability that the steel rail welding seam has cracks and predicting the steel rail welding seam to have the cracks by the model; phi FN The probability that the steel rail welding seam has no crack and the model predicts the steel rail welding seam has the crack is represented; phi FP Representing the probability of the model predicting that the rail weld has no cracks when the rail weld has cracks; phi TN And (5) representing the probability that the steel rail weld joint does not have cracks and predicting the steel rail weld joint by the model.
(2) Depth similarity index
Aiming at the similarity index of the depth of the steel rail weld, the method is used for judging the difference between the obtained weld crack depth and the actual crack depth when the model has crack damage to the steel rail weld. I.e. whether the model can exhibit smaller prediction errors. Meanwhile, when no crack exists in the rail weld, the prediction of the crack depth of the rail weld for the part of the model should be close to 0. According to the function executed by the index, the depth similarity index mathematical expression can be designed as follows:
wherein S is a test data set;
abs—taking absolute value to calculate;
p (|S) -the proportion of data in the test set state is · can be damage or damage;
depth pre -model predicting the depth of the lesion;
depth gt -actual lesion depth in the presence of lesions;
depth max -maximum lesion depth in the presence of lesions in the dataset;
phi-the trained rail weld crack monitoring model.
The calculation of the model depth similarity index in the formula (5-2) is divided into three parts, and when the model predicted damage type is lossless and the model predicted damage type is the same as the actual situation, DAI is 0; when the model predicts the damage and is the same as the actual situation, DAI is the proportion of damage data to the total data set multiplied by the proportion of the absolute value of the difference value between the predicted damage depth and the actual damage depth to the actual damage depth; when the model predicted damage type is different from the actual situation, DAI is the predicted damage type data, which is the ratio of damage data to the total data set multiplied by the ratio of predicted damage depth to the maximum damage depth in the data set.
The DAI evaluation index can effectively flatten the depth prediction error into each of the same type of data. Meanwhile, the DAI index is combined with the damage classification error of the model, the probability of the data type in the data set is used as the weight of the crack depth error under different prediction results of the model, the error precision of the crack depth prediction under different types is reasonably shared, and the influence degree of the damage depth error value under the condition of damage type prediction error on the DAI index is increased. The DAI shows the stability of the model for crack damage prediction, and the smaller the DAI is, the more accurate and stable the model for crack damage prediction is proved.
(3) Index of positioning accuracy
Aiming at the positioning accuracy index of the steel rail welding seam, analysis should be carried out from three aspects of crack length, width and positioning accuracy. And (3) regarding the crack length and width evaluation index, the same as the DAI, multiplying the ratio of the crack length and width predicted value to the actual crack length and width value by the ratio of corresponding crack state data in the training set. Aiming at judging the crack positioning precision, calculating Euclidean distance by utilizing three points of coordinates of two ends and coordinates of a central point of the predicted crack and coordinates of three points of the crack in actual conditions, and dividing the calculated result by the distance between the longest two points in a 144 multiplied by 72 rectangular range of sensor layout, namelyAnd meanwhile, referring to a DAI calculation weight distribution mode, taking the proportion of the damage type data in the test set as a weight, and carrying out weighted calculation on the three-point Euclidean distance result. The LAI calculation formula obtained by analysis is as follows:
in the method, in the process of the invention,predicting a coordinate point space formed by two ends of the crack position and the center;
P pre -model prediction point coordinates;
P gt -real point coordinates;
-with an index weight space in the DAI;
eu (·) Eu-Euclidean distance calculation;
phi-the trained rail weld crack monitoring model.
The LAI index shows inaccuracy of the model on the damage positioning capability, and according to the calculation process of the formula (5-3), the larger the error is, the larger the LAI index is, which indicates that the model is inaccurate in positioning crack damage.
(4) Modal component accuracy index
Aiming at the accuracy index of the Lamb wave modal component, the essence is to measure the information similarity between two groups of time series data, and a dynamic time planning algorithm (Dynamic Time Warping, DTW) is introduced to the calculation of the accuracy index of the modal component through the research of T Rakthanmanon et al on the similarity of the time series data. The DTW compares the similarity of the two sequences into a typical function optimization problem, and the algorithm uses a time normalization function W (n) meeting a certain condition to describe the time correspondence between the test sequence and the reference sequence, so as to solve the normalization function corresponding to the minimum accumulated distance when the two sequences are matched. When the durations of the same salient features of the two sequences are different, the DTW can effectively find the data points of the two time sequences, align the data and then calculate the euclidean distance between the sequences.
Therefore, the IMFAI calculation formula after DTW is introduced is as follows:
in the method, in the process of the invention,-a data space consisting of three modal components after single-channel Lamb wave data modal decomposition of the steel rail weld;
imf pred -model prediction of modal components;
imf gt -an actual modal component;
phi-the trained rail weld crack monitoring model.
The IMFAI evaluates the similarity between three modal components and real modal components generated by the model, reflects the modal decomposition capacity of the model on rail weld Lamb wave multi-modal superposition data, and the larger the IMFAI is, the stronger the modal decomposition capacity of the model is.
The MDDs 2 student model has inferior performance in various indexes compared with the MDDs 1 teacher model due to simpler model structure and model parameter scale. MDDs 1-Eb7 obtain the best parameter results in the MDDs 1 model due to its strong parameter quantity and model fitting capability; the MDCDs2-Eb3 in the MDCDs2 has optimal effects in both CAIrecall, DAI and IMFAI, and shows relatively more balanced and excellent parameter results.
In order to evaluate the performance degradation rate indexes of MDDs 2 when different teacher models are learned, the design provides a model performance degradation rate calculation formula as follows:
wherein, D is a performance decay rate index;
Φ * -a trained MDCDs1 or MDCDs2 model;
CAI * model-CAI index;
DAI * model-DAI index;
LAI * model-LAI index;
IMFAI * IMFAI index of model x;
the decay rate of the MDCDs2 model compared with the MDCDs1 model under four tasks can be obtained through the calculation of the formulas (5-5) to (5-8). For the calculation formula of the decay rate index, the numerator is the first stage model index minus the second stage model index, and the denominator is the index with larger ratio before and after decay. Tables 5-3 show the performance degradation rates of MDCDs2 relative to MDCDs1 for different indices in the real data test set for different EfficientNet model architectures for MDCDs 2:
with the improvement of the architecture complexity and the parameter quantity of the model based on the Efficient, the task execution capacity of the MDDs 1 is gradually improved, and various indexes can prove that the MDDs 1 model can show better learning effect with the improvement of the model complexity. However, as the scale and the parameter number of the MDCDs2 model are unchanged, as the scale of the MDCDs1 is increased, the fitting effect of the MDCDs1 model on the Lamb wave data of the rail weld seam gradually reaches the upper limit of the fitting capability of the MDCDs2 model and exceeds the fitting capability of the MDCDs2, at this time, compared with Eb1, eb3, eb5 and Eb7, when the network model architecture is extracted from the characteristics of the teacher model MDCDs1 and refers to Eb3, the student model MDCDs2 can achieve the best learning effect, the model capability degradation rate is the lowest in the four indexes, and when the scale of the MDCDs1 model is continuously increased, the data characteristic distribution rule learned by the teacher model MDCDs1 can not be understood by the student model MDCDs2 any more, so that the situation of the sudden increase of the learning capability degradation rate of the MDCDs2 model occurs, and the task execution effect of the MDCDs2 is reduced.
MDCD operation efficiency and cost performance index analysis
The MDCDs1 model is based on different EfficientNet model architectures, and the model NoP index, FLPs index and the running speed of the model under GPU NVIDIA Tesla P and GPU NVIDIA GeForce 1060 aiming at single batch input of a real data test set. As model NoP increases, model fleps and run time also increase and run efficiency decreases.
As shown in fig. 2, 3 and 4, the running time of different models under two GPUs, the running time of models with the same parameters under different GPUs, the four index change trends of different models, and the four index change trends of different parameter models are visualized. In the whole, the model operation efficiency gradually decreases along with the increase of the model scale, and meanwhile, when the model scale increases to 3.0M of MDDs 2-Eb3 along with the increase of the parameters, each index of the model shows stepwise jump growth, so that the maximum data distribution knowledge learned by MDDs 2 per unit parameter can be preliminarily judged.
To accurately judge the learning efficiency of each model, namely: the data distribution quantity learned per unit parameter is designed into an operation cost performance index CP Φ The cost performance of the model is evaluated, and the calculation formula of the running cost performance index is as follows:
in the formula, CAI Φ -an injury classification accuracy index of the model;
DAI Φ -model depth prediction accuracy index of the model;
LAI Φ -a positioning accuracy index of the model;
IMFAI Φ -a modal component accuracy index of the model;
phi-the trained rail weld crack monitoring model;
Params Φ model parameters, unit: millions of people;
Params max -maximum model parameters in all models, unit: millions of people.
The cost performance of the model is calculated by adopting the method, so that the cost performance index of the model is always in nonlinear negative correlation with the parameter quantity of the model, and is in linear positive correlation with the CAI and IMFAI indexes of the model, and is in linear negative correlation with the DAI and LAI indexes of the model. Simultaneous Params max The data scale of each index influencing the model cost performance index is balanced, so that the CP cost performance index can nonlinearly show the cost performance relationship between different parameter models and four output indexes in the real data test set of the rail weld crack damage.
When the teacher model MDCDs1 adopts an EfficientNet-b3 architecture, the student model MDCDs2 can exert the parameter fitting capability to the maximum extent, and the best model is obtained through learning, wherein the maximum cost performance index is 0.89. In the whole, the MDCDs2 model has higher cost performance indexes when facing different teacher models, so that the compression strategy of the teacher-student model can effectively improve the overall execution efficiency of the model when aiming at rail weld Lamb wave modal decomposition and rail weld crack damage monitoring tasks. When the model is deployed in a production environment, the cost performance evaluation index calculation method can be modified according to the production environment requirement to select the most suitable model.
MDDs 1 and MDDs 2 are compared with indexes of the current advanced deep learning and traditional digital signal processing method on the related tasks of the rail weld crack damage real data set. From tables 5-6, the MDCDs1-Eb7 model shows the effect similar to the precision and accuracy of the current advanced method in each task, but the parameter quantity of the MDCDs1-Eb7 is too large, the cost performance is too low, and the MDCDs1-Eb7 model is not suitable for being applied to scenes with higher requirements on real-time performance; the MDCDs2-Eb3 achieves the effect close to the current common algorithm under the condition of using fewer parameters, can simultaneously execute 4 tasks, has high operation efficiency, and is suitable for being applied to scenes with high real-time requirements.
And performing effect display on the model MDCDs2-Eb3 with the highest cost performance index, selecting a group of data in the test set, and observing and comparing the difference between the real data in the test set and the model output data.
FIG. 6 shows graphs comparing the ASD and VMD methods of tables 5-6 with the results of the MDCDs2-Eb3 signal modality decomposition. The left column shows the results of modal decomposition of the three methods, and the right column shows the absolute value errors of the results of modal decomposition of the three methods and the results of modal decomposition of the data of the rail weld crack damage test set. As can be seen from the left column, the visual observation effect of the VMD in the result prediction is poor, the method gathers the main modal information in the IMF1 for the modal decomposition of the signal, the dispersion in the IMF2 is less, and the IMF3 is a noise signal; by observing the absolute errors of the modal decomposition result of the third line IMF3 and the modal decomposition of IMF3, it can be seen that the ASD has larger noise information doped with the modal decomposition effect, and the modal decomposition effects of IMF1 and IMF2 are similar to MDCDs2-Eb 3.
The model is trained in advance by the rail weld crack damage simulation data, so that the model learns more sufficient crack damage position distribution, and the real rail weld crack damage data is mainly distributed at four positions of the rail weld, so that the real data is learned to knowledge from simulation data to be adjusted, and the model is restrained to identify that cracks mainly appear at the positions of the root of the weld, but the cracks still cannot be restrained completely to appear on the ground of the rail top or the top surface of the rail bottom. This phenomenon can be avoided by enriching the real data set. Therefore, the model is relatively accurate in predicting the crack position when the crack exists, and the model has high accuracy in predicting the position where the crack exists truly although the model has small deviation and false detection in predicting the crack position.
The method comprises the steps of comparing MDCDs2-Eb3 with advanced models and algorithms capable of executing crack positioning tasks to predict cracks at the rail top and bottom of the root of a weld joint on the left side of a steel rail, taking the upper left corner as the origin of a coordinate axis, carrying out transverse statistics on the coordinate axis of the width by using probability points with probability values larger than 0.75, taking the widest position of the continuous points as a crack width prediction result, and enabling the crack width to be 4mm in the real case. From the figure, the visual effect on the MDCDs2-Eb3 crack positioning task is close, the central positions are [21,15], and the widths are about 4mm; the prediction result of Model1 has deviation of central position on the length and width coordinate axes, the predicted central position is [25,6], and the width is about 8mm; model2 has a similar effect to MDCDs2-Eb3, with center position [24,18], and width of about 4mm; model3 had an excessively large prediction range for lesion localization, and had a predicted center position of [25,6] and a width of about 10mm.
Fig. 7 shows a width prediction result fitting curve, a scattered point contrast graph (upper left and upper right) and an absolute value error line graph (lower left and lower right) between MDCDs2-Eb3 and an executable rail weld crack damage positioning task model and algorithm, wherein the fitting curve can approximately represent the distribution characteristics of data learned by the model and the relation between crack length and width. The upper left image in the graph draws the comparison between MDDs 2-Eb3 and real data (group Truth), and it can be seen that the MDDs 2-Eb3 do not learn the characteristic distribution rule of the data completely and effectively, and when the crack length is short, the prediction result of the model on the crack width deviates from the real situation, the maximum error can reach 0.73mm, and as the crack length increases, the prediction of the model on the crack width gradually approaches the real situation, and the minimum error is 0.04mm; the upper right image of the graph depicts the comparison curves and scatter plots of the Model 1-Model 3, MDCDs2-Eb3 and the actual conditions in tables 5-6, and the waveform trend and lower right error plot can be seen that the crack length-width curves generated by Model1 are closest to the actual conditions, the minimum error and the maximum error are respectively 0.04mm and 0.71mm, the crack length-width curves generated by Model3 are most different from the actual conditions, and the minimum error and the maximum error are respectively 0.03mm and 0.95mm.
Tables 5-7 show a representative comparison of the predicted and actual results of the model for the presence and depth of a rail weld crack. As can be seen from the table, when the welding seam has cracks, the prediction of the existence probability of the model to the welding seam cracks is relatively accurate, the accuracy of the model to the crack prediction can be increased by properly adjusting the threshold value of the prediction probability as can be seen from the tables 5-5, and the accuracy of the model to the crack prediction is 76%. And carrying out traversal statistics on the depth prediction result, wherein when a crack exists, the model prediction result is relatively accurate, the error range is +/-1.5 mm, and when the crack does not exist, the prediction error of the model on the crack depth result is relatively large, and the error range is 0-3 mm.
The depth prediction results between the MDDs 2-Eb3 and the executable rail weld crack damage depth prediction task model and algorithm fit curves, scatter point contrast graphs (upper left and upper right) and absolute value error line graphs (lower left and lower right). The upper left image in fig. 8 depicts the comparison between MDCDs2-Eb3 and the real data (Ground Truth), and it can be seen that the MDCDs2-Eb3 relatively effectively learns the nonlinear relationship between the crack depth and the length in the real situation, and the crack depth prediction error is 0.01mm and 3.57mm at the minimum and the maximum respectively, and it can be seen through the lower left error graph that as the crack depth increases, the crack depth prediction error of the MDCDs2-Eb3 increases. The upper right image in FIG. 8 depicts the comparative and scatter plots of Model4, model5, MDCDs2-Eb3 and the actual conditions in tables 5-6, and the observed waveform trend shows that the crack length-width curves generated by Model4 are closer to the actual conditions than by Model5, the maximum and minimum errors of Model4 are 0.15mm and 6.38mm, respectively, and the maximum and minimum errors of Model5 are 0.03mm and 6.94mm, respectively.
The above-described embodiments are merely preferred embodiments for fully explaining the present invention, and the scope of the present invention is not limited thereto. Equivalent substitutions and modifications will occur to those skilled in the art based on the present invention, and are intended to be within the scope of the present invention. The protection scope of the invention is subject to the claims.
Claims (2)
1. The monitoring method for performance analysis and comparison on the steel rail weld data set based on MDCD is characterized by comprising the following steps:
s1: analyzing Lamb wave structures in the steel rail welding seams;
s2: monitoring Lamb wave data characteristics on rail weld crack damage;
s3: the method comprises the steps of structurally designing a deep learning network model of a generator and a discriminator of an countermeasure generation network, wherein the first stage is based on a deep learning neural network which is complex in design structure and more in parameters and can effectively extract and process data characteristics so as to meet the requirement on deep characteristic extraction of input data, and the second stage is based on a teacher and student theory in transfer learning to design a network structure model with simpler design, so that the characteristic extraction capacity and the appointed data output capacity of the first stage are learned;
s4: evaluating the model training effects of the two stages, comparing the performance degradation rate of the second-stage model relative to the first-stage model, and simultaneously comparing the execution effects of the two-stage training model on different tasks with the current advanced and popular deep learning algorithm and the traditional digital signal processing algorithm;
in order to accurately judge the learning efficiency of each model, an operation cost performance index CP is designed Φ The cost performance of the model is evaluated, and the calculation formula of the running cost performance index is as follows:
in the formula, CAI Φ -a model lesion classification accuracy CAI index;
DAI Φ -model depth prediction accuracy DAI index of the model;
LAI Φ -a localization accuracy LAI index of the model;
IMFAI Φ -a modal component accuracy IMFAI index of the model;
phi-the trained rail weld crack monitoring model;
Params Φ model parameters, unit: millions of people;
Params max -the largest model of all modelsParameter number, unit: millions of people.
2. The method for monitoring the performance analysis and comparison of MDCD-based on the rail weld data set according to claim 1, wherein the step S4 comprises: s41: and defining the result evaluation indexes of the two-stage model after training, ensuring that the indexes can fully reflect the performance of the model on the related rail weld crack damage monitoring task, and displaying each index of the two-stage model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110728637.2A CN113591270B (en) | 2021-06-29 | 2021-06-29 | MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110728637.2A CN113591270B (en) | 2021-06-29 | 2021-06-29 | MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113591270A CN113591270A (en) | 2021-11-02 |
| CN113591270B true CN113591270B (en) | 2024-01-19 |
Family
ID=78245174
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110728637.2A Active CN113591270B (en) | 2021-06-29 | 2021-06-29 | MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113591270B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175434A (en) * | 2019-06-19 | 2019-08-27 | 西南交通大学 | A kind of rail fastener system injury detection method based on convolutional neural networks |
| CN111402209A (en) * | 2020-03-03 | 2020-07-10 | 广州中国科学院先进技术研究所 | U-Net-based high-speed railway steel rail damage detection method |
| AU2020101011A4 (en) * | 2019-06-26 | 2020-07-23 | Zhejiang University | Method for identifying concrete cracks based on yolov3 deep learning model |
| CN111598845A (en) * | 2020-04-26 | 2020-08-28 | 北京科技大学 | Pavement crack detection and positioning method based on deep learning and NEO-6M positioning module |
| CN112069647A (en) * | 2020-07-20 | 2020-12-11 | 南京航空航天大学 | Model modeling method for describing anisotropic nonlinear mechanical behavior of ceramic-based fiber bundle composite material |
| CN112581430A (en) * | 2020-12-03 | 2021-03-30 | 厦门大学 | Deep learning-based aeroengine nondestructive testing method, device, equipment and storage medium |
| CN113034478A (en) * | 2021-03-31 | 2021-06-25 | 太原科技大学 | Weld defect identification and positioning method and system based on deep learning network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9506836B2 (en) * | 2012-10-09 | 2016-11-29 | The Boeing Company | Methods and systems for structural health monitoring |
-
2021
- 2021-06-29 CN CN202110728637.2A patent/CN113591270B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175434A (en) * | 2019-06-19 | 2019-08-27 | 西南交通大学 | A kind of rail fastener system injury detection method based on convolutional neural networks |
| AU2020101011A4 (en) * | 2019-06-26 | 2020-07-23 | Zhejiang University | Method for identifying concrete cracks based on yolov3 deep learning model |
| CN111402209A (en) * | 2020-03-03 | 2020-07-10 | 广州中国科学院先进技术研究所 | U-Net-based high-speed railway steel rail damage detection method |
| CN111598845A (en) * | 2020-04-26 | 2020-08-28 | 北京科技大学 | Pavement crack detection and positioning method based on deep learning and NEO-6M positioning module |
| CN112069647A (en) * | 2020-07-20 | 2020-12-11 | 南京航空航天大学 | Model modeling method for describing anisotropic nonlinear mechanical behavior of ceramic-based fiber bundle composite material |
| CN112581430A (en) * | 2020-12-03 | 2021-03-30 | 厦门大学 | Deep learning-based aeroengine nondestructive testing method, device, equipment and storage medium |
| CN113034478A (en) * | 2021-03-31 | 2021-06-25 | 太原科技大学 | Weld defect identification and positioning method and system based on deep learning network |
Non-Patent Citations (1)
| Title |
|---|
| 结构裂纹损伤的Lamb波层析成像监测与评估研究;王强;胥静;王梦欣;宋春晓;;机械工程学报(06);第30-36页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113591270A (en) | 2021-11-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112232400B (en) | Stainless steel weld ultrasonic defect detection method based on depth feature fusion | |
| CN110389170B (en) | Train component crack damage detection method and system based on Lamb wave imaging | |
| CN111896625B (en) | Rail damage real-time monitoring method and monitoring system thereof | |
| Baxter et al. | Delta T source location for acoustic emission | |
| Zhang et al. | Deep convolutional neural network probability imaging for plate structural health monitoring using guided waves | |
| CN109649432B (en) | System and method for monitoring integrity of steel rail of cloud platform based on guided wave technology | |
| CN112816556B (en) | Defect detection method, device, equipment and storage medium | |
| Sun et al. | Machine learning for ultrasonic nondestructive examination of welding defects: A systematic review | |
| Monaco et al. | Guided waves based SHM systems for composites structural elements: statistical analyses finalized at probability of detection definition and assessment | |
| CN107782785B (en) | Underground pipeline safety assessment method | |
| Hensman et al. | Spatial scanning for anomaly detection in acoustic emission testing of an aerospace structure | |
| Amirafshari et al. | Estimation of weld defects size distributions, rates and probability of detections in fabrication yards using a Bayesian theorem approach | |
| US10317367B2 (en) | Eddy-current flaw detector and eddy-current flaw detection method | |
| CN113591270B (en) | MDCD-based monitoring method for performance analysis and comparison on steel rail weld data set | |
| Schoefs et al. | The αδ method for modelling expert judgement and combination of non-destructive testing tools in risk-based inspection context: application to marine structures | |
| CN112525923A (en) | Shared juice platform pipeline inner wall detection method and system and readable storage medium | |
| US20230084562A1 (en) | Non-destructive inspection method and system based on artificial intelligence | |
| Yaacoubi et al. | A model-based approach for in-situ automatic defect detection in welds using ultrasonic phased array | |
| Ikeda et al. | Hammering acoustic analysis using machine learning techniques for piping inspection | |
| Im et al. | Multi-modal diagnostic method for detection of concrete crack direction using light-section method and hammering test | |
| CN111474300B (en) | Structure local defect detection method based on space-time regression model | |
| Iwata et al. | AI-aided Hammering Test System to Automatically Generate Anomaly Maps. | |
| Riahi et al. | Differentiation of leakage and corrosion signals in acoustic emission testing of aboveground storage tank floors with artificial neural networks | |
| Kim et al. | Automated data evaluation in phased-array ultrasonic testing based on A-scan and feature training | |
| Zhao et al. | Automated quantification of small defects in ultrasonic phased array imaging using AWGA-gcForest algorithm |
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 |