CN113239730B - Method for automatically eliminating structural false modal parameters based on computer vision - Google Patents
Method for automatically eliminating structural false modal parameters based on computer vision Download PDFInfo
- Publication number
- CN113239730B CN113239730B CN202110384984.8A CN202110384984A CN113239730B CN 113239730 B CN113239730 B CN 113239730B CN 202110384984 A CN202110384984 A CN 202110384984A CN 113239730 B CN113239730 B CN 113239730B
- Authority
- CN
- China
- Prior art keywords
- modal
- false
- mode
- order
- structural
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Abstract
The invention discloses a computer vision-based method for automatically eliminating structural false modal parameters. Drawing a modal shape drawing by utilizing the existing data of the structure to be measured; manually calibrating according to the false mode and the real mode order and making a data set of the structure to be measured; training a vibration mode image classifier by using a data set of a structure to be tested; inputting a response signal to be detected of the structure into a modal parameter solver; solving to obtain the recognition result of the doping of the real modal parameter and the false modal parameter of the structure to be detected; carrying out effective classification of each order; and automatically classifying the modal parameters of the structure to be detected by using an automatic classifier. The method is used for solving the problems that the accurate result of modal parameter identification is mainly manually judged, the manual selection mode is low in efficiency in the face of massive structural monitoring data, the online automatic analysis of the structural modal is difficult to realize, and the real-time early warning and other functions of the structural health monitoring system are seriously influenced.
Description
Technical Field
The invention relates to the field of structural health monitoring and computer vision, in particular to a method for automatically eliminating structural false modal parameters based on computer vision.
Background
Major engineered structures are inevitably damaged during service, and the damage threatens the normal use and safety of the structures. The structural health monitoring can utilize a sensor network and data management analysis equipment which are structurally installed to establish an early warning network, so that the real-time sensing, identification and diagnosis of structural load, damage and safety state are realized, and the service safety of a building structure can be effectively guaranteed.
The identification of structural modal parameters is a classic inverse problem of structural dynamics, and the structural modal parameters (frequency, vibration mode and damping ratio) are identified through input and output data of structural actual measurement. The modal parameters represent the dynamic characteristics of the structure, are only related to the physical parameters and the mechanical model of the structure, have important significance in structural health monitoring, and are the basis of structural damage identification, model correction and safety evaluation. Currently, the existing modal parameter identification methods can be generally divided into two types: an experimental modal analysis method and a modal parameter identification method under environmental excitation. The traditional experimental modal analysis method needs to carry out manual active excitation on a structure, obtain a frequency response function or an impulse response function according to the relation between input and output, and further identify and obtain modal parameters. However, large structures have difficulty achieving effective manual active excitation. Therefore, the modal parameter identification method under the environment excitation without inputting data is more widely applied to actual engineering. The modal parameter identification method of the environmental excitation mainly comprises a frequency domain decomposition method, an NExT + ERA method and a random subspace method. The method assumes that the environmental excitation is ideal white noise, and can realize effective identification of the structure modal parameters only by processing the output response data of the structure, thereby greatly saving the monitoring time and cost, simplifying the monitoring process and not influencing the normal use of the structure.
However, in practical applications, since the environmental excitation does not satisfy the ideal white noise assumption, some spurious modes often exist in the mode parameter identification result under the environmental excitation. The real mode of the structure is the inherent property of the structure, and when the structure is not damaged, the real mode of the structure can exist stably. However, spurious modes do not have a stable physical meaning, and they occur randomly in the recognition result. At present, the accurate result of modal parameter identification is mainly judged manually, and in the face of massive structural monitoring data, the manual selection mode is low in efficiency, the online automatic analysis of structural modes is difficult to realize, and the functions of real-time early warning and the like of a structural health monitoring system are seriously influenced.
Disclosure of Invention
The invention provides a computer vision-based automatic elimination method for structure false modal parameters, which is used for solving the problems that the accurate result of modal parameter identification is mainly manually judged, the efficiency of a manual selection mode is low in the face of massive structure monitoring data, the online automatic analysis of the structure modal is difficult to realize, and the functions of real-time early warning and the like of a structure health monitoring system are seriously influenced.
The invention is realized by the following technical scheme:
a computer vision-based automatic removing method for structural false modal parameters comprises the following steps:
step 1: performing modal analysis on existing monitoring data of the structure to be detected, and drawing a modal shape drawing;
step 2: manually calibrating the modal shape image drawn in the step 1 according to the false mode and the real mode order and manufacturing the modal shape image into a structural data set to be tested;
and step 3: training a vibration mode image classifier by using the structural data set to be tested in the step 2;
and 4, step 4: inputting a response signal to be detected of the structure into a modal parameter solver;
and 5: solving to obtain a recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected by using the modal parameter solver in the step 4;
step 6: performing effective classification of each order on the recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected obtained in the step 5 by using the vibration mode image classifier trained in the step 3;
and 7: and (4) outputting the modal parameters corresponding to the false modal image distinguished in the step (6) and the real modal of each order, and utilizing an automatic classifier to realize the elimination of the false modal parameters of the structure to be detected and the automatic classification of the real modal parameters according to the order.
Further, the step 1 specifically includes determining a modality to be identified; carrying out frequency domain analysis on the structural vibration response data, and determining the frequency spectrum range of the modal to be identified; performing a filtering operation on the data; and selecting a modal parameter identification method under the environment excitation to perform modal identification on the filtered data.
Further, the step 2 of creating the data set specifically includes performing modal parameter identification on the existing data to draw a vibration mode image, dividing the drawn vibration mode image into two parts, wherein the first part is used as a data source of a vibration mode image training verification set, and the second part is used as a data source of a vibration mode image test set; manually labeling and classifying all the modes, and dividing the modes into false mode vibration types and k +1 type labels of the front k-order mode vibration types; and randomly selecting the vibration mode images labeled and classified by using an algorithm, and respectively making the vibration mode images into a training verification set and a test set.
Further, the step 4 of training the modal parameter solver specifically includes that the modal parameter solver applies a modal parameter identification method under any environment excitation, and the modal parameter solver solves the modal parameter identification result by initially obtaining a modal parameter identification result of a real mode and a false mode in a mixed manner.
Further, the vibration mode image classifier in step 3 is specifically trained to use any one of supervised convolutional neural network classification models to realize effective identification and classification of the vibration mode image.
Further, the implementation of the automatic classification of the modal parameters in step 7 is specifically that, firstly, a priori knowledge is used to perform preliminary screening on a modal parameter identification result of a mixture of a real modality and a false modality, that is, only the modality on the stability graph satisfying the stability condition of the formula (1) and the empirical damping ratio threshold formula (2) is retained, and further classification and identification are performed on the modality;
the modes with negative damping ratio and high damping ratio are false modes;
where i is the assumed order of the structural system in the case of modal parameter recognition under environmental excitation, fiFor modal frequencies, xi, obtained in corresponding orderiFor the modal damping ratio found at the corresponding order,
corresponding to the modal shape vector obtained in the order,
in order to measure the empirical frequency tolerance of the structure,
for the empirical damping ratio tolerance of the structure to be measured,
for the empirical MAC value tolerance of the structure to be tested,
and obtaining the maximum threshold value of the empirical damping ratio of the structure to be measured.
The invention has the beneficial effects that:
the invention can customize a modal analysis system for the structure to be tested, and realizes the long-term online health monitoring of the structure.
The invention adopts the computer vision and deep neural network to automatically eliminate false modal parameters, realize the automatic identification of the modal parameters and improve the intelligent level of the structural health monitoring system.
Drawings
FIG. 1 is an algorithm framework diagram of the present invention.
FIG. 2 is a flow chart of an exemplary algorithm of the present invention.
FIG. 3 is a power spectrum of an exemplary bridge sensor vibration data of the present invention.
FIG. 4 is a diagram of the vertical modalities of the present invention for manually determining the desired identification by the random subspace approach.
FIG. 5 is a production mode image dataset of the present invention.
FIG. 6 is a block diagram of mode shape image data enhancement of the present invention.
Fig. 7 is a diagram of a residual learning unit of the present invention.
Fig. 8 is a diagram of the network structure of the ResNet34 of the present invention.
FIG. 9 is a training process loss graph of the present invention.
FIG. 10 is a training process validation set accuracy chart of the present invention.
FIG. 11 is a test set confusion matrix diagram of the present invention.
Fig. 12 is a six-order mode recognition result MAC value before each day for 7 months to 9 months according to the present invention.
FIG. 13 is a frequency value of the six-order modal recognition results of the present invention taken from month 7 to month 9 before each day.
FIG. 14 is the damping ratio for the six-order mode identification results of the invention before each day from 7 months to 9 months.
FIG. 15 is a flow chart of a method of the present invention.
Fig. 16 is a graph of the modal recognition results of 8/10/2020.
Fig. 17 is a standard modality chart for evaluating a modality recognition result according to the present invention.
Fig. 18 is a graph of the 3 rd order modal parameters of the invention on days 7, 11.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
A computer vision-based automatic removing method for structural false modal parameters comprises the following steps:
step 1: performing modal analysis on existing monitoring data of the structure to be detected, and drawing a modal shape drawing;
step 2: manually calibrating the modal shape image drawn in the step 1 according to the false mode and the real mode order and manufacturing the modal shape image into a structural data set to be tested;
and step 3: training a vibration mode image classifier by using the structural data set to be tested in the step 2; the vibration mode image classifier is any one of supervised depth convolution classification networks; the network utilizes a manually marked modal image training set to learn the capability of classifying real modes and false modes of each order; if the required identification frequency contains k-order real modes, the mode image training set needs to contain k +1 labels, wherein all false modes are classified into the same type of labels. The generalization capability of the network can be improved by adopting data enhancement skills during deep convolutional classification network training;
and 4, step 4: inputting a response signal to be detected of the structure into a modal parameter solver; a modal parameter solving method (such as a random subspace method, a NExT + ERA method and the like) under any environment excitation is utilized;
and 5: solving to obtain a recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected by using the modal parameter solver in the step 4;
step 6: performing effective classification of each order on the recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected obtained in the step 5 by using the vibration mode image classifier trained in the step 3;
and 7: and (4) outputting the modal parameters corresponding to the false modal image distinguished in the step (6) and the real modal of each order, and utilizing an automatic classifier to realize the elimination of the false modal parameters of the structure to be detected and the automatic classification of the real modal parameters according to the order.
Further, the step 1 specifically includes determining a modality to be identified; carrying out frequency domain analysis on the structural vibration response data, and determining the frequency spectrum range of the modal to be identified; performing a filtering operation on the data; and selecting a modal parameter identification method under the environment excitation to perform modal identification on the filtered data.
Further, the step 2 of creating the data set specifically includes performing modal parameter identification on the existing data to draw a vibration mode image, dividing the drawn vibration mode image into two parts, wherein the first part is used as a data source of a vibration mode image training verification set, and the second part is used as a data source of a vibration mode image test set; manually labeling and classifying all the modes, and dividing the modes into false mode vibration types and k +1 type labels of the front k-order mode vibration types; and randomly selecting the vibration mode images labeled and classified by using an algorithm, and respectively making the vibration mode images into a training verification set and a test set.
Further, the step 4 of training the modal parameter solver specifically includes that the modal parameter solver applies a modal parameter identification method under any environment excitation, and the modal parameter solver solves the modal parameter identification result by initially obtaining a modal parameter identification result of a real mode and a false mode in a mixed manner.
The modal parameter identification method under any environment excitation is applied, and the method comprises but is not limited to a random subspace method and a NExT + ERA method.
Further, the vibration mode image classifier in step 7 is specifically trained to use any one of supervised convolutional neural network classification models to realize effective identification and classification of the vibration mode image.
Further, the implementation of the automatic classification of the modal parameters in step 3 is specifically that, firstly, a priori knowledge is used to perform preliminary screening on a modal parameter identification result of a mixture of a real mode and a false mode, that is, only the mode satisfying the stability condition of the formula (1) and the empirical damping ratio threshold value formula (2) on the stability graph is retained, and further classification and identification are performed on the mode; the stable structure does not usually have negative damping ratio and high damping ratio, so the modes with the negative damping ratio and the high damping ratio are false modes;
where i is the assumed order of the structural system in the case of modal parameter recognition under environmental excitation, fiFor modal frequencies, xi, obtained in corresponding orderiFor the modal damping ratio found at the corresponding order,
corresponding to the modal shape vector obtained in the order,
in order to measure the empirical frequency tolerance of the structure,
for the empirical damping ratio tolerance of the structure to be measured,
for the empirical MAC value tolerance of the structure to be tested,
and obtaining the maximum threshold value of the empirical damping ratio of the structure to be measured.
Example 2
The structural vibration response data of a certain bridge from 5 months to 9 months is obtained, and the bridge is provided with 14 vertical vibration sensors which are uniformly distributed on 7 sections of the bridge in pairs. The example plans to make a vibration mode image data set by using data of 5 and 6 months, and finally realizes automatic identification of the first 6-order vertical modal parameters of bridge data of 7, 8 and 9 months. The process flow is shown in FIG. 2.
Determining the modality to be identified
Firstly, frequency domain analysis is needed to be carried out on the bridge vibration response data, and the frequency spectrum range of the mode to be identified is determined. As shown in fig. 3, it can be found from the power spectrum of the vibration data of the bridge sensor that the low-frequency modes of the first orders are included in the frequency domain range of 0Hz to 0.5 Hz. Then, in order to make the interested frequency domain range more prominent, the data is filtered to remove the unnecessary components in the signal except 0Hz to 0.5 Hz. And finally, selecting a modal parameter identification method under the environment excitation to perform modal identification on the filtered data, wherein a random subspace algorithm (replaceable) is adopted in the demonstration example. Finally, the stochastic subspace algorithm identifies the first 6 th order vertical modal process as shown in fig. 4. According to the random subspace algorithm theory, as the assumed order of the structure dynamic model is increased, the real mode stably exists, and the false mode randomly appears. With this modal characteristic, modal parameter identification is performed for the model at each order, assuming that the structural power system has different orders. And drawing all the obtained modal parameters on a stable graph with the frequency as an abscissa and the assumed model order as an ordinate. Each point on the stable graph is a first-order mode and comprises frequency, damping ratio and vibration mode, wherein the point meeting the preset conditions is called a stable point, an axis formed by the stable points is called a stable axis, and the mode corresponding to the stable point of the minimum system order on the stable axis is the real mode of the system as known from the random subspace algorithm theory. The stable diagram shown in fig. 4 has 8 stable axes, but the stable axis mode shapes around 0.3721Hz and 0.4890Hz are torsional mode shapes, and the present example only discusses the automatic identification of the vertical mode shape, and therefore, the above is omitted. Finally, the frequency, the two-dimensional mode shape image and the three-dimensional mode shape image of the first 6 th order vertical mode shape of the bridge are shown in fig. 4. Theoretically, the learning classification of the convolution network can be realized for the two-dimensional vibration mode image or the three-dimensional vibration mode image, but the problem that the visual angle covers part of the vibration mode image exists when the three-dimensional image is used, and the two-dimensional vibration mode image is simpler. Therefore, in this example, a two-dimensional mode image is used as an image to be recognized and classified.
Realization of modal parameter solver
As shown in fig. 2, the vibration response data to be analyzed needs to be solved by a modal parameter solver, and a modal parameter identification result of a mixture of a real mode and a false mode is obtained preliminarily. The modal parameter solver can apply any modal parameter identification method under environmental excitation, such as a random subspace method, a NExT + ERA method, and the like. The covariance drive random subspace method is based on a discrete time random state space model of the system, input and noise items are considered as ideal white noise, a discrete system state space matrix is identified by utilizing the statistical characteristics of the white noise and combining matrix Singular Value Decomposition (SVD) and the like, and finally modal parameters of the system are identified and obtained. The principle of the covariance-driven stochastic subspace approach is not described in detail here. And (3) solving modal parameters by using a set of input modal parameter solver for 2 hours according to daily vibration data of the bridge in the period of 7-9 months and using a random subspace driven by covariance, wherein the maximum assumed order of the bridge dynamic model is 120.
Producing a vibro-graphic data set
The present example proposes to create a vibration pattern image dataset using data of 5 and 6 months. In order to obtain a training sample, the existing data needs to be subjected to modal parameter identification to draw a vibration mode image, and the random subspace method is still selected to carry out modal parameter identification on the data of 5 and 6 months. As shown in fig. 5, a random subspace method is used to perform mode identification on data from 5 months 1 to 6 months 20 days, and from 6 months 21 to 6 months 30 days, respectively, so as to draw a two-dimensional image with a mode shape of 64 pixels in length and width, where the former is used as a data source of a mode shape image training verification set, and the latter is used as a data source of a mode shape image test set. And then, manually labeling and classifying all the modes, and dividing the modes into seven types of labels including false mode shapes and the first 6 orders of mode shapes. Finally, randomly selecting the vibration mode images labeled and classified by an algorithm, wherein ten thousand vibration mode images in each type of the training verification set are randomly divided by taking 80% as the training set and 20% as the verification set; one thousand images per type of mode shape were collected. The training set is a data sample of the network learning image characteristics, and the verification set can monitor whether the network is over-fitted in the training process, so that the generalization capability of the network is preliminarily evaluated. The test set may ultimately evaluate the classification performance of the network.
Appropriate data enhancement processing (alternatives) can be performed when the data set is called, as shown in fig. 6. Random brightness change, tone change, contrast change, random horizontal inversion, random proper distortion deformation and the like can be carried out on a vibration mode image, the data enhancement methods are randomly combined, and finally, a data enhanced image is obtained. The diversity of various samples is increased by using a data enhancement technology, the influence of non-vibration shape characteristics such as color, brightness and the like can be reduced in the training process, so that the network can pay more attention to the essential characteristics of the vibration shape, and the performance and generalization capability of model classification can be improved.
Training vibration type image classifier
The Convolutional Neural Network (CNN) performs convolution calculation on an image by using a convolution kernel, so that the translation invariant feature of the image is extracted while calculation parameters are reduced, and the Convolutional Neural Network (CNN) becomes a common network type of an image processing task. With the rapid development of convolutional neural networks, the network structure is also deeper and deeper, and some novel and wonderful network structures are also proposed and applied, such as: VGGNet, InceptionNet, ResNet, EfficientNet, and the like.
The invention can use any kind of supervised convolutional neural network classification model to realize the effective identification and classification of the vibration mode image, and the most classical ResNet network (replaceable) is selected in the embodiment. While increasing the depth of the network can theoretically give the network the ability to extract more complex feature patterns, deepening the network does not actually improve the accuracy of the network. When the network depth is increased, the network accuracy is easy to saturate or even decline, which is called the degradation problem of the deep network. Hokeming et al studied the cause of network degradation and proposed a ResNet network architecture. At the heart of the ResNet network is a residual learning unit, as shown in fig. 7. Assuming that the input of the residual learning unit is x and the output is h (x), the residual learning unit should keep the approximation between the output h (x) and the input x to avoid the network degradation phenomenon. For this purpose, the residual learning unit adds a skip path in addition to the conventional network processing path, and directly adds the input x to the result f (x) of the network processing path, so that the learning target of the network is shifted from the output h (x) to learn the residual function f (x) ═ h (x) -x. The network structure used in this example is ResNet34, and is shown in fig. 8.
The experimental conditions of the present example are as follows: windows 10, 64-bit operating system, calls the pytore framework using Python programming language. The computer configuration comprises a notebook computer, a GeForce GTX 1660ti video card and a 6G video memory; intel (R) core (TM) i7-9750 processor, main frequency 2.60GHz, memory 16 GB. The model selects cross entropy as a loss function, the learning rate is set to be 0.01, the size of a training batch is set to be 128, and the number of experimental iterations is set to be 30. The training set loss curve and the verification set accuracy in the final training process are shown in fig. 9 and 10, and it can be seen that the classification accuracy of the network on the verification set vibration mode images is greater than 98.50%, and the false modal shape and the modal shape of each order can be effectively distinguished. And finally, selecting the 23 rd iteration network parameters with the highest verification set accuracy rate of 99.94% as a final classification model. The classification accuracy of the final classification model on the test set vibration pattern image is 95.97%, and the test set classification confusion matrix is shown in fig. 11.
Implementing automatic classification of modal parameters
As can be seen from the stable graph of fig. 4, since the maximum assumed order of the bridge model system is 120 orders, the number of true modes and the number of false modes obtained by the random subspace method are redundant, and particularly, the number of false modes is too large. Although classification of the mode shape images can be selected for all mode identification results theoretically, in order to increase the operation efficiency and reduce the consumption of operation resources, firstly, appropriate priori knowledge is applied to primarily screen the random subspace identification modes, that is, only the mode satisfying the formula (1) stability condition and the empirical damping ratio threshold formula (2) on the stable graph is reserved, and further classification and identification are performed on the mode. The stable structure does not usually exhibit negative and high damping ratios, and therefore is generally considered a spurious mode for negative and high damping ratio modes.
0＜ζi＜0.01 (2)
In order to verify that the effect of analyzing the screened modes is approximately the same as that of analyzing all the modes, randomly selecting vibration data of a certain day to carry out classification test on the two mode analysis modes. Finally, data of a day of 2020, 8, 10 and the like are selected, and the analysis results are shown in table 1. It can be seen from the table that when the maximum assumed order of the bridge dynamic model is 120, the total number of modes finally obtained by the random subspace method is as many as 46005, and 82.6% of the total number of modes are spurious modes, which is consistent with the conjecture obtained from the stable graph of fig. 4. The operational efficiency is greatly reduced if all the modes are analyzed. However, when the mode identification result is preliminarily screened by using the stable graph principle and the prior knowledge, the number of the modes to be analyzed is reduced to 7360, and a sufficient number of real modes are reserved, so that the operation efficiency is greatly improved, and the subsequent analysis process is not obviously influenced.
TABLE 12020 years 8 months 10 days modal Classification results
For each screened mode, the algorithm will save its frequency, damping and mode shape and will draw a mode shape image. And then, inputting the modal shape image into a shape image classifier to determine the class to which the modal shape image belongs, and outputting the softmax value corresponding to the seven neural units of the network full-connection layer as a probability value classified into the class. Since the stochastic subspace approach assumes a maximum order of 120 for the bridge dynamics model, and the 24-hour day data is divided into 12 groups, the recognition of the true mode at each order in the day is redundant. And sorting the real modal recognition results of each order of each day from large to small according to the probability values classified into the categories, selecting the first 5 recognition results to output, automatically generating a table, and taking the table as a final recognition result, such as the table 2 which is the final recognition result of 8 months and 10 days in 2020.
Discussion of automatic recognition results of 6-order vertical modal parameters from 7 months to 9 months ago;
and automatically identifying the vertical modal parameters of 6 orders before 7 months to 9 months, wherein the first 6 orders of modals on other dates are successfully identified except that the 3 rd modality on 26 days of 9 months, the 3 rd modality on 27 days of 9 months and the 3 rd modality on 30 days of 9 months are not successfully identified in the identification result. A group of more standard first 6-order modes are randomly selected from the training set to evaluate the automatic identification result of the mode parameters from 7 months to 9 months, and the selected 6 standard modes are shown in table 3.
Fig. 17 shows 5 results of each-step daily Modal parameter recognition, and since the present invention is based on the automatic Modal parameter recognition implemented by mode image classification, the result with the highest mode assessment confidence Criterion (MAC) value is selected from the 5 recognition results for evaluation. The calculation formula of the MAC value is shown in formula (3).
Wherein phii，φjAre respectively the vibration mode vector phii *Is the conjugate transpose of the mode shape vector. The calculated MAC value is in the range of 0 to 1 and closer to 1 indicates that the two mode vectors are more similar. Fig. 12 shows the MAC values of the first 6-order mode shapes recognized in months 7 to 9 and the corresponding standard mode shapes, and it can be seen that most of the recognition results have MAC values of 0.8 or more, a small number of 3-order modes have lower MAC values, and particularly, the MAC value of the 3-order mode on day 7 and 11 is only 0.29. The 3 rd order modal parameters for 7 months and 11 days are shown in table 4.
As can be seen from fig. 18, the lower half (downstream side) of the 3 rd order mode shape on this day is similar to the standard shape, while the upper half is different from the standard shape. However, it can be seen from the frequency that the five recognition results are all around the 3 rd order standard mode frequency 0.2268HZ, so that they can be determined to belong to the third order real mode. Such a modality may be referred to as a "real modality with defects", possibly due to imperfections in the local sensor data. In other recognition results from month 7 to month 9, the true mode with defects also appears, but as can be seen from fig. 12, the generalization capability of the mode image classifier is enhanced by virtue of the image enhancement operation in the convolutional network training process, and the classification can be correctly recognized even if the true mode with defects appears. Fig. 13 shows frequency values of the modal recognition results of six orders before 7 to 9 months, and it can be seen that the frequency of each order is relatively stable, which also illustrates that the algorithm can realize effective classification of the modes of each order. Fig. 14 shows damping ratio values of the six-order mode recognition results from 7 to 9 months ago, and the damping ratio recognition results show certain discreteness, which are mainly due to the following three reasons: (1) when the bridge vibrates, the surrounding air interacts with the bridge vibration to form additional pneumatic damping, and therefore the damping identification result based on the measured vibration signal comprises structural damping and pneumatic damping. The pneumatic damping is closely related to the flow field wind speed and the shape of the bridge structure. Due to the fact that the shape of the bridge structure is complex, the field wind speed is complex and changeable and random (the wind speed and the wind direction change time, and the wind field space is uneven), the damping ratio identification result presents certain discreteness. (2) The damping cause of the bridge structure is complex, and the influence factors are numerous, which is also the important reason for the randomness of the damping. (3) The stochastic subspace approach assumes the environmental excitation as white noise, and the actual environmental load does not completely conform to such an assumption. In addition, the value of the system order also affects the accuracy of the damping ratio. The value is too large, so that more false modes are caused; and the value is small, so that the real mode of the structure cannot be identified.
Claims (6)
1. A method for automatically eliminating structural false modal parameters based on computer vision is characterized by comprising the following steps:
step 1: performing modal analysis on existing monitoring data of the structure to be detected, and drawing a modal shape drawing;
step 2: manually calibrating the modal shape image drawn in the step 1 according to the false mode and the real mode order and manufacturing the modal shape image into a structural data set to be tested;
and step 3: training a vibration mode image classifier by using the structural data set to be tested in the step 2;
and 4, step 4: inputting a response signal to be detected of the structure into a modal parameter solver;
and 5: solving to obtain a recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected by using the modal parameter solver in the step 4;
step 6: performing effective classification of each order on the recognition result of doping of the real modal parameters and the false modal parameters of the structure to be detected obtained in the step 5 by using the vibration mode image classifier trained in the step 3;
and 7: and (4) outputting the modal parameters corresponding to the false modal image distinguished in the step (6) and the real modal of each order, and utilizing an automatic classifier to realize the elimination of the false modal parameters of the structure to be detected and the automatic classification of the real modal parameters according to the order.
2. The method for automatically eliminating the structural false modal parameters based on the computer vision according to claim 1, wherein the step 1 is to determine a modal to be identified; carrying out frequency domain analysis on the structural vibration response data, and determining the frequency spectrum range of the modal to be identified; performing a filtering operation on the data; and selecting a modal parameter identification method under the environment excitation to perform modal identification on the filtered data.
3. The method for automatically eliminating the structure false modal parameters based on the computer vision as claimed in claim 1, wherein the step 2 of creating the data set specifically includes performing modal parameter recognition on existing data to draw a vibration mode image, and dividing the drawn vibration mode image into two parts, wherein the first part is used as a data source of a vibration mode image training verification set, and the second part is used as a data source of a vibration mode image testing set; manually labeling and classifying all the modes, and dividing the modes into false mode vibration types and k +1 type labels of the front k-order mode vibration types; and randomly selecting the vibration mode images labeled and classified by using an algorithm, and respectively making the vibration mode images into a training verification set and a test set.
4. The method for automatically eliminating the structural false modal parameters based on the computer vision as claimed in claim 1, wherein the step 4 uses a modal parameter solver, specifically, the modal parameter solver applies a modal parameter recognition method under any environmental excitation, and a modal parameter recognition result of mixing a real mode and a false mode is obtained preliminarily through solving by the modal parameter solver.
5. The method for automatically eliminating the structural false modal parameters based on the computer vision as claimed in claim 1, wherein the vibration mode image classifier of the step 3 is specifically trained to use any one of supervised convolutional neural network classification models to realize effective identification and classification of the vibration mode image.
6. The method according to claim 1, wherein the step 7 of automatically classifying the structural false modal parameters is to firstly apply a priori knowledge to preliminarily screen a modal parameter identification result of a mixture of real modes and false modes, that is, only the modes on the stability graph that satisfy the stability condition of the formula (1) and the empirical damping ratio threshold formula (2) are retained, and further classification and identification are performed on the modes; the modes with negative damping ratio and high damping ratio are false modes;
where i is the assumed order of the structural system in the case of modal parameter recognition under environmental excitation, fiFor modal frequencies, xi, obtained in corresponding orderiFor the modal damping ratio found at the corresponding order,
corresponding to the modal shape vector obtained in the order,
in order to measure the empirical frequency tolerance of the structure,
for the empirical damping ratio tolerance of the structure to be measured,
for the empirical MAC value tolerance of the structure to be tested,
and obtaining the maximum threshold value of the empirical damping ratio of the structure to be measured.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110384984.8A CN113239730B (en) | 2021-04-09 | 2021-04-09 | Method for automatically eliminating structural false modal parameters based on computer vision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110384984.8A CN113239730B (en) | 2021-04-09 | 2021-04-09 | Method for automatically eliminating structural false modal parameters based on computer vision |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113239730A CN113239730A (en) | 2021-08-10 |
| CN113239730B true CN113239730B (en) | 2022-04-05 |
Family
ID=77127961
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110384984.8A Active CN113239730B (en) | 2021-04-09 | 2021-04-09 | Method for automatically eliminating structural false modal parameters based on computer vision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113239730B (en) |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006317394A (en) * | 2005-05-16 | 2006-11-24 | Tokyo Institute Of Technology | Rigidity identifying system of elastic support |
| CN101894097A (en) * | 2010-07-23 | 2010-11-24 | 哈尔滨工业大学 | Maneuvering target tracking method with organic combination of Kalman filtering and empirical mode decomposition |
| CN103996056A (en) * | 2014-04-08 | 2014-08-20 | 浙江工业大学 | Tattoo image classification method based on deep learning |
| CN104200004A (en) * | 2014-07-28 | 2014-12-10 | 东北大学 | Optimized bridge damage identification method based on neural network |
| CN104200265A (en) * | 2014-07-28 | 2014-12-10 | 东北大学 | Improved bridge damage identification method based on neural network |
| CN105787655A (en) * | 2016-02-24 | 2016-07-20 | 西安工业大学 | Superhigh-layer structure modal parameter identification method |
| CN106446812A (en) * | 2016-09-13 | 2017-02-22 | 西安科技大学 | Driving state recognition method based on approximate entropy template matching |
| CN106777763A (en) * | 2016-12-31 | 2017-05-31 | 大连理工大学 | A kind of accurate method of discrimination of engineering structure false mode |
| CN106840256A (en) * | 2017-01-17 | 2017-06-13 | 赵玉 | A kind of civil engineering structure health detecting system |
| CN107609291A (en) * | 2017-09-22 | 2018-01-19 | 哈尔滨工业大学 | A kind of false mode elimination method based on Density Clustering |
| CN110889430A (en) * | 2019-10-24 | 2020-03-17 | 中国科学院计算技术研究所 | News image detection method, system and device based on multi-domain visual features |
| CN111079444A (en) * | 2019-12-25 | 2020-04-28 | 北京中科研究院 | Network rumor detection method based on multi-modal relationship |
| CN111831790A (en) * | 2020-06-23 | 2020-10-27 | 广东工业大学 | False news identification method based on low threshold integration and text content matching |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10990298B2 (en) * | 2018-10-22 | 2021-04-27 | International Business Machines Corporation | Implementing data requests with quality of service information |
-
2021
- 2021-04-09 CN CN202110384984.8A patent/CN113239730B/en active Active
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006317394A (en) * | 2005-05-16 | 2006-11-24 | Tokyo Institute Of Technology | Rigidity identifying system of elastic support |
| CN101894097A (en) * | 2010-07-23 | 2010-11-24 | 哈尔滨工业大学 | Maneuvering target tracking method with organic combination of Kalman filtering and empirical mode decomposition |
| CN103996056A (en) * | 2014-04-08 | 2014-08-20 | 浙江工业大学 | Tattoo image classification method based on deep learning |
| CN104200004A (en) * | 2014-07-28 | 2014-12-10 | 东北大学 | Optimized bridge damage identification method based on neural network |
| CN104200265A (en) * | 2014-07-28 | 2014-12-10 | 东北大学 | Improved bridge damage identification method based on neural network |
| CN105787655A (en) * | 2016-02-24 | 2016-07-20 | 西安工业大学 | Superhigh-layer structure modal parameter identification method |
| CN106446812A (en) * | 2016-09-13 | 2017-02-22 | 西安科技大学 | Driving state recognition method based on approximate entropy template matching |
| CN106777763A (en) * | 2016-12-31 | 2017-05-31 | 大连理工大学 | A kind of accurate method of discrimination of engineering structure false mode |
| CN106840256A (en) * | 2017-01-17 | 2017-06-13 | 赵玉 | A kind of civil engineering structure health detecting system |
| CN107609291A (en) * | 2017-09-22 | 2018-01-19 | 哈尔滨工业大学 | A kind of false mode elimination method based on Density Clustering |
| CN110889430A (en) * | 2019-10-24 | 2020-03-17 | 中国科学院计算技术研究所 | News image detection method, system and device based on multi-domain visual features |
| CN111079444A (en) * | 2019-12-25 | 2020-04-28 | 北京中科研究院 | Network rumor detection method based on multi-modal relationship |
| CN111831790A (en) * | 2020-06-23 | 2020-10-27 | 广东工业大学 | False news identification method based on low threshold integration and text content matching |
Non-Patent Citations (4)
| Title |
|---|
| Blind Modal Identification of Output-Only Structures in Time-Domain Based on Complexity Pursuit;Yang Y;《Earthquake Engineering &》;20131231;第1885-1905页 * |
| 环境激励下结构模态参数识别的机器学习方法;刘大伟;《中国优秀硕士学位论文全文数据库 工程科技II辑》;20210215;C034-405 * |
| 环境激励下钢板组合梁桥模态参数识别;李超等;《安徽建筑》;20200325(第03期);第96-98页 * |
| 通讯塔钢结构动力特性的数值计算与现场测试;周兆明等;《测试技术学报》;20200117(第01期);第61-67页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113239730A (en) | 2021-08-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Pan et al. | LiftingNet: A novel deep learning network with layerwise feature learning from noisy mechanical data for fault classification | |
| CN108665460B (en) | Image quality evaluation method based on combined neural network and classified neural network | |
| CN110297479B (en) | Hydroelectric generating set fault diagnosis method based on convolutional neural network information fusion | |
| CN110636066B (en) | Network security threat situation assessment method based on unsupervised generative reasoning | |
| CN109389171B (en) | Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology | |
| CN108873401B (en) | Liquid crystal display response time prediction method based on big data | |
| CN114169377A (en) | G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment | |
| CN112070720A (en) | Transformer substation equipment defect identification method based on deep learning model | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption | |
| CN110502989A (en) | A kind of small sample EO-1 hyperion face identification method and system | |
| Sun et al. | Data-driven fault diagnosis method based on second-order time-reassigned multisynchrosqueezing transform and evenly mini-batch training | |
| CN110458071B (en) | DWT-DFPA-GBDT-based optical fiber vibration signal feature extraction and classification method | |
| Wang et al. | Distortion recognition for image quality assessment with convolutional neural network | |
| CN117056678B (en) | Machine pump equipment operation fault diagnosis method and device based on small sample | |
| Gao et al. | Image quality score distribution prediction via alpha stable model | |
| CN113109782B (en) | Classification method directly applied to radar radiation source amplitude sequence | |
| CN113239730B (en) | Method for automatically eliminating structural false modal parameters based on computer vision | |
| Chandler et al. | Discrimination of locally stationary time series based on the excess mass functional | |
| DE112014006799T5 (en) | Method for measuring a user behavior consistency level based on a complex correspondence system | |
| CN111652238B (en) | Multi-model integration method and system | |
| CN116994044A (en) | Construction method of image anomaly detection model based on mask multi-mode generation countermeasure network | |
| CN116309493A (en) | Method and system for detecting defects of textile products | |
| CN112463387B (en) | Method for identifying deep learning model on local server based on GPU space-time resource consumption | |
| CN115099135A (en) | Improved artificial neural network multi-type operation power consumption prediction method | |
| CN114818945A (en) | Small sample image classification method and device integrating category adaptive metric learning |
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 |