A bridge cable anomaly identification and location method that is insensitive to sensor failures

Abstract

The present invention discloses a bridge cable anomaly identification and positioning method that is insensitive to sensor failures, and the steps are as follows: (1) extraction of a group cable force characteristic vector based on monitoring data; (2) establishment of an online evaluation model for inclined cables based on local variability detection of the cable force characteristic vector; and (3) isolation of abnormal components in the cable force characteristic vector. The present invention constructs a group cable force characteristic vector that reflects the stability correlation of vehicle-induced cable forces between different cables; establishes a bridge cable anomaly identification method based on variability detection of the group cable force characteristic vector, and realizes online monitoring of inclined cables. Based on the isolation method of abnormal cable force components, abnormal cables can be accurately located. Finally, the method is insensitive to sensor failures. In the event of a cable force sensor failure, the model does not need to be retrained and can continue to effectively perform online monitoring and identification of inclined cable force anomalies. Therefore, the present invention has high engineering application value in the field of bridge cable force anomaly assessment.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge structure performance evaluation, and in particular relates to a bridge cable anomaly identification and positioning method that is insensitive to sensor failures. Background Art

[0002] As the primary supporting member of a cable-stayed bridge, the cable-stay is responsible for transmitting the dead load and traffic load of the superstructure to the main tower and foundation. However, the cable-stay is under high stress for a long period of time and is affected by the coupling of alternating stress from vehicle loads, environmental effects, and abnormal vibrations. These effects make the cable-stay extremely susceptible to damage such as wire breakage, fatigue, deterioration of the protective layer, and corrosion of the steel wire. As the inevitable damage accumulates, the cable-stay deteriorates further, resulting in a decrease in the tensile strength of the cable and even cable breakage. Once a cable-stay breaks, it will inevitably cause changes in the mechanical properties of the cable-stayed bridge and a loss of the bridge's bearing capacity, thereby affecting the safety and stability of the entire bridge. Therefore, monitoring and performance evaluation of the cable-stay is crucial to ensuring the safe operation of cable-stayed bridges.

[0003] Currently, stay-cable condition assessment methods based on monitoring data primarily include direct and indirect methods. Direct methods diagnose stay-cables by monitoring cable forces or cable stress changes, while indirect methods utilize bridge deck strain, rotational influence lines, girder deflections, and other parameters to indirectly monitor cable condition. However, indirect methods often require monitoring data from multiple bridge components or locations, which may not always be available in actual engineering deployments. Furthermore, the data analysis process is more complex, limiting their practical application. Furthermore, extensive research has been conducted on direct methods. Dan Danhui et al. proposed a health assessment method for cable-stayed bridges based on a group cable force dissimilarity measure. However, this method requires a corresponding finite element model to extract the influence matrix. Li Shunlong et al. proposed a method for assessing the condition of stay-cables based on pattern recognition of the upstream and downstream vehicle-induced cable force ratios. However, this method uses a Gaussian mixture model to statistically model the vehicle-induced cable force ratio, making it incapable of real-time online monitoring of the cables. Furthermore, this method can only identify damaged cable pairs, not the specific damaged cable. Peng Zhen et al. proposed a long-term cable condition assessment method based on the matched cable tension ratio of two cables on the same side. This method uses the slope of the least squares fit line of the matched cable tension ratio of the two cables over a period of time as a damage-sensitive feature. Similarly, this method cannot provide real-time cable condition assessment. Patent publication number CN115048998A proposes a method for identifying and locating cable force anomalies in cable-stayed bridges based on monitoring data, enabling online cable condition assessment and anomaly warning. Furthermore, this method achieves online anomaly location in cables by constructing an anomaly isolation indicator. However, if the cable force sensor of a cable fails during online monitoring, this method becomes inapplicable and requires model retraining. In summary, existing cable condition assessment methods based on monitoring data often suffer from difficulties in ensuring real-time performance and accurately locating abnormal cables. Furthermore, sensor failure is an inevitable issue in health monitoring systems, but existing methods rarely address the applicability of bridge cable anomaly identification methods in the presence of cable force sensor failures.

[0004] Therefore, the present invention proposes a bridge cable anomaly identification and positioning method that is insensitive to sensor failures. While realizing real-time online monitoring and anomaly positioning of the cable force, the method also has good robustness against cable force sensor failures. That is, in the event of a cable force sensor failure, the cable can continue to be monitored online without retraining the model. Summary of the Invention

[0005] The purpose of the present invention is to provide a bridge cable anomaly identification and positioning method that is insensitive to sensor failure. The specific technical solution is as follows:

[0006] Step 1. Extraction of group cable force feature vector based on monitoring data.

[0007] (1.1) Extract vehicle-induced cable forces. Set the moving window length to 15 minutes and select cable force monitoring data within a window length for kernel density estimation fitting. Use the cable force corresponding to the maximum probability density point extracted from the kernel density estimation model as the representative value of the static load cable force during that period. Set the window moving step length to 5 minutes and repeat the above steps to update the representative values ​​of the static load cable force in different time periods. Use spline interpolation, median filtering, and sliding average to extract the trend term of the extracted static load cable force representative value. Subtract the static load cable force trend term from the monitored original cable force data to obtain the vehicle-induced cable force.

[0008] (1.2) Cable tension periods with only a single peak are considered the vehicle-induced cable force for a single vehicle crossing the bridge and are identified and extracted. First, the peak and valley points of the vehicle-induced cable force are extracted using the differential method. Cable tension periods are selected where both sides of the peak decrease monotonically to the corresponding valley point. The valley points on both sides should be less than a certain threshold, set at 3 kN. Second, the width of the extracted peak cable force period for a single vehicle crossing the bridge should be no less than 4 seconds, and the peak value should not be less than 8 kN.

[0009] (1.3) According to the case of a single vehicle crossing the bridge, there is a time lag in the appearance of the corresponding vehicle-induced cable force peak when the vehicle passes through different cables in sequence. The time difference between the vehicle-induced peak cable forces of two adjacent inclined cables is less than 2s. i Find the peak value T of the cable force response under the same vehicle load v (p i ).

[0010] (1.4) Construct the group cable force characteristic vector x under the single vehicle crossing bridge condition i , which includes the vehicle-induced peak cable forces of n cables.

[0011] x i =[T v (p1),T v (p2),...,T v (p n )] T (1)

[0012] Step 2. Establishment of an online evaluation method for stay cables based on the detection of local variability of cable force eigenvectors

[0013] (2.1) According to step 1, the cable force monitoring data of the inclined cable in the normal state are extracted to form a group cable force feature vector, and the set X(t)∈R m×n As the training set, m is the number of extracted cable force feature vectors, and n is the dimension of the cable force feature vector.

[0014] (2.2) For the training set X(t)∈R m×n Each cable force eigenvector x i , use formula (2) to calculate and other n-1 cable force eigenvectors x j Then, according to formula (3), the distance d i,j Arrange in ascending order and take the x corresponding to the first k Euclidean distances j Form the nearest neighbor subset N k (x i ).

[0015] d i,j =||x i -x j ||2,j=1,...,n;j≠i (2)

[0016] d i,1 ≤d i,2 ≤…≤d i,k (3)

[0017] N k (x i )={x i,1 ,x i,2 ,...,x i,k}∈R m×n (4)

[0018] Where: x i,k is the cable force eigenvector x i The kth nearest neighbor vector of ; k is the number of nearest neighbor subsets, with all feature vectors x in the training set i The range and standard deviation of the calculated warning index are used as the target, and the number of subsets corresponding to the minimum range and standard deviation is selected through the cross-validation method as the k value used in this method;

[0019] (2.3) Determine the nearest neighbor set G according to formula (5). Determine the set-based nearest path s as shown in formula (6), for all 1≤j≤k, p j+1 is a set {p1,p2,...,p j}in {p j+1 ,p j+2 ,...,p k+1 If the nearest neighbor is not unique, the selection is made in the order of the elements in the nearest neighbor set G.

[0020] G=N k (x i )∪{x i}={x i ,x i,1 ,x i,2 ,...,xi,k} (5)

[0021] s=<p1,p2,...,p k+1 >,p j ∈G (6)

[0022] Where: The nearest neighbor set G and the nearest path s based on the set contain the same cable force eigenvector, that is, In addition, the first element of the nearest neighbor set G and the nearest path s is the same, that is, p1 = x i .

[0023] The nearest neighbor calculation method based on Mahalanobis distance is used to determine the distance between the vector and the set, and then determine the shortest path s. The calculation method is shown in formula (7). For any given x∈P, if y∈Q satisfies dist(x,y)=dist(P,Q), then y can be said to be the nearest neighbor sample of set P in set Q. This can be used to find the shortest distance between a sample point and a set.

[0024] dist(P,Q)=min{d M (x,y):x∈P&y∈Q} (7)

[0025] in

[0026] d M (x,y)=(xy) T ·S -1 ·(xy) (8)

[0027] In formula (8), S is the covariance matrix, and the cable force eigenvectors in the training set are used to calculate according to formula (9).

[0028]

[0029] Where: N is the number of cable force feature vectors in the training set, is the mean vector of all cable force eigenvectors in the training set;

[0030] (2.4) Based on the determined cable force characteristic vector x i Its nearest neighbor N k (x i ) further determines the nearest trajectory based on the set <e1,e2,...,e k >, where e j For the connection set o j With the eigenvector p j+1 The shortest connection distance.

[0031] e j =(o j ,pj+1 ),o j ∈{p1,...,p j} (10)

[0032] Among them, e j The determination of needs to meet the requirements of formula (11).

[0033] dist(e j )=dist(o j ,p j+1 )=min{d M (o ji ,p j+1 ):o ji ∈o j &1≤i≤j} (11)

[0034] (2.5) Define the cable force characteristic vector x i Isolation Index DI k (x i ), and its calculation formula is shown in formula (12).

[0035]

[0036] (2.6) Repeat steps (2.2)-(2.5) to calculate the nearest neighbor subset {N k (o)|o∈N k (x i )}. Then, calculate each cable force eigenvector x according to formula (13): i Local variation index DI k (x i )

[0037]

[0038] (2.7) The generalized extreme value distribution (GEV) is used in combination with the maximum block method to fit the distribution of the local variation index in the training set, where the GEV distribution expression is shown in formula (14).

[0039]

[0040] The maximum block method selects the optimal number of blocks b by applying the goodness-of-fit metric through the KS hypothesis test, and then the set of local variation indices Θ = [DI k (x1),DI k (x2),...,DI k (x m)] is divided into b equal-sized non-overlapping blocks, and the maximum value of each block is extracted for GEV distribution fitting. In order to determine the warning threshold τ under a certain significance level α α , it is necessary to calculate the extreme quantiles of the GEV distribution based on the estimated GEV parameters and according to formula (15).

[0041]

[0042] Where: ξ, σ, μ represent the shape, scale and location parameters of the GEV distribution;

[0043] (2.8) During the online monitoring phase, repeat steps 1 and 2 to calculate each new cable force characteristic vector z i Local variation index DI k (z i ) and compare it with the warning threshold set in step (2.7), so as to perform online evaluation and abnormal warning of the inclined cable;

[0044] Step 3. Isolation of abnormal cable force components of cable force eigenvector

[0045] (3.1) In the online monitoring stage, when a new cable force characteristic vector z calculated in (2.8) i Local variation index DI k (z i ) exceeds the set warning threshold, the cable force characteristic vector z is reduced in sequence i A tension component in the k (z i ), when the calculated local variation index drops below the threshold, it indicates that the corresponding inclined cable of the cable force component has an abnormality.

[0046] Beneficial effects of the present invention:

[0047] 1. The cable-stayed bridge assessment method developed in this paper, based on detecting local variability in cable-force eigenvectors, enables real-time online condition assessment of cables. Furthermore, the method offers clear interpretability, ensuring its applicability to other cable-stayed bridges.

[0048] 2. The bridge cable anomaly identification method proposed in the present invention has a certain robustness against cable force sensor failures. In the event of a sensor failure, that is, the cable force feature vector loses a cable force component, the bridge cables can still be monitored online and abnormal warnings can be issued without the need for retraining the model.

[0049] 3. Based on the fact that this method has good robustness against cable tension sensor failures, that is, when a cable tension component is lost in the cable tension characteristic vector, the warning threshold can be left unchanged and online monitoring of the bridge cables can continue. The present invention further proposes a method for isolating abnormal components in the cable tension characteristic vector. This method is easy to understand and operate, and can accurately locate abnormal cables. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 This is the online status assessment result of the stay cable when the cable force sensor is intact;

[0052] Figure 3 This is the abnormal location result of the inclined cable;

[0053] Figure 4 The online status assessment results of the stay cable when the cable force sensor fails;

[0054] Figure 5 The figure shows the comparison of model training results under normal and faulty conditions of the cable force sensor. DETAILED DESCRIPTION

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and a calculation example.

[0056] The method for identifying and locating abnormal cable force in cable-stayed bridges of the present invention is divided into three steps: "extraction of cable force characteristic vectors of group cables based on monitoring data", "establishment of an online evaluation model for cable-stayed cables based on detection of local variability of cable force characteristic vectors" and "isolation of abnormal components in cable force characteristic vectors". The implementation process of the method is as follows: Figure 1 shown.

[0057] In this specific numerical example, the cable force monitoring data of a cable-stayed bridge in service in China for ten days is used for verification. Here, the monitoring data of the five cables SJS08-SJS12 on the upstream side are taken as an example. First, the vehicle-induced cable force is extracted according to step (1.1), and then the vehicle-induced cable force peak period under the condition of a single vehicle crossing the bridge is identified and extracted according to step (1.2). Finally, step (1.3) is performed to match the cable force peaks between different cables, thereby obtaining the cable group force characteristic vector. Among them, the cable force characteristic vector X1∈R of SJS08-SJS12 for ten days is extracted. 5×5750,Afterwards, a bridge cable anomaly recognition model is established. It is known that one of the cables SJS08-SJS12 is damaged on the tenth day. Therefore, the cable force feature vector on the tenth day is used as the test set, and the remaining cable force feature vectors are divided into training set and validation set in a ratio of 4:1, that is, 1-4366 is the training set, 4367-5457 is the validation set, and 5458-5750 is the test set.

[0058] According to step 2, an online evaluation model for stay cables based on the detection of local variability of cable force eigenvectors is established. The training set is used to train the model and set the threshold according to step (2.7). The validation set is used to verify the effectiveness of the proposed method. The test set is further used to evaluate the abnormality recognition effect of the proposed method under the damaged state of the stay cables. Figure 2 The abnormal identification result of the inclined cable, the isolation index DI calculated from the test set k (x i ) exceed the set threshold.

[0059] Follow the isolation procedure of the abnormal cable force components in step 3 and reduce the cable force characteristic vector z i and recalculate the local variation index. Figure 3 It shows that when the cable force component of cable SJS11 is missing, most of the local variation indices calculated by the test set drop below the threshold, indicating that an anomaly has occurred in the cable force component corresponding to the inclined cable SJS11.

[0060] In order to verify that the bridge cable anomaly identification method proposed in this invention has a certain robustness to the cable force sensor failure, it is assumed that the SJS08 sensor fails on the tenth day. Therefore, the cable force component monitored by the faulty sensor (i.e., SJS08) is excluded and the feature vector X2∈R is reconstructed. 4×5750 Similarly, the cable force eigenvector from the tenth day was used as the test set. The remaining cable force eigenvectors were divided into training and validation sets at a ratio of 4:1, with 1-4366 as the training set, 4367-5457 as the validation set, and 5458-5750 as the test set. The isolation index was calculated using the test set after the faulty sensor was removed, and the threshold value obtained from the training data set when the sensor was in normal condition was used for early warning. Figure 4 This indicates that after a sensor failure, the cable anomaly recognition method proposed in this invention can continue to use the threshold value obtained from the training of the dataset under the normal state of the sensor to issue an early warning, without the need to retrain the model. Furthermore, the model was retrained according to the traditional method, and the training results of the two cases were compared. Figure 5This indicates that the calculated isolation indices for the two scenarios are highly correlated, and the magnitude of change in the warning indices is essentially the same. Therefore, the model's warning threshold does not need to be updated. In summary, the proposed bridge cable anomaly identification method exhibits robustness against cable tension sensor failures, enabling continued online monitoring and anomaly warning of bridge cables without requiring model retraining.

Claims

1. A bridge cable anomaly identification and positioning method that is insensitive to sensor failure, characterized in that: Here are the steps: Step 1. Extraction of group cable force feature vector based on monitoring data (1.1) Extracting vehicle-induced cable forces: First, set the moving window step size to 15 minutes and select cable force monitoring data within a window length for kernel density estimation fitting. The cable force corresponding to the maximum probability density point extracted from the kernel density estimation model is used as the representative value of the static load cable force within that period. Then, the moving window step size is set to 5 minutes, and the above steps are repeated to update the representative values ​​of static cable forces at different time periods. The static cable force trend term is extracted from the extracted representative values ​​of static cable forces using spline interpolation, median filtering, and sliding average. The static cable force trend term is subtracted from the monitored original cable force data to obtain the vehicle-induced cable force. (1.2) The cable force period with only a single peak is regarded as the vehicle-induced cable force in the case of a single vehicle crossing the bridge and is identified and extracted: First, the peak and valley points of the vehicle-induced cable force are extracted by the difference method. The cable force period is selected in which both sides of the peak decrease monotonically to the corresponding valley point. The valley points on both sides should be less than a certain threshold, which is set to 3kN. Second, the width of the peak period of the vehicle-induced cable force in the case of a single vehicle crossing the bridge is extracted to be no less than 4s and the peak value is no less than 8kN. (1.3) According to the case of a single vehicle crossing the bridge, there is a time lag in the appearance of the corresponding vehicle-induced cable force peaks when the vehicle passes through different cables in sequence. The time difference between the vehicle-induced peak cable forces of two adjacent cables is less than 2s. i Find the peak value T of the cable force response under the same vehicle load v (p i ); (1.4) Construct the group cable force characteristic vector x under the condition of a single vehicle crossing the bridge i , which includes the vehicle-induced peak cable forces of n cables: x i =[T v (p1),T v (p2),...,T v (p n )] T (1) Step 2. Establishment of an online evaluation method for stay cables based on the detection of local variability of cable force eigenvectors (2.1) According to step 1, the cable force monitoring data of the inclined cable in the normal state are extracted to form a group cable force feature vector, and the set X(t)∈R m×n As the training set, where m is the number of extracted cable force feature vectors and n is the dimension of the cable force feature vector; (2.2) For the training set X(t)∈R m×n Each cable force eigenvector x i , use formula (2) to calculate and other n-1 cable force eigenvectors x j The Euclidean distance between them; then according to formula (3) i,j Arrange in ascending order and take the x corresponding to the first k Euclidean distances j Form the nearest neighbor subset N k (x i ); d i,j =||x i -x j ||2,j=1,...,n;j≠i (2) d i,1 ≤d i,2 ≤…≤d i,k (3) N k (x i )={x i,1 ,x i,2 ,...,x i,k }∈R m×n (4) Where: x i,k is the cable force eigenvector x i The kth nearest neighbor vector of ; k is the number of nearest neighbor subsets, with all feature vectors x in the training set i The calculated range and standard deviation of the early warning index are used as the target, and the number of subsets corresponding to the minimum range and standard deviation is selected through the cross-validation method as the k value to be used; (2.3) According to formula (5), determine the nearest neighbor set G, and determine the set-based nearest path s as shown in formula (6). For all 1≤j≤k, p j+1 is a set {p1,p2,...,p j }in {p j+1 ,p j+2 ,...,p k+1 }; if the nearest neighbor is not unique, it is selected according to the order of the elements in the nearest neighbor set G; G=N k (x i )∪{x i }={x i ,x i,1 ,x i,2 ,...,x i,k } (5) s=<p1,p2,...,p k+1 >,p j ∈G (6) Where: The nearest neighbor set G and the nearest path s based on the set contain the same cable force eigenvector, that is, In addition, the first element of the nearest neighbor set G and the nearest path s is the same, that is, p1 = x i ; The nearest neighbor calculation method based on Mahalanobis distance is used to determine the distance between the vector and the set, and then determine the nearest path s. The calculation method is shown in formula (7). For any given x∈P, if y∈Q satisfies dist(x,y)=dist(P,Q), then y is the nearest neighbor sample of set P in set Q, thereby finding the shortest distance between a sample point and a set. dist(P,Q)=min{d M (x,y):x∈P&y∈Q} (7) in, d M (x,y)=(x-y) T ·S -1 ·(x-y) (8) In formula (8), S is the covariance matrix, and the cable force eigenvectors in the training set are used to calculate according to formula (9); Where: N is the number of cable force feature vectors in the training set, is the mean vector of all cable force eigenvectors in the training set; (2.4) Based on the determined cable force characteristic vector x i and its nearest neighbor N k (x i ) further determines the nearest trajectory based on the set <e1,e2,...,e k >, where e j For the connection set o j With the eigenvector p j+1 The shortest connection distance; E j =(o j ,p j+1 ),O j ∈{p1,...,p j } (10) Among them, e j The determination of needs to meet the requirements of formula (11); dist(e j )=dist(o j ,p j+1 )=min{d M (o ji ,p j+1 ):o ji ∈o j &1≤i≤j} (11) (2.5) Define the cable force characteristic vector x i Isolation Index DI k (x i ), and its calculation formula is shown in formula (12): (2.6) Repeat steps (2.2)-(2.5) to calculate the nearest neighbor subset {N k (o)|o∈N k (x i )}; Then, calculate each cable force eigenvector x according to formula (13): i Local variation index DI k (x i ); (2.7) The generalized extreme value distribution GEV is used in combination with the maximum block method to fit the distribution of the local variation index in the training set. The GEV distribution expression is shown in formula (14): The maximum block method selects the optimal number of blocks b by applying the goodness-of-fit metric through the KS hypothesis test, and then the set of local variation indices calculated from the training set Θ = [DI k (x1),DI k (x2),...,DI k (x m )] is divided into b equal-sized non-overlapping blocks, and the maximum value of each block is extracted for GEV distribution fitting; in order to determine the warning threshold τ under a certain significance level α α , it is necessary to calculate the extreme quantiles of the GEV distribution based on the estimated GEV parameters and according to formula (15); Where: ξ, σ, μ represent the shape, scale and location parameters of the GEV distribution; (2.8) During the online monitoring phase, repeat steps 1 and 2 to calculate each new cable force characteristic vector z i Local variation index DI k (z i ) and compare it with the warning threshold set in step (2.7), so as to perform online evaluation and abnormal warning of the inclined cable; Step 3. Isolation of abnormal cable force components of cable force eigenvector (3.1) In the online monitoring stage, when a new cable force characteristic vector z calculated in (2.8) i Local variation index DI k (z i ) exceeds the set warning threshold, the cable force characteristic vector z is reduced in sequence i A tension component in the k (z i ), when the calculated local variation index drops below the threshold, it indicates that the corresponding inclined cable of the cable force component has an abnormality.

Images