Intelligent defect analysis method based on bridge vibration data

Abstract

The invention relates to the technical field of bridge safety, in particular to an intelligent defect analysis method based on bridge vibration data, and the method comprises the steps: collecting vibration signals through an acceleration sensor, and carrying out the grouping recording and storage; performing single-point failure test processing and continuous failure test processing on the bridge vibration data; calculating frequency domain features, time domain features and statistical features of the bridge vibration data, inputting the frequency domain features, the time domain features and the statistical features into an anomaly recognition model, and outputting a defect detection result; and carrying out visualization processing on the defect detection result and the bridge structure part with the corresponding serial number. The method is used for improving the reliability and accuracy of a bridge vibration monitoring system and effectively improving the intelligent defect analysis capability of bridge vibration data, so that more accurate support is provided for maintenance and safety of a bridge.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge safety, and in particular to an intelligent defect analysis method based on bridge vibration data. Background Art

[0002] The safety and durability of bridge structures are crucial to the stable operation of public infrastructure. In recent years, with the aging of bridges and the increase in traffic loads, bridge health monitoring and defect identification have become key issues in bridge management and maintenance. Bridge vibration data is an important means of structural health monitoring (SHM). Through various sensors installed on the bridge, the dynamic response of the bridge under traffic loads and environmental influences is continuously monitored. By analyzing these vibration data, the structural status and potential defects of the bridge can be identified.

[0003] Traditional bridge defect detection methods mainly rely on manual inspection and regular maintenance. This method is not only costly and inefficient, but also unable to monitor the health of the bridge in real time. There may be a large amount of noise and outliers in the vibration data. How to effectively remove these interfering information in the data preprocessing stage and improve the accuracy of defect identification is still a problem that needs to be solved urgently. Summary of the invention

[0004] In order to solve the problems in the above-mentioned background technology, the present invention provides an intelligent defect analysis method based on bridge vibration data, which can measure the weight of the lower limbs, clarify the pressure detection distribution points, and remind patients when the weight is too heavy, thereby helping patients to walk underground with confidence and facilitating the rehabilitation of their lower limbs.

[0005] The present invention adopts the following technical solutions:

[0006] An intelligent defect analysis method based on bridge vibration data, the method comprising:

[0007] Several acceleration sensors are symmetrically arranged at the main beam, supporting point, mid-span, pier top and bridge tower of the bridge. Vibration signals are collected by the acceleration sensors and recorded and stored in groups. The same numbering rules are adopted to number the bridge structure parts with the same type of characteristic information and the corresponding acceleration sensors and the collected bridge vibration data.

[0008] Perform single-point failure inspection and continuous failure inspection on bridge vibration data to check whether there are sensor failures or abnormal readings in the data and identify possible continuous failure data in the data;

[0009] Calculating the frequency domain features, time domain features and statistical features of the bridge vibration data, inputting the frequency domain features, time domain features and statistical features into an abnormality recognition model, and outputting defect detection results;

[0010] The defect detection results and the corresponding numbered bridge structure parts are visualized to display bridge vibration data and analysis results in real time.

[0011] Furthermore, the single point failure detection processing method includes:

[0012] The bridge vibration data collected from the acceleration sensor is tested by a type of SVM model, and the training method of the type of SVM model is:

[0013] Collecting raw data from the acceleration sensor and normalizing the raw data to ensure that the data are on the same scale;

[0014] The original data is converted into a feature vector set {x1, x2, ..., x n}, the number of data points in the feature vector set is n, and each x i is a feature vector,

[0015] Construct the original form objective function, the formula of the objective function f(x) is:

[0016]

[0017] Where w is the normal vector of the hyperplane, ν is a hyperparameter used to control the flexibility of the model, between 0 and 1, and ξ is a slack variable used to allow some data points to be outside the hyperplane w;

[0018] Set the constraints of the original objective function. The constraint formula of the original objective function is:

[0019] <w,φ(x i )>≥1-ξ i ;

[0020] Among them, φ(x i ) is the data point x i High-dimensional feature maps;

[0021] A dual form objective function is constructed, and the formula of the dual form objective function g(x) is:

[0022]

[0023] Among them, α i , α j represents the Lagrange multiplier, K(x i ,x j ) represents a kernel function, and the kernel function is expressed as:

[0024]

[0025] σ represents the bandwidth parameter, which is used to control the width of the Gaussian distribution;

[0026] Set the constraint conditions of the dual form objective function. The formula of the constraint conditions of the dual form objective function is:

[0027] Calculate the bias term ρ:

[0028] A decision function is constructed, and the bridge vibration data to be detected is normalized and converted into data point x. The decision function formula is:

[0029]

[0030] Among them, sign is the sign function, K(x i ,x) is the similarity calculated by the kernel function; if k(x)=1, the data point x is considered to be a normal point, and the bridge vibration data is normal data; if k(x)=-1, the data point x is considered to be an abnormal point, and the bridge vibration data is abnormal data.

[0031] Furthermore, the method for continuous failure inspection processing includes:

[0032] Use sliding windows to check the continuity of vibration data for several bridges and identify sections with missing data;

[0033] The time series prediction model ARIMA is used to predict the bridge vibration data under normal conditions, and the deviation between the actual bridge vibration data and the predicted bridge vibration data is compared;

[0034] The bridge vibration data is identified by using the recurrent neural network model LSTM trained with the bridge vibration data under normal conditions;

[0035] Set dynamic thresholds to detect continuous abnormal data and dynamically adjust the thresholds based on historical bridge vibration data;

[0036] Recover continuously failed data through a deep neural network model.

[0037] Furthermore, the forecasting model ARIMA includes an autoregressive part, a differential part and a sliding average part, and the training method of the forecasting model ARIMA includes:

[0038] Perform differential processing on bridge vibration data to make it stable;

[0039] Determine the order of the autoregressive part, the difference part and the moving average part respectively; among them, the formula of the autoregressive model is:

[0040] Y t =φ1Y t-1+φ2Y t-2 +…+φ p Y t-p +∈ t ;

[0041] Among them, p represents the order of the autoregressive term, Y t represents the observed value at time point t, φ i represents the autoregressive coefficient, ∈ t represents the white noise term;

[0042] The difference part is used to transform the non-stationary sequence in the autoregressive part into a stationary sequence. The formula of the difference part is:

[0043] Δ d Y t =(1-B) d Y t ;

[0044] Among them, Δ represents the differential operation, B represents the lag operator, and BY t =Y t-1 , d is the difference order;

[0045] The formula for the sliding average part is:

[0046] Y t =μ+∈ t +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q ;

[0047] Among them, q represents the order of the sliding average term, μ represents the mean of the sequence, and θ i represents the sliding mean coefficient, ∈ t represents the white noise term;

[0048] The formula of the forecasting model ARIMA is:

[0049] φ(B)Δ d Y t =θ(B)∈ t ;

[0050] Where φ(B)=1-φ1B-φ2B 2 -…-φ p B p ; θ(B)=1+θ1B+θ2B 2 +…+θ q B q .

[0051] Furthermore, the method for calculating the frequency domain characteristics, time domain characteristics and statistical characteristics of the bridge vibration data includes:

[0052] The time domain signal of the bridge vibration data is converted into a frequency domain signal by Fourier transform, and the power spectrum density of the frequency domain signal is calculated to represent the power per unit frequency range. The calculation formula of the power spectrum density is:

[0053] S x (f) = |X(f)| 2 ;

[0054] Where X(f) is the frequency domain signal of the bridge vibration data;

[0055] Calculate the mean, variance and standard deviation of the time domain signal of bridge vibration data to describe the basic statistical properties of bridge vibration data;

[0056] The autocorrelation function is calculated to measure the similarity of the time series to its own lagged version. The formula of the autocorrelation function is:

[0057] Among them, τ represents the lag time, μ represents the mean, and σ represents the standard deviation;

[0058] The kurtosis is calculated to describe the sharpness of the data distribution. The calculation formula of the kurtosis is:

[0059]

[0060] The skewness is calculated to describe the symmetry of the data distribution. The calculation formula of the skewness is:

[0061]

[0062] The time dependence in the time series data is captured by the autocorrelation function, which is formulated as:

[0063] Among them, R k represents the autocorrelation function value of delay k, x i represents the data value at the i-th time point, x i+1 represents the data value at the i+kth time point, μ represents the mean, and N represents the total number of data points.

[0064] Furthermore, the training method of the anomaly recognition model includes:

[0065] The frequency domain features, time domain features and statistical features of the historical records are converted into original feature data sets, and standardized to obtain the data matrix X to ensure that the frequency domain features, time domain features and statistical features have the same scale. The standardized formula is:

[0066]

[0067] in represents the standardized eigenvalue, x i represents the original eigenvalue, μ i represents the mean of the feature, σ i represents the standard deviation of the feature;

[0068] The feature dimension is reduced by principal component analysis, and the covariance matrix between the features is calculated. The covariance matrix represents the linear relationship between different features of the data. The calculation formula of the covariance matrix is:

[0069]

[0070] Where n represents the number of samples in the original feature dataset, X T Denote the transpose of the data matrix X. For the covariance matrix C, find the eigenvalues ​​λ and eigenvectors v such that:

[0071] Cv=λv;

[0072] According to the size of the eigenvalue λ, the eigenvector v corresponding to the first k largest eigenvalues ​​is selected so that the first k largest eigenvalues ​​can explain 95% of the total variance;

[0073] Let W be the eigenvector corresponding to the first k largest eigenvalues, and the transformed data matrix X new for:

[0074] X new =XW.

[0075] Furthermore, the training method of the anomaly recognition model also includes:

[0076] Through the K-fold cross validation method, the transformed data matrix X new Divide into k subsets X containing the same number of samples k , for each subset, use the current subset as the validation set The remaining k-1 subsets are used as training sets The training set is trained by the KNN model based on the k-nearest neighbor algorithm, and the validation set corresponding to the training set is used for prediction. The accuracy of each subset is calculated. The output of each subset is set to The calculation formula of the accuracy is:

[0077]

[0078] in, Represents the indicator function, which has a value of 1 when the prediction is correct and 0 otherwise; the average of the performance indicators of all subsets is taken as the performance indicator of the final model:

[0079]

[0080] Select the KNN model parameters that correspond to the performance metric that is closest to the final model.

[0081] Furthermore, an electronic device includes:

[0082] processor;

[0083] a memory for storing processor-executable instructions;

[0084] The processor is configured to call the instructions stored in the memory to execute the above method.

[0085] Furthermore, a computer-readable storage medium stores computer program instructions, which implement the above method when executed by a processor.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] The present invention provides an intelligent defect analysis method based on bridge vibration data. By performing single-point failure inspection and continuous failure inspection on the bridge vibration data, the reliability and accuracy of the bridge vibration monitoring system are improved. The method calculates the frequency domain characteristics, time domain characteristics and statistical characteristics of the bridge vibration data, inputs the frequency domain characteristics, time domain characteristics and statistical characteristics into an abnormality recognition model, and outputs defect detection results. The defect detection results and the corresponding numbered bridge structure parts are visualized to realize real-time display of bridge vibration data and analysis results. By combining with an abnormality recognition algorithm, in-depth analysis and processing are performed to improve the effectiveness of feature selection and feature extraction, thereby improving the accuracy of the analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0089] Figure 1 This is a schematic diagram of an intelligent defect analysis method based on bridge vibration data of the present invention. DETAILED DESCRIPTION

[0090] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0091] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.

[0092] An intelligent defect analysis method based on bridge vibration data, the method comprising:

[0093] Several acceleration sensors are symmetrically arranged at the main beam, support point, mid-span, pier top and bridge tower of the bridge. The vibration signals are collected by the acceleration sensors and recorded and stored in groups. The same numbering rules are adopted to number the bridge structure parts with the same type of characteristic information and the corresponding acceleration sensors and the collected bridge vibration data; ensure that the collected vibration data covers the response of the entire bridge structure and can effectively reflect the overall health status of the bridge; each acceleration sensor collects the vibration signal of the bridge under the influence of vehicle load, wind force, earthquake and other environmental factors in real time; the collected vibration data is grouped according to the bridge structure parts and characteristic information. For example, all sensor data of the main beam part are divided into one group, and the data of the support point part are divided into another group, and so on. Each group of data can be more conveniently compared and analyzed in the same category; through numbering, the working status of each sensor can be tracked.

[0094] Single-point failure inspection and continuous failure inspection are performed on bridge vibration data to check whether there are sensor failures or abnormal readings in the data and to identify possible continuous failure data in the data; the data of the acceleration sensor may be affected by its own failure, environmental noise or external interference, resulting in data failure. In order to ensure the accuracy and reliability of the data, single-point failure inspection and continuous failure inspection are usually required for bridge vibration data. These two inspection methods can help identify sensor failures or abnormal readings and detect possible continuous failure areas in the data.

[0095] The frequency domain features, time domain features and statistical features of the bridge vibration data are calculated, and the frequency domain features, time domain features and statistical features are input into the anomaly recognition model to output defect detection results; the frequency domain features, time domain features and statistical features of the bridge vibration data can effectively reflect the dynamic response of the bridge structure and help identify potential structural defects. The anomaly recognition model uses an optimized K-nearest neighbor algorithm for identification, and constructs the frequency domain features, time domain features and statistical features into a feature vector for input. The K-nearest neighbor algorithm learns the vibration characteristics of the bridge under normal conditions, classifies the newly collected bridge vibration data, identifies abnormal situations, and classifies different types of defects to help carry out targeted maintenance and repair.

[0096] The defect detection results and the corresponding numbered bridge structure parts are visualized, and the bridge vibration data and analysis results are displayed in real time; the vibration data, detection results and structure part numbers are integrated, and a simplified plan view or three-dimensional model of the bridge is used to intuitively display the layout and numbering of various parts of the bridge. A real-time vibration data graph is drawn for each sensor to show the changing trend of the vibration data. The detection results are displayed on the vibration data graph with different colors or marks to distinguish normal data from abnormal data, and the status prompts of various structural parts of the bridge are displayed.

[0097] Specifically, the single point failure detection process is used to detect abnormal data points caused by sensor failure or external interference, effectively helping the bridge monitoring system to filter out unreliable data and ensure the accuracy of the analysis results. The single point failure detection process method includes:

[0098] The bridge vibration data collected from the acceleration sensor is tested by a type of SVM model, and the training method of the type of SVM model is:

[0099] Collecting raw data from the acceleration sensor and normalizing the raw data to ensure that the data are on the same scale and avoid the influence of feature value range differences on model training;

[0100] The original data is converted into a feature vector set {x1, x2, ..., x n}, the number of data points in the feature vector set is n, and each x i is a feature vector,

[0101] Construct the original form objective function, the formula of the objective function f(x) is:

[0102]

[0103] Where w is the normal vector of the hyperplane, ν is a hyperparameter used to control the flexibility of the model, between 0 and 1, and ξ is a slack variable used to allow some data points to be outside the hyperplane w;

[0104] Set the constraints of the original objective function. The constraint formula of the original objective function is:

[0105] <w,φ(x i )>≥1-ξ i ;

[0106] Among them, φ(x i ) is the data point x i High-dimensional feature maps;

[0107] A dual form objective function is constructed, and the formula of the dual form objective function g(x) is:

[0108]

[0109] Among them, α i , α j represents the Lagrange multiplier, K(x i ,x j ) represents a kernel function, and the kernel function is expressed as:

[0110]

[0111] σ represents the bandwidth parameter, which is used to control the width of the Gaussian distribution;

[0112] Set the constraint conditions of the dual form objective function. The formula of the constraint conditions of the dual form objective function is:

[0113] Calculate the bias term ρ:

[0114] A decision function is constructed, and the bridge vibration data to be detected is normalized and converted into data point x. The decision function formula is:

[0115]

[0116] Among them, sign is the sign function, K(x i,x) is the similarity calculated by the kernel function; if k(x)=1, the data point x is considered to be a normal point, and the bridge vibration data is normal data; if k(x)=-1, the data point x is considered to be an abnormal point, and the bridge vibration data is abnormal data; by constructing a hyperplane in a high-dimensional space to separate normal and abnormal data, and combining the kernel function, the model can handle nonlinear boundary problems. After the new bridge vibration data is input into the model, it is normalized and the state of each bridge vibration data is judged by the decision function, which provides a reliable means of monitoring bridge vibration data and helps to timely detect data anomalies caused by sensor failure or external interference.

[0117] Specifically, the continuous failure detection process is used to identify continuous data failures caused by sensor failure, data transmission interruption or external interference, so as to ensure the integrity and reliability of the data. The method of the continuous failure detection process includes:

[0118] Use sliding windows to check the continuity of vibration data for several bridges and identify sections with missing data;

[0119] The time series prediction model ARIMA is used to predict the bridge vibration data under normal conditions, and the deviation between the actual bridge vibration data and the predicted bridge vibration data is compared;

[0120] Bridge vibration data is identified by using a recursive neural network model LSTM that is trained with bridge vibration data under normal conditions. LSTM is a recursive neural network (RNN) that is suitable for processing and predicting time series data, especially capturing the long-term dependency of data. The LSTM model is trained using bridge vibration data under normal conditions so that the model can learn the normal pattern of bridge vibration data. The trained LSTM model is used to analyze real-time bridge vibration data, and the mean square error is calculated to measure the deviation between the actual data and the LSTM predicted data. If the loss value of the bridge vibration data exceeds the set threshold, it is marked as abnormal data.

[0121] Set a dynamic threshold to detect continuous abnormal data, dynamically adjust the threshold based on historical bridge vibration data; use historical normal bridge vibration data to calculate the dynamic threshold. Dynamically adjust the threshold based on data change rate, fluctuation amplitude and other characteristics, and use the dynamic threshold to detect continuous abnormal data segments in bridge vibration data. If the deviation of multiple consecutive points in the data segment exceeds the dynamic threshold, the detection formula is:

[0122] Threshold=μ t +κσ t ;

[0123] μ t and σ tare the mean and standard deviation of the historical data at time t, respectively. κ is an adjustment parameter used to control the sensitivity of the threshold. A dynamic threshold is used to detect continuous abnormal data segments in bridge vibration data. If the deviation of multiple consecutive points in the data segment exceeds the dynamic threshold, it is marked as continuous abnormal data.

[0124] The continuous failure data is recovered through the deep neural network model. The deep neural network is a neural network with a multi-layer structure that can learn the complex nonlinear relationship of data. It is suitable for the task of recovering failed data. The DNN model is trained with complete normal data and partially simulated failure data to learn how to recover complete data from partially lost data. For the detected continuous failure data segments, they are input into the trained DNN model to recover the missing data.

[0125] Specifically, the forecasting model ARIMA includes an autoregressive part, a differential part and a moving average part, and the training method of the forecasting model ARIMA includes:

[0126] Perform differential processing on the bridge vibration data to make it stable; use historical normal bridge vibration data to train the ARIMA model;

[0127] Determine the order of the autoregressive part, the difference part and the moving average part respectively; among them, the formula of the autoregressive model is:

[0128] Y t =φ1Y t-1 +φ2Y t-2 +…+φ p Y t-p +∈ t ;

[0129] Among them, p represents the order of the autoregressive term, Y t represents the observed value at time point t, φ i represents the autoregressive coefficient, ∈ t represents the white noise term;

[0130] The difference part is used to transform the non-stationary sequence in the autoregressive part into a stationary sequence. The formula of the difference part is:

[0131] Δ d Y t =(1-B) d Y t ;

[0132] Among them, Δ represents the differential operation, B represents the lag operator, and BY t =Y t-1 , d is the difference order;

[0133] The formula for the sliding average part is:

[0134] Y t =μ+∈ t +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q ;

[0135] Among them, q represents the order of the sliding average term, μ represents the mean of the sequence, and θ i represents the sliding mean coefficient, ∈ t represents the white noise term;

[0136] The formula of the forecasting model ARIMA is:

[0137] φ(B)Δ d Y t =θ(B)∈ t ;

[0138] Where φ(B)=1-φ1B-φ2B 2 -…-φ p B p ; θ(B)=1+θ1B+θ2B 2 +…+θ q B q .

[0139] The Autoregressive Integrated Moving Average (ARIMA) model is a statistical model used for time series analysis that can capture linear trends, seasonality, and randomness in time series. The input bridge vibration data is differentially processed, and the time point data is predicted using the ARIMA model to compare the deviation between the predicted data and the actual data.

[0140] Specifically, the method for calculating the frequency domain characteristics, time domain characteristics and statistical characteristics of the bridge vibration data includes:

[0141] The time domain signal of the bridge vibration data is converted into a frequency domain signal by Fourier transform, and the power spectrum density of the frequency domain signal is calculated to represent the power per unit frequency range. The calculation formula of the power spectrum density is:

[0142] S x (f)=|X(f)| 2 ;

[0143] Where X(f) is the frequency domain signal of the bridge vibration data;

[0144] Calculate the mean, variance and standard deviation of the time domain signal of bridge vibration data to describe the basic statistical properties of bridge vibration data;

[0145] The autocorrelation function is calculated to measure the similarity of the time series to its own lagged version. The formula of the autocorrelation function is:

[0146] Among them, τ represents the lag time, μ represents the mean, and σ represents the standard deviation;

[0147] The kurtosis is calculated to describe the sharpness of the data distribution. The calculation formula of the kurtosis is:

[0148]

[0149] The skewness is calculated to describe the symmetry of the data distribution. The calculation formula of the skewness is:

[0150]

[0151] The time dependence in the time series data is captured by the autocorrelation function, which is formulated as:

[0152] Among them, R k represents the autocorrelation function value of delay k, x i represents the data value at the i-th time point, x i+1 represents the data value at the i+kth time point, μ represents the mean, and N represents the total number of data points.

[0153] Specifically, the training method of the anomaly recognition model includes:

[0154] The frequency domain features, time domain features and statistical features of the historical records are converted into original feature data sets, and standardized to obtain the data matrix X to ensure that the frequency domain features, time domain features and statistical features have the same scale. The standardized formula is:

[0155]

[0156] in represents the standardized eigenvalue, x i represents the original eigenvalue, μ i represents the mean of the feature, σ i represents the standard deviation of the feature;

[0157] The principal component analysis is used to reduce the feature dimension, convert the high-dimensional data into a low-dimensional space, retain the main variability in the data, and calculate the covariance matrix between the features. The covariance matrix represents the linear relationship between different features of the data. The calculation formula of the covariance matrix is:

[0158]

[0159] Where n represents the number of samples in the original feature dataset, XT Denote the transpose of the data matrix X. For the covariance matrix C, find the eigenvalues ​​λ and eigenvectors v such that:

[0160] Cv=λv;

[0161] According to the size of the eigenvalue λ, the eigenvector v corresponding to the first k largest eigenvalues ​​is selected so that the first k largest eigenvalues ​​can explain 95% of the total variance;

[0162] Let W be the eigenvector corresponding to the first k largest eigenvalues, and the transformed data matrix X new for:

[0163] X new =XW;

[0164] This reduces the computational complexity and improves the generalization ability of the model.

[0165] Specifically, the training method of the anomaly recognition model also includes:

[0166] Through the K-fold cross validation method, the transformed data matrix X new Divide into k subsets X containing the same number of samples k , for each subset, use the current subset as the validation set The remaining k-1 subsets are used as training sets The training set is trained by the KNN model based on the k-nearest neighbor algorithm, and the validation set corresponding to the training set is used for prediction. The accuracy of each subset is calculated. The output of each subset is set to The calculation formula of the accuracy is:

[0167]

[0168] in, Represents the indicator function, which has a value of 1 when the prediction is correct and 0 otherwise; the average of the performance indicators of all subsets is taken as the performance indicator of the final model:

[0169]

[0170] Select the KNN model parameters corresponding to the performance index closest to the final model; the training method of KNN is to define the abnormal category and set the standard for excessive vibration: set a threshold, and when the vibration value of the abnormal data point exceeds the threshold, the vibration is considered to be too large; use frequency domain features, time domain features and statistical features as the input of the KNN model, select Euclidean distance as the distance metric of KNN, and select the K value with the highest average classification accuracy through K-fold cross validation. For each abnormal data point to be classified, calculate its distance from each data point in the training set and find its closest K neighbors; count the categories of these K neighbors, and classify the abnormal data points into the most categories, so as to identify whether the vibration of the bridge vibration data is too large.

[0171] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0172] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein.

[0173] The present invention is further described above with the aid of specific embodiments. However, it should be understood that the specific description here should not be construed as limiting the essence and scope of the present invention. Various modifications made to the above embodiments by ordinary technicians in the field after reading this specification are all within the scope of protection of the present invention.

Claims

1. An intelligent defect analysis method based on bridge vibration data, characterized in that: The method comprises, Several acceleration sensors are symmetrically arranged at the main beam, supporting point, mid-span, pier top and bridge tower of the bridge. Vibration signals are collected by the acceleration sensors and recorded and stored in groups. The same numbering rules are adopted to number the bridge structure parts with the same type of characteristic information and the corresponding acceleration sensors and the collected bridge vibration data. Perform single-point failure inspection and continuous failure inspection on bridge vibration data to check whether there are sensor failures or abnormal readings in the data and identify possible continuous failure data in the data; Calculating the frequency domain features, time domain features and statistical features of the bridge vibration data, inputting the frequency domain features, time domain features and statistical features into an abnormality recognition model, and outputting defect detection results; The defect detection results and the corresponding numbered bridge structure parts are visualized to display bridge vibration data and analysis results in real time.

2. The method according to claim 1, characterized in that The single point failure detection processing method includes: The bridge vibration data collected from the acceleration sensor is tested by a type of SVM model, and the training method of the type of SVM model is: Collecting raw data from the acceleration sensor and normalizing the raw data to ensure that the data are on the same scale; The original data is converted into a feature vector set {x1, x2, ..., x n }, the number of data points in the feature vector set is n, and each x i is a feature vector, Construct the original form objective function, the formula of the objective function f(x) is: Where w is the normal vector of the hyperplane, ν is a hyperparameter used to control the flexibility of the model, between 0 and 1, and ξ is a slack variable used to allow some data points to be outside the hyperplane w; Set the constraints of the original objective function. The constraint formula of the original objective function is: <w,φ(x i )>≥1-ξ i ; Among them, φ(x i ) is the data point x i High-dimensional feature maps; A dual form objective function is constructed, and the formula of the dual form objective function g(x) is: Among them, α i , α j represents the Lagrange multiplier, K(x i ,x j ) represents a kernel function, and the kernel function is expressed as: σ represents the bandwidth parameter, which is used to control the width of the Gaussian distribution; Set the constraint condition of the dual form objective function. The formula of the constraint condition of the dual form objective function is: Calculate the bias term ρ: A decision function is constructed, and the bridge vibration data to be detected is normalized and converted into data point x. The decision function formula is: Among them, sign is the sign function, K(x i ,x) is the similarity calculated by the kernel function; if k(x)=1, the data point x is considered to be a normal point, and the bridge vibration data is normal data; if k(x)=-1, the data point x is considered to be an abnormal point, and the bridge vibration data is abnormal data.

3. The method according to claim 1, characterized in that The method of continuous failure inspection processing includes: Use sliding windows to check the continuity of vibration data for several bridges and identify sections with missing data; The time series prediction model ARIMA is used to predict the bridge vibration data under normal conditions, and the deviation between the actual bridge vibration data and the predicted bridge vibration data is compared; The bridge vibration data is identified by using the recurrent neural network model LSTM trained with the bridge vibration data under normal conditions; Set dynamic thresholds to detect continuous abnormal data and dynamically adjust the thresholds based on historical bridge vibration data; Recover continuously failed data through a deep neural network model.

4. The method according to claim 3, characterized in that The forecasting model ARIMA includes an autoregressive part, a differential part and a sliding average part. The training method of the forecasting model ARIMA includes: Perform differential processing on the bridge vibration data to make it stable; use historical normal bridge vibration data to train the ARIMA model; Determine the order of the autoregressive part, the difference part, and the moving average part; the formula of the autoregressive model is: Among them, p represents the order of the autoregressive term, Y t represents the observed value at time point t, represents the autoregressive coefficient, ∈ t represents the white noise term; The difference part is used to transform the non-stationary sequence in the autoregressive part into a stationary sequence. The formula of the difference part is: Δ d AND t (1-B) d AND t ; Among them, Δ represents the differential operation, B represents the lag operator, and BY t =Y t-1 , d is the difference order; The formula for the sliding average part is: Y t =μ+∈ t +θ1∈ t-1 +θ2∈ t-2 +…+θ q ∈ t-q ; Among them, q represents the order of the sliding average term, μ represents the mean of the sequence, and θ i represents the sliding mean coefficient, ∈ t represents the white noise term; The formula of the forecasting model ARIMA is: in .

5. The method according to claim 1, characterized in that The method for calculating the frequency domain characteristics, time domain characteristics and statistical characteristics of the bridge vibration data includes: The time domain signal of the bridge vibration data is converted into a frequency domain signal by Fourier transform, and the power spectrum density of the frequency domain signal is calculated to represent the power per unit frequency range. The calculation formula of the power spectrum density is: S x (f)=|X(f)| 2 ; Where X(f) is the frequency domain signal of the bridge vibration data; Calculate the mean, variance and standard deviation of the time domain signal of bridge vibration data to describe the basic statistical properties of bridge vibration data; The autocorrelation function is calculated to measure the similarity of the time series to its own lagged version. The formula of the autocorrelation function is: Among them, τ represents the lag time, μ represents the mean, and σ represents the standard deviation; The kurtosis is calculated to describe the sharpness of the data distribution. The calculation formula of the kurtosis is: The skewness is calculated to describe the symmetry of the data distribution. The calculation formula of the skewness is: The time dependence in the time series data is captured by the autocorrelation function, which is formulated as: Among them, R k represents the autocorrelation function value of delay k, x i represents the data value at the i-th time point, x i+1 represents the data value at the i+kth time point, μ represents the mean, and N represents the total number of data points.

6. The method according to claim 1, characterized in that The training method of the anomaly recognition model includes: The frequency domain features, time domain features and statistical features of the historical records are converted into original feature data sets, and standardized to obtain the data matrix X to ensure that the frequency domain features, time domain features and statistical features have the same scale. The standardized formula is: in represents the standardized eigenvalue, x i represents the original eigenvalue, μ i represents the mean of the feature, σ i represents the standard deviation of the feature; The feature dimension is reduced by principal component analysis, and the covariance matrix between the features is calculated. The covariance matrix represents the linear relationship between different features of the data. The calculation formula of the covariance matrix is: Where n represents the number of samples in the original feature dataset, X T Denote the transpose of the data matrix X. For the covariance matrix C, find the eigenvalues ​​λ and eigenvectors v such that: Cv=λv; According to the size of the eigenvalue λ, the eigenvector v corresponding to the first k largest eigenvalues ​​is selected so that the first k largest eigenvalues ​​can explain 95% of the total variance; Let W be the eigenvector corresponding to the first k largest eigenvalues, and the transformed data matrix X new for: X new =XW。 7. The method according to claim 6, characterized in that The training method of the anomaly recognition model also includes: Through the K-fold cross validation method, the transformed data matrix X new Divide into k subsets X containing the same number of samples k , for each subset, use the current subset as the validation set The remaining k-1 subsets are used as training sets The training set is trained by the KNN model based on the k-nearest neighbor algorithm, and the validation set corresponding to the training set is used for prediction. The accuracy of each subset is calculated. The output of each subset is set to The calculation formula of the accuracy is: in, Represents the indicator function, which has a value of 1 when the prediction is correct and 0 otherwise; the average of the performance indicators of all subsets is taken as the performance indicator of the final model: Select the KNN model parameters that correspond to the performance metric that is closest to the final model.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images