Mass fatigue damage monitoring data anomaly identification method based on time series data reconstruction and adaptive threshold

Abstract

The invention relates to the technical field of bridge long-term health monitoring, and discloses a mass fatigue damage monitoring data abnormity identification method based on time series data reconstruction and an adaptive threshold, and the method comprises the steps: obtaining normal data from original bridge strain data of a monitoring target, and training a time series data reconstruction model based on a deep learning network; performing high-precision reconstruction on possible abnormal data in the original bridge strain data of the monitoring target by using a time sequence data reconstruction model to obtain a transient energy response decision coefficient of data of each time period in a plurality of time periods on a frequency domain and a strain amplitude response decision coefficient of data of each time period on a time domain; and then clustering the mahalanobis distances between the data, and dynamically and accurately identifying real abnormal data in the possible abnormal data of the monitoring target by combining a time domain-frequency domain two-dimensional adaptive abnormal threshold. According to the method, the abnormal recognition of the mass fatigue damage monitoring data can be efficiently and precisely realized, and the generalization ability is relatively high.

Description

Technical Field

[0001] The present invention relates to the technical field of long-term health monitoring of bridges, and in particular to a method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds. Background Art

[0002] The complex welding structure and local stress conditions lead to prominent fatigue problems of orthotropic steel bridge decks under repeated vehicle loads. Compared with numerical analysis and indoor tests, on-site monitoring can truly reflect the fatigue response caused by variable-amplitude vehicle loads while taking into account the random traffic distribution and structural status of each lane. The fatigue life reliability analysis of welding details based on on-site monitoring data is more in line with reality and has higher accuracy. However, the data collected by the monitoring system often contains various types of data signal anomalies caused by sensor failures, system failures and environmental influences. These anomalies will mask the real information contained in the data and cause the fatigue life assessment to deviate from the true value. Therefore, it is necessary to improve the availability and effectiveness of data through data anomaly identification.

[0003] In recent years, the detection of abnormalities in time series signals based on deep learning methods has been widely used in many industrial fields. Some methods convert time series signals into image signals and build image heterogeneity classification models based on deep neural networks to identify abnormalities. Some methods deeply analyze data, identify possible data abnormality types, create data abnormality labels, and build multi-label classification deep neural network models to achieve abnormality recognition.

[0004] These methods have encountered major obstacles when applied to large-scale fatigue monitoring data sets. On the one hand, the sampling frequency of fatigue strain monitoring of orthotropic steel bridge decks is generally not less than 100 Hz, and the number of samples for only one channel can reach 8.64 million per day. Conventional computers cannot meet the computing power and storage requirements for deep learning after converting fatigue strain time signals into image signals. On the other hand, due to the randomness of vehicle loads, when lighter vehicles pass by, the strain response signal-to-noise ratio is low, and when overweight vehicles pass by, the strain may have extreme value responses. Deep learning methods may lead to overfitting of data with low signal-to-noise ratios and insufficient model generalization ability. Overweight vehicles only account for a very small part of all passing vehicles. Deep learning is difficult to learn the characteristics of high strain responses from limited data, which may cause the model to mistakenly judge high strain response data as outliers, while such strains contribute much more to the degree of fatigue damage than low strains. In addition, the performance of sensors and acquisition equipment may gradually decline, resulting in the emergence of new types of anomalies, and fixed anomaly thresholds may lead to a decrease in the accuracy of anomaly judgment.

[0005] Therefore, a new method for anomaly identification of massive fatigue damage monitoring data is urgently needed. Summary of the invention

[0006] The present invention provides a method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds, so as to solve the defect that the prior art fails to efficiently and accurately identify abnormal data when facing massive fatigue damage monitoring data.

[0007] A method for constructing a time series data reconstruction model, comprising: Obtaining the original bridge strain data of the monitoring target; Preprocessing the original bridge strain data of the monitoring target to obtain normal data in the original bridge strain data; According to the normal data in the original bridge strain data, a time series data reconstruction model of the monitoring target is trained.

[0008] According to a method for constructing a time series data reconstruction model provided by the present invention, the original bridge strain data of the monitoring target is preprocessed to obtain normal data in the original bridge strain data, including: Subdividing the original bridge strain data of the monitoring target into multiple data segments according to a preset time period; The number of stress cycles and the maximum amplitude response of each data segment in multiple data segments are counted by using the rain flow counting method; According to the stress cycle number and the maximum amplitude response of each data segment in the multiple data segments, a mean value and a standard deviation of the stress cycle number and a mean value and a standard deviation of the maximum amplitude response of each data segment in the multiple data segments are obtained; According to the mean and standard deviation of the number of stress cycles and the mean and standard deviation of the maximum amplitude response of each data segment in multiple data segments, the original bridge strain data of the data segments whose number of stress cycles is within one times the standard deviation of their mean and whose maximum amplitude response is within one times the standard deviation of their mean are taken as normal data in the original bridge strain data.

[0009] According to a method for constructing a time series data reconstruction model provided by the present invention, the original bridge strain data of the monitoring target is preprocessed to obtain normal data in the original bridge strain data, including: Using adaptive modal decomposition to remove the influence of daily temperature changes on normal data in the original bridge strain data; and / or, Wavelet packet analysis is used to reduce the noise of normal data in the original bridge strain data.

[0010] According to a method for constructing a time series data reconstruction model provided by the present invention, the deep learning network of the time series data reconstruction model includes a maximum overlap discrete wavelet transform module, a convolution module, a channel weight allocation module, an attention mechanism module, and a fully connected module, wherein the function used in the maximum overlap discrete wavelet transform module is a db6 wavelet function; the convolution module includes three groups of convolution layers and Batchnorm normalization layers connected in series, the convolution kernel sizes are set to 7, 5, and 3 respectively, the convolution step sizes are all 1, and the number of output channels of each convolution layer is 32, 64, and 128 respectively; the channel weight allocation module includes three groups of first global maximum pooling (GlobalMaxPooling) layers connected in series, first self-attention (se The self-attention layer adopts a multi-head attention with a size of 16, and the output channel is 128; the attention mechanism module is composed of a bi-LSTM layer, a Layernorm normalization layer, a second global maximum pooling layer, a second self-attention layer, a second sigmoid activation layer and a multiplication layer in series, wherein the bi-LSTM layer has an output channel of 128, and the second self-attention layer adopts a multi-head attention with a size of 16, and the output channel is 128; the fully connected module includes 2 layers of fully connected layers in series, and the number of output channels is 64, 32, and 1 respectively, which is used to map features to the target output domain.

[0011] According to a method for constructing a time series data reconstruction model provided by the present invention, normal data in the original bridge strain data is used to train a deep learning network of the time series data reconstruction model. The input size of the deep learning network of the time series data reconstruction model is 1 dimension, 2000 data points, the loss function of the time series data reconstruction model is a root mean square error function, the optimization function of the time series data reconstruction model is an Adam optimization function, and the regularization factor of the time series data reconstruction model is 0.0002.

[0012] The present invention provides a method for identifying anomalies in massive fatigue damage monitoring data, comprising: The original bridge strain data of the monitoring target and the normal data in the original bridge strain data obtained according to the construction method of the time series data reconstruction model described in any one of the above items are used to obtain possible abnormal data in the original bridge strain data of the monitoring target, wherein the possible abnormal data in the original bridge strain data of the monitoring target includes possible abnormal data in multiple time periods; According to possible abnormal data in the original bridge strain data of the monitoring target, a time series data reconstruction model is obtained by any of the above-mentioned time series data reconstruction model construction methods to obtain reconstructed data; According to the possible abnormal data and the reconstructed data in the original bridge strain data of the monitoring target, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period and the corresponding reconstructed data are obtained; According to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods is obtained; According to the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of each time period in multiple time periods, the Mahalanobis distances between the determination coefficients of the transient energy response in the frequency domain and the determination coefficients of the strain amplitude response in the time domain of all time periods in multiple time periods are clustered to obtain the real abnormal data among the possible abnormal data in the original bridge strain data of the monitoring target.

[0013] According to a method for identifying abnormalities in massive fatigue damage monitoring data provided by the present invention, the original bridge strain data of the monitoring target and normal data in the original bridge strain data obtained by the method for constructing a time series data reconstruction model according to any one of the above items, and possible abnormal data in the original bridge strain data of the monitoring target are obtained, including: According to the construction method of the time series data reconstruction model described in any of the above items, the original bridge strain data of the monitoring target and the normal data in the original bridge strain data are obtained, and the data other than the normal data in the original bridge strain data of the monitoring target are used as possible abnormal data in the original bridge strain data of the monitoring target.

[0014] According to a method for identifying abnormalities in massive fatigue damage monitoring data provided by the present invention, the method obtains the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of the monitoring target and the reconstructed data corresponding to the possible abnormal data in the original bridge strain data of each time period in multiple time periods, including: According to the possible abnormal data and reconstructed data in the original bridge strain data of the monitoring target, using the first expression, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period and its corresponding reconstructed data are obtained; The first expression includes: , , In the first expression, Indicates The transient energy response determination coefficient of the original bridge strain data and its corresponding reconstructed data in the frequency domain for each period is: Indicates The strain amplitude response determination coefficient of the original bridge strain data and its corresponding reconstructed data in the time domain, Represents the normalized The original bridge strain data for each period, Represents the normalized The reconstructed data corresponding to the original bridge strain data for each period, Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Indicates The number of sampling points of possible abnormal data in the original bridge strain data of each period, Represents a frequency domain number.

[0015] According to a method for identifying abnormalities in massive fatigue damage monitoring data provided by the present invention, the method obtains the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods according to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, including: According to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods is obtained by using the second expression; The second expression includes: , , In the second expression, Indicates The Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain for each time period, Indicates The transient energy response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the frequency domain, Indicates The strain amplitude response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the time domain, express and The covariance of express and The matrix, Z = ], express.

[0016] According to a method for identifying abnormalities in massive fatigue damage monitoring data provided by the present invention, the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain of each time period in multiple time periods are clustered according to the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain of all time periods in the multiple time periods, so as to obtain the real abnormal data in the possible abnormal data in the original bridge strain data of the monitoring target, including: According to the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of each time period in the multiple time periods, the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of all time periods in the multiple time periods are clustered to obtain abnormal clustering clusters; The possible abnormal data in the original bridge strain data of the monitoring target corresponding to the abnormal cluster is taken as the real abnormal data in the possible abnormal data in the original bridge strain data of the monitoring target.

[0017] The present invention provides a method for identifying abnormalities in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds. Normal data is obtained from the original bridge strain data of the monitoring target and used to train a time series data reconstruction model based on a deep learning network. There is no need to convert the fatigue strain time signal into an image signal. The time series data reconstruction model directly learns the time domain characteristics and frequency domain characteristics of the time series to achieve efficient and high-precision time series data reconstruction. The time series data reconstruction model is then used to reconstruct the possible abnormal data in the original bridge strain data of the monitoring target with high precision. Then, based on the possible abnormal data and the reconstructed data in the original bridge strain data, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in multiple time periods are obtained. These are used as abnormal evaluation indicators, and then the Mahalanobis distance between the two abnormal evaluation indicators is clustered. The real abnormal data in the possible abnormal data of the monitoring target is dynamically and accurately identified by combining the dual-dimensional indicators of the time domain and the frequency domain. The present invention can efficiently and accurately realize the abnormal identification of massive fatigue damage monitoring data, and has strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or 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 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.

[0019] Figure 1 A schematic flow chart of a method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds provided by the present invention.

[0020] Figure 2 An example diagram showing how to use the adaptive modal decomposition method to remove the impact of daily temperature changes on normal data in the original bridge strain data.

[0021] Figure 3 A structural diagram of a deep learning network of a time series data reconstruction model provided by the present invention.

[0022] Figure 4 (a) shows the input value and reconstructed value of normal data in the bridge strain data of the 242th period, Figure 4 (b) shows the frequency domain characteristics obtained by continuous wavelet transform of the input value and reconstructed value of the original possible abnormal data in the bridge strain data of the 242th period, and Figure 4 (c) shows the correlation between the input value and reconstructed value of the original possible abnormal data in the bridge strain data of the 672th period in the frequency domain (left) and time domain (right).

[0023] FIG. 5( a ) shows an example diagram of the distribution of the determination coefficient of the strain amplitude response in the time domain, and FIG. 5( b ) shows an example diagram of the distribution of the determination coefficient of the transient energy response in the frequency domain. FIG6(a) shows an example diagram of the clustering result of the Mahalanobis clustering between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain, and FIG6(b) shows the non-abnormal data (black line) and the real abnormal data (red line) among the possible abnormal data. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0025] Combine the following Figure 1 - Figure 6 describes the method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds provided by the present invention. It should be noted that the execution subject of the method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds provided by the present invention can be any network side device / terminal side device that meets the technical requirements, such as a model building device, a data anomaly identification device, etc.

[0026] Figure 1 The present invention provides a flow chart of a method for identifying anomalies in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds. Before performing the method for identifying anomalies in massive fatigue damage monitoring data on a monitoring target, a time series data reconstruction model of the monitoring target needs to be constructed first.

[0027] The method for constructing a time series data reconstruction model of a monitoring target includes the following steps: S101. Acquire original bridge strain data of a monitoring target, wherein the original bridge strain data may be original bridge strain data of the monitoring target within a period of time (eg, one day), which is time series data.

[0028] S102, preprocessing the original bridge strain data of the monitoring target to obtain normal data in the original bridge strain data. In this embodiment, S102 may include: The original bridge strain data of the monitoring target is subdivided into multiple data segments according to a preset time period (e.g., 1 hour); The number of stress cycles and the maximum amplitude response of each data segment in multiple data segments are counted by using the rain flow counting method; According to the stress cycle number and the maximum amplitude response of each data segment in the multiple data segments, a mean value and a standard deviation of the stress cycle number and a mean value and a standard deviation of the maximum amplitude response of each data segment in the multiple data segments are obtained; According to the mean and standard deviation of the number of stress cycles and the mean and standard deviation of the maximum amplitude response of each data segment in the multiple data segments, the original bridge strain data of the data segment whose number of stress cycles is within one times the standard deviation of its mean and whose maximum amplitude response is within one times the standard deviation of its mean is taken as the normal data in the original bridge strain data; Adaptive modal decomposition is used to remove the influence of daily temperature changes on normal data in the original bridge strain data. Daily temperature changes mainly affect the average strain change, and have little effect on the stress spectrum caused by vehicle loads, so they have little effect on fatigue damage caused by vehicle loads. Eliminating the strain response trend changes caused by daily temperature effects can improve the learning efficiency of deep learning and the accuracy of data reconstruction (see Figure 2 ); Wavelet packet analysis is used to reduce the noise of normal data in the original bridge strain data.

[0029] S103. According to the normal data in the original bridge strain data (in this embodiment, the normal data is divided into training data and test data at a ratio of 8:2), a time series data reconstruction model of the monitoring target is trained.

[0030] See also Figure 3In this embodiment, the deep learning network of the time series data reconstruction model includes a maximum overlap discrete wavelet transform module, a convolution module, a channel weight allocation module, an attention mechanism module, and a fully connected module, wherein the maximum overlap discrete wavelet transform module is used to expand the time series data according to the frequency domain characteristics, and the function adopted is the db6 wavelet function; the convolution module is used to strengthen the local characteristics of the time series data in the time dimension, and the convolution module includes three groups of convolution layers and Batchnorm normalization layers connected in series, and the convolution kernel sizes are set to 7, 5, and 3 respectively, and the convolution step size is 1. The number of output channels of each convolution layer is 32, 64, and 128 respectively; the channel weight allocation module is used to dynamically adjust the weight coefficient of each channel, and the channel weight allocation module includes three groups of first global maximum pooling (GlobalMaxPooling) layers connected in series, first self-attention (se The self-attention layer adopts a multi-head attention with a size of 16 and an output channel of 128. The attention module is used to capture and remember long-term dependencies in time series data and optimize global dependencies, so as to improve the model's focus on high-strain responses. The attention mechanism module is composed of a bi-LSTM layer, a Layernorm normalization layer, a second global maximum pooling layer, a second self-attention layer, a second sigmoid activation layer and a multiplication layer in series, wherein the bi-LSTM layer has an output channel of 128, and the second self-attention layer adopts a multi-head attention with a size of 16 and an output channel of 128. The fully connected module includes two fully connected layers in series, with the number of output channels being 64, 32 and 1 respectively, which are used to map features to the target output domain. The input size of the deep learning network of the time series data reconstruction model is 1 dimension and 2000 data points. The loss function of the time series data reconstruction model is the root mean square error function. The optimization function of the time series data reconstruction model is the Adam optimization function. The regularization factor of the time series data reconstruction model is 0.0002.

[0031] The time series data reconstruction model uses the maximum overlap discrete wavelet transform to expand the time series data according to the frequency domain characteristics while maintaining the time series morphology of the data. Through the fusion of channel weight allocation and the bidirectional attention mechanism and the self-attention mechanism, the data features are extracted in a targeted manner to optimize the problems of overfitting of the low strain response segment and underfitting of the high strain response segment.

[0032] After the time series data reconstruction model is constructed, the method for identifying anomalies in massive fatigue damage monitoring data provided by the present invention can be used to identify anomalies in massive fatigue damage monitoring data of the monitoring target, including the following steps: S104. Obtain possible abnormal data in the original bridge strain data of the monitoring target according to the original bridge strain data of the monitoring target obtained in S101 and S102 and the normal data in the original bridge strain data. Specifically, in this embodiment, the data other than the normal data in the original bridge strain data of the monitoring target is used as the possible abnormal data in the original bridge strain data of the monitoring target, wherein the possible abnormal data in the original bridge strain data of the monitoring target includes possible abnormal data of multiple time periods (for example, one hour is used as the length of a time period).

[0033] S105. According to possible abnormal data in the original bridge strain data of the monitoring target, a time series data reconstruction model is constructed through S101-S103 to obtain corresponding reconstructed data.

[0034] S106. According to the possible abnormal data and the corresponding reconstructed normal data in the original bridge strain data of the monitoring target, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data and the corresponding reconstructed data in the original bridge strain data of each time period in multiple time periods are obtained by using the first expression. The first expression includes: , , In the first expression, Indicates The transient energy response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the frequency domain, Indicates The strain amplitude response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the time domain, Represents the normalized Possible abnormal data in the original bridge strain data of each period, Represents the normalized The reconstructed data corresponding to the possible abnormal data in the original bridge strain data of each period, Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Indicates The number of sampling points of possible abnormal data in the original bridge strain data of each period, Represents a frequency domain number.

[0035] S107, according to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, using the second expression, obtain the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods. The second expression includes: , , In the second expression, Indicates The Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain for each time period, Indicates The transient energy response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the frequency domain, Indicates The strain amplitude response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the time domain, express and The covariance of express and The matrix, Z = ], .

[0036] S108. Cluster the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain for all time periods in the multiple time periods according to the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain for each time period in the multiple time periods to obtain abnormal clustering clusters, and use the possible abnormal data in the original bridge strain data of the monitoring target corresponding to the abnormal clustering clusters as the true abnormal data in the possible abnormal data in the original bridge strain data of the monitoring target.

[0037] In this embodiment, fatigue details of an orthotropic rigid bridge deck bridge are monitored, and the frequency is 200Hz. The data volume of a single channel for one day is 17.28 million data points, and the original bridge strain data of an orthotropic rigid bridge deck bridge is obtained. The data segments are subdivided into hours, and each data segment includes 720,000 data points. The normal data of 7 hours in the original bridge strain data are selected to establish the training set and the test set training time series data reconstruction model, and the ratio of the two is 8:2. The possible abnormal data in the original bridge strain data is used as the input value (original value) of the time series data reconstruction model, and the reconstruction value corresponding to the input value is obtained. The input value and the reconstruction value are transformed by continuous wavelet transform using the db6 wavelet function, and the time-frequency characteristics of the input value and the reconstruction value of the possible abnormal data in the original bridge strain data are obtained as shown in Figure 4. The distribution of the transient energy response determination coefficient of the possible abnormal data in the original bridge strain data and its corresponding reconstruction data in the frequency domain and the distribution of the strain amplitude response determination coefficient in the time domain are obtained as shown in Figure 5. The clustering result of clustering the Mahalanobis distance is shown in Figure 6.

[0038] The present invention provides a method for identifying abnormalities in massive fatigue damage monitoring data based on time series data reconstruction and adaptive thresholds. Normal data is obtained from the original bridge strain data of the monitoring target and used to train a time series data reconstruction model based on a deep learning network. The possible abnormal data in the original bridge strain data of the monitoring target is reconstructed with high precision using the time series data reconstruction model. Then, based on the possible abnormal data in the original bridge strain data and the corresponding reconstructed data, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in multiple time periods are obtained. These are used as abnormal evaluation indicators, and then the Mahalanobis distance between the two abnormal evaluation indicators is clustered. The real abnormal data in the possible abnormal data of the monitoring target is dynamically and accurately identified by combining the dual-dimensional indicators of the time domain and the frequency domain. The present invention can realize the abnormality identification of massive fatigue damage monitoring data efficiently and accurately, and has strong generalization ability.

Claims

1. A method for constructing a time series data reconstruction model, characterized in that: include: Obtaining the original bridge strain data of the monitoring target; Preprocessing the original bridge strain data of the monitoring target to obtain normal data in the original bridge strain data; According to the normal data in the original bridge strain data, a time series data reconstruction model of the monitoring target is trained.

2. The method for constructing a time series data reconstruction model according to claim 1, characterized in that: The preprocessing of the original bridge strain data of the monitoring target to obtain normal data in the original bridge strain data includes: Subdividing the original bridge strain data of the monitoring target into multiple data segments according to a preset time period; The number of stress cycles and the maximum amplitude response of each data segment in multiple data segments are counted by using the rain flow counting method; According to the stress cycle number and the maximum amplitude response of each data segment in the multiple data segments, a mean value and a standard deviation of the stress cycle number and a mean value and a standard deviation of the maximum amplitude response of each data segment in the multiple data segments are obtained; According to the mean and standard deviation of the number of stress cycles and the mean and standard deviation of the maximum amplitude response of each data segment in multiple data segments, the original bridge strain data of the data segments whose number of stress cycles is within one times the standard deviation of their mean and whose maximum amplitude response is within one times the standard deviation of their mean are taken as normal data in the original bridge strain data.

3. The method for constructing a time series data reconstruction model according to claim 2, characterized in that: The preprocessing of the original bridge strain data of the monitoring target to obtain normal data in the original bridge strain data includes: Using adaptive modal decomposition to remove the influence of daily temperature changes on normal data in the original bridge strain data; and / or, Wavelet packet analysis is used to reduce the noise of normal data in the original bridge strain data.

4. The method for constructing a time series data reconstruction model according to any one of claims 1 to 3, characterized in that: The deep learning network of the time series data reconstruction model includes a maximum overlap discrete wavelet transform module, a convolution module, a channel weight allocation module, an attention mechanism module, and a fully connected module. The maximum overlap discrete wavelet transform module uses the db6 wavelet function; the convolution module includes three groups of convolution layers and Batchnorm normalization layers connected in series, the convolution kernel sizes are set to 7, 5, and 3 respectively, the convolution step size is 1, and the number of output channels of each convolution layer is 32, 64, and 128 respectively; the channel weight allocation module includes three groups of first global maximum pooling layers, first self-attention layers, and first sigmoid activation layers connected in series for dynamic The weights of 128 channels are adjusted dynamically, the self-attention layer adopts a multi-head attention of size 16, and the output channel is 128; the attention mechanism module is composed of a bi-LSTM layer, a Layernorm normalization layer, a second global maximum pooling layer, a second self-attention layer, a second sigmoid activation layer and a multiplication layer in series, where the bi-LSTM layer has an output channel of 128, and the second self-attention layer adopts a multi-head attention of size 16, and the output channel is 128; the fully connected module includes 2 layers of fully connected layers in series, and the number of output channels is 64, 32, and 1 respectively, which is used to map features to the target output domain.

5. The method for identifying anomalies in massive fatigue damage monitoring data according to claim 4 is characterized in that: The input size of the deep learning network of the time series data reconstruction model is dimension and 2000 data points. The loss function of the time series data reconstruction model is the root mean square error function. The optimization function of the time series data reconstruction model is the Adam optimization function. The regularization factor of the time series data reconstruction model is 0.0002.

6. A method for identifying anomalies in massive fatigue damage monitoring data, characterized in that: include: The original bridge strain data of the monitoring target and the normal data in the original bridge strain data obtained according to the construction method of the time series data reconstruction model according to any one of claims 1 to 5 are used to obtain possible abnormal data in the original bridge strain data of the monitoring target, wherein the possible abnormal data in the original bridge strain data of the monitoring target includes possible abnormal data in multiple time periods; According to possible abnormal data in the original bridge strain data of the monitoring target, a time series data reconstruction model is obtained by the method for constructing a time series data reconstruction model according to any one of claims 1 to 5 to obtain reconstructed data; According to the possible abnormal data and the reconstructed data in the original bridge strain data of the monitoring target, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period and the corresponding reconstructed data are obtained; According to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods is obtained; According to the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of each time period in multiple time periods, the Mahalanobis distances between the determination coefficients of the transient energy response in the frequency domain and the determination coefficients of the strain amplitude response in the time domain of all time periods in multiple time periods are clustered to obtain the real abnormal data among the possible abnormal data in the original bridge strain data of the monitoring target.

7. The method for identifying anomalies in massive fatigue damage monitoring data according to claim 6 is characterized in that: The original bridge strain data of the monitoring target and the normal data in the original bridge strain data obtained by the method for constructing the time series data reconstruction model according to any one of claims 1 to 5 are used to obtain possible abnormal data in the original bridge strain data of the monitoring target, wherein the possible abnormal data in the original bridge strain data of the monitoring target includes possible abnormal data of multiple time periods, including: According to the original bridge strain data of the monitoring target and the normal data in the original bridge strain data obtained by the construction method of the time series data reconstruction model according to any one of claims 1 to 5, the data other than the normal data in the original bridge strain data of the monitoring target are used as possible abnormal data in the original bridge strain data of the monitoring target.

8. The method for identifying anomalies in massive fatigue damage monitoring data according to claim 7, characterized in that: The method of clustering the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain of all the time periods in the multiple time periods according to the Mahalanobis distances between the transient energy response determination coefficients in the frequency domain and the strain amplitude response determination coefficients in the time domain of each time period in the multiple time periods, and obtaining the real abnormal data in the possible abnormal data in the original bridge strain data of the monitoring target, includes: According to the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of each time period in the multiple time periods, the Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain of all time periods in the multiple time periods are clustered to obtain abnormal clustering clusters; The possible abnormal data in the original bridge strain data of the monitoring target corresponding to the abnormal cluster is taken as the real abnormal data in the possible abnormal data in the original bridge strain data of the monitoring target.

9. The method for identifying anomalies in massive fatigue damage monitoring data according to claim 8, characterized in that: The method of obtaining the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in multiple time periods and the corresponding reconstructed data thereof according to the possible abnormal data and the reconstructed data in the original bridge strain data of the monitoring target includes: According to the possible abnormal data and reconstructed data in the original bridge strain data of the monitoring target, using the first expression, the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period and its corresponding reconstructed data are obtained; The first expression includes: ， ， In the first expression, Indicates The transient energy response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the frequency domain, Indicates The strain amplitude response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the time domain, Represents the normalized Possible abnormal data in the original bridge strain data of each period, Represents the normalized The reconstructed data corresponding to the possible abnormal data in the original bridge strain data of each period, Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Express After continuous wavelet analysis, the transient energy response is The total instantaneous energy response obtained by integrating in the upper frequency domain is Indicates The number of sampling points of possible abnormal data in the original bridge strain data of each period, Represents a frequency domain number.

10. The method for identifying anomalies in massive fatigue damage monitoring data according to claim 9, characterized in that: The method of obtaining the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods according to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data thereof, includes: According to the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of the possible abnormal data in the original bridge strain data of each time period in the multiple time periods and the corresponding reconstructed data, the Mahalanobis distance between the transient energy response determination coefficient in the frequency domain and the strain amplitude response determination coefficient in the time domain of each time period in the multiple time periods is obtained by using the second expression; The second expression includes: ， ， In the second expression, Indicates The Mahalanobis distance between the determination coefficient of the transient energy response in the frequency domain and the determination coefficient of the strain amplitude response in the time domain for each time period, Indicates The transient energy response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the frequency domain, Indicates The strain amplitude response determination coefficient of the possible abnormal data in the original bridge strain data of each period and its corresponding reconstructed data in the time domain, express and The covariance of express and The matrix, Z = ], express.

Images