Bridge deflection monitoring response anomaly detection method based on deep convolution auto-encoder

Abstract

The invention discloses a bridge deflection monitoring response anomaly detection method based on a deep convolution auto-encoder, and the method comprises the steps: collecting a bridge deflection monitoring response D (t) within a period of time t, separating components D1 (t), D2 (t) and D3 (t) through an ideal filter according to a frequency cut-off lower limit and a frequency cut-off upper limit, selecting a window through employing a moving average method, separating D2 (t) into D21 (t) and D22 (t), and carrying out the detection of the bridge deflection monitoring response D (t). The method comprises the following steps: constructing a sample data set of bridge deflection monitoring response in a normal state, obtaining deflection response components in the normal state according to the method, designing a deep convolutional network (DCNN), taking time-history response pictures of the deflection response components as input, training four information feature extraction models, constructing a classification model through reconstruction errors, and constructing a classification model. Actually-measured deflection response components are obtained from actually-measured deflection response Dm (t) according to the method and are respectively input into an information feature extraction model, and the actually-measured deflection response components are respectively divided into a normal state and an abnormal state through reconstruction errors.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge deformation measurement, and more specifically, to a method for detecting abnormal response of bridge deflection monitoring based on a deep convolutional autoencoder. Background Art

[0002] As of now, there are more than 1 million highway bridges in China, and the number of bridges built is still growing rapidly every year. The number of trucks is 11.6666 million, the tonnage of trucks is 169.6733 million tons, and the average truck load is 145,435 tons. It can be seen that the highway transportation infrastructure continues to face the burden of heavy-duty transportation, which also makes many bridges that have been in service for many years face operational safety risks. With the increase in traffic volume and the increase in the service life of bridges, bridge structures may be damaged and degraded due to environmental erosion, traffic load, material aging, etc. If these damages cannot be detected and repaired in time, they are likely to affect driving safety, shorten the service life of bridges, and even cause sudden damage and collapse of bridges.

[0003] At present, most bridge inspection methods use radar interferometry technology, which has a measurement accuracy of millimeters. Radar interferometry technology is used to measure bridge deflection with high precision, so as to reflect the deformation of the bridge based on the measured bridge deflection. However, when a vehicle passes by the bridge or is affected by factors such as wind, the deflection of the bridge will change dynamically, and radar interferometry technology will be affected by the dynamic changes, resulting in inaccurate bridge deformation measurement results reflected by the measured bridge deflection.

[0004] Therefore, there is an urgent need for a bridge health monitoring system to check the condition of the bridge at various times and under various circumstances, improve inspection efficiency and accuracy, provide real-time data on the health of the bridge structure, and help engineers make better maintenance decisions. The monitoring system usually includes a variety of sensors, such as displacement sensors, strain sensors, accelerometers, temperature sensors, etc., which are used to monitor key parameters of the bridge, such as the vertical acceleration and deflection of the bridge. Summary of the invention

[0005] The present invention aims to overcome at least one defect in the prior art and provide a bridge deflection monitoring response anomaly detection method based on a deep convolutional autoencoder to solve the problem of inaccurate bridge deflection measurement when affected by factors such as vehicles or wind.

[0006] The technical method adopted by the present invention is a bridge deflection monitoring response abnormality detection method based on a deep convolutional autoencoder, comprising the following steps: Step 100: Collect the bridge deflection monitoring response over a period of time t D ( t ); Step 200: Design a series of ideal filters and set the frequency cutoff lower limit according to the characteristics of the bridge w l and frequency cutoff upper limit w u , the ideal filter is based on w l and w u , separate the components D 1( t ), D 2( t )and D 3( t ); Step 300: Use the moving average method to select a window D 2( t ) is separated into D 21 ( t )and D 22 ( t ); Step 400: Constructing the monitoring response of the bridge deflection under normal conditions D ( t ) sample data set, the deflection response components under normal conditions are obtained according to the above method. D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ); Step 500: Design a deep convolutional network DCNN to D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ) as input, and train the information feature extraction models DCNN1 and DCNN 21 , DCNN 22 and DCNN3, building a classification model by reconstructing the error; Step 600: Convert the measured deflection response D m (t) The four components obtained by the above separation method D m 1( t ), Dm 21 ( t ), D m 22 ( t )and D m 3( t ), input the trained DCNN1 and DCNN respectively 21 , DCNN 22 In CNN and DCNN3, the reconstruction error is used to divide them into normal state and abnormal state respectively.

[0007] The high-precision sensor collects the bridge deflection monitoring response data within a period of time t D ( t ), ensuring the accuracy and integrity of the data; designing an ideal filter based on the structural characteristics and dynamic behavior of the bridge and setting the upper frequency cutoff limit w u and the lower frequency cutoff w l , the low-frequency response of the bridge deflection response is separated by a filter D 1( t ), intermediate frequency response D 2( t ) and high frequency response D 3( t ) to achieve refined signal processing.

[0008] In order to deeply understand the frequency characteristics of the bridge deflection response, the step 100 is specifically to collect the bridge deflection monitoring response within a time range t. D ( t ), obtain spectrum information through fast Fourier transform F ( w ). FFT technology can decompose complex time domain signals into a series of simple sine waves or cosine waves, each of which corresponds to a specific frequency component. By converting the response into frequency spectrum information, the distribution law of the bridge deflection response at different frequencies can be reflected. By observing and analyzing the frequency spectrum information F(w), the dominant frequency, secondary frequency and potential abnormal frequency components in the bridge deflection response can be identified, which provides strong support for subsequent signal separation, feature extraction and anomaly detection.

[0009] In order to separate the signal components, the ideal filter in step 200 is based on w l and w u , separate the components D 1( t ), D 2(t )and D 3( t ) Specifically, the frequency w < w l The components are separated into F 1( w ), the frequency w l ≤ w ≤ w u The components are separated into F 2( w ), the frequency w > w u The components are separated into F 3( w ), and the components are transformed by inverse Fourier transform F 1( w ), F 2( w )and F 3( w ) are converted into time domain signals, and we get D 1( t ), D 2( t )and D 3( t ). For frequency w < w l The low-pass ideal filter is designed to separate the low-frequency components from the original spectrum information to form a new spectrum component. F 1( w ); for frequency w l ≤ w ≤ w u The components of the spectrum are obtained by designing a bandpass ideal filter to extract the signal within this frequency range and form a spectrum component. F 2( w ) ; for frequency w > w u The high-pass ideal filter is designed to separate the high-frequency noise component from the original spectrum information to form the spectrum component F 3( w ) After completing the fine separation of the spectral components, the spectral components are transformed by inverse Fourier transform. F 1( w ), F 2( w )and F 3(w ) are converted back to time domain signals, which not only retains the phase information of the original signal, but also ensures that the time domain signal D 1( t ), D 2( t )and D 3( t )’s accuracy and completeness.

[0010] Said D 1( t )+ D 21 ( t ) represents the low-frequency pseudo-static component caused by long-term deflection, annual temperature difference, monthly temperature difference, daily temperature difference, etc., represents the medium-frequency pseudo-static component caused by average wind field and congested traffic, and D 22 ( t ) represents the high-frequency pseudo-static component caused by sparse traffic, D 3( t ) represents the high-frequency dynamic components caused by traffic dynamics, environmental vibration and noise. The determination of high-frequency pseudo-static components is due to the irregularity and instantaneity of traffic flow, which shows more frequent and rapid fluctuation characteristics, thus showing the characteristics of high-frequency pseudo-static in dynamic analysis; high-frequency dynamic components are not only high in frequency and fast in change, but also often accompanied by significant energy transfer and dynamic response, which has an important impact on the overall stability and safety of the structure.

[0011] In order to efficiently process complex data, the deep convolutional network DCNN in step 500 includes an encoder module, a space-channel self-attention module, and a decoding module. The encoder module extracts low-dimensional feature representations of input data through multi-layer convolution; the space-channel self-attention module uses the attention mechanism to dynamically adjust feature weights in two dimensions, space and channel, so as to capture and strengthen key information and improve the model's ability to recognize complex feature patterns; the decoding module maps the features processed by the encoding and self-attention mechanisms back to the original data space, and gradually restores the spatial resolution and detail information of the data through upsampling and inverse convolution operations, and finally outputs high-quality prediction or classification results.

[0012] In order to achieve feature extraction and dimensionality reduction, the encoder module includes 4 convolution submodules, including convolution layer, downsampling layer, and activation function. The convolution layer uses multi-core convolution operations to deeply explore the local adjustment of input data, generate feature maps, and provide key information for subsequent processing; the downsampling layer uses pooling operations (such as maximum pooling or average pooling) to reduce the size of feature maps and data dimensions, while retaining important feature information and enhancing the robustness and generalization ability of the model; the activation function introduces nonlinear activation functions and nonlinear characteristics to the model, so that the network can learn and represent complex nonlinear relationships and improve the expressiveness of the model.

[0013] In order to gradually restore and refine the feature map information, the decoding module includes three deconvolution submodules, including a deconvolution layer, an upsampling layer, and an activation function. The deconvolution layer maps the feature map from low resolution back to high resolution step by step through the deconvolution operation (transposed convolution), while extracting and enhancing the feature information to provide data for subsequent image reconstruction or feature recovery; the upsampling layer uses an upsampling algorithm (such as bilinear interpolation, nearest neighbor interpolation, or upsampling network) to increase the size of the feature map, restore image details, and maintain the consistency and coherence of the features; the activation function selects a nonlinear activation function (such as ReLU, Leaky ReLU, PReLU, or ELU, etc.) to introduce nonlinear characteristics into the decoding process, enhance the expression ability of the model, and make the decoded image or feature more realistic and natural.

[0014] The determination condition of the normal state in step 600 is that when the four classification modules all have normal results, the deflection response is determined to be in a normal state.

[0015] In order to distinguish different abnormal types, the abnormal state determination condition of step 600 is: (1) When DCNN1, DCNN 21 , DCNN 22 The abnormality of DCNN1 and DCNN3 is considered as abnormal state 1-4, that is, abnormal state 1 is that DCNN1 is abnormal, and abnormal state 2 is that DCNN 21 An exception occurs, and abnormal state 3 is DCNN 22 An abnormality occurs. Abnormal state 4 means that DCNN3 has an abnormality. (2) When DCNN1, DCNN 21 , DCNN 22 The combination of DCNN1 and DCNN3 is considered abnormal state 5-10, that is, abnormal state 5 is DCNN1 and DCNN 21 Both are abnormal, abnormal state 6 is DCNN1 and DCNN 22 Both DCNN1 and DCNN3 are abnormal. Abnormal state 7 is abnormal. Abnormal state 8 is DCNN21 and DCNN 22 All of them are abnormal, abnormal state 9 is DCNN 21 Both DCNN and DCNN3 are abnormal, and abnormal state 10 is DCNN 22 and DCNN3 all showed abnormalities; (3) When DCNN1, DCNN 21 , DCNN 22 The abnormal combination of DCNN1, DCNN3 is considered as abnormal state 11-14, that is, abnormal state 11 is DCNN1, DCNN 21 and DCNN 22 All of them are abnormal, abnormal state 12 is DCNN1, DCNN 21 and DCNN3 are abnormal, and abnormal state 13 is DCNN1, DCNN 22 Both DCNN and DCNN3 are abnormal, and abnormal state 14 is DCNN 21 , DCNN 22 and DCNN3 all showed abnormalities; (4) When DCNN1, DCNN 21 , DCNN 22 and DCNN3 are all abnormal, which is considered abnormal state 15. When abnormal state 15 occurs, it means that all modules are abnormal and emergency measures need to be taken for intervention.

[0016] The abnormal types corresponding to the 15 abnormal states are: abnormal state 1 corresponds to abnormal temperature difference, abnormal state 2 corresponds to abnormal interference and abnormal noise, abnormal state 3 corresponds to data loss and low resolution, abnormal state 4 corresponds to abnormal heavy vehicle, abnormal state 5 corresponds to abnormal temperature difference, abnormal interference and abnormal noise, abnormal state 6 corresponds to abnormal temperature difference, data loss and low resolution, abnormal state 7 corresponds to abnormal temperature difference and abnormal heavy vehicle, abnormal state 8 corresponds to abnormal interference, abnormal noise, data loss and low resolution, abnormal state 9 corresponds to abnormal interference, abnormal noise and abnormal heavy vehicle, abnormal state 10 corresponds to data loss, low resolution and abnormal heavy vehicle, abnormal state 11 corresponds to abnormal temperature difference, abnormal noise, data loss and low resolution, abnormal state 12 corresponds to abnormal temperature difference, abnormal interference, abnormal noise and abnormal heavy vehicle, abnormal state 13 corresponds to abnormal temperature difference, data loss, low resolution and abnormal heavy vehicle, abnormal state 14 corresponds to abnormal interference, abnormal noise, data loss, low resolution and abnormal heavy vehicle, abnormal state 15 corresponds to abnormal temperature difference, abnormal interference, abnormal noise, data loss, abnormal heavy vehicle and sensor failure.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Real-time monitoring to determine the type of abnormality Through the above method, real-time monitoring of abnormal bridge deflection data can be achieved, and the data loss, abnormal interference, abnormal heavy vehicles, abnormal temperature difference, abnormal noise, sensor failure, low resolution and other situations in the actual situation can be judged according to the abnormalities of different frequency components in the data. It can provide key method decisions for data analysis and abnormal state evaluation of bridge deflection response health monitoring, which is of great significance for ensuring the safety of bridge operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings of the present invention are only used for illustrative purposes and cannot be understood as limiting the present invention. In order to more clearly illustrate the technical methods of the embodiments of the present invention, the drawings required for describing the embodiments are 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 creative work.

[0019] Figure 1 This is a flow chart of a method for detecting abnormal response of bridge deflection monitoring based on a deep convolutional autoencoder according to the present invention.

[0020] Figure 2 This is a technical roadmap for a bridge deflection monitoring response anomaly detection method based on a deep convolutional autoencoder of the present invention.

[0021] Figure 3 This is an architecture diagram of a deep convolutional autoencoder of the present invention.

[0022] Figure 4 It is the deflection diagram and spectrum diagram of the bridge deflection monitoring response D(t) under normal state of the present invention.

[0023] Figure 5 It is the deflection diagram and spectrum diagram of the bridge deflection monitoring response Dm(t) under abnormal state of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical methods 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 are within the scope of protection of the present invention.

[0025] Example 1 like Figure 1 and Figure 2 As shown, this embodiment provides a bridge deflection monitoring response abnormality detection method based on a deep convolutional autoencoder, including: Step 100: Collect the bridge deflection monitoring response over a period of time t D ( t ); Step 200: Design a series of ideal filters and set the frequency cutoff lower limit according to the characteristics of the bridge w l and frequency cutoff upper limit w u , the frequency w < w l The components are separated into D 1( t ), the frequency w l ≤ w ≤ w u The components are separated into D 2( t ), the frequency w > w u The components of are separated into components using an ideal filter D 3( t ); Step 300: Further adopt the moving average method and select a suitable window size to D 2( t ) is separated into D 21 ( t )and D 22 ( t ), representing D 2( t )’s quasi-static and dynamic components; Step 400: Constructing the monitoring response of the bridge deflection under normal conditions D ( t ) and obtain the deflection response components under normal conditions according to the above method. D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ); Step 500: Design a deep convolutional autoencoder architecture, including an encoder module, a spatial-channel self-attention module, and a decoder module. D 1( t ), D 21 (t ), D 22 ( t )or D 3( t ) as input, and train the models DCNN1 and DCNN that can extract the information features of the time-course response images. 21 , DCNN 22 and DCNN3, building a classification model by reconstructing the error; Step 600: Convert the measured deflection response D m ( t ) The four components obtained according to the above separation method D m 1( t ), D m 21 ( t ), D m 22 ( t )and D m 3( t ), input the trained DCNN1 and DCNN respectively 21 , DCNN 22 In CNN and DCNN3, the reconstruction error is used to divide them into normal and abnormal states.

[0026] In this embodiment, step 100 specifically includes collecting the bridge deflection monitoring response within a period of time t. D ( t ), obtain spectrum information through fast Fourier transform F ( w ).

[0027] Step 200 includes step 210 of separating spectral components and step 220 of converting the signal into a time domain. Step 210: Separate the spectrum components, specifically, various filters are used to separate the spectrum components obtained in step 100. F ( w ) is filtered, and the low-pass filter filters the frequency w < w l The components are separated into spectral components F 1( w ), the bandpass filter will w l ≤ w ≤ w u The components are separated into spectral components F 2(w ), the high-pass filter will w > w u The components are separated into spectral components F 3( w ).

[0028] Step 220: Time domain signal conversion, specifically, using inverse Fourier transform to transform the three components F 1( w ), F 2( w )and F 3( w ) are converted into time domain signals, and we get D 1( t ), D 2( t )and D 3( t ).

[0029] Step 400: construct the bridge deflection monitoring response under normal conditions D ( t ) and obtain the deflection response components under normal conditions according to the above method. D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ), specifically, D 1( t )+ D 21 ( t ) represents the low-frequency pseudo-static component caused by long-term deflection, annual temperature difference, monthly temperature difference, daily temperature difference, etc., represents the medium-frequency pseudo-static component caused by average wind field and congested traffic, D 22 ( t ) represents the high-frequency pseudo-static component caused by sparse traffic, D 3( t ) represents the high-frequency dynamic components caused by traffic dynamics, environmental vibration and noise.

[0030] In this embodiment, if Figure 3As shown, step 500: design a deep convolutional autoencoder architecture, including an encoder module, a spatial-channel self-attention module, and a decoder module. The encoder module includes 4 convolution submodules, each of which includes a convolution layer, a downsampling layer, and an activation function. The convolution submodule is used to extract image features of the picture. The convolution layer uses multi-core convolution operations to deeply explore the local adjustment of the input data and generate feature maps to provide key information for subsequent processing. The downsampling layer uses pooling operations (such as maximum pooling or average pooling) to reduce the size of the feature map and reduce the data dimension, while retaining important feature information and enhancing the robustness and generalization ability of the model. The activation function introduces a nonlinear activation function to introduce nonlinear characteristics to the model, so that the network can learn and represent complex nonlinear relationships and improve the expression ability of the model. The spatial-channel self-attention module uses the attention mechanism to dynamically adjust the feature weights in the two dimensions of space and channel, thereby capturing and strengthening key information, so that the network can pay more attention to the pictures and picture areas of a certain channel, enhance its ability to recognize picture mutations, and improve the model's ability to recognize complex feature patterns. The decoder module includes three deconvolution modules, each of which contains a deconvolution layer, an upsampling layer and an activation function for reconstructing images. The deconvolution layer uses a deconvolution operation (transposed convolution) to gradually map the feature map from low resolution back to high resolution, while extracting and enhancing feature information to provide data for subsequent image reconstruction or feature recovery. The upsampling layer uses an upsampling algorithm (such as bilinear interpolation, nearest neighbor interpolation or upsampling network) to increase the size of the feature map and restore image details while maintaining feature consistency and coherence. The activation function selects a nonlinear activation function (such as ReLU, Leaky ReLU, PReLU or ELU, etc.) to introduce nonlinear characteristics to the decoding process, enhance the expression ability of the model, and make the decoded image or feature more realistic and natural. D 1( t ), D 21 ( t ), D 22 ( t )or D 3( t ) as input, and train the models DCNN1 and DCNN that can extract the information features of the time-course response images. 21 , DCNN 22 and DCNN3, building a classification model through reconstruction error.

[0031] In this embodiment, combined with Figure 4 and Figure 5 As shown, step 600: convert the measured deflection response D m ( t) The four components obtained according to the above separation method D m 1(t), D m 21 ( t ), D m 22 ( t )and D m 3( t ), input the trained DCNN1 and DCNN respectively 21 , DCNN 22 In DCNN3 and DCNN4, the reconstruction error is used to classify the deflection response into two states: normal and abnormal. The judgment condition for the normal state is that when all four classification modules have normal results, the deflection response is judged to be in a normal state; the judgment condition for the abnormal state is that (1) When DCNN1, DCNN 21 , DCNN 22 and DCNN3 are considered abnormal states 1-4 if they are abnormal alone; (2) When DCNN1, DCNN 21 , DCNN 22 The combination of DCNN3 and DCNN3 is considered abnormal 5-10; (3) When DCNN1, DCNN 21 , DCNN 22 The abnormal combination of DCNN3 and DCNN3 is considered as abnormal state 11-14; (4) When DCNN1, DCNN 21 , DCNN 22 and DCNN3 are considered abnormal status 15.

[0032] When abnormality 15 occurs, it means that all modules are abnormal and emergency measures need to be taken for intervention. These 15 abnormal states correspond to different abnormal types, and the specific corresponding situations are shown in the following table.

[0033]

[0034] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A bridge deflection monitoring response anomaly detection method based on deep convolutional autoencoder, characterized in that: The following steps are involved: Step 100: Collect the bridge deflection monitoring response over a period of time t D ( t ); Step 200: Design a series of ideal filters and set the frequency cutoff lower limit according to the characteristics of the bridge w l and frequency cutoff upper limit w u , the ideal filter is based on w l and w u , separate the components D 1( t ), D 2( t )and D 3( t ); Step 300: Use the moving average method to select a window D 2( t ) is separated into D 21 ( t )and D 22 ( t ) ; Step 400: Constructing the monitoring response of the bridge deflection under normal conditions D ( t ) sample data set, the deflection response components under normal conditions are obtained according to the above method. D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ); Step 500: Design a deep convolutional network DCNN to D 1( t ), D 21 ( t ), D 22 ( t )and D 3( t ) as input, and train the information feature extraction models DCNN1 and DCNN 21 , DCNN 22 and DCNN3, building a classification model by reconstructing the error; Step 600: Convert the measured deflection response D m (t) The four components obtained by the above separation method D m 1( t ), D m 21 ( t ), D m 22 ( t )and D m 3( t ), input the trained DCNN1 and DCNN respectively 21 , DCNN 22 In CNN and DCNN3, the reconstruction error is used to divide them into normal state and abnormal state respectively.

2. According to claim 1, a bridge deflection monitoring response anomaly detection method based on deep convolutional autoencoder is characterized in that: The step 100 specifically includes collecting the bridge deflection monitoring response within a time range t. D ( t ), obtain spectrum information through fast Fourier transform F ( w ).

3. According to claim 2, a bridge deflection monitoring response anomaly detection method based on deep convolutional autoencoder is characterized in that: In step 200, the ideal filter is based on w l and w u , separate the components D 1( t ), D 2( t )and D 3( t ) Specifically, the frequency w < w l The components are separated into F 1( w ), the frequency w l ≤ w ≤ w u The components are separated into F 2( w ), the frequency w > w u The components are separated into F 3( w ), and the components are transformed by inverse Fourier transform F 1( w ), F 2( w )and F 3( w ) are converted into time domain signals, and we get D 1( t ), D 2( t )and D 3( t ).

4. The method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder according to claim 3 is characterized in that: Said D 1( t )+ D 21 ( t ) represents the low-frequency pseudo-static component caused by long-term deflection, annual temperature difference, monthly temperature difference, daily temperature difference, etc., represents the medium-frequency pseudo-static component caused by average wind field and congested traffic, and D 22 ( t ) represents the high-frequency pseudo-static component caused by sparse traffic, D 3( t ) represents the high-frequency dynamic components caused by traffic dynamics, environmental vibration and noise.

5. According to claim 4, a method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder is characterized in that: The deep convolutional network DCNN in step 500 includes an encoder module, a space-channel self-attention module, and a decoding module.

6. According to claim 5, a method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder is characterized in that: The encoder module includes four convolution submodules, and the convolution submodule includes a convolution layer, a downsampling layer, and an activation function.

7. According to claim 6, a method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder is characterized in that: The decoding module includes three deconvolution submodules, and the deconvolution submodule includes a deconvolution layer, an upsampling layer, and an activation function.

8. The method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder according to claim 7 is characterized in that: The determination condition of the normal state in step 600 is that when the four classification modules all have normal results, the deflection response is determined to be in a normal state.

9. The method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder according to claim 8 is characterized in that: The abnormal state determination condition of step 600 is: (1) When DCNN1, DCNN 21 , DCNN 22 The abnormality of DCNN1 and DCNN3 is considered as abnormal state 1-4, that is, abnormal state 1 is that DCNN1 is abnormal, and abnormal state 2 is that DCNN 21 An exception occurs, and abnormal state 3 is DCNN 22 An abnormality occurs. Abnormal state 4 means that DCNN3 has an abnormality. (2) When DCNN1, DCNN 21 , DCNN 22 The combination of DCNN1 and DCNN3 is considered abnormal state 5-10, that is, abnormal state 5 is DCNN1 and DCNN 21 Both are abnormal, abnormal state 6 is DCNN1 and DCNN 22 Both DCNN1 and DCNN3 are abnormal. Abnormal state 7 is abnormal. Abnormal state 8 is DCNN 21 and DCNN 22 All of them are abnormal, abnormal state 9 is DCNN 21 Both DCNN and DCNN3 are abnormal, and abnormal state 10 is DCNN 22 and DCNN3 all showed abnormalities; (3) When DCNN1, DCNN 21 , DCNN 22 The abnormal combination of DCNN1, DCNN3 is considered as abnormal state 11-14, that is, abnormal state 11 is DCNN1, DCNN 21 and DCNN 22 All of them are abnormal, abnormal state 12 is DCNN1, DCNN 21 and DCNN3 are abnormal, and abnormal state 13 is DCNN1, DCNN 22 Both DCNN and DCNN3 are abnormal, and abnormal state 14 is DCNN 21 , DCNN 22 and DCNN3 all showed abnormalities; (4) When DCNN1, DCNN 21 , DCNN 22 and DCNN3 are considered abnormal if all four are abnormal.

10. The method for detecting abnormal response of bridge deflection monitoring based on deep convolutional autoencoder according to claim 9 is characterized in that: The abnormal types corresponding to the 15 abnormal states are: abnormal state 1 corresponds to abnormal temperature difference, abnormal state 2 corresponds to abnormal interference and abnormal noise, abnormal state 3 corresponds to data loss and low resolution, abnormal state 4 corresponds to abnormal heavy vehicle, abnormal state 5 corresponds to abnormal temperature difference, abnormal interference and abnormal noise, abnormal state 6 corresponds to abnormal temperature difference, data loss and low resolution, abnormal state 7 corresponds to abnormal temperature difference and abnormal heavy vehicle, abnormal state 8 corresponds to abnormal interference, abnormal noise, data loss and low resolution, abnormal state 9 corresponds to abnormal interference, abnormal noise and abnormal heavy vehicle, abnormal state 10 corresponds to data loss, low resolution and abnormal heavy vehicle, abnormal state 11 corresponds to abnormal temperature difference, abnormal noise, data loss and low resolution, abnormal state 12 corresponds to abnormal temperature difference, abnormal interference, abnormal noise and abnormal heavy vehicle, abnormal state 13 corresponds to abnormal temperature difference, data loss, low resolution and abnormal heavy vehicle, abnormal state 14 corresponds to abnormal interference, abnormal noise, data loss, low resolution and abnormal heavy vehicle, abnormal state 15 corresponds to abnormal temperature difference, abnormal interference, abnormal noise, data loss, abnormal heavy vehicle and sensor failure.

Images