Anomaly detection method for bridge structure health monitoring data based on deep learning

Abstract

The present invention discloses a method for detecting anomalies in bridge structure health monitoring data based on deep learning. The method abandons the traditional method of over-reliance on data analysis in the time domain. Instead, the method visualizes the data in the time domain, frequency domain and Gram angle field (GAF) and then superimposes the data into a three-channel composite image through an image processing algorithm. The composite image of the three channels is input into a VGG16 neural network model that incorporates a channel attention mechanism for processing. The VGG16 neural network model that incorporates a channel attention mechanism can dynamically adjust the weight of each channel according to image features and perform dimensionality reduction processing, thereby improving the accuracy of abnormal data recognition and reducing the complexity of the model. In addition, the neural network model that incorporates a channel attention mechanism in the present application is not affected by trend, seasonality, periodicity and random disturbances in traditional time series, and has a strong feature extraction capability and good robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge monitoring, and in particular relates to a method for detecting anomaly in bridge structure health monitoring data based on deep learning. Background Art

[0002] The construction of bridges in my country has developed rapidly and the number is huge. According to the "Statistical Bulletin on the Development of the Transportation Industry in 2022" issued by the Ministry of Transport, by the end of 2022, there were 1.0332 million highway bridges in my country with a total length of 85.7649 million meters, including 8,816 super-large bridges and 159,600 large bridges. In recent years, the development trend of bridge structures is that the span is getting longer and longer and the structure is getting more and more complex. There are more and more potential risk points in the use of bridge structures. Because the service life of bridges is usually decades or even hundreds of years, coupled with poor working environment, improper maintenance, serious vehicle overloading, degradation of building materials and natural disasters, bridge structures are prone to damage, aging, local damage and other problems, and even serious accidents, causing great losses of social wealth and threatening people's lives. If bridges can be maintained and inspected regularly throughout the year, the operating conditions of bridges can be improved and the probability of bridge accidents can be reduced. The structural health monitoring system provides a new solution for bridge maintenance, which mainly obtains structural responses and extracts damage-sensitive indicators by using periodic sensor arrays.

[0003] A complete bridge health monitoring system usually includes hundreds of sensors to monitor various indicators. Therefore, a large bridge health monitoring system collects a huge amount of data. Since bridge structures are mostly in the wild, sensors are inevitably affected by natural factors such as high temperature, high humidity, and complex electromagnetic environment. From data collection to data transmission, and then from data transmission to data storage, each process and each link may cause data abnormalities. These abnormal data will cause false alarms in the system and even affect the evaluation results of the bridge structure health. Therefore, it is very necessary to identify abnormal data.

[0004] In order to solve the above problems, our unit has designed a bridge structure health monitoring data anomaly recognition method based on image classification. In this method, the influence of data imbalance on the accuracy of abnormal data detection is solved by integrating the focal loss function. The Mobile ViT network realizes the complementary advantages of the convolutional layer in the convolutional neural network and the self-attention mechanism in the Transformer. It uses lightweight convolution to reduce the amount of training, and uses the sub-attention mechanism for global feature extraction, improving the characteristics of ordinary convolution operations to intelligently extract local features, and the shortcomings of Transformer, such as large number of parameters and difficulty in training; that is, although this method improves the accuracy of abnormal data detection, it still uses the existing traditional methods to limit the processed data to the time domain perspective. The time series shows trend, seasonality, periodicity and random disturbance in the time series, that is, the time domain features are not enough to distinguish the categories of abnormal data well. Modeling it from the perspective of the time domain has limitations and lacks wide applicability.

[0005] In this regard, based on this, this application proposes a new deep learning-based anomaly detection method for bridge structure health monitoring data. Summary of the invention

[0006] The purpose of the present invention is to provide a bridge structure health monitoring data anomaly detection method based on deep learning, which divides the original time series data into several parts, and visualizes them in the time domain, frequency domain and Gram angle field (GAF) respectively, and then superimposes the time domain image, frequency domain image and GAF ​​image into a three-channel image, and uses the VGG16 neural network model integrated with the channel attention mechanism to dynamically adjust the weight coefficient of each channel according to the feature description, thereby improving the accuracy of abnormal data recognition and detection.

[0007] In order to solve the above problems, the present application provides a method for detecting anomalies in bridge structure health monitoring data based on deep learning, which comprises the following steps:

[0008] S1: Generate image dataset

[0009] Obtain the bridge structure health monitoring data and divide the monitoring data into several parts. At the same time, perform image visualization processing on the monitoring data of each part in the time domain, frequency domain and Gram angle field (GAF), and then superimpose the three images corresponding to each part to form a three-channel composite image.

[0010] S2: Selected training set

[0011] Acquire the health monitoring data of the bridge structure with known health status and perform the above step S1 on it, then label the formed three-channel composite image according to the data features, and use the monitoring data as a training set;

[0012] S3: Constructing a VGG16 network with channel attention mechanism as an anomaly detection model

[0013] The anomaly detection model includes a convolution module, a channel attention mechanism module, a fully connected layer and a Softmax classifier, the convolution module includes multiple groups, the channel attention mechanism module is connected in series between two adjacent convolution modules; the convolution module consists of at least two convolution layers and one maximum pooling layer;

[0014] S4: Training anomaly detection model

[0015] The synthetic images in the training set are input into the VGG16 network with the channel attention mechanism for deep recognition and learning to achieve model training;

[0016] S5: Output the test results

[0017] Obtain the health monitoring data of the bridge to be monitored, and input the monitoring data into the trained anomaly detection model after executing the above S1, and output the detection results through the Softmax classifier.

[0018] As a preferred solution of the present application, in step S3, the channel attention mechanism module includes a maximum pooling layer and an average pooling layer arranged in parallel, and a shared layer and an activation layer connected in series therewith; by performing average pooling and maximum pooling operations on the input features, different feature descriptions can be obtained respectively, and the shared layer can compress the different feature descriptions. Finally, the different feature descriptions can be combined element by element and a channel attention mechanism graph can be generated through the activation layer.

[0019] As a preferred solution of the present application, the activation layer adopts a Sigmoid activation function.

[0020] As a preferred solution of the present application, the shared layer includes a global pooling layer, a fully connected layer, a RELU activation layer, a fully connected layer and a Sigmoid activation layer connected in series in sequence. The shared layer of this structure can reduce the dimension of features and extract important information.

[0021] As a preferred solution of the present application, in step 3, the filter of the convolutional layer is 3×3, and the step size is 1; the filter of the maximum pooling layer is 2×2, and the step size is 2; the fully connected layer includes three layers connected in sequence, and the number of channels of each fully connected layer is 4096, 4096 and 1000 respectively.

[0022] As a preferred solution of the present application, in step S2, the data features include health status and distribution trend, and the data health status includes abnormal data and normal data in various situations.

[0023] As a preferred solution of the present application, in steps S1 and S2, the acquired bridge structure health monitoring data should satisfy the time sequence;

[0024] When obtaining a time domain image, the one-dimensional data in each separated segment is plotted on the time axis to obtain a frequency domain image;

[0025] When obtaining frequency domain images, Fourier transform algorithm is used to convert data from time domain to frequency domain, and then the graph is drawn with frequency as the horizontal axis and amplitude as the vertical axis;

[0026] The following steps are involved in obtaining the Gram Corner Field GAF ​​image:

[0027] SS1: Normalization processing. Before normalization processing, the data that meets the time order is defined as the time series X = {x 1 ,x 2 ,…,x i ,…x n}, where n is the total time point, i is the time point, i∈[1,n];

[0028] The time series X = {x 1 ,x 2 ,…,x i ,…x n} is scaled to the range [-1,1] by:

[0029]

[0030] SS2: Polar coordinate transformation, using polar coordinate transformation to convert time series data into vectors, specifically:

[0031]

[0032] Where, t i is the timestamp, N is the constant factor for generating the regularized polar coordinate space, is the phase angle, r is the polar coordinate half pound, is the standardized and scaled X;

[0033] SS3: Calculate the Gram matrix and obtain the Gram matrix through the inner product form defined by the Gram angular field;

[0034] SS4: Image generation, converting the Gram matrix into a two-dimensional image.

[0035] As a preferred solution of the present application, at least the bridge structure health monitoring data in the training set is acceleration data.

[0036] As a preferred solution of the present application, in step S1, the specific method for acquiring the three-channel image includes: using the opencv image processing algorithm to map the time domain image, frequency domain image, and GAF ​​image corresponding to each part of the data to the three RGB channels respectively, so as to realize the merging and visualization processing of the image data.

[0037] Compared with the prior art, the advantage of the deep learning-based bridge structure health monitoring data anomaly detection method of the present application is that: the present application abandons the traditional over-reliance on data analysis in the time domain, but instead visualizes the data in the time domain, frequency domain and Gram angle field (GAF) respectively, and then superimposes them into a three-channel composite image through an image processing algorithm, and inputs the three-channel composite image into a VGG16 neural network model that incorporates a channel attention mechanism for processing. The VGG16 neural network model that incorporates a channel attention mechanism can dynamically adjust the weights of each channel according to image features, that is, extract channel features in a targeted manner, thereby improving the accuracy of abnormal data recognition. In addition, the neural network model that incorporates a channel attention mechanism of the present application is not affected by trends, seasonality, periodicity and random disturbances in traditional time series, and has strong feature extraction capabilities and good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the overall concept framework of the detection method provided in an embodiment of the present invention.

[0039] Figure 2 A flow chart of a detection method provided in an embodiment of the present invention.

[0040] Figure 3 The time domain graphs under various data features in the training set provided by the embodiment of the present invention.

[0041] Figure 4 Frequency domain spectra of various data features in the training set provided by the embodiment of the present invention.

[0042] Figure 5 The Gram angular field GASF map under various data features in the training set provided by the embodiment of the present invention.

[0043] Figure 6 A schematic diagram of the VGG16 network model structure incorporating a channel attention mechanism provided in an embodiment of the present invention.

[0044] Figure 7 A schematic diagram of the structure of the channel attention mechanism module provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described in detail below in conjunction with specific implementations and with reference to the accompanying drawings. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope of the present invention and its application.

[0046] like Figure 1 As described above, a schematic diagram of the overall idea framework of a bridge structure health monitoring data anomaly detection method based on deep learning provided in this embodiment, that is, the method in this embodiment first obtains the health monitoring data of the bridge to be monitored, and then divides it into several parts, and the data of each part is respectively subjected to image visualization processing in the time domain, frequency domain and Gram angle field GAF, and the time domain graph, frequency domain graph and GAF ​​graph are superimposed using an image processing algorithm to form a composite image containing three RGB color channels, and finally the three-channel composite image is input into a VGG16 neural network model that integrates a channel attention mechanism, and the model is used to identify and detect abnormal data and classify and output the abnormal data.

[0047] like Figure 2 As shown, it is a flowchart of the specific steps and methods provided in this embodiment, and the method comprises:

[0048] S1: Generate image dataset

[0049] The bridge structure health monitoring data is obtained and divided into several parts. At the same time, the monitoring data of each part is respectively visualized in the time domain, frequency domain and Gram angle field (GAF), and then the three images corresponding to each part are superimposed to form a three-channel composite image. It can be understood that the specific category of the monitoring data can be one or more data that can be collected by the bridge health monitoring system, such as acceleration data, wind speed data, strain data, etc. The present embodiment preferably uses acceleration data. When there are multiple types of data, each type of data is processed in sequence.

[0050] In this embodiment, the monitoring data can be separated according to a specified time length, for example, one segment is separated every hour, and it is ensured that the data in each segment is collected by the same sensor. Of course, in other embodiments of the present application, it can also be separated according to bytes or data length, etc., depending on the actual situation. It can be understood that the segmented data can be understood as window data without overlap between two adjacent windows.

[0051] S2: Selected training set

[0052] Obtain the structural health monitoring data of the bridge with known health status and perform the above step S1 on it first, then label the formed three-channel composite image according to the data features, and use the monitoring data as a training set; in this embodiment, the data features mainly include the health status and distribution trend of the data, wherein the health status of the data includes abnormal data and normal data in various situations, and the abnormal situations preferably include missing, sub-minimum, outlier, out-of-range oscillation and drift in this embodiment; it can be understood that in this embodiment, the specific category of the structural health monitoring data of the bridge as a training set can be one or more data that can be collected by the bridge health monitoring system, such as acceleration data, wind speed data, strain data, etc., and this embodiment is preferably acceleration data, that is, this embodiment is a process of identifying and detecting abnormal data of the data obtained by the acceleration sensor. Specifically, this embodiment selects the acceleration data collected by the structural health monitoring system of a large-span cable-stayed bridge as the data used in the sample of this patent, and the database is provided by the First National Structural Health Monitoring Competition Committee. The main span of the cable-stayed bridge is 1,088 meters long, the side span is 300 meters long, and the bridge tower is 306 meters high. It was completed in 2008 and has a total of 38 accelerometers installed on the bridge deck and the tower.

[0053] In this embodiment, in steps S1 and S2, the acquired bridge structure health monitoring data should satisfy the time sequence.

[0054] The acquired monitoring data that meets the time sequence are processed for image visualization in the time domain, frequency domain and Gram angle field (GAF). The specific processing methods include:

[0055] When obtaining a time domain image, the one-dimensional data in each separated segment is plotted on the time axis to obtain a time domain image; Figure 3 As shown, it is the time domain graph under various data features in the training set provided by this embodiment.

[0056] When obtaining frequency domain images, the Fourier transform algorithm is used to convert data from the time domain to the frequency domain, and then the frequency domain graph is drawn with frequency as the horizontal axis and amplitude as the vertical axis. The specific Fourier transform algorithm is:

[0057]

[0058]

[0059] In the formula, A K is the Fourier coefficient, which is a complex number, and its absolute value |A K | is called the amplitude spectrum; x n is the nth data of the time series X with a length of N; f k is the frequency.

[0060] like Figure 4 As shown, it is the frequency domain spectrum of various data features in the training set provided by this embodiment.

[0061] The following steps are involved in obtaining the Gram Corner Field GAF ​​image:

[0062] SS1: Normalization processing. Before normalization processing, the data that meets the time order is defined as the time series X = {x 1 ,x 2 ,…,x i ,…x n}, where n is the total time point, i is the time point, i∈[1,n];

[0063] The time series X = {x 1 ,x 2 ,…,x i ,…x n} is scaled to the range [-1,1] by:

[0064]

[0065] SS2: Polar coordinate transformation, using polar coordinate transformation to convert time series data into vectors, specifically:

[0066]

[0067] Where, t i is the timestamp, N is the constant factor for generating the regularized polar coordinate space, is the phase angle, r is the polar coordinate half pound, is the standardized and scaled X;

[0068] SS3: Calculate the Gram matrix. Since it is difficult to distinguish Gaussian noise from the inner product output data defined by the Gram matrix, this embodiment uses two unique inner product forms with penalty terms defined by the Gram angular field to obtain the Gram matrix to eliminate the influence of Gaussian noise.

[0069] The two inner product forms defined by the Gram angular field are:

[0070] <x i ,x j >=cos(φ i +φ j )

[0071] <x i ,x j >=sin(φ i -φ j )

[0072] The above two different inner product definition forms can respectively obtain two different Gram angle fields: GASF and GADF, as shown in the following formulas:

[0073]

[0074]

[0075] In this embodiment, the Gram angle field GASF is preferably used.

[0076] SS4: Image generation, converting the GASF angle matrix into a two-dimensional image, that is, generating the Gram angular field GASF image.

[0077] like Figure 5 As shown, the Gram angle field GASF map under various data features in the training set provided in this embodiment.

[0078] In this embodiment, each acquired image is 32 bits and has a resolution of 224×224, which can represent the image features of the data.

[0079] When acquiring a three-channel image, the opencv image processing algorithm is used to map the time domain image (atlas), frequency domain image, and GAF ​​image corresponding to each part of the data to the three RGB channels respectively, so as to realize the merging and visualization of the image data, and then obtain a three-channel composite image.

[0080] S3: Constructing a VGG16 network with channel attention mechanism as an anomaly detection model

[0081] The anomaly detection model includes a convolution module, a channel attention mechanism module, a fully connected layer and a Softmax classifier. The convolution module contains multiple groups, and the channel attention mechanism module is connected in series between two adjacent convolution modules, such as Figure 6 shown.

[0082] The convolution module consists of at least two convolution layers and one maximum pooling layer. The preferred convolution module in this embodiment includes a total of five groups, each of which consists of two convolution layers and one maximum pooling layer, wherein the convolution layers are all 3×3 filters with a stride of 1. In this embodiment, the number of channels of each convolution layer from front to back in the five groups of convolution modules are preferably 64, 128, 256, 512, and 512 respectively.

[0083] This embodiment can form a deep feature extraction and dimensionality reduction module by combining the convolution layer and the maximum pooling layer. After the synthetic image is input into the network, the convolution layer is responsible for extracting the features of each channel in the input synthetic image data, and then the maximum pooling layer performs dimensionality reduction sampling on the features extracted by the convolution layer. This can achieve the goal of retaining important feature information while reducing computational overhead and overfitting risks, thereby improving the performance of the neural network model.

[0084] The channel attention mechanism module can identify and divide the importance of each channel. During the identification process, different tasks select different channels for enhancement or suppression. Specifically, the image processed by the previous convolution module is input into the channel attention mechanism module. The channel attention mechanism module uses the image as input feature and obtains the features of each channel dimension and the importance of each channel. It then multiplies the image features processed by the previous convolution module to adjust the weight of each channel, so that the weight coefficient of the important channel becomes larger and the weight of the secondary channel becomes smaller. This achieves the extraction and identification of the main features and improves the accuracy of the model in identifying abnormal data.

[0085] The channel attention mechanism module includes a maximum pooling layer and an average pooling layer set in parallel, as well as a shared layer and an activation layer connected in series, such as Figure 7 As shown in the figure; the average pooling layer is used to aggregate the spatial information of the input features to effectively reduce the amount of data, while retaining the spatial structure information of the original features and enhancing the representativeness and stability of the features. It can be understood that the input feature is the image processed by the previous convolution module. The maximum pooling layer is used to select the input features, screen out the features with better classification distinction, and find clues to the target features; by performing average pooling and maximum pooling operations on the input features, the input features can be aggregated in spatial dimensions, and the feature description information obtained by the two pooling operations is input into the shared layer, and the compression of the different feature descriptions is realized in the shared layer, which reduces the parameter overhead and the complexity of the model. At the same time, it is also beneficial to reduce the fitting of the model. Finally, the different feature descriptions can be combined element by element to generate a channel-channel attention mechanism diagram through the activation layer; the channel-channel attention mechanism diagram can adjust the weights of each channel by multiplying it with the image processed by the previous convolution module, that is, the input feature, thereby realizing dynamic adjustment of the weights of different channels.

[0086] In this embodiment, the activation layer preferably uses a Sigmoid activation function, and the activation size is R C / r×1×1 , where r represents the reduction ratio.

[0087] In this embodiment, the shared layer is a multilayer perceptron plus a hidden layer. Specifically, the shared layer includes a global pooling layer, a fully connected layer, a RELU activation layer, a fully connected layer and a Sigmoid activation layer connected in series in sequence. The global pooling layer is used to perform pooling operations on the feature map, the fully connected layer is used to connect neurons, the RELU activation layer is used to introduce nonlinearity, and the Sigmoid activation layer is used to output probability values. The shared layer of this structure can reduce the dimension of features and extract important information, enhance the expression and generalization capabilities of the model, and at the same time can also reduce the fitting degree of the neural network model and improve the robustness.

[0088] The following is a channel attention mechanism diagram M obtained by the channel attention mechanism module provided in this embodiment: c The formula is:

[0089]

[0090] Where σ is the sigmoid activation function; W 0 ∈R C / r×C With W 1 ∈R C×C / r N is the multilayer perceptron weight, which is shared by two inputs, W 0 is the weight of the RELU function.

[0091] The maximum pooling layer is a 2×2 filter with a stride of 2.

[0092] The fully connected layer consists of three layers connected in sequence, and the number of channels of each fully connected layer is 4096, 4096, and 1000 respectively.

[0093] The Softmax classifier contains 1000 channels, which can convert the output of the last layer of the neural network, that is, the fully connected layer in this embodiment, into the probabilities of each category. In this embodiment, Softmax preferably outputs seven data in the end, and the sum of the seven data is 1, for example [0.1, 0.3, 0.6, 0, 0, 0, 0]. The seven data are the classification of the six abnormal data mentioned above (missing, sub-minimum value, outlier, out-of-range oscillation, drift, trend) and normal data.

[0094] S4: Training anomaly detection model

[0095] The synthetic images in the training set are input into the VGG16 network that integrates the channel attention mechanism for deep recognition and learning to achieve model training.

[0096] S5: Output the test results

[0097] Obtain the health monitoring data of the bridge to be monitored, and input the monitoring data into the trained anomaly detection model after executing the above S1, output the detection results through the Softmax classifier, and determine the type of anomaly based on the detection results, such as data missing or data drift.

[0098] In this embodiment, the health monitoring data of the bridge to be monitored is preferably acceleration data. The specific use principle of this embodiment is described below using acceleration data as the data to be monitored:

[0099] First, the acquired acceleration data is divided into several parts according to the duration. At the same time, the monitoring data of each part is visualized in the time domain, frequency domain and Gram angle field GAF ​​respectively. Then, the three images corresponding to each part are superimposed to form a three-channel composite image. Finally, the three-channel composite image is input into the network (or model) that integrates the channel attention mechanism and the neural network VGG16. In this model, the composite image is first subjected to feature extraction and dimensionality reduction processing through the first convolution module, and then enters the channel attention mechanism and uses the channel attention mechanism to obtain the dimension and importance of each channel in the image, and the obtained dimension and importance are multiplied with the features extracted by the convolution module. In this way, the weight of each channel in the image changes according to the importance of the feature, and the cycle is repeated until completion. It can be seen that after the time series data is visualized in the time domain, frequency domain, and GAF ​​respectively and superimposed to form a combined image, the main features of each channel in the input three-channel composite image can be weighted and reduced in dimension with the cooperation of the convolution module and the channel attention mechanism module. This reduces the parameter overhead on the one hand, and enhances the representativeness of the features on the other hand, so that the neural network can identify and extract the channel features in a targeted manner, thereby improving the accuracy and efficiency of the system abnormal data identification and detection. In addition, the neural network model incorporating the channel attention mechanism in this embodiment is not affected by the trend, seasonality, periodicity, and random disturbances in the traditional time series, and has a strong feature extraction capability and good robustness.

[0100] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several improvements can be made without departing from the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of the claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for detecting abnormality in bridge structure health monitoring data based on deep learning, characterized in that The following steps are involved: S1: Generate image dataset Obtain the bridge structure health monitoring data and divide the monitoring data into several parts. At the same time, perform image visualization processing on the monitoring data of each part in the time domain, frequency domain and Gram angle field (GAF), and then superimpose the three images corresponding to each part to form a three-channel composite image. S2: Selected training set Acquire the health monitoring data of the bridge structure with known health status and perform the above step S1 on it, then label the formed three-channel composite image according to the data features, and use the monitoring data as a training set; S3: Constructing a VGG16 network with channel attention mechanism as an anomaly detection model The anomaly detection model includes a convolution module, a channel attention mechanism module, a fully connected layer and a Softmax classifier. The convolution module includes multiple groups. The channel attention mechanism module is connected in series between two adjacent convolution modules. The convolution module consists of at least two convolution layers and one maximum pooling layer. The convolution layer and the maximum pooling layer are combined to form a deep feature extraction and dimensionality reduction module. S4: Training anomaly detection model The synthetic images in the training set are input into the VGG16 network with the channel attention mechanism for deep recognition and learning to achieve model training; S5: Output the test results Obtain the health monitoring data of the bridge to be monitored, and input the monitoring data into the trained anomaly detection model after executing the above S1, and output the detection results through the Softmax classifier; In step S3, the channel attention mechanism module includes a maximum pooling layer and an average pooling layer arranged in parallel, and a shared layer and an activation layer connected in series therewith; different feature descriptions can be obtained respectively by performing average pooling and maximum pooling operations on the input features, and the shared layer can compress the different feature descriptions, and finally the different feature descriptions can be combined element by element to generate a channel attention mechanism graph through the activation layer; The shared layer includes a global pooling layer, a fully connected layer, a RELU activation layer, a fully connected layer and a Sigmoid activation layer connected in series. The shared layer of this structure can reduce the dimension of features and extract important information; In steps S1 and S2, the acquired bridge structure health monitoring data should satisfy the time sequence; When obtaining a time domain image, the one-dimensional data in each separated segment is plotted on the time axis to obtain a frequency domain image; When obtaining frequency domain images, Fourier transform algorithm is used to convert data from time domain to frequency domain, and then the graph is drawn with frequency as the horizontal axis and amplitude as the vertical axis; The following steps are involved in obtaining the Gram Corner Field GAF ​​image: SS1: Normalization processing. Before normalization processing, the data that meets the time order is defined as a time series. , where For the total time point, For time point, [1,n]; The time series Scaling to the range [-1, 1] is done by: ； SS2: Polar coordinate transformation, using polar coordinate transformation to convert time series data into vectors, specifically: ， In the formula, is the timestamp, is the constant factor for the regularized polar coordinate generation space, is the phase angle, r is the polar coordinate half pound, After normalization and scaling ; SS3: Calculate the Gram matrix and obtain the Gram matrix through the inner product form defined by the Gram angular field; SS4: Image generation, converting the Gram matrix into a two-dimensional image.

2. The method for detecting abnormality in bridge structure health monitoring data based on deep learning according to claim 1, characterized in that: The activation layer adopts the Sigmoid activation function.

3. The method for detecting abnormality in bridge structure health monitoring data based on deep learning according to claim 1, characterized in that: In step 3, the filter of the convolutional layer is , with a step size of 1; the filter of the maximum pooling layer is , the step size is 2; the fully connected layer includes three layers connected in sequence, and the number of channels of each fully connected layer is 4096, 4096 and 1000 respectively.

4. The method for detecting abnormality in bridge structure health monitoring data based on deep learning according to claim 1, characterized in that: In step S2, the data features include health status and distribution trends, and the health status includes abnormal data and normal data in various situations.

5. The method for detecting abnormality in bridge structure health monitoring data based on deep learning according to claim 1, characterized in that: At least the bridge structure health monitoring data in the training set is acceleration data.

6. The method for detecting abnormality in bridge structure health monitoring data based on deep learning according to claim 1, characterized in that: In step S1, the specific method for obtaining the three-channel composite image includes: using the opencv image processing algorithm to map the time domain image, frequency domain image, and GAF ​​image corresponding to each part of the data to the three channels of R, G, and B respectively, so as to realize the merging and visualization processing of the image data.

Images