A method and device for diagnosing abnormal structural states considering the influence of complex environments

Abstract

A method and device for diagnosing abnormal structural states considering complex environmental impacts. The method includes: A. Collecting structural reference state data under time-varying environmental impacts and constructing a training data set; B. Designing and training a convolutional autoencoder to convert high-dimensional original data into low-dimensional latent features; C. Designing and training a conditional diffusion probability model and introducing an attention mechanism to fuse and analyze multi-source data to establish a conditional probability model for reference state monitoring data under different environmental impacts; D. Calculating reference state statistical indicators and using the reference state statistical indicators to construct an outlier identification model SVDD; E. Inputting a test state data set to diagnose whether the test state is an abnormal state. The present invention only requires a structural reference state data set and does not need to rely on structural physical models such as finite elements, and realizes high-precision diagnosis of abnormal structural states under time-varying environmental impacts.

Description

Technical Field

[0001] The present invention relates to a method and device for diagnosing abnormal structural states considering the influence of complex environments. The conditional probability deep learning model ATT-CLDPM is used to identify the benchmark state monitoring data patterns under different environmental influences, and statistical indicators are constructed to measure the deviation of the test state from the benchmark state data pattern. The outlier identification model SVDD is used to identify the statistical indicators deviating from the benchmark state, effectively diagnosing the abnormal structural states under time-varying environmental influences, belonging to the field of civil structure health monitoring. Background Art

[0002] Civil engineering structures are affected by multiple factors such as loads, natural disasters, material aging, and environmental effects for a long time, and gradually show degradation of structural performance, leading to potential safety hazards. In order to effectively monitor and evaluate the health status of structures, data-driven structural anomaly diagnosis methods have received extensive attention. Such methods can accurately identify abnormal structural states by extracting and analyzing the statistical patterns of monitoring data, thus providing support for early warning of structural damage and failure.

[0003] Deep learning methods are commonly used data-driven structural anomaly diagnosis methods, which are divided into supervised learning methods and unsupervised learning methods. Such methods require data of normal and abnormal structural states as training samples, and extract signal features of normal and abnormal structural states therefrom to classify and identify the test state. However, structural anomalies usually do not exist in the early stage of the structure and are quite rare during the long service life of the structure. Therefore, it may be impossible to obtain monitoring data related to abnormal structural states. To overcome this difficulty, unsupervised learning methods that do not require labeled data of abnormal structural states are applied to anomaly diagnosis.

[0004] In recent years, unsupervised learning methods have gradually become a research hotspot. These methods do not rely on labeled data of abnormal states, but learn the data patterns in the benchmark state, and diagnose abnormal behaviors by obtaining the data pattern of the test state and the data pattern of the benchmark state. Therefore, they have important applications in structural health monitoring. However, most existing methods fail to effectively consider the changes in environmental conditions, which play a crucial role in actual monitoring. Changes in environmental conditions, such as temperature, humidity, wind speed, etc., often affect the distribution of structural responses, and then lead to fluctuations in monitoring data. If these environmental factors are ignored, the anomaly diagnosis method may misjudge the data pattern changes caused by environmental changes and cannot accurately diagnose the abnormal structural states. To overcome this difficulty, unsupervised learning methods that only require normal state data are applied to anomaly diagnosis. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention proposes a method and device for diagnosing abnormal structural states considering the influence of complex environments, realizing high-precision diagnosis of abnormal structural states under the influence of time-varying environments.

[0006] The first aspect of the present invention relates to a method for diagnosing abnormal structural states considering the influence of complex environments, including the following steps:

[0007] A. Collect structural reference state data under the influence of time-varying environments, and perform preprocessing to construct a data set for deep learning model training;

[0008] B. Design and train a Convolutional Auto-Encoder (CAE), convert high-dimensional raw data to low-dimensional latent features, use convolutional kernels and deconvolutional kernels to extract key temporal features and correlation features of multi-measurement point signals of the structure, and select the optimal latent variable dimension and CAE hidden layer dimension according to the minimum convergence value of the loss function to optimize the CAE network, effectively using a small number of key features to represent the raw data;

[0009] C. Introduce a U-net with an attention mechanism (Attention U-net, ATT-Unet) network, design and train an Attention-guided Conditional Latent Diffusion Probabilistic Model (ATT-CLDPM), fuse and analyze multi-source data, establish a conditional probability model for reference state monitoring data under different environmental influences, and select the optimal ATT-Unet hidden layer dimension according to the minimum convergence value of the loss function to realize data pattern representation under different environmental conditions;

[0010] D. Use the statistical index data point set in the reference state to construct a Support Vector Data Description (SVDD) model, and construct a reference state domain representing the reference state;

[0011] E. Input the test state data set, and judge whether the test state is an abnormal state according to the relationship between the statistical index points and the reference state domain.

[0012] Further, step A specifically includes:

[0013] A1. Use sensing technology to collect structural reference state data for a long time, including multi-measurement point response data reflecting the mechanical behavior of the structure and data on environmental conditions;

[0014] A2. Preprocess the structural reference state data, and use the maximum-minimum normalization method to scale the data to the range of [0,1], so as to enhance the generalization ability of the deep learning model;

[0015] A3. Detrend the data collected by the sensor with drift faults to ensure the validity of the data;

[0016] A4. Slice the preprocessed data using a time window with a fixed step size, and construct a structural reference state dataset using the sliced samples, which is divided into a training set and a validation set.

[0017] Further, step B specifically includes:

[0018] B1. The CAE neural network is divided into two parts: an encoder and a decoder;

[0019] B2. The encoder gradually reduces the dimension of the original data through multiple convolutional layers and converts it into low-dimensional hidden features;

[0020] B3. The decoder increases the dimension of the low-dimensional hidden features through multiple transposed convolutional layers and reconstructs them into high-dimensional original data;

[0021] B4. When designing the CAE network framework, take the latent variable dimension and the hidden layer dimension to optimize the CAE network as important design parameters, so as to adjust the network framework according to the loss function later;

[0022] B5. Take the mean square error between the input original data and the output reconstructed data as the loss function of the CAE. The training objective of the CAE can be expressed as:

[0023]

[0024] Among them, represents the CAE loss function, x is the input original data, is the output reconstructed data, ε represent the decoder and the encoder respectively;

[0025] B6. For CAE networks with different latent variable dimensions c latent during training, the number of iterations uses the optimal number of iterations, and select the optimal latent variable dimension c latent corresponding to the minimum convergence value of the validation set damage function;

[0026] B7. For CAE networks with different hidden layer dimensions c CAE during training, the number of iterations uses the optimal number of iterations, and select the optimal hidden layer dimension c CAE corresponding to the minimum convergence value of the validation set damage function;

[0027] B8. Design the optimal CAE network using the optimal latent variable dimension c latent and the optimal hidden layer dimension c CAE and train the optimal CAE network using the training dataset to convert the original data into low-dimensional hidden features.

[0028] Furthermore, step C specifically includes:

[0029] C1. Establish an ATT-CLDPM to identify the monitoring data patterns under different environmental conditions. This model is mainly divided into a forward process and a reverse process;

[0030] C2. In the forward process, according to the Markov chain hypothesis, in the Markov chain at step T, zero-mean Gaussian noise with variance β is gradually added to the latent features, and the original latent features are transformed into standard Gaussian noise. The noise-added latent feature z at step t t can be expressed as t

[0031]

[0032] where z0 is the original latent feature, z0 = ε(x), is a hyperparameter and can be expressed as

[0033]

[0034] C3. The reverse process is to predict the noise component through ATT-Unet, and use the environmental condition variable ec and the initial latent feature z0 as conditional constraints to gradually remove the noise component in the standard Gaussian noise and gradually restore it to the initial latent feature. This process can be expressed as

[0035]

[0036] where α t = 1 - β t , represents the noise predicted by ATT-Unet;

[0037] C4. When designing the ATT-Unet network framework, introduce the multi-head attention mechanism and attention gate to perform fusion analysis on data in different dimensions;

[0038] C5. The input of the multi-head attention includes the environmental condition variable and the predicted noisy feature of the structural response at step t, forming a fusion feature. The output of the multi-head attention feature is;

[0039]

[0040] where MA represents the multi-head attention mechanism, W A , is a learnable linear transformation matrix, d represents the dimension of the environmental condition variable, and softmax(·) is the softmax activation function, softmax(x i ) = exp(x i ​) / ∑ i exp(x i );

[0041] C6. Resize the output of the multi-head attention to form the fused feature z ec , and concatenate it with the predicted noisy feature at the t-th step conditional feature z0, and input it into the subsequent steps;

[0042] C7. The subsequent steps of the ATT-Unet are to first perform convolution on the input feature to obtain the deep feature, and then use upsampling to extract the predicted noise parameters from the deep feature. In order to retain both the deep feature and the shallow feature, a skip connection with an attention mechanism is added between different layers, that is, the feature in the previous convolutional layer and the upsampled feature pass through the attention gate, which can be expressed as

[0043]

[0044] where, is the output of the attention gate, α i is the attention weight, ψ, W v , W g are learnable linear transformation matrices;

[0045] C8. When designing the ATT-Unet network framework, the hidden layer dimension is used as an important design parameter for subsequent adjustment of the network framework according to the loss function;

[0046] C9. Take the mean square error of the noise predicted by the ATT-Unet and the actual noise ε t added at the t-th step as the loss function of the ATT-Unet, and the training objective can be expressed as:

[0047]

[0048] where, represents the loss function of the ATT-Unet;

[0049] C10. For the ATT-Unet network with different hidden layer dimensions c ATT-Unet , during training, the number of iterations adopts the optimal number of iterations, and select the optimal hidden layer dimension c ATT-Unet corresponding to the minimum convergence value of the validation set damage function.

[0050] Furthermore, step D specifically includes:

[0051] D1. Obtain the mean square error MSE and the original-reconstructed signal ratio ORSR of the anomaly statistical indicators according to the input and output data of the ATT-CLDPM deep learning. The two anomaly statistical indicators can be expressed as:

[0052]

[0053] where x i is the i-th input sample of ATT-CLDPM, is the i-th output sample of ATT-CLDPM, and m is the total number of samples;

[0054] D2. Construct an outlier statistical index point set S = {s i | s i = (MSE i , ORSR i ). Substitute it into the SVDD model to construct a hyper-sphere plane with the minimum volume as the boundary of the reference state domain. The reference state domain can enclose all points corresponding to the reference states and exclude the points corresponding to the abnormal states. The construction of the hyper-sphere plane with the minimum volume can be expressed as:

[0055]

[0056] where R represents the radius of the hyper-sphere, C represents the penalty coefficient, ξ i represents the i-th slack variable, l represents the number of slack variables, c represents the coordinates of the center of the hyper-sphere, and Φ(s i ) represents the eigenvector obtained after mapping s i to the high-dimensional space;

[0057] D3. Construct the SVDD model by using different combinations of the penalty coefficient and the Gaussian kernel coefficient, and select the appropriate boundary of the reference state domain.

[0058] Furthermore, step E specifically includes:

[0059] E1. Substitute the data set collected under the structural test state into the trained ATT-CLDPM to obtain two outlier statistical indexes MSE and ORSR;

[0060] E2. Judge whether the outlier statistical index point corresponding to the test data set is inside the decision boundary through the decision function. The decision function can be expressed as:

[0061] f(s) = R 2 - ||Φ(s) - c|| (9)

[0062] where S represents the outlier statistical index point in the test state. If f(s) is less than 0, the index point is outside the decision boundary, indicating that the structure is in an abnormal state at this time.

[0063] The second aspect of the present invention relates to a structural abnormal state diagnosis device considering the influence of complex environments, including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement a structural abnormal state diagnosis method considering the influence of complex environments according to the present invention.

[0064] The third aspect of the present invention relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements a structural abnormal state diagnosis method considering the influence of complex environments according to the present invention.

[0065] The innovations of the present invention are as follows: First, based on the Markov chain hypothesis, the conditional probability distribution of monitoring data is modeled using CLDPM, which helps to effectively identify data patterns under different environmental conditions. By introducing latent representations, the original data is mapped to a low-dimensional feature space, thereby improving the accuracy and efficiency of abnormal diagnosis. Second, to comprehensively analyze structural responses and environmental variables, an attention mechanism is introduced to form ATT-Unet and integrated into CLDPM to achieve the fusion analysis of multi-source data such as structural responses and environments. Third, two abnormally sensitive statistics are selected to measure the difference between the distribution of monitoring data and the distribution modeled by ATT-CLDPM, and outliers in the statistical data are identified through the SVDD method, thereby realizing the diagnosis of structural abnormal states under the influence of complex environments.

[0066] The working principle of the present invention is: using the probability-based deep learning model CLDPM for the probability characterization of data distributions can better capture the variability of complex data patterns. By integrating conditional probability constraints, the formed conditional probability model can establish the conditional dependence relationship between structural response data and environmental variables, thereby accurately diagnosing structural abnormalities in the complex and changeable structural service environment.

[0067] The advantages of the present invention are: it can perform fusion analysis on structural response data and environmental data under complex environmental conditions, establish a conditional probability distribution model to identify the monitoring data representing the structural abnormal state, without relying on accurate structural modeling and without obtaining various structural abnormal state data, having the advantage of simple operation, and can realize high-precision and high-efficiency diagnosis of structural abnormal states. Brief Description of the Drawings

[0068] Figure 1 is a flowchart of the method of the present invention;

[0069] Figure 2 is a schematic diagram of the monitored structure size of the present invention;

[0070] Figure 3 is a schematic diagram of the ATT-LCDPM framework of the present invention;

[0071] Figure 4 It is the statistical index of the normal state of the monitoring structure of the present invention and the result diagram of the reference state domain;

[0072] Figure 5 It is the schematic diagram of the abnormal state of the monitoring structure of the present invention;

[0073] Figure 6 The result diagram of the state diagnosis of the monitoring structure of the present invention.

[0074] Figure 7 It is the schematic diagram of the device of the present invention. Detailed implementation manners

[0075] The following further elaborates on the present invention in conjunction with the attached drawings.

[0076] Embodiment 1

[0077] As Figure 1 shown, this embodiment relates to a method for diagnosing the abnormal state of a structure considering the influence of complex environments, specifically including the following steps:

[0078] A. Collect the reference state data of the structure under the influence of time-varying environmental conditions, and perform preprocessing to construct a data set for deep learning model training; specifically including:

[0079] A1. Monitor the truss bridge structure. As Figure 2 shown, arrange 8 acceleration sensors on the lower chord beam of the truss bridge, and perform multiple samplings on the acceleration of the truss bridge structure under the influence of time-varying environmental conditions at a sampling frequency of 256 Hz. The duration of a single sampling is 32 s, and record the environmental conditions at each sampling. Since this structure is sensitive to temperature, the temperature value reflecting the environmental conditions is synchronously recorded each time during sampling, and the temperature change range is from 10 °C to 30 °C;

[0080] A2. Preprocess the collected reference state data, and use the maximum-minimum normalization method to scale the data to the range of [0,1], so as to enhance the generalization ability of the deep learning model

[0081] A3. Perform detrending processing on the data collected by the sensors with drift faults to ensure the validity of the data;

[0082] A4. Use a time window with a fixed step size of 1024 to slice the preprocessed data, and use the sliced samples to construct a reference state data set of the structure, which is divided into a training set and a validation set according to a ratio of 19:1.

[0083] B. Design and train a Convolutional Auto-Encoder (CAE) to transform high-dimensional raw data into low-dimensional latent features, extract the key temporal features and correlation features of the structural multi-measurement point signals using convolutional and deconvolutional kernels, and select the optimal latent variable dimension and CAE hidden layer dimension based on the minimum convergence value of the loss function to optimize the CAE network, effectively representing the raw data using a small number of key features; specifically including:

[0084] B1. The CAE neural network is divided into two parts: an encoder and a decoder;

[0085] B2. The encoder gradually reduces the dimension of the raw data through multiple convolutional layers and transforms it into low-dimensional latent features. The encoder parameters are shown in Table 1:

[0086] Table 1 Encoder design parameters

[0087]

[0088] Note: Conv = Convolution; BN = Batch Normalization; LReLU = LeakyReLU activation function;

[0089] B3. The decoder upsamples the low-dimensional latent features through multiple deconvolutional layers and reconstructs them into high-dimensional raw data. The decoder parameters are shown in Table 2:

[0090] Table 2 Decoder design parameters

[0091]

[0092] Note: DeConv = Transposed Convolution;

[0093] B4. When designing the CAE network framework, take the latent variable dimension and hidden layer dimension to optimize the CAE network as important design parameters for subsequent adjustment of the network framework according to the loss function;

[0094] B5. Use the mean square error between the input raw data and the output reconstructed data as the loss function of the CAE, and minimize the loss function through the Adam algorithm for neural network optimization training to achieve network training;

[0095] B6. For CAE networks with different latent variable dimensions c latent during training, use the optimal number of iterations, and select the optimal latent variable dimension c corresponding to the minimum convergence value of the validation set loss function latent ;

[0096] B7. For CAE networks with different hidden layer dimensions c CAEFor the CAE network, the optimal number of iterations is adopted during training, and the optimal hidden layer dimension c corresponding to the minimum convergence value of the validation set damage function is selected. CAE ;

[0097] B8. Utilize the optimal latent variable dimension c latent and the optimal hidden layer dimension c CAE to design the optimal CAE network, and use the training dataset to train the optimal CAE network to convert the original data into low-dimensional latent features.

[0098] C. Introduce a U-net with an attention mechanism (Attention U-net, ATT-Unet) network, design and train an attention-guided conditional latent diffusion probabilistic model (Attention-guided Conditional Latent Diffusion Probabilistic Model, ATT-CLDPM), fuse and analyze multi-source data, establish a conditional probability model for the benchmark state monitoring data under different environmental impacts, and select the optimal ATT-Unet hidden layer dimension according to the minimum convergence value of the loss function to achieve data pattern characterization under different environmental conditions;

[0099] C1. Establish an ATT-CLDPM to identify the monitoring data patterns under different environmental conditions. This model is mainly divided into a forward process and a reverse process;

[0100] C2. In the forward process, according to the Markov chain hypothesis, in the Markov chain of T steps, zero-mean Gaussian noise with a variance of β t is gradually added to the latent features to convert the original latent features into standard Gaussian noise;

[0101] C3. The reverse process is to predict the noise component through ATT-Unet, and use the environmental condition variable ec and the initial latent feature z0 as conditional constraints to gradually remove the noise component in the standard Gaussian noise and gradually restore it to the initial latent feature;

[0102] C4. When designing the ATT-Unet network framework, introduce a multi-head attention mechanism and attention gates to fuse and analyze data of different dimensions. The ATT-Unet network parameters are shown in Table 3:

[0103] Table 3 Decoder design parameters

[0104]

[0105]

[0106]

[0107]

[0108] Note: Concat = concatenation; ATT-Gate = attention gate;

[0109] C5. The input of the multi-head attention includes environmental condition variables and the noisy features of the predicted structural response at the t-th step, forming fused features;

[0110] C6. After resizing the output of the multi-head attention, fused feature z is formed ec , and it is concatenated with the noisy features condition feature z0 at the t-th step of prediction and input into subsequent steps;

[0111] C7. The subsequent steps of the ATT-Unet are to first perform convolution on the input features to obtain deep features, and then use upsampling to extract the predicted noise parameters from the deep features. In order to retain both deep features and shallow features simultaneously, skip connections with an attention mechanism are added between different layers, that is, the features in the previous convolutional layer and the upsampled features are passed through the attention gate;

[0112] C8. When designing the ATT-Unet network framework, the hidden layer dimension is used as an important design parameter so as to adjust the network framework according to the loss function subsequently;

[0113] C9. The mean square error between the noise predicted by the ATT-Unet and the actually added noise ε t at the t-th step is used as the loss function of the ATT-Unet, and the loss function is minimized through the Adam algorithm for neural network optimization training to achieve network training;

[0114] C10. For the ATT-Unet network with different hidden layer dimensions c ATT-Unet , the optimal number of iterations is adopted during training, and the optimal hidden layer dimension c corresponding to the minimum convergence value of the validation set damage function is selected ATT-Unet .

[0115] D. A support vector data description model (SVDD) is constructed using the statistical index data points set in the reference state to construct a reference state domain representing the reference state; specifically including:

[0116] D1. The mean square error MSE of the abnormal statistical index and the original-reconstructed signal ratio ORSR are obtained according to the input and output data of the ATT-CLDPM deep learning;

[0117] D2. An abnormal statistical index point set is constructed and substituted into the SVDD model to construct a hyper-sphere plane with the smallest volume as the boundary of the reference state domain. This reference state domain can enclose all the points corresponding to the reference state and can exclude the points corresponding to the abnormal state;

[0118] D3. Construct an SVDD model by using different combinations of penalty coefficients and Gaussian kernel coefficients, and select an appropriate decision boundary for abnormal states.

[0119] E. Input the test state data set, and determine whether the test state is an abnormal state according to the relationship between the statistical index points and the benchmark state domain; specifically including:

[0120] E1. The test state includes 4 test states: In test state 1, no abnormality appears in the structure, and the temperature change range is from 10°C to 30°C; in test state 2, no abnormality appears in the structure, and the temperature change range is from 5°C to 35°C; in test state 3, the cross-section of a single lower chord member of the structure is reduced by 20%, and the temperature change range is from 5°C to 35°C; in test state 4, the cross-sections of two lower chord members of the structure are reduced by 20%, and the temperature change range is from 5°C to 35°C. Test state 1 and test state 2 are normal states, and test state 3 and test state 4 are abnormal states. The members with cross-section reduction are Figure 5 marked in light color; Substitute the data set collected under the structural test state into the trained ATT-CLDPM to obtain two abnormal statistical indicators MSE and ORSR;

[0121] E2. Determine whether the abnormal statistical index points corresponding to the test data set are inside the decision boundary through the decision function, and the abnormal state diagnosis results are presented in Figure 6 and the diagnosis accuracy rates are sorted out in Table 4.

[0122] Table 4 Structural state diagnosis accuracy rate

[0123]

[0124] Example 2

[0125] As Figure 7 , this example relates to a structural abnormal state diagnosis device considering the influence of complex environments, including a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the structural abnormal state diagnosis method considering the influence of complex environments in Example 1.

[0126] Example 3

[0127] This example relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the structural abnormal state diagnosis method considering the influence of complex environments in Example 1.

[0128] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for diagnosing the abnormal state of a structure considering the influence of complex environments, comprising the following steps: A. Collecting the structural reference state data under time-varying environmental influences, performing preprocessing, and constructing a data set for deep learning model training; B. Designing and training a Convolutional Auto-Encoder (CAE), converting high-dimensional original data into low-dimensional latent features, extracting the key time-series features and correlation features of the multi-measurement point signals of the structure by using convolutional kernels and deconvolutional kernels, and selecting the optimal latent variable dimension and CAE hidden layer dimension according to the minimum convergence value of the loss function to optimize the CAE network, effectively using a small number of key features to represent the original data; C. Introducing a U-net with an attention mechanism (Attention U-net, ATT-Unet) network, designing and training an attention-guided conditional latent diffusion probabilistic model (Attention-guided Conditional Latent Diffusion Probabilistic Model, ATT-CLDPM), fusing and analyzing multi-source data, establishing a conditional probability model for the reference state monitoring data under different environmental influences, and selecting the optimal ATT-Unet hidden layer dimension according to the minimum convergence value of the loss function to realize data pattern representation under different environmental conditions; specifically including: C1. Establishing an ATT-CLDPM to identify the monitoring data patterns under different environmental conditions, and this model is mainly divided into a forward process and a reverse process; C2. In the forward process, according to the Markov chain hypothesis, in the Markov chain at step T, zero-mean Gaussian noise with variance β is gradually added to the latent features, transforming the original latent features into standard Gaussian noise. The noise-added latent feature z at step t t can be expressed as t as follows where z0 is the original hidden feature z0 = E(x), and the hyperparameter can be expressed as C3. The reverse process is to predict the noise component through ATT-Unet, and using the environmental condition variable ec and the initial latent feature z0 as conditional constraints, gradually removing the noise component in the standard Gaussian noise and gradually restoring it to the initial latent feature, and this process can be expressed as where α t = 1 - β t , denotes the noise predicted by ATT-Unet; C4. When designing the ATT-Unet network framework, introducing a multi-head attention mechanism and attention gates to fuse and analyze data of different dimensions; C5. The input of the multi-head attention includes the environmental condition variable and the noisy feature of the predicted structural response at the t-th step, forming a fused feature, and the output of the multi-head attention feature is; Among them, MA represents the multi-head attention mechanism, W A , is a learnable linear transformation matrix, d represents the dimension of the environmental condition variable, and softmax(·) is the softmax activation function softmax(x i ) = exp(x i ) / ∑ i exp(x i ); C6. Resize the output of the multi-head attention to form the fused feature z ec , and concatenate it with the predicted noisy feature at the t-th step and the conditional feature z0, and input them into the subsequent steps; C7. The subsequent steps of ATT-Unet are to first perform convolution on the input features to obtain deep features, and then use upsampling to extract the predicted noise parameters from the deep features. In order to retain both deep features and shallow features, add skip connections with an attention mechanism between different layers, that is, passing the features in the previous convolutional layer and the upsampled features through an attention gate, which can be expressed as Among them, is the output of the attention gate, and α i is the attention weight, and ψ, W v , W g are learnable linear transformation matrices; C8. When designing the ATT-Unet network framework, taking the hidden layer dimension as an important design parameter for subsequent adjustment of the network framework according to the loss function; C9. Take the mean square error between the noise predicted by ATT-Unet and the actual noise ε added at the t-th step t as the loss function of ATT-Unet, and the training objective can be expressed as: min L ATT-Unet = ||ε t - ε θ (z t , t, z0, ec)|| (6) Among them, L ATT-Unet represents the loss function of ATT-Unet; C10. For the ATT-Unet network with different hidden layer dimensions c ATT-Unet during training, the number of iterations is the optimal number of iterations, and the optimal hidden layer dimension c corresponding to the minimum convergence value of the validation set damage function is selected ATT-Unet ; D. Using the statistical index data point set in the reference state to construct a Support Vector Data Description (SVDD) model, and constructing a reference state domain representing the reference state; E. Inputting the test state data set, and judging whether the test state is an abnormal state according to the relationship between the statistical index points and the reference state domain.

2. The structural abnormal state diagnosis method considering complex environmental impacts according to claim 1, characterized in that, Step A specifically includes: A1. Long-term collect structural reference state data using sensing technology, including multi-measurement point response data reflecting the structural mechanical behavior and data on environmental conditions; A2. Preprocess the structural reference state data and scale the data to the range of [0, 1] using the maximum-minimum normalization method to enhance the generalization ability of the deep learning model; A3. Detrend the data collected by sensors with drift faults to ensure the validity of the data; A4. Slice the preprocessed data using a time window with a fixed step size, and use the sliced samples to construct a structural reference state data set, which is divided into a training set and a validation set.

3. The structural abnormal state diagnosis method considering complex environmental impacts according to claim 1, characterized in that Step B specifically includes: B1. The CAE neural network is divided into two parts: an encoder and a decoder; B2. The encoder gradually reduces the dimension of the original data through multiple convolutional layers and converts it into low-dimensional hidden features; B3. The decoder increases the dimension of the low-dimensional hidden features through multiple deconvolutional layers and reconstructs them into high-dimensional original data; B4. When designing the CAE network framework, take the latent variable dimension and the hidden layer dimension to optimize the CAE network as important design parameters so as to adjust the network framework according to the loss function subsequently; B5. Take the mean square error between the input original data and the output reconstructed data as the loss function of the CAE. The training objective of the CAE can be expressed as: Among them, L CAE represents the CAE loss function, x is the original input data, is the reconstructed output data, and D and E represent the decoder and encoder respectively; B6. For CAE networks with different latent variable dimensions c latent During training, the number of iterations is set to the optimal number of iterations, and the optimal latent variable dimension c corresponding to the minimum convergence value of the validation set damage function is selected latent ; B7. For CAE networks with different hidden layer dimensions c CAE During training, the optimal number of iterations is adopted, and the optimal hidden layer dimension c corresponding to the minimum convergence value of the validation set damage function is selected CAE ; B8. Utilize the optimal latent variable dimension c latent and the optimal hidden layer dimension c CAE Design the optimal CAE network, and use the training data set to train the optimal CAE network to convert the original data into low-dimensional latent features.

4. A method for diagnosing structural abnormal states considering complex environmental impacts as described in claim 1, characterized in that, Step D specifically includes: D1. Obtain the anomaly statistical metrics mean square error MSE and original-reconstructed signal ratio ORSR according to the input and output data of the ATT-CLDPM deep learning. The two anomaly statistical metrics can be expressed as: where x i is the i-th input sample of ATT-CLDPM, is the i-th output sample of ATT-CLDPM, and m is the total number of samples; D2. Construct the abnormal statistical index point set S = {s i | s i = (MSE i , ORSR i )}, substitute it into the SVDD model to construct a hyper-sphere plane with the smallest volume as the boundary of the reference state domain. This reference state domain can enclose all the points corresponding to the reference states and exclude the points corresponding to the abnormal states. The construction of the hyper-sphere plane with the smallest volume can be expressed as: where R represents the radius of the hypersphere, C represents the penalty coefficient, and ξ i represents the i-th slack variable, l represents the number of slack variables, c represents the coordinates of the hypersphere center, and Φ(s i ) represents the eigenvector obtained after mapping s i to a high-dimensional space; D3. Construct an SVDD model using different combinations of penalty coefficients and Gaussian kernel coefficients and select an appropriate boundary of the reference state domain.

5. The structural abnormal state diagnosis method considering the influence of complex environment according to claim 1, characterized in that, Step E specifically includes: E1. Substitute the data set collected under the structural test state into the trained ATT-CLDPM to obtain two anomaly statistical metrics MSE and ORSR; E2. Judge whether the anomaly statistical metric point corresponding to the test data set is inside the decision boundary through the decision function. The decision function can be expressed as: f(s) = R 2 -||Φ(s) - c|| (9) Where S represents the anomaly statistical metric point in the test state. If f(s) is less than 0, the metric point is outside the decision boundary, indicating that the structure is in an abnormal state at this time.

6. A structural abnormal state diagnosis device considering the influence of complex environments, characterized in that, Including a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement a structural abnormal state diagnosis method according to any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, A program is stored thereon. When the program is executed by a processor, it implements a structural abnormal state diagnosis method according to any one of claims 1-5.

Images